KevinZai OpenClaw Production Patterns
Battle-tested patterns extracted from a production system running 38 AI agents across 12 workspaces. These patterns work standalone — no OpenClaw dependency required.
1. COMP System (Commit-Operate-Message-Prove)
Every agent action follows this protocol:
COMMIT → Declare what you're about to do (in task tracker)
OPERATE → Execute the task
MESSAGE → Report results (to requester + comms log)
PROVE → Provide evidence (screenshot, test output, diff)
Why it works: Creates an audit trail. Any agent can pick up where another left off. Eliminates "I thought I did that" failures.
How to apply: Before any multi-step task, write your plan to tasks/todo.md (COMMIT). Execute each step (OPERATE). Update the task with results (MESSAGE). Include verification output (PROVE).
2. Daily Micro-Commits
Commit frequently with small, atomic changes:
- Every function/component gets its own commit
- Commit message references the task ID
- Never batch more than 30 minutes of work
Why: Enables easy rollback, clear git blame, and lets other agents understand changes.
3. Atomic Messaging Protocol
When agents communicate:
- Every message includes: sender, recipient, action, context, expected response
- Every send pairs with a log entry (audit trail)
- Messages are self-contained (recipient shouldn't need to read prior context)
How to apply in Claude Code: When using subagents, give complete context in the prompt. Don't assume the subagent has your conversation history.
4. Cost-Aware Routing
Route tasks to the cheapest capable model:
| Task Type | Model Tier | Examples |
|---|---|---|
| Simple transforms | Free (Ollama, CF Workers) | Format conversion, template fill |
| Code generation | $ (Flash, Groq) | Boilerplate, tests, docs |
| Complex coding | $$ (Sonnet) | Architecture, debugging, review |
| Deep reasoning | $$$ (Opus) | System design, security audit |
Rule: Always try the cheaper tier first. Escalate only if quality is insufficient.
5. Escalation Boundaries
Clear rules for when to escalate:
- Stay at current level: Task is within scope, no blockers
- Escalate up: Task needs 2+ agents, spans multiple sessions, needs human approval
- Escalate to human: Security concerns, destructive operations, budget exceeded, architectural decisions
6. Role-Based Permissions
Each agent/mode has explicit boundaries:
- What files it can read/write
- What commands it can execute
- What external services it can call
- What budget it can spend
Apply via: Claude Code's permissions.allow and permissions.deny in settings.json.
7. Profile Injection
Load context based on the task, not the session:
- Workspace-specific instructions (like modes in CC Commander)
- Task-specific tool access
- Time-bounded permissions (night mode = more autonomy)
8. Session Handoff Protocol
When a session runs out of context or needs fresh start:
- Save state:
/ccc-save-sessionor auto via pre-compact hook - Write handoff note: what's done, what's next, any blockers
- New session reads: CLAUDE.md → tasks/todo.md → session file → resume
This is exactly what CC Commander's hooks automate: context-guard warns at 70%, pre-compact saves before compaction, session-coach gives periodic guidance.
Integration with CC Commander Components
| OpenClaw Pattern | CC Commander Equivalent |
|---|---|
| COMP system | tasks/todo.md + /verify |
| Micro-commits | auto-checkpoint hook |
| Cost routing | Mode system (normal/yolo/night) |
| Escalation | confidence-gate hook |
| Session handoff | pre-compact hook + /ccc-save-session |
| Role permissions | Mode-based permission configs |