Loading ToolBench
A background observer for Claude Code sessions that provides read-only tools for workspace analysis, file searching, and command execution within a sandboxed environment
Four tools with complete schemas and descriptions, but significant gaps in parameter documentation and error guidance. Read, Grep, and Glob have clear, concise descriptions (40-60 chars) and proper JSON schemas with required fields marked. However, most parameters lack descriptions: Grep's 'pattern' and 'path' are undescribed; Glob's 'pattern' and 'path' lack guidance on format/constraints. Run tool has an empty enum (no allowlisted commands visible), making it non-functional. No output schemas documented. Error handling is minimal, tools return generic 'error: <message>' without recovery guidance. Naming is clear and verb-based (Read, Grep, Glob, Run), following conventions. No parameter validation rules, ranges, or format constraints stated in descriptions.
Find files by glob pattern in the observed workspace.
Search file contents in the observed workspace with a regular expression.
Read a file from the observed workspace. Read-only.
Run one of the allowlisted commands, in the workspace root. The command must be given exactly as listed, and only commands granted to the calling detector will execute.
Run tool has empty enum, no allowlisted commands visible in schema, making the tool non-functional. The enum array is always empty regardless of grants.
Grep and Glob parameters lack descriptions. 'pattern' parameter has no guidance on regex syntax, glob format, or constraints. 'path' parameter undescribed, unclear if it accepts absolute paths, workspace-relative paths, or both.
No output schemas documented. LLMs cannot predict what Read, Grep, or Glob return, field names, types, structure, or limits. Prevents downstream tool chaining and forces LLM guessing.
Error handling returns generic 'error: <message>' with no recovery guidance. LLM cannot determine if error is retryable, user-fixable, or fatal. No actionable next steps provided.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
Read tool's 'offset' and 'limit' parameters lack constraints. No stated minimum/maximum values, default behavior, or bounds checking guidance. LLM could pass negative or absurdly large values.