Feedback wanted: Playwright + LLM trace fields

[aaronlab]

BrowserTrace has a focused Playwright + LLM debugging guide:

https://aaronlab.github.io/browsertrace/playwright-llm-debugging.html

The current low-friction path is to call run.snapshot(page, ...) around the points where an LLM chooses the next browser action. That records the Playwright URL and screenshot plus optional model_input and model_output.

For public sharing, use:

browsertrace export <run_id> --public -o public.html

--public omits prompt/model I/O, screenshots, and URLs from the standalone HTML export.

I want feedback from people building LLM-guided Playwright scripts or browser agents:

• Should BrowserTrace add a higher-level Playwright + LLM helper beyond run.snapshot?
• What fields should it capture by default?
• Do you want selector/action, DOM snippet, accessibility tree text, network errors, console errors, prompt/messages, model output, retries, or extracted data?
• Are individual redaction flags enough, or should public-safe export support more fields?

Please comment with the shape of a real failure you have debugged and which context was missing from logs or Playwright Trace Viewer.

[aaronlab]

Added a no-network Playwright + LLM-shaped example in commit 401a482:

• examples/playwright_llm_loop_example.py
• README and examples/README.md now link it
• docs/playwright-llm-debugging.html now mentions it from the source-checkout path

The example records a failed checkout flow with:

• action labels: observe page -> choose selector -> click model-selected button
• model_input.messages
• page URL, DOM snippet, accessibility tree text
• console/network error slots
• model_output.action and model_output.selector
• agent_stack="playwright+llm", selector, and retry metadata
• a failed step when button.checkout.primary is disabled

Still looking for feedback on whether this should stay as a documented pattern around run.step / run.snapshot, or become a higher-level Playwright + LLM helper.

[laolaoqi]

I can contribute to this integration effort. I have experience with Playwright, LLM tracing pipelines, and Browser Use debugging patterns.

I can help define the trace fields that would capture Playwright + LLM agent failures effectively:

• Step-by-step Playwright action logs (navigation, click, type, screenshot)
• LLM reasoning traces per action
• Screenshot evidence at failure points
• Network request/response pairs for API-calling agents
• Console errors and page state snapshots

Timeline: Field definitions within 24h, integration PR within 48h of discussion.
Skills: Python, Playwright, LLM agent tracing, Browser Use
Payment: USDT/USDC TRC20
Wallet: 0x6c21794E27C34E93cd0049918d04Da58FC72595c

Happy to start a discussion thread on the specific fields that would be most useful.

[aaronlab]

Thanks for offering.

This repository is not offering paid work or bounties. Please only proceed if you want to contribute as a normal voluntary open-source contribution.

For this feedback issue, the most useful contribution is concrete field feedback rather than a broad integration PR: a real Playwright + LLM failure shape, which fields were missing from the trace, and any privacy/safety constraints around screenshots, network data, console logs, or model output. If you want to propose a code change later, please first keep the scope narrow and discuss it here.