fix(optimizer+daemon): 统计口径与日志刷屏 — cron读jobs键(57任务)、daemon逐行分类(error 60295→~50)、compaction空转刷10万行停刷、深度思考日志改单行

scripts/daemon.py

@@ -118,6 +118,7 @@ def log_reasoning_step(step_type, message, data=None):
 COMPACTION_MODEL = "glm-4-flash"
 COMPACTION_RATIO_THRESHOLD = 0.80  # Grok IntraCompactionTrigger 思路：token ratio > 80% 触发摘要
 COMPACTION_MIN_TURNS = 10           # 至少 10 条 pattern 才压缩
+COMPACTION_MIN_INTERVAL = 300       # 距上次实际压缩至少 5 分钟（2026-09-06 防每 30s tick 空转刷日志）
 _compaction_last_run = 0             # 上次压缩时间戳
 _compaction_count = 0               # 累计压缩次数
 
@@ -2399,7 +2400,9 @@ def deep_think(ctx, state, changes, journal, solutions_lib):
     if not result:
         return {"evaluation_previous_goal": "LLM调用失败", "memory": "上次调用失败", "next_goal": "重试"}, ""
 
-    log(f"  深度思考 ({tokens}t): {result[:200]}")
+    # 2026-09-06 修复：result 含换行时 log() 会把整段 LLM 输出折成多行续行
+    # （每 2 分钟一次 deep_tick → 数十万续行），改为压缩成单行再写日志。
+    log(f"  深度思考 ({tokens}t): {' '.join(result[:200].split())}")
 
     # Parse JSON output
     reflection_dict = {"evaluation_previous_goal": "", "memory": "", "next_goal": ""}
@@ -2712,7 +2715,13 @@ def main_loop():
             # ── Grok Build Compaction 系统：自动压缩检查 ───────────────────
             state_for_compact = {"patterns": state.get("_patterns", []), "observations": state.get("_observations", [])}
             should_cp, cp_reason = _should_compact(ctx, state_for_compact)
-            if should_cp:
+            # 2026-09-06 修复：旧代码只要 _should_compact 返回 True（journal>150 常驻成立——
+            # journal_entry→trim_journal 裁剪到 JOURNAL_MAX=200，永远 >150）就 log+尝试压缩，
+            # 而 patterns 为空时 _trigger_compaction 必然早退 → 每 30s light tick 刷一行
+            # "🗜️ Compaction 触发"，50 天刷了 10.3 万行（daemon.log 58MB 主源之一）。
+            # 现在：patterns 不足 10 条不空跑；距上次实际压缩 <5 分钟不重复尝试。
+            if should_cp and len(state_for_compact.get("patterns", [])) >= COMPACTION_MIN_TURNS \
+                    and time.time() - _compaction_last_run >= COMPACTION_MIN_INTERVAL:
                 log(f"  🗜️ Compaction 触发: {cp_reason}")
                 _trigger_compaction(ctx, state_for_compact)
             

scripts/optimizer.py

@@ -48,17 +48,45 @@ def collect_daemon_metrics():
         metrics["learned_solutions"] = ctx.get("learned_count", 0)
     
     # 从 daemon.log 提取崩溃和错误
+    # ⚠️ 2026-09-06 修复：
+    #   1) 只统计日志尾部 2MB（≈最近1-2天），避免把历史(如 07-25 NVML 6145 次崩溃)当"当前崩溃"；
+    #   2) 旧代码用子串 count("Error"/"error"/"失败") 会把 deep_think 的 LLM 输出文本
+    #      (如 soulful 行"沮丧情绪（失败）") 也数进去 → 报出 60295 个假错误。
+    #      改为逐行分类：只统计 [DAEMON] 时间戳前缀行（daemon 自己的 log() 写的行），
+    #      跳过 LLM 输出续行；崩溃=行首"❌ 崩溃"，错误=⚠️/[call_llm] 失败标记行。
     log_file = D + "/daemon.log"
     if os.path.exists(log_file):
-        with open(log_file) as f:
+        with open(log_file, "rb") as f:
-            content = f.read()
+            f.seek(0, 2)  # 到末尾
-        metrics["crashes"] = content.count("❌ 崩溃")
+            size = f.tell()
-        metrics["errors"] = content.count("Error") + content.count("error") + content.count("失败")
+            f.seek(max(0, size - 2 * 1024 * 1024))  # 尾部 2MB
-        # 模型调用次数
+            content = f.read().decode("utf-8", errors="ignore")
-        metrics["model_calls"] = content.count("深度思考")
+        line_re = re.compile(r"^\[DAEMON\] \d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2} (.*)$")
-        # 从日志提取 token 数
+        call_re = re.compile(r"^\[DAEMON\] \d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}\s+深度思考 \((\d+)t\)")
-        token_matches = re.findall(r'\((\d+)t\)', content)
+        err_words = ("失败", "异常", "错误", "failed", "Failed", "超时", "timeout", "Timeout")
-        metrics["model_tokens"] = sum(int(t) for t in token_matches)
+        crashes = errors = model_calls = model_tokens = 0
+        for line in content.splitlines():
+            m = line_re.match(line)
+            if not m:
+                continue  # LLM 输出续行等非 daemon 行，跳过
+            cm = call_re.match(line)
+            if cm:
+                model_calls += 1
+                model_tokens += int(cm.group(1))
+                continue
+            msg = m.group(1).lstrip()
+            if msg.startswith("❌ 崩溃"):
+                crashes += 1
+            elif (msg.startswith(("⚠️", "❌")) or msg.startswith("[call_llm]")) \
+                    and any(w in msg for w in err_words):
+                errors += 1
+            elif "Traceback" in msg or re.search(r"\b(Error|ERROR|Exception)\b", msg):
+                errors += 1
+        metrics["crashes"] = crashes
+        metrics["errors"] = errors
+        # 模型调用次数 + token 数（只从"深度思考 (Nt)"调用行取，N=该次调用 tokens）
+        metrics["model_calls"] = model_calls
+        metrics["model_tokens"] = model_tokens
     
     return metrics
 
@@ -74,24 +102,31 @@ def collect_cron_metrics():
     }
     
     # 从 cron/jobs.json 读取
+    # ⚠️ 2026-09-06 修复：jobs.json 顶层是 dict {"jobs": [57个任务], "updated_at": ...}，
+    #    旧代码把 dict.values() 当任务列表 → total_jobs 只数到顶层键个数(2)，
+    #    57 个真实任务全部漏计（报告显示 "2任务 0成功"）。
     jobs_file = HERMES + "/cron/jobs.json"
     if os.path.exists(jobs_file):
         with open(jobs_file) as f:
             try:
-                jobs = json.load(f)
+                data = json.load(f)
-                # 处理不同格式
+                if isinstance(data, dict):
-                if isinstance(jobs, dict):
+                    jobs = data.get("jobs", [])
-                    jobs = [v for v in jobs.values()]
+                    # 兜底: 兼容 {job_id: job} 形态
-                elif isinstance(jobs, list):
+                    if not isinstance(jobs, list) and jobs:
-                    pass
+                        jobs = list(jobs.values())
-                
+                elif isinstance(data, list):
+                    jobs = data
+                else:
+                    jobs = []
+
                 metrics["total_jobs"] = len(jobs)
                 for j in jobs:
                     if isinstance(j, dict):
-                        status = j.get("last_status", "")
+                        status = j.get("last_status") or ""
                         if status == "ok":
                             metrics["ok_jobs"] += 1
-                        elif status and status != "ok":
+                        elif status in ("error", "failed"):
                             metrics["failed_jobs"] += 1
                         if j.get("no_agent"):
                             metrics["no_agent_jobs"] += 1