Evidence | agentlane

The evidence folder

Every lane has a folder in agentlane's state, ~/.local/state/agentlane/lanes/<id>/evidence. Hooks and exec commands get its path in AGENTLANE_EVIDENCE. It survives release, and goes seven days after the lane ends, with the rest of the lane's files.

npx agentlane evidence list

Each evidence command prints the path of what it wrote, so an agent can cite the file in its report. When the lane holds a device and a browser, --device and --browser pick which one.

Screenshots

npx agentlane evidence screenshot

The same picture as screenshot, kept in the evidence folder: a JPEG from a device, a PNG from a browser.

Video

npx agentlane evidence video start
# drive the change
npx agentlane evidence video stop

On a simulator it records with simctl, as H.264. On an emulator it uses screenrecord, which Android caps at 180 seconds. In a browser it records a WebM of the page. A video suits a change a person will review.

Traces

npx agentlane evidence trace start
npx agentlane evidence trace stop

What a trace holds depends on what it traces:

Target	The trace
A Playwright browser	Playwright's action trace, which `npx playwright show-trace` opens. It runs from `up` on, so a `stop` without a `start` covers the session so far.
An agent-browser browser	A DevTools performance trace, between a `start` and a `stop`.
A device	What agentlane saw: the driver's log, the flows that ran, the screenshots they kept, and the device's log over the traced time.

Logs

npx agentlane evidence logs --since 5m

One bundle of the lane's logs over a window: the services, the hooks, the driver, the device's log and the browser console. --since takes minutes, seconds or hours, such as 90s or 2h.

What a failure keeps

Evidence of a failure is kept without anyone asking:

A failed run, and a failed command under exec on a browser lane, keep their own bundle: the trace, a screenshot, the screen or the page, and the driver's log or the console.
maestro test under exec or in a hook writes its whole output to the evidence folder, its takeScreenshot files included, in a folder per run: <stamp>-maestro-test-<pid>, ending in -failed when the test failed.

Each Maestro run is also linked into Maestro's own ~/.maestro/tests, so a script that reads that folder keeps working. That folder holds the runs of every lane on the machine, though, so a script that runs in a lane should read $AGENTLANE_EVIDENCE.

The project's Playwright tests can write there too: point outputDir at a folder under AGENTLANE_EVIDENCE, as Browsers for web apps shows, and a failed run's trace survives the next run and the release.