cdp-visual-evidence-verification

name	cdp-visual-evidence-verification
description	Use when verifying CDP screenshots and vision findings.
tags	["cdp","visual","verification","screenshot","vision","infocard"]

CDP Visual Evidence Verification

Purpose

Screenshots and vision-model findings are evidence only after verification. This skill covers two failure classes observed in production (2026-08-12, Qwen-MM-Plugins infocard run):

1. Screenshot misdelivery: the capture command returned a DIFFERENT page than the target.
2. Vision false positives: the vision model reported defects that did not exist in the artifact.

Both failures, if uncaught, either block a valid publish or pass an invalid one. This skill is the cross-check protocol.

1. Screenshot identity verification (shared CDP 9222)

On a shared --cdp 9222 Chrome, agent-browser screenshot can capture a DIFFERENT tab than the one open returned. Observed wrong captures: a ChatGPT 404 login page and a Trigger.dev preview page, both while the target card was open in another tab.

Byte-size is NOT identity. A wrong-page capture and the correct card measured identical 281167 bytes in the same session. Never infer page identity from file size, dimensions, or a successful exit code.

Mandatory protocol before trusting any screenshot

1. Bind the session with --pin-tab from the FIRST command; keep the same --session + --cdp 9222 + --pin-tab flags on open, set viewport, and screenshot.
2. Runtime.evaluate on the pinned targetId (via CDP or --json eval):
   • location.href must match the target URL
   • document.title must match the target title
   • scrollWidth == clientWidth (mechanical no-horizontal-overflow evidence; NOT visual pass)
3. Only after identity confirms, analyze/save the screenshot.

Finding the right target

Use CDP Target.getTargets to enumerate type="page" entries. Multiple tabs may share the same title; match on URL prefix, not title alone. Iframe targets (type="iframe") cannot be screenshotted directly.

2. Vision findings are hypotheses, not verdicts

A vision model report of a typo, layout asymmetry, missing element, or color mismatch must be checked against the artifact BEFORE acting. Observed false positives (2026-08-12):

Vision report	Reality after check
ARCHITURE spelling error (critical)	HTML contained correct ARCHITECTURE (grep confirmed)
Blue quote block "right edge flush / asymmetric" (major)	getBoundingClientRect(): gapLeft=gapRight=14, symmetric

Verification mapping

Vision claim	Check method
Text typo / wrong copy	grep -o "PATTERN" docs/<slug>.html (or search_files)
Layout asymmetry / element position	getBoundingClientRect() on element AND parent; compare left/right gaps
Missing / truncated content	DOM presence check (document.querySelector), element count
Color / theme mismatch	grep CSS variables in the file against the declared theme

Record only verified defects as critical/major. Do not block publish on an unverified claim, and do not dismiss a real defect without the same check.

3. Vision API failure handling

• 503 chat_admission_busy / structure_limit: infrastructure, not content. Wait (20–30s) and retry. Do not treat as a verdict.
• 400 invalid_request_error on image_url: image too large for the API. Shrink to ≤1400px width, JPEG quality ~82, then retry. Full-page mobile captures (e.g. 375×11457) MUST be downscaled before analysis.
• Screenshot captured but analysis failed: keep the screenshot file; retry analysis after resizing.
• Differentiated retries: after the retry budget is exhausted, record VISUAL_PENDING — never promote static/DOM checks to visual PASS.

4. Build timeout ≠ build failure

npm run build (or any long build) can exceed the terminal cap (e.g. 600s) while still completing successfully. After a timeout, verify before re-running:

grep -c "<slug>" _index.yaml   # index contains the card?
ls -la docs/<slug>.html*        # artifact mtimes current?

If both confirm, the build finished despite the timeout; proceed to commit/push. Re-running is a fallback, not the first step.

Pitfalls

• Never git remote set-url with a hand-typed placeholder token — tool output redaction hides the real value; overwriting pushurl with a placeholder silently breaks the next push. Restore the token from ~/.git-credentials via a script that READS the file (Python/shell), never echo it: set BOTH origin (fetch) and --push origin URLs, because set-url --push alone leaves fetch on SSH → Permission denied (publickey).
• agent-browser tabs command can hang on a busy shared browser; prefer CDP Target.getTargets via browser_cdp.
• After visual work, close ONLY the owned preview tabs (Target.closeTarget with exact targetId), then re-enumerate to confirm absence.

References

• references/2026-08-12-qwen-mm-plugins-evidence.md — 完整实录：截错 tab、视觉误报、vision API 失败形态、build 超时、git 凭据恢复（含命令）

Related

• infocard-publish-sop (user-owned): visual gate, screenshot manifest, publish discipline
• browser-core (user-owned): Tab ownership and lifecycle contract
• any2card (user-owned): visual evidence production for cards