Two tools with complete input schemas and descriptions, but both have significant gaps in documentation quality and parameter descriptions. Tool naming follows verb_noun convention (create_, get_), which is strong. However, descriptions are functional but lack depth regarding when to use each tool, error recovery guidance, and parameter format expectations. Input schemas are well-structured (Pydantic models with types), but parameter descriptions within the schema are inconsistent in quality. The create_llamator_run tool has a complex nested schema but several parameters lack actionable constraints (e.g., 'Optional API key' doesn't explain format or security implications). The get_llamator_run tool is simpler but has minimal documentation on expected output structure. Neither tool includes documented output schemas, which violates pattern:tool requirements. Error handling is present in implementation (_await_job_completion timeout, _build_error_notice) but not surfaced in the tool description to guide LLM recovery.

• Add explicit output schema documentation to both tools. Example: 'Returns LlamatorRunToolResponse with fields: {status: str, job_id: str, result: {aggregated: dict[str, dict[str, int]]}, error: {error_type: str, message: str} | null}'
• Move api_key out of tool parameters. Instead, configure OpenAI/compatible endpoints via server environment (LLAMATOR_MODEL_API_KEY, etc.) and reference by endpoint name in tool parameters (e.g., 'model_config_name': 'openai-test'). Document in tool description: 'Model credentials are configured server-side; pass the model config name, not the API key.'
• Upgrade 'preset_name' from description hint to formal enum schema constraint. List valid presets: ['all', 'rus', 'owasp:llm01', ...]. Include in description: 'Built-in preset name from the list of enum values.'
• Add timeout and retry guidance to create_llamator_run description: 'Polls job status for up to [timeout_seconds] seconds (default 300). Timeout errors are retryable; call again with the same job_id to resume waiting. Job completion status: SUCCEEDED or FAILED.'
• Add error recovery examples to both tools: 'If job not found, call with a different job_id or create a new run. If timeout occurs, call get_llamator_run(job_id) to check status without re-submitting. If results are incomplete, check run_config.enable_reports and artifacts_path settings.'
• Document 'artifacts_path' behavior: 'Relative path where LLAMATOR stores artifacts (reports, logs). Directory created automatically if missing. Path must be relative (no absolute paths, no /. or ../ traversal). Artifacts retrievable via [artifacts_download_url if enabled].'