Knitbrain provides 37 well-named tools with mostly clear descriptions and complete input schemas across the board. Naming follows consistent verb_noun conventions (optimize, retrieve, search, read, classify, run, record, save, load, compose, etc.), which is excellent. All tools have descriptions ranging from 80-200+ characters, meeting the 10-1024 baseline. All input schemas visible in e2e-tools.mjs and src/mcp/tools.ts include type definitions and descriptions for parameters. However, several limitations prevent a higher score: (1) Output schemas are NOT documented, responses appear to be free-form text or JSON without published structure, forcing LLMs to parse unstructured data; (2) Error handling is minimal, tools lack recovery guidance and categorization (retryable vs fatal); (3) Many tools bundle related concerns (e.g., knitbrain_run orchestrates classify + compose + propose in one call, violating single-responsibility); (4) Composition gaps, tools like run_loop and run_workflow appear to do complex multi-step work internally rather than exposing granular steps; (5) Missing pagination support on list/search tools (e.g., brain_search, search_learnings, activity_log have no limit/offset beyond activity_log's limit param). The tool set is ambitious and coherent for a 'loop engineering' system, but lacks the precision and defensiveness of A-grade production tools.

• Document output schemas for all tools. Publish JSON Schema definitions for responses, especially for search and meter tools. Include field names, types, and example structures in tool descriptions.
• Add structured error responses. Define a standard error format: {error_code, message, recovery_hint, retryable: boolean}. Provide actionable guidance (e.g., 'User not found. Try search_users() first.').
• Split knitbrain_run into granular tools: classify_task, compose_workflow, propose_agents (three separate calls, composable). Agents can then reuse classify without running the full pipeline.
• Add pagination to all search/list tools. Accept 'limit' (default 20, max 100) and 'cursor' or 'offset' parameters. Return total_count and next_cursor in responses.
• Clarify prerequisite semantics. Document what 'prior classification' means: is it stored in session state? Scoped to the request? Provide an example of how agents should chain calls (e.g., classify_task → record_learning → compose_workflow).
• Add parameter range constraints. For numeric params (max_iters, wall_clock_s, limit), state min/max in descriptions (e.g., 'max_iters: 1-100').
• Return chaining IDs in all responses. If a tool call should be followed by another, include the relevant IDs in the response (e.g., compose_workflow should return a workflow_id for load_workflow).
• Add timeout handling to long-running tools. knitbrain_run and knitbrain_run_loop should document timeout behavior and return partial progress on deadline.