Capturing portable evidence

The goal is an artifact a human (and the review agent) can open and believe. Pick the lightest capture that proves the criterion, and prefer engine-level capture so it works the same locally and headless/cloud. UI capture uses agent-browser (agent-browser.md) on the 端 you chose (web / electron; backend/CLI → cli).

Evidence type → how to capture

screenshot — UI state

Screenshots via CDP (the browser engine) need no real display:

# web (named session)
agent-browser --session app open "<url>"
agent-browser --session app click @e1 # act, after snapshot -i to get refs
agent-browser --session app screenshot ./proof/after.png

# desktop (Electron, over CDP)
agent-browser --cdp 9222 screenshot ./proof/desktop.png

Avoid macOS screencapture / osascript unless the criterion is explicitly native/local-only — they don't run in the cloud.

dom_snapshot — structure / content proof

When the proof is "this element/text exists with these values", a DOM dump is smaller and more assertable than a pixel screenshot:

agent-browser --session app eval "document.querySelector('[data-testid=list]').outerHTML" > ./proof/list.html

Upload as --type dom_snapshot --file ./proof/list.html.

text — stdout, logs, computed values

Backend / CLI behavior is best proven by the command output itself — upload inline, no file:

lh verify submit --operation "$LOBE_OPERATION_ID" --item "$CHECK_ITEM_ID" --type text \
  --content "$(your-cli command --json)" \
  --by cli --desc "command reports success after the change"

A network capture (HAR) or a request/response pair from a full-stack run also uploads as text (--file ./proof/capture.har or --content).

transcript — conversation / request log

A saved agent conversation, request/response pair, or event log file: --type transcript --file ./proof/run.jsonl.

gif / video — behavior over time

Required whenever a criterion asserts change over time (streaming output, a ticking timer, a loading→loaded transition, an animation, a multi-step flow) — a static screenshot cannot prove these. Record a clip and upload --type gif / --type video --file …. Full recipes (portable CDP frame-sequence → MP4/GIF, OS screen recording, GIF-vs-MP4): recording.md.

Provenance (--by)

Tag how the artifact was produced so the reviewer can weigh it:

• agent-browser — captured through agent-browser (screenshots, DOM, eval)
• cdp — captured directly via Chrome DevTools Protocol
• cli — command stdout / computed text
• program — produced by a script or test you ran

Headless / cloud portability

The decisive constraint per 端 is how a screenshot is captured: engine-level capture (CDP) needs no display; OS-level capture is macOS-only.

Checklist:

• ✅ agent-browser screenshot / eval (DOM, console) — engine-level, headless-safe
• ✅ --type text --content … — pure text, always works
• ❌ screencapture, osascript, native-app screen recording — macOS-only, not cloud-safe

When a run must stay cloud-portable, prefer CDP-based evidence over OS-level capture wherever both exist.

Don't leak secrets

Evidence is visible to reviewers. Never capture a screenshot or text dump that contains a live token, cookie, password, or other secret — see auth.md.