Brain server has 8 tools with descriptions and basic parameter schemas, but significant gaps prevent higher scores. All tools have descriptions (good), but most parameters lack type definitions in the visible schema, only string types are declared. No output schemas are documented. Parameter descriptions are present but generic. Error handling is minimal: tools shell a binary and return text verbatim, with no guidance on recovery or error classification. The server prioritizes fidelity to engine output over LLM usability, which conflicts with agent tool design principles.

• Add JSON Schema output definitions for all tools. E.g., brain_match should document: {type: 'object', properties: {survivors: {type: 'array', items: {type: 'object', properties: {case_id, score, action, ...}}}, discard_log: {type: 'string'}, ...}}.
• Convert parameter descriptions to LLM-optimized format (50 - 150 chars). E.g., 'stage: The situation lifecycle stage (one of: initial, escalated, resolved). Omit if not stated by operator.' Include valid values inline or reference situation_vocab.
• Add enum constraints to parameters with fixed vocabularies. E.g., verdict: {type: 'string', enum: ['hit', 'miss', 'partial'], description: '...'}. action: {type: 'string', enum: [from ACTION_TYPES], description: '...'}.
• Wrap subprocess calls with error classification. Catch non-zero exit codes and stderr, then return structured errors: {error: 'invalid_input', message: '...', recovery: 'Call situation_vocab() first to see valid values.', retryable: false}.
• Document result limits and pagination. E.g., 'brain_match returns up to 20 ranked survivors. If more exist, the discard log will indicate truncation. Call with additional constraints to narrow results.'