Loading ToolBench
Let one coding-agent CLI consult another through Codex, Claude Code, OpenCode, or Antigravity.
17 tools with mostly complete schemas and descriptions, but significant gaps in parameter documentation and error handling guidance. Tool names follow verb_noun convention well (orchestrator_consult, orchestrator_review, orchestrator_workflow_*). Descriptions range 50-150 chars, meeting baseline minimums. However, many parameters lack detailed constraints (e.g., 'agent' param accepts string but no enum of valid agents despite orchestrator_list_consult_agents existing). No output schemas documented. Error handling is minimal, no recovery guidance, no categorization of retryable vs fatal errors. Confirmation tokens appear in destructive operations but lack explanation of how to obtain them. Missing: pagination guidance for list tools, field-level descriptions for complex objects, and actionable error messages.
Cancel an in-flight review before it completes.
Consult another vendor's coding agent -- Codex, Claude Code, OpenCode, or experimental Antigravity -- running under its own account.
Delete a consultation and all its history.
Delete a review and all its findings.
Delete a workflow and all its history.
Combine findings from all reviewers into a final report.
Get details of a specific review.
Parameter 'agent' in orchestrator_consult and orchestrator_workflow_run_step accepts free-form string with no enum constraint. Should reference valid agents from orchestrator_list_consult_agents or declare allowed values explicitly.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., consultation_id, review_id, workflow_id) without knowing response structure.
Confirmation tokens (confirm_token) appear in destructive operations but lack explanation of how to obtain them or what they represent. Parameter description should clarify the approval workflow.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
List the agents available for consultation, with their capabilities and models.
List all reviews, optionally filtered by status.
Plan a review of work by external agents without sending anything yet.
Send a planned review to external agents after user approval.
Cancel an in-flight workflow.
Plan the next step in a workflow without executing it.
Record a step that the host (user) executed themselves.
Run a planned workflow step through an agent.
Start a workflow: research, implementation, and review phases with the host deciding when to advance.
Get the current status of a workflow and what actions are allowed next.
List tools (orchestrator_list_reviews, orchestrator_list_consult_agents) lack pagination parameters (limit, offset/cursor). No guidance on result limits or how to handle large datasets.
No error handling guidance in tool descriptions. LLMs cannot determine if errors are retryable, user-fixable, or fatal. Missing recovery suggestions (e.g., 'If agent not found, call orchestrator_list_consult_agents').