AI supervisor for Claude Code that reviews implementation plans and completed code
Two tools with reasonable naming and explicit schemas, but descriptions are generic and parameters lack guidance on format/validation. Both tools are correctly registered with input schemas in internal/mcp/server.go, making them verifiable. However, descriptions do not explain WHEN to use each tool, what prerequisites exist, or what data structures are returned. The consult_tech_lead tool's 'context' parameter has no type specification (just 'object'), and review_code's 'files' parameter similarly lacks detail. Error handling is present but minimal, no guidance on recovery or next steps for LLMs.
Consult the tech lead before showing a plan to the user. Required for all implementation plans.
Review completed code implementation. Call after finishing implementation for a quick sanity check.
Parameter descriptions lack format/validation guidance. 'context' in consult_tech_lead has no type; 'files' in review_code is described as 'Map of filename to content or diff' but lacks explicit schema. LLMs cannot infer whether to pass a flat object or nested structure.
Tool descriptions are too generic (44 and 54 chars respectively). They state WHAT but not WHEN to call or any prerequisites. 'Consult the tech lead before showing a plan' does not explain what 'decision' means or whether it blocks further action. 'Review completed code implementation' does not clarify if review must pass before deployment or is purely advisory.
No documented output schema. Tool handlers return CallToolResult with Content array, but no specification of what fields the LLM should expect (Decision, Reason, Guidance, RevisedPlan, RevisedFiles for consult_tech_lead). LLMs cannot plan downstream actions without knowing response structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 48 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 40 | 2024-11-05+ | v1 |
Error handling returns only a generic text message in CallToolResult without categorization. When tech_lead.Review() fails, the LLM gets 'Error consulting tech lead: <err>' with no guidance on whether to retry, ask the user, or abandon. Does not follow recovery-guide pattern.
consult_tech_lead naming mixes advisory and gating semantics. 'Consult' suggests optional advice, but description says 'Required for all implementation plans', a permission-gating or review-approval tool. Rename to clarify (e.g., review_implementation_plan, get_implementation_approval).
review_code 'files' parameter is ambiguous, unclear if agent should pass full content or diffs, and data structure (flat keys vs nested) is not specified. Leads to LLM guessing and malformed calls.