Loading ToolBench
A Stack-Overflow-style knowledge base for Claude Code debugging sessions. Captures problem spans from Claude Code tool use, drafts solutions via LLM, stores entries in SQLite with vector embeddings, and exposes search/retrieval tools via MCP.
Three tools with clear naming (search, get, feedback) and reasonable descriptions (100-150 chars). All have input schemas with type definitions. However, parameter descriptions are sparse or missing entirely. The 'limit' parameter in search lacks a description. The 'notes' parameter in feedback lacks a description. Output schemas are not documented in the visible code. Error handling guidance is absent, no recovery hints or categorization. The tools follow basic structure but lack the polish expected of production-grade agent tools.
Record whether a Remembug entry helped resolve the current problem. Increments the confirmation count when helpful=true.
Fetch the full text of a single Remembug entry by id.
Search the local Remembug knowledge base for entries matching a natural-language query. Use this BEFORE reasoning from scratch about a new error — there may already be a captured solution.
Parameter descriptions missing or incomplete. 'limit' in search has no description. 'notes' in feedback has no description. 'entry_id' in get and feedback lack context (is this a UUID, integer, or string format?). LLMs cannot infer parameter semantics from names alone.
Output schemas not documented. The visible code shows input schemas but no return type documentation. LLMs need to know what fields to expect (e.g., does search return entry_id, title, snippet, relevance_score?) to plan downstream calls and extract data correctly.
No error handling guidance. Tools lack recovery hints (e.g., 'If entry not found, try search() with a broader query'). No error categorization (retryable vs. user-fixable vs. fatal). Agents cannot determine next steps on failure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Parameter 'limit' in search lacks constraints. No minimum/maximum specified. Unbounded integers let LLMs pass absurd values (e.g., limit=999999) that could break performance or timeout.
Tool descriptions lack WHEN-to-use guidance. 'Search the local Remembug knowledge base' tells WHAT but not WHEN. Should state: 'Call this first when debugging a new error to check if a solution already exists.' This helps LLMs prioritize tool selection.