Loading ToolBench
A dispatcher that watches the htask board for ready work, brings a worker pane up in Herdr for each ready task, delivers the task's goal, tracks the worker, and stops at review.
Critical gaps across all dimensions. Tool definitions lack proper JSON Schema with type information, schemas are inferred from test code rather than explicitly registered. Descriptions exist but are domain-specific and lack LLM-optimization guidance (e.g., 'dispatch' describes Herdr workflow, not what the tool does for an agent). No parameter type definitions visible in schema registration. No error handling guidance, no output schema documentation, no security considerations. The 'stop' tool is marked CLI-only and should not be exposed over MCP at all.
Reserve a task and spawn a worker to claim and work it. The dispatcher reserves the task and returns at once; the worker claims the task itself and the dispatcher tracks it until review.
Report the dispatcher's health and configuration: version, contract, socket, state directory, config directory, base pane, worker count, tick interval, bindings, log path, verification lane, gate lane, Herdr API status, proxy status, quota, accounts, profiles, worker deaths, layout, and worker configuration.
Stream events from the dispatcher: task spawned, worker claimed, worker submitted, worker died, task finished, task failed, task parked, task released, or binding re-adopted.
Report the dispatcher's state: how many workers are live, how many tasks are reserved, and which profile each worker launched with.
Stop the dispatcher daemon. This is a CLI-only verb and not served over MCP.
No visible JSON Schema with type definitions for any tool. Schemas scored 0 across all tools, cannot verify parameter types, constraints, or output structure.
Descriptions are domain-centric (Herdr workflow terminology) rather than LLM-optimized. 'dispatch' describes task reservation and worker spawning in Herdr terms, not what an agent should do with it. Descriptions lack WHAT/WHEN/WHY structure for agent selection.
'stop' tool is marked 'CLI-only verb and not served over MCP' in its own description, yet it is listed as an MCP tool. This tool should be removed from the MCP interface entirely.
No error handling guidance. Tools return state/events but no documentation of failure modes, recovery steps, or error classification (retryable vs fatal).
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 35 | <=2025-11-25 | v2 |
No output schema documentation. 'doctor' returns 'health and configuration' but structure is undefined. 'events' streams events but event schema is not documented. LLMs cannot plan downstream actions without knowing response structure.