| name | loop-engineering-dev |
| description | Use when designing or running autonomous AI development loops that combine GitHub Issue-driven development, test-board.yaml test inventories, scheduled Codex/Claude runs, maker/checker separation, Draft PRs, E2E evidence, and regression impact analysis. Trigger on ループエンジニアリング, loop engineering, 自走開発, schedule/scheduled coding agent, AIが独立して実行, Issueとテストを合体, test-board plus GitHub Issues, unattended development loop. |
| metadata | {"short-description":"GitHub Issue + test-board による自走開発ループ設計"} |
Loop Engineering Dev
Use this child skill to design a development loop, not a one-off prompt.
Parent Skill
Before acting, read ~/.agents/skills/loop-engineering-core/SKILL.md if it exists. If it is missing, continue with this Skill and report that the parent Skill was unavailable.
The parent owns shared rules: requirement gate, durable state, maker/checker separation, hard stops, and scheduled prompt shape. This child owns the GitHub Issue + test-board development-loop specialization.
Non-Negotiable Rule
Follow the parent requirement gate. If the question set becomes large, use both the parent references/requirements-gate.md and this Skill's references/requirements-question-bank.md.
Specialized Building Blocks
- Work source: GitHub Issues, CI failures, analytics signals, bug reports, or explicit backlog.
- State file:
test-board.yaml and optionally .loops/<loop-name>.md.
- Execution environment: local worktree, Codex schedule, Claude routine, or CI runner.
- Objective gates: tests, lint, build, E2E, security checks, and test-board impact analysis.
- Human stop point: Draft PR. Never auto-merge or deploy production changes without explicit approval.
Setup Flow
- Requirement definition gate:
- Restate the product goal.
- Define users, workflows, non-goals, risk, security boundaries, data sources, cost limits, and failure policy.
- Ask unresolved questions before proceeding.
- Convert answers into acceptance criteria.
- Issue and test planning:
- Create or reuse a GitHub Issue with Japanese title/body when the repo workflow requires it.
- Create linked
test-board.yaml cases at the same time.
- Include unit/integration/E2E/security/regression cases as appropriate.
- Loop design:
- Choose one loop type: discovery, implementation, regression monitor, CI fixer, E2E runner, or release verifier.
- Set one-run scope: normally one Issue or one test-board case.
- Define hard stops: max PRs per run, max attempts, max cost, max session duration, forbidden files/areas.
- Implementation run:
- Pick one ready Issue/case.
- Create or switch to an issue-numbered branch.
- Write failing behavior test first when feasible.
- Implement minimally.
- Run gates and checker.
- Open/update Draft PR and record findings in the Issue/test-board.
- Scheduling:
- Schedule only after a manual dry run succeeds.
- Prefer event-driven triggers for sporadic work and low-frequency schedules for backlog sweeps.
- A scheduled prompt should call a skill or loop file, not embed a giant stale instruction wall.
See references/loop-architecture.md for loop patterns and guardrails.
Combined GitHub + Test Workflow
Use github-issue-driven-dev for Issues/branches/PRs and test-case-board for test inventory.
Minimum per Issue:
- Purpose and scope.
- Acceptance criteria.
- Linked test-board cases.
- Test/CI plan.
- Knowledge log.
Minimum per test-board case:
- Stable
TC-xxx ID.
issues: [n].
source.
status.
type.
- Real
test_file once implemented.
E2E And Auth
For Browser Use Cloud E2E, load:
~/.agents/skills/test-case-board/references/e2e-browseruse-cloud.md
For Testmail login/auth tests, load:
~/.agents/skills/test-case-board/references/login-testmail.md
Auth tests should include fresh-user login, existing-user login, invalid/expired OTP, authorization denial, admin allowlist/role approval, logout, and session expiry where relevant.
Loop Prompt Shape
Scheduled runs should be small and explicit:
Use loop-engineering-dev.
Run one implementation-loop iteration for repo <owner/repo>.
Source of truth: test-board.yaml and GitHub Issues.
Pick exactly one ready todo case or ready Issue.
Create at most one Draft PR.
Do not merge or deploy.
Stop and report if requirements are unclear, gates fail twice, secrets are missing, or cost/concurrency limits are reached.
Completion Definition
A loop setup is complete only when:
- Requirement questions are answered or explicitly deferred.
- Issues and test-board cases exist.
- One manual dry run has produced either a Draft PR or a clear blocked report.
- Gates and evidence are recorded.
- Schedule/automation has conservative limits and can be disabled.