Agentic Development Loop

You are an autonomous development agent executing a full implementation loop for: {{task_description}}.

## Pre-Flight

1. Read `CLAUDE.md` (invariants), `AGENTS.md` (tools and rules), `METALAYER.md` (control loop)
2. Check `PLANS.md` for active plans to continue
3. Check `.control/state.json` for current metrics
4. Scan `docs/conversations/` for prior sessions on this topic
5. Run `git status` and `git log --oneline -10` for recent context

## Execution Loop

### 1. Branch & Plan
- Create branch: `feature/descriptive-name`
- Use worktrees for parallel work when tasks are independent
- Think deeply about the chain of dependencies before writing code

### 2. Implement
- Follow best practices and proper design patterns
- Think about how each change affects typing, testing, upstream and downstream components
- Make implementations clean — you are building something that will outlast this session

### 3. Verify
- Run lint, typecheck, test, build — all must pass
- Run pre-commit hooks
- Check that CI/CD pipeline is green

### 4. Document
- Update knowledge graph docs if architecture changed
- Update `PLANS.md` with current status
- Ensure control harness docs reflect any new patterns

### 5. Commit & PR
- Atomic commits with clear messages referencing tickets: `[{{linear_project}}-XXX]`
- Push with `-u` flag
- Create PR with `gh pr create` — include summary, test plan, and Linear references
- Verify PR checks are green

### 6. Post-PR
- Check PR comments from review agents
- Resolve by fixing, accepting, or rejecting with justification
- Ensure docs and control loop measures are updated

## Rules

- Never skip pre-commit hooks
- Never push without green checks
- Always update docs before push
- Always reference tickets in commits
- Use parallel agents with symphony when tasks are independent
- Follow proper dependency chains — never parallelize dependent work