Loading ToolBench
A Swift-based agent framework with tool execution, capability management, and human-in-the-loop approval workflows
Aria defines 2 tools with partial schema coverage. coreai_aria_proof has a complete input schema and description, but propose_action, despite being a critical HITL (human-in-the-loop) tool, lacks output schema documentation and has a vague description that doesn't explain the approval workflow or what happens after the proposal. Neither tool has error handling guidance, recovery paths, or actionable error messages. The propose_action tool is particularly concerning: it accepts a free-form payloadJSON string parameter with no validation hints, format constraints, or examples of valid JSON structures. Descriptions are present but generic (194 chars baseline; these average ~120 chars). No tool declares permissions, audit requirements, or idempotency guarantees. Tool composition is minimal (only 2 tools), limiting agent capability.
Returns the deterministic Core AI through Aria proof marker.
Propose an action for the user to approve BEFORE it happens (e.g. sending an email, booking an event). You cannot perform these actions directly — this is the only way. After calling propose_action, STOP and wait for the user's decision; never claim the action was taken. Provide the action kind, a short title, optional detail to review, and the action arguments as a JSON object string in payloadJSON.
propose_action lacks output schema and approval workflow documentation. Description does not explain what the agent should do after calling the tool (wait for user decision, poll for result, etc.) or what the response structure is. This is a critical HITL tool and requires explicit guidance.
payloadJSON parameter in propose_action accepts free-form JSON string with no validation hints, format constraints, or examples. Description says 'JSON object string' but does not specify required/optional keys, data types, or valid structures. LLMs will hallucinate invalid JSON.
No error handling guidance in either tool. Descriptions do not explain failure modes, recovery paths, or what the LLM should do if a call fails. E.g., what happens if propose_action is called with invalid JSON in payloadJSON?
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2025-06-18+ | v2 |
No permission declarations or audit trail guidance. propose_action is a side-effecting tool (WRITE risk) but does not document required scopes, who can approve actions, or how calls are logged for compliance.
coreai_aria_proof description is vague ('Returns the deterministic Core AI through Aria proof marker'). Does not explain WHEN to call it, what 'proof marker' means, or what the agent should do with the returned marker. Lacks context for LLM tool selection.