Agent Visual Evidence Manifest

Prove a PR's screenshots, GIFs, and recordings are bound to a real, branch-current, daemon-backed run — not a reused mock wearing a "LIVE" label.

Use This For

• Gating a PR's visual-evidence artifacts before merge: does every screenshot/GIF/recording carry a complete provenance manifest?
• Catching a stale or reused artifact whose manifest commit no longer matches the PR's branch HEAD.
• Catching an artifact whose manifest never says whether its data is real, fixture, or mock.
• Verifying an operator-control-panel PR's proof set actually covers the required state inventory, not just the happy path.
• Deciding whether a fixture/mock-labeled artifact is acceptable (it is, if honestly declared) versus an undeclared one (it never is).

Do Not Use This For

• How to capture a screenshot without interrupting the operator's desktop — headless Playwright, screencapture -x -l, capture harnesses (port-daddy-agent-skill's visual-evidence doctrine).
• Designing the shape of the Work Receipt an artifact ultimately attaches to (agent-work-receipt-designer).
• The broad product-level question of whether a multi-agent authoring surface clears a dogfood/Potemkin bar (multi-agent-authoring-product-bar).

The Proof-Manifest Gate

flowchart TD
  A[PR carries visual artifacts] --> B{Any artifacts attached?}
  B -->|No| Z[FAIL: no-artifacts]
  B -->|Yes| C[Check each manifest's six fields]
  C --> D{All fields present?}
  D -->|No| E[FAIL: manifest-missing-provenance-field]
  D -->|Yes| F{sourceLabel declared?}
  F -->|No| G[FAIL: undeclared-source-label]
  F -->|Yes, real/fixture/mock| H{manifest.commit == branch HEAD?}
  H -->|No| I[FAIL: commit-mismatch]
  H -->|Yes| J{isControlPanelPr?}
  J -->|Yes| K{All 7 states covered?}
  K -->|No| L[FAIL: control-panel-state-coverage-incomplete]
  K -->|Yes| M[PASS]
  J -->|No| M

1. Collect every artifact's manifest. For each screenshot, GIF, or recording attached to the PR, gather daemonPort, runId, transcriptHeadHash, agentNodeId, commit, and sourceLabel. An empty artifact set is never safe — it fails closed, not open.
2. Check field completeness per artifact. Any of the six fields missing, null, or empty-string means the artifact cannot yet count as proof; see references/manifest-field-semantics.md for what each field proves and why partial credit isn't a thing.
3. Check the source label specifically. real, fixture, and mock are all acceptable answers — an absent one is not. This gets its own check because it is the single field that separates honest illustration from disguised fakery.
4. Check every manifest's commit against the PR's branch HEAD. A manifest bound to an earlier commit is stale proof, even if every other field is filled in — the surface it depicts may no longer exist.
5. If the PR touches the operator control-panel surface, check state coverage. Union every artifact's demonstrated state against the required set (active, historical, blocked, stale, gate, interrupt, receipt); see references/control-panel-state-coverage.md for what each state must show and why partial coverage hides the bugs that matter most.
6. Score and gate. Any critical finding blocks pass regardless of score; score below 75 also blocks even with zero criticals (defense in depth for medium/high accumulation).
7. Route capture-technique and receipt-schema questions elsewhere. This gate only judges the manifest's structure and consistency — it does not (and cannot, on its own) prove a real-labeled artifact wasn't staged; that needs the adversarial canary described in sandboxed-adversarial-test-harness.

Output Contract

A passing proof-manifest audit reads a spec with:

• branchCommit: the PR branch's current HEAD commit, checked against every artifact.
• isControlPanelPr: whether full state coverage applies.
• statesCovered: the union of states this PR's artifacts demonstrate.
• artifacts[]: each { file, manifest: { daemonPort, runId, transcriptHeadHash, agentNodeId, commit, sourceLabel } }.

Use scripts/proof_manifest_audit.mjs to audit a proof-manifest-spec JSON and return { pass, score, findings, recommendations }.

Anti-Patterns

Zero Or Incomplete Provenance

Novice: Open the PR with "see attached" screenshots and no manifest at all, or a manifest that only bothers to name the file path. Expert: Every artifact carries all six provenance fields before the PR is marked ready — an empty artifact set or a half-filled manifest is treated as a hole in the proof, not a formality to backfill later. Detection: proof_manifest_audit.mjs fires no-artifacts (critical) when the artifact list is empty, and manifest-missing-provenance-field (critical) per artifact missing any of daemonPort/runId/transcriptHeadHash/agentNodeId/commit/sourceLabel.

The Undeclared Mock

Novice: Ship a polished screenshot of a stubbed UI state without saying so — it reads as "LIVE" because nothing says otherwise. Expert: Label every artifact real, fixture, or mock honestly. A mock-labeled illustration is legitimate; a real-shaped manifest with the label silently missing is not — that omission is the exact failure mode a red-team review calls "a mock or visual artifact fak[ing] the hardest part." Detection: proof_manifest_audit.mjs fires undeclared-source-label (critical) whenever an artifact's sourceLabel is missing, independent of whether its other five fields are complete.

Stale Commit Or Partial State Coverage

Novice: Reuse last week's "it's live" GIF because recapturing feels slower than shipping, or attach five polished happy-path screenshots and call a control-panel PR done. Expert: Every manifest's commit matches the PR's current branch HEAD, and a control-panel PR's artifact set is checked against the full required-state list — active, historical, blocked, stale, gate, interrupt, receipt — not just the states that happened to screenshot well. Detection: proof_manifest_audit.mjs fires commit-mismatch (critical) when manifest.commit !== branchCommit, and control-panel-state-coverage-incomplete (critical) when isControlPanelPr is true but statesCovered is missing any required state.

References

File	Load When
references/manifest-field-semantics.md	Need to know what each of the six manifest fields proves, or whether a fixture/mock label is acceptable versus an absent one.
references/control-panel-state-coverage.md	Need to decide whether a PR is control-panel-scoped, or which of the seven required states it still hasn't demonstrated.
examples/expected-output.md	Need to see a bad manifest set audited, then the same set fixed and passing.
templates/output-template.md	Need a reusable per-artifact manifest template to fill in while assembling a PR body.
schemas/proof-manifest-spec.schema.json	Need to validate a proof-manifest-spec JSON payload's structure before auditing it.
scripts/proof_manifest_audit.mjs	Need deterministic scoring of a PR's visual-evidence proof-manifest completeness and coverage.
agents/openai.yaml	Need a subagent descriptor for delegated proof-manifest auditing.

Layout QA gate (mechanical — run before shipping)

Before calling any rendered page, artifact, dashboard, deck, or component done, run the mechanical overflow/collision checker. It renders the page headlessly and flags text-vs-text collisions, clipped/ellipsis-truncated elements, text escaping its container, and horizontal page scroll — the visual defects a screenshot hides and that only appear at a specific width or in one theme.

Resolve layout-overflow-guard from the active skill catalog before running it. The command below shows the standard Claude install path; use the path reported by your harness. If the skill is absent, install or sync it instead of skipping this gate.

python3 ~/.claude/skills/layout-overflow-guard/scripts/check_layout.py <file-or-url> \
  --widths 1280,1100,860,720,390 --themes light,dark

You do not need to read check_layout.py — invoke it with the Bash tool and act on its report and exit code (non-zero = a defect). The script's source never enters your context; only its findings do. Drive it to zero violations across every width and both themes before you ship. Full detail: the layout-overflow-guard skill.

<!-- BEGIN BUNDLE INDEX (auto: index_references.py) -->

Skill Bundle Index

Every file in this skill, and when to open it. Auto-generated; run scripts/index_references.py --fix.

root

• CHANGELOG.md — Agent Visual Evidence Manifest — Changelog — - Initial skill creation - Core proof-manifest-gate process defined - Reference files and deterministic proof_manifest_audit script added
• README.md — Agent Visual Evidence Manifest — Verify that every visual-evidence artifact (screenshot/GIF/recording) attached to a PR carries a provenance manifest binding it to real daem

agents/

• agents/openai.yaml — openai (data/schema)

examples/

• examples/expected-output.md — Example Output: Agent Visual Evidence Manifest — Scenario: an agent finishes a control-panel change, attaches two screenshots to the PR, and reuses a GIF from last week's branch for the "li
• examples/sample-input.json — sample input (data/schema)

references/

• references/control-panel-state-coverage.md — Control-Panel State Coverage — Use this when you need to know which states a control-panel PR's proof-artifact set must cover, and why partial coverage (e.g.
• references/manifest-field-semantics.md — Manifest Field Semantics — Use this when you need to know what each provenance field actually proves, why all six are required, and why an honest mock/fixture labe

schemas/

• schemas/proof-manifest-spec.schema.json — proof manifest spec.schema (data/schema)

scripts/

• scripts/proof_manifest_audit.mjs

templates/

• templates/output-template.md — Proof Manifest Template — Fill in one block per artifact before opening or marking ready a PR that carries visual evidence.

<!-- END BUNDLE INDEX -->