🧩 Capabilities

Rust-based semantic vector search over vault .md files; enables retrieve_memory / store_memory style queries across sessions

Cron-driven pipeline: L1.5 auto-intake, dedup by hash, merge assistant, promote pending→patterns; persistent cross-session knowledge maintenance

Janitor computes vitality scores, decays stale patterns, bumps accessed ones; distillation-rules guide LLM quality promotion; prevents knowledge rot

Session-flush pipeline: compacted context tail → LLM distillation → wiki.md auto-ingest with salience tagging; closes the context→memory write path

In-process sentence-transformers (BAAI/bge-small-en-v1.5) on prime1:18900; 27489 chunks from 2142 memory files; LAN-accessible for voice-bot snippet injection

run_eval.py --staging + direct_file_recall(); JANITOR_VAULT_OVERRIDE for isolated janitor runs; measures PRE/POST recall without QMD drift; closes Rex Phase 0 autoharness blocker

Full Paw→Javus→Rex→Paw review loop over Discord — Javus and Rex post directly, Paw receives via allowBots; no human bridge needed

roll-call.sh orchestrates ready-check across all agents before meetings; writes a short-lived token that gates meeting-brief dispatch

4-pane Javus dispatch pool with file-signal state (.running/.idle/.unknown); symlinks to vault agent-runtime; dynamic round-robin with drain queue

[ack:TOKEN] echo-back rule ensures Javus task receipt is observable; .dispatched/.running/.done/.failed state files in ~/.claude/state/javus-inbox/ close the task lifecycle

Three-layer stale prevention: Paw 5-min SLA after Rex verdict, check-rex-pending.sh cron watchdog, last_progress_at in active-task.json

claude setup-token generates long-lived OAuth token (~1yr); javus-channels-shadow.sh auto-injects CLAUDE_CODE_OAUTH_TOKEN on startup; eliminates manual /login

LaunchAgent fires every 5 min; checks active-task.json last_progress_at; alerts to task thread if running >30 min with no progress update

BullMQ-semantics SQLite queue for Paw→Javus dispatch; crash-safe (pending→dispatched→running→completed/failed); send-to-javus-minions.sh producer + minions-consumer.sh lifecycle

Adds role column to minions.db; send-to-rex-minions.sh gives Paw→Rex the same crash-resilient queue as Javus; Rex Discord listener retired, Rex tmux-only

send-to-zen-minions.sh (role=zen, dev:4.1) extends minions queue to Zen (WeChat/Discord creative tasks); shim-lock guard prevents collision with Hermes-Zen dev:4.1 sharing

Adds 'cancelled' status + cancel_reason column (v7 schema migration); minions-cancel.sh with exit-2 sentinel; idempotent cancel; late complete on cancelled job → reject

Node.js HTTP server on 127.0.0.1:7860 bridges Claude Code PermissionRequest hooks to Discord; Roger can approve/deny tool calls remotely via 'approve <id>'

Pre-compaction snapshot hook captures session tail before runtime trims context; snapshot filename matches janitor intake format for auto-promotion; prevents knowledge loss on context overflow

Migrates all OpenClaw cron jobs from anthropic API billing to Max subscription via launchd + cron-dispatch.sh → send-to-javus.sh → claude -p; eliminates API cost for automated tasks

Adds --via paw-pane to cron-dispatch.sh routing cron tasks to dev:1.1 via tmux; briefs persist in state dir for async pane pickup; offloads cron to dedicated pane

error_intake.py records behavioral errors into error-bank.json; janitor escalates Tier-3 patterns for autoresearch; foundation of all G1/G2 governance

Called async from minions-consumer.sh cmd_complete; Flow A: rex job done → parse verdict → append to task-verdicts.jsonl; Flow B: javus job (GRR auto-ack) → dispatch lightweight Rex review for corpus signal

Adds swimlane field (system_logic/external_integration/human_decision/ai_judgment) to error-bank; backfilled 270 rows; janitor reports per-lane failure distribution; tier3 thresholds now per-swimlane

dispatch_tier3_autoresearch() in memory_janitor.py; 7-day cooldown per error category; writes inbox entries for Paw pickup; closes G1 loop: errors automatically escalate to research

Samples ai_judgment errors from error-bank (category-stratified, 24h filter, AUDIT_MIN_COUNT>=5); builds Rex brief with 5-step protocol per event; writes JSONL audit trail; daily cron 09:00 UTC

Paw-native OD artifact generator: 4 layers (official-system.ts + tokens.css + SKILL.md + deck-framework.ts); claude CLI -p Max sub backend; outputs 16KB HTML + 72KB PDF decks with Linear design tokens

Backward-decomposition dispatch format: each sub-goal has an immediately verifiable concrete assertion; per-checkpoint persist; inspired by BES/GraphFlow/grandmap batch-100 crash lesson

Self-contained HTML dashboards over 12,182-event JSONL stream; swimlane timeline, phase flow, decision chains, graph explorer; living visualization of the Paw system's own evolution