# 2026-01-26

## Session: Claude Code + Clawdbot Integration Analysis

### What Happened
- William asked me to analyze his Claude Code setup (`~/.claude/`)
- Did deep dive on architecture: multi-agent hierarchy, state files, skills, workflows, etc.
- Discussed tradeoffs between Claude Code (heavy governance) vs Clawdbot (flat/agile)

### Key Insights
- Claude Code has good patterns: guardrails, KB facts, component registry, tiered model delegation
- But also overhead: 10+ state files at init, rigid orchestration layer, non-executable workflows
- Clawdbot is leaner: on-demand skills, simple memory files, cron/heartbeats

### Changes Made

#### Clawdbot Improvements
1. **TOOLS.md** — Added full homelab infrastructure knowledge (k8s cluster, nodes, IPs, services, namespaces, AI stack, workstation details, key repos)
2. **USER.md** — Updated with William's role (cloud engineer), expertise (k8s, GitOps), technical preferences, working style
3. **AGENTS.md** — Added comprehensive guardrails section (blocked commands, confirm-first commands, safe paths, sensitive paths)

#### Claude Code Improvements
1. **component-registry.json** — Removed TODO markers (`/README`, `README` agent)
2. **personal-assistant.md** — Rewrote initialization section:
   - Split into "Required at Session Start" (3 files) vs "Load On-Demand" (8 files)
   - Added fast-path routing section (skip orchestration for simple requests)
   - Added confidence-based routing (high/medium/low → act/confirm/ask)
3. **future-considerations.json** — Added fc-061 (PA fast-path routing)

### Decisions
- Don't merge the two systems wholesale — cherry-pick good patterns
- Claude Code for CLI/coding work, Clawdbot for messaging/assistant
- Each tool for what it's good at

### LLM Routing Strategy (added later in session)
- Created LLM-ROUTING.md with comprehensive guide
- Three CLIs available: claude, opencode (GitHub Copilot), gemini
- Tiered approach: haiku/flash (cheap) → sonnet (balanced) → opus/gpt-5.2 (powerful)
- GitHub Copilot models are "free" with subscription — use for one-shot tasks
- Gemini for long context (1M+ tokens)
- Updated TOOLS.md and AGENTS.md with LLM routing info
- Updated Claude Code's model-policy.json with CLI quick-reference

### Google Workspace Access
- Tested Gmail/Calendar access methods
- `gog` CLI is better than Claude Code's Python scripts (cleaner, more flexible)
- Gmail: ✅ working via `gog gmail search`
- Calendar: ✅ authorized via `gog auth add ... --services calendar` (shows no events next 7 days)
- Tasks: ✅ authorized via `gog auth add ... --services tasks`; can list task lists + tasks
- Updated TOOLS.md to prefer gog for Workspace operations

### Integrations
- WhatsApp gateway connected (2026-01-26)

### Claude Code Hooks (observed)
- hooks configured: UserPromptSubmit → prompt-context.py; SessionStart → session-start.sh; PreCompact → pre-compact.sh; SessionEnd → session-end.sh; PreToolUse(Bash|Write|Edit) → guardrail.py

### Next Steps
- Decide whether to add Google Voice support via a separate tool/server (not supported by gog)
- Could add heartbeat-like proactive checks to Claude Code
- Could port rag-search skill concept to Clawdbot
- Monitor if fast-path routing actually helps in practice
- Test LLM delegation in practice