Loading ToolBench
Message bus for AI agent coordination — agent-to-agent consult via claude -p
Partyline defines 5 tools with clear action verbs (consult, list, roundtable, roundtable_transcript, roundtable_sessions). All tools have descriptions (avg ~80 chars) and Zod-based input schemas with type annotations. However, descriptions are terse and lack context on WHEN to use each tool or what downstream actions they enable. Parameter descriptions are present but minimal (e.g., 'The registered agent name to consult' lacks guidance on how to discover valid agent names). Output schemas are not formally documented, responses are text-only, which forces LLMs to parse unstructured agent responses. Error handling returns plain text errors without recovery guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all tools being read-only. Schemas are well-formed but descriptions fall short of the 50-200 char LLM-optimized baseline.
Consult a named agent by spawning a claude -p session in that agent's working directory.
List all registered agents and their descriptions.
Run a round-table discussion among multiple registered agents on a topic.
List recent round-table sessions with metadata.
Retrieve the full transcript for a saved round-table session.
Output schemas not documented. All tools return plain text responses; LLMs cannot parse structured data or chain results to downstream tools. No documented return types or field names.
Descriptions lack WHEN-to-use context. 'Consult a named agent' does not explain when to use consult vs roundtable, or how to discover valid agent names. LLMs cannot disambiguate tool selection.
No tool annotations. All 5 tools are read-only but lack readOnlyHint annotation. Agents cannot infer safety/idempotency from tool metadata.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | <=2025-11-25 | v2 |
Error handling returns plain text without recovery guidance. 'Error: <message>' does not tell LLM whether to retry, ask user, or abort. No error classification.
Parameter descriptions are minimal. 'Comma-separated list of registered agent names' lacks format validation rules or examples of valid agent names. LLMs may pass invalid formats.