Loading ToolBench
A governance kernel implementation for constrained autonomous systems. Provides policy evaluation, intent management, world state inspection, reconciliation control, and governance decision-making through a REST API and subprocess service boundary.
GAP Kernel exposes 12 tools via FastAPI HTTP endpoints with basic JSON schemas. However, critical quality gaps undermine production readiness: most tool descriptions are extremely terse (10-30 chars, well below the 50-200 char baseline for LLM optimization); parameter descriptions are minimal or absent; output schemas are not documented; error handling provides no recovery guidance; and tool naming lacks consistent verb-noun patterns. The governance domain is sophisticated, but the tool interface does not meet production standards for agent interaction. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications in metadata.
Retrieve audit status and signed lineage records
Create a new intent with objectives and constraints
Evaluate a strategy proposal against governance constraints
Execute a stored request with optional approval
Retrieve a stored request by ID
Health check endpoint
Ingest entity state into the world model
Tool descriptions are critically short (10-35 chars). LLM-optimized descriptions should be 50-200 chars explaining WHAT, WHEN, and WHY. Current descriptions like 'Health check endpoint' and 'Retrieve audit status' lack context for agent selection.
Output schemas are not documented. Agents cannot plan downstream calls or extract required fields (e.g., does 'propose' return a proposal_id? Does 'evaluate' return a decision or a list of constraints?). This forces agents to guess and wastes tokens on error recovery.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk metadata (READ_ONLY vs WRITE). Agents cannot distinguish safe tools from destructive ones without reading descriptions. This violates current MCP spec (2026-07-28) tool annotation pattern.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | <=2025-11-25 | v2 |
Submit a proposal for governance evaluation and authorization
Request reauthorization for a stored request
Resolve an escalated governance decision
Submit a proposal review
Trigger a reconciliation cycle
Parameter descriptions are minimal or absent. 'approval' in execute is described as 'Optional human approval signature and metadata' but does not explain format, required fields, or how to construct it. 'constraints' in create_intent lacks any description of structure or semantics.
No error handling guidance. Tools do not document what errors are retryable, user-fixable, or fatal. No recovery hints (e.g., 'If proposal rejected, call reauthorize() with updated constraints'). Agents cannot self-correct on failures.