Never get stuck INSIDE a turn. The turn-based guardrails don't see spirals or hangs.

Hermes' native tool_loop_guardrails only fire ACROSS turns. The two ways work dies are inside one turn: a reasoning spiral (you think in circles and never emit) and a poll-hang (one tool call blocks forever on a server that never exits). The mcp-watchdog server gives you deterministic, model-free checks for exactly these. Use them — they are cheap and they don't call the model.

Per-turn discipline

• Reason in ≤3-4 sentences, then ACT. Execution turns are not for deep thinking — that is what caused the spiral. (High reasoning effort is reserved for planning; see workflow-effort-routing.)
• If a turn's reasoning is getting long or circular, call mcp_hermes_max_watchdog_check_spiral(recent_thinking_text=<your last reasoning>). If spiral_detected is true: STOP thinking, write a STUCK SUMMARY, and go to the recovery ladder (workflow-stuck-detect-reset) — revert + replan. Do NOT push through with more reasoning.

Long-running operations (the poll-hang fix)

• NEVER block on a process that may not return. Start it backgrounded, then call mcp_hermes_max_watchdog_check_stall(tool_name, elapsed_s, expecting_heartbeat, last_heartbeat_age_s) once.
  • hung: true → it is silent past its budget; kill it and try another approach.
  • waiting: true → it is heartbeating/serving; that is SUCCESS, leave it running, move on.
• Do this exactly once per process. Never poll-loop. (See workflow-long-running-processes.)

Progress & budget (every few subtasks)

• Call mcp_hermes_max_watchdog_check_progress(task_id, signals={files_touched, tests_passing, checkpoints, turn}) every few subtasks. If no_progress is true (deltas all zero over N), you are stalling — go to the recovery ladder instead of grinding.
• At task start, call mcp_hermes_max_watchdog_start_task_budget(task_id, wall_clock_s, max_turns, usd_cap). Periodically check_budget(task_id, turns_used, usd_spent). On budget_exceeded: checkpoint cleanly, write a STUCK SUMMARY, and STOP — do not run over the budget.

On ANY watchdog flag (spiral / hung / no_progress / budget_exceeded)

1. Stop. Do not make another edit.
2. Snapshot then revert: mcp_hermes_max_checkpoint_snapshot_state(task_id, ...) is for the GOOD state; when stuck call mcp_hermes_max_checkpoint_revert_to_last_green() to land on known-good code, then restore_state(task_id) to bring the PLAN back.
3. Replan from the STUCK SUMMARY + clean tree (the recovery ladder in workflow-stuck-detect-reset).

Graceful degradation

If the watchdog server is down, its calls fail — that is fine. Skip the self-check and fall back to Hermes' native turn-based guardrails + the ≤3-4-sentence rule. Never let a watchdog outage block you.