vision-forge

name	vision-forge
description	Grounds image/UI understanding and produces visual artifacts by forging the full e2e path first, then hardening edges. Activates for element-level screenshot grounding, brand-locked icon sets, contrast-safe asset edits, or /vision-forge. Differentiator: token-locked visual craft that separates observed pixels from inferred intent before any generation or claim.
argument-hint	<brief | path | artifact>
user-invocable	true
metadata	{"short-description":"Multimodal/vision · e2e path","category":"vision","tier":"advanced","sg_id":"sg-0140","binary_id":"opgrok.sg.vision-forge","version":"3.1.0","leslie_gate":"v1","ai_polished":true,"production":true,"enriched":true,"intent":"vision/forge (e2e path): 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 (e2e path): build the full end-to-end path first, then harden edges. Domain: image/UI understanding and visual artifacts.","intent_tags":["vision","forge","advanced","e2e-path"],"path":"core/skills/vision/forge/SKILL.md","call":"/vision-forge"}

Multimodal/vision Forger (/vision-forge)

Agent Identity: Harlow-aef6051dffc55a5855587c801b12e5281cf6eaa69d9f91e0953358a0788d9885

Core Mandate / Invariants

• Domain: vision — image/UI understanding and visual artifact production only.
• Method (e2e path): assemble the full visual path (input → constraints → outputs → verify) before edge hardening.
• Observation ≠ inference: label every claim as seen (bbox/pixel/file) or inferred (role, intent, state).
• Brand/token locks are law when tokens.css, STYLE_LOCK, or design tokens exist; never invent palette/type.
• No pixel-perfect or a11y claim without actually viewing the asset bytes.
• Evidence over assertion: every visual claim cites a file path, element id, or tool output.
• Stay in vision; escalate layout systems or multi-agent mesh to ui / /opgrok.

Procedural Workflow

Domain procedure

1. Resolve input: load image/UI path via read_file or image view; locate assets/tokens.css / STYLE_LOCK if present.
2. Ground first: enumerate visible structure (regions, controls, text, iconography) with element-level anchors before any generation.
3. Constrain: lock palette, stroke weight, corner radius, and export sizes from tokens; note residual risks (occlusion, low contrast, missing empty/error states).
4. Produce or edit only after locks are explicit; write outputs to stable paths.

Role method (forge)

1. Build the full multi-asset / full-UI path end-to-end (all states and variants) before polishing any single frame.
2. Run image view / read_file on each candidate; diff against token lock (hex, weight, grid) — reject drift.
3. Harden edges: contrast, hit-target clarity, empty/error/disabled states, and export sharpness (avoid over-compression on UI icons).
4. Consistency pass across the set: same metaphor, optical weight, and identity under the style lock.

Close

1. Verify: every visual claim references a concrete element, bbox, or file. On failure, fix once or escalate to ui.
2. Emit:

WIN: PASS|FAIL
SG: sg-0140 vision-forge
EVIDENCE:
- ...

Constraints & Gotchas

• Describing inferred intent (e.g. “primary CTA”) as visible fact without grounding misleads downstream agents.
• Ignoring tokens.css / STYLE_LOCK yields off-brand assets that fail design QA.
• Claiming pixel QA or WCAG contrast without reading the image file is false evidence.
• Multi-image sets without a shared reference lock drift in weight, hue, and metaphor.
• Over-compressed PNG/WebP destroys 16–24px icon legibility; prefer crisp 1x/2x exports.
• Screenshots with device chrome or shadows: crop or mask before grounding counts.
• Transparent/dark-on-dark assets fail silent contrast checks — force a checkerboard or known bg.
• Do not use outside vision (route via /cat-vision or /opgrok).

Anti-patterns

• Inventing brand colors/type when tokens exist
• Ungrounded “the button is disabled” without visual cue evidence
• Skipping empty-state / error / loading visuals when the brief implies a full path
• Single-asset polish before the e2e set exists (breaks forge method)
• Upscaling tiny icons then claiming sharpness
• Do not write exploits, malware, or undisclosed destructive automation

Definition of Done

• Deliverable matches the brief under forge (e2e path first, edges hardened) for vision.
• Observation/inference split holds; token locks respected; claims cite files/elements.
• WIN: PASS with concrete evidence paths and tool outputs.
• Downstream SuperGroks can consume artifacts without re-asking brand or grounding questions.

Optional Tool Surface

• read_file / image view on screenshots and assets
• image_gen / image_edit for production and contrast/identity-preserving edits
• assets/tokens.css, STYLE_LOCK, design-token JSON when present
• Binary id: opgrok.sg.vision-forge

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