Loading ToolBench
Governance-first, multi-agent operational mediation layer for clinicians (Google ADK + Gemini). Structures and triages inbound clinical communication before it reaches the clinician.
TensorNode defines 5 tools with clear, governance-focused descriptions emphasizing read-only/proposal-only semantics. All tools have input schemas with typed parameters and descriptions. However, output schemas are not explicitly documented in the visible code, parameter descriptions lack constraint details (ranges, enums, formats), and error handling guidance is absent. Tool names use 'draft_' prefix (non-standard verb_noun pattern) and descriptions are concise but could better explain when/why to use each tool vs. alternatives. No tool annotations (readOnlyHint, etc.) present despite all being read-only.
List free appointment slots. Read-only — books nothing.
Fetch non-diagnostic continuity context for a patient. Read-only.
Look up vetted, non-diagnostic FAQ/reassurance content. Read-only.
Compose reply/notification text. Returns a DRAFT — sends nothing.
Propose a slot. PROPOSAL ONLY — does not book the appointment.
Output schemas not documented. Tools return Pydantic models (FaqEntry, PatientContext, AvailabilitySlot, SlotProposal) but no schema definitions visible in tool registration. LLMs cannot plan downstream calls without knowing return structure.
Parameter constraints missing. 'max_results' and 'limit' lack min/max bounds; 'days' has default=7 but no range; 'preferred_hour' (0-23) documented in description but not as schema constraint; 'from_date' format not enforced. Unbounded or under-constrained params invite invalid LLM inputs.
Tool naming uses 'draft_' prefix instead of standard verb_noun pattern (get_, search_, list_, propose_). Non-standard prefix obscures intent and conflicts with Arcade naming baseline (90% of A+ tools start with action verb). Rename to search_faq, get_patient_context, list_calendar_slots, propose_appointment, compose_notification.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | <=2025-11-25 | v2 |
No error handling guidance. Tools lack recovery hints (e.g., 'If patient_ref not found, try search_patients() first'). No categorization of errors as retryable vs. user-fixable. Agents cannot self-correct on failures.
No tool annotations despite governance-critical semantics. All tools are read-only/proposal-only but lack readOnlyHint or idempotentHint annotations. Agents cannot infer safety properties from schema alone.