visual-verifier

name	visual-verifier
description	Rendered-output verification station for the swarm post-flight eval. After a maker subagent produces a visual deliverable (UI/route/site), the station captures a full-page screenshot via agent-browser, then an independent verifier model (not equal to author, per aiewf P0-2) reads the screenshot against seed acceptance criteria + project DESIGN.md tokens + prior screenshots. Match marks the task complete; mismatch emits a structured visual diff for the maker's next iteration. The verifier is the exit condition, not the maker. Recurring visual failure modes graduate to instincts.yaml via the extract-patterns gate (ZOU-452).
compatibility	Created for Zo Computer
metadata	{"author":"marlandoj.zo.computer","linear":"ZOU-493","aiewf":"SIL-13","depends_on":"consensus-gate (reviewer-independence.ts), agent-browser, extract-patterns"}

Visual Self-Verification Station

When to use

The swarm post-flight eval (Stage 2) invokes this station when a seed task is flagged visual: true. The station runs AFTER text-based acceptance criteria pass — it is a second-order check that catches rendered-output failures text eval structurally misses (wrong palette, broken layout, hydration flash, overlapping elements, default Tailwind tokens instead of project DESIGN.md).

Architecture (confirmed defaults)

• Multimodal read OUTSIDE the gate (q1=A): the station's verifier script sends the screenshot to a vision-capable model and gets a text verdict + structured diff. The consensus-gate is NOT modified — only the reviewer-independence.ts helper library is reused for the ≠author constraint.
• Single verifier (q2): one multimodal model reads the screenshot. A multi-verifier panel is available behind VISUAL_VERIFIER_PANEL=1 for future activation but defaults off.

Scripts

scripts/capture.ts

Captures a full-page screenshot of a URL via agent-browser.

bun scripts/capture.ts --url "http://localhost:3099/my-route" --output /tmp/screenshot.png

scripts/verify.ts

Independent verifier — reads the screenshot image, compares against references, returns a structured verdict.

bun scripts/verify.ts \
  --screenshot /tmp/screenshot.png \
  --criteria "The page uses the project's amethyst+gold palette, not default zinc" \
  --design-md /home/workspace/Projects/my-project/01-brand/DESIGN.md \
  --author "hf:zai-org/GLM-5.2" \
  --label "my-route"

scripts/station.ts

Full pipeline: capture → verify → emit verdict. Used by the post-flight harness.

bun scripts/station.ts \
  --url "http://localhost:3099/my-route" \
  --criteria "..." \
  --design-md "..." \
  --author "..." \
  --label "..." \
  --project-dir "..."

Config flags

Flag	Default	Effect
VISUAL_VERIFIER	ON (1)	Master switch; 0 skips the station entirely
VISUAL_VERIFIER_MODEL	gpt-4o	Multimodal model for the screenshot read
VISUAL_VERIFIER_PANEL	OFF (0)	If ON, runs a multi-verifier panel via consensus-gate helpers
VISUAL_VERIFIER_HYDRATE_MS	3000	Sleep after agent-browser open before screenshot
VISUAL_VERIFIER_CONF_THRESHOLD	0.7	Minimum verifier confidence to act on a mismatch

Compounding

Every mismatch is logged to Projects/<project>/visual-failures.jsonl. The existing extract-patterns gate (ZOU-452) reads this log on its scheduled run and applies the 4-criteria test; a pattern with >=2 confirmed occurrences graduates to instincts.yaml. The station does NOT write instincts directly — it feeds the gate.

Rollback

VISUAL_VERIFIER=0 disables the station. Non-visual deliverables are byte-identical to today (the visual flag is absent by default). Full revert: remove the Stage 2 plug-in call, delete this skill, remove the visual flag from the seed schema.