agent-kit

Experimental Claude Code setup for running coding agents mostly on their own.

Flow

Small changes:

/shape <idea>          validate, spar on scope, write specs/<slug>.md
/ship <repo> <slug>    acceptance tests by a separate agent, implement, exercise, PR, independent review
/guard <repo>#<pr>     loop on CI, review comments and rebases until green
/land <repo>#<pr>      mechanical checks + fresh-context risk verdict

Multi-slice projects:

/plan-project <repo> <project>   grill, architecture, slice specs with dependencies and file ownership
/run-project <repo> <project>    parallel slice agents in worktrees, PRs into dev, integration check
                                 after each merge, one release PR dev → main

/new-project scaffolds a repo with AGENTS.md, CI (actions pinned to SHAs), pre-commit, Dependabot and branch protection.

Guardrails

• Acceptance tests are written by a different agent than the code, committed first, and protected: hooks/acceptance-guard.sh blocks edits to them, hooks/acceptance-stop.sh runs them before a session may stop, and /land diffs them against the baseline commit.
• Review is never done by the session that wrote the code.
• /land hard-blocks on auth, billing, license, customer-data and agent-config paths, foreign authors and forks.
• hooks/scan-output.sh flags hidden bidi/zero-width characters and known injection or exfiltration markers in content the agent reads.
• Agents never merge to main or post PR comments on their own.
• /run-project can give clear, no-risk slices to a free model via OpenCode. It can only edit the slice's files; Claude tests and opens the PR.

Setup on a new machine

cp config.example.json ~/.config/agent-kit/config.json   # owner and projects folder
./install.sh

Then merge hooks/settings.example.json into ~/.claude/settings.json. /plan-project and /shape also use grilling, domain-modeling, codebase-design and prototype from mattpocock/skills (npx skills add https://github.com/mattpocock/skills --skill <name>).