Loading ToolBench
Hybrid recall system for Claude that exposes session search, symbol grounding, and code/database schema indexing via MCP tools. Provides semantic + lexical search over conversation history, environment variable validation, database schema grounding, and documentation retrieval.
claude-dejavu provides 7 read-only tools with consistent naming (all prefixed 'dejavu_') and reasonable descriptions (80 - 200 chars). However, critical gaps undermine production readiness: (1) Input schemas are visible but lack formal type/description structure in the source code provided, parameters are documented in descriptions but not in explicit JSON Schema format; (2) descriptions are adequate but generic, lacking LLM-optimized guidance on WHEN to use each tool vs. alternatives; (3) no output schemas documented, LLMs cannot plan downstream calls or extract structured data; (4) no error handling guidance, tools return results but do not explain failure modes or recovery paths; (5) all tools are read-only (low risk) but lack idempotent/readOnly annotations. The server is functional for basic session/search workflows but falls short of production-grade tool design.
Tool failures with truncated context. Returns error logs and diagnostics from prior sessions.
On-demand full content retrieval for specific turn IDs. Returns complete turn content for a given session_id and turn_id.
Sessions that touched files matching a path substring. Returns sessions filtered by file path patterns.
Best matching session + ready-to-run resume command. Finds the most relevant prior session and returns a command to resume it.
Cheap hybrid search returning gists + IDs. Two-pass design: returns gists + IDs (~80 tok/hit). Performs hybrid Weaviate search with fallback to PostgreSQL full-text search. Scope-aware with default scope = workspace (if cwd ∈ one) else current project.
List recent sessions in current scope. Returns metadata for recent sessions filtered by scope.
No output schemas documented. LLMs cannot infer what fields each tool returns, forcing them to guess at downstream field names and risk broken tool chains.
Input schemas lack formal JSON Schema structure with explicit type and description fields. Parameters are documented in prose descriptions but not in machine-parseable schema format.
No error handling guidance. Tools do not explain failure modes (e.g., 'session not found', 'invalid scope'), recovery steps, or whether errors are retryable.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | 2026-07-28+ | v2 |
List defined workspaces. Returns all configured workspaces with their metadata.
Descriptions lack LLM-optimized guidance on WHEN to use each tool. E.g., dejavu_search vs dejavu_resume vs dejavu_sessions are all session-discovery tools but lack clear differentiation.
No tool annotations (readOnlyHint, idempotentHint). All tools are read-only and safe to retry, but this is not declared in the tool definition.