openclaw/workspace/notes/AI_ACTION_PLAN.md

AI_ACTION_PLAN.md — Implementation Phases

---

Phase 1 — Backbone

• Install Redis as system service (native redis-server.service, enabled, running since Jan 25)
• Enable Redis persistence (AOF + RDB) and local-only bind
• Enforce Redis auth (requirepass) and update clients (password in james/.env)
• Enable systemd auto-restart + health checks
• Define channels (james.events.*, james.decisions, james.actions, james.audit)
• Implement audit logging (audit logger in james/tools/audit_logger.py)
• Test pub/sub flow (verified 2026-02-04)

Deliverable: Redis running, events can be published/subscribed

Redis persistence + auto-restart steps (draft):

1. Update james/redis/redis.conf:
   • appendonly yes
   • appendfsync everysec
   • save 900 1, save 300 10, save 60 10000
   • requirepass <strong-pass>
   • bind 127.0.0.1
2. Ensure volume mounts in docker-compose for /data (AOF/RDB files).
3. In james/systemd/redis.service:
   • Restart=always
   • RestartSec=3
   • StartLimitIntervalSec=0
   • ExecStartPre=/usr/bin/docker compose -f /home/alex/clawd/james/docker-compose.yml up -d redis
   • ExecStop=/usr/bin/docker compose -f /home/alex/clawd/james/docker-compose.yml down
4. Reload systemd and enable:
   • systemctl daemon-reload
   • systemctl enable --now redis.service
5. Validate:
   • redis-cli -a <pass> ping
   • kill container, confirm auto-restart

---

Phase 1.5 — Event Schema Validation

• Define JSON schema for universal events
• Add schema checks to publishers (watcher + publish tool + analyst done; mail agent pending)

Deliverable: Invalid events rejected before publish

---

Phase 2 — Node Watchers

• Deploy watcher-agent on all nodes
• Publish system events (CPU, RAM, disk, Docker) (basic watcher in james/agents/watcher/)
• Add event rate limiting + batching (default 30s)
• Add TTL for low-priority events
• Verify multi-node awareness
• Add GPU metrics (if available)

Deliverable: System events flowing into Redis from all nodes

---

Phase 3 — Analyst Agent

• Connect Qwen via Ollama
• Implement JSON-only reasoning
• Publish decisions to james.decisions
• Test classification accuracy
• Enforce Ollama format: json + schema validation retry on parse errors

Deliverable: Events → Analyst → Structured decisions

---

Phase 3.5 — Analyst Hardening

• Build prompt test suite (edge cases + false positives)
• Capture accuracy/confidence baselines
• Add retry policy for malformed JSON (max 2 retries)

Deliverable: Stable, measurable analyst output

---

Phase 4 — Policy Engine

• Encode POLICY.md rules
• Enforce approval logic (internal vs external)
• Integrate audit trail
• Test forbidden action blocking
• Implement approval timeouts via Redis TTL (james.pending default deny)

Deliverable: Actions gated by policy, logged to audit

---

Phase 4.5 — Policy Dry-Run

• Add dry-run mode (log allow/deny without enforcement)
• Review dry-run logs and tune rules

Deliverable: Tuned policy rules with low false positives

---

Phase 5 — Discord Control Plane

• Build approval bot
• Reaction-based action gating (👍/👎)
• Role-based permissions
• Status dashboard

Deliverable: Human can approve/deny actions via Discord

---

Phase 6 — Home Assistant Integration

• Status sensors (sensor.james_node_*)
• Push notifications
• Critical alert automations
• Optional TTS for urgent events

Deliverable: Home Assistant aware of James state, sends notifications

---

Phase 7 — Mail Agent

• IMAP/Gmail fetch
• Classification (action required, invoice, security, FYI)
• Draft replies
• Approval loop
• Add mail.draft.review step with mandatory human approval
• Pre-filter sensitive patterns (secrets/PII)

Deliverable: Daily inbox scan, human-approved replies

---

Phase 8 — Voice Layer

• Cloud model integration (GPT-4 or Claude for voice)
• Briefing and escalation mode
• Daily system summary
• Optional ElevenLabs TTS
• Hybrid TTS: local for routine, cloud for complex

Deliverable: Natural language interface for status/briefings

---

Phase 9 — Security + Metrics

• Redis network isolation (localhost/VPN only)
• Audit log immutability (append-only)
• Metrics: events/hour, confidence dist, approval latency, false positives, cloud tokens

Deliverable: Hardened ops + visibility

---

Success Criteria

• ✅ Zero unaudited external actions
• ✅ Sub-1000 cloud tokens per day
• ✅ Multi-node awareness
• ✅ Human-approved communications
• ✅ Reliable event classification

---

Current State (as of 2026-02-04)

• Home Assistant running (192.168.1.24:8123)
• Ollama + Qwen2.5-coder available
• Docker game servers operational
• Discord integration exists (clawdbot)
• Google OAuth configured (Gmail/Calendar)

Next immediate step: Phase 1 (Redis backbone)