agent-debuglisted

# Agent Debug Diagnose a failing or stuck agent run through capture, root-cause classification, contained recovery, and a structured introspection report — instead of retrying blindly. ## Usage ``` /agent-debug # Diagnose the current stuck/failing run /agent-debug --report-only # Skip recovery, just produce the introspection report ``` ## Behavior ### Phase 1: Capture Record the failure before touching anything: - error type, message, last tool call sequence - what the agent was trying to accomplish - signs of context pressure: duplicated plans, oversized pasted output, repeated near-identical prompts - environment assumptions to verify: cwd, branch, expected files, service state ### Phase 2: Diagnose Match the failure to a known pattern before changing anything: | Symptom | Likely cause | Check | | --- | --- | --- | | Same tool call repeated | loop with no exit condition | last N tool calls for repetition | | Reasoning degrading, output rambling | context overflow from duplicated notes/logs | recent context for low-signal bulk | | Connection refused / timeout | wrong port, service down | service health, URL, port | | 429 / rate limit | retry storm, no backoff | call frequency and spacing | | File missing after write | wrong cwd, race, branch drift | actual path, `git status`, file existence | | "Fixed" but still failing | wrong hypothesis | isolate the exact failing case, re-derive | Ask: is this logic, state, environment, or policy failure? De