Loading ToolBench
Minimal single agent demo with basic file tools and MCP support
Mini Agent provides 7 tools with complete input schemas and descriptions, but quality is uneven. bash, read_file, write_file, and edit_file have detailed, actionable descriptions (100-200 chars) with proper parameter types and constraints. However, record_note, recall_notes, and get_skill lack depth: descriptions are generic, parameter guidance is minimal, and output schemas are undocumented. No tool declares error recovery guidance or categorizes errors as retryable/fatal. Tool composition is reasonable (file ops, bash, notes, skills) but lacks idempotency guarantees and chaining metadata. Naming follows verb_noun convention well (read_file, write_file, edit_file, bash). Risk annotations (IRREVERSIBLE, READ_ONLY, WRITE) are present but not integrated into parameter validation or error handling.
Execute shell commands in foreground or background. For terminal operations like git, npm, docker, etc. DO NOT use for file operations - use specialized tools. Parameters: - command (required): PowerShell command to execute - timeout (optional): Timeout in seconds (default: 120, max: 600) for foreground commands - run_in_background (optional): Set true for long-running commands (servers, etc.) Tips: - Quote file paths with spaces: cd "My Documents" - Chain dependent commands with semicolon: git add . ; git commit -m "msg" - Use absolute paths instead of cd when possible - For background commands, monitor with bash_output and terminate with bash_kill Examples: - git status - npm test - python -m http.server 8080 (with run_in_background=true)
Perform exact string replacement in a file. The old_str must match exactly and appear uniquely in the file, otherwise the operation will fail. You must read the file first before editing. Preserve exact indentation from the source.
Get complete content and guidance for a specified skill, used for executing specific types of tasks
Read file contents from the filesystem. Output always includes line numbers in format 'LINE_NUMBER|LINE_CONTENT' (1-indexed). Supports reading partial content by specifying line offset and limit for large files. You can call this tool multiple times in parallel to read different files simultaneously.
Recall all previously recorded session notes. Use this to retrieve important information, context, or decisions from earlier in the session or previous agent execution chains.
Output schemas undocumented. Tools return ToolResult objects but LLMs cannot see what fields to expect (success, content, error). Downstream tool chaining requires explicit field documentation.
No error recovery guidance. Tools lack descriptions of failure modes, retryability, or next steps. E.g., bash timeout or file not found errors do not guide the LLM on recovery.
record_note and recall_notes lack actionable descriptions. 'Record important information' and 'Recall all previously recorded session notes' are vague. When should the LLM use these vs. inline context? What is the retention scope?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Record important information as session notes for future reference. Use this to record key facts, user preferences, decisions, or context that should be recalled later in the agent execution chain. Each note is timestamped.
Write content to a file. Will overwrite existing files completely. For existing files, you should read the file first using read_file. Prefer editing existing files over creating new ones unless explicitly needed.
get_skill description references 'list_skills' tool that does not exist in the tool set. This breaks discoverability and forces the LLM to guess skill names.
bash tool accepts 'command' parameter but description says 'PowerShell command', inconsistent with typical bash/shell semantics. Unclear if this is bash, sh, or PowerShell.