Claude Code skills for handing a single GitHub issue to an agent, implementing it in an isolated worktree, and getting it back review-ready — uncommitted, with a second-pass code review.
The human owns the commit and the PR. The agents never do.
agent-implement-laravel/
SKILL.md # Laravel/Sail-specific: worktree, implement, validate, review
agent-implement-generic/
SKILL.md # stack-agnostic: detects the project's own tooling
~/.claude/skills/agent-implement-laravel and
~/.claude/skills/agent-implement-generic are symlinks into this repo — edit
the skills here, not the installed copies.
Laravel/Sail-specific. Takes one well-scoped issue and runs it end to end:
- Provision — isolated worktree at
worktrees/<project>-issue-<N>on branchfix/issue-<N>, withsail up -dandnpm run devrunning. - Implement — stays in scope; escalates to the
grill-with-docsskill when acceptance criteria, approach, or technical detail are ambiguous. - Self-verify — runs the
tddskill to red-green the fix before handing off. - Validate —
./vendor/bin/pint,./vendor/bin/phpstan analyse, and./vendor/bin/sail artisan testmust be clean. - Review pass — spawns an Opus agent that chains into the
/code-reviewskill via theSkilltool (rather than inlining its instructions), fixes every Critical/Important finding, then re-runs validation. - Stop — leaves everything uncommitted.
Same flow, stack-agnostic. Detects the project's own dev/start command
(docker-compose.yml, Makefile, package.json scripts, Procfile, …) and
its own format/lint/test tooling via agent-state test/agent-state lint
(Jest, Vitest, Mocha, pytest, PHPUnit, cargo, go, ESLint, golangci-lint,
clippy, PHPStan, ruff, etc.) instead of assuming Laravel.
None currently — the hermetic Herdr-delegate suite and the agent-diff-reviewer
recall eval were both removed along with the code they tested. /code-review is
Anthropic's shipped skill, not something this repo evaluates.
agent-implement-laravel is Laravel-specific (Sail, Pint, PHPStan, Vite). Use
agent-implement-generic for any other stack.