ADR 0005: Review context loading

Status

Accepted.

Context

A review is only as good as the context behind it. Domain skills selected by label, as in ADR 0002, cover neither the programming language actually touched nor the reviewed repository's own conventions, and any added context-gathering step must never be allowed to break the claim itself.

Decision

review.claim assembles three sources of context and serves all of them in the composed review task. Language skills are auto-detected from the pull request's changed files: detectLanguages (core/languages.ts) maps file extensions against a table of 12 languages, TypeScript, JavaScript, Python, Go, Rust, Haskell, Java, Kotlin, Swift, Scala, C/C++, and Solidity, with no label required, and the matched skills/lang/<name>.md checklists land in instructions.languages. Repository context is fetched from the reviewed repo itself at the pinned SHA (gatherRepoContext, core/repo-context.ts): exact files (AGENT.md, AGENTS.md, CLAUDE.md, .claude/CLAUDE.md) plus a shallow, one-level listing of .claude/, .codex/, and .claude/skills/* for markdown files, bounded to a 10-file cap and a 64 KB total size cap, with missing paths skipped silently. All three sources are wrapped in try/catch inside claimReview; any detection or fetch failure degrades to an empty array, languages: [], repoContext: [], and never throws, so a context error can never fail a claim.

Consequences

Reviews get auto-detected language guidance and repo-specific conventions with zero configuration, and the same context reaches every host because it travels inside the composed task rather than living in per-host files. The caps are a deliberate ceiling: a repository with more relevant markdown than the caps allow gets a partial, best-effort slice, and the orchestration skill's instruction to also read the local checkout is the intended supplement for whatever the API fetch missed or capped.