Skip to main content

Contributing a New Coding Agent Host

Host support is an evidence-backed set of claims, not one manifest or one transcript parser. A contribution can support only the slices the host exposes, as long as partial and unavailable coverage stays explicit.

The complete, canonical workflow lives in the repository:

Define the support boundary first

SliceRequired decision
Native contractHost/version and primary source verified
ShellNative, source-local, generated, or none
Configured assetsAvailable, partial, or unavailable scopes
Session evidenceAvailable, partial, or unavailable fields and events
Shared registrationOnly capabilities that are actually implemented
OutputExisting Canvas, HTML, Markdown, or a justified new mode
PackagingPublic npm, runtime bundle, source-only, or none

A shell does not establish session support, and a parser does not establish native Skill discovery. Record each claim in a dated spec with stable acceptance ids and an evidence route.

Contribution workflow

  1. Read the host's versioned, primary contract for manifests, paths, configuration precedence, workspace identity, events, and privacy boundaries.
  2. Keep any host shell thin; canonical judgment remains in shared Skills, models, references, templates, and capability-owned scripts.
  3. Add configured-asset and session adapters independently. Use sanitized fixtures, reject foreign-workspace evidence, and represent missing fields as unobserved rather than zero.
  4. Propagate the host id only through registries whose capabilities exist. Prefer an explicit unsupported error to fallback through another host.
  5. Verify focused tests, the full suite, native-host smoke paths, package boundaries, and Windows/macOS/Linux behavior proportional to the change.
  6. Update the adapter matrix and submit Story/Spec/Test/Risk evidence through the repository pull request template.

Worked examples

  • PR #6 — Qwen Code demonstrates why native source checks, environment precedence, privacy, and case-insensitive filesystem coverage must supplement synthetic tests.
  • PR #22 — GitHub Copilot demonstrates spec-led separation of shell, assets, sessions, registries, documentation, and evidence-honesty boundaries.

Pull requests can change during review. Use them as worked examples, then defer to the current repository instructions and canonical guide.