Loading ToolBench
Local-first control plane for evidence-linked decision memory, governed tasks, deterministic evals, and human-approved actions.
Server has well-structured tool definitions with complete input schemas and descriptions for all 6 tools. Naming follows verb_noun convention (memory_search, memory_get, plan_generate, governance_review, eval_report, capture_run). Input schemas are properly typed with required fields and constraints (e.g., limit 1-50). Output schemas are documented with oneOf patterns for success/error cases. However, parameter descriptions are minimal (10-30 chars), lacking context on WHEN to use each tool or dependencies between them. Error handling is present but generic. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Three workflow tools (plan_generate, governance_review, eval_report) have nearly identical signatures and purposes, unclear when to call which one.
Run the full deterministic workflow on capture input and return the complete result.
Run the workflow on capture input and return the evaluation result.
Run the workflow on capture input and return the governance gate result.
Fetch a single durable memory run by its stable runId.
Search durable memory runs by keyword across title, summary, entities, and memory text.
Run the deterministic workflow on capture input and return the governed execution plan projection.
Three workflow tools (plan_generate, governance_review, eval_report) accept identical input schemas and lack descriptions distinguishing when to call each. LLMs cannot determine which tool to select without explicit guidance on differences.
Parameter descriptions are minimal (10-30 chars). 'Source meeting/project text to capture' for 'text' param lacks context on expected format, length, or content structure. Descriptions should be 50-200 chars with format hints.
No tool annotations present. All tools are read-only but lack readOnlyHint annotation. capture_run and governance_review appear to have side effects (workflow execution) but lack destructiveHint or idempotentHint.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2025-06-18+ | v2 |
Error handling returns generic ExpectedInputError codes without recovery guidance. 'invalid_limit' error does not suggest next steps. Per pattern:recovery-guide, errors should include actionable recovery hints.
capture_run, plan_generate, governance_review, and eval_report all accept optional 'title', 'source', 'capturedAt', 'participants' fields but lack descriptions explaining their purpose or impact on workflow behavior.