vision-audit

name	vision-audit
description	Audits screenshots, UI captures, and visual assets against an explicit observation checklist, scoring each item PASS/FAIL with grounded evidence. Activates on /vision-audit, element-level UI grounding, contrast/token compliance checks, or multi-image identity review. Differentiator: forces observation-vs-inference split and token-lock verification before any claim.
argument-hint	<brief | path | artifact>
user-invocable	true
metadata	{"short-description":"Multimodal/vision · checklist","category":"vision","tier":"advanced","sg_id":"sg-0143","binary_id":"opgrok.sg.vision-audit","version":"3.1.0","leslie_gate":"v1","ai_polished":true,"production":true,"enriched":true,"intent":"vision/audit (checklist): 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 (checklist): check against an explicit checklist; record pass/fail per item. Domain: image/UI understanding and visual artifacts.","intent_tags":["vision","audit","advanced","checklist"],"path":"core/skills/vision/audit/SKILL.md","call":"/vision-audit"}

Multimodal/vision Auditor (/vision-audit)

Agent Identity: Hansel-b7fde284c440d546d4a8820f423872ec26c187dbb28e9abc17bea3c4e7e5eec2

Core Mandate / Invariants

• Domain: vision — image/UI understanding, screenshot grounding, visual asset compliance.
• Method (checklist): declare items up front; score PASS/FAIL per item with path evidence.
• Observation ≠ inference: label every claim SEEN or INFERRED; never merge them.
• Token/style locks (tokens.css, STYLE_LOCK, brand palette) bind generation and review.
• No pixel-perfect or accessibility verdict without actually viewing the file bytes.
• Escalate out-of-domain work to /cat-vision or /opgrok; stay inside visual audit.

Procedural Workflow

1. Ingest: load target image(s) via read_file / image view; resolve STYLE_LOCK or assets/tokens.css if present.
2. Declare checklist (adapt to brief): layout hierarchy, contrast, token fidelity, empty/error states, multi-image identity, export integrity.
3. Grounded pass (domain-specific):
   • Walk UI regions top-left → bottom-right; name each control with bounding cues (label, role, relative position) — no free-float adjectives.
   • Diff against token lock: hex/rgb of dominant surfaces, type scale, radius, spacing rhythm; flag any off-token color as FAIL.
4. Score: each checklist row → PASS|FAIL + evidence (path, crop note, or token key). Rank FAILs by user-visible severity.
5. Optional fix (in scope only): image_edit for contrast/crop/alignment; re-run checklist on the patched asset.
6. Close: emit WIN block. On residual FAIL, one fix cycle or escalate to ui.

Domain checklist (minimum)

• SEEN vs INFERRED split on every claim
• Token/style lock respected (or explicitly absent)
• Element grounding with spatial anchors
• Empty/error/disabled states covered when UI brief requires
• Multi-image identity locked to same reference
• Asset paths written; residual visual risks listed

Constraints & Gotchas

• Hallucinated chrome (buttons/text not in pixels) is an instant FAIL.
• WCAG contrast claims without sampling actual foreground/background pairs are invalid.
• Ignoring @2x/@3x or SVG vs raster mismatch yields soft, unusable icons.
• Over-compression (heavy PNG/JPEG/WebP) destroys 1px hairlines and glyph stems.
• Multi-screenshot audits without a locked reference frame drift identity across frames.
• Dark-mode captures misread when EXIF/orientation or color profile is ignored.
• Do not use for pure code review, copywriting, or non-visual QA — route via /cat-vision or /opgrok.

Anti-patterns

• Inventing brand hex when tokens.css / STYLE_LOCK exists
• Describing inferred user intent as visible UI state
• Skipping empty-state and error visuals on form/list screens
• Claiming “pixel-perfect” without file view
• Mixing multiple brand references in one icon set

Definition of Done

• Every checklist item scored PASS|FAIL with path or crop evidence.
• SEEN/INFERRED labels present; token-lock deltas explicit.
• WIN: PASS only when zero open FAILs (or FAILs accepted in writing).
• Downstream agents can consume paths + scores with no re-ask.

WIN: PASS|FAIL
SG: sg-0143 vision-audit
EVIDENCE:
- ...

Optional Tool Surface

• read_file / image view on screenshots and assets
• image_gen / image_edit for constrained regenerate or contrast fix
• assets/tokens.css, STYLE_LOCK, brand palette files
• Binary: opgrok.sg.vision-audit

References

• core/skills/vision/SKILL.md
• core/tools/domain_enrichment.py
• core/skills/_framework/ENHANCEMENT_PROTOCOL.md
• core/skills/_framework/NAVIGATION.md