When stuck, do NOT thrash in a polluted context. Summarize, reset, retry with fresh eyes.

Small-model failures are usually CONTEXT-POLLUTION failures: the context has filled with failed attempts, contradictory state, and dead ends, and the model can no longer find the thread. More turns in that context make it worse. The fix is a clean reset, not more attempts.

STUCK is detected TWO ways — use both:

• Across turns (Hermes-native): same error 3 times, 5+ turns with no verifiable progress, or you cannot clearly state the current code state. The native tool_loop_guardrails also trip here.
• Within a turn (mcp-watchdog — the signals the turn-based guardrails can't see):
  • mcp_hermes_max_watchdog_check_spiral(recent_thinking_text=...) → spiral_detected (reasoning loop)
  • mcp_hermes_max_watchdog_check_stall(...) → hung (a tool call silent past its budget)
  • mcp_hermes_max_watchdog_check_progress(task_id, signals=...) → no_progress (deltas all zero)
  • mcp_hermes_max_watchdog_check_budget(task_id, ...) → budget_exceeded Any of these firing means STUCK — act on it the same as a native trip. (If the watchdog is down, the calls just fail; fall back to the across-turns signals — graceful degradation.)

When stuck, climb the RECOVERY LADDER in order — only go to the next rung if the current one doesn't unblock you. Don't skip to escalate, and don't thrash on one rung:

1. STOP. Do not make another edit. Write a STUCK SUMMARY: (a) the goal, (b) what is verifiably TRUE right now (what works, what files exist, last green checkpoint), (c) exactly what is failing and the precise error, (d) the 2-3 approaches already tried that did NOT work.
2. incremental-verify — re-run the verifier on the last-known-good scope to confirm what is actually green vs your assumption.
3. checkpoint-revert — mcp_hermes_max_checkpoint_revert_to_last_green() to land on known-good code (it stashes any dirty tree first, nothing is lost), then mcp_hermes_max_checkpoint_restore_state(task_id) to bring back the PLAN + decision notes (so the reset is lossless — you reset CONTEXT, not your understanding of the goal).
4. replan — RESET CONTEXT: fresh attempt with ONLY the STUCK SUMMARY + reverted-clean code + restored PLAN.md. Drop the failed-attempt history — it's noise now.
5. alternative-approach — try a DIFFERENT approach than the ones in the summary.
6. decompose-further — if the subtask is too big to land green, split it smaller and retry the smallest piece.
7. fresh-subagent — hand a clean, well-scoped restatement to an isolated sub-agent (no polluted context).
8. escalate — a genuinely-hard, well-scoped subproblem → mcp_hermes_max_escalation_escalate (if enabled); a missing DECISION or a repeated wall after reset → ping the human (Telegram) with the STUCK SUMMARY and wait. This is the "loop overnight then ping me" contract.
9. abort-clean — if even escalation can't proceed, stop on the last green checkpoint with the STUCK SUMMARY recorded. A clean stop beats a broken tree.