test(browser): regression tests for issue #723 error clarity

[sena-labs]

Summary

• Add three regression tests to tests/test_browser_agent_regressions.py that guard against reintroducing the confusing "Task reached step limit without completion. Last page: about:blank" error reported in Problem with the browsing tool #723.
• Each test patches get_runtime to simulate a different failure mode of the Playwright-based browser tool and asserts that the returned message is specific, actionable, and does not contain the old "step limit" / "about:blank" wording.

Background

Issue #723 reported that every browser task since v0.9.4 silently failed with:

Task reached step limit without completion. Last page: about:blank.
The browser agent may need clearer instructions on when to finish.

Root cause: the removed browser-use library sent an anthropic-beta: structured-outputs-2025-11-13 header to OpenRouter, which proxied requests to providers (Amazon Bedrock, Google Vertex) that reject it. After three consecutive API errors, browser-use exhausted its step budget and returned the generic fallback message above — giving users no indication that the real problem was an API rejection.

The fix (commit 983d431a) replaced the browser-use agent with the native Playwright _browser plugin, which propagates concrete error messages:

Failure mode	Current message
Runtime cannot start	Browser runtime unavailable: <reason>
Action call fails	Browser <action> failed: <reason>
Unknown action keyword	Unknown browser action: <name>

The existing test_legacy_browser_dependency_is_removed test already verifies that browser-use and _browser_agent are gone. This PR adds complementary tests that verify the new tool's error-handling contracts, so the old vague wording cannot be reintroduced undetected.

Tests added

Test name	What it verifies
test_browser_tool_returns_clear_error_when_runtime_unavailable	get_runtime raises → message contains "runtime unavailable", not "step limit"
test_browser_action_failure_names_the_failed_action	runtime.call raises → message names the failing action and contains "failed"
test_browser_unknown_action_returns_informative_message	Unknown action keyword → message names the keyword, not a blank failure

Test plan

pytest tests/test_browser_agent_regressions.py \
  -k "test_browser_tool_returns_clear_error_when_runtime_unavailable \
      or test_browser_action_failure_names_the_failed_action \
      or test_browser_unknown_action_returns_informative_message \
      or test_legacy_browser_dependency_is_removed" \
  -v

All four tests pass. No other tests are modified.

Closes #723

🤖 Generated with Claude Code