Summary

Cron jobs running payload.kind=agentTurn against local models intermittently time out and/or hit context truncation failures, even when sessionTarget: "isolated" and payload.maxMessages/max_tokens are set.

Environment

• OpenClaw CLI: 2026.2.19-2 (45d9b20)
• Cron job: Aicia Gmail Monitor v2
  • schedule: every 5m (everyMs=300000)
  • sessionTarget: isolated
  • model: local/zai-org/glm-4.7-flash
  • payload.timeoutSeconds: 120
  • maxRunSeconds: 120
  • payload.maxMessages: 10
  • payload.max_tokens: 512

Observed behavior

From openclaw cron runs --id <jobId>:

• Frequent errors: Error: cron: job execution timed out (typically at ~60000ms before we increased limits)
• Also observed (returned as summary text):
  • Cannot truncate prompt with n_keep (11643) >= n_ctx (8192)
  • Connection error.
• Some runs show unexpectedly huge prompt usage under a different provider/model:
  • provider: ollama, model: qwen3.5:397b
  • input_tokens: ~50k

Expected behavior

• sessionTarget: isolated should prevent any history/prompt bloat across runs.
• payload.maxMessages and payload.max_tokens should cap request size deterministically.
• A job configured for a specific model should not silently execute under a different provider/model (or at least should log/flag this explicitly).

Questions / suspected issues

1. Is maxMessages and/or max_tokens currently enforced for cron agentTurn jobs?
2. Why can an isolated cron run still end up with a prompt that triggers n_keep >= n_ctx?
3. Under what conditions will cron execute with a different provider/model than configured?
4. Should payload.timeoutSeconds automatically set maxRunSeconds (they can diverge today)?

Repro (approx)

1. Create cron job:
   • sessionTarget: isolated
   • payload.kind=agentTurn
   • model local/zai-org/glm-4.7-flash
   • short message instructing to run a local script
2. Run at a tight interval (1–5 minutes)
3. Observe timeouts and occasional context/truncation errors in run history.

Why this matters

This looks like the root cause behind "cron congestion" / recurring timeouts for local-model automation workloads.