A Claude Code skill for contributing to open source projects with mandatory human approval gates and self-improvement.
/tackle <issue> # Start working on an issue
/tackle --help # Show all commands
Required:
- Research - Fetches upstream CONTRIBUTING.md, open issues, and PRs. Caches for fast reuse.
- Duplicate Detection - Searches for existing PRs that address your issue before you start coding.
- Plan - Creates an implementation plan based on upstream conventions.
- 🚧 GATE: Plan Review - Stops and waits for your approval before coding.
- Implement - Writes code following detected project style.
- Validate - Runs tests, checks that changes are isolated to a single concern.
- 🚧 GATE: Pre-Submit Review - Creates a draft PR for your review. Stops and waits for approval.
- Submit - Marks the draft PR ready for maintainer review.
- Reflect - Reviews the run for friction points and proposes skill improvements.
- Mandatory approval gates - No code written or PR submitted without explicit human sign-off
- Upstream-aware - Checks for existing fixes before duplicating work
- Convention-following - Parses CONTRIBUTING.md and applies commit style, test requirements, PR format
- Self-improving - Logs issues and detects patterns across runs; proposes fixes after 2+ occurrences
- Session-resilient - State persists via beads molecules; resume after compaction or new session
After researching upstream issues and PRs, the plan is presented for approval:
Ask for rationale and the agent explains tradeoffs:
After implementation and validation, a draft PR is created for review:
The draft PR on GitHub:
After approval, CI is verified before marking the PR ready for review:
Built on ideas from: