agent-introspection-debugginglisted

# CIEL ADAPTATION: Agent Introspection Debugging (Self-Diagnosis) This skill formalizes the "Self-Debug" protocol for CIEL sub-agents. It provides a systematic 4-phase loop to diagnose and recover from repetitive failures, tool-call loops, and reasoning drift before escalating to a human. ## The 4-Phase Loop ### 1. Failure Capture Before retrying, the agent MUST record the failure state in a structured format: - **Session Goal**: What was the intended outcome? - **Last Meaningful Tool**: Which tool call preceded the failure? - **Error Signature**: Exact error message or pattern seen. - **Environment State**: cwd, branch, and any relevant system assumptions. ### 2. Root-Cause Diagnosis Match the capture against common failure patterns: - **Looping**: Repeated same tool call with same inputs. - **Drift**: Reasoning has pivoted away from the original goal. - **Stale State**: Working on files or assumptions that no longer match the filesystem. - **Quota/Resource**: 429s, memory exhaustion, or context overflow. ### 3. Contained Recovery Take the smallest reversible action to change the diagnosis surface: - **Restate**: Re-initialize the task goal in a fresh context. - **Verify**: Re-read the filesystem or run a diagnostic command (e.g., `ls`, `git status`). - **Narrow**: Shrink the scope to a single failing test or file. - **Escalate**: If the failure is high-risk (e.g., persistent auth failure), stop and ask the human. ### 4. Introspection Report Log the debugging se