implement

---
name: implement
description: Implement a single planned issue, optionally inside a bounded autonomous loop. Use to execute one unit of an approved plan.
---

<!-- GENERATED from commands/implement.md by bin/build-codex-skills.py — do not edit here. -->

Implement the work described in an issue or task.

## Arguments
- **issue**: Issue ID (e.g., Linear identifier like `CAR-123`) or path to a local plan `.md` file — required
- **--autonomous**: run inside the built-in **bounded loop** — iterate implement→verify until the acceptance criteria and tests pass, **or** the cap (3) is reached, then **escalate**. Off by default (a single pass).

## Steps
1. Read the issue or task file.
   - If a Linear ID is provided and Linear MCP tools are available, read the Linear issue.
   - If the issue has a parent, read it for broader context and design decisions.
   - If a local `.md` path is provided, read that file.
   - If a matching Claude Code Task exists (find it via `TaskList`), read it with `TaskGet` for the latest state and dependencies.
2. Mark work as started:
   - If Linear is available, set the issue status to **"In Progress"**.
   - If a Claude Code Task tracks this work, `TaskUpdate` it to `in_progress`. This is what makes the work resumable across sessions (see `/resume`).
3. Implement the changes described in the issue, following the acceptance criteria.
4. Mark work complete:
   - `TaskUpdate` the Claude Code Task to `completed` **only** when the work is fully done (no failing checks, no partial implementation). If blocked, keep it `in_progress` and note the blocker.
   - If Linear is available, advance the issue status per your team's flow.
5. After implementation, present a summary of what was changed and which files were touched.

## Autonomous mode (`--autonomous`)
When `--autonomous` is set, wrap steps 3–4 in the **bounded convergence loop** (the same cap-3 mechanism `/run` Phase 4 uses):
1. Implement the changes.
2. **Verify in a separate verifier subagent** (never self-verify — see the rule in `run.md`): run the project's tests and `/review-code` against the diff; collect failures/findings.
3. Clean → mark the task `completed` and stop. Not clean → fix and repeat from step 1.
4. **Cap: 3 iterations.** If failures remain after 3, **escalate**: stop, summarize what's blocking and the fixes attempted each iteration, and hand back. **Never** mark the task `completed` past unresolved failures, and never auto-push.

Sensitive work still runs `/secure` and stops at the final commit/push approval. Without `--autonomous`, do a single pass (steps 1–5 above) and stop.

## Notes
- Follow all project conventions from CLAUDE.md or similar project documentation.
- Only make changes described in the issue. Keep solutions simple and focused.
- If the issue description is ambiguous or missing information, ask the user before writing code.