Five tools with clear naming (verb_noun pattern: cp_start_*, cp_track_*, cp_get_*) and reasonable descriptions (80-180 chars). All tools have input schemas with Zod validation and type definitions. However, output schemas are undocumented, responses are JSON-stringified text without declared field structure. Parameter descriptions are present but generic; several lack actionable constraints (e.g., 'Custom session ID' doesn't explain format or length). No error handling guidance, no recovery paths, and no documentation of what happens on invalid inputs. Tool composition is sound (single responsibility), but output chaining is weak, cp_get_budget and cp_get_run_summary return opaque JSON strings rather than structured objects with typed fields.

• Document output schemas explicitly. Define cp_get_budget return as {runId: string, budget: {used: number, limit: number, percentUsed: number}} and cp_get_run_summary as {runId: string, sessionId: string, toolCalls: Array<{toolName: string, inputTokens: number, outputTokens: number, durationMs: number}>, totalTokens: number}. Return structured objects, not JSON strings.
• Add numeric constraints to parameters. contextLimit: 'Token limit (1000 - 1000000, default 200000)', durationMs: 'Duration in milliseconds (0 - 3600000)', baselineOverheadTokens: 'Tokens (0 - 500000)'. Use min/max in Zod schema and repeat in description.
• Add error handling documentation. Document: 'If sessionId not found, returns error 404 with message "Session not found. Call cp_start_session first." If runId invalid, returns 400 with "Invalid runId format."' Include recovery hints for each error class.
• Validate and sanitize 'arguments' and 'output' in cp_track_tool_call. Set max JSON depth (e.g., 10 levels), max size (e.g., 100KB), and reject circular references. Return clear error: 'Arguments exceed 100KB. Truncate or split into multiple calls.'
• Add pagination to cp_get_run_summary. Accept optional limit (1 - 1000, default 100) and offset parameters. Return {toolCalls: [...], total: number, hasMore: boolean}. Document: 'Large runs may return paginated results. Use offset to fetch subsequent batches.'