Keddy provides 12 well-named, read-only tools with clear descriptions and basic parameter documentation. Naming follows verb_noun convention consistently (keddy_search_*, keddy_get_*). All descriptions are present and substantive (100+ chars), exceeding the 10-char minimum. However, parameter documentation is incomplete: many optional parameters lack descriptions or constraints. Output schemas are not documented in the source, responses are formatted via helper functions (textResult, jsonResult) but the actual shape of returned data is not formally specified. No error handling guidance is visible (no recovery hints, retry classification, or validation error messages). Tool descriptions could be more specific about use cases and when to prefer one over another (e.g., when to use keddy_get_session_skeleton vs keddy_get_session). Security is sound (read-only semantics, no credentials), and composition is strong, tools chain well (search → get details → transcript). The codebase is production-grade (TypeScript, Zod for schema validation), but the tool definitions themselves lack the rigor expected of a 70+ score.

• Document output schemas for all tools. Create a TypeScript interface or Zod schema for each tool's response and include it in the tool description or a separate schema reference. Example: 'Returns {session_id: string, created_at: ISO8601, model: string | null, total_tokens: number, error_count: number, file_count: number}[]'.
• Add min/max constraints to numeric parameters. Example: 'limit (1 - 100, default 20)' instead of just 'Max results (default 20)'. Use Zod .min().max() or parameter descriptions to make bounds explicit.
• Convert free-form parameters to enums where values are known. For keddy_search_by_file, change operation from string to enum ['read', 'write', 'edit', 'all']. For keddy_get_recent_plans, change status to enum ['draft', 'active', 'completed'].
• Add error handling guidance to all tool descriptions. Document expected failure modes and recovery steps. Example: 'If session_id is invalid, returns error 'Session not found'. If you have a partial ID, call keddy_search_sessions first.'
• Explicitly guide tool selection with dependency hints. Example: keddy_get_session_skeleton description should say 'Call this first if you want a lightweight overview. If you need the full transcript or detailed segments, call keddy_get_session next.'
• Add pagination documentation. If tools support offset/limit, state the total result count in the response. Example: 'Returns {results: [...], total: number, has_more: boolean}' so LLMs know when to fetch the next page.
• Validate parameter ranges at runtime and return actionable error messages. Example: If limit > 100, return 'Limit must be 1 - 100; using 100' instead of a generic validation error. This helps LLMs self-correct.