* feat(tools): add Google Custom Search as alternative to Brave Search
Adds google_search tool using Google Custom Search API as an alternative
to the existing web_search tool (Brave Search).
Changes:
- Add google_search_tool with full implementation
- Register Google credentials (GOOGLE_API_KEY, GOOGLE_CSE_ID)
- Register tool in tools/__init__.py
- Add README with setup instructions
Closes#793
* test(tools): add unit tests for google_search tool
Adds 7 tests mirroring web_search_tool test patterns:
- Missing API key error handling
- Missing CSE ID error handling
- Empty query validation
- Long query validation
- num_results clamping
- Default parameters
- Custom language/country parameters
All tests pass.
* refactor(tools): add multi-provider support to web_search tool
BREAKING CHANGE: None - backward compatible. Brave remains default.
- Add Google Custom Search as alternative provider in web_search
- Add 'provider' parameter: 'auto' (default), 'google', 'brave'
- Auto mode tries Brave first for backward compatibility
- Remove separate google_search_tool (consolidated into web_search)
- Update tests to cover multi-provider functionality (13 tests)
- Update README documentation
Users with BRAVE_SEARCH_API_KEY: No changes needed
Users with GOOGLE_API_KEY + GOOGLE_CSE_ID: Can use provider='google'
Users with both: Brave preferred by default, use provider='google' to force
Closes#793
* feat(tools): fixed readme
---------
Co-authored-by: Mustafa Abdat <abdamus@hilti.com>
Centralized _get_api_key in prompts.py to support OpenAI, Cerebras, and Groq via environment variables while maintaining Anthropic support through CredentialManager.
Previously, when exports/ was missing or empty, the bash glob
`exports/*/` would not match anything and the loop would silently
do nothing. The job would pass without actually validating anything,
which was misleading.
Now the job:
- Explicitly checks if exports/ directory exists
- Uses nullglob to handle empty directories properly
- Logs clear messages when skipping validation
- Reports the number of agents validated when successful
Fixes#887
The "Available Tools" table listed `execute_command` but the actual
registered name is `execute_command_tool`. This aligns the docs with
the runtime name in __init__.py and the tool's own README.
Fixes#901
- Add output_model field to NodeSpec for specifying Pydantic model
- Add max_validation_retries field (default: 2) for retry configuration
- Add validation_errors field to NodeResult for error tracking
- Implement validate_with_pydantic() in OutputValidator
- Implement format_validation_feedback() for LLM retry prompts
- Auto-generate JSON schema from Pydantic model for response_format
- Add retry loop that feeds validation errors back to LLM
- Add 28 comprehensive tests covering all new functionality
Fernando Mano