Thanks for helping improve the delego Claude Code plugin. Contributions to the skills and agents are welcome — more policy templates, broker reviewers, and operating skills make the toolkit more useful.
Important: skills and agents are instructions that run in other people's Claude Code sessions. A contribution is a prompt, not just documentation — review it as you would code that executes on someone else's machine.
- Fork the repo and open a PR from a branch in your fork.
- Keep
python scripts/validate.pygreen (CI runs it): the manifests parse, and every skill/agent has valid frontmatter, a kebab-casename, a realdescription, and unique agent names. - Fill in the PR template, including the AI-assistance disclosure.
- Correct delego semantics. Fail-closed; the evaluation order (forbidden → rules → default deny); the invariants (no LLM in the decision path, no credential custody, fingerprint + intent binding, single-use approvals, an append-only signed audit). Don't describe behaviour delego doesn't have.
- Honest caveats. Don't oversell — e.g. the audit chain does not detect truncation of the most recent receipts, and the policy can't see the URL query string (until protocol 0.3).
- Safe instructions. No destructive or credential-exfiltrating guidance, and nothing that bypasses delego.
- A precise
description. It drives auto-delegation — specific enough to fire on the right requests, not so broad it hijacks unrelated ones.
Changes to the plugin / marketplace manifests or the bundled MCP server config are maintainer-reviewed — they affect every install. Open an issue first.
AI assistants are welcome tools, but AI-generated skills/agents get stricter review (they're prompts that run for others). Disclose AI use in the PR template, expect closer scrutiny, and own every line — "the model wrote it" is not a defence. Fork + PR + green CI; no fast path.
Apache-2.0. By contributing you agree your contribution is licensed under it.