From 4e03c1b10669d8ae6ca54cb595d49d654e926a89 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=B0=8F=E5=94=AF=20A06?= <xiaowei-a06@local>
Date: Thu, 10 Sep 2026 18:14:39 +0800
Subject: [PATCH] =?UTF-8?q?skill+config:=20agnes=202.0=E2=86=922.5=20?=
 =?UTF-8?q?=E6=B8=85=E7=90=86=20+=202B=2044K=C3=972=20=E5=8F=8C=E6=A7=BD?=
 =?UTF-8?q?=E5=90=8C=E6=AD=A5?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- config.yaml: llama-local context_length 90112→45056 (44K槽); 9处 agnes-2.0-flash→2.5
- 11 skill bump version: 现行引用(配置/路由/代码默认值)改 2.5; 历史审计记录保留
- SOUL: 路由表修正 (npc=网上免费deepseek·sensenova, dsh=本地llama-local)
- session-to-zhiyi.py: MAX_INPUT_CHARS 70000→32000 (44K槽适配)
- 注: skills/kanban-router|kanban-routing 被 .gitignore 'kanban*' 误伤未纳入追踪(既有问题)
---
 SOUL.md                                       | 16 ++++++++++----
 config.yaml                                   | 21 +++++++++----------
 skills/chinese-platform-extraction/SKILL.md   |  5 +++--
 .../references/cron-script-limitations.md     |  2 +-
 skills/creative/agnes-ai/SKILL.md             | 11 +++++-----
 skills/creative/agnes-ai/references/agnes.py  |  2 +-
 .../agnes-ai/references/agnes_subprocess.py   |  2 +-
 skills/cron-ops/SKILL.md                      |  7 ++++---
 skills/devops/hermes-debug/SKILL.md           |  6 +++---
 skills/devops/llm-gateway-ops/SKILL.md        | 13 ++++++------
 skills/devops/provider-tiering/SKILL.md       | 11 +++++-----
 .../model-distribution-five-forms.md          |  6 +++---
 .../media/free-ai-media-generators/SKILL.md   |  4 ++--
 .../references/agnes-ai-verified.md           |  4 ++--
 .../scripts/agnes-check.py                    |  2 +-
 skills/openclaw/openclaw/SKILL.md             |  8 +++----
 skills/zhiyi/zhiyi/SKILL.md                   |  8 +++----
 17 files changed, 70 insertions(+), 58 deletions(-)

diff --git a/SOUL.md b/SOUL.md
index 2f3ee81d..d24dd4ea 100755
--- a/SOUL.md
+++ b/SOUL.md
@@ -108,6 +108,13 @@ skill 是我的法术
 3. **搜不到** → 根据经验自建 skill（skill_manage create，含验收与 commit），不许回到裸终端裸干模式
 4. 与自动路由/看板并行不悖：路由定执行者，技能定怎么干
 
+### 8. 本地 2B 自动下沉准则（2026-09-10 牧尘定案）
+**高频、结构化、可验证的活 → 默认下沉本地 2B；云模型只留低频、高质敏、创意的活。**
+- ✅ 代码/脚本/bug → opencode+2B（`--auto --dir`）；调研/分析/报告 → ao 角色串行（`ao-yaml-2b-runner.py`）；复杂任务 → serial-4agent 流水线；蒸馏/喂织忆 → 本地 2B；kanban 代码 → 自动巡检。
+- ❌ 2B 不做：闲聊/冷门知识、高难度翻译、大型工程重构、无人值守长链、面向客户质敏产出、SOUL/重要文档定稿。
+- 铁律：2B 禁止要求 JSON 输出（只写自评壳丢正文）→ 一律纯文本正文；opencode 必须 `--auto --dir`；验收硬检查代码化（2B 会假 pass）；返修上限 2 次。
+- 细节/触发词/费用账 → 加载 **`local-2b-task-triage`** skill。判断顺序：先想能不能下沉 2B，再想用哪个执行器，最后才考虑云。
+
 ---
 
 ## 🎯 任务执行铁律
@@ -116,10 +123,10 @@ skill 是我的法术
 
 | 关键词 | 路由 |
 |--------|------|
-| 调研/分析/竞品/报告/选型/方案/设计 | research (agnes-2.0-flash) |
-| 代码/实现/写/改/bug/插件/部署/飞书 | default (MiniMax-M3) |
-| npc/云端编码/CodeBuddy/cnb/替我上班 | npc (云端) |
-| dsh/模型测试/本地模型/llama | dsh（额度尽时 fallback default） |
+| 调研/分析/竞品/报告/选型/方案/设计 | research (agnes-2.5-flash) |
+| 代码/实现/写/改/bug/插件/部署/飞书 | default (deepseek-v4-flash) |
+| npc/网上免费智能/替我上班 | npc (网上免费 deepseek · sensenova) |
+| dsh/本地模型/模型测试/llama | dsh (本地 llama-local/minicpm5-2b) |
 | 并行/并发/swarm/分身 | Swarm 模式 |
 
 **禁止**：默认自己干 / 默认 default profile / 问牧尘"用什么模式"。
@@ -324,6 +331,7 @@ python3 ~/.hermes/scripts/kanban-route.py "目标" --swarm    # 并行
 ---
 
 *v4.3：2026-09-06。核心原则新增第 7 条「组合技能铁律」（强制）：任何任务必先查技能库并组合执行；技能库没有→优先搜索下载安装；搜不到→凭经验自建。牧尘直接指令，已同步 MEMORY + 织忆 distilled。*
+*v4.4：2026-09-10。核心原则新增第 8 条「本地 2B 自动下沉准则」：高频/结构化/可验证的活默认下沉本地 2B，云只留质敏/创意。配套新 skill `local-2b-task-triage`（决策表+铁律+费用账）。牧尘 09-10 直接指令。*
 *v4.2：2026-09-03 深夜。记忆规范加写入分级铁律（[已验证]/[推断]/[偏好]；推断禁写 MEMORY/USER、禁覆盖已验证旧记录；冲突不静默）。配套：织忆插件 commit 读 conflicts 返回 warning。*
 *v4.1：2026-09-03。记忆规范加路由总纲指针 + §5 扩为记忆调用主动触发铁律（memory_search/graph/write 硬触发词）。*
 *v4.0：2026-09-03。SOUL.md 从 36966 字节 → ~19000 字节（context 限制内）。自治能力章节抽出到 `~/.hermes/docs/SOUL-autonomy.md`，遵循索引/详情分离原则。*
diff --git a/config.yaml b/config.yaml
index 1d6d4913..6d906aca 100644
--- a/config.yaml
+++ b/config.yaml
@@ -12,10 +12,9 @@ providers:
   agnes:
     base_url: https://apihub.agnes-ai.com/v1
     cost_factor: 0.0
-    default_model: agnes-2.0-flash
+    default_model: agnes-2.5-flash
     key_env: AGNES_API_KEY
     models:
-    - agnes-2.0-flash
     - agnes-2.5-flash
     rate_limit: 1000
     timeout: 60
@@ -43,7 +42,7 @@ providers:
     api_key: 0ExNiLblJvIWBDpkS50fwOBw4MmqLyKdHJK5iQtlw9dOMWBP
     base_url: http://127.0.0.1:3000/v1
     cost_factor: 0.0
-    default_model: agnes-2.5-flash
+    default_model: nvidia/nemotron-3-super-120b-a12b
     models:
     - nvidia/nemotron-mini-4b-instruct
     - openai/gpt-oss-120b
@@ -104,7 +103,7 @@ providers:
     - minicpm5-2b
     rate_limit: 100
     timeout: 300
-    context_length: 90112
+    context_length: 45056   # 2026-09-10 双槽 44K×2
   gmi-cloud:
     api_key: eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpZCI6ImFlZGEwZTJmLTI1N2ItNGEwYi04NzBhLWUwZjI3MzQ0NWM2YSIsInNjb3BlIjoiaWVfbW9kZWwiLCJwcm9kdWN0IjoiSUUiLCJvd25lcklkIjoiOTYxZGY1MWUtMjAyMi00ZDZjLWI2YTgtNWMxN2M5NzA1NjEyIn0.7v6kPl4VcNHtOcU3hN2FAMGVEuDxvym1h9FwXZBaD7M
     base_url: https://api.gmi-serving.com/v1
@@ -118,7 +117,7 @@ providers:
         Gecko) Chrome/120.0 Safari/537.36
 fallback_providers:
 - provider: agnes
-  model: agnes-2.0-flash
+  model: agnes-2.5-flash
 - provider: zhipu
   model: glm-4-flash
 - provider: sensenova
@@ -259,7 +258,7 @@ auxiliary:
       timeout: 90
   compression:
     provider: agnes
-    model: agnes-2.0-flash
+    model: agnes-2.5-flash
     context_length: 512000
     fallback_chain:
     - provider: zhipu
@@ -284,7 +283,7 @@ auxiliary:
     timeout: 30
   mcp:
     provider: agnes
-    model: agnes-2.0-flash
+    model: agnes-2.5-flash
     base_url: ''
     api_key: ''
     timeout: 30
@@ -302,7 +301,7 @@ auxiliary:
     timeout: 120
   kanban_decomposer:
     provider: agnes
-    model: agnes-2.0-flash
+    model: agnes-2.5-flash
     base_url: ''
     api_key: ''
     timeout: 180
@@ -443,7 +442,7 @@ memory:
   user_char_limit: 1375
   provider: zhiyi
 delegation:
-  model: agnes-2.0-flash
+  model: agnes-2.5-flash
   provider: agnes
   base_url: ''
   api_key: ''
@@ -473,7 +472,7 @@ moa:
       reference_models:
       - model: deepseek-v4-flash
         provider: deepseek
-      - model: agnes-2.0-flash
+      - model: agnes-2.5-flash
         provider: agnes
       - model: glm-4-flash
         provider: zhipu
@@ -593,7 +592,7 @@ cron:
   provider: ''
   wrap_response: true
   default_deliver: origin,feishu:oc_81f6df701c872a1122f32080e366543f
-  fallback_model: agnes-2.0-flash
+  fallback_model: agnes-2.5-flash
   fallback_provider: agnes
   gateway_required: true
 kanban:
diff --git a/skills/chinese-platform-extraction/SKILL.md b/skills/chinese-platform-extraction/SKILL.md
index 79fe3a46..f635bea7 100644
--- a/skills/chinese-platform-extraction/SKILL.md
+++ b/skills/chinese-platform-extraction/SKILL.md
@@ -1,9 +1,10 @@
 ---
 name: chinese-platform-extraction
-version: "1.0.0"
+version: "1.0.1"
+date: 2026-09-10
 description: "Use when extracting content from WeChat or Zhihu."
 metadata:
-  version: "1.0.0"
+  version: "1.0.1"
   author: "小唯"
   category: "productivity"
 tags: ["web-scraping", "chinese-platforms", "wechat", "zhihiu", "content-extraction"]
diff --git a/skills/chinese-platform-extraction/references/cron-script-limitations.md b/skills/chinese-platform-extraction/references/cron-script-limitations.md
index 0610c070..58acd343 100644
--- a/skills/chinese-platform-extraction/references/cron-script-limitations.md
+++ b/skills/chinese-platform-extraction/references/cron-script-limitations.md
@@ -28,7 +28,7 @@ Cron 系统将整个 `script` 字段视为文件路径，不解析空格分隔
 cron:
   model: glm-4-flash
   model_provider: zhipu
-  fallback_model: agnes-2.0-flash
+  fallback_model: agnes-2.5-flash
   fallback_provider: agnes
 ```
 
diff --git a/skills/creative/agnes-ai/SKILL.md b/skills/creative/agnes-ai/SKILL.md
index 2e923f69..416087a7 100644
--- a/skills/creative/agnes-ai/SKILL.md
+++ b/skills/creative/agnes-ai/SKILL.md
@@ -1,7 +1,8 @@
 ---
 name: agnes-ai
 description: "Agnes AI 全模态 API — 文本/图像/视频生成。Base URL: https://apihub.agnes-ai.com/v1，OpenAI 兼容格式"
-version: 1.4.0
+version: 1.4.1
+date: 2026-09-10
 author: 小唯
 tags: [ai, image-generation, video-generation, text-generation, 免费]
 homepage: https://platform.agnes-ai.com
@@ -27,7 +28,7 @@ homepage: https://platform.agnes-ai.com
 | 模型 | 用途 | 备注 |
 |------|------|------|
 | `agnes-1.5-flash` | 文本生成 | 旧版 |
-| `agnes-2.0-flash` | 文本生成 | ✅ 推荐 |
+| `agnes-2.5-flash` | 文本生成 | ✅ 推荐 |
 | `agnes-2.5-flash` | 文本生成 | 新版，✅ 可用 |
 | `agnes-2.5-pro-alpha` | 文本生成 | ⚠️ 思考模型，content: null，不适合做主模型；**Terminal-Bench v2.1 67%**（2026-08-15 文章：能定位 Django 50万行代码里 asgi.py 异步 bug），适合复杂工程分析 |
 | `agnes-image-2.0-flash` | 图像生成 | |
@@ -45,7 +46,7 @@ KEY=$(grep AGNES_API_KEY ~/.hermes/.env | grep -v "^#" | head -1 | cut -d= -f2-)
 curl -s -X POST "https://apihub.agnes-ai.com/v1/chat/completions" \
   -H "Authorization: Bearer *** \
   -H "Content-Type: application/json" \
-  -d '{"model":"agnes-2.0-flash","messages":[{"role":"user","content":"say hi in 3 words"}],"max_tokens":20}'
+  -d '{"model":"agnes-2.5-flash","messages":[{"role":"user","content":"say hi in 3 words"}],"max_tokens":20}'
 ```
 
 ### ✅ bash 脚本（视频生成推荐）
@@ -93,7 +94,7 @@ KEY=$(grep AGNES_API_KEY ~/.hermes/.env | grep -v "^#" | head -1 | cut -d= -f2-)
 curl -s -X POST "https://apihub.agnes-ai.com/v1/chat/completions" \
   -H "Authorization: Bearer *** \
   -H "Content-Type: application/json" \
-  -d '{"model":"agnes-2.0-flash","messages":[{"role":"user","content":"say hello"}],"max_tokens":20}'
+  -d '{"model":"agnes-2.5-flash","messages":[{"role":"user","content":"say hello"}],"max_tokens":20}'
 ```
 
 ⚠️ **中文输入有 bug**（`model=None`），用英文提问。
@@ -184,6 +185,6 @@ GCS_URL = https://storage.googleapis.com/agnes-aigc/aigc/videos/YYYY/MM/DD/{vide
 ```bash
 model.provider=custom
 model.base_url=https://apihub.agnes-ai.com/v1
-model.default=agnes-2.0-flash
+model.default=agnes-2.5-flash
 model.api_key=sk-7k9...2ikW
 ```
\ No newline at end of file
diff --git a/skills/creative/agnes-ai/references/agnes.py b/skills/creative/agnes-ai/references/agnes.py
index 1767792c..01c06dc9 100755
--- a/skills/creative/agnes-ai/references/agnes.py
+++ b/skills/creative/agnes-ai/references/agnes.py
@@ -40,7 +40,7 @@ def generate_video(prompt, duration=5, poll_interval=5, max_wait=300):
         time.sleep(poll_interval)
     return ""
 
-def generate_text(prompt, model="agnes-2.0-flash", max_tokens=200):
+def generate_text(prompt, model="agnes-2.5-flash", max_tokens=200):
     r = post("/chat/completions", {"model": model, "messages": [{"role": "user", "content": prompt}], "max_tokens": max_tokens})
     return r.json().get("choices", [{}])[0].get("message", {}).get("content", "")
 
diff --git a/skills/creative/agnes-ai/references/agnes_subprocess.py b/skills/creative/agnes-ai/references/agnes_subprocess.py
index a3533267..823577c0 100755
--- a/skills/creative/agnes-ai/references/agnes_subprocess.py
+++ b/skills/creative/agnes-ai/references/agnes_subprocess.py
@@ -43,7 +43,7 @@ def generate_image(prompt, size='1024x1024', model='agnes-image-2.0-flash'):
     })
     return d.get('data', [{}])[0].get('url', '')
 
-def generate_text(prompt, model='agnes-2.0-flash', max_tokens=200):
+def generate_text(prompt, model='agnes-2.5-flash', max_tokens=200):
     d = curl_post('/chat/completions', {
         'model': model,
         'messages': [{'role': 'user', 'content': prompt}],
diff --git a/skills/cron-ops/SKILL.md b/skills/cron-ops/SKILL.md
index 92989cf1..c5cd71f0 100644
--- a/skills/cron-ops/SKILL.md
+++ b/skills/cron-ops/SKILL.md
@@ -1,9 +1,10 @@
 ---
 name: cron-ops
-version: "1.0.0"
+version: "1.0.1"
+date: 2026-09-10
 description: "Use when managing cron jobs or configuring fallback chains."
 metadata:
-  version: "1.0.0"
+  version: "1.0.1"
   author: "小唯"
   category: "devops"
 tags: ["cron", "fallback", "retry", "scheduled-tasks", "运维"]
@@ -21,7 +22,7 @@ Hermes 定时任务的配置、fallback、重试和故障排查。
 cron:
   model: glm-4-flash          # 主模型
   model_provider: zhipu        # 主 provider
-  fallback_model: agnes-2.0-flash  # fallback 模型
+  fallback_model: agnes-2.5-flash  # fallback 模型
   fallback_provider: agnes     # fallback provider
   provider: auto               # auto = 按 model_provider 走
   gateway_required: true       # gateway 在才 fire
diff --git a/skills/devops/hermes-debug/SKILL.md b/skills/devops/hermes-debug/SKILL.md
index 72153f3c..f751b699 100644
--- a/skills/devops/hermes-debug/SKILL.md
+++ b/skills/devops/hermes-debug/SKILL.md
@@ -2,8 +2,8 @@
 tags: [devops]
 name: hermes-debug
 description: Hermes 故障排查工具 — 常见问题诊断命令、模型配置、Gateway 状态检查、多Profile排查、NewAPI DEGRADED错误。牧尘专用。
-version: 1.14.0
-date: 2026-08-01
+version: 1.14.1
+date: 2026-09-10（agnes 2.0→2.5 字样清理）
 ---
 
 # Hermes 故障排查工具
@@ -215,7 +215,7 @@ fallback_providers: '["mimo", "deepseek", "opencode-free"]'
 # ✅ 正确：YAML 列表，每项是 {provider, model} 字典（base_url/key 从 providers 段解析）
 fallback_providers:
   - provider: agnes
-    model: agnes-2.0-flash
+    model: agnes-2.5-flash
   - provider: zhipu
     model: glm-4-flash
   - provider: sensenova
diff --git a/skills/devops/llm-gateway-ops/SKILL.md b/skills/devops/llm-gateway-ops/SKILL.md
index 3d2d2d94..40eb56a0 100644
--- a/skills/devops/llm-gateway-ops/SKILL.md
+++ b/skills/devops/llm-gateway-ops/SKILL.md
@@ -1,6 +1,7 @@
 ---
 name: llm-gateway-ops
-version: "1.0.0"
+version: "1.0.1"
+date: 2026-09-10
 description: LLM 网关运维 — OmniRoute/NewAPI 类 AI 网关的评估、部署、API 调用、观测体系、cron 切换。含网关选型方法论、npm 安装坑、urllib 不兼容、SSE 读取模式、systemd 常驻、看门狗接入。
 ---
 
@@ -275,9 +276,9 @@ agnes_block = """  agnes:
     key_env: AGNES_API_KEY
     base_url: https://apihub.agnes-ai.com/v1
     cost_factor: 0.0
-    default_model: agnes-2.0-flash
+    default_model: agnes-2.5-flash
     models:
-      - agnes-2.0-flash
+      - agnes-2.5-flash
       - agnes-2.5-flash
     rate_limit: 1000
     timeout: 60
@@ -299,7 +300,7 @@ d = json.load(open('~/.hermes/cron/jobs.json'))  # 先 cp 备份
 jobs = d if isinstance(d, list) else d.get('jobs', [])
 for j in jobs:
     if 'newapi' in (j.get('provider') or '') or 'gpt-oss' in (j.get('model') or ''):
-        j['provider'] = 'agnes'; j['model'] = 'agnes-2.0-flash'
+        j['provider'] = 'agnes'; j['model'] = 'agnes-2.5-flash'
 json.dump(d, open('~/.hermes/cron/jobs.json','w'), ensure_ascii=False, indent=2)
 ```
 改完 `cronjob list` 确认生效（model/provider 字段变化），再 `cronjob run <id>` 实测一个任务（投研简报实测通过）。
@@ -308,7 +309,7 @@ json.dump(d, open('~/.hermes/cron/jobs.json','w'), ensure_ascii=False, indent=2)
 
 ```ini
 Environment=LLM_ENDPOINT=https://apihub.agnes-ai.com/v1/chat/completions
-Environment=LLM_MODEL=agnes-2.0-flash
+Environment=LLM_MODEL=agnes-2.5-flash
 Environment=LLM_API_KEY=<AGNES 完整 key>
 ```
 ⚠️ 只改 MODEL 会 401（endpoint 还指着 NewAPI）。改后 `systemctl --user daemon-reload && systemctl --user restart zhiyid`，验证 `/api/v1/health` + `cat /proc/$(pgrep -f zhiyid-new)/environ | tr '\0' '\n' | grep LLM_`。
@@ -334,7 +335,7 @@ Environment=LLM_API_KEY=<AGNES 完整 key>
 
 | 条件 | 示例 | 做法 |
 |------|------|------|
-| provider 有免费额度且 API 稳定 | Agnes（agnes-2.0-flash） | 直接用 `AGNES_API_KEY` 调 `apihub.agnes-ai.com` |
+| provider 有免费额度且 API 稳定 | Agnes（agnes-2.5-flash） | 直接用 `AGNES_API_KEY` 调 `apihub.agnes-ai.com` |
 | provider 的 API 格式与 OpenAI 不完全兼容 | 某些国内模型 | 直调避免兼容层损失 |
 | 网关路由可能选错模型/超时 | 自动路由 cron 504 问题 | 钉死到 provider 原生端点 |
 | 需要推理模型的 reasoning_content 字段 | gpt-oss 系列 | 网关可能吞掉该字段 |
diff --git a/skills/devops/provider-tiering/SKILL.md b/skills/devops/provider-tiering/SKILL.md
index 9067bb36..b3e4ad29 100644
--- a/skills/devops/provider-tiering/SKILL.md
+++ b/skills/devops/provider-tiering/SKILL.md
@@ -1,7 +1,8 @@
 ---
 name: provider-tiering
 description: Provider 分层策略 — 免费模型做常规工作，付费模型用于战略决策。含 NewAPI 免费 NVIDIA 模型库、模型健康巡检系统、cron job 模型 override、delegate_task 模型路由、MOA 配置规范。牧尘专用。
-version: 3.7.0
+version: 3.7.1
+date: 2026-09-10（agnes 2.0→2.5 字样清理）
 tags: [model-cost, provider-strategy, newapi, deepseek, tiering, health-check, auto-heal]
 category: devops
 trigger: 任何涉及模型选型、 provider 配置、 cron job 创建、 delegate_task 派发、新工作流搭建的场景
@@ -251,7 +252,7 @@ print(f'稳定: {d[\"stable\"]}, 不稳定: {d[\"unstable\"]}, 死: {d[\"dead\"]
 **API Key**：`AGNES_API_KEY`（~/.hermes/.env，51 字符完整 key，obsidian key.md 备份）
 **成本因子**：`cost_factor: 0.0`（免费）
 **日额限制**：文本 50万 token/天，图像 500张/天，视频 150条/天
-**模型**：`agnes-2.0-flash`（文本，**推荐主力**，稳定）、`agnes-2.5-flash`（文本，质量略高）、`agnes-image-2.0/2.1-flash`（图像）、`agnes-video-v2.0`（视频）
+**模型**：`agnes-2.5-flash`（文本，**推荐主力**）、`agnes-image-2.0/2.1-flash`（图像）、`agnes-video-v2.0`（视频）
 
 #### 为什么接入（2026-08-17 实测，牧尘指示）
 
@@ -264,9 +265,9 @@ NewAPI 免费模型不稳定：`openai/gpt-oss-120b` 3 次调用 1 次空响应
     key_env: AGNES_API_KEY
     base_url: https://apihub.agnes-ai.com/v1
     cost_factor: 0.0
-    default_model: agnes-2.0-flash
+    default_model: agnes-2.5-flash
     models:
-      - agnes-2.0-flash
+      - agnes-2.5-flash
       - agnes-2.5-flash
     rate_limit: 1000
     timeout: 60
@@ -462,7 +463,7 @@ agent-default-model:
     "api": "openai-completions",
     "timeoutSeconds": 120,
     "models": [
-      {"id": "agnes-2.0-flash", "name": "Agnes 2.0 Flash", "reasoning": true,
+      {"id": "agnes-2.5-flash", "name": "Agnes 2.5 Flash", "reasoning": true,
        "input": ["text"], "contextWindow": 1048576, "maxTokens": 131072,
        "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0}}
     ]
diff --git a/skills/devops/provider-tiering/references/model-distribution-five-forms.md b/skills/devops/provider-tiering/references/model-distribution-five-forms.md
index 68773640..8b0676f6 100644
--- a/skills/devops/provider-tiering/references/model-distribution-five-forms.md
+++ b/skills/devops/provider-tiering/references/model-distribution-five-forms.md
@@ -13,7 +13,7 @@
 
 | # | 分体 | 本质 | 主模型 | 备用模型 |
 |:-|:-----|:-----|:------|:--------|
-| 1 | **cron 任务** | 定时自动化 | agnes-2.0-flash (11个) / deepseek-v4-flash (2个) | — |
+| 1 | **cron 任务** | 定时自动化 | agnes-2.5-flash (11个) / deepseek-v4-flash (2个) | — |
 | 2 | **delegate_task** | 并行任务分身 | 继承父 session | — |
 | 3 | **OpenClaw** | 独立网关代理 | deepseek-v4-flash (sensenova) | glm-5.2 (sensenova) |
 | 4 | **opencode** | 编码代理 | deepseek-v4-flash (sensenova) | glm-5.2 |
@@ -64,7 +64,7 @@
 **distill-model-watchdog.py CANDIDATE_POOL**：
 ```python
 CANDIDATE_POOL = [
-    "agnes-2.0-flash",           # ⭐ Agnes 优先
+    "agnes-2.5-flash",           # ⭐ Agnes 优先
     "deepseek-v4-flash",         # 商汤免费，1M ctx
     "glm-5.2",                   # 商汤免费，1M ctx
     "sensenova-6.8-flash-lite",  # 商汤免费，轻量多模态
@@ -76,7 +76,7 @@ CANDIDATE_POOL = [
 CONFIG_DECLARED_MODELS = [
     "deepseek-v4-flash",         # 商汤免费 1M ctx ⭐
     "glm-5.2",                   # 商汤免费 1M ctx
-    "agnes-2.0-flash",           # Agnes 免费 128K ctx
+    "agnes-2.5-flash",           # Agnes 免费 128K ctx
     "sensenova-6.8-flash-lite",  # 商汤免费 262K ctx
 ]
 ```
diff --git a/skills/media/free-ai-media-generators/SKILL.md b/skills/media/free-ai-media-generators/SKILL.md
index 21eceaa1..f09b97f5 100755
--- a/skills/media/free-ai-media-generators/SKILL.md
+++ b/skills/media/free-ai-media-generators/SKILL.md
@@ -4,10 +4,10 @@ description: 免费 AI 图片/视频生成平台汇总 — Pollinations、Huggin
 trigger: 生成图片 / 画图 / text2image / AI画图 / 免费出图 / 视频生成 / Agnes
 trigger_fallback: true
 author: 小唯
-version: 2.0.0
+version: 2.0.1
 tags: [image-generation, video-generation, free-api, pollinations, huggingface, agnes-ai]
 created: 2026-05-08
-updated: 2026-06-03
+updated: 2026-09-10（agnes 2.0→2.5 字样清理）
 ---
 
 # 免费 AI 图片/视频生成平台汇总
diff --git a/skills/media/free-ai-media-generators/references/agnes-ai-verified.md b/skills/media/free-ai-media-generators/references/agnes-ai-verified.md
index 524336ad..ebd635f9 100755
--- a/skills/media/free-ai-media-generators/references/agnes-ai-verified.md
+++ b/skills/media/free-ai-media-generators/references/agnes-ai-verified.md
@@ -16,7 +16,7 @@
 | 模型 | 用途 |
 |------|------|
 | `agnes-1.5-flash` | 文本生成（旧版） |
-| `agnes-2.0-flash` | 文本生成（新版） |
+| `agnes-2.5-flash` | 文本生成（新版，2.0 已升级） |
 | `agnes-image-2.0-flash` | 图像生成 |
 | `agnes-image-2.1-flash` | 图像生成（新版） |
 | `agnes-video-v2.0` | 视频生成（异步） |
@@ -48,7 +48,7 @@ def get(path):
 ### 文本
 ```python
 r = post("/chat/completions", {
-    "model": "agnes-2.0-flash",
+    "model": "agnes-2.5-flash",
     "messages": [{"role": "user", "content": "say hello in 5 words"}],
     "max_tokens": 50
 })
diff --git a/skills/media/free-ai-media-generators/scripts/agnes-check.py b/skills/media/free-ai-media-generators/scripts/agnes-check.py
index d7daf05c..586c19bf 100755
--- a/skills/media/free-ai-media-generators/scripts/agnes-check.py
+++ b/skills/media/free-ai-media-generators/scripts/agnes-check.py
@@ -17,7 +17,7 @@ import requests
 # 1. 文本
 print("=== 文本测试 ===")
 r = requests.post(f"{BASE}/chat/completions",
-    json={"model":"agnes-2.0-flash","messages":[{"role":"user","content":"say hello in 3 words"}],"max_tokens":20},
+    json={"model":"agnes-2.5-flash","messages":[{"role":"user","content":"say hello in 3 words"}],"max_tokens":20},
     headers=h, timeout=20)
 print("  状态:", r.status_code, r.json().get("choices", [{}])[0].get("message", {}).get("content", r.text[:80]))
 
diff --git a/skills/openclaw/openclaw/SKILL.md b/skills/openclaw/openclaw/SKILL.md
index 43ba6a2c..1a121bfb 100644
--- a/skills/openclaw/openclaw/SKILL.md
+++ b/skills/openclaw/openclaw/SKILL.md
@@ -2,9 +2,9 @@
 tags: [openclaw, mcp, hermes-integration]
 name: openclaw
 description: "OpenClaw 日常维护与 Hermes MCP 集成 — skill 管理、配置修复、飞书频道、诊断命令、MCP Server 接入"
-version: 1.5.0
+version: 1.5.1
 author: 小唯
-date: 2026-07-12
+date: 2026-09-10（agnes 2.0→2.5 字样清理）
 ---
 
 readiness_status: available
@@ -439,13 +439,13 @@ cfg = json.load(open(p))
 # 1. 改 defaults
 cfg['agents']['defaults']['model'] = {
     'primary': 'mimo/mimo-v2.5-pro',
-    'fallbacks': ['sensenova/deepseek-v4-flash', 'agnes/agnes-2.0-flash']
+    'fallbacks': ['sensenova/deepseek-v4-flash', 'agnes/agnes-2.5-flash']
 }
 # 2. 改每个 agent
 for a in cfg['agents']['list']:
     a['model'] = {
         'primary': 'mimo/mimo-v2.5-pro',
-        'fallbacks': ['sensenova/deepseek-v4-flash', 'agnes/agnes-2.0-flash']
+        'fallbacks': ['sensenova/deepseek-v4-flash', 'agnes/agnes-2.5-flash']
     }
 json.dump(cfg, open(p, 'w'), indent=2, ensure_ascii=False)
 ```
diff --git a/skills/zhiyi/zhiyi/SKILL.md b/skills/zhiyi/zhiyi/SKILL.md
index 8af3aa0f..687fef91 100644
--- a/skills/zhiyi/zhiyi/SKILL.md
+++ b/skills/zhiyi/zhiyi/SKILL.md
@@ -2,8 +2,8 @@
 tags: [zhiyi]
 name: zhiyi
 description: "织忆 (MemoryWeave) 聚合技能 — API 客户端 + 开发工作流 + 运维规范。含 commit/recall API、数据架构、部署验证、Go 方法论。"
-version: 11.44
-updated: 2026-09-09（蒸馏输入粒度→88K 窗口利用率追查；正本路径 src/memoryweave；24-38K 大请求真凶仍未定位）
+version: 11.45
+updated: 2026-09-10（agnes 2.0→2.5 字样清理）
 
 ---
 
@@ -1181,7 +1181,7 @@ python3 -c "from plugins.memory.zhiyi import HermesZhiYiMemoryProvider; \
 
 | 维度 | 事实 |
 |------|------|
-| 当前实际蒸馏模型 | `agnes-2.0-flash`（付费，~2.6s/调用，reasoning_tokens 占大头） |
+| 蒸馏模型（**2026-09-05 审计时**） | `agnes-2.0-flash`（付费，~2.6s/调用）→ **2026-09-10 已切 `local/minicpm5-2b`（本地2B，免费）** |
 | 看门狗健康判 | ✅ agnes 探针 OK（最近 30min 都 silent） |
 | local/qwen7b 当前状态 | **健康**（llama-server pid 578442，8080 OK，24 t/s，1.1s 出 JSON）—— watchdog 没用而已 |
 | 成本 | agnes 是付费，local 0 成本；当前 agnes 1 调用 ~$0.0003 量级（小但有） |
@@ -1204,7 +1204,7 @@ python3 -c "from plugins.memory.zhiyi import HermesZhiYiMemoryProvider; \
 3. **手动测试 local → 切回 agnes 的路径**：改 `LLM_MODEL=local/qwen7b` + `LLM_ENDPOINT=http://127.0.0.1:8080/v1/chat/completions` + `LLM_API_KEY=local-key` → `systemctl --user daemon-reload && restart zhiyid` → journal 验证 `flush START: model=local/qwen7b` + `LLM entities: N`。
 4. **审计蒸馏链路的标准 6 步**（避免再"凭印象答"）：① systemd unit Environment 真值 → ② /proc/<pid>/environ 真值（防 unit 没 reload）→ ③ cron 看门狗 output 最近非空条目 → ④ cron 巡检 _heal_distill_models 配置 → ⑤ llama-server /health + 当前 CPU/GPU → ⑥ zhiyid sidecar 日志最近一次 `flush START: model=...` → 三方一致才算"健康"。
 
-**修正点**：之前版本（v11.37）写的"zhiyid.service 蒸馏已切本地 llama-server`local/qwen7b`(0成本)" **2026-09-05 06:40 后失效**，当前实际是 agnes-2.0-flash。本段是事实真相。
+**修正点**：之前版本（v11.37）写的"zhiyid.service 蒸馏已切本地 llama-server`local/qwen7b`(0成本)" **2026-09-05 06:40 后失效**，当时实际是 agnes-2.0-flash。⚠️ **2026-09-10 最新状态：zhiyid 已切 `local/minicpm5-2b`（本地2B，LLM_ENDPOINT=127.0.0.1:8080）**。本段保留为历史审计记录。
 
 ---