kxkm_clown/TODO.md

TODO

Session 2026-03-26 — TTS Evaluation lot-30 [done]

• lot-30-pocket-tts: evaluation complete — Pocket TTS v1.1.1 (Kyutai, MIT, 100M params, CPU 6x realtime, ~200ms first chunk, voice cloning 5s ref). VERDICT: SURVEILLER. Bloquant: EN only, FR planifie sans date (issue #118). Stack actuel optimal: Qwen3-TTS :9300 + Chatterbox :9200 + Piper fallback. Action: Watch GitHub release + issue #118. Spike immediat quand FR annonce. Voir docs/LOT030_TTS_EVALUATION.md.

Session 2026-03-26 — RAG audit + Journald monitoring [done]

• lot-28-rag-config: audit confirmé — 4 env vars impl. in rag.ts (lines 13-16), startup model check in init(), lot marked [done] in PLAN.md
• lot-29 journald: scripts/journald-monitor.sh created (TUI colored dots, --watch, --lines, --service flags)
• lot-29 journald: ops/v2/journald-alerts.conf created (INI config, restart policies, alert conditions per service)

Session 2026-03-26 — Timeline UI + Composition Tests [done]

• lot-547: Timeline model v1 (already in composition-store.ts: Track, TimelineClip, TimelineMarker, TimelineModelV1)
• lot-548: Waveform timeline UI v1 (TimelineView.tsx + CSS, routed at #timeline/:id)
• lot-549: Composition tests renfort (20 tests, 0 fail — added setActiveComposition, BPM clamping, markers sort, multi-track clips)
• lot-201: Schema memory persona v2 (done in prior session via feat(personas): migrate runtime memory to v2-local store)
• lot-202: Policy engine extraction/summarization/pruning (done via feat(personas): add configurable memory policy engine)
• lot-205: Retirer fallback legacy personas.json / persona-*.json (aucun fichier legacy sur disque)
• lot-206: Supprimer aliases createInMemory*Repo (déjà absent — migration convergée)
• P11 routes extraction: routes/personas.ts, routes/node-engine.ts, routes/chat-history.ts, app-middleware.ts

Session 2026-03-25 — Personas runtime V2

• Migrer le store local runtime personas vers data/v2-local/* en fichier par persona
• Conserver fallback de lecture legacy (personas.json, persona-*.json) pour migration douce
• Corriger les fuites de références mutables dans repos in-memory personas/sources/feedback/proposals
• Ajouter tests de non-régression migration + immutabilité (create-repos.test.ts)
• Aligner scripts smoke.js et build.js sur KXKM_LOCAL_DATA_DIR=data/v2-local
• Mettre les artefacts legacy build derrière flag explicite BUILD_INCLUDE_LEGACY_PERSONA_V1=1
• Exposer le mode fallback API comme local au lieu de memory
• Recaler les factories canoniques sur createLocal*Repo avec alias de compat
• Push distant des commits locaux (bloqué dans cette session: transport git/gh indisponible)
• lot-201: Implémenter schema memory persona v2 (working + archival blocks)
• lot-202: Ajouter policy engine extraction/summarization/pruning configurable
• lot-203: Benchmark runtime memory local vs approche Mem0 — verdict: garder le store local (latence <5ms, 0 deps externes, Mem0 trop lourd pour l'échelle actuelle) — voir docs/LOT203_MEMORY_BENCHMARK.md
• lot-204: Spike training OpenCharacter + PCL sur 5 personas — GO conditionnel. RTX 4090 suffisant. Pipeline: dialogues synthétiques Ollama → SFT QLoRA Qwen2.5-7B Unsloth → DPO PCL-style → GGUF Ollama. 7-12h total. Voir docs/LOT204_TRAINING_SPIKE.md. Suite: lot-204a (enrichir profils + script génération), lot-204b (SFT), lot-204c (export + eval).
• lot-205: Retirer le fallback legacy personas.json / persona-*.json après fermeture de fenêtre de migration
• lot-206: Supprimer les alias createInMemory*Repo après convergence complète des imports/tests

P0 Critical (sécurité & stabilité)

• Fix bash injection dans node-engine-runtimes.js — validation runtimeId/nodeType + timeout 30min
• Ajouter timeout sur les appels Ollama — 15s metadata, 5min chat streaming
• Validation des entrées sur les messages WebSocket — 64KB max frame, 8192 chars text, type checks
• Rate limiting par user/IP sur chat — rate-limit.js + 30 msg/min par IP dans WebSocket
• escapeHtml dédupliqué → utils.js
• normalizeAuth consolidé dans admin-api.js
• ensureSeedGraphs guard flag ajouté
• finishRun comptage d'artifacts sans JSON parse
• recoverRunnableRuns double-read corrigé

P1 V1 Quality

• Migrer vers le SDK officiel ollama-js — fait en P7 (ollama.js réécrit)
• Ajouter un audit logging pour les actions admin — audit-log.js + intégré dans http-api.js et server.js
• Implémenter l'analyse image/audio dans attachment-pipeline.js — stubs factory avec adapter slot
• Corriger la validation d'origine postMessage — déjà en place (personas.js:1476)
• Ajouter la déduplication de requêtes dans admin-api.js — fait en P7 (deduplicatedFetch)
• Node Engine : validation de tri topologique — déjà en place (cycle detection dans runner)
• Node Engine : timeout d'exécution par nœud — 10min default via NODE_ENGINE_STEP_TIMEOUT_MS

P2 V2 Domaines

• Schéma Postgres + migrations + repos typés (packages/storage) — session, persona, graph, run repos
• Module auth réel (packages/auth) — crypto.scrypt, token gen, extractSessionId, validateLoginInput
• Logique domaine chat (packages/chat-domain) — ChatMessage, ChatSession, compactHistory, channel validation
• Logique domaine persona (packages/persona-domain) — validatePersonaUpdate, aggregateFeedback, computePersonaDiff
• Brancher les repos Postgres dans apps/api — async createApp(), fallback in-memory si pas de DATABASE_URL

P3 Node Engine V2

• Porter registry → packages/node-engine (15 node types, 7 familles)
• Porter graph ops (topologicalSort, validateEdgeContracts, collectNodeInputs)
• Porter run state machine (createRun, RunStep, resolveFinalStatus)
• Porter queue logic (createQueueState, enqueue, dequeue, canDequeue)
• Runtime definitions (5 runtimes)
• Isoler les runtimes avec sandboxing approprié — fait en P8 (sandbox.ts)
• Adaptateurs d'entraînement réels (LoRA, QLoRA, SFT) — fait en P8 (training.ts)
• Brancher le runner V2 dans apps/worker — poll loop, stub executors, graceful shutdown

P4 Frontend V2

• API client centralisé (api.ts)
• 9 composants React (Header, Login, Nav, PersonaList, PersonaDetail, NodeEngineOverview, GraphDetail, RunStatus, ChannelList)
• Routing hash-based + responsive CSS
• Interface chat React (WebSocket live) — fait en P7 (Chat.tsx + useWebSocket.ts)
• Éditeur visuel Node Engine (React Flow) — fait en P7 (NodeEditor.tsx + EngineNode.tsx)

P5 TUI & Ops

• TUI health-check (V1+V2+Ollama+disk+memory)
• TUI queue-viewer (runs, statuses)
• TUI persona-manager (overview)
• Log rotation (--dry-run, --max-age-days)
• Packages/tui: ansi, statusDot, formatTable, drawBox

P6 Migration

• Matrice de parité V1 → V2 — scripts/parity-check.js (persona, graph, channel, API shape checks)
• Scripts de migration de données — scripts/migrate-v1-to-v2.js (personas, graphs, runs → Postgres, --dry-run support)
• Smoke tests pour V2 — scripts/smoke-v2.js (22 tests, 5 catégories, npm run smoke:v2)
• Procédure de rollback — scripts/rollback-v2.js (drop/truncate tables with confirmation, --yes, --tables filter)

P0+ Sécurité V1 (deep analyse 2026-03-16)

• Path traversal dans storage.js — sanitisation session IDs + boundary check memory paths
• Path traversal dans persona-registry.js / persona-store.js — safeFsId() helper
• Path traversal dans attachment-store.js — sanitisation IDs + boundary check
• SSRF dans web-tools.js — blocage localhost, IPs privées, .local/.internal
• Response body limit dans web-tools.js — truncation 2 MB
• Log injection dans storage.js — sanitisation paramètre role
• Crash JSONL corrompu dans storage.js — try/catch par ligne
• Map mutation during iteration dans sessions.js — collect then process
• Session leak /msg dans commands.js — clé stable au lieu de Date.now()
• Unbounded userRateLimits dans chat-routing.js — pruning > 200 entries
• Session pruning O(n) dans admin-session.js — throttle 60s

P7 Intégration avancée

• Migrer vers ollama-js SDK officiel — ollama.js réécrit avec Ollama class, même interface
• Chat WebSocket React live — hook useWebSocket + composant Chat IRC, auto-reconnect
• Éditeur visuel Node Engine avec React Flow — NodeEditor.tsx + EngineNode.tsx, 7 familles colorées
• Déduplication requêtes GET dans admin-api.js — deduplicatedFetch transparent
• Repos Postgres pour persona sources/feedback/proposals — 3 tables + repos + fallback in-memory
• CI/CD GitHub Actions — .github/workflows/ci.yml (check V1+V2)
• Deep analyse finale V1+V2 — 14 modules V1 vérifiés, 3 fixes TS V2, intégrité confirmée

P8 Production Readiness

• Adaptateurs training réels (TRL + Unsloth pour LoRA/DPO) — packages/node-engine/src/training.ts + worker intégré
• Sandboxing runtimes Node Engine (containers/VM) — packages/node-engine/src/sandbox.ts (none/subprocess/container)
• Turborepo pour build orchestration monorepo — turbo.json + scripts alignés + CI mis à jour
• Tests unitaires V2 avec node:test + supertest — 102 tests, 46 suites, 6 packages + API integration
• Tests React avec Vitest + RTL — 33 tests, 6 composants (Header, Login, Nav, PersonaList, RunStatus, ChannelList)
• Créer le repo GitHub privé — https://github.com/electron-rare/kxkm_clown

P9 Code Quality (simplify review)

• Triple filter → single-pass loop dans node-engine.js:deriveAsyncMeta
• Duplicate sanitization extraite dans attachment-store.js:sanitizeId
• Double loadModelIndex() éliminé dans node-engine-store.js:registerDeployment

P10 Lot 11 — Consolidation & Feature Parity

Phase A — Analyse & Recherche

• Deep analyse code V1+V2 (agent en cours)
• Veille OSS mise à jour (agent en cours)
• Recherche HuggingFace (agent en cours)

Phase B — Correctifs sécurité (deep analyse)

• P0 SEC-01 Path traversal node-engine-runner.js — reject absolute paths + rootDir boundary check
• P0 SEC-04 V2 login role self-assignment — viewer par défaut, admin via ADMIN_TOKEN
• P1 BUG-06 Health endpoint leaking DATABASE_URL — remplacé par storageMode string
• P1 BUG-02 Timeout promise leak node-engine-runner.js — AbortSignal cancel
• P1 SEC-03 Attachments sans auth — ajout requireAdminNetwork middleware
• Compilation + 119 tests OK après correctifs

Phase C — Feature Parity V2

• Recovery on crash worker — recoverStaleRuns() + worker startup recovery
• Cancel support — requestCancel() repo + shouldCancel callback worker + API endpoint
• Tab completion chat V2 — fuzzy matching 108+ commands + @mentions
• Commandes slash V2 — parseSlashCommand, resolveCommand, generateHelpText + 11 commandes + 17 tests
• Mémoire conversationnelle V2 — ConversationMemory, addToMemory, buildLlmContext, clearMemory
• Status strip admin V2 — GET /api/v2/status (personas, graphs, runs, queue)
• Subnet gate V2 — CIDR middleware /api/v2/admin/* avec ADMIN_SUBNET env
• Retention sweep V2 — deleteOlderThan() repo + POST /api/v2/admin/retention-sweep
• Export HTML V2 — GET /api/v2/export/html avec download attachment
• Upload fichiers V2 — bouton upload base64 dans Chat.tsx, accept image/audio/text/pdf/json/csv
• Tab completion chat V2 — nicks + commandes slash, cycling Tab, reset auto

Phase D — Déploiement & Docs

• Docker — Dockerfile (multi-stage Node 22 alpine) + docker-compose.yml (5 services) + .dockerignore
• Documentation utilisateur — UserGuide.tsx (8 sections accordion, lot 472)
• Performance profiling — deep analysis API (21 findings) + frontend (15 findings), 9 fixes applied

P11 Lot 17 — Deep Audit & Refactoring

Phase A — Analyse & documentation

• Script TUI deep-audit.js (security, perf, complexity, deps) — ops/v2/deep-audit.js
• Veille OSS enrichie 2026-03-17 (10 nouvelles catégories) — docs/OSS_WATCH_2026-03-16.md
• Diagrammes Mermaid (Context Store, Docker, Inter-persona) — docs/ARCHITECTURE.md
• AGENTS.md refondu (matrice 10 agents, Mermaid routing, pipeline) — docs/AGENTS.md
• PLAN.md consolidé avec lots 17-19
• TODO.md consolidé avec backlog Phase 6+
• Deep analyse — API 21 findings + frontend 15 findings, 9 fixes applied

Phase B — Refactoring code

• P1 ws-chat.ts: extraction modules (523 LOC core + 5 modules extracted)
  • ws-upload-handler.ts (241 LOC)
  • ws-conversation-router.ts (583 LOC)
  • ws-commands-chat.ts (885 LOC)
  • ws-commands-generate.ts (1841 LOC)
  • ws-commands-info.ts (876 LOC)
• P1 app.ts: extraction — server.ts 300 LOC + app.ts 131 LOC (already modular)
  • Extraire routes/personas.ts
  • Extraire routes/node-engine.ts
  • Extraire routes/chat.ts (→ routes/chat-history.ts)
  • Extraire middleware/auth.ts (→ app-middleware.ts)
• P2 writeFileSync — composition-store async (lot 461)
• P2 console.log — all migrated to pino structured logging
• P2 React.memo sur Chat, ChatHistory, VoiceChat, NodeEditor — ChatMessage already memo'd
• P2 Lazy load: React.lazy + Suspense pour routes lourdes — 17 lazy routes + WaveformPlayer

Phase C — Infrastructure

• SearXNG dans docker-compose — kxkm_clown-searxng-1 :8080 (healthy)
• MinerU/Docling dans docker-compose — kxkm_clown-docling-1 :9400 (healthy)
• Spike BGE-M3 embeddings — default changed to bge-m3 (lot 471)
• Déployer deep-audit.js sur kxkm-ai — cron 3am daily (lot 469)
• Créer utilisateur Discord Pharmacius — kxkm_clown-discord-bot-1 (up 4 days)

Phase D — Nouveaux node types

• music_generation node (ACE-Step 1.5 + AI Bridge 17 backends)
• voice_clone — XTTS-v2 via AI Bridge /generate/voice-clone (lot 504)
• document_extraction node (Docling :9400)

P12 Lot 18 — Voice & MCP (futur)

• XTTS-v2 voice cloning par persona
• LLMRTC WebRTC streaming (TypeScript, VAD, barge-in)
• MCP SDK integration — 6 MCP tools (chat, personas, search, status, music_generate, ai_bridge_health)
• PCL + OpenCharacter pipeline fine-tune
• Chatterbox TTS evaluation — deployed as fallback TTS, remote backend on :9100

P13 Lot 19 — Music & Creative (futur)

• ACE-Step 1.5 production — AI Bridge /generate/music-ai + openDIAW.be AceStep instrument
• /compose command (prompt → musique) — lots 320+, full composition pipeline
• Flux 2 dans ComfyUI (blocked — needs 12GB checkpoint download)
• A2A Protocol — Agent Card + JSON-RPC /a2a endpoint (lot 502)

Lot 20 - Deep Analyse Continue & Execution Chainee [en cours]

A faire maintenant:

• Poursuivre extraction modulaire — ws-chat 523 LOC + 5 modules
• Decouper app.ts — already 300+131 LOC modular
• Instrumentation perf — Prometheus /metrics + ws_message latency (lot 501)
• SearXNG :8080 + Docling :9400 (Docker compose, healthy)

Fait sur ce lot:

• Extraction modulaire du bloc upload/analyse de ws-chat.ts (ws-upload-handler.ts).
• Refonte UI Minitel depuis la racine du site (public/index.html, public/styles.css, public/app.js).
• Deep audit execute et relance apres correctifs.
• Corrections context-store appliquees et validees.
• check:v2 et test:v2 au vert.
• Correctif anti-decrement TTS negatif (ttsActive).
• Cleanup opportuniste des sessions expirees (mode memory).
• Purge des logs vides/obsoletes ops/v2/logs.
• Script ops/v2/run-deep-cycle.sh ajoute et execute.
• Tests apps/api/src/context-store.test.ts ajoutes et valides.
• Scoring de dette technique integre dans ops/v2/deep-audit.js.
• Verification JSON dette: score 78/100 (niveau high).

Phase Session 2026-03-19/20 — Lots 24-94

Summary

71 lots completed (lot-24 through lot-94). Major achievements:

• Structured logging: pino JSON logs, 43 console.log replaced, 0 remaining
• WS protocol hardening: Promise chain per-connection, seq numbers, FIFO ordering, 6 integration tests
• Frontend perf: 17 lazy routes (-53% initial JS), React.memo + useCallback stabilization
• CRT effect: CSS-only (scanlines, vignette, phosphor glow, boot animation 0.8s)
• Systemd services: kxkm-tts + kxkm-lightrag, auto-restart, deploy.sh migrated from tmux
• Tool-calling: llama3.1:8b-instruct for Sherlock, benchmark 3 models
• Qwen3-TTS: 0.6B deployed :9300, 9 speakers, on-demand VRAM
• Docling + Reranker: PDF parsing :9400, BGE reranker :9500
• 19 chat commands: /help /nick /who /personas /web /clear /status /model /persona /reload /export /compose /imagine /voice /memory /context /rag /stats /uptime
• Markdown rendering: marked + DOMPurify in chat
• Smart routing: 5 topic domains (music, philosophy, tech, arts, science)
• Dynamic context: 4k-32k adaptive window
• Perf instrumentation: 6 labels, p50/p95/p99 endpoint
• Error telemetry: 16 error labels, structured error logging
• Zod validation: 19 schemas on all API routes
• MIME magic bytes: file-type validation, SAFE_MIMES allowlist
• WS reconnect: exponential backoff 1s-30s, seq gap detection
• CRT boot animation: modem dial, phosphor warmup, scanline reveal
• Chat virtualization: react-window, variable row heights
• 425 tests, 0 failures across all packages
• 12 services in production docker-compose
• MCP server: 4 tools, stdio transport
• Discord bridge: text + voice bot
• Timing-safe token: crypto.timingSafeEqual for admin auth
• Audio size limit: 50MB enforced
• Architecture docs: 4 Mermaid diagrams
• Health check: bash TUI, 19 service checks

Future work (lots 95-100)

• P1 E2E Playwright — 14 tests instruments + commands + DAW (lot 401)
• P2 Persona DPO — scripts/dpo-pipeline.js (lot 510)
• P2 lot-97: Multi-channel support (/join, /channels, channel selector UI)
• P3 lot-98: File sharing — POST/GET/DELETE /api/v2/media/shared (lot 536)
• P2 Mobile responsive — safe-area-inset, 44px tap targets, 380px/landscape
• P3 lot-100: Public demo mode — guest_N auto-nick, read-only until /nick

P14 Lot 24 — Deep Analyse 3 + Reactivity [done]

Phase A — Fixes live session 2026-03-19

• Cookie Secure retire (COOKIE_SECURE env, HTTP fonctionne)
• ADMIN_TOKEN=kxkm dans docker-compose + AdminPage champ password
• MediaExplorer fix reponse API ({ok,data} wrapper)
• Historique 20 derniers messages a la connexion WS [HH:MM]
• Streaming chunks temps reel (type "chunk", curseur clignotant)
• Personas paralleles (Promise.all)
• SearXNG JSON active + auto web_search (Sherlock)
• pickResponders detecte mots-cles web → Sherlock
• Timestamps HH:MM sur tous messages
• TTS retire du chat
• Delai inter-persona 2s → 500ms, timeout Ollama 5min → 2min
• /compose progress ticker (feedback 5s, elapsed time, timeout handler)
• /imagine progress ticker (feedback 5s)
• Admin endpoints verifies OK (overview 5ms, personas 33, analytics 326 msgs)
• /compose duration parsing (5-120s, plus hardcode 30s)
• tts-server.py JSON parsing securise (try-catch)
• Audio size limit 50MB (Python + Node)

Phase B — Analyse approfondie

• Analyse code complete: 33 personas, 8 services, 15+ node types, 135+ tests
• 10 findings prioritaires identifies (P0 securite → P3 docs)
• Veille OSS — reference_oss_audio_2026.md + reference_mcp_agents_2026.md
• Audit docs — TODO.md 244 done, FEATURE_MAP + ARCHITECTURE updated
• Fix tests — 265 pass, 0 fail (lot 506)

Phase C — Livrables

• PLAN.md mis a jour (lots 21-29, statuts corriges)
• TODO.md mis a jour (P14)
• Memoire projet mise a jour
• ARCHITECTURE.md — DAW/AI Bridge subgraph added (lot 507)
• README.md — 17+ services, 112+ cmds, openDIAW.be, MCP, A2A (lot 507)
• Script diagnostic TUI — scripts/health-check.sh (19 checks)
• docs/OSS — 25+ projets documentes (Chatterbox, Stable Audio, RAVE, etc.)

Phase D — Prochaines priorites

• P1 Structured logging — pino, all console.log migrated
• P2 Tests integration — integration.test.ts (7 tests, perf/a2a/mcp/rag)
• P2 lot-28: RAG configurable — 4 env vars (CHUNK_SIZE, MIN_SIMILARITY, MAX_RESULTS, EMBEDDING_MODEL)
• P2 lot-29: Systemd units — 12 user units on kxkm-ai
• P2 lot-29 (journald): journald-monitor.sh TUI (colored dots, --watch, --lines, --service) + ops/v2/journald-alerts.conf (INI restart policy, alert conditions, Discord)
• P3 lot-27: CRT effects — CSS barrel distortion + vignette + phosphor glow (lot 538)

Phase Session 2026-03-20 (lots 128-143)

• /changelog, /version commands
• /dice, /roll, /flip fun commands
• /ban, /unban moderation
• /mute, /unmute per-client persona filter
• @mention notification + @persona tab-complete
• Idle auto-disconnect (30min warn, 35min kick)
• Streaming think:false fix
• /whisper private persona, /w alias
• Code block rendering (triple backtick)
• /history, /search, /react commands
• /invite persona, /time, /session commands
• Connection quality indicator (reconnect status)
• IP connection limit (max 5/IP)
• Ollama model fallback (qwen3:4b)
• Persona cooldown (3s inter-persona)
• qwen3.5:9b migration (256K ctx, adaptive thinking)
• ComfyUI local :8189
• Auto-detect image/music generation in natural language

Lot 424-427 — Deep Analysis Fixes (session 2026-03-21)

• Persona memory cache (30s TTL, LRU 50 entries) — ws-conversation-router.ts
• RAG reranker circuit breaker (2s timeout, skip after 2 consecutive fails for 60s)
• Silent catch → logger.warn on channel state save
• Upload binary detection (null bytes in first 512 bytes)
• JSON streaming chunk size guard (100KB max)
• Mention regex depth limit (10 max per message)
• Mascarade circuit breaker exponential backoff (5s/15s/60s)
• Accessibility: aria-labels on vote/reaction/copy/pin buttons
• Accessibility: role="listbox" + role="option" on autocomplete dropdowns
• VoiceChat audio queue cap (10 max, prevents memory leak)
• Reconnect countdown UX ("dans Xs, tentative N/20")
• Wavesurfer.js waveform player in chat audio messages
• Tab completion V2 (fuzzy match 108+ commands + @mentions)
• Mobile responsive deep pass (touch targets 44px, 360px breakpoint)
• E2E Playwright tests (commands + DAW)
• Feature map + Architecture Mermaid updated with DAW/AI Bridge
• TTS sentence-boundary chunking + cleanForTTS
• /translate /tr /collab /persona-create /radio /summarize /mood commands
• openDIAW.be: 9 instruments (Drone/Grain/Glitch/Circus/Honk/Magenta/AceStep/KokoroTTS/Piper)
• AI Bridge: 5 new ffmpeg instrument endpoints
• DawAIPanel: Kokoro TTS + Instruments AI sections

Backlog (from analysis)

• Chat virtualization — 500-msg cap + React.memo (adequate)
• Route prefetch — requestIdleCallback for Compose/Imagine/DAW/Gallery
• WS batching — React built-in batching + processingChain serialization
• Form labels htmlFor — PersonaList (lot 465)
• Admin health — error details + AI Bridge card (lot 463)
• AI suggestion ranking — startsWith priority in command completion
• Media search — text filter on title/source/type in MediaGallery (lot 539)
• Maps LRU — userStats 1000->500, channelSeq 30min cleanup
• Context store — parallel Promise.all for summary + entries
• ws-chat.ts extraction — 523 LOC + 5 modules (4980 LOC total)
• Chatterbox TTS — already primary backend (chatterbox-remote)
• Transformers.js — BrowserSTT whisper-tiny ONNX (lot 432)
• Sound design endpoint — /generate/sound-design (lot 433)
• Claude Agent SDK — scripts/claude-agent.js (5 tools, interactive mode, lot 533)
• IndexTTS-2 emotion control (theater-grade expressiveness)
• RAVE real-time audio style transfer (IRCAM)

Session 2026-03-24 — Lots 534-540+

Fait

• Lot 531 — Context parallel load, command ranking, mobile CSS, route prefetch

• Lot 532 — mascarade SSE streaming + thinking field extraction

• Lot 533 — Grafana + Prometheus monitoring stack, enriched metrics

• Lot 534 — /deepresearch multi-step search agent (OpenSeeker-inspired)

• Lot 535 — mascarade thinking field fix (Python + Node.js double safety net)

• Lot 536 — Grafana dashboard 16 panels + Prometheus multi-machine

• Lot 537 — Cloudflare Tunnel kxkm-ai (remplace lots 538-540)

URLs live

• https://kxkm.saillant.cc — 3615 J'ai pété (kxkm-ai GPU)
• https://kxkm-api.saillant.cc — API directe
• https://kxkm-mascarade.saillant.cc — mascarade LLM orchestrator
• https://kxkm-comfy.saillant.cc — ComfyUI (image gen)
• https://kxkm-tower.saillant.cc — instance tower (hub)
• https://kxkm-grafana.saillant.cc — Grafana dashboards
• https://mascarade.saillant.cc — mascarade tower (Authentik SSO)

Execution Queue 2026-03-24 (P0/P1) [done]

• lot-541: Regex intent extraction fix + 13 tests non-regression
• lot-542: Async ffmpeg (~35 calls unblocked from event loop)
• lot-543: Worker index.ts unified on worker-runtime.ts (583→155 LOC)
• lot-544: Chat render memoization (wordCount + searchMatches useMemo)

Execution Queue 2026-03-24 (P1/P2) [in-progress]

• lot-545: Sync statuts PLAN/TODO/docs/AGENTS (lots 95-100, 178-200)
• lot-546: lot-178 ACE-Step direct API pour /compose duration fiable
• lot-547: lot-180 timeline model v1 (tracks/clips/markers + schema)
• lot-548: lot-194 waveform timeline UI v1 (desktop + mobile)
• lot-549: composition tests renfort (store/remix/export/timeline)