Cerebellum has 4 tools with clear naming and basic descriptions, but critical gaps in schema documentation, parameter descriptions, and error handling. The server provides minimal context about what parameters actually do and how tools compose. No visible input/output schema documentation in the source beyond basic type hints. Parameters lack range constraints, enum definitions, and detailed guidance. Error handling is absent, no recovery guides or categorized error responses visible. The tool set is coherent (semantic_search, list_recent, stats, capture) and verb-noun naming is correct, but parameter-level documentation and schema rigor fall well short of production baselines.

• Add JSON Schema output declarations for all four tools. For semantic_search, document: { results: [{ id, similarity_score, content, created_at }], total_matched }. For list_recent: { thoughts: [{ id, content, type, created_at }], total_count }. For capture: { success, thought_id, message }. For stats: { total_thoughts, breakdown_by_type, top_topics, top_people }.
• Add per-parameter descriptions with constraints: semantic_search.limit (integer, 1 - 100, default 10); list_recent.days (integer, 1 - 365, default 7) and list_recent.limit (integer, 1 - 100, default 20); capture.type (enum: observation|task|idea|reference|people|preference|veto, optional, defaults to 'observation').
• Document the happy path and error cases for each tool. E.g., capture should state: 'Returns thought_id on success. Returns error with 'invalid_type' if type is not one of the 7 allowed values. Returns 'storage_error' if DB write fails, retry after 5 seconds or check system logs.'
• Add a 'Composition Guide' section in server instructions explaining which tools chain together. E.g., 'Use semantic_search to find relevant past thoughts (returns thought_id). Use capture to save a new insight. Use stats to understand your thinking patterns across all captures.'
• Define what fields semantic_search and list_recent return, and ensure capture can accept at least one ID type (e.g., related_thought_id) to link new captures to old ones. Document this in the capture parameter list.
• Add min/max constraints to all numeric parameters: query/content string lengths; days range (1 - 365); limits (1 - 100); ensure defaults are safe (e.g., limit=10, not limit=10000).