Changelog

Changelog

Full version history. For detailed commit-level changes, see CHANGELOG.md.

---

v4.0.0 — Cognitive Core & Autonomous Systems Edition (2026-05-29)

Architecture

• Cognitive Core — Replaced implicit ReAct loops with a deterministic async state machine.
• Explicit Execution Graphs — Tasks are now decomposed into persistent Plan, PlanStep, Observation, and Reflection models.
• Dynamic Plugin System — Tools can now be hot-loaded from a plugins/ directory without core code changes.
• Durable Episodic Memory — Every plan step, tool outcome, and agent reflection is persisted in SQL for audit and learning.

Security & Safety

• Tool Policy Engine — Centralized gatekeeper that evaluates RiskLevel and PermissionProfile before any tool execution.
• Local Sandbox — Multiprocessing-based Python executor for Docker-free environments; auto-detects Docker and falls back gracefully.
• Protected Processes — Policy engine blocks termination of critical system processes (explorer.exe, etc.).

Features

• Self-Aware Tools — Added manage_plugins (add/remove capabilities at runtime) and search_codebase (agent can inspect its own logic).
• Network Diagnostic Suite — New tools for IP discovery, host pinging, and connectivity checks.
• Proactive Health Monitoring — Autonomous loop now detects and alerts on critical CPU, RAM, and Battery events.
• Structured LLM Mode — Routing hint that strictly enforces JSON output for high-precision tool argument extraction.

Reliability

• Verification Layer — Explicit state for verifying tool outputs before proceeding to next plan steps.
• Async Queue Manager — Offloads long-running tool executions to background Redis-backed workers.

---

v3.2.1 — Security Hardening & Observability Edition (2026-04-30)

Security

• SEC-01 — Prompt injection resistance: <user_task> XML boundaries + [BLOCKED:TOKEN] neutralisation for ReAct control tokens
• SEC-02 — WhatsApp X-Hub-Signature-256 HMAC verification; forged payloads → HTTP 403
• SEC-03 — Telegram MASTER_TELEGRAM_CHAT_ID allowlist; unknown senders → "Unauthorized."
• SEC-06 — SECRET_KEY auto-generates cryptographically-secure ephemeral key; no more "fallback" literal
• CQ-01/02 — copy_file, move_file, create_folder enforce SEARCH_ALLOWED_DIRS via _assert_in_allowed_dirs()

Reliability

• ARCH-04 — asyncio.Lock() prevents Qdrant client FileLock collision during concurrent startup
• DR-01 — Autonomous loop task stored + done-callback registered; stop() cancels it
• DR-02 — AgentOrchestrator.shutdown() implemented; no more zombie asyncio tasks on restart
• DR-03 — APScheduler uses shutdown(wait=True); in-flight jobs complete before event loop closes
• DR-05 — IPC token corruption (non-UTF-8, empty, OS error) caught specifically with CRITICAL log before regeneration

Observability

• /api/v1/health/live — liveness probe (always 200)
• /api/v1/health/ready — readiness probe (checks DB, Redis, Qdrant, LLM; returns 503 with per-dependency map)
• amadeus_tool_duration_seconds{tool_name,success} Histogram — per-tool latency
• amadeus_tool_executions_total{tool_name,result} Counter — per-tool result breakdown
• amadeus_memory_errors_total{operation} Counter — Qdrant failure visibility

Architecture

• IMessagingAdapter Protocol + InboundMessage dataclass in src/infra/messaging/protocols.py
• P6-T7: Memory deduplication via uuid5(session_id:role:text) — idempotent upserts
• P7-Chaos03: QueueFullError caught in Telegram background handler → friendly "busy" reply
• PC-02: SSE word-by-word → sentence-chunk streaming (~15 words, 4× fewer sleep calls)

Tests

• 20 new unit tests: test_security_hardening.py + test_agent_reliability.py

---

v3.2.0 — Production Architecture Edition (2026-04-29)

Architecture

• AmadeusService decomposed from 1,381-line God-Object → 5 focused sub-services
• ToolRegistry exclusively built by DI container — no double-registration
• ConversationMessage model unified across the codebase

Performance

• Semantic router index built in thread pool — no blocking startup
• asyncio.get_event_loop() → asyncio.get_running_loop() throughout
• execution_history bounded to deque(maxlen=500)

Security

• eval() replaced with restricted AST math evaluator
• /chat/history locked behind current_active_user
• Sentry send_default_pii=False

---

v3.1.0 — Semantic Router Edition (2026-04-22)

Added

• Zero-Training Semantic Tool Router — replaces sklearn SVM classifier, hot-pluggable, cosine similarity threshold 0.50. No retraining ever required when adding new tools.
• Hybrid Workspace Indexer — BM25 + dense vector retrieval fused via Reciprocal Rank Fusion. Incremental builds, context-augmented chunking. Preserves code identifiers like AUTH_UUID_7392 as single BM25 tokens.
• Flash Memory Cache — Tier-1 NumPy float32 ring buffer (100 entries, ~307 KB RAM, threshold 0.85). ~1 µs vs ~5 ms Qdrant round-trip.

Security

• Docker network isolation — Postgres (5432) and Redis (6379) ports no longer host-exposed.
• SECRET_KEY auto-generation via Setup_Amadeus.bat.
• .dockerignore tightened to exclude .env files and model weights.

---

v3.0.0 — Clean Adapter Edition (2026-04-19)

Added

• LLMAdapter abstract base class — all providers conform to a typed interface.
• LlamaCppAdapter with multi-turn memory via ConversationContext.
• OllamaAdapter with OLLAMA_ENABLED flag and safe connection check.

Fixed

• Qdrant UUIDv5 upsert collision bug.
• sklearn version mismatch on Python 3.11+.
• LlamaCppAdapter exception type correctness (LLMError not Exception).

Quality

• Coverage gate raised to 80%.
• bandit 0 HIGH gate added to CI.

---

v2.0.0 — Clean Architecture Rewrite (2026-04-17)

Added

• Full Clean Architecture rewrite: core / app / infra / api layers.
• Multi-LLM router with Redis quota tracking (Groq → Gemini → OpenAI fallback chain).
• FastAPI REST + WebSocket API with JWT auth (HS256).
• Rate limiting via SlowAPI (keyed by JWT sub).
• Telegram, WhatsApp, Email integrations.
• Docker + Railway deployment with multi-stage Dockerfile.
• Alembic migrations.
• APScheduler proactive loop.

---

← Known-Limitations-and-Roadmap | Home →