Loading ToolBench
Local-first multi-agent task orchestration engine and worktree supervisor with Waymark continuity. Provides MCP tools for task management, git worktree provisioning, and progress tracking.
Arbiter MCP defines 7 tools with complete input schemas and descriptions. Naming follows verb_noun convention (claim, checkpoint, discover, complete, fail, recover, status). Descriptions are substantive (80-180 chars), explaining WHAT and WHEN. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Error handling is present but generic (errorResult/jsonResult wrappers). No tool annotations (readOnlyHint/destructiveHint) despite clear WRITE vs READ_ONLY risk classifications. Parameter descriptions are adequate but lack format/constraint details (e.g., 'path' accepts any string; no validation guidance). Lease epoch and worker_id patterns suggest distributed coordination, but this complexity is not surfaced in descriptions to guide agent reasoning.
Record an intermediate progress milestone and refresh the worker lease heartbeat with Waymark telemetry.
Claim the next available READY task from Arbiter's DAG and provision an isolated worktree with Waymark.
Complete an active task, seal its Waymark trajectory, commit worktree changes, and unblock downstream DAG tasks.
Discover syntax symbols in a file inside the caller's active leased worktree.
Mark a task as failed and abandon its active Waymark trajectory.
Inspect or recover an orphaned Waymark lock in a task's worktree.
Query task queue status or details of a specific task.
Output schemas not documented. LLMs cannot predict response structure (e.g., arbiter_claim_task returns task_id, worktree_path, lease_epoch, but this is inferred from code, not declared in tool definition). Breaks downstream tool chaining and forces agents to guess field names.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications (WRITE vs READ_ONLY). Agents cannot distinguish safe-to-retry operations from destructive ones without explicit hints.
Parameter descriptions lack format/constraint guidance. 'path' accepts any string with no validation rules stated. 'lease_epoch' is described as 'monotonic' but no guidance on what happens if stale. Agents cannot self-correct invalid inputs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Error handling is generic (errorResult wrapper). No recovery guidance. If arbiter_claim_task fails with 'No ready tasks', agent does not know whether to retry, wait, or escalate. No error classification (retryable vs fatal).
arbiter_recover_lock description is vague ('Inspect or recover an orphaned Waymark lock'). Does not explain when to call it, what 'orphaned' means, or what 'force' does. Agents cannot determine if this is a recovery tool or a diagnostic tool.