Loading ToolBench
An interactive terminal assistant that helps users complete various tasks
6 tools with mixed quality. AgentTool and ProcessStart have detailed schemas with good parameter descriptions (8-12 params each, all typed). AskUserQuestion and ProcessRead/List/Stop are simpler but well-described. All tools have descriptions (194-250 chars, within baseline 34-392). However, naming lacks verb clarity (AgentTool, AskUserQuestion are nouns; ProcessStart/List/Read/Stop are better). No output schemas documented. Error handling guidance absent. No tool annotations (readOnlyHint, destructiveHint). Risk levels assigned but not reflected in tool definitions.
Delegate tasks to subagents for autonomous execution. Supports single tasks, batch tasks, background execution, context forking, and worktree isolation.
Ask the user one or more questions with predefined options and collect their answers with optional annotations.
List all background processes, optionally including exited, failed, and stopped processes.
Read log output from a background process by process_id, with support for stream selection, tail, offset, and limit.
Start a background process with optional startup timeout, wait conditions, environment variables, and label.
Stop a background process by process_id with optional forceful termination.
No output schemas documented. LLMs cannot infer what fields to expect from tool responses, forcing unstructured parsing and increasing hallucination risk.
AgentTool and AskUserQuestion use noun-based names instead of verb_noun pattern. 'AgentTool' is vague; 'delegate_to_subagent' or 'run_subagent' would clarify intent. 'AskUserQuestion' should be 'ask_user' or 'request_user_input'.
No error handling guidance. Tools declare risk levels (WRITE, READ_ONLY, REVERSIBLE) but provide no recovery hints. E.g., ProcessStart timeout should suggest 'increase startup_timeout_ms or check wait_for conditions'.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2025-06-18+ | v2 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk levels are metadata but not exposed in tool definitions for LLM planning.
ProcessStart 'wait_for' parameter lacks format guidance. Description says 'Output substrings' but does not specify case sensitivity, regex support, or timeout behavior if none match.