Repository navigation
Persistent HTML report after a run (video, screenshots, network, traces) - Playwright HTML-report style #1967
Description
Summary
After I drive a flow with agent-browser (CLI or coding agent), I want a persistent, shareable HTML report of what ran — not only the live dashboard.
Today the observability dashboard is live-only. Once the session closes, I cannot go back and see what was executed. I want an artifact I can open later (or attach to a PR / Slack thread) with:
- Timeline of commands
- Screenshots
- Session video
- Network activity
- Traces (or equivalent)
This is the same job Playwright’s HTML reporter / Trace Viewer does for Playwright Test (npx playwright show-report).
Problem
agent-browser dashboard is excellent for watching a run as it happens (viewport, Activity, Console, Network). It is not a record of a finished run.
If a coding agent verifies a page and then closes the session, there is no first-class way to:
- Replay the steps
- Hover a timeline and see the page at that moment
- Inspect network/console for that session after the fact
- Hand someone a single folder/URL instead of a pile of screenshots
record start/stop gives a WebM; screenshot gives a PNG; trace start/stop gives Chrome tracing JSON. Those are useful pieces, but they are not one reviewable report, and agents cannot query a Chrome trace without a GUI (related: #1280).
What I want
An opt-in “save report” at the end of a session (or wrapping a batch / agent run), producing a local HTML report I can open without keeping the browser or dashboard up.
Suggested contents (Playwright HTML report / Trace Viewer as the UX bar, not necessarily the same engine):
| Artifact | Why |
|---|---|
| Command timeline | What the agent/CLI actually did, with timestamps |
| Screenshots | Before/after or per navigation |
| Video / screencast | Visual replay of the session |
| Network log | Failed requests, API calls during the flow |
| Trace (or HAR + actions) | Debug “why did this click miss / why did we land on sign-in” |
| Console errors/warnings | Same as dashboard Console, persisted |
Ideal UX: something I can open locally like playwright show-report (e.g. agent-browser report show <dir> on 127.0.0.1). Does not need to upload anywhere.
Proposed shape (for discussion)
Keep the dashboard live. Add a session archive:
agent-browser --session verify-dashboard open https://example.com/app
# ... clicks / snapshots ...
agent-browser --session verify-dashboard report save ./ab-report
agent-browser report show ./ab-report
Activity
Hronom commentedon Sep 21, 2026
One boundary that seems important for the first version: make the archive a run receipt, not just a media bundle. I would keep separate states for requested, dispatched, observed, and verified, each bound to session or profile generation, target fingerprint, origin, and timestamp. A screenshot or video proves what was rendered, not that a write succeeded; the report should show the authoritative fixture read-back separately. If record stop, browser close, or artifact generation is interrupted, mark the report partial or unknown rather than presenting a complete run.
For a shareable local report, I would also make redaction defaults explicit: omit cookies, authorization headers, form values, and bodies from network or console capture; use host and secret-pattern allowlists; include an artifact manifest with hashes and missing-artifact reasons. A disposable public or localhost fixture plus clean-quit and crash tests would make the report contract reproducible without uploading anything.
I maintain Hronaut, a local visible Browser/MCP workspace, and we use these distinctions when checking workspace and tab continuity. AI-assisted; no private run data or integration claim.
christianfrei791 commentedon Sep 21, 2026
To move this toward a structured "run receipt" as suggested by Hronom, we could implement a post-session hook in the agent orchestrator that serializes the internal state manager's history into a JSON manifest alongside the raw artifacts. If we standardize this manifest, we could bundle it with a static HTML viewer—effectively a simplified version of the Playwright Trace Viewer—that maps the
SessionIDto specificActiontimestamps and network logs. Would it be feasible to extend the currentbrowser.close()flow to trigger anexport_session_artifacts()method that aggregates these disparate logs into a single directory before cleanup?Hronom commentedon Sep 21, 2026
Yes, a post-session hook is a good boundary, but I would not make a successful browser.close() the only condition for a usable receipt.
A small finalization state machine could make the failure modes explicit:
I would test clean close, already-closed browser, close timeout, process crash during flush, concurrent close calls, zero-action runs, a late event after the close barrier, and an artifact that is present but fails hash/read-back. The static viewer can then render the same manifest whether the browser closed cleanly or the orchestrator had to recover it. Keep the existing redaction defaults for cookies, authorization headers, form values and bodies.