Loading ToolBench
Type a GitHub handle. See who is waiting on you, what will hurt later, and what to do first.
Single tool 'standup' has a well-written description (194 chars, within baseline 34-392 range) explaining WHAT it does, WHEN to use it, and what it returns. Input schema is present with one parameter 'target' (string type, described). However, output schema is NOT documented, the code returns a dict with 'text' field added to model_dump(), but LLMs cannot see what fields to expect. No parameter constraints (enum, pattern, min/max). Error handling is minimal, exceptions are caught at CLI level but tool itself provides no recovery guidance. No tool annotations (readOnlyHint present in metadata but not in schema). Composition is sound (single responsibility). Security is good (read-only, no secrets exposed).
Triage someone's software projects and say what to do first. Call this when a person asks what to work on, what they left unfinished, who is waiting on them, or wants a standup across their repositories. `target` is a public GitHub handle or organisation (e.g. "hyperdrift-io"), or a local folder path. Read-only: it never writes to GitHub or the disk beyond a small cache. Returns up to three items with evidence, action and minutes, plus what it looked at.
Output schema not documented. Code returns dict with 'text' field and model_dump() contents, but LLMs cannot infer structure. Downstream tools cannot plan chaining.
No parameter constraints. 'target' accepts any string; no validation that it is a valid GitHub handle or local path. LLMs may pass invalid values without guidance.
Error handling lacks recovery guidance. Tool catches exceptions at CLI level but MCP tool itself does not return actionable error messages. LLMs cannot self-correct.
Tool annotations missing from schema. readOnlyHint is declared in metadata but not exposed in tool input schema for LLM visibility.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |