v5.4.0 — 🧠 Real Framework

Phase 9 release. Lifts TITAN from "agent that runs" to "agent that
remembers, doesn't fabricate, doesn't leak, and survives flooded inboxes."
Eight tracks across two co-working agents (Claude + Kimi K2.6) landed in a
single bundle.

728 deterministic tests pass in 7.80 s. Typecheck + UI build clean.

Created by Tony Elliott aka djtony707.

Added

Track B — Memory upgrades

• src/memory/index.ts (NEW) — inverted-index TF-IDF keyword search.
  Sub-50 ms query at 5 000 episodes (verified via tests/unit/memory-index.test.ts).
  Wired into searchMemory, addEpisode, enforceMemoryBounds.
• addEpisode now accepts { awaitEntities: true } to close the entity-
  extraction race window — replaces the brittle "wait 100 ms and hope" pattern.
• enforceMemoryBounds now prunes entities by salience score
  (typeWeight × (1 + episodeRefs + facts)) instead of FIFO lastSeen.
  Identity entities (person, project) survive log floods that previously
  evicted them.
• memory.vectorSearchEnabled default flipped false → true. Silent
  fallback contract preserved — installs without Ollama running don't break.

Track D — Fabrication guard

• src/safety/fabricationGuard.ts (NEW) — pattern detection across six
  categories (file_write / file_edit / file_delete / shell_run /
  web_action / tool_used). All patterns require first-person voice and
  cross-check against tool history.
• verifyFileWriteClaim(path, expected?) — SHA-256 hash check, lenient
  trailing-whitespace compare, doesn't throw on invalid paths.
• buildNudgeMessage(findings) — blunt corrective text for the next turn.

Track A — Sub-agent safety (Kimi)

• Stall detection (3 identical responses → bail).
• Loop detection (identical tool+args fingerprint → bail).
• Per-tool error wrapping (returns ToolResult instead of throwing).
• Tool output summarization (>10K chars truncated with marker).
• Graceful degradation (all tools fail → early bail with error summary).

Track C — Self-improvement activation (Kimi)

• Checkpoint-before-mutation + auto-rollback on score drop.
• Rate limiting (max 1 mutation/hour via canMutate/recordMutation).

Fixed

Operational session leak (Phase 9 hotfix)

• TITAN PC v5.3.2 accumulated 755 in-memory sessions in 29 min. Root
  cause: every endpoint that internally calls processMessage with a
  templated channel name (autoresearch-trigger-${type},
  twilio-call-${callSid}, initiative-fix, monitor, mesh,
  deliberation, eval) created a unique cache key under
  ${channel}:${userId}:${agentId} — and all sessions shared the same
  30 min idle TTL. At ~26 sessions/min creation rate, the 30 min window
  buffered 750+ entries before the first one expired.
• src/agent/session.ts: new isEphemeralChannel(channel) classifier.
  Persistent allowlist (webchat, voice, discord, telegram, slack, …)
  keeps the full SESSION_TIMEOUT_MS (30 min); everything else gets
  EPHEMERAL_TTL_MS (5 min).
• LRU cap on ephemeral cache: EPHEMERAL_MAX_ACTIVE = 100. Beyond that,
  oldest-by-lastActive get evicted.
• 7-day idle DB purge in cleanupStaleSessions so the store doesn't grow
  forever when sessions never get re-messaged.
• src/gateway/server.ts: cleanup interval shortened 5 min → 60 s so
  ephemeral 5 min TTL evicts within ~1 min of expiry.
• New POST /api/sessions/sweep endpoint with {channel?, channelPrefix?, idleMs?, force?} body for live operational drain — no service restart.
  Default sweep closes every ephemeral; force: true includes persistent.

Pre-existing typecheck errors

• src/skills/builtin/gepa.ts: await in a non-async function — wrapped
  daemon registration in void (async () => { … })().
• src/skills/builtin/self_improve.ts: SELF_IMPROVE_DIR used at line
  20 before its export const declaration at line 111. Hoisted const to
  top, re-exported via export { SELF_IMPROVE_DIR }.

Test fixture stale-date trap

• tests/integration/smoke.test.ts "Session Listing" idle fixture used
  hardcoded 2026-04-13 dates — past the new 7-day idle purge threshold.
  Wrapped in vi.useFakeTimers + vi.setSystemTime('2026-04-13T11:00Z')
  so the test stays valid as wall-clock time advances. Added
  debouncedSave: vi.fn() to the memory/memory.js mock.

Tests

• tests/unit/memory-index.test.ts (NEW, 24): tokenization, ranking,
  IDF dampening, idempotent add, removal, vocabulary, performance budget.
• tests/unit/memory-vector.test.ts (NEW, 5): schema default, embedding-
  model default, signature exports.
• tests/unit/memory-salience-pruning.test.ts (NEW, 10): survival
  ordering, recency tiebreak, identity protection, fallback weights.
• tests/unit/fabrication-guard.test.ts (NEW, 26): all six categories,
  multi-finding responses, no false fires on tool-backed claims, third-
  person prose ignored, buildNudgeMessage format.
• tests/unit/fabrication-verify.test.ts (NEW, 10): existence, empty-
  file flagging, lenient compare, hash determinism.
• tests/unit/session-cleanup.test.ts (NEW, 21): per-channel TTL split,
  LRU cap eviction order, sweep filters, 7-day idle purge.
• tests/unit/subagent-safety.test.ts (NEW, 11): stall/loop detection,
  per-tool error wrap, output truncation, graceful degradation.
• tests/unit/self-improve.test.ts (NEW, 7): checkpoint/restore + rate
  limiting.

---