What problem does it solve?
Adapter-path ClaudeEval in runtime mode can silently fail or loop, producing missing tool results, hallucinated answers, or no final notebook output, leaving users unable to understand where the turn pipeline broke.
Core Features & Use Cases
- Symptom-driven fault isolation: Maps common runtime anomalies (tool_call text leaking, repeated web search, “過去の検索結果が見えない”, missing final text, stuck phases) to the most likely internal responsibility area.
- Turn and event trace triage: Guides users to inspect ClaudeTurnTrace for the correct lifecycle sequence and detect abnormal endings (e.g., tool loops, budget exhaustion, finalize/text skip patterns).
- ConversationState and tool-result verification: Helps verify whether tool calls/results were recorded, whether prompts include rich results, and whether the final proposal text exists even when notebook rendering fails.
- AsyncToolExec pipeline diagnosis: Provides targeted checks for polling tick registration, AsyncToolExec progress, finalize/continuation scheduling, and lingering polling-tick residues.
Quick Start
Run the skill by asking your AI to diagnose a “ClaudeRuntime adapter tool-flow malfunction” using ClaudeRateLimitStatus[], ClaudeTurnTrace[$ClaudeLastRuntimeId], and ConversationState.Messages to pinpoint the failing stage.
