feat(evals): evidence roll S1 — photo-roll collection view + sticky PR publisher + skill

[benjaminshafii]

What

Slice S1 of the Evidence Roll program (spec: prds/evidence-roll/evidence-roll-program.md, included). Reuses the existing photoRoll proof primitives (#3322/#3359) and adds the two missing adapters — nothing new is captured, only rendered:

• @openwork/evidence (zero new deps): tolerant scanRolls reader over evals/results/rolls/*/roll.json (legacy flow-lane dirs readable read-only), renderCollectionHtml (roll-of-rolls: cards with pass/fail badges, thumbnails, links to each roll's own index.html), renderPrMarkdown (claim → screenshot gallery with per-expectation PASS/FAIL), publishPr (Vercel Blob upload — env token → Infisical fallback, lifted from the deprecated runner/reporters/pr.ts, not imported — and a single sticky <!-- photo-roll --> comment, edited in place).
• CLI: pnpm --dir evals run roll [-- --open] · pnpm --dir evals run publish:pr -- --pr <n> [--roll <dir|name>] [--dry-run].
• Skill .opencode/skills/pr-photo-roll: judgment only (pick the roll, only roll.json-attributed images), mechanics in the CLI.
• fraimz: photoRoll additionally records best-effort gitSha/branch (additive; readers tolerate absence).
• Reserved (schema/docs only, S2): serialized surfaces + handoff entries for held-Daytona human testing; S3: snapshot-per-frame resumable rolls.

Flow lane untouched (evals/runner/** unchanged).

Tests (commands + results)

pnpm --dir evals run test        # 108 pass / 0 fail (incl. real-writer compatibility test)
pnpm --dir evals run typecheck   # clean

Real end-to-end: headless hosts.chrome() rendered the generated collection page → screenshot() → validate() (OpenAI vision, both expectations PASS) → photoRoll → collection rebuilt → the sticky gallery comment below was posted by the new publisher itself (Blob-hosted images).

Known env finding (out of scope): hosts.desktop() local spawn on macOS brought up CDP but no page target in 120s (twice); app specs' home remains the Daytona nightly.

Proof format

Spec acceptance suites + the self-published gallery comment below, per program-owner decision (no fraimz flow lane).

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
openwork-app	Ready	Preview	Aug 1, 2026 8:15pm
openwork-den	Ready	Preview	Aug 1, 2026 8:15pm
openwork-den-worker-proxy	Ready	Preview	Aug 1, 2026 8:15pm
openwork-diagnostics	Ready	Preview	Aug 1, 2026 8:15pm
openwork-landing	Ready	Preview, v0	Aug 1, 2026 8:15pm

[benjaminshafii]

Photo roll — evidence-roll-collection-proof

✅ 1/1 frames passed · 0 failed · 0 unvalidated · 2 expectations passed · 0 failed

1. A page titled as a photo roll collection lists at least one roll card

• ✅ PASS A page titled as a photo roll collection lists at least one roll card — The page is titled “Photo rolls” and displays one bordered roll card for “evidence roll smoke.”
• ✅ PASS A roll card shows a pass/fail style summary badge — The roll card displays the gold status badge “UNVALIDATED.”

---

Roll created 2026-08-01T19:35:16.906Z · Source: evals/results/rolls/2026-08-01T19-35-16-906Z-evidence-roll-collection-proof/roll.json · Repro: pnpm --dir evals run publish:pr -- --pr 3417 --roll 2026-08-01T19-35-16-906Z-evidence-roll-collection-proof

[diff-warden]

Warden security clearance: clear. No new security issues found in this diff (4b09c908500d9a0ebfd7baf360a42b677e54f88d). Automated clearance satisfies the required-review gate only — a human still reviews and merges. Analysis run