Loading ToolBench
Automatic Codex plan review in Claude Code; on-demand in local MCP hosts; human approval stays yours
Single tool 'review_plan' has a well-structured input schema with 5 parameters (1 required, 4 optional) and clear descriptions. Tool name uses verb_noun pattern (review_plan). Description is 189 chars, within baseline (p10=34, p90=392). However, output schema (ReviewPlanToolOutputShape) is not visible in source, only referenced by import. Tool annotations present (readOnlyHint, idempotentHint, destructiveHint, openWorldHint). No parameter enums despite 'verdict' field likely having constrained values (approve|revise|reject). Error handling and recovery guidance absent from description.
Review a development/architecture plan with peer LLMs and return a consolidated verdict (approve | revise | reject) with prioritized findings. Runs configured reviewer CLIs and writes a local session log to ~/.inspectrum/sessions/.
Output schema not visible in source code. ReviewPlanToolOutputShape is imported but definition not provided. Cannot verify return structure, field types, or whether verdict enum is properly constrained.
Tool description does not document error cases or recovery guidance. No mention of what happens if reviewer CLIs fail, timeout, or return invalid JSON. LLM has no guidance on retry strategy or fallback.
Parameter 'focus' and 'context' lack guidance on expected format or length. Description is generic ('Specific areas to focus the review on'). No constraint on token budget or max length, risking prompt injection or context overflow.
Tool writes to local filesystem (~/.inspectrum/sessions/) but description does not warn about side effects or idempotency. Annotation sets idempotentHint=false, but no guidance on whether repeated calls with same input produce same session or new ones.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2025-06-18+ | v2 |
No pagination or result limiting documented. If consolidated review contains many findings, response could be very large. No mention of max findings returned or truncation behavior.