Loading ToolBench
Four tools with clear action verbs and reasonable descriptions (avg 180 chars), but critical gaps in schema completeness and parameter documentation. create_repl has empty input schema ({}), install_packages and run_code lack parameter descriptions, and no output schemas are documented. Error handling is present but generic. No tool annotations (readOnlyHint/destructiveHint). Naming is good (verb_noun pattern), but parameter-level documentation is sparse. Composition is sound, tools chain logically, but schema rigor falls short of production baseline.
Create a new REPL session. Returns a session ID that can be used to identify the REPL. Initializes a virtual environment and starts a dedicated REPL process for this session. The following packages are already available: - numpy - pandas - matplotlib - scipy - scikit-learn - seaborn - plotly
Install the given packages in the code interpreter session using pip.
Reset the REPL session: terminate kernel, delete all cells and state.
Execute code in the specified session, with optional parent cell reference if you want to continue an existing thread of execution. If you leave it blank, it will start a new session. IMPORTANT: If you want to see a matplotlib figure, you need to call `display_figure(fig)` in your code, NOT `plt.show()` or `plt.savefig()`. `display_figure()` is a custom function that has already been imported into the session by the helper process. Returns the execution results and the new cell ID.
create_repl has empty input schema ({}). No parameters documented, no output schema defined.
install_packages and run_code lack descriptions for input parameters (session_id, packages, code, parent_cell_id). LLMs cannot infer parameter meaning from names alone.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., cell_id from run_code result).
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). reset_repl is marked DESTRUCTIVE but lacks annotation; install_packages and run_code are WRITE but unmarked.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | <=2025-11-25 | v2 |
Error handling in _process_response_for_mcp converts errors to text but does not guide LLM recovery. No retryable/user-fixable/fatal classification.