Самодиагностика и восстановление AI-агента

★ 7.9 · ai-agents

agent-introspection-debugging is a Claude Code skill that implements a structured four-phase self-debugging workflow for AI agent failures: failure capture, root-cause diagnosis, contained recovery, and a human-readable introspection report. It activates when an agent is looping on the same tools, hitting maximum tool-call limits, experiencing context growth or prompt drift, or encountering a mismatch between expected and actual environment state. Rather than retrying blindly, the agent fills in Failure Capture and Recovery Action templates, matches the failure against a pattern table covering loops, context overflow, ECONNREFUSED, 429 quota exhaustion, stale diffs, and wrong hypotheses, then applies the smallest reversible corrective action. The session ends with a structured Self-Debug Report covering the failure, root-cause hypothesis, recovery taken, and evidence of outcome. It integrates with verification-loop, continuous-learning-v2, council, and workspace-surface-audit for downstream steps.