Low-Visibility Fix

Audit existing mobile UI (WeChat mini-program / H5 / app) for FIELD low-visibility conditions — low light, glare, gloves, wet hands, vibration — and emit a structured, implementer-ready handoff DOCUMENT SET (findings by severity + prioritized fix recommendations). Runs a deterministic analyzer plus a bounded visual/browser pass; scopes to specific page(s)/component(s) for cheap multi-round re-runs. NEVER edits the target — it hands documents to another agent to apply. Use when someone has an existing field/industrial UI and wants it checked for contrast, tap-target size, color-only state, or icon-only buttons, and wants a fix plan. Triggers: "审计这个界面在低光/眩光/戴手套下是否清晰可点， 给我修改建议文档", "audit this field UI and hand me a fix plan", "check only <page/component> for low-visibility issues", "$low-visibility-fix". Not for: editing/fixing the files directly (this skill produces docs, an implementer applies them); designing a new UI from scratch; disability/screen-reader a11y (colorblind, low-vision); generic CSS review with no low-visibility angle.

Low-Visibility UI Audit → Handoff Docs

Audit existing mobile UI for field low-visibility conditions — environmental (low light, glare) + physical (gloves, wet hands, vibration) — and emit a structured handoff document set for another agent to implement. Field conditions only; disability/a11y populations are out of scope — redirect them.

This skill never edits the target. Its stable, must-always-succeed deliverable is the doc set (audit.json + report.md), written only to the out dir. It is built to run many rounds on a scoped subset — keep each run bounded (targeted pages/components, capped scan).

Protocol — analyze, document, hand off (never edit the target)

1. Scope — confirm a field low-visibility audit of an EXISTING mobile UI; else redirect (new design / a11y / generic CSS = out of scope). Read the targeted request (--pages, --selector); with none, a bounded default scan.
2. Preflight — locate the in-scope files for the targeted scope only: H5 .html/.css, or mini-program .wxml/.wxss. Pick the visual-pass render method (rules/visual-pass.md); if none is available, degrade — still emit docs.
3. Analyze — python3 scripts/audit.py <target> [--pages a,b] [--selector .x] [--css sheet.wxss]. Deterministic findings thresholded by design-tokens.json.
4. Visual pass (bounded) — render the targeted page(s) at a mobile viewport, resolve the needs_judgment long tail by visual reasoning, label visual findings as estimates. Never fabricate a ratio.
5. Synthesize — audit.py writes the doc set to the out dir (default <target>/.lv-audit/). Confirm audit.json validates and every section is present, including the zero-findings "clean" doc.
6. Report — summarize findings by severity, point at the doc paths + residual needs_judgment, suggest the next-round scope. Do not edit the target.

Output goes only to the out dir (writing into the target source tree → OUTPUT_ONLY refusal). No network beyond rendering the target. Refuse to recommend lowering contrast/target below the field threshold for a field-critical control; offer a compliant alternative. Severity, interpretation, doc-set spec, and degraded mode → the modules below.

Modules & scripts

Path	When
rules/audit-protocol.md	Step 3–4: severity model, reading findings, completing needs_judgment, degraded mode, adversarial holds
rules/handoff-docs.md	Step 5: doc-set sections + how fixes are written as recommendations (not applied)
rules/visual-pass.md	Step 2/4: per-target render method, what the visual pass catches, the page cap, graceful degradation
references/design-tokens.json	threshold values (single source of truth)
references/field-conditions.md	why each threshold exists; standards citations
assets/handoff-doc.template.md	the doc-set shape an implementer agent reads
assets/fix-snippets.html	compliant component patterns referenced by recommendations (never applied)
scripts/audit.py <target> [--pages a,b] [--selector .x] [--css f] [--out d] [--input-mode static|visual_estimate] [--json]	scope → analyze → emit doc set; never edits the target (exit 3 = empty scope)
scripts/analyze.py <file> [--css f] [--selector .x] [--viewport-px n]	deterministic findings JSON for one file (H5 or WXML+WXSS)
evals/run_all.py	regression + behavioral harness — every adversarial edge bound to a case