@desplega.ai/agent-swarm 1.92.0 → 1.92.2

package/src/be/seed-scripts/index.ts

@@ -22,6 +22,9 @@ import { getScript, upsertScriptByName } from "../scripts/db";
 import { extractArgsJsonSchema } from "../scripts/extract-schema";
 import { typecheckScript } from "../scripts/typecheck";
 import type { Seeder, SeedItem } from "../seed/types";
+import bootTriageSrc from "./catalog/boot-triage.ts" with { type: "text" };
+// @ts-expect-error Bun text imports return the raw source string for bundling standalone scripts.
+import catalogReportSrc from "./catalog/catalog-report.ts" with { type: "text" };
 import compoundInsightsSrc from "./catalog/compound-insights.ts" with { type: "text" };
 import dateResolveSrc from "./catalog/date-resolve.ts" with { type: "text" };
 import fetchReadableSrc from "./catalog/fetch-readable.ts" with { type: "text" };
@@ -30,6 +33,7 @@ import groupCountSrc from "./catalog/group-count.ts" with { type: "text" };
 import jsonQuerySrc from "./catalog/json-query.ts" with { type: "text" };
 import linearIssueSrc from "./catalog/linear-issue.ts" with { type: "text" };
 import memoryDedupCheckSrc from "./catalog/memory-dedup-check.ts" with { type: "text" };
+import memoryEvalSrc from "./catalog/memory-eval.ts" with { type: "text" };
 import opsCatalogAuditSrc from "./catalog/ops-catalog-audit.ts" with { type: "text" };
 import scheduleHealthSrc from "./catalog/schedule-health.ts" with { type: "text" };
 import slackThreadFlattenSrc from "./catalog/slack-thread-flatten.ts" with { type: "text" };
@@ -50,6 +54,14 @@ export type SeedScript = {
 // module's default export, so the cast restores the real shape.
 const asText = (s: unknown): string => s as string;
 
+const CATALOG_REPORT_IMPORT_RE = /^import\s+\{[^}]*\}\s+from "\.\/catalog-report";\n\n?/m;
+
+function bundleCatalogReport(source: string): string {
+  const helper = asText(catalogReportSrc);
+  if (!CATALOG_REPORT_IMPORT_RE.test(source)) return source;
+  return `${helper}\n\n${source.replace(CATALOG_REPORT_IMPORT_RE, "")}`;
+}
+
 export const SEED_SCRIPTS: SeedScript[] = [
   {
     name: "gh-pr-snapshot",
@@ -103,7 +115,7 @@ export const SEED_SCRIPTS: SeedScript[] = [
       "Scan recently failed swarm tasks and cluster them by failure reason, agent or schedule to surface recurring problems.",
     intent:
       "Find patterns in swarm task failures — which agent, schedule or error keeps breaking — for a reliability review.",
-    source: asText(taskFailureAuditSrc),
+    source: bundleCatalogReport(asText(taskFailureAuditSrc)),
   },
   {
     name: "memory-dedup-check",
@@ -150,7 +162,7 @@ export const SEED_SCRIPTS: SeedScript[] = [
       "Per-schedule failure rate check over recent tasks — flags schedules with failure rates above a configurable threshold.",
     intent:
       "Find unhealthy schedules that keep failing — for daily compounding, reliability reviews, or ops triage.",
-    source: asText(scheduleHealthSrc),
+    source: bundleCatalogReport(asText(scheduleHealthSrc)),
   },
   {
     name: "tool-usage",
@@ -166,7 +178,15 @@ export const SEED_SCRIPTS: SeedScript[] = [
       "All-in-one swarm-wide daily ops snapshot: task completion/failure summary, real failure clusters (excludes superseded/cancelled bookkeeping), schedule health flags, tool usage top-25, memory health/pollution stats, seed-script candidate tool triplets, and a per-agent breakdown. Aggregates across ALL agents via direct read-only SQL.",
     intent:
       "Single-call daily compounding Phase 0 helper — replaces ~25 raw tool roundtrips with one compressed JSON result covering every agent. For daily evolution, self-scripting candidates, ops reviews, or heartbeat context.",
-    source: asText(compoundInsightsSrc),
+    source: bundleCatalogReport(asText(compoundInsightsSrc)),
+  },
+  {
+    name: "memory-eval",
+    description:
+      "3-axis memory quality evaluation: carry-forward context (do follow-up tasks retrieve useful memories from prior tasks?), follow preferences (are CLAUDE.md/IDENTITY.md/SOUL.md/TOOLS.md memories retrieved and useful?), and stay current (what fraction of retrieved memories are fresh vs stale?). Outputs a baseline report to agent-fs + a swarm Page.",
+    intent:
+      "Measure memory system health across OpenAI Dreaming-inspired axes — before/after baseline for architecture changes, blog-post numbers, daily quality monitoring.",
+    source: asText(memoryEvalSrc),
   },
   {
     name: "ops-catalog-audit",
@@ -174,7 +194,15 @@ export const SEED_SCRIPTS: SeedScript[] = [
       "Audit-as-code catalog check for schedules, workflows, and prompt/template drift. Clusters actionable findings by goal and can publish an authed HTML report page.",
     intent:
       "Re-run the ops inventory audit in one call: duplicate/dead schedules, code-work routing risks, enabled workflow fixtures, structured-output gate gaps, prompt registry drift, stale hosts, and systemDefault skill duplicates.",
-    source: asText(opsCatalogAuditSrc),
+    source: bundleCatalogReport(asText(opsCatalogAuditSrc)),
+  },
+  {
+    name: "boot-triage",
+    description:
+      "Post-restart heartbeat triage snapshot: deploy restart PR context, recent real failures, stuck offline-agent work, orphaned tasks, and superseded tasks missing resume children.",
+    intent:
+      "Run immediately after a swarm restart to gather deterministic boot triage data in one read-only call before the Lead decides what to retry, cancel, or escalate.",
+    source: asText(bootTriageSrc),
   },
 ];
 

package/src/be/seed-skills/index.ts

@@ -25,12 +25,6 @@ import kvStorageContent from "../../../templates/skills/kv-storage/content.md" w
 };
 import pagesConfig from "../../../templates/skills/pages/config.json" with { type: "text" };
 import pagesContent from "../../../templates/skills/pages/content.md" with { type: "text" };
-import scheduledTaskResilienceConfig from "../../../templates/skills/scheduled-task-resilience/config.json" with {
-  type: "text",
-};
-import scheduledTaskResilienceContent from "../../../templates/skills/scheduled-task-resilience/content.md" with {
-  type: "text",
-};
 import scriptWorkflowsConfig from "../../../templates/skills/script-workflows/config.json" with {
   type: "text",
 };
@@ -77,7 +71,6 @@ const BUILT_IN_SKILL_SOURCES = [
   { config: artifactsConfig, body: artifactsContent },
   { config: kvStorageConfig, body: kvStorageContent },
   { config: pagesConfig, body: pagesContent },
-  { config: scheduledTaskResilienceConfig, body: scheduledTaskResilienceContent },
   { config: scriptWorkflowsConfig, body: scriptWorkflowsContent },
   { config: swarmScriptsConfig, body: swarmScriptsContent },
   { config: workflowIterateConfig, body: workflowIterateContent },

package/src/be/skill-sync.ts

@@ -8,10 +8,11 @@
  * This runs on the API side — workers call it via POST /api/skills/sync-filesystem.
  */
 
+import type { Dirent } from "node:fs";
 import { existsSync, mkdirSync, readdirSync, rmSync, writeFileSync } from "node:fs";
 import { homedir } from "node:os";
-import { join } from "node:path";
-import { getAgentSkills } from "./db";
+import { dirname, join } from "node:path";
+import { getAgentSkills, getSkillFiles } from "./db";
 
 export interface SkillSyncResult {
   synced: number;
@@ -29,11 +30,68 @@ export interface SkillSyncResult {
  */
 const SWARM_MARKER_FILE = ".swarm-managed";
 
+function reconcileManagedSkillFiles(skillDir: string, currentRelativeFiles: Set<string>): number {
+  if (!existsSync(join(skillDir, SWARM_MARKER_FILE))) return 0;
+
+  let removed = 0;
+
+  const walk = (dir: string, relativeDir = ""): boolean => {
+    let entries: Dirent[];
+    try {
+      entries = readdirSync(dir, { withFileTypes: true });
+    } catch {
+      return false;
+    }
+
+    let hasEntries = false;
+    for (const entry of entries) {
+      const relativePath = relativeDir ? `${relativeDir}/${entry.name}` : entry.name;
+      const fullPath = join(dir, entry.name);
+
+      if (entry.isDirectory()) {
+        const childHasEntries = walk(fullPath, relativePath);
+        if (!childHasEntries) {
+          try {
+            rmSync(fullPath, { recursive: true, force: true });
+          } catch {
+            hasEntries = true;
+          }
+        } else {
+          hasEntries = true;
+        }
+        continue;
+      }
+
+      if (
+        relativePath === "SKILL.md" ||
+        relativePath === SWARM_MARKER_FILE ||
+        currentRelativeFiles.has(relativePath)
+      ) {
+        hasEntries = true;
+        continue;
+      }
+
+      try {
+        rmSync(fullPath, { force: true });
+        removed++;
+      } catch {
+        hasEntries = true;
+      }
+    }
+
+    return hasEntries;
+  };
+
+  walk(skillDir);
+  return removed;
+}
+
 /**
  * Sync agent's installed skills to the filesystem.
  *
  * For simple skills (content in DB): writes SKILL.md to ~/.claude/skills/<name>/
- * For complex skills (isComplex=true): skipped here (handled by npx in entrypoint)
+ * For DB-backed complex skills: writes SKILL.md plus bundled skill_files rows.
+ * Legacy complex skills without skill_files remain handled by npx in entrypoint.
  */
 export function syncSkillsToFilesystem(
   agentId: string,
@@ -44,6 +102,7 @@ export function syncSkillsToFilesystem(
   const home = homeOverride ?? homedir();
   const errors: string[] = [];
   let synced = 0;
+  let removed = 0;
 
   // Directories to write to
   const skillDirs: string[] = [];
@@ -67,7 +126,8 @@ export function syncSkillsToFilesystem(
 
   for (const skill of skills) {
     if (!skill.isActive || !skill.isEnabled) continue;
-    if (skill.isComplex) continue; // Complex skills handled by npx
+    const bundledFiles = skill.isComplex ? getSkillFiles(skill.id) : [];
+    if (skill.isComplex && bundledFiles.length === 0) continue; // Legacy complex skills handled by npx
     if (!skill.content) continue;
 
     // Sanitize skill name to prevent path traversal (strip /, .., and non-safe chars)
@@ -75,6 +135,9 @@ export function syncSkillsToFilesystem(
     if (!safeName) continue;
 
     writtenNames.add(safeName);
+    const currentBundledFilePaths = new Set(
+      bundledFiles.filter((file) => !file.isBinary).map((file) => file.path),
+    );
 
     for (const baseDir of skillDirs) {
       const skillDir = join(baseDir, safeName);
@@ -83,14 +146,36 @@ export function syncSkillsToFilesystem(
 
       try {
         mkdirSync(skillDir, { recursive: true });
+        removed += reconcileManagedSkillFiles(skillDir, currentBundledFilePaths);
         writeFileSync(skillFile, skill.content, "utf-8");
         writeFileSync(markerFile, "", "utf-8");
         synced++;
       } catch (err) {
-        errors.push(
-          `${skill.name} -> ${skillDir}: ${err instanceof Error ? err.message : "Unknown error"}`,
+        const msg = err instanceof Error ? err.message : "Unknown error";
+        errors.push(`${skill.name} -> ${skillDir}: ${msg}`);
+        console.error(
+          `[skill-sync] Failed to write SKILL.md for ${skill.name} to ${skillDir}: ${msg}`,
         );
       }
+
+      for (const file of bundledFiles) {
+        if (file.isBinary) {
+          console.log(`[skill-sync] Skipping binary skill file ${skill.name}/${file.path}`);
+          continue;
+        }
+
+        const targetPath = join(skillDir, file.path);
+        try {
+          mkdirSync(dirname(targetPath), { recursive: true });
+          writeFileSync(targetPath, file.content, "utf-8");
+        } catch (err) {
+          const msg = err instanceof Error ? err.message : "Unknown error";
+          errors.push(`${skill.name}/${file.path} -> ${targetPath}: ${msg}`);
+          console.error(
+            `[skill-sync] Failed to write bundled file ${skill.name}/${file.path} to ${targetPath}: ${msg}`,
+          );
+        }
+      }
     }
   }
 
@@ -98,7 +183,6 @@ export function syncSkillsToFilesystem(
   // present). Leaves user-installed personal skills alone — important on
   // local dev where ~/.codex/skills holds skills the user installed
   // outside the swarm.
-  let removed = 0;
   for (const baseDir of skillDirs) {
     if (!existsSync(baseDir)) continue;
 

package/src/commands/runner.ts

@@ -3127,6 +3127,7 @@ async function checkCompletedProcesses(
   state: RunnerState,
   role: string,
   apiConfig?: ApiConfig,
+  cancelledSignaled?: Set<string>,
 ): Promise<void> {
   const completedTasks: Array<{
     taskId: string;
@@ -3161,6 +3162,9 @@ async function checkCompletedProcesses(
   // Remove completed tasks from the map and ensure they're marked as finished
   for (const { taskId, result, cursorUpdates, workingDir, credentialInfo } of completedTasks) {
     state.activeTasks.delete(taskId);
+    vcsDetectedTasks.delete(taskId);
+    vcsCheckTimestamps.delete(taskId);
+    cancelledSignaled?.delete(taskId);
 
     if (apiConfig) {
       removeActiveSession(apiConfig, taskId);
@@ -4122,7 +4126,7 @@ export async function runAgent(config: RunnerConfig, opts: RunnerOptions) {
 
         // Wait if at capacity (though unlikely on fresh startup)
         while (state.activeTasks.size >= state.maxConcurrent) {
-          await checkCompletedProcesses(state, role, apiConfig);
+          await checkCompletedProcesses(state, role, apiConfig, cancelledSignaled);
           await Bun.sleep(1000);
         }
 
@@ -4341,7 +4345,7 @@ export async function runAgent(config: RunnerConfig, opts: RunnerOptions) {
     await pingServer(apiConfig, role);
 
     // Check for completed processes first and ensure tasks are marked as finished
-    await checkCompletedProcesses(state, role, apiConfig);
+    await checkCompletedProcesses(state, role, apiConfig, cancelledSignaled);
 
     // Live HARNESS_PROVIDER reconciliation. Re-fetches `swarm_config` (overlaid
     // on env) and swaps the adapter if the resolved provider changed —

package/src/heartbeat/templates.ts

@@ -27,17 +27,19 @@ Goal: Review system status and your standing orders, take action if needed.
 
 ## Instructions
 1. **Read your HEARTBEAT.md** — run \`read /workspace/HEARTBEAT.md\` to get the latest standing orders (the snapshot above may be slightly stale).
-2. Review the system status above for anything that needs attention (stalled tasks, idle workers with available work, anomalies).
-3. **CRITICAL — Reboot failure triage:** Failures with reason "worker session not found" or "worker session heartbeat is stale" indicate tasks that were INTERRUPTED by a server restart. These are NOT "expected auto-cleanup" — they represent work that was lost mid-execution. For each such failure:
+2. **Prune tracked HEARTBEAT items FIRST.** Active Blockers + Watch Items + Open Discussion combined must stay at ≤10 items; 20 is the absolute max only when genuinely super busy. This cap does not apply to evergreen Standing Orders / Governance / Playbook-index reference sections. Before adding anything, re-check every tracked item against its lift trigger and remove anything resolved, stale, or past its trigger date. Lift incident detail to memory; never keep it inline in HEARTBEAT.md. Every new tracked item must include an explicit lift trigger + date; if it has no removal condition, do not add it. If the tracked list is already at/over cap, prune before (or instead of) adding — the cap is binding.
+3. **Run seeded heartbeat data-gathering.** Use \`script-run\` with global script \`Heartbeat Audit\` and pass the current HEARTBEAT.md text as \`heartbeatMarkdown\`. It covers Rules #10/#13/#15/#16/#17: resolved stale PRs, pool-target risk schedules, schedule/provider failure clusters, and whether daily-blocker-digest ran today. The Slack thread-reply check (Rule #11) stays Lead-side; the script runtime has no Slack token.
+4. Review the system status above plus the \`Heartbeat Audit\` result for anything that needs attention (stalled tasks, idle workers with available work, anomalies).
+5. **CRITICAL — Reboot failure triage:** Failures with reason "worker session not found" or "worker session heartbeat is stale" indicate tasks that were INTERRUPTED by a server restart. These are NOT "expected auto-cleanup" — they represent work that was lost mid-execution. For each such failure:
    - Check what the task was (via \`get-task-details\` with the task ID from the failure)
    - If a retry task was auto-created (tagged \`reboot-retry\`), verify it is progressing
    - If no retry exists and the work is still needed, re-create the task
    - Do NOT dismiss these as "expected" or "auto-cleanup"
-4. Review your standing orders for any periodic checks or actions.
-5. If something needs attention — take action now using your available tools (create tasks, post to Slack, cancel stuck tasks, etc.).
-6. If everything looks healthy and no standing orders are actionable — complete this task with a brief "All clear" summary. You may NOT say "All clear" if reboot-related failures exist that haven't been triaged.
-7. Do NOT create another heartbeat-checklist task — the system handles scheduling.
-8. **Update your standing orders** — After every heartbeat check, edit \`/workspace/HEARTBEAT.md\` directly. Add new patterns you noticed (recurring failures, workers needing attention), remove resolved items. This is your live operational runbook — keep it current.`,
+6. Review your standing orders for any periodic checks or actions.
+7. If something needs attention — take action now using your available tools (create tasks, post to Slack, cancel stuck tasks, etc.).
+8. If everything looks healthy and no standing orders are actionable — complete this task with a brief "All clear" summary. You may NOT say "All clear" if reboot-related failures exist that haven't been triaged.
+9. Do NOT create another heartbeat-checklist task — the system handles scheduling.
+10. **Update HEARTBEAT.md only after pruning.** Keep it current, but keep tracked items capped: remove resolved items, add only dated lift-triggered items, and lift detail to memory instead of growing the file.`,
   variables: [
     {
       name: "system_status",
@@ -73,22 +75,24 @@ The API server has just restarted (deployment, pod rotation, or crash). An aggre
 {{heartbeat_content}}
 
 ## Instructions
-1. **Triage reboot-interrupted work FIRST.** If the "Reboot-Interrupted Work" section above lists tasks:
+1. **Prune tracked HEARTBEAT items FIRST.** Active Blockers + Watch Items + Open Discussion combined must stay at ≤10 items; 20 is the absolute max only when genuinely super busy. This cap does not apply to evergreen Standing Orders / Governance / Playbook-index reference sections. Before adding anything, re-check every tracked item against its lift trigger and remove anything resolved, stale, or past its trigger date. Lift incident detail to memory; never keep it inline in HEARTBEAT.md. Every new tracked item must include an explicit lift trigger + date; if it has no removal condition, do not add it. If the tracked list is already at/over cap, prune before (or instead of) adding — the cap is binding.
+2. **Run seeded boot triage.** Use \`script-run\` with global script \`boot-triage\` to gather deploy-restart PR context, recent real failures, stuck offline-agent work, orphaned pending/offered tasks, and superseded tasks missing resume children in one read-only call.
+3. **Triage reboot-interrupted work FIRST.** If the "Reboot-Interrupted Work" section above or the \`boot-triage\` result lists tasks:
    - For each task: verify the retry is progressing via \`get-task-details\` with the retry task ID
    - If a retry failed or is stuck, re-create the task manually
    - If the work is no longer needed, cancel the retry task
    - You MUST address every item — do NOT skip this section
-2. **Verify supersede + resume worked end-to-end.** Worker crashes / OOMs are recovered via supersede (parent → \`superseded\`) + a fresh \`taskType=resume\` child created by the heartbeat sweep (DES-523). Sanity check:
+4. **Verify supersede + resume worked end-to-end.** Worker crashes / OOMs are recovered via supersede (parent → \`superseded\`) + a fresh \`taskType=resume\` child created by the heartbeat sweep (DES-523). Sanity check:
    - List recent \`superseded\` tasks: \`list-tasks status=superseded\` (last ~hour).
    - For each, confirm a child task with \`taskType=resume\` and a non-terminal status exists. If a superseded task is missing its resume child, the work is silently dropped — recreate the task manually.
    - Look for \`in_progress\` tasks older than 5 min on agents that show as offline — the sweep should have caught them. If any remain, recreate as needed.
-3. **Check orphaned tasks.** If the "Orphaned Tasks" section lists pending/offered tasks assigned to offline workers, re-assign or cancel them.
-4. Review agent status — are all expected workers online? If not, note which are missing.
-5. Review your standing orders for any post-reboot checks.
-6. Take action using your available tools.
-7. Complete this task with a summary of what you found and what actions you took. Include the status of each reboot-interrupted task.
-8. Do NOT create another boot-triage task — this is a one-off event.
-9. **Update your standing orders** — If the reboot revealed a pattern worth monitoring (e.g., frequent restarts, specific tasks that keep failing), add a standing order to HEARTBEAT.md via \`update-profile\` with \`heartbeatMd\`.`,
+5. **Check orphaned tasks.** If the "Orphaned Tasks" section or \`boot-triage\` result lists pending/offered tasks assigned to offline workers, re-assign or cancel them.
+6. Review agent status — are all expected workers online? If not, note which are missing.
+7. Review your standing orders for any post-reboot checks.
+8. Take action using your available tools.
+9. Complete this task with a summary of what you found and what actions you took. Include the status of each reboot-interrupted task.
+10. Do NOT create another boot-triage task — this is a one-off event.
+11. **Update HEARTBEAT.md only after pruning.** If the reboot revealed a pattern worth monitoring, add it only as a dated lift-triggered tracked item and only if the cap still holds after pruning. Lift incident detail to memory instead of growing HEARTBEAT.md.`,
   variables: [
     {
       name: "system_status",

package/src/http/index.ts

@@ -43,12 +43,17 @@ import { handleHeartbeat } from "./heartbeat";
 import { handleInboxState } from "./inbox-state";
 import { handleIntegrations } from "./integrations";
 import { handleKv } from "./kv";
-import { handleMcp } from "./mcp";
+import {
+  closeIdleMcpTransports,
+  DEFAULT_MCP_TRANSPORT_IDLE_TIMEOUT_MS,
+  handleMcp,
+  type McpTransportActivity,
+} from "./mcp";
 import { handleMcpBridge } from "./mcp-bridge";
 import { handleMcpOAuth, startMcpOAuthPendingGc, stopMcpOAuthPendingGc } from "./mcp-oauth";
 import { handleMcpServers } from "./mcp-servers";
-import { handleMcpUser } from "./mcp-user";
-import { handleMemory } from "./memory";
+import { closeIdleMcpUserTransports, handleMcpUser } from "./mcp-user";
+import { handleMemory, startMemoryGc, stopMemoryGc } from "./memory";
 import { handleMetrics } from "./metrics";
 import { handlePageProxy } from "./page-proxy";
 import { handlePages } from "./pages";
@@ -99,12 +104,15 @@ const globalState = globalThis as typeof globalThis & {
   __transports?: Record<string, StreamableHTTPServerTransport>;
   __transportsUser?: Record<string, StreamableHTTPServerTransport>;
   __sessionUsers?: Record<string, string>;
+  __transportActivity?: McpTransportActivity;
+  __transportActivityUser?: McpTransportActivity;
   __sigintRegistered?: boolean;
   __apiGcInterval?: ReturnType<typeof setInterval>;
   __runId?: string;
 };
 
 const API_GC_INTERVAL_MS = 5 * 60 * 1000;
+const MCP_TRANSPORT_IDLE_TIMEOUT_MS = DEFAULT_MCP_TRANSPORT_IDLE_TIMEOUT_MS;
 
 type GcCapableGlobal = typeof globalThis & { gc?: () => void };
 
@@ -130,11 +138,25 @@ function startApiGcInterval() {
 
   const gc = (globalThis as GcCapableGlobal).gc;
   if (typeof gc !== "function") {
-    console.log("[HTTP] Explicit GC unavailable; start API with --expose-gc to enable sweeps");
-    return;
+    console.log("[HTTP] Explicit GC unavailable; idle MCP transport sweeps remain enabled");
   }
 
   const interval = setInterval(() => {
+    const closedOwnerTransports = closeIdleMcpTransports(transports, transportActivity, {
+      idleTimeoutMs: MCP_TRANSPORT_IDLE_TIMEOUT_MS,
+      label: "MCP",
+    });
+    const closedUserTransports = closeIdleMcpUserTransports(
+      transportsUser,
+      sessionUsers,
+      transportActivityUser,
+      { idleTimeoutMs: MCP_TRANSPORT_IDLE_TIMEOUT_MS },
+    );
+    if (closedOwnerTransports > 0 || closedUserTransports > 0) {
+      console.log(
+        `[HTTP] Closed ${closedOwnerTransports} owner MCP and ${closedUserTransports} user MCP idle transport(s)`,
+      );
+    }
     scheduleApiGc("periodic API sweep");
   }, API_GC_INTERVAL_MS);
   interval.unref?.();
@@ -151,6 +173,8 @@ const transports: Record<string, StreamableHTTPServerTransport> = globalState.__
 const transportsUser: Record<string, StreamableHTTPServerTransport> =
   globalState.__transportsUser ?? {};
 const sessionUsers: Record<string, string> = globalState.__sessionUsers ?? {};
+const transportActivity: McpTransportActivity = globalState.__transportActivity ?? {};
+const transportActivityUser: McpTransportActivity = globalState.__transportActivityUser ?? {};
 
 const httpServer = createHttpServer(async (req, res) => {
   const startTime = performance.now();
@@ -291,8 +315,8 @@ const httpServer = createHttpServer(async (req, res) => {
         () => handleSessions(req, res, pathSegments, queryParams),
         () => handleInboxState(req, res, pathSegments, queryParams),
         () => handleTaskTemplates(req, res, pathSegments, queryParams),
-        () => handleMcp(req, res, transports),
-        () => handleMcpUser(req, res, transportsUser, sessionUsers),
+        () => handleMcp(req, res, transports, transportActivity),
+        () => handleMcpUser(req, res, transportsUser, sessionUsers, transportActivityUser),
       ];
 
       try {
@@ -334,6 +358,8 @@ globalState.__httpServer = httpServer;
 globalState.__transports = transports;
 globalState.__transportsUser = transportsUser;
 globalState.__sessionUsers = sessionUsers;
+globalState.__transportActivity = transportActivity;
+globalState.__transportActivityUser = transportActivityUser;
 
 async function shutdown() {
   console.log("Shutting down HTTP server...");
@@ -362,6 +388,9 @@ async function shutdown() {
   // Stop MCP OAuth pending-session garbage collector
   stopMcpOAuthPendingGc();
 
+  // Stop memory expired-row garbage collector
+  stopMemoryGc();
+
   if (globalState.__apiGcInterval) {
     clearInterval(globalState.__apiGcInterval);
     delete globalState.__apiGcInterval;
@@ -372,6 +401,7 @@ async function shutdown() {
     console.log(`[HTTP] Closing transport ${id}`);
     transport.close();
     delete transports[id];
+    delete transportActivity[id];
   }
 
   for (const [id, transport] of Object.entries(transportsUser)) {
@@ -379,6 +409,7 @@ async function shutdown() {
     transport.close();
     delete transportsUser[id];
     delete sessionUsers[id];
+    delete transportActivityUser[id];
   }
 
   // Close all active connections forcefully
@@ -522,6 +553,18 @@ httpServer
 
     // Start MCP OAuth pending-session garbage collector (5-min tick)
     startMcpOAuthPendingGc();
+
+    // Start expired-memory garbage collector (1-hour tick, immediate first run)
+    startMemoryGc();
+
+    // Background backfill: re-embed any agent_memory rows with wrong-dimension
+    // embeddings (e.g. 1536d instead of 512d). Non-blocking, idempotent, no-op
+    // when the DB is clean. See src/be/memory/boot-reembed.ts.
+    import("../be/memory/boot-reembed")
+      .then(({ runBootReembed }) => runBootReembed())
+      .catch((err) => {
+        console.error("[boot-reembed] startup backfill failed (non-fatal):", err);
+      });
   })
   .on("error", (err) => {
     console.error("HTTP Server Error:", err);

package/src/http/mcp-user.ts

@@ -5,6 +5,7 @@ import { isInitializeRequest } from "@modelcontextprotocol/sdk/types.js";
 import { resolveUserByToken } from "@/be/users";
 import { createUserServer } from "@/server-user";
 import type { User } from "@/types";
+import { closeIdleMcpTransports, type McpTransportActivity, markMcpTransportActivity } from "./mcp";
 
 function unauthorized(res: ServerResponse): true {
   res.writeHead(401, { "Content-Type": "application/json" });
@@ -32,6 +33,7 @@ export async function handleMcpUser(
   res: ServerResponse,
   transports: Record<string, StreamableHTTPServerTransport>,
   sessionUsers: Record<string, string>,
+  sessionActivity: McpTransportActivity = {},
 ): Promise<boolean> {
   const sessionId = req.headers["mcp-session-id"] as string | undefined;
 
@@ -57,16 +59,19 @@ export async function handleMcpUser(
 
     if (sessionId && transports[sessionId]) {
       transport = transports[sessionId];
+      markMcpTransportActivity(sessionActivity, sessionId);
     } else if (!sessionId && isInitializeRequest(body)) {
       transport = new StreamableHTTPServerTransport({
         sessionIdGenerator: () => randomUUID(),
         onsessioninitialized: (id) => {
           transports[id] = transport;
           sessionUsers[id] = user.id;
+          markMcpTransportActivity(sessionActivity, id);
         },
         onsessionclosed: (id) => {
           delete transports[id];
           delete sessionUsers[id];
+          delete sessionActivity[id];
         },
       });
 
@@ -74,6 +79,7 @@ export async function handleMcpUser(
         if (transport.sessionId) {
           delete transports[transport.sessionId];
           delete sessionUsers[transport.sessionId];
+          delete sessionActivity[transport.sessionId];
         }
       };
 
@@ -92,11 +98,13 @@ export async function handleMcpUser(
     }
 
     await transport.handleRequest(req, res, body);
+    markMcpTransportActivity(sessionActivity, transport.sessionId);
     return true;
   }
 
   if (req.method === "GET" || req.method === "DELETE") {
     if (sessionId && transports[sessionId]) {
+      markMcpTransportActivity(sessionActivity, sessionId);
       await transports[sessionId].handleRequest(req, res);
       return true;
     }
@@ -109,3 +117,18 @@ export async function handleMcpUser(
   res.end("Method not allowed");
   return true;
 }
+
+export function closeIdleMcpUserTransports(
+  transports: Record<string, StreamableHTTPServerTransport>,
+  sessionUsers: Record<string, string>,
+  sessionActivity: McpTransportActivity,
+  options: { now?: number; idleTimeoutMs?: number } = {},
+): number {
+  return closeIdleMcpTransports(transports, sessionActivity, {
+    ...options,
+    label: "user MCP",
+    onClose: (id) => {
+      delete sessionUsers[id];
+    },
+  });
+}