Loading ToolBench
An AI-powered proactive assistant that monitors emails and calendar events, processes them with LLM analysis, and suggests actionable next steps. Integrates with Google Gmail and Calendar via Composio MCP servers.
Single tool with critical definition gaps. COMPOSIO_INITIATE_CONNECTION has a basic description but lacks depth on when/why to use it, prerequisites, and return value structure. Input schema present but minimal (only 'tool' param with generic string type, no enum of valid apps). No output schema documented. Description is 156 chars (within range) but does not explain what happens after redirect, error cases, or idempotency. Tool name uses verb but is overly long (27 chars). No parameter validation guidance. Error handling in code is ad-hoc (JSON parsing fallback, basic string extraction) with no structured error responses to guide LLM recovery.
Initiates OAuth connection with a Composio-integrated service (e.g., Gmail, Google Calendar). Returns a redirect URL for user authentication.
No output schema documented. LLM cannot predict response structure (redirect_url, error format, success indicator). Code shows ad-hoc JSON parsing with fallback string extraction, indicating response is unstructured.
Input parameter 'tool' lacks enum constraint. Should enumerate valid Composio apps (gmail, google_calendar, etc.) instead of free-form string. LLM will hallucinate invalid app names.
Description does not explain return value, prerequisites (Composio API key?), or what LLM should do with redirect_url. Missing: 'When to use' context, idempotency guarantee, error recovery guidance.
Parameter 'tool' description is generic ('The name of the app/service'). Does not specify format, valid values, or examples. LLM lacks actionable guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2026-07-28+ | v2 |
Error handling is unstructured. Code attempts JSON parsing, falls back to substring search for 'https://backend.composio.dev/api/v3/s/', and prints to stdout. No structured error response to guide LLM on retry, user action, or fatal conditions.