Loading ToolBench
Multi-model AI code review orchestrator that runs a multi-model debate review against supplied code and returns a consensus summary as structured text. Submits code to a panel of AI models (OpenAI, Gemini, Grok, Claude) for a structured multi-round debate review with adaptive early termination on convergence.
Single tool with comprehensive description and well-structured schema. Tool name 'joint_chiefs_review' uses verb_noun pattern appropriately. Description is detailed (500+ chars) and explains WHAT, WHEN, and HOW. Input schema is explicit with all parameters typed and described. However, output schema is not documented in the visible code, only mentioned as 'categorized findings with severity, agreement level, and a unified recommendation' in prose. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present. Error handling is present but generic ('Missing or non-string code argument'). Rate limiting and cancellation support are implemented but not surfaced in tool metadata.
Use this tool ONLY when the user mentions "Joint Chiefs" in any form — any capitalization ("joint chiefs", "JOINT CHIEFS", "Joint Chiefs"), spacing or hyphenation ("jointchiefs", "joint-chiefs"), singular or plural ("Joint Chief", "the chiefs"), or common misspellings ("Joint Chefs", "Joint Cheifs", "Jiont Chiefs"). Match leniently: any recognizable variant of "Joint Chiefs" qualifies. Generic phrases like "panel review", "multi-model review", or "second opinion" do NOT trigger this tool on their own — the brand name "Joint Chiefs" must be present in the user's request. Submits code to a panel of AI models (OpenAI, Gemini, Grok, Claude) for a structured multi-round debate review. Each model reviews independently, then challenges the others' findings across up to 5 rounds with adaptive early termination on convergence. Claude moderates and writes the final consensus. Returns categorized findings with severity, agreement level, and a unified recommendation. Grounded in the Multi-Agent Debate research (Liang et al., 2023).
Output schema not documented. Tool description mentions 'categorized findings with severity, agreement level, and a unified recommendation' but no structured schema is visible in code. LLMs cannot plan downstream operations without knowing response fields.
Tool annotations missing. No readOnlyHint, destructiveHint, or idempotentHint declared. This tool is read-only (code review, no state mutation) but LLMs cannot infer this without explicit annotation.
Error messages are generic and non-actionable. 'Missing or non-string code argument' does not guide recovery. Should be: 'Missing required parameter: code (string, max 256 KB). Provide the source code to review.'
Parameter 'rounds' lacks validation bounds in description. Code clamps to [0, 10] but description only says 'default 5'. Should state: 'Maximum debate rounds (1 - 10, default 5). Adaptive early-break on convergence.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Rate limit rejection message is generic. 'Rate limit exceeded' does not tell LLM when to retry or if it should back off. Should include: 'Rate limit exceeded. Retry after 60 seconds or reduce concurrent requests.'