Repository navigation
v3.3.0
3.3.0 is about precision — the reviewer stops letting loose words draw big conclusions. A comment command can now force a re-review without label permissions, but only for people verified to have them; Addresses #N is an implementation reference, not a merge promise; a requirement is in scope because the change actually touches it, not because "src" appears in both sentences; and a partial review now says which kind of gap it hit instead of one generic shrug.
Plus the v2 parity oracle — the scaffolding that proved the v2→v3 port faithful — is retired now that its job is done.
Force a re-review with a comment
/ai-reviewcommand (#952, closes #914): a newrereview-commandinput (default/ai-review, empty disables) makes a PR comment whose body is exactly that command force a fresh full review — no label UI permissions needed. The trigger is deliberately narrow:- Bare command only. Trailing non-whitespace text (e.g.
/ai-review smart) never triggers, so the argument namespace stays reserved. - Permission is verified, not assumed. The label trigger is self-authorizing (only triage+ can label); the comment path resolves the commenter's permission through the forge API and requires triage-or-higher, failing closed on lookup errors, unrecognized roles, or malformed responses. Forgejo collaborators map through the same rule.
- Races lose nothing. A comment run is pinned to the head it checked out; if a push raced it, the comment run skips and the
synchronizerun covers the new head. - Fork PRs, closed PRs, and drafts no-op (forks get a pointer to
ai-review-fork), and a forge outage during the pre-checkout gate drops the trigger rather than turning every comment into a red job.
- Bare command only. Trailing non-whitespace text (e.g.
Read the PR's words the way GitHub does
Addresses #Nis an implementation reference (#956, closes #953): it's now recognized alongsideImplements/Part of/Refsas non-closing, so the review reads the referenced issue's acceptance criteria instead of ignoring it — and past-tense closing forms (closed/fixed/resolved) only count at line start, so prose like "covered by closed #479" mid-sentence stops hijacking scope.- Requirement-trace scope needs real evidence (#958, closes #957):
subjectTouched()used to treat any two single-word lexical hits as scope overlap, which dragged unrelatedAGENTS.mdstandards (untrusted-content, fork-privilege, credential transport) into a PR's requirement trace via words likedocs,request, ornever. Scope now requires a multi-word phrase match or an identifier-shaped term (sourceSha,SQLite) — generic words and bare acronyms (PR,API) no longer establish it. Linked-issue requirements and entries without provenance stay in scope, unchanged (#935 contract).
Partial reviews say why
- Trace gap ≠ execution gap (#955, closes #954): a review that finishes its investigation but can't verify one requirement's trace used to publish the same generic "did not resolve every required check" prose as one that stopped reading early. A new
incomplete-reasonoutput (andincomplete_reasonmarker key) distinguishesexecutionfromrequirement_trace(andboth/none), and the published prose matches the actual gap. Complete runs omit the key, so marker bytes don't change.
Test scaffolding
- The v2 parity oracle is retired (#948): its fixtures live on as v3 snapshot tests. v3 is now its own reference, not a shadow of v2.
Upgrading
Pin to v3.3.0, or the floating v3 tag. Nothing breaks; the only new input is rereview-command, and it only acts if your workflow listens for issue_comment created events on pull requests.
Full changelog: v3.2.0...v3.3.0
Generated changelog (release-please)
Features
Bug Fixes
- precheck: recognize Addresses and scope past-tense closing keywords (#953) (#956) (5a659f7)
- requirement-trace: require strong subject evidence for scope (#957) (#958) (bcae3d3)
- test: list the dogfood context-window inputs as overrides (#951) (f32389b)
- v3: split trace gap from execution gap (#954) (#955) (f891b14)