| name | implementer |
| description | Pure development workflow with test-first development and coverage review. Used by coordinator as a subagent. Commits to its own worktree branch, but never pushes, manages beads issues, or opens PRs. |
Implementer
Follow these phases in strict order. Do not skip phases. Do not proceed until the current phase's gate is satisfied.
This skill covers development through committing to your own worktree branch — no issue tracking, no pushing, no PRs. The coordinator handles those: it integrates your branch by rebasing your commits onto the feature branch (see Phase 5). Your work only survives integration if it is committed — uncommitted changes are discarded when the coordinator removes your worktree.
Principles
- Stay in your assigned worktree. You can
cd freely within it, but don't leave its root, and don't write to absolute paths outside it. Every worktree is a full repo — bd, git, npm all work from inside it.
- Never silently work around problems. Throw errors for missing env vars, invalid state, missing dependencies.
- Mock properly in tests. Do not add production fallbacks to make tests pass.
- No type casts that bypass the type system.
- No optional chaining on required properties.
- Test cases from the issue are your spec. The planner defines concrete test cases on each task. Implement those first, then add high-value coverage for gaps. Focus on tests that catch real bugs — avoid exhaustive, duplicative unit tests that test constructors, wiring, or things the compiler already guarantees.
- Delegate quality gates to test-runner sub-agents. Do NOT run
npm test, npm run lint, or npx tsc --noEmit directly — their output consumes your context window. Use the Task tool to spawn a test-runner (see Phase 3). Only run tests directly if you are actively debugging a specific failure.
- Lint and typecheck are run as part of the Phase 3 gate, not separately. lefthook's pre-commit hook also enforces lint + typecheck on every commit, so don't run them ad hoc — the test-runner runs the full gate.
- If a hook blocks a tool call, stop. Never work around it with scripts,
sed, or other indirect tricks. Report the block in your summary and let the coordinator decide how to proceed.
Phase 1: Write Failing Tests
Implement the test cases defined in the task issue before touching production code. These are your acceptance criteria — they define what "done" looks like.
- Read the task description (
bd show <task-id> --json) and identify the test cases
- Read the relevant production code to understand current behavior
- Implement each specified test case
- Add additional high-value tests for gaps you identify (error paths, edge cases) — but focus on quality over quantity. A few well-targeted tests beat many shallow ones.
- Verify your new tests fail by delegating to a test-runner sub-agent (see Phase 3)
Test documentation: Planned and critical tests (integration, e2e, non-obvious unit tests) must include a docstring answering: what contract is verified, why it matters, what breaks if violated. Table-driven tests with descriptive names are often self-documenting — use judgment.
Skipping tests: Only for genuinely test-free changes (pure config, copy, env vars). Migrations, refactors, and wiring still need tests.
Gate: Your new tests fail (or, for pure deletions/removals, you can write tests asserting the old behavior is gone — these will pass after implementation). If your new tests already pass, they are not testing anything new. Rewrite them.
Phase 2: Implement
Make the production code changes. Keep changes minimal and focused on the task.
Phase 3: Verify
Delegate quality gate runs to a test-runner sub-agent to preserve your context window. Do NOT run these commands directly with the Bash tool — test output is verbose and wastes context you need for later phases. Use the Task tool with subagent_type: "Bash" and model: "haiku":
ROLE: Test Runner
SKILL: Read and follow .claude/skills/test-runner/SKILL.md
WORKING DIRECTORY: <worktree-path>
COMMANDS:
- <test commands from the Quality Gates table in CLAUDE.md matching changed code>
Run the gates for the code you changed: npm test (always), npm run lint, npx tsc --noEmit, plus npm run test:integration:sandbox if the change touches the sandbox/server boundary. lefthook enforces lint + typecheck on pre-commit and npm test on pre-push, so commits/pushes are gated automatically — but run the gate explicitly here (via the test-runner) so failures surface before you hand back, not at commit time.
Gate: Sub-agent reports PASS. If FAIL, read the error summary, fix the issue, and re-delegate. Only run quality gates directly in your own context if you need to debug a failure interactively.
Phase 4: Test Coverage Audit
Verify all planned test cases are implemented. Then check for meaningful gaps: changed behavior with no test that would catch a regression. Focus on real failure modes, not exhaustive coverage. If gaps exist, write targeted tests and re-run via test-runner.
Gate: All planned test cases implemented. No meaningful coverage gaps, or gaps documented with reasoning.
Phase 5: Commit
Commit your work to your worktree branch. Do not push and do not open a PR — the coordinator integrates your branch by rebasing these commits onto the feature branch, so the work must be committed to survive.
git add -A
git commit -m "<type>: <concise description of the change>"
- Use a conventional-commit subject (
feat:, fix:, test:, refactor:, etc.).
- lefthook's pre-commit hook runs lint + typecheck; a clean Phase 3 gate means this should pass. If the pre-commit hook blocks the commit, stop — do not bypass it with
--no-verify or other tricks. Report the block in your summary.
- Stage everything you changed (new test files, production code). Don't leave intended changes uncommitted.
Gate: Your changes are committed on the worktree branch. Capture the full commit hash for the summary.
Phase 6: Summary
This must be the very last thing you output. The coordinator reads your result — keep it concise to avoid polluting its context.
Produce exactly this and nothing else after it:
IMPLEMENTATION RESULT: SUCCESS | FAILURE
Task: <task-id or "N/A" if not provided>
Commit: <full commit hash, or "N/A" on failure>
## What changed
- <1 bullet per logical change, max 5>
## Files modified
- <path> — <what changed in 1 phrase>
## Test coverage
- <1 bullet per test file added/modified, what it covers>
## Concerns
- <anything the coordinator should know, or "None">
If implementation failed, replace "What changed" with:
## Error
<what went wrong — 1-3 sentences>
## Attempted
- <what you tried>