Loading ToolBench
Local WebSocket server for AgenBuild AI Code Assistant to enable real-time file syncing with security controls for path traversal prevention.
The server implements a single destructive tool (file-update) with a complete input schema but significant gaps in description quality, parameter documentation, error handling, and safety measures. The tool schema is present and properly typed, but lacks guidance on when/why to use it, expected behavior, and recovery paths. No output schema is documented. The parameter description mentions path correction and traversal prevention but lacks actionable error guidance for the LLM. Critical security and composition concerns are unaddressed in the tool definition itself.
Receives file updates from the AI and writes them to the local project directory. Supports file creation, update, and deletion with path traversal prevention and intelligent path correction.
Destructive operation lacks confirmation/dry-run pattern. Tool can delete arbitrary files with no safety gate or preview capability. LLMs can trigger permanent data loss without confirmation.
No output schema documented. LLM cannot reason about what the tool returns (success indicators, error details, affected file list). Breaks downstream tool composition and error recovery.
Error responses are implementation-level console logs, not LLM-actionable guidance. When a path traversal attempt fails, the LLM sees no structured response explaining why or what to do. 'Denied attempt to access a file outside project directory' needs recovery guidance.
Tool name 'file-update' is ambiguous. Does it create, update, or delete? Name should be 'batch_write_files' or 'update_and_delete_files' to clarify multi-responsibility. Current name invites misuse.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 8 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 29 | - | v1 |
Tool description (193 chars) mentions path correction and traversal prevention but does not explain WHEN the LLM should use this tool, what success looks like, or failure modes. Missing: 'Returns object with per-file success/failure indicators.'
Parameter 'files' lacks sub-property descriptions. LLM cannot distinguish between valid string content and null for deletion without reading implementation. Schema should include inline examples or description: 'Object where keys are file paths (string) and values are file content (string) or null to delete.'
No rate limiting or batch size constraints documented. LLM could pass thousands of files in one call, overwhelming filesystem or timing out. Tool should enforce a max file count per request.
No idempotence guarantee documented. If the LLM retries on ambiguous failure, will duplicate file writes cause corruption? Tool description should state: 'Idempotent: retrying the same request with same files/content is safe.'
Security annotation missing. Tool is marked DESTRUCTIVE but MCP toolAnnotations feature is not implemented. Should include 'destructiveHint: true' in tool registration to signal risk to downstream clients.