Loading ToolBench
Lightweight multi-model advisor for Claude Code — parallel Codex + Gemini consultation via MCP
Single tool 'consult' has a complete input schema with 6 parameters, all typed and described. Tool description is detailed (Chinese, ~180 chars) and explains use cases and limitations. However, output schema is not documented, callers cannot predict the structure of formatAsMarkdown() results. Parameter descriptions are concise but lack format/constraint details (e.g., timeout_sec range, cwd validation). No error recovery guidance in descriptions. Tool is read-only (safe), but lacks idempotent/destructive annotations.
并发调用 Codex、Gemini、Claude CLI，获取三个模型的独立视角。适合架构选型、方案设计、技术决策、复杂调研。调用方担任 Judge 盲评综合输出。不适合：typo 修复、单行改动、只读查询。
Output schema not documented. formatAsMarkdown() result structure is opaque to callers, they cannot predict fields, types, or pagination. LLMs cannot plan downstream operations or extract data reliably.
Parameter constraints missing. timeout_sec lacks min/max bounds (0 = unlimited, but no upper limit stated). cwd has no validation guidance (absolute vs relative, existence check?). LLMs may pass invalid values.
No error recovery guidance. If all providers fail, error message lists failures but does not suggest next steps (retry? check credentials? reduce timeout?). LLM has no recovery path.
Tool annotations missing. No readOnlyHint, destructiveHint, or idempotentHint declared. Although 'consult' is read-only and safe to retry, this is not formally signaled to the MCP client.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Naming: 'consult' is a noun, not a verb. Verb-first naming (e.g., 'get_expert_opinion', 'query_models') aligns with LLM conventions and makes intent clearer at a glance.