agent-visual-evidence-manifest

name	agent-visual-evidence-manifest
description	Verify every visual-evidence artifact (screenshot/GIF/recording) attached to a PR carries a provenance manifest binding it to real daemon-backed truth — daemon port, run id, transcript head hash, agent node id, commit, and an honest real/fixture/mock source label — and that operator-control-panel PRs cover the required set of proof states. Use when gating a PR's proof artifacts before merge, auditing whether a "LIVE" visual claim is backed by a real daemon run versus a reused or fixture-backed mock, or defining the required state-coverage set for a control-panel change. NOT for the capture technique itself — headless Playwright, `screencapture -x -l`, non-interruptive capture (use port-daddy-agent-skill's visual-evidence doctrine), designing the receipt body schema an artifact attaches to (use agent-work-receipt-designer), or the broader dogfood/Potemkin product-quality bar (use multi-agent-authoring-product-bar).
license	Apache-2.0
allowed-tools	Read,Write,Edit,Bash,Grep,Glob
metadata	{"category":"Agent & Orchestration","tags":["proof-manifest","visual-evidence","provenance","control-panel","fail-closed"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"pairs-with":[{"skill":"agent-work-receipt-designer","reason":"A sealed receipt attaches artifacts as evidence; this skill is what proves an attached artifact is genuinely daemon-backed before the receipt cites it."},{"skill":"sandboxed-adversarial-test-harness","reason":"Shares the fail-closed, prove-it-positively framing — an adversarial canary is the only way to catch a \"real\"-labeled artifact that is actually a disguised mock, which this manifest audit cannot detect on its own."},{"skill":"multi-agent-authoring-product-bar","reason":"This skill's manifest gate is one narrow, provable input into that skill's broader audit of whether a control-panel feature is real or Potemkin."}],"io-contract":{"kind":"deliverable","consumes":["[Truncated]","[Truncated]"],"produces":["[Truncated]","[Truncated]"]}}

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.