Review Lens

Review Lens

Coordinate, do not perform specialist passes. Public API dominates library review: prioritize what consumers can construct, implement, match, store and depend on across releases. Risk changes attention within each pass, never the required roster.

Procedure

1. Establish facts once using shared context; reuse matching caller-supplied facts. Inventory affected packages/configurations and scan changed public surface first. Read package comparison to establish comparison scope before API/docs extraction.
2. Dispatch all ten required specialists, even for small, docs-only, naming-only or manifest-only changes. Read worker isolation before dispatch; give each fresh worker its permitted factual handoff and the coverage-record contract below. No inline or combined specialist passes.
3. Merge by root cause/fix, including existing discussion. Keep the strongest supported evidence; resolve contradictions with the owners and decisive evidence, not repeated whole passes. Read the findings contract when merging; preserve it in intermediate and final output. Consolidate public-surface coverage and limitations, then enforce the publication gate below. Apply the contract's automatic clean/nit-only approval rule only to a complete merged result, not just newly posted comments; retain delivery's mode, ownership and finding-refresh restrictions.
4. Refresh target/head immediately before delivery. A review stays pinned to its snapshot unless a complete review permits the descendant refresh below. Head movement after incomplete coverage requires a fresh review; other movement requires a fresh review or blocked result, not stale publication.
5. Deliver once: after the roster passes either the completion or publication gate, dispatch one fresh review-delivery worker with merged findings, coverage manifest and authorized mode. Supply the combined verdict for a complete review; for an incomplete review supply the internal blocked status and blocked-area diagnostics, never a public verdict. Local/report-only work stays in chat. Finish with shared-context cleanup.

Best-effort finding refresh

Only after the complete roster passes its completion gate may the coordinator inspect reviewedHead..currentHead and current source to re-evaluate existing merged findings. Never discover new findings or claim full coverage of new commits. Require unchanged target/base, an ancestor reviewed head, and the complete exact delta/current source. Retargeting, rewritten/non-descendant history or incomplete evidence requires a fresh review or blocked result.

Classify every merged finding:

Record findingRefresh: reviewed/current heads, target/base, every classification and inspected delta reference. Keep the original manifest pinned. Force COMMENT-only/no ADO vote regardless of original verdict. State that full Review Lens coverage ended at the reviewed head; only existing findings were best-effort re-evaluated through current head. Blocked refresh evidence cannot authorize publication.

Required coverage

Every row is mandatory on every invocation, including output-only API and docs retrieval. Each owns its area, not another full review; assign cross-area root causes one owner. Direct focused requests retain only their requested workflow.

review-public-api stays output-only/report-only: supply package/configuration, pinned revisions, execution permission, matching artifacts and factual packageComparison, never source, manifests, source diffs, docs text or other reviewers' findings. Its isolated docs-filtering stage is mandatory even for an empty applicable report; return only the filtered area result for merging.

The isolated docs worker supplies data, never findings/verdict. Route its bundle to API-design/consistency or the isolated API filter, never to the output-only API worker as candidate evidence. Reuse matching captures/bundles, not workers; reuse saves builds, not required passes.

Coverage manifest, completion and publication gates

Keep one returned coverageManifest record per required skill: skill, actual workerId, exact pinned snapshot, status, concise evidence/artifact reference. Never invent IDs, substitute coordinator passes or invent findings. No unrelated probes to fill rows.

• completed: worker finished its scoped procedure, including required extraction/comparison/filtering, and returned findings/data or explicit no-findings. Supported one-sided package comparisons can complete under the package-comparison contract.
• not-applicable: dispatched worker established no applicable surface from stated, permitted evidence. Confirmed absence of Rust library packages can qualify API/docs; small/docs-only Rust changes alone cannot.
• blocked: required evidence, permission, tools, isolation or dependency is unavailable; failed workers and missing output also block, never succeed.

Set reviewComplete=true only when all ten records match one reviewed snapshot and are completed or evidence-backed not-applicable.

Set reviewPublishable=true when all ten specialists were dispatched, all ten records match one reviewed snapshot, and at least one record is completed. blocked records do not prevent publication of results from completed areas. Missing records, skipped specialists, snapshot mismatch, or a run with no completed area remains non-publishable. Treat snapshot currency separately: revalidate it immediately before delivery and never persist it as part of the reviewPublishable claim.

When reviewPublishable=true and reviewComplete=false, deliver the review as COMMENT/no ADO vote regardless of the findings' severity. Lead the public summary with a prominent warning that coverage is incomplete, list every blocked area with its concise diagnostic, and state that the verdict is withheld. Do not describe unassessed areas as clean or claim complete current-head coverage. Complete reviews and valid finding refreshes retain their existing delivery rules.