[cabibbz]

You're right — I never fixed it. The last change asked the reviewer politely to cite paths, and this run is proof either the ask didn't deploy or the reviewer ignored it. Prompts are hopeful; schema fields are gates. This is the gate version.

Structural change: filesReferenced is required

The output schema (both JSON and Zod) now requires a filesReferenced: string[] field on every review. The reviewer commits to which attached paths it relied on — empty is a commitment that none bore on the decision, not something inferred from what the prose happened not to name.

The count comes from that field, not from prose

fileEvidence no longer text-searches. It reads the reviewer's own list, ignores duplicates, normalises leading ./, and shows exactly what the reviewer committed to. The old heuristic could have missed a citation the reviewer wrote as overlay.cpp in a code fence, and it couldn't distinguish "partial citation" from "none".

The gate rejects fabrication

A path in filesReferenced that was not among the packet's attachments is rejected as GATE_FABRICATED_PATH before release. The reviewer cannot claim to have read files it wasn't given. Backwards-safe: older packets without a recorded attachment list are not subjected to the check retroactively.

What every layer now does

Layer	What changed
sandbox/review-schema.json	filesReferenced required
lib/types.ts (Zod)	Same, both review and opaque variants
sandbox/*-policy.md	Explicit paragraph: the field is required, the phone counts from it, a fabrication rejects the review
lib/gate.ts	Renders under ATTACHED FILES USED:, merges with attribution, rejects fabricated paths
lib/sandbox-runtime.ts	Passes attached paths to the gate; the mock references real attachments so the whole pipe is exercised in tests
lib/research.ts	fileEvidence takes the reviewer's list, not review text
components/dashboard/review-parsing.ts	Parses the new section
components/dashboard.tsx	Card fed from the structured field, with raw-JSON fallback for the Bob-Regress case

Tests

• Gate refuses an unattached path (with the leading-./ allowance verified)
• Gate skips the check when no attachment list was recorded, so old packets still release
• fileEvidence takes only paths that were actually attached, handles duplicates, normalises ./
• Every policy states the field is required and warns fabrication rejects the review
• e2e: the mock references its real attachments, the review renders ATTACHED FILES USED:, that section survives onto the answered job
• Schema field-order test updated
• Browser at 390×844: FILES READ | 2 of 2 (green), card lists the referenced files

122 tests (121 pass / 1 Windows-only skip), lint, build, e2e all clean.

Honest limits

Older reviews in your history keep their existing text — the card will read Bob-Regress as no field found. Any reviewer that returns a review missing filesReferenced will now fail schema parse and be blocked with GATE_INVALID_SCHEMA — that's stricter than before, and intentional: it's what makes this a gate rather than an ask.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
sol-review-gate	Ready	Preview	Jul 29, 2026 7:57pm