Review only. Accept one target plus selected IDs/URLs, unresolved threads, comments since a ref/date, one author, or all comments, with an optional verified receipt.
- Verify target revision/head and comment ID/updated-time fingerprints. Reuse unchanged selected comment bodies and code evidence; retrieve new/edited comments, necessary surrounding chronology, changed target/diff/checks, and relationship-bearing context only when the claim depends on it. For selected PR review threads, preserve/paginate thread ID, replies, resolution, outdated/path/line state; prefer configured connector/API and use GitHub GraphQL when flat comments omit thread state. Never infer unresolved from flat comments alone. If metadata cannot establish freshness, retrieve the affected material fully.
- Use one
explorer by default. Fetch metadata first. Soft-cap each injected fragment near 1,000 tokens; split/expand intentionally and never exceed 10,000 unless no safe alternative exists. Inject exact snippets rather than complete histories/logs unless conflicts require expansion. Preserve author, URL/ID, location, state, reviewed revision/head, and cutoff.
- Cache evidence, not conclusions. Resolve each claim against current code, decisions, later comments, linked authoritative sources, and docs. Give exact evidence plus fingerprints—not prior conclusions—to a fresh high-reasoning read-only
default reviewer.
- Classify valid defects/questions, deeper implications, duplicates, conflicts, misunderstandings, suggestions, obsolete requests, and satisfied comments. Deduplicate centrally while mapping every selected comment. Current requirements and code outrank stale comment context.
Render fixer-ready P# defects and N# | sources | satisfied|obsolete|duplicate:P#|unsupported|suggestion|superseded|conflict:Q# | evidence dispositions. Put only unresolved decisions under Please answer: as direct Q# questions; no NEXT. Re-evaluate affected comments after answers.
Output findings/questions or NO COMMENTS, plus a receipt containing target revision/head, selected comment fingerprints/cutoff, linked evidence fingerprints, inspected locations, and result IDs. After emitting the receipt, automatically run $github-review-learnings with that receipt; it may record only settled, evidence-bounded learnings and must not change the review result.
1---2name: evaluate-github-comments3description: Determine which selected issue or PR comments still require action, without changing anything.4---56Review only. Accept one target plus selected IDs/URLs, unresolved threads, comments since a ref/date, one author, or all comments, with an optional verified receipt.781. Verify target revision/head and comment ID/updated-time fingerprints. Reuse unchanged selected comment bodies and code evidence; retrieve new/edited comments, necessary surrounding chronology, changed target/diff/checks, and relationship-bearing context only when the claim depends on it. For selected PR review threads, preserve/paginate thread ID, replies, resolution, outdated/path/line state; prefer configured connector/API and use GitHub GraphQL when flat comments omit thread state. Never infer unresolved from flat comments alone. If metadata cannot establish freshness, retrieve the affected material fully.92. Use one `explorer` by default. Fetch metadata first. Soft-cap each injected fragment near 1,000 tokens; split/expand intentionally and never exceed 10,000 unless no safe alternative exists. Inject exact snippets rather than complete histories/logs unless conflicts require expansion. Preserve author, URL/ID, location, state, reviewed revision/head, and cutoff.103. Cache evidence, not conclusions. Resolve each claim against current code, decisions, later comments, linked authoritative sources, and docs. Give exact evidence plus fingerprints—not prior conclusions—to a fresh high-reasoning read-only `default` reviewer.114. Classify valid defects/questions, deeper implications, duplicates, conflicts, misunderstandings, suggestions, obsolete requests, and satisfied comments. Deduplicate centrally while mapping every selected comment. Current requirements and code outrank stale comment context.1213Render fixer-ready `P#` defects and `N# | sources | satisfied|obsolete|duplicate:P#|unsupported|suggestion|superseded|conflict:Q# | evidence` dispositions. Put only unresolved decisions under `Please answer:` as direct `Q#` questions; no `NEXT`. Re-evaluate affected comments after answers.1415Output findings/questions or `NO COMMENTS`, plus a receipt containing target revision/head, selected comment fingerprints/cutoff, linked evidence fingerprints, inspected locations, and result IDs. After emitting the receipt, automatically run `$github-review-learnings` with that receipt; it may record only settled, evidence-bounded learnings and must not change the review result.
Tristan-Hauber