create-walle 0.10.1 → 0.10.2

package/template/coding-task-manager/server/boot.js

@@ -15,6 +15,7 @@
 
 const fs = require('node:fs');
 const path = require('node:path');
+const { classifyUserProvenance } = require('../../shared/conversation-provenance');
 const { createReadPoolBreaker } = require('../lib/readpool-breaker');
 const { offthreadFallback } = require('../lib/resume-admission');
 
@@ -36,6 +37,39 @@ function semverCompare(a, b) {
   return 0;
 }
 
+// WP-3 (design doc §5 D-7): a tier-3 AI title is generated only from a first message
+// shared/conversation-provenance.js rates `human`. The harness/provider envelopes that also
+// persist as role="user" (<local-command-caveat>, <task-notification>, "# AGENTS.md
+// instructions for", a compaction summary, …) are skipped instead, so the name falls back to
+// the agent title or the default label. One predicate, two callers: this module's background
+// job and server.js's /api/sessions/generate-titles route.
+function isHumanPromptCandidate(provider, firstMessage) {
+  const text = String(firstMessage || '');
+  if (!text) return false;
+  // The classifier's only provider-dependent rule is openai-only, and agent_sessions.provider
+  // is durably 'claude' | 'openai' (transcript-store maps codex -> openai) — same normalization
+  // as lib/migrations/v25-conversation-provenance.js.
+  const normalized = String(provider || '').toLowerCase() === 'openai' ? 'openai' : 'claude';
+  return classifyUserProvenance({ provider: normalized, text }).provenance === 'human';
+}
+
+// Boilerplate stripping for the title prompt itself — this is cleaning text, not classifying a
+// row (eligibility is isHumanPromptCandidate's verdict). Only the `This session is being
+// continued…` (compaction) line was dropped against the classifier: an eligible row can never
+// start with it. The resume forms below are not in the classifier's envelope table, and a
+// `[Request interrupted …]` prefix only matches the classifier when the whole message is the
+// marker, so they still need stripping here.
+function cleanFirstMessage(msg) {
+  if (!msg) return '';
+  return msg
+    .replace(/^Continue the conversation[^\n]*\n*/i, '')
+    .replace(/^Resume:?\s*/i, '')
+    .replace(/^Summary:?\s*\n?/i, '')
+    .replace(/^The summary below covers[^\n]*\n*/i, '')
+    .replace(/^\[(?:Request interrupted|Tool use|Error|Retrying)[^\]]*\]\s*/g, '')
+    .trim();
+}
+
 function createBootJobs(deps) {
   const {
     installRoot, // path.join(__dirname, '..') from server.js — package.json lives here
@@ -68,6 +102,10 @@ function createBootJobs(deps) {
 
   async function checkForUpdates(source = 'scheduled') {
     updateState.currentVersion = getCurrentVersion();
+    // CTM_DISABLE_UPDATE_CHECKS=1 (dev.sh, the render harness) disables the check for EVERY caller,
+    // including GET /api/updates/check?refresh=1 the page sends after load -- not only server.js's
+    // timers -- so an isolated instance never reads the public registry (B19).
+    if (env.CTM_DISABLE_UPDATE_CHECKS === '1') return updateState;
     trackUpdateTelemetry(telemetry, 'check_started', {
       source,
       currentVersion: updateState.currentVersion,
@@ -166,11 +204,100 @@ function createBootJobs(deps) {
     });
   }
 
+  // --- WP-6: one-shot agent-title backfill (design doc session-name-agent-title-precedence
+  // §§1e/6/8) ---
+  // Every row CTM already had a transcript for, but never an agent name for, gets one read of its
+  // transcript tail (≤64 KiB, off the main thread, in the read pool) and one apply through
+  // lib/agent-title-apply.js on the db-owner. Once per DB: the settings marker is written only
+  // after a COMPLETE walk, so a run cut short by an unavailable pool resumes on the next boot
+  // (already-named rows have left the candidate set by then, so the retry is cheap, not a redo).
+  // Never blocks boot: it runs from the 'auto-title' scheduler job, behind the same
+  // _canRunStartupMaintenance admission every other maintenance job uses.
+  const AGENT_TITLE_BACKFILL_MARKER = 'agent_title_backfill_v1';
+  const AGENT_TITLE_BACKFILL_TIMEOUT_MS = 30000;
+  const AGENT_TITLE_BACKFILL_PAGE_ROWS = 20;
+  // 40 pages = 800 rows in one boot's walk — 5× the §1c candidate census (164). Past that the run
+  // leaves the marker unset and the remainder waits for the next boot rather than holding the job.
+  const AGENT_TITLE_BACKFILL_MAX_PAGES = 40;
+  // Deliberately NOT env-tunable (no new flag for a one-shot walk): unlike the auto-title job
+  // above, this runs at most once per DB, so a dropped run costs one boot, not a steady state.
+  const _agentTitleBackfillBreaker = createReadPoolBreaker({ degradeAfter: 3, cooldownMs: 60000 });
+
+  /**
+   * @returns {Promise<{skipped?: string, scanned?: number, applied?: number, pages?: number,
+   *   marker?: string, error?: string}>} never throws — the caller is a title-generation job.
+   */
+  async function runAgentTitleBackfill() {
+    // Dependency-injected seams (this module is require()-able standalone — see the header). A
+    // harness that stubs dbModule/offthreadRead without the backfill surface simply skips.
+    if (typeof offthreadRead !== 'function' || typeof dbModule?.getSetting !== 'function'
+      || typeof dbModule?.setSetting !== 'function'
+      || typeof dbModule?.applyAgentTitleBackfillRows !== 'function') {
+      return { skipped: 'no_seam' };
+    }
+    try {
+      const marker = String(dbModule.getSetting(AGENT_TITLE_BACKFILL_MARKER, '') || '');
+      if (marker) return { skipped: 'done', marker };
+      const skipIds = [];
+      let scanned = 0;
+      let applied = 0;
+      let pages = 0;
+      let complete = false;
+      while (pages < AGENT_TITLE_BACKFILL_MAX_PAGES) {
+        const page = await offthreadRead({
+          breaker: _agentTitleBackfillBreaker,
+          op: 'backfillAgentTitles',
+          payload: { dbPath: dbModule.getDbPath(), limit: AGENT_TITLE_BACKFILL_PAGE_ROWS, skipIds },
+          timeoutMs: AGENT_TITLE_BACKFILL_TIMEOUT_MS,
+          onSuccess: (res) => (res && Array.isArray(res.rows) ? {
+            rows: res.rows,
+            scanned: Number(res.scanned) || res.rows.length,
+            scannedIds: Array.isArray(res.scannedIds) ? res.scannedIds : [],
+          } : null),
+          onFail: () => null,
+        });
+        if (!page) return { skipped: 'read_pool_unavailable', scanned, applied, pages };
+        pages += 1;
+        scanned += page.scanned;
+        if (page.rows.length) {
+          const outcome = await detachOwnerMutation(
+            'agent-title-backfill',
+            () => dbModule.applyAgentTitleBackfillRows(page.rows),
+          );
+          applied += Number(outcome?.applied) || 0;
+        }
+        for (const id of page.scannedIds) skipIds.push(id);
+        // A short page is the end of the walk; a page that scanned nothing cannot advance it.
+        if (page.scanned < AGENT_TITLE_BACKFILL_PAGE_ROWS) { complete = true; break; }
+        if (!page.scannedIds.length) break;
+      }
+      if (!complete) return { skipped: 'page_cap', scanned, applied, pages };
+      // '<ISO time>:<applied>/<scanned>' — the marker's value is also the run's receipt.
+      const value = `${new Date().toISOString()}:${applied}/${scanned}`;
+      await detachOwnerMutation(
+        'agent-title-backfill-marker',
+        () => dbModule.setSetting(AGENT_TITLE_BACKFILL_MARKER, value),
+      );
+      logger.log(`[agent-title-backfill] scanned=${scanned} applied=${applied} pages=${pages}`);
+      return { scanned, applied, pages, marker: value };
+    } catch (error) {
+      // REPORT: the marker stays unset, so the next boot's job run retries the walk.
+      logger.warn('[agent-title-backfill] failed:', error.message);
+      return { skipped: 'failed', error: error.message };
+    }
+  }
+
   async function runAutoTitleGeneration() {
     try {
       if (Date.now() < _autoTitleDisabledUntil) return; // Backoff after auth failure
       // WP-2 (D-1): no CTM-side credential preflight — a credential-less setup
       // fails callClaudeAsync below and the catch block's 1-hour backoff applies.
+      // WP-6: the one-shot agent-title backfill rides this job — it is this module's only
+      // scheduler-registered body (lib/session-jobs.js 'auto-title', gated by
+      // _canRunStartupMaintenance), so it inherits the existing background/maintenance admission
+      // instead of competing with boot or the interactive lane. The settings marker makes every
+      // later run a no-op, and a failure here never fails the title generation below.
+      await runAgentTitleBackfill();
 
       // Find sessions needing titles: have messages but no title in sessions table.
       // Runs on the read-pool worker (see _autoTitleCandidatesOffThread) to keep the
@@ -179,23 +306,13 @@ function createBootJobs(deps) {
 
       if (needTitles.length === 0) return;
 
-      // Filter out CTM internal sessions
-      const eligible = needTitles.filter(s => !isCtmInternalSession(s.first_message));
+      // Filter out CTM internal sessions, and any first message that is harness/provider
+      // boilerplate rather than a human prompt (D-7) — a wrapper text titled by the LLM is the
+      // bug this tier used to have.
+      const eligible = needTitles.filter(s => !isCtmInternalSession(s.first_message)
+        && isHumanPromptCandidate(s.provider, s.first_message));
       if (eligible.length === 0) return;
 
-      // Strip resume/continuation boilerplate from first messages
-      function cleanFirstMessage(msg) {
-        if (!msg) return '';
-        return msg
-          .replace(/^This session is being continued from a previous conversation[^\n]*\n*/i, '')
-          .replace(/^Continue the conversation[^\n]*\n*/i, '')
-          .replace(/^Resume:?\s*/i, '')
-          .replace(/^Summary:?\s*\n?/i, '')
-          .replace(/^The summary below covers[^\n]*\n*/i, '')
-          .replace(/^\[(?:Request interrupted|Tool use|Error|Retrying)[^\]]*\]\s*/g, '')
-          .trim();
-      }
-
       const prompt = `Generate short, descriptive titles (max 60 chars) for these Claude Code sessions.
 Each session has an ID and the user's first message. Create a title that summarizes what the user was working on.
 
@@ -220,7 +337,7 @@ Return ONLY a JSON array of objects with "id" and "title" fields. No markdown fe
         let generated = 0;
         for (const t of titles) {
           if (t.id && t.title) {
-            await detachOwnerMutation('auto-generated-session-title', () => dbModule.setSessionTitleNew(t.id, t.title, false));
+            await detachOwnerMutation('auto-generated-session-title', () => dbModule.setSessionTitleNew(t.id, t.title, false, { source: 'ai_title', status: 'generated' }));
             generated++;
           }
         }
@@ -259,8 +376,10 @@ Return ONLY a JSON array of objects with "id" and "title" fields. No markdown fe
     updateStateIsStale,
     refreshUpdateState,
     runAutoTitleGeneration,
+    runAgentTitleBackfill, // WP-6: also exposed so a resume/repair path can re-walk on demand
+    AGENT_TITLE_BACKFILL_MARKER,
     healStartupTaskSessionIds,
   };
 }
 
-module.exports = { createBootJobs, readPackageVersion, semverCompare };
+module.exports = { createBootJobs, readPackageVersion, semverCompare, isHumanPromptCandidate, cleanFirstMessage };