Skip to content

CHANGELOG.md

Latest commit

 

History

History
133 lines (118 loc) · 7.14 KB

CHANGELOG.md

File metadata and controls

133 lines (118 loc) · 7.14 KB

Changelog

All notable changes to cato-daemon are documented here. Format follows Keep a Changelog. Versioning follows Semantic Versioning.

[0.2.0] — 2026-03-06

Summary

Major feature release. Adds 20+ new subsystems across 6 phases of the Cato roadmap (Phases A, B, H, I, J, and gap fixes). All 1271 tests pass. Four independent Kraken audits: 94%, 91%, 90%, 98% confidence — all APPROVED.

Added

Phase A — Token Reduction Infrastructure

  • HOT/COLD skill file split: <!-- COLD --> delimiter in all skill .md files; context builder only loads HOT section (≤300 tokens) by default
  • Query tier classifier (cato/orchestrator/query_classifier.py): routes queries to TIER_A (cheapest), TIER_B (mid), or TIER_C (fan-out all) based on keywords, token count, code blocks, file paths, and confidence
  • Memory distiller (cato/core/distiller.py): compresses conversation history every 20 turns into 400-token SQLite chunks with semantic embeddings
  • SlotBudget (cato/core/context_builder.py): explicit token allocation across 4 tiers (identity 2000, memory 4000, workspace 4000, overflow 2000)
  • HybridRetriever (cato/core/retrieval.py): BM25 + cosine similarity hybrid search over MemorySystem chunks (0.4 BM25 + 0.6 semantic)
  • ContextGate (cato/core/context_gate.py): admission control that enforces slot budgets before context is assembled

Phase B — Top 10 Skills

  • SchedulerDaemon (cato/core/schedule_manager.py): YAML-based cron scheduler with per-schedule asyncio tasks, overlap prevention, hot-reload
  • SessionCheckpoint (cato/core/session_checkpoint.py): token-aware context anchoring — checkpoints session state when usage > 80% of limit
  • MemorySystem (cato/core/memory.py): hybrid BM25 + sentence-transformer long-term memory with ANN index (hnswlib) opt-in at 5000+ chunks
  • Mem0 Fact Store: store_fact/load_top_facts/apply_decay on MemorySystem — persistent fact store with confidence decay
  • Knowledge Graph (cato/core/memory.py): kg_nodes/kg_edges SQLite tables, entity extraction (files, @mentions, CamelCase, ALLCAPS), multi-hop traversal via recursive CTE
  • WebSearchTool (cato/tools/web_search.py): query classifier (code / academic / news / general), multi-engine routing, result deduplication
  • GitHubTool (cato/tools/github_tool.py): REST API wrapper for issues, PRs, file content, repo search — all read-only, no auth required for public
  • PythonExecutor (cato/tools/python_executor.py): sandboxed subprocess execution with dangerous-pattern blocking, timeout, matplotlib capture
  • SkillImprovementCycle (cato/orchestrator/skill_improvement_cycle.py): correction detection, structured ledger, versioned skill backups with SHA-256
  • ClawFlows (cato/orchestrator/clawflows.py): declarative multi-step workflow engine with step chaining, variable substitution, retry logic
  • ContextPool (cato/core/context_pool.py): pre-warmed context assembly for low-latency skill invocations

Phase H — Safety Foundation

  • ReversibilityRegistry (cato/audit/reversibility_registry.py): singleton registry of tool reversibility scores (0.0 reversible → 1.0 irreversible), blast radius classification, custom tool registration
  • ActionGuard (cato/audit/action_guard.py): pre-action gate that enforces confirmation rules based on reversibility × autonomy level thresholds
  • LedgerMiddleware (cato/audit/ledger.py): hash-chained, Ed25519-signed tamper-evident record of every agent action; verify_chain() detects both prev_hash linkage breaks and field-level mutations (re-hash verification)
  • DelegationToken / TokenStore (cato/auth/token_store.py): SQLite-backed scoped delegation tokens with spending ceilings, expiry, revocation
  • TokenChecker (cato/auth/token_checker.py): pre-action scope check against active delegation tokens with wildcard category matching
  • AuditLog (cato/audit/audit_log.py): SHA-256 hash-chained append-only audit log (migrated from cato/audit.py to cato/audit/ package)

Phase I — Epistemic Layer

  • EpistemicMonitor (cato/orchestrator/epistemic_monitor.py): premise extraction from agent reasoning, confidence tracking, gap detection, interrupt budget enforcement, unresolved gap persistence
  • DisagreementSurfacer (cato/orchestrator/disagreement_surfacer.py): multi-model output comparison via Jaccard distance + confidence stdev, 4-way disagreement classification, consensus/minority view identification
  • ContradictionDetector (cato/memory/contradiction_detector.py): Jaccard-gated fact contradiction detection (TEMPORAL / SOURCE / PREFERENCE / FACTUAL), duplicate pair prevention, resolution tracking, health summary

Phase J — Memory & Temporal

  • DecisionMemory (cato/memory/decision_memory.py): structured decision record with premises, confidence, ledger linkage, outcome quality scoring, overconfidence profile, reliable pattern extraction
  • OutcomeObserver (cato/memory/outcome_observer.py): background asyncio task that polls open decisions and applies timeout outcomes; configurable per-action-type observation windows
  • HabitExtractor (cato/personalization/habit_extractor.py): passive interaction observer that infers soft constraints from rejection/acceptance patterns; get_soft_constraints() for skill prompt injection
  • VolatilityMap (cato/context/volatility_map.py): URL/resource type to volatility score mapping; domain override support
  • TemporalReconciler (cato/context/temporal_reconciler.py): wake-up protocol that identifies stale dependencies and generates a structured WakeupBriefing on daemon restart
  • AnomalyDetector (cato/monitoring/anomaly_detector.py): multi-domain weak-signal monitor with baseline tracking, cross-source anomaly detection, self-calibrating false-positive suppression, prediction record keeping

Fixed

  • verify_chain() field re-hash (Kraken 2A Issue 1): verify_chain() now re-computes each row's hash from all fields and compares against stored record_hash, detecting any field-level mutation (not just prev_hash breaks)
  • OutcomeObserver configurable windows (Kraken 2B Issue 1): constructor now accepts optional observation_windows: dict[str, float] parameter; backward-compatible (defaults to module-level _OBSERVATION_WINDOWS)

Tests

  • Total: 1271 passed, 1 skipped, 0 failed
  • New test files: 30 (covering all new subsystems)
  • New E2E smoke test: tests/test_e2e_full_pipeline.py (108 tests covering all Phase A–J integrations end-to-end with tmp_path isolation)

[0.1.0] — 2026-02-15

Added

  • Initial release
  • Core agent loop with asyncio WebSocket streaming
  • Encrypted vault (AES-256-GCM) with cato vault set/get
  • Hard spend caps ($1 session / $20 monthly)
  • Telegram and WhatsApp adapters
  • Patchright browser automation (Conduit)
  • Multi-model coding agent (Claude + Codex + Gemini fan-out)
  • Talk Page UI (cato/ui/coding_agent.html)
  • Basic audit logging
  • CLI process pool for warm model invocations
  • 445 tests passing
⚡