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).
else redirect (new design / a11y / generic CSS = out of scope). Read the
targeted request (--pages, --selector); with none, a bounded default scan.
.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.
[--css sheet.wxss]. Deterministic findings thresholded by design-tokens.json`.
resolve the needs_judgment long tail by visual reasoning, label visual
findings as estimates. Never fabricate a ratio.
audit.py writes the doc set to the out dir (default audit.json validates and every section is
present, including the zero-findings "clean" doc.
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.
| 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 | visual_estimate] [--json]` | scope → analyze → emit doc set; never edits the target (exit 3 = empty scope) |
scripts/analyze.py | deterministic findings JSON for one file (H5 or WXML+WXSS) | |
evals/run_all.py | regression + behavioral harness — every adversarial edge bound to a case |
共 1 个版本