[mathisdittrich]

browser-use/browser-use#6398 (merged 2026-11-14) brings the judge back in the open-source library: after each run that took a step, one low-effort request to the agent's own model fills result.judgement (plus one text-only retry if the model rejects the images with an HTTP 4xx, per #6403). The open-source docs still said the judge was removed (use_judge ignored, ground_truth / judgement() → result.success).

Changes, all facts from the #6398 diff (MIGRATING.md, README.md, agent/service.py, agent/judge.py, agent/views.py, legacy.py, telemetry/):

• all-parameters: new judge (default True) and judge_context (default None) rows; use_judge removed from the ignored list (it is now an alias, popped from kwargs, no longer in IGNORED).
• output: judgement field, cost_usd includes the judge, new "Judge" section with the Judgement fields.
• migrating: use_judge / ground_truth kept as judge / judge_context; judge_llm dropped; judgement() is now the result.judgement attribute; browser_use.agent.judge JudgementResult / SimpleJudgeResult point to Judgement.
• observability / telemetry: cost covers the judge; agent_event carries the scrubbed judgement (schema 5), only whether judge_context is set is sent as configuration; the verdict is scrubbed text, so short or partial quotes of the context can remain.

⚠️ Merge after the next browser-use release. #6398 merged after 1.0.4 (PyPI 2026-11-13 20:56 UTC), and the 1.0.4 wheel has neither judge nor judge_context, so pip install browser-use users would see params that do not exist yet. (The docs-oss-params CI installs browser-use from GitHub main, so it already passes.)

Not documented (unsure): the exact warning text a failed judge logs, and whether the CLI/MCP server expose the judgement.

🤖 Generated with Claude Code

---

Summary by cubic

Updates the open-source docs to reflect that browser-use 1.0 judges every run that took a step, filling result.judgement with the verdict.

• Documents judge (default True) and judge_context, plus result.judgement and its fields; gotchas are also written on success when a site rejected, errored or blocked the agent, and if the model rejects images with an HTTP 4xx error that mentions image, the judge retries once with text alone.
• use_judge and ground_truth stay as aliases, judge_llm is dropped, and judgement() becomes the result.judgement attribute.
• The judge reads the run as redacted text capped at 120k characters, plus up to 4 screenshots (the run's last distinct ones and the final page) when vision is on; a failed judge request leaves result.judgement None and never changes the run's own result.
• Notes the judge's cost in cost_usd and the scrubbed verdict in telemetry, with the 300-character caps that apply per field; judge_context itself is never sent, though partial quotes of it can survive in the scrubbed verdict.

Merge after the next browser-use release. browser-use/browser-use#6398 merged after 1.0.4, and PyPI 1.0.4 has neither judge nor judge_context; the parameter check CI installs from main and passes.

^{Written for commit fc46e1d. Summary will update on new commits.}

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
browser-use	🟢 Ready	View Preview	Nov 14, 2026, 9:21 AM

💡 Tip: Enable Automations to automatically generate PRs for you.

[cubic-dev-ai]

All reported issues were addressed across 5 files

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 1 file (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 2 files (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 2 files (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>