Skip to content

STATUS.md

Latest commit

 

History

History
113 lines (84 loc) · 11.1 KB

STATUS.md

File metadata and controls

113 lines (84 loc) · 11.1 KB
Raw

Robin — Implementation Status

Snapshot: 2026-06-16 Build state: typecheck + lint clean (Phase D code); full suite green aside from the known network-dependent flakes (spotify ×2, ebird, recall)

Note: the "What's running" / model detail below is stale (predates the Opus-4.8-only switch and Phases A–D); treat subsystem specifics as historical until a full refresh.

Phase D — domain-gated memory ingestion (2026-06-16)

Dev/engineering noise is filtered at the ingestion boundary by inverting an unbounded dev blocklist into a closed personal-domain allowlist (system/brain/memory/domains.ts, 11 domains). The biographer extracts only personal-domain claims/entities (tagging each), pure-dev sessions are skipped at capture (looksPureDev), and the belief-promotion gate rejects explicit non-personal domains (NULL grandfathered). Runtime kill-switch: biographer.domainGating in policies.yaml. Retroactive cleanup: robin memory degate [--apply] [--llm] (reversible reject, dry-run default) — one-time run culled ~92 engineering candidates from the legacy backlog while preserving Kevin's ventures/projects. Spec/plan: docs/design/2026-06-16-domain-gated-memory-*.

What's running

Robin runs as a single-process daemon on macOS via launchd (KeepAlive). Current reference setup: M5 Max 64 GB, Ollama with qwen3.5:35b-a3b (MoE, 3B active params) for reasoning/summarize and qwen3-embedding:8b for embeddings.

The full loop is operational: Claude Code sessions are captured via a session-end hook, the biographer extracts entities and relations, embedder indexes content for vector search, the dream job runs nightly for prediction resolution + journal generation, and 18 integrations tick on their cron schedules.

Subsystems

Memory + cognition

  • Event store: append-only events + events_content with 4096-dim embeddings (sqlite-vec, Matryoshka)
  • Recall: hybrid FTS5 + vector search, graceful degradation when embedder unavailable
  • Entity graph: ~5000+ entities, ~14000+ relations, extracted by the biographer
  • Beliefs: topic-keyed claim supersession via believe / recall_belief (with ROW_NUMBER ordering for deterministic latest-per-topic). Belief lifecycle (P1–P4): provenance classification (first-party/inferred/third-party/external/unknown), verified_at freshness tracking, confidence decay for inferred claims, selective tagging in session-start primer (only suspect beliefs annotated)
  • Predictions: confidence tracking + Brier calibration (computed at resolution in resolve_prediction), resolved by the dream job
  • Corrections: correction log with optional topic link; topic-linked corrections auto-retract the contradicted belief nightly (the ONE sanctioned auto-repair path)

Cognition jobs

  • biographer.run — multi-tick entity/relation extraction from captured sessions. Chunks sessions at 10k chars, processes ≤10 chunks per tick, persists progress in biographer_progress. Circuit-breaks on unreachable LLM (no empty markers on Ollama outage).
  • embedder.run — deferred embedding of events_content rows (every minute, single-flight Ollama)
  • dream.run — nightly at 03:00 local: prediction resolution, metrics rollup, journal generation, belief-freshness scan (flags stale beliefs + bounded resolver re-query), corrections→belief replay (auto-retracts topic-linked corrected beliefs)

Surfaces

  • robin-core MCP (16 tools): recall, remember, believe, recall_belief, find_entity, get, list, predict, record_correction, audit, explain, health, metrics, journal, power, skill
  • robin-extension MCP (~14 tools): per-integration action dispatchers + run, integration_status, ingest, related_entities, resolve_prediction, check_action, update
  • CLI: init, doctor, daemon, status, pause/resume, incognito, offline/online, db, import, reindex, upgrade, publish, published, reauth, integrations, hooks, mcp, beliefs backfill-provenance, ingest-archive
  • HTTP: health endpoint + session-end hook receiver (port 41273)

Skills

  • System skills (ship with package): skill-authoring, memory-curation, web-research
  • User skills: user-data/extensions/skills/<name>/ (gitignored, user-shadows-system)
  • Catalog embedded in the skill MCP tool description for progressive disclosure

Integrations

  • 9 built-in: gmail, google_calendar, github, linear, chrome, weather, finance_quote, claude_code (session capture), notify
  • 9 user extensions: whoop, spotify, ebird, lunch_money, letterboxd, lrc, nhl, shipments, spotify_write
  • Hot-reload via chokidar file watcher

Scheduler reliability

  • Cron self-re-arms after completion (Bug C fix — cron never silently dies)
  • In-process lease reaper (60s interval, Bug B fix)
  • Dead-worker lease recovery at boot (controlled-restart path)
  • Heartbeat monitor → 30-min sustained-CRITICAL gate → exit(1) for launchd respawn (Bug A fix)
  • withTimeout on every LLM call (Bug E/F fix — no handler can hang indefinitely)

Test coverage

413 tests across the full stack. Highlights:

  • Biographer: multi-tick, user-shadows-system, circuit-breaker, entity filter, disambiguation
  • Skills: loader, MCP tool, built-in validity
  • Scheduler: cron re-arm, lease recovery, dead-worker recovery
  • Memory: recall modes, FTS5 sanitization, belief supersession, embedding pipeline
  • Integration: loader, scheduler-glue, init-failure resilience
  • End-to-end: foundation smoke (init → doctor → daemon → job → stop)

Known gaps (deferred)

  • Interactive robin init — TTY prompts + OAuth device flow + model pulling. --yes (non-interactive) covers daily use.
  • APFS snapshots for robin db backup — current VACUUM-INTO backup works; APFS needs elevated permissions.
  • Job retention/pruning — completed job rows accumulate (~1500/day from embedder). No auto-prune yet.
  • Multi-account integrations — one instance per integration name.

Recent changes (2026-06-11) — Phase C: memory-quality pack

Spec: docs/design/2026-06-10-trust-feedback-memory-design.md §C1–C4; plan: docs/design/2026-06-11-memory-quality-plan.md.

  • Belief canonicalization (C1): canonicalizeTopic() (negation/modifier stripping) inside believe() — the single write choke point — with a cross-slug claim-similarity gate (levenshtein < 0.4; false merges worse than duplicates); recall_belief canonical-first lookup symmetry; one-time sweep robin beliefs canonicalize [--apply] (retraction-based collapse, belief.canonicalize audit events). Live sweep 2026-06-11: 23 groups → 11 merged, 12 conservatively skipped.
  • Risk-weighted freshness (C2): runBeliefFreshness re-queries the top-N stale heads by score (uncertainty + over-age + correction history) instead of the first-N lottery; same maxRequeries spend.
  • Claim dead-letter retry (C3): claim_failures table (migration 026); biographer dead-letters extraction timeouts/validation failures (verbatim chunk); the nightly dream pass retries (max 3 attempts, 5/pass), prunes exhausted rows after 30d, and fires/resolves a Phase-A backlog alert (>10 open).
  • Entity profile staleness (C4): profile_generated_at (migration 027, backfilled); stamped on every profile write (upsert, dream, merge); dream's spare summary budget regenerates stale profiles of active entities; find_entity/get/related_entities serve >30-day profiles as deterministic relation summaries (profile_stale: true), never as current truth.

Previous changes (2026-06-11) — Phase B: agentic outcome loop

Spec: docs/design/2026-06-10-trust-feedback-memory-design.md §B1–B5; plan: docs/design/2026-06-11-agentic-outcome-loop-plan.md.

  • Structured outcomes: every agent handler (A–L) requests a shared outputFormat envelope (outcome/changes/impact/notes, +2 maxTurns headroom); runner-entry persists outcome/impact/structured_json/verified onto the run's agent_usage row (migration 025).
  • Deterministic verification (no LLM): per-handler post-condition checks (system/agent/verifiers.ts); a did-work claim that fails its verifier records outcome-mismatch and fires a Phase-A alert (auto-resolved by the handler's next verified run).
  • Handler B ingest-only write: research briefs land in memory as research.brief events via the ingest MCP action (default mode + write builtins denied — plan mode blocks allowlisted MCP writes).
  • Autonomous K worktree isolation: write-to-repo handlers get the same throwaway-worktree flow as robin agent (kept on changes, pruned otherwise).
  • Adaptive dispatch: deterministic pre-checks skip the SDK spawn when a handler has nothing to do; 3 consecutive failures bench a handler for 3 rotations + Phase-A alert; bench clears only on observed post-bench success. State in user-data/state/runtime/agent-runner-adaptive.json.
  • ROI surface: robin metrics --agents + MCP metrics {agents:true} — per-handler runs, spend, outcome distribution, last did-work.
  • Learning records generalized: all autonomous handlers write agent-runs/<ts>-<handler>.md (no-op and pre-flight-capped runs skipped).

Previous changes (2026-06-10) — Phase A: trust & alerting

  • Persistent alerts table (migration 024, dedup-by-open-(source,key), severity escalation, auto-resolve, ack); integration staleness/skip-streak/degraded-stream invariants with power-state suppression; tick/job/crash-loop capture; check timeouts + overlap guard; surfaces: robin alerts, MCP alerts tool, doctor freshness table, morning-brief health section.

Previous changes (2026-05-25)

  • Belief lifecycle (P1–P4): provenance policy module (provenance.ts), recall enrichment (confidence/provenance/age on belief hits), selective suspect-tagging in primer, formation gate (external claims routed out, class-thresholded promotion), nightly freshness scan with bounded resolver re-query, topic-linked corrections→belief auto-retraction. Migrations 013 (belief_candidates.provenance) + 014 (corrections.topic). 25 files, ~1700 lines, 2 migrations. Key safety fix: explicit-supersede writes append instead of same-day upsert (preserves history + prevents self-referencing supersession chain).
  • ingest-archive: bulk text archive → recall (sha-deduped, paragraph-chunked, deferred embedding). CLI robin ingest-archive <dir> --source=name.
  • CI fixes: .npmrc auto-install-peers (lockfile mismatch), build-essential for native modules, pre-existing formatter/import lint errors resolved.

Previous changes (2026-05-23)

  • Skills system: MCP-surfaced skill tool with catalog-in-description, system + user skills, 3 seeded system skills
  • Biographer multi-tick: sessions resume across cron ticks via biographer_progress table (migration 006)
  • Biographer circuit breaker: LLM-unreachable → abort without advancing cursor (prevents empty-marker corruption)
  • CHUNK_CHARS raised 6k → 10k, MAX_CHUNKS_PER_TICK 4 → 10
  • Belief enumerate: ROW_NUMBER window function for deterministic same-timestamp ordering
  • LLM dispatcher: withTimeout on every invoke/embed + per-call timeoutMs override
  • MAX_SESSION_BODY_CHARS raised 200k → 1M (multi-tick makes large sessions safe)
  • Docs reorganization: CONTRIBUTING/RUNBOOK/SECURITY moved to docs/, ARCHITECTURE.md + PUBLISHING.md created