agent-policy-kit

Repo-local operating policy and skills for coding agents.

This package installs one shared agent policy and skill set into a project so Codex, Amp, Pi, and Claude Code can use the same workflows.

One-liner clone + install into the current repo from the latest release:

tmp="$(mktemp -d)" && git clone --depth 1 --branch v0.2.0 https://github.com/luisalima/agent-policy-kit.git "$tmp/agent-policy-kit" && "$tmp/agent-policy-kit/scripts/install.sh" --target .

To install from main instead:

tmp="$(mktemp -d)" && git clone --depth 1 https://github.com/luisalima/agent-policy-kit.git "$tmp/agent-policy-kit" && "$tmp/agent-policy-kit/scripts/install.sh" --target .

From this checkout:

./scripts/install.sh --target /path/to/repo

Install only selected agent layouts:

./scripts/install.sh --target /path/to/repo --codex
./scripts/install.sh --target /path/to/repo --amp
./scripts/install.sh --target /path/to/repo --pi
./scripts/install.sh --target /path/to/repo --claude

Flags can be combined:

./scripts/install.sh --target /path/to/repo --pi --claude

Re-run the installer to update already installed skill folders:

./scripts/install.sh --target /path/to/repo

Default install:

AGENTS.md
CLAUDE.md
.agents/skills/<skill>/SKILL.md
.agents/skills/<skill>/USER.md (optional local extension)
.claude/skills/<skill>/SKILL.md
.claude/skills/<skill>/USER.md (optional local extension)

Agent mapping:

• Codex: AGENTS.md + .agents/skills/
• Amp: AGENTS.md + .agents/skills/
• Pi: AGENTS.md + .agents/skills/
• Claude Code: CLAUDE.md + .claude/skills/

The installer does not overwrite existing AGENTS.md or CLAUDE.md files. It appends or updates only the managed agent-policy-kit block when needed. Installed skill folders are refreshed on every run so targets receive skill updates from the package. Bundled SKILL.md files are package-managed and may be overwritten during refresh. Put local skill instructions in USER.md beside the installed SKILL.md; the installer preserves that file on normal installs.

• opentasks: upstream skill from https://github.com/luisalima/opentasks-skill for managing docs/tasks/ task and question tracking.
• linters: conservative workflow for adding or improving project-appropriate lint commands.
• security-review: focused security review after substantial security-impacting implementation changes.
• threat-model: threat modeling before security-relevant architecture work.
• tdd: test-first implementation workflow.
• done: definition-of-done checklist.
• adr: architectural decision record workflow.
• spike: declared exploratory-work workflow.
• slice: vertical-slice planning workflow.

In the target repo, run the opentasks bootstrap command if you want durable task/question tracking:

/opentasks bootstrap

For Pi, invoke the same skill as:

/skill:opentasks bootstrap

Use the security skills at different points in the work:

/threat-model <scope>
/security-review <branch or diff>

Threat modeling is intended before architecture work that changes trust boundaries, auth, privileged automation, deployment topology, sensitive data flows, or third-party integrations. Security review is intended after substantial implementation work with plausible security impact.

See CONTRIBUTING.md. CI runs installer smoke tests, git diff --check, skill metadata validation, release-version validation, and ShellCheck.

MIT. See LICENSE.