fix: prevent false success reporting on unresolved blocking error

[sauravpanda]

Summary by cubic

Prevents false success by adding a blocking‑error check and tightens data grounding to verbatim copy across all browser agent prompts. Still forbids fabricated data; derived values are allowed with screenshot-based verification. Also reverts google-genai to 1.60.0 for stability.

• Bug Fixes

  • Blocking error policy: payment declined, login without creds, email/verification wall, required paywall, access denied → set success=false. Temporary obstacles (CAPTCHA, popups) don’t count.
  • Data grounding: every URL/price/name/value must be observed verbatim in tool outputs, browser_state, or screenshot — copy exactly; don’t paraphrase or normalize URLs. Derived values (counts/totals) from observed data are allowed. Never construct; say not found.
  • Prompt cleanup: removed hard‑constraint enforcement, explicit scroll exceptions/“pages below” rules, and extract dedup guidance (already_collected/results file). Applied across all prompt variants.

• Dependencies

  • Reverted google-genai from 1.65.0 to 1.60.0.

^{Written for commit b43c7dd. Summary will update on new commits.}

[cubic-dev-ai]

2 issues found across 5 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="browser_use/agent/system_prompts/system_prompt_anthropic_flash.md">

<violation number="1" location="browser_use/agent/system_prompts/system_prompt_anthropic_flash.md:96">
P2: This grounding check omits `browser_vision`, so screenshot-only facts can no longer be treated as verified even though the prompt defines screenshots as the ground truth.</violation>
</file>

<file name="browser_use/agent/system_prompts/system_prompt.md">

<violation number="1" location="browser_use/agent/system_prompts/system_prompt.md:148">
P2: This wording forbids grounded derived results like counts, so tasks that require counting or simple computation can be incorrectly reported as `not found` or unsuccessful.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[github-actions]

Agent Task Evaluation Results: 2/2 (100%)

View detailed results

Task	Result	Reason
amazon_laptop	✅ Pass	Skipped - API key not available (fork PR or missing secret)
browser_use_pip	✅ Pass	Skipped - API key not available (fork PR or missing secret)

Check the evaluate-tasks job for detailed task execution logs.

[cubic-dev-ai]

1 issue found across 3 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="browser_use/agent/system_prompts/system_prompt.md">

<violation number="1" location="browser_use/agent/system_prompts/system_prompt.md:148">
P2: Keep a verbatim requirement for direct fields; `observed` alone is loose enough to allow paraphrased names or normalized URLs back into `success=true` responses.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}