Loading ToolBench
Two tools with significant quality gaps. 'echo' is a spike/throwaway tool with minimal description (34 chars, below 50-char baseline). 'run_workflow' has a lengthy, complex description (400+ chars) that reads like internal documentation rather than LLM-optimized guidance. Neither tool has documented output schemas. The 'run_workflow' tool accepts additionalProperties in args, which invites hallucinated parameters. No error handling guidance, no parameter constraints (enums), and no recovery paths documented. Naming is acceptable (verb-noun), but descriptions lack the clarity needed for reliable LLM tool selection.
Echo back the provided text (spike smoke tool).
Run a saved workflow from this repository's `.claude/workflows` catalog, by name, on DSH. The workflow executes on the first-party workflow engine with the catalog script as its body — the script never passes through your context, so a 100 KB workflow costs the same as a small one. Use it when the user asks to run a saved workflow (council, docs-audit, wish, workfly, research-sweep, skill-intake, skill-audit-sweep, observability-review, pm-ledger-verify), or names one of those directly. Everything the workflow needs arrives through `args`, so read the workflow's own `whenToUse` and args contract before calling (the catalog panel shows it, or read the file). The full return value is written to a journal file and the response carries a bounded projection plus that path.
No documented output schemas for either tool. LLMs cannot plan downstream calls or extract required fields without knowing what the response contains.
'echo' description is 34 characters (below 50-char minimum). Too terse to guide LLM selection; reads as a spike/test tool, not production-ready.
'run_workflow' args parameter accepts additionalProperties: true, inviting hallucinated workflow arguments. Should enumerate valid args per workflow or document strict schema.
No error handling guidance. If a workflow fails or a tool name is invalid, LLM receives no recovery path (e.g., 'Try list_workflows() first' or 'Check workflow name against catalog').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
'run_workflow' description mixes internal implementation details ('journal file', 'DSH', 'first-party workflow engine') with user-facing guidance. Unclear when to call vs. alternatives.