MCP server that exposes a review database via stdio, providing tools to manage code review comments, search symbols, and read code indices
Conductor MCP has well-structured tool definitions with clear naming (verb_noun pattern) and comprehensive descriptions (avg 180 chars, within 10-1024 baseline). All 9 tools have input schemas with typed parameters and descriptions. However, output schemas are not documented, LLMs cannot predict return structure. Parameter descriptions lack format constraints (e.g., line_start is 1-based but not stated in description). Error handling is minimal; no recovery guidance or actionable error messages visible. Tool composition is sound (single responsibility, clear chains via comment_id), but missing pagination for list-like tools and no confirmation pattern for destructive operations (resolve_comment, reply_to_comment).
Leave an inline self-review comment on a file/line range in the current branch's diff. Author is set to 'claude' and it appears inline in the Conductor diff view. High-signal, low-frequency tool: flag only what a human reviewer genuinely needs — tricky logic, deliberate tradeoffs, or spots you are unsure about — not routine changes the diff already makes obvious.
Get the summary of changes for the current branch's diff.
Get full details of a review comment and all its replies. Use the comment ID (or prefix) from get_pending_comments.
List unresolved (pending) review comments. By default, only comments for the current git branch are returned. Set all_branches=true to see comments across all branches. Use get_comment_thread to read full details and replies for a specific comment.
Read the definition and documentation of a specific code symbol from the semantic index.
Add a reply to a review comment. Author is automatically set to 'claude'.
Output schemas not documented. LLMs cannot predict return structure for any tool, forcing them to guess at downstream field names and types. This breaks tool chaining and wastes tokens on clarification.
Destructive operations (resolve_comment, reply_to_comment, create_comment) lack confirmation or dry-run pattern. Agents can irreversibly modify review state without safeguards.
Parameter descriptions lack format constraints. 'line_start' is 1-based but not stated; 'kind' accepts 'Issue' or 'Note' but no enum constraint; 'comment_id' accepts prefix but matching logic undocumented.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | <=2025-11-25 | v2 |
Mark a review comment as resolved.
Search for code symbols (functions, types, etc.) in the repository's semantic index. Returns matching symbols with their locations and confidence scores.
Set the summary of changes for the current branch's diff.
No error handling guidance. Tools return errors but do not tell LLMs what to do next (retry, ask user, abort). No categorization of retryable vs fatal errors.
get_pending_comments and search_symbols lack pagination. No limit, offset, or cursor parameters. Large result sets will blow context window.