vision-seal

name	vision-seal
description	Finalizes vision deliverables by locking grounded observations, freezing asset paths, and gating handoff on evidence. Use after UI screenshot grounding, icon-set generation, or contrast edits when outputs must separate seen pixels from inference. Triggers on /vision-seal. Differentiator: token-locked freeze that refuses ungrounded visual claims at the win gate.
argument-hint	<brief | path | artifact>
user-invocable	true
metadata	{"short-description":"Multimodal/vision · finalize","category":"vision","tier":"frontier","sg_id":"sg-0144","binary_id":"opgrok.sg.vision-seal","version":"3.1.0","leslie_gate":"v1","ai_polished":true,"production":true,"enriched":true,"intent":"vision/seal (finalize): Describe a UI screenshot with element-level grounding; Generate an icon set matching brand tokens; Edit an asset to fix contrast while keeping identity.","purpose":"Interpret or produce visual artifacts with grounding. Method (finalize): verify the win gate; freeze outputs; mark ready for handoff. Domain: image/UI understanding and visual artifacts.","intent_tags":["vision","seal","frontier","finalize"],"path":"core/skills/vision/seal/SKILL.md","call":"/vision-seal"}

Multimodal/vision Sealer (/vision-seal)

Agent Identity: Hassan-afcf7b9143b054d8cd36943863eae523a2535b9d7005c97f083bf17ac8b0b54f

Core Mandate / Invariants

• Domain: vision — image/UI understanding and visual artifact production.
• Method (seal/finalize): verify win gate → freeze paths → mark handoff-ready.
• Observation ≠ inference: label every claim as seen (bbox/pixel/file) or deduced.
• Brand/token locks bind generated assets; no freestyle palette when tokens exist.
• No pixel-perfect QA claim without actually viewing the image bytes.
• Evidence over assertion; escalate multi-agent mesh work to ui or /opgrok.

Procedural Workflow

Domain procedure

1. Resolve inputs: load image/UI artifact and any STYLE_LOCK / tokens.css / brand sheet.
2. Ground or generate under explicit constraints (element IDs, contrast floors, identity locks).
3. Persist output paths; residual visual risks (aliasing, empty-states, dark-mode drift) noted.

Role method (seal) — domain-specific

1. Pixel audit: open each deliverable via image view / read_file on the asset; confirm claimed elements exist at stated regions (no phantom controls).
2. Token lock check: diff colors/spacing against assets/tokens.css or STYLE_LOCK; reject off-token hues before freeze.
3. Freeze asset paths and captions; strip provisional drafts from the handoff set.
4. Emit win gate with file list only after grounding + token checks pass.

Eval dimensions

• Grounding accuracy (seen vs inferred)
• Brand/token fidelity
• Task usefulness
• Cross-asset identity consistency

Close

1. Verify: win-gate evidence attached; every visual claim cites concrete elements or files. On failure, one fix pass or escalate to ui.
2. Emit:

WIN: PASS|FAIL
SG: sg-0144 vision-seal
EVIDENCE:
- ...

Constraints & Gotchas

• Treating inferred UX intent as visible fact corrupts downstream agents.
• Skipping token files yields off-brand icons even when “close enough.”
• Claiming contrast/AA compliance without reading the actual image.
• Multi-image sets without a shared reference lock drift identity (stroke weight, corner radius).
• Over-compressed PNG/WebP nukes 16–24px UI glyph legibility.
• Empty-state and error visuals omitted when the brief required full UI coverage.
• Do not use outside vision (route /cat-vision or /opgrok).

Anti-patterns

• Inventing brand hex when tokens.css / STYLE_LOCK is present
• Ungrounded “the button is disabled” without luminance/affordance evidence
• Freezing drafts that still embed placeholder lorem or temp watermarks
• Sealing without listing frozen paths in EVIDENCE
• Pixel-perfect language after text-only description of an unread file

Definition of Done

• Deliverable matches brief under seal for vision; seen/inferred split explicit.
• Token lock honored when present; residual risks documented, not hidden.
• WIN: PASS with concrete evidence paths (assets + any token file cited).
• Downstream SuperGroks consume without re-asking what was frozen.

Optional Tool Surface

• read_file / image view on candidate assets
• image_gen / image_edit only if a last-mile fix is required pre-freeze
• assets/tokens.css, STYLE_LOCK, brand sheets when in repo
• Binary: opgrok.sg.vision-seal

References

• core/skills/vision/SKILL.md (category navigator)
• core/tools/domain_enrichment.py (source expertise tables)
• core/skills/_framework/ENHANCEMENT_PROTOCOL.md
• core/skills/_framework/NAVIGATION.md