agent-introspection-debugging

A four-phase self-debugging loop (failure capture → root-cause diagnosis → contained recovery → introspection report) for AI agent runs that are looping, burning tokens, or drifting from the task. Provides a pattern table (tool-call loops, context overflow, `ECONNREFUSED`/429s, stale diffs, wrong hypotheses) and forces the agent to verify world state, shrink scope, and run one discriminating check before retrying.