Problem

Built-in diagnostics stuck-session recovery only supports abort-drain (abort_embedded_run) and release_lane after stuckSessionAbortMs. Before that threshold, stalled embedded runs log recovery=none / observe_only.

For production assistants, aborting a long tool-heavy turn is often the wrong first move. Operators want a configurable ladder:

1. Steer — sessions.steer / interrupt active run with a recovery prompt
2. Nudge — sessions.send if still no user-visible progress
3. Abort — sessions.abort / abortAndDrainEmbeddedPiRun only as last resort

Today this requires a separate cron/heartbeat agent, which is fragile (model may ignore instructions, spam channels if misconfigured, etc.).

Proposed config

{
  diagnostics: {
    stuckSessionWarnMs: 120000,
    stuckSessionAbortMs: 720000,
  stuckSessionRecovery: {
      mode: "steer-then-nudge-then-abort", // | "abort" (current behavior)
      steerMessage: "…",  // optional template
      nudgeMessage: "…",
      recheckMs: 30000,
    }
  }
}

Expected behavior

• On session.stalled / eligible stalled_agent_run, run steer first (same path as sessions.steer / chat abort+inject).
• Re-check progress after recheckMs; if still stalled, sessions.send nudge.
• Re-check again; if still stalled and age ≥ stuckSessionAbortMs, fall back to current abort-drain.
• Emit structured diagnostic events: recovery.requested / recovery.completed with action: steer|nudge|abort.
• Never deliver recovery messages to external channels unless the session itself would reply there.

Workaround in use

Cron heartbeat every 4m executing the ladder via openclaw gateway call sessions.steer|abort + sessions_send, with gateway abort at 12m as final fallback.

Environment

OpenClaw 2026.5.22, multi-agent gateway (Discord/Telegram), embedded Pi runs, lossless-claw context engine.