Repo-aware directive generator with optional MCP server. Exposes OpenAgent pipeline as MCP tools for analyzing repositories and generating implementation directives.
OpenAgentMCP exposes 6 tools with clear verb-noun naming (analyze_repo, assess_repo, run_tests, get_test_log, list_repos, get_repo). All tools have descriptions (avg ~150 chars, within 10-1024 baseline). Input schemas are present with typed parameters. However, parameter descriptions are sparse or missing in several tools (e.g., run_tests lacks detail on test_cmd resolution logic; get_repo has minimal context). Output schemas are documented in docstrings but not formally declared in JSON Schema. Error handling is basic, run_tests returns structured error dicts, but recovery guidance is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels (analyze_repo is WRITE, others READ_ONLY). Missing pagination for list_repos despite caching. Overall: solid foundation, but lacks LLM-optimization rigor and formal output schema declarations.
Analyze a repository and generate a directive. Runs the full pipeline: Scanner → Assessor → Writer → StateWriter. Writes docs/state/current.md in the target repo.
Assess a repository's current state without generating a directive. Runs Scanner + Assessor only. Does not write any files. Cheaper and faster than analyze_repo — useful for status checks.
Return cached context for a single repo by directory name. Does not trigger a scan — reads cache only.
Read the async test log and check process status.
Discover all git repos one level under root_path. Results are cached for OPENAGENT_CACHE_TTL_SECONDS (default 300s). refresh=True forces a full re-scan regardless of cache state.
Spawn the repo's test suite as a background process. If test_cmd is empty, resolves from AGENT_CONTRACT.md or detects from repo contents. Returns within one second.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite clear risk levels. analyze_repo and run_tests are destructive (WRITE); others are read-only. LLMs cannot infer safety without explicit hints.
Output schemas not formally declared in JSON Schema format. Docstrings describe return dicts, but LLMs cannot parse unstructured docstring descriptions reliably. Missing typed field definitions.
Parameter descriptions incomplete or vague. run_tests: 'test_cmd' lacks detail on resolution fallback logic. get_repo: 'name' description is minimal ('Directory name of the repository'). list_repos: 'max_repos' lacks rationale for 50/200 bounds.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
list_repos returns cached results but lacks pagination support (no limit/offset in response, no next_cursor). Caching TTL is server-side only; no client-side cache control hints.
Error handling lacks recovery guidance. run_tests returns 'No test command found' but does not suggest how to add test_cmd to AGENT_CONTRACT.md or what format is expected.