The Lyceum server has severe definition quality gaps. Tool descriptions are present but vague and lack actionable context for LLM selection. Input schemas are partially visible but lack proper type definitions and parameter descriptions. No output schemas are documented. Tools like 'challenge', 'rebut', and 'compactContext' have generic descriptions that don't explain WHEN to use them or WHAT they return. The 'options' parameter appears in multiple tools but is never defined, LLMs cannot infer its structure. Error handling is absent from all tool definitions. No tool declares its risk profile or side effects clearly in the description itself (risk labels exist in metadata but not in the description text LLMs read). The server conflates multiple concerns (e.g., 'challenge' scans claims, plans, AND code edits in one tool).

• Define the 'options' parameter schema explicitly for challenge, compactContext, and compress. Specify required keys, types, and valid values (e.g., {flaw_classes: string[], max_unblocked: number}).
• Add output schemas to all tools. Example for 'challenge': {flaws: [{type: string, location: string, severity: 'low'|'medium'|'high'}], summary: string}.
• Expand descriptions to 100-200 chars with WHEN/WHY context. Example: 'Detects reasoning flaws in claims and plans. Call this BEFORE accepting an agent's conclusion to catch logical gaps, false assumptions, or incomplete reasoning.'
• Split 'challenge' into three tools: challenge_claim, challenge_plan, challenge_code. Each has one responsibility and clearer naming.
• Add error handling guidance to 'stop' and 'danger'. Example: 'If stop() times out (>1000ms), the action may have already executed, verify state manually. If danger() returns true, stop() is automatically invoked; no manual call needed.'