Loading ToolBench
Reference factory for orchestrating Claude managed agents — open-source production-grade software pipeline with file-based skill overlays for personal recipes
Single tool 'consult_advisor' has a well-structured schema with required parameters and clear descriptions. However, the tool's purpose is narrow (escalation-only), and the description, while detailed at ~280 chars, reads more as usage guidance than a functional tool description. Input schema is present with proper types and descriptions for both 'question' and 'context' parameters. No output schema is documented. The tool follows a READ_ONLY risk classification but lacks explicit error handling guidance or recovery paths for when the advisor is unavailable or rate-limited.
LAST RESORT escalation to a senior Opus advisor. Only use when ALL of these are true: (1) the decision is irreversible or expensive to undo, (2) you have already exhausted the context you have, (3) a mistake here will block downstream work. Do NOT use for routine judgment calls, style preferences, implementation choices, or to "check your work." Each call is expensive and capped per session — spend wisely.
No output schema documented. LLMs cannot plan downstream actions or extract structured data from advisor responses.
No error handling guidance. Description lacks recovery paths for rate limits, unavailability, or invalid escalation criteria.
Tool description emphasizes constraints ('do NOT use for...') rather than positive capability. Reframe as 'Use when...' to improve LLM selection.
No pagination or result limits documented. If advisor returns large context or multi-part guidance, no mechanism to handle it.
Single tool limits composition. No complementary tools for follow-up questions, clarification, or advisor feedback loops.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |