Open Work Agent is meant to become a high-quality public open-source template. Prefer current best practices, boring maintainable tooling, clear docs, strong tests, and small reviewed slices over speed. Build each part as if future contributors will learn from it, extend it, and trust it.
Work in digestible chunks.
Each chunk should be small enough to reason about before code is written. The default loop is:
- Pick the next ordered GitHub issue.
- Restate what problem this task solves.
- Break it into the smallest useful implementation slice.
- Propose a simple system design.
- Explain the design in plain language.
- Wait for explicit acceptance.
- Implement only the accepted slice.
- Verify it and update the relevant docs or backlog state.
Avoid big autonomous loops, broad rewrites, and large generated code drops. The goal is a thoughtful kit built one understandable piece at a time.
Every implementation task should start with the smallest meaningful failing test or test update.
Unit tests protect harness contracts, policy decisions, and pure runtime behavior. E2E tests protect important app-visible workflows. Tests should prove behavior a maintainer would care about, not only exercise files for coverage.
Use pnpm check as the full local quality gate before committing.
Use conventional commits. Commit completed work locally, but do not push unless explicitly asked.
Adopt SemVer through Changesets after the first coherent public scope is complete. Until then, keep release-specific automation out of the critical path.