openclaw - 💡(How to fix) Fix Feature Request: Restore `agents.defaults.llm.idleTimeoutSeconds` config key

Summary

agents.defaults.llm.idleTimeoutSeconds was removed from the config schema in v2026.4.26 (openclaw doctor --fix labels it as a retired key). I would like to request restoring this config option.

Why this is needed

This knob is meaningfully different from models.providers.<id>.timeoutSeconds:

• models.providers.<id>.timeoutSeconds controls the total request lifetime (wall clock)
• idleTimeoutSeconds controls the maximum silence between streamed tokens/chunks (resets on every received token)

These solve different problems. For local/self-hosted models, especially large models with heavy prefill, the model may need 120s+ before the first token. In that case:

• increasing provider timeoutSeconds does not solve the issue if the internal idle watchdog still aborts at 120s
• users need an explicit way to raise the idle timeout without globally changing the default for everyone

Before removal, this config worked

• Users were successfully using agents.defaults.llm.idleTimeoutSeconds in v2026.4.15 (see #68461 with "idleTimeoutSeconds": 360000, #64541 with "idleTimeoutSeconds": 180)
• It was retired in v2026.4.26 — doctor --fix removes it as a retired key (see #73331)
• By v2026.5.12, the schema rejects it with: agents.defaults: Unrecognized key: "llm"

Proposed solution

Restore a user-configurable override, for example:

{
  "agents": {
    "defaults": {
      "llm": {
        "idleTimeoutSeconds": 300
      }
    }
  }
}

This preserves the current default behavior (120s hardcoded DEFAULT_LLM_IDLE_TIMEOUT_SECONDS) for everyone else, while giving local-model operators a necessary escape hatch.

I am not asking to increase the global default for all providers — that was correctly rejected in #76122. I am asking to restore a user-configurable override that existed before and was removed without a feature-level discussion.

Removing this knob left local-model users without a viable config-level workaround for slow first-token prefill.

Related

• #59678 — agents.defaults.timeoutSeconds ignored due to internal idle timeout
• #63200 — Idle-stream timeout breaks local models with heavy context (no config workaround)
• #65576 — Cron runs disable LLM idle watchdog when no explicit config
• #76122 — PR to increase default (rejected; review confirms provider timeout ≠ idle timeout)
• #64541 — User used agents.defaults.llm.idleTimeoutSeconds = 180 before removal
• #68461 — User used agents.defaults.llm.idleTimeoutSeconds = 360000 before removal