@securecoms/agent-bridge 0.28.0 → 0.28.1

package/README.md

@@ -798,6 +798,23 @@ If you embed the bridge as a library (`new BridgeAgent(cfg, deps)`), semantic re
 
 ---
 
+## Release notes: 0.28.1
+
+<!-- [COPY, NEW] this section. Gate: Eric sign-off before npm publish (release PR). -->
+
+0.28.1 makes semantic memory recall more reliable. Install it in place of
+0.28.0 and restart the bridge. `@securecoms/agent-runtime` 0.11.0 is unchanged.
+
+- **After a restart.** When the model files are already on the machine, the
+  bridge loads the owner's private-chat memory at startup, so the first
+  messages after a restart get recall. Without the files, the model still
+  downloads on the owner's first private message, as before.
+- **Time limit.** The bridge now waits up to 600 ms for a typed message's
+  memory search and up to 250 ms during a voice call (it was 150 ms for both).
+- **Idle cache.** A memory cache in use is no longer dropped after 30 minutes.
+- **`/healthz`.** New `recall_cache_bytes`: the bytes of this bridge's memory
+  caches, a number only.
+
 ## Release notes: 0.28.0
 
 0.28.0 adds semantic memory in the owner's private chat. Install it together

package/dist/bin.js

@@ -30879,6 +30879,8 @@ function rankRecall(query, cache, o) {
 
 // src/memory/recall/recall-book.ts
 var DIMS = 384;
+var RECALL_QUERY_TIMEOUT_TEXT_MS = 600;
+var RECALL_QUERY_TIMEOUT_VOICE_MS = 250;
 var RecallBook = class _RecallBook {
   constructor(d) {
     this.d = d;
@@ -30932,7 +30934,6 @@ var RecallBook = class _RecallBook {
   static IDLE_MS = 30 * 6e4;
   static TOP_K = 5;
   static TEXT_MAX_CHARS = 2e3;
-  static QUERY_TIMEOUT_MS = 150;
   caches = /* @__PURE__ */ new Map();
   entries = /* @__PURE__ */ new Map();
   warming = /* @__PURE__ */ new Set();
@@ -30945,6 +30946,15 @@ var RecallBook = class _RecallBook {
   cacheWarm(conversationId) {
     return this.caches.has(conversationId);
   }
+  /**
+   * Task 27 Step 1: the bytes the warm vector caches hold (a number, content-free; /healthz `recall_cache_bytes`).
+   * The backing buffer, not the view: a warm that skipped unreadable rows installs a shorter view of a full buffer.
+   */
+  cacheBytes() {
+    let n = 0;
+    for (const c of this.caches.values()) n += c.vecs.buffer.byteLength;
+    return n;
+  }
   /**
    * §7.3 (H2): snapshot first, open outside the mutex, install under it. Never on a turn's path. Never rejects:
    * a read failure leaves the cache cold (fail closed: no recall).
@@ -31018,8 +31028,8 @@ var RecallBook = class _RecallBook {
    * §7.1: compute at ingest. Snapshot the capture FIRST (H2); `receipt`
    * replaces its purge and mode epochs with the ones captured when the turn
    * was received (Task 16 fix H-2), so a purge or a mode change after receipt
-   * refuses too. 150 ms bound on the query embed; a cold cache starts a warm
-   * and gives nothing.
+   * refuses too. `timeoutMs` bounds the query embed (the caller passes its turn kind's bound); a cold cache
+   * starts a warm and gives nothing.
    */
   async prepare(p) {
     const fresh = this.d.store.recallCapture(p.conversationId);
@@ -31036,6 +31046,7 @@ var RecallBook = class _RecallBook {
       void this.warm(p.conversationId);
       return;
     }
+    cache.lastUse = now;
     const ac = new AbortController();
     let timer;
     const q = await Promise.race([
@@ -31044,12 +31055,11 @@ var RecallBook = class _RecallBook {
         timer = setTimeout(() => {
           ac.abort();
           r(null);
-        }, p.timeoutMs ?? _RecallBook.QUERY_TIMEOUT_MS);
+        }, p.timeoutMs);
       })
     ]);
     clearTimeout(timer);
     if (!q || q.length !== DIMS) return;
-    cache.lastUse = this.nowMs();
     const ranked = rankRecall(
       q,
       { vecs: cache.vecs, dims: DIMS, messageIds: cache.messageIds, recallIds: cache.recallIds },
@@ -33147,6 +33157,13 @@ var BridgeAgent = class _BridgeAgent {
   recallStageEpochs = /* @__PURE__ */ new Map();
   /** 7b Task 16: in-flight fire-and-forget index taps; close() waits for them (bounded). */
   recallTapWork = /* @__PURE__ */ new Set();
+  /**
+   * §13.5 live check (2026-10-09, cold start): set when an apply turns recall on (the boot apply included), cleared
+   * once the maintenance tick has started the DM prewarm, or when recall is off at a tick.
+   */
+  recallPrewarmPending = false;
+  /** The DM prewarm in flight (one at a time); close() waits for it (bounded). */
+  recallPrewarmWork = null;
   /**
    * Final-review item 2: conversations whose `audience` skip was already logged in this process. A room (or any
    * conversation the owner-DM gate refuses) refuses every turn the same way, so the line is written once per
@@ -35346,7 +35363,9 @@ var BridgeAgent = class _BridgeAgent {
         await this.notes.exclusive(async () => {
           if (this.store.applyRecallMode("semantic", mem.recallDays).purged > 0) this.onRecallRetentionChangedLocked();
         });
+        const turnedOn = !this.recallConfig.on;
         this.recallConfig = { on: true };
+        if (turnedOn) this.armRecallPrewarm();
         if (this.cfg.modelSource === "download" && !this.modelDownloadNoticeLogged) {
           this.modelDownloadNoticeLogged = true;
           writeServiceLog("info", "memory.model_download_enabled");
@@ -35367,6 +35386,54 @@ var BridgeAgent = class _BridgeAgent {
       writeServiceLog("warn", "memory.recall_mode_failed", {}, e);
     }
   }
+  /**
+   * §13.5 live check (2026-10-09, cold start). After a restart in semantic mode turn 1 got no recall (the embedder
+   * was not ready: ensureStarted, return) and turn 2 none either (no cache: a background warm, return). So when an
+   * apply turns recall on (the boot apply included) and the model files are on disk, the embedder is started here,
+   * off the turn path; the
+   * maintenance tick warms the DM caches once it is ready (runRecallPrewarm). Synchronous and never throws:
+   * ensureStarted only schedules the start, so boot never waits for it.
+   */
+  armRecallPrewarm() {
+    try {
+      if (!this.recallProviderOn()) return;
+      if (!this.embedder.filesPresent()) return;
+      this.recallPrewarmPending = true;
+      this.embedder.ensureStarted();
+    } catch (e) {
+      writeServiceLog("warn", "memory.recall_prewarm_failed", {}, e);
+    }
+  }
+  /**
+   * The maintenance tick's half of the cold-start prewarm: once the embedder is ready (no cache RAM is spent on an
+   * embedder that may never come up), warm each DM conversation, one at a time, detached from the tick. DM is the
+   * persisted is_dm === true, the rule the membership sweep uses (audit r1 M6); a vault channel is never one. The
+   * prewarm only opens vectors through RecallBook.warm, which snapshots the capture first and installs nothing
+   * across a clear, a mode change or a purge (H2, M-A3). It indexes nothing and prepares nothing. Recall turned off
+   * before or during it stops it (I1, Q1). Never throws.
+   */
+  runRecallPrewarm() {
+    if (!this.recallPrewarmPending || this.recallPrewarmWork || this.closingDown) return;
+    if (this.embedder.state().state !== "ready") return;
+    this.recallPrewarmPending = false;
+    this.recallPrewarmWork = (async () => {
+      try {
+        const dms = this.store.memoryConversationIds().filter((c) => !this.vaultChannelIds.has(c) && this.store.getChannelMeta(c)?.isDm === true);
+        for (const c of dms) {
+          if (this.closingDown || !this.recallProviderOn()) return;
+          await this.recall.warm(c);
+        }
+      } catch (e) {
+        writeServiceLog("warn", "memory.recall_prewarm_failed", {}, e);
+      }
+    })().finally(() => {
+      this.recallPrewarmWork = null;
+    });
+  }
+  /** @internal test seam: resolves once the DM prewarm in flight, if any, has settled. */
+  async settleRecallPrewarm() {
+    await this.recallPrewarmWork;
+  }
   /** A retention change or a prune removed rows: drop the vector caches. Called only inside notes.exclusive. */
   onRecallRetentionChangedLocked() {
     this.recall.dropAllCachesLocked();
@@ -35376,8 +35443,8 @@ var BridgeAgent = class _BridgeAgent {
     this.recall.dropAllLocked();
   }
   /**
-   * 7b spec §7.1, §7.2: Q1 (config half) and Q2 here, then the 150 ms bound
-   * inside RecallBook.prepare. Called beside prepareNotesForTurn, before the
+   * 7b spec §7.1, §7.2: Q1 (config half) and Q2 here, then the turn kind's
+   * bound inside RecallBook.prepare (600 ms text, 250 ms voice; ruling 2026-10-09). Called beside prepareNotesForTurn, before the
    * turn is staged. The receipt epochs (Task 16 fix H-2) narrow the capture:
    * the in-cycle receipt, else the persisted staging marks; marks that exist
    * but lack an epoch give no recall (fail closed). Never throws.
@@ -35389,7 +35456,8 @@ var BridgeAgent = class _BridgeAgent {
       if (marks !== void 0 && (marks.purgeEpoch === void 0 || marks.modeEpoch === void 0)) return;
       const receipt = marks === void 0 ? void 0 : { purgeEpoch: marks.purgeEpoch, modeEpoch: marks.modeEpoch };
       const exclude = this.store.recentMemoryMessageIds(t.channelId, 50);
-      await this.recall.prepare({ messageId: t.messageId, conversationId: t.channelId, query: t.text, excludeIds: exclude, ...receipt ? { receipt } : {} });
+      const timeoutMs = t.kind === "voice" ? RECALL_QUERY_TIMEOUT_VOICE_MS : RECALL_QUERY_TIMEOUT_TEXT_MS;
+      await this.recall.prepare({ messageId: t.messageId, conversationId: t.channelId, query: t.text, excludeIds: exclude, timeoutMs, ...receipt ? { receipt } : {} });
     } catch (e) {
       writeServiceLog("warn", "memory.recall_prepare_failed", {}, e);
     }
@@ -35551,8 +35619,10 @@ var BridgeAgent = class _BridgeAgent {
     }
     if (!this.recallProviderOn()) {
       this.reembedHot = false;
+      this.recallPrewarmPending = false;
       return;
     }
+    this.runRecallPrewarm();
     if (!this.reembedHot && n % 60 !== 0) return;
     try {
       this.reembedHot = await this.reembed.tick() > 0;
@@ -37344,7 +37414,10 @@ var BridgeAgent = class _BridgeAgent {
             const turnText = parsed.text;
             const isCallStart = turnText.trim() === VOICE_CALL_START_MARKER;
             const coalescible = this.voiceCoalesceWindowMs > 0 && m.voiceMode === true && speaker !== void 0 && parsed.transcribed === true && voiceGuidance !== void 0 && !isCallStart;
-            if (!coalescible) await this.prepareRecallForTurn({ channelId: m.channelId, messageId: m.messageId, text: parsed.text, authorization: turnAuthorization });
+            if (!coalescible) {
+              const kind2 = m.voiceMode === true || parsed.transcribed === true ? "voice" : "text";
+              await this.prepareRecallForTurn({ channelId: m.channelId, messageId: m.messageId, text: parsed.text, authorization: turnAuthorization, kind: kind2 });
+            }
             let fastPathRoomBlocked = false;
             if (!coalescible && fastPathEligible && meta && meta.is_dm === false) {
               const reserved = this.fastReplyReserved.get(m.channelId) ?? 0;
@@ -37428,7 +37501,7 @@ var BridgeAgent = class _BridgeAgent {
                 const [firstId, ...tailIds] = win.ids;
                 if (tailIds.length) this.voiceCoalescedTails.set(firstId, tailIds);
                 const text = win.parts.join(" ");
-                await this.prepareRecallForTurn({ channelId, messageId: firstId, text, authorization: turnAuthorization });
+                await this.prepareRecallForTurn({ channelId, messageId: firstId, text, authorization: turnAuthorization, kind: "voice" });
                 const merged = { ...held, text };
                 void recordInboundMemory(this.memory, channelId, senderId, text, firstId, "voice", win.purgeEpoch);
                 let roomBlocked = false;
@@ -38174,6 +38247,7 @@ var BridgeAgent = class _BridgeAgent {
       retired_drained: this.provider.retiredDrainedCount(),
       // INV-CUSTODY-DURABLE §5.C.3: custody health.
       expired_drained: this.provider.expiredDrainedCount(),
+      recall_cache_bytes: this.recall.cacheBytes(),
       custody_persist_failing: this.provider.custodyHealth().persistFailing,
       custody_persist_failures: this.provider.custodyHealth().persistFailures,
       custody_read_failing: this.provider.custodyHealth().readFailing,
@@ -38307,6 +38381,16 @@ var BridgeAgent = class _BridgeAgent {
       ]);
       if (timer) clearTimeout(timer);
     }
+    if (this.recallPrewarmWork) {
+      let timer;
+      await Promise.race([
+        this.recallPrewarmWork,
+        new Promise((resolve3) => {
+          timer = setTimeout(resolve3, _BridgeAgent.GRACEFUL_DRAIN_MS);
+        })
+      ]);
+      if (timer) clearTimeout(timer);
+    }
     this.embedder.close?.();
     if (this.replyWork.size) {
       let timer;

package/dist/bridge-agent.d.ts

@@ -1192,6 +1192,13 @@ export declare class BridgeAgent {
     private readonly recallStageEpochs;
     /** 7b Task 16: in-flight fire-and-forget index taps; close() waits for them (bounded). */
     private readonly recallTapWork;
+    /**
+     * §13.5 live check (2026-10-09, cold start): set when an apply turns recall on (the boot apply included), cleared
+     * once the maintenance tick has started the DM prewarm, or when recall is off at a tick.
+     */
+    private recallPrewarmPending;
+    /** The DM prewarm in flight (one at a time); close() waits for it (bounded). */
+    private recallPrewarmWork;
     /**
      * Final-review item 2: conversations whose `audience` skip was already logged in this process. A room (or any
      * conversation the owner-DM gate refuses) refuses every turn the same way, so the line is written once per
@@ -2001,13 +2008,33 @@ export declare class BridgeAgent {
     private applyPayloadMemory;
     /** Spec §5.3 (H3): mode state lives in the store; level-triggered after every refetch. Never throws. */
     private applyRecallModeFromPayload;
+    /**
+     * §13.5 live check (2026-10-09, cold start). After a restart in semantic mode turn 1 got no recall (the embedder
+     * was not ready: ensureStarted, return) and turn 2 none either (no cache: a background warm, return). So when an
+     * apply turns recall on (the boot apply included) and the model files are on disk, the embedder is started here,
+     * off the turn path; the
+     * maintenance tick warms the DM caches once it is ready (runRecallPrewarm). Synchronous and never throws:
+     * ensureStarted only schedules the start, so boot never waits for it.
+     */
+    private armRecallPrewarm;
+    /**
+     * The maintenance tick's half of the cold-start prewarm: once the embedder is ready (no cache RAM is spent on an
+     * embedder that may never come up), warm each DM conversation, one at a time, detached from the tick. DM is the
+     * persisted is_dm === true, the rule the membership sweep uses (audit r1 M6); a vault channel is never one. The
+     * prewarm only opens vectors through RecallBook.warm, which snapshots the capture first and installs nothing
+     * across a clear, a mode change or a purge (H2, M-A3). It indexes nothing and prepares nothing. Recall turned off
+     * before or during it stops it (I1, Q1). Never throws.
+     */
+    private runRecallPrewarm;
+    /** @internal test seam: resolves once the DM prewarm in flight, if any, has settled. */
+    settleRecallPrewarm(): Promise<void>;
     /** A retention change or a prune removed rows: drop the vector caches. Called only inside notes.exclusive. */
     private onRecallRetentionChangedLocked;
     /** Markdown mode deleted every row: drop the caches and the per-message map. Called only inside notes.exclusive. */
     private onRecallPurgedAllLocked;
     /**
-     * 7b spec §7.1, §7.2: Q1 (config half) and Q2 here, then the 150 ms bound
-     * inside RecallBook.prepare. Called beside prepareNotesForTurn, before the
+     * 7b spec §7.1, §7.2: Q1 (config half) and Q2 here, then the turn kind's
+     * bound inside RecallBook.prepare (600 ms text, 250 ms voice; ruling 2026-10-09). Called beside prepareNotesForTurn, before the
      * turn is staged. The receipt epochs (Task 16 fix H-2) narrow the capture:
      * the in-cycle receipt, else the persisted staging marks; marks that exist
      * but lack an epoch give no recall (fail closed). Never throws.
@@ -2877,6 +2904,9 @@ export declare class BridgeAgent {
         retired_drained: number;
         /** INV-CUSTODY-DURABLE §5.C.6: rows silently drained past expires_at. */
         expired_drained: number;
+        /** 7b Task 27 Step 1: bytes held by this bridge's warm recall vector caches. A number, content-free; the fleet
+         *  RAM script (scripts/ops/semantic-memory-ram-budget.mjs) sums it with the shared embedder's RSS. */
+        recall_cache_bytes: number;
         /** INV-CUSTODY-DURABLE §5.C.3: a decrypted, un-acked row is memory-only
          *  (its durable write failed); a restart now would lose it. */
         custody_persist_failing: boolean;

package/dist/index.js

@@ -30369,6 +30369,8 @@ function rankRecall(query, cache, o) {
 
 // src/memory/recall/recall-book.ts
 var DIMS = 384;
+var RECALL_QUERY_TIMEOUT_TEXT_MS = 600;
+var RECALL_QUERY_TIMEOUT_VOICE_MS = 250;
 var RecallBook = class _RecallBook {
   constructor(d) {
     this.d = d;
@@ -30422,7 +30424,6 @@ var RecallBook = class _RecallBook {
   static IDLE_MS = 30 * 6e4;
   static TOP_K = 5;
   static TEXT_MAX_CHARS = 2e3;
-  static QUERY_TIMEOUT_MS = 150;
   caches = /* @__PURE__ */ new Map();
   entries = /* @__PURE__ */ new Map();
   warming = /* @__PURE__ */ new Set();
@@ -30435,6 +30436,15 @@ var RecallBook = class _RecallBook {
   cacheWarm(conversationId) {
     return this.caches.has(conversationId);
   }
+  /**
+   * Task 27 Step 1: the bytes the warm vector caches hold (a number, content-free; /healthz `recall_cache_bytes`).
+   * The backing buffer, not the view: a warm that skipped unreadable rows installs a shorter view of a full buffer.
+   */
+  cacheBytes() {
+    let n = 0;
+    for (const c of this.caches.values()) n += c.vecs.buffer.byteLength;
+    return n;
+  }
   /**
    * §7.3 (H2): snapshot first, open outside the mutex, install under it. Never on a turn's path. Never rejects:
    * a read failure leaves the cache cold (fail closed: no recall).
@@ -30508,8 +30518,8 @@ var RecallBook = class _RecallBook {
    * §7.1: compute at ingest. Snapshot the capture FIRST (H2); `receipt`
    * replaces its purge and mode epochs with the ones captured when the turn
    * was received (Task 16 fix H-2), so a purge or a mode change after receipt
-   * refuses too. 150 ms bound on the query embed; a cold cache starts a warm
-   * and gives nothing.
+   * refuses too. `timeoutMs` bounds the query embed (the caller passes its turn kind's bound); a cold cache
+   * starts a warm and gives nothing.
    */
   async prepare(p) {
     const fresh = this.d.store.recallCapture(p.conversationId);
@@ -30526,6 +30536,7 @@ var RecallBook = class _RecallBook {
       void this.warm(p.conversationId);
       return;
     }
+    cache.lastUse = now;
     const ac = new AbortController();
     let timer;
     const q = await Promise.race([
@@ -30534,12 +30545,11 @@ var RecallBook = class _RecallBook {
         timer = setTimeout(() => {
           ac.abort();
           r(null);
-        }, p.timeoutMs ?? _RecallBook.QUERY_TIMEOUT_MS);
+        }, p.timeoutMs);
       })
     ]);
     clearTimeout(timer);
     if (!q || q.length !== DIMS) return;
-    cache.lastUse = this.nowMs();
     const ranked = rankRecall(
       q,
       { vecs: cache.vecs, dims: DIMS, messageIds: cache.messageIds, recallIds: cache.recallIds },
@@ -32897,6 +32907,13 @@ var BridgeAgent = class _BridgeAgent {
   recallStageEpochs = /* @__PURE__ */ new Map();
   /** 7b Task 16: in-flight fire-and-forget index taps; close() waits for them (bounded). */
   recallTapWork = /* @__PURE__ */ new Set();
+  /**
+   * §13.5 live check (2026-10-09, cold start): set when an apply turns recall on (the boot apply included), cleared
+   * once the maintenance tick has started the DM prewarm, or when recall is off at a tick.
+   */
+  recallPrewarmPending = false;
+  /** The DM prewarm in flight (one at a time); close() waits for it (bounded). */
+  recallPrewarmWork = null;
   /**
    * Final-review item 2: conversations whose `audience` skip was already logged in this process. A room (or any
    * conversation the owner-DM gate refuses) refuses every turn the same way, so the line is written once per
@@ -35096,7 +35113,9 @@ var BridgeAgent = class _BridgeAgent {
         await this.notes.exclusive(async () => {
           if (this.store.applyRecallMode("semantic", mem.recallDays).purged > 0) this.onRecallRetentionChangedLocked();
         });
+        const turnedOn = !this.recallConfig.on;
         this.recallConfig = { on: true };
+        if (turnedOn) this.armRecallPrewarm();
         if (this.cfg.modelSource === "download" && !this.modelDownloadNoticeLogged) {
           this.modelDownloadNoticeLogged = true;
           writeServiceLog2("info", "memory.model_download_enabled");
@@ -35117,6 +35136,54 @@ var BridgeAgent = class _BridgeAgent {
       writeServiceLog2("warn", "memory.recall_mode_failed", {}, e);
     }
   }
+  /**
+   * §13.5 live check (2026-10-09, cold start). After a restart in semantic mode turn 1 got no recall (the embedder
+   * was not ready: ensureStarted, return) and turn 2 none either (no cache: a background warm, return). So when an
+   * apply turns recall on (the boot apply included) and the model files are on disk, the embedder is started here,
+   * off the turn path; the
+   * maintenance tick warms the DM caches once it is ready (runRecallPrewarm). Synchronous and never throws:
+   * ensureStarted only schedules the start, so boot never waits for it.
+   */
+  armRecallPrewarm() {
+    try {
+      if (!this.recallProviderOn()) return;
+      if (!this.embedder.filesPresent()) return;
+      this.recallPrewarmPending = true;
+      this.embedder.ensureStarted();
+    } catch (e) {
+      writeServiceLog2("warn", "memory.recall_prewarm_failed", {}, e);
+    }
+  }
+  /**
+   * The maintenance tick's half of the cold-start prewarm: once the embedder is ready (no cache RAM is spent on an
+   * embedder that may never come up), warm each DM conversation, one at a time, detached from the tick. DM is the
+   * persisted is_dm === true, the rule the membership sweep uses (audit r1 M6); a vault channel is never one. The
+   * prewarm only opens vectors through RecallBook.warm, which snapshots the capture first and installs nothing
+   * across a clear, a mode change or a purge (H2, M-A3). It indexes nothing and prepares nothing. Recall turned off
+   * before or during it stops it (I1, Q1). Never throws.
+   */
+  runRecallPrewarm() {
+    if (!this.recallPrewarmPending || this.recallPrewarmWork || this.closingDown) return;
+    if (this.embedder.state().state !== "ready") return;
+    this.recallPrewarmPending = false;
+    this.recallPrewarmWork = (async () => {
+      try {
+        const dms = this.store.memoryConversationIds().filter((c) => !this.vaultChannelIds.has(c) && this.store.getChannelMeta(c)?.isDm === true);
+        for (const c of dms) {
+          if (this.closingDown || !this.recallProviderOn()) return;
+          await this.recall.warm(c);
+        }
+      } catch (e) {
+        writeServiceLog2("warn", "memory.recall_prewarm_failed", {}, e);
+      }
+    })().finally(() => {
+      this.recallPrewarmWork = null;
+    });
+  }
+  /** @internal test seam: resolves once the DM prewarm in flight, if any, has settled. */
+  async settleRecallPrewarm() {
+    await this.recallPrewarmWork;
+  }
   /** A retention change or a prune removed rows: drop the vector caches. Called only inside notes.exclusive. */
   onRecallRetentionChangedLocked() {
     this.recall.dropAllCachesLocked();
@@ -35126,8 +35193,8 @@ var BridgeAgent = class _BridgeAgent {
     this.recall.dropAllLocked();
   }
   /**
-   * 7b spec §7.1, §7.2: Q1 (config half) and Q2 here, then the 150 ms bound
-   * inside RecallBook.prepare. Called beside prepareNotesForTurn, before the
+   * 7b spec §7.1, §7.2: Q1 (config half) and Q2 here, then the turn kind's
+   * bound inside RecallBook.prepare (600 ms text, 250 ms voice; ruling 2026-10-09). Called beside prepareNotesForTurn, before the
    * turn is staged. The receipt epochs (Task 16 fix H-2) narrow the capture:
    * the in-cycle receipt, else the persisted staging marks; marks that exist
    * but lack an epoch give no recall (fail closed). Never throws.
@@ -35139,7 +35206,8 @@ var BridgeAgent = class _BridgeAgent {
       if (marks !== void 0 && (marks.purgeEpoch === void 0 || marks.modeEpoch === void 0)) return;
       const receipt = marks === void 0 ? void 0 : { purgeEpoch: marks.purgeEpoch, modeEpoch: marks.modeEpoch };
       const exclude = this.store.recentMemoryMessageIds(t.channelId, 50);
-      await this.recall.prepare({ messageId: t.messageId, conversationId: t.channelId, query: t.text, excludeIds: exclude, ...receipt ? { receipt } : {} });
+      const timeoutMs = t.kind === "voice" ? RECALL_QUERY_TIMEOUT_VOICE_MS : RECALL_QUERY_TIMEOUT_TEXT_MS;
+      await this.recall.prepare({ messageId: t.messageId, conversationId: t.channelId, query: t.text, excludeIds: exclude, timeoutMs, ...receipt ? { receipt } : {} });
     } catch (e) {
       writeServiceLog2("warn", "memory.recall_prepare_failed", {}, e);
     }
@@ -35301,8 +35369,10 @@ var BridgeAgent = class _BridgeAgent {
     }
     if (!this.recallProviderOn()) {
       this.reembedHot = false;
+      this.recallPrewarmPending = false;
       return;
     }
+    this.runRecallPrewarm();
     if (!this.reembedHot && n % 60 !== 0) return;
     try {
       this.reembedHot = await this.reembed.tick() > 0;
@@ -37094,7 +37164,10 @@ var BridgeAgent = class _BridgeAgent {
             const turnText = parsed.text;
             const isCallStart = turnText.trim() === VOICE_CALL_START_MARKER;
             const coalescible = this.voiceCoalesceWindowMs > 0 && m.voiceMode === true && speaker !== void 0 && parsed.transcribed === true && voiceGuidance !== void 0 && !isCallStart;
-            if (!coalescible) await this.prepareRecallForTurn({ channelId: m.channelId, messageId: m.messageId, text: parsed.text, authorization: turnAuthorization });
+            if (!coalescible) {
+              const kind2 = m.voiceMode === true || parsed.transcribed === true ? "voice" : "text";
+              await this.prepareRecallForTurn({ channelId: m.channelId, messageId: m.messageId, text: parsed.text, authorization: turnAuthorization, kind: kind2 });
+            }
             let fastPathRoomBlocked = false;
             if (!coalescible && fastPathEligible && meta && meta.is_dm === false) {
               const reserved = this.fastReplyReserved.get(m.channelId) ?? 0;
@@ -37178,7 +37251,7 @@ var BridgeAgent = class _BridgeAgent {
                 const [firstId, ...tailIds] = win.ids;
                 if (tailIds.length) this.voiceCoalescedTails.set(firstId, tailIds);
                 const text = win.parts.join(" ");
-                await this.prepareRecallForTurn({ channelId, messageId: firstId, text, authorization: turnAuthorization });
+                await this.prepareRecallForTurn({ channelId, messageId: firstId, text, authorization: turnAuthorization, kind: "voice" });
                 const merged = { ...held, text };
                 void recordInboundMemory(this.memory, channelId, senderId, text, firstId, "voice", win.purgeEpoch);
                 let roomBlocked = false;
@@ -37924,6 +37997,7 @@ var BridgeAgent = class _BridgeAgent {
       retired_drained: this.provider.retiredDrainedCount(),
       // INV-CUSTODY-DURABLE §5.C.3: custody health.
       expired_drained: this.provider.expiredDrainedCount(),
+      recall_cache_bytes: this.recall.cacheBytes(),
       custody_persist_failing: this.provider.custodyHealth().persistFailing,
       custody_persist_failures: this.provider.custodyHealth().persistFailures,
       custody_read_failing: this.provider.custodyHealth().readFailing,
@@ -38057,6 +38131,16 @@ var BridgeAgent = class _BridgeAgent {
       ]);
       if (timer) clearTimeout(timer);
     }
+    if (this.recallPrewarmWork) {
+      let timer;
+      await Promise.race([
+        this.recallPrewarmWork,
+        new Promise((resolve) => {
+          timer = setTimeout(resolve, _BridgeAgent.GRACEFUL_DRAIN_MS);
+        })
+      ]);
+      if (timer) clearTimeout(timer);
+    }
     this.embedder.close?.();
     if (this.replyWork.size) {
       let timer;

package/dist/memory/recall/recall-book.d.ts

@@ -9,6 +9,14 @@
 import type { RecallCapture, RecallRow, RuntimeInboxRewrite, SqliteMlsStateStore } from "@agentvault/connect-mls-store";
 import { type ServiceLog } from "../../service-log.js";
 import type { Embedder } from "../embedder/client.js";
+/**
+ * Spec §7.1, §20 E1.4 (architect ruling, owner-approved 2026-10-09): the per-turn-kind wall-clock bound on the
+ * query embed. A typed turn waits up to 600 ms; a coalesced voice turn at window close up to 250 ms. A result
+ * after the bound is discarded (fail closed: no recall), never attached late. The cutoff cancels the job at the
+ * shared embedder. Each kind's value is this one line.
+ */
+export declare const RECALL_QUERY_TIMEOUT_TEXT_MS = 600;
+export declare const RECALL_QUERY_TIMEOUT_VOICE_MS = 250;
 export type IndexSkip = "mode" | "audience" | "voice" | "length" | "secret" | "stamp_missing" | "cleared" | "epoch" | "purged" | "duplicate" | "store";
 export interface RecallTapCapture {
     readonly generation: number | undefined;
@@ -54,13 +62,17 @@ export declare class RecallBook {
     private static readonly IDLE_MS;
     private static readonly TOP_K;
     private static readonly TEXT_MAX_CHARS;
-    private static readonly QUERY_TIMEOUT_MS;
     private readonly caches;
     private readonly entries;
     private readonly warming;
     private same;
     private nowMs;
     cacheWarm(conversationId: string): boolean;
+    /**
+     * Task 27 Step 1: the bytes the warm vector caches hold (a number, content-free; /healthz `recall_cache_bytes`).
+     * The backing buffer, not the view: a warm that skipped unreadable rows installs a shorter view of a full buffer.
+     */
+    cacheBytes(): number;
     /**
      * §7.3 (H2): snapshot first, open outside the mutex, install under it. Never on a turn's path. Never rejects:
      * a read failure leaves the cache cold (fail closed: no recall).
@@ -81,15 +93,15 @@ export declare class RecallBook {
      * §7.1: compute at ingest. Snapshot the capture FIRST (H2); `receipt`
      * replaces its purge and mode epochs with the ones captured when the turn
      * was received (Task 16 fix H-2), so a purge or a mode change after receipt
-     * refuses too. 150 ms bound on the query embed; a cold cache starts a warm
-     * and gives nothing.
+     * refuses too. `timeoutMs` bounds the query embed (the caller passes its turn kind's bound); a cold cache
+     * starts a warm and gives nothing.
      */
     prepare(p: {
         messageId: string;
         conversationId: string;
         query: string;
         excludeIds: ReadonlySet<string>;
-        timeoutMs?: number;
+        timeoutMs: number;
         receipt?: {
             purgeEpoch: number;
             modeEpoch: number;

package/dist/version.d.ts

@@ -1 +1 @@
-export declare const BRIDGE_VERSION = "0.28.0";
+export declare const BRIDGE_VERSION = "0.28.1";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@securecoms/agent-bridge",
-  "version": "0.28.0",
+  "version": "0.28.1",
   "type": "module",
   "engines": {
     "node": ">=22.5.0"