@modusensus/dsh-mneme 0.8.5 → 0.8.6

package/lib/index.js

@@ -43,7 +43,15 @@ export { Config };
 // unit-testable; the extractor only ever sees a callLLM(messages, options)
 // => Promise<string>. The route always carries a real provider/model (dsh-llm
 // GenerateOptions requires both) — never a bare stream.
-export function createEntityStreamAdapter({ llm, agentDefaultModel, logger }) {
+//
+// Issue #250: it also accounts for itself. Entity extraction is the background
+// LLM path that runs on every memory write, yet it never reached
+// llm_audit_logs — the adapter only ever held an llm handle, so it had no way
+// to call service.saveLlmAudit. It now takes service (+ config for the shared
+// llmAudit.enabled gate) and writes one row per stream attempt, matching the
+// contract of runAuditedLlm in dream.js: a rejected effort attempt records its
+// own error row and the retry records its own success row.
+export function createEntityStreamAdapter({ llm, agentDefaultModel, logger, service, config }) {
   return async function streamEntityText(messages, options = {}) {
     let route = {};
     if (options.provider) route.provider = options.provider;
@@ -58,8 +66,38 @@ export function createEntityStreamAdapter({ llm, agentDefaultModel, logger }) {
       } catch { /* fall through to whatever route we already have */ }
     }
     const effort = options.reasoningEffort;
+    // 没有解析出 provider/model 就没有可归属的模型——与 dream 的 resolveRoute
+    // 无路由早退同口径，那种情况不写审计行。
+    const modelId = route.provider && route.model ? `${route.provider}:${route.model}` : "";
     const tryStream = (withEffort) => {
       let text = "";
+      let inputTokens = 0;
+      let outputTokens = 0;
+      const startedAt = Date.now();
+      const timestamp = new Date(startedAt).toISOString();
+      // 记账是 best-effort：写审计行失败只 warn，绝不反噬抽取本身（CONTRIBUTING
+      // 的 fail-safe 硬约定）。
+      const writeAudit = (status, errorMessage) => {
+        if (config?.llmAudit?.enabled === false || !modelId || typeof service?.saveLlmAudit !== "function") return;
+        try {
+          service.saveLlmAudit({
+            timestamp,
+            trigger_source: "entityExtract",
+            operation_type: "entity_extract",
+            model_id: modelId,
+            input_tokens: inputTokens,
+            output_tokens: outputTokens,
+            total_tokens: inputTokens + outputTokens,
+            cost_usd: 0,
+            duration_ms: Date.now() - startedAt,
+            status,
+            error_message: errorMessage,
+            related_memory_ids: []
+          });
+        } catch (auditError) {
+          logger?.warn?.(`dsh-mneme: entity extraction llm audit write failed: ${String(auditError)}`);
+        }
+      };
       return (async () => {
         for await (const chunk of llm.stream({
           ...route,
@@ -68,11 +106,26 @@ export function createEntityStreamAdapter({ llm, agentDefaultModel, logger }) {
           messages
         })) {
           if (chunk.type === "text-delta" && typeof chunk.text === "string") text += chunk.text;
-          if (chunk.type === "finish" && (chunk.reason?.kind === "error" || chunk.reason?.kind === "aborted")) return undefined;
+          // DSH 的 StreamChunk 契约把用量嵌在 chunk.usage（TokenUsage）；兼容直接
+          // 平铺在 chunk 上的旧形态。归一放在这里，下面只需认平铺形状——与
+          // dream.js 的 streamText 同一处理。
+          if (chunk.type === "usage") {
+            const u = chunk.usage ?? chunk;
+            const i = u.input_tokens ?? u.inputTokens ?? u.prompt_tokens ?? u.promptTokens;
+            const o = u.output_tokens ?? u.outputTokens ?? u.completion_tokens ?? u.completionTokens;
+            if (Number.isFinite(i)) inputTokens = i;
+            if (Number.isFinite(o)) outputTokens = o;
+          }
+          if (chunk.type === "finish" && (chunk.reason?.kind === "error" || chunk.reason?.kind === "aborted")) {
+            writeAudit("error", "llm stream aborted or errored");
+            return undefined;
+          }
         }
+        writeAudit("success", null);
         return text;
       })().catch((err) => {
         logger?.warn?.(`[dsh-mneme] entity extraction llm stream failed: ${String(err)}`);
+        writeAudit("error", String(err?.message ?? err));
         return undefined;
       });
     };
@@ -273,6 +326,9 @@ export const apply = (ctx, config) => {
         dimension: cfg.localEmbedDimension,
         device: cfg.localEmbedDevice,
         batchSize: cfg.localEmbedBatchSize,
+        // 池化方式必须与模型的训练口径一致（BGE 系 = CLS）。它既进 embed() 的调用，
+        // 也进 modelHash —— 池化改了就是换向量空间，既有索引会被判失配并重建。
+        pooling: cfg.localEmbedPooling,
         cacheDir: cfg.embedModelCacheDir,
         runtimeDir: cfg.runtimeDir,
         // #188：embedModelMirror 接成 transformers 的下载镜像（此前死配置）。
@@ -403,6 +459,7 @@ export const apply = (ctx, config) => {
       minIntervalMs: (cfg.dreamMinIntervalMinutes ?? 0) * 60000,
       logger: ctx.logger,
       semantic: { embedder, vectorIndex },
+      lastRunAtSeed: store.lastDreamRunAt("auto"),
       onRun: () => (dream ? dream.runDream(ctx, service, cfg) : Promise.resolve({ ok: true, skipped: true }))
     });
     service.setDreamHook(() => dream.maybeSchedule(service));
@@ -419,6 +476,7 @@ export const apply = (ctx, config) => {
       service,
       config: cfg,
       logger: ctx.logger,
+      lastRunAtSeed: store.lastDreamRunAt("sleep"),
       onRun: (signal) => (sleep ? runSleep(ctx, service, cfg, ctx.logger, { embedder, vectorIndex }, signal) : Promise.resolve({ ok: true, skipped: true }))
     });
     service.setSleepHook(() => sleep.noteWrite());
@@ -442,7 +500,11 @@ export const apply = (ctx, config) => {
       const streamEntityText = createEntityStreamAdapter({
         llm: ctx.llm,
         agentDefaultModel: ctx.agentDefaultModel,
-        logger: ctx.logger
+        logger: ctx.logger,
+        // Issue #250: the adapter needs service to write its llm_audit_logs row
+        // (it never had it) and config for the shared llmAudit.enabled gate.
+        service,
+        config: cfg
       });
       service.setEntityExtractor((memory) =>
         extractEntities(memory, { store, config: cfg, callLLM: streamEntityText, logger: ctx.logger })

package/lib/local-embedder.js

@@ -21,6 +21,32 @@ function modelHash(model) {
   return `${model}#${hashString(model)}`;
 }
 
+/**
+ * Model families whose embedding is the CLS token rather than a mean over tokens.
+ *
+ * BGE (BAAI/bge-*) is trained that way: the model's own `1_Pooling/config.json`
+ * declares `pooling_mode_cls_token: true` / `pooling_mode_mean_tokens: false`, and
+ * the official README says "select the last hidden state of the first token" plus
+ * L2 normalization. transformers.js's `feature-extraction` defaults to `mean`, so
+ * feeding it a BGE model silently produces slightly-off vectors — no error, just a
+ * worse ranking. This table is the extension point: a family goes in only once its
+ * pooling is confirmed against the model's own `1_Pooling` (or equivalent) config.
+ */
+const CLS_POOLED_MODEL_PATTERNS = [/bge/i];
+
+/**
+ * Resolve the pooling to embed with.
+ *
+ * `auto` (default) picks by model family — BGE → `cls`, everything else → `mean`,
+ * so models that were never affected keep their exact previous behavior. An
+ * explicit `cls` / `mean` wins over the family table.
+ */
+export function resolvePooling(pooling, model) {
+  const mode = String(pooling ?? "auto").trim().toLowerCase();
+  if (mode === "cls" || mode === "mean") return mode;
+  return CLS_POOLED_MODEL_PATTERNS.some((re) => re.test(String(model ?? ""))) ? "cls" : "mean";
+}
+
 /**
  * Lazy default loader: dynamic import keeps module load cheap.
  *
@@ -71,8 +97,9 @@ function tensorToRows(tensor) {
 
 /**
  * ONNX text embedder backed by transformers.js (onnxruntime-node underneath).
- * Runs fully offline with mean pooling + L2 normalization for BERT-style
- * models like bge-small-zh. `engineFactory` is injectable for tests.
+ * Runs fully offline; pooling is resolved per model family (BGE → CLS, else mean —
+ * see resolvePooling) and vectors are L2-normalized. `engineFactory` is injectable
+ * for tests.
  */
 export class LocalEmbedder {
   constructor(opts = {}) {
@@ -87,6 +114,10 @@ export class LocalEmbedder {
     // （含测试）保持裸 fetch 的现状，配置默认值由 index.js 传进来。
     this.resilientModelDownload = opts.resilientModelDownload === true;
     this.useDtype = opts.useDtype || "q8";
+    // 池化：配置值（auto/cls/mean）→ 生效值（cls/mean）。解析结果既进 embed()，
+    // 也进 modelHash —— 换池化就是换向量空间，旧索引必须被判失配后重建。
+    this.poolingMode = String(opts.pooling ?? "auto").trim().toLowerCase();
+    this.pooling = resolvePooling(this.poolingMode, this.model);
     // 模型镜像下载源（embedModelMirror，#188 补充）：空 = transformers 默认。
     this.remoteHost = String(opts.remoteHost ?? "").trim();
     this.logger = opts.logger ?? null;
@@ -114,19 +145,19 @@ export class LocalEmbedder {
     this.extractor = await this.engineFactory("feature-extraction", this.model, options);
     this.ready = true; // service reads this to flush queued re-embeds
     this.logger?.info?.(
-      `[dsh-mneme] local embedder ready: ${this.model} (dim=${this._dimension}, device=${this.device})`
+      `[dsh-mneme] local embedder ready: ${this.model} (dim=${this._dimension}, device=${this.device}, pooling=${this.pooling})`
     );
     return this;
   }
 
-  /** Embed many texts with mean pooling; chunks at batchSize. */
+  /** Embed many texts with the model's pooling; chunks at batchSize. */
   async embed(texts) {
     if (!Array.isArray(texts)) throw new TypeError("embed expects an array of strings");
     if (!this.extractor) throw new Error("LocalEmbedder not initialized");
     const out = [];
     for (let i = 0; i < texts.length; i += this.batchSize) {
       const chunk = texts.slice(i, i + this.batchSize);
-      const tensor = await this.extractor(chunk, { pooling: "mean", normalize: true });
+      const tensor = await this.extractor(chunk, { pooling: this.pooling, normalize: true });
       out.push(...tensorToRows(tensor));
     }
     return out;
@@ -142,7 +173,14 @@ export class LocalEmbedder {
   }
 
   get modelHash() {
-    return modelHash(this.model);
+    // Pooling is part of the vector space: a CLS index cannot be queried with mean
+    // vectors (or vice versa). Keep the historical `model#hash` shape for `mean` so
+    // indexes that were never affected are not invalidated needlessly; every other
+    // pooling gets a fingerprint of its own, which makes the index-consistency gate
+    // flag the stale mean-space index and rebuild it.
+    return this.pooling === "mean"
+      ? modelHash(this.model)
+      : modelHash(`${this.model}@${this.pooling}`);
   }
 
   dispose() {

package/lib/settings.js

@@ -172,8 +172,10 @@ const FEATURE_FLAG_ENUMS = {
   summarizeDedupeMode: ["off", "title", "vector"],
   // Issue #126：sleep 冲突阶段的动作集（conflict 默认 = 现状；full = 六分支）。
   sleepActionSet: ["conflict", "full"],
-  // Issue #125：dream 候选集构造方式（window 默认 = 现状；hybrid 并入向量组）。
-  dreamCandidateMode: ["window", "hybrid"]
+    // Issue #125：dream 候选集构造方式（window 默认 = 现状；hybrid 并入向量组）。
+    dreamCandidateMode: ["window", "hybrid"],
+    // 本地嵌入池化：auto = 按模型族判定（BGE → cls），可显式覆盖为 cls / mean。
+    localEmbedPooling: ["auto", "cls", "mean"]
 };
 const FEATURE_FLAG_STRING_MAX = 200;
 

package/lib/store.js

@@ -1532,6 +1532,22 @@ export function createStore(path) {
     return rows.map(toDreamRun);
   }
 
+  // Issue #89：上次实际开跑时刻（epoch ms），从审计表恢复——调度器的 lastRunAt
+  // 只活在内存里，进程重启即归零，闸门对新实例放行 → 重启后立刻连发。审计表
+  // 本来就逐 run 落库（failed/degraded 也算 run），直接读它就是事实源，零迁移。
+  function lastDreamRunAt(runType = null) {
+    const row = runType
+      ? db.prepare(
+          "SELECT created_at FROM dream_runs WHERE run_type = ? ORDER BY created_at DESC, id LIMIT 1"
+        ).get(runType)
+      : db.prepare(
+          "SELECT created_at FROM dream_runs ORDER BY created_at DESC, id LIMIT 1"
+        ).get();
+    if (!row?.created_at) return 0;
+    const ms = Date.parse(row.created_at);
+    return Number.isFinite(ms) ? ms : 0;
+  }
+
   /**
    * Latest ruling-rule version seen on the audit trail. policy_epoch is a config
    * value stamped onto each run by the caller; reading the newest row's epoch
@@ -2486,6 +2502,7 @@ export function createStore(path) {
     needsEmbedding,
     searchVector,
     saveDreamRun,
+    lastDreamRunAt,
     getDreamRun,
     listDreamRuns,
     getLatestPolicyEpoch,

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@modusensus/dsh-mneme",
   "description": "Cross-session memory plugin for DeepSeek Harness with autoDream consolidation: SQLite store, Markdown mirrors, 9 model tools, automatic injection, session summarization, user profile/rules, custom slash commands, vector (semantic) search, and a Web GUI panel",
-  "version": "0.8.5",
+  "version": "0.8.6",
   "license": "MIT",
   "repository": {
     "type": "git",

package/src/config.js

@@ -248,6 +248,18 @@ export const Config = z.object({
   localEmbedDimension: z.natural().default(512),
   localEmbedDevice: z.union([z.const("cpu"), z.const("gpu")]).default("cpu"),
   localEmbedBatchSize: z.natural().min(1).max(64).default(8),
+  // 本地嵌入的池化方式（**bug 修复，不是新能力**）：BGE 系是按 CLS 池化训练的——
+  // 模型自带的 `1_Pooling/config.json` 明确写着 `pooling_mode_cls_token: true` /
+  // `pooling_mode_mean_tokens: false`，官方 README 也是「select the last hidden state of
+  // the first token」+ L2 normalize；而 transformers.js 的 `feature-extraction` 默认 mean。
+  // ⇒ 此前本地嵌入对 BGE 系一直用错池化：**不报错、只是向量系统性偏差**，检索质量静默受损。
+  // 'auto' = 按模型族判定（BGE → cls，其余 → mean，未受影响的模型行为不变）；也可显式钉住。
+  // ⚠️ 池化决定向量空间，改它会改变 modelHash（见 local-embedder.js），既有索引会被判失配并重建。
+  localEmbedPooling: z.union([
+    z.const("auto"),
+    z.const("cls"),
+    z.const("mean")
+  ]).default("auto"),
 
   // Ollama embedder.
   ollamaBaseUrl: z.string().default("http://localhost:11434"),
@@ -488,11 +500,18 @@ export const Config = z.object({
   }).default({}),
 
   // --- LLM audit trail (Bug8) ------------------------------------------------
-  // Records every background LLM call (autoDream consolidation + summary,
-  // autoSummarize compression) into llm_audit_logs: tokens, duration, status
-  // and which trigger produced it. Failures are recorded as status=error and
-  // never block the feature. retentionDays bounds the table: older rows are
-  // purged on boot.
+  // Records every background LLM call into llm_audit_logs: tokens, duration,
+  // status and which trigger produced it. Failures are recorded as
+  // status='error' and never block the feature. retentionDays bounds the table:
+  // older rows are purged on boot.
+  //
+  // Covered trigger_source values: autoDream (consolidation + summary),
+  // autoSummarize (compression), sleep (conflict + pattern) and entityExtract
+  // (issue #250 — the last two were structurally missing the hook, not a
+  // deliberate narrowing). entityExtract runs on every memory write, so row
+  // growth scales with write volume; retentionDays caps the ceiling but not the
+  // rate. If that turns out too fast, sample by operation_type or give it a
+  // separate retention — do not add a second gate: this flag is the single one.
   llmAudit: z.object({
     enabled: z.boolean().default(true),
     retentionDays: z.natural().min(1).max(3650).default(90)

package/src/dream/sleep.js

@@ -49,12 +49,22 @@ function parseJsonArray(text) {
   }
 }
 
-/** Same stream consumption contract as dream.js. */
-async function streamText(ctx, options, onStreamError) {
+/**
+ * Same stream consumption contract as dream.js.
+ *
+ * `onUsage`（optional, issue #250）receives any usage chunk for token accounting.
+ * dream.js 那份一直有它；sleep 这份副本没有，导致 conflict / pattern 两条链路的
+ * token 从未进 llm_audit_logs（面板因此只看到 autoDream 与 autoSummarize）。
+ */
+async function streamText(ctx, options, onUsage, onStreamError) {
   if (!ctx?.llm?.stream) return undefined;
   let text = "";
   for await (const chunk of ctx.llm.stream(options)) {
     if (chunk.type === "text-delta" && typeof chunk.text === "string") text += chunk.text;
+    // DSH 的 StreamChunk 契约把用量嵌在 chunk.usage（TokenUsage：inputTokens /
+    // outputTokens / totalTokens）；兼容直接平铺在 chunk 上的旧形态。归一放在
+    // 这里，调用侧的 reporter 只需认平铺形状——与 dream.js:239 同一处理。
+    if (chunk.type === "usage" && typeof onUsage === "function") onUsage(chunk.usage ?? chunk);
     if (chunk.type === "finish" && (chunk.reason?.kind === "error" || chunk.reason?.kind === "aborted")) {
       // Same rc.1 error-as-finish-chunk behavior as dream.js — surface the
       // cause instead of discarding it.
@@ -67,6 +77,79 @@ async function streamText(ctx, options, onStreamError) {
   return text;
 }
 
+/**
+ * Issue #250：给 sleep 的一条 LLM 调用记一笔 llm_audit_logs（token / 时长 / 状态 /
+ * 触发源）。与 dream.js 的 runAuditedLlm 同口径但各自独立一份——dream 那两个
+ * checker 的调用形状绑定着它自己的排程细节，硬抽共享抽象只会收敛回 dream 的形态。
+ *
+ * 记账是 best-effort：写审计行失败只 warn，绝不反噬 sleep 本体（CONTRIBUTING 的
+ * fail-safe 硬约定）。`body(reportUsage)` 负责真实消费流，usage chunk 经
+ * reportUsage 上报。
+ *
+ * 审计诚实性：流被中止/报错时 body 返回 undefined，记 status='error'；流式成功但
+ * 输出不可用时 `spec.auditError` 给出原因，同样记 error——否则同一轮调用会在
+ * 「phase 报 failed」与「audit 报 success」之间自相矛盾（CONTRIBUTING 的审计诚实性
+ * 硬约定）。`auditError` 必须是**纯判据函数**：调用方的控制流不能依赖它的副作用，
+ * 因为审计关闭时这里整段早退、钩子根本不会被调用（dream 正是那样踩的，见 PR 正文）。
+ */
+async function auditedSleepLlm(ctx, service, config, logger, spec, body) {
+  if (config?.llmAudit?.enabled === false || typeof service?.saveLlmAudit !== "function") {
+    return body(() => {});
+  }
+  const startedAt = Date.now();
+  const timestamp = new Date(startedAt).toISOString();
+  let inputTokens = 0;
+  let outputTokens = 0;
+  let status = "success";
+  let errorMessage = null;
+  try {
+    const result = await body((usage) => {
+      if (!usage) return;
+      const i = usage.input_tokens ?? usage.inputTokens ?? usage.prompt_tokens ?? usage.promptTokens;
+      const o = usage.output_tokens ?? usage.outputTokens ?? usage.completion_tokens ?? usage.completionTokens;
+      if (Number.isFinite(i)) inputTokens = i;
+      if (Number.isFinite(o)) outputTokens = o;
+    });
+    if (result === undefined) {
+      status = "error";
+      const streamErr = typeof spec.streamError === "function" ? String(spec.streamError() ?? "") : "";
+      errorMessage = streamErr ? `llm stream aborted or errored (${streamErr})` : "llm stream aborted or errored";
+    } else if (typeof spec.auditError === "function") {
+      // 流式成功但输出不可用，仍然是一次失败的调用——记 error，不与调用方的
+      // failed 返回值打架。
+      const message = spec.auditError(result);
+      if (message) {
+        status = "error";
+        errorMessage = message;
+      }
+    }
+    return result;
+  } catch (error) {
+    status = "error";
+    errorMessage = String(error?.message ?? error);
+    throw error;
+  } finally {
+    try {
+      service.saveLlmAudit({
+        timestamp,
+        trigger_source: spec.triggerSource,
+        operation_type: spec.operationType,
+        model_id: spec.modelId,
+        input_tokens: inputTokens,
+        output_tokens: outputTokens,
+        total_tokens: inputTokens + outputTokens,
+        cost_usd: 0,
+        duration_ms: Date.now() - startedAt,
+        status,
+        error_message: errorMessage,
+        related_memory_ids: spec.relatedMemoryIds ?? []
+      });
+    } catch (auditError) {
+      logger?.warn?.(`dsh-mneme sleep: llm audit write failed: ${String(auditError)}`);
+    }
+  }
+}
+
 /** LLM route (Issue #25): explicit sleepProvider/Model wins, then the dream
  *  route as a shared explicit fallback, then the agent default model. Sleep
  *  can pin a cheaper model for its bulk passes without disturbing the dream
@@ -232,7 +315,19 @@ async function phaseConflicts(ctx, service, config, logger, runId, semantic = nu
   const conflictPrompt = actionSet === "full" ? STR.prompts.conflictFull[language] : STR.prompts.conflict[language];
   const runConflict = (withEffort) => {
     conflictStreamFailure = "";
-    return streamText(ctx, {
+    return auditedSleepLlm(ctx, service, config, logger, {
+      triggerSource: "sleep",
+      operationType: "sleep_conflict",
+      modelId: `${route.provider}:${route.model}`,
+      // 候选对的双方 id：面板的活动流用 related_memory_ids 解析「沉淀条数」。
+      relatedMemoryIds: samePairs.flatMap((p) => [p.a.id, p.b.id]),
+      streamError: () => conflictStreamFailure,
+      // 纯判据：输出里没有 JSON 数组 ⇒ 这轮是失败的调用，审计记 error。**不要**把
+      // 解析结果存进闭包给下面的控制流用——审计关闭时 auditedSleepLlm 整段早退、
+      // 这个钩子不会被调用，控制流会跟着失效（dream 的既有坑）。代价是审计开启时
+      // 多解析一次，换来两条路径彻底解耦。
+      auditError: (text) => (parseJsonArray(text) ? null : "no json array in llm output")
+    }, (reportUsage) => streamText(ctx, {
     provider: route.provider,
     model: route.model,
     purpose: "sleep-conflict",
@@ -242,7 +337,7 @@ async function phaseConflicts(ctx, service, config, logger, runId, semantic = nu
       { role: "system", content: [{ type: "text", text: conflictPrompt }], source: { kind: "plugin", plugin: "dsh-mneme" } },
       { role: "user", content: [{ type: "text", text: listText }], source: { kind: "plugin", plugin: "dsh-mneme" } }
     ]
-  }, (reason) => { conflictStreamFailure = describeStreamFailure(reason); });
+  }, reportUsage, (reason) => { conflictStreamFailure = describeStreamFailure(reason); }));
   };
   const text = await withEffortFallback(ctx, sleepEffort, () => runConflict(true), () => runConflict(false), () => conflictStreamFailure);
   if (text === undefined) {
@@ -385,7 +480,16 @@ async function phasePatterns(ctx, service, config, logger, runId, signal = null)
   let patternStreamFailure = "";
   const runPattern = (withEffort) => {
     patternStreamFailure = "";
-    return streamText(ctx, {
+    return auditedSleepLlm(ctx, service, config, logger, {
+      triggerSource: "sleep",
+      operationType: "sleep_pattern",
+      modelId: `${route.provider}:${route.model}`,
+      // 本轮扫描的记忆 id：与 dream 的 relatedMemoryIds 同义，供面板解析沉淀条数。
+      relatedMemoryIds: memories.map((m) => m.id),
+      streamError: () => patternStreamFailure,
+      // 纯判据，同 conflict 阶段：不把解析结果存进闭包供控制流使用。
+      auditError: (text) => (parseJsonArray(text) ? null : "no json array in llm output")
+    }, (reportUsage) => streamText(ctx, {
     provider: route.provider,
     model: route.model,
     purpose: "sleep-pattern",
@@ -395,7 +499,7 @@ async function phasePatterns(ctx, service, config, logger, runId, signal = null)
       { role: "system", content: [{ type: "text", text: STR.prompts.pattern[language].replace("N", String(maxPatterns)) }], source: { kind: "plugin", plugin: "dsh-mneme" } },
       { role: "user", content: [{ type: "text", text: listText }], source: { kind: "plugin", plugin: "dsh-mneme" } }
     ]
-  }, (reason) => { patternStreamFailure = describeStreamFailure(reason); });
+  }, reportUsage, (reason) => { patternStreamFailure = describeStreamFailure(reason); }));
   };
   const text = await withEffortFallback(ctx, sleepEffort, () => runPattern(true), () => runPattern(false), () => patternStreamFailure);
   if (text === undefined) {
@@ -507,6 +611,9 @@ function deriveStatus(phases) {
  */
 export async function runSleep(ctx, service, config, logger, semantic = null, signal = null) {
   const runId = randomUUID();
+  // Issue #89：开跑时刻——审计行 created_at 的基准（= 冷却种子），run 耗时
+  // （LLM 各 phase）不应计入重启后的冷却窗口。
+  const sleepStartedAt = Date.now();
   const phases = {};
   const attempt = async (name, fn) => {
     if (signal?.aborted) return; // user resumed activity — stop before next phase
@@ -543,6 +650,9 @@ export async function runSleep(ctx, service, config, logger, semantic = null, si
   try {
     service.saveDreamRun({
       id: runId,
+      // Issue #89：审计行记开跑时刻而非完成时刻——run 耗时（含 LLM 各 phase）
+      // 不该计入重启后的 sleep 冷却种子。
+      created_at: new Date(sleepStartedAt).toISOString(),
       status,
       provider: route?.provider,
       model: route?.model,
@@ -582,10 +692,14 @@ export function createSleepScheduler({
   onRun = null,
   now = () => Date.now(),
   setTimeoutFn = setTimeout,
-  clearTimeoutFn = clearTimeout
+  clearTimeoutFn = clearTimeout,
+  // Issue #89 同族：冷却时刻由调用方从 dream_runs 审计表恢复（run_type='sleep'）——
+  // 内存变量进程重启即归零，冷却闸对新实例放行 → 重启后立即进入 sleep 冷却盲区。
+  // 测试注入 0（默认）即保持旧行为。
+  lastRunAtSeed = 0
 }) {
   let lastWriteAt = now();
-  let lastRunAt = 0;
+  let lastRunAt = lastRunAtSeed;
   let running = false;
   let disposed = false;
   let idleTimer = null;