openclaw - 💡(How to fix) Fix Cron isolated sessions fail with LiveSessionModelSwitchError when agent primary model differs from payload.model [1 participants]

Bug Report

Summary

Cron jobs with sessionTarget: "isolated" and payload.kind: "agentTurn" fail immediately with LiveSessionModelSwitchError when the payload.model field specifies a different model than the agent's configured model.primary. This makes the model field in cron payloads effectively non-functional.

Environment

• OpenClaw version: 2026.3.28 (f9b1079)
• Runtime: Linux x64, Node v22.22.1
• Channel: Telegram (primary) + Discord (secondary)

Root Cause Analysis

The issue stems from a mismatch between how isolated cron sessions are created vs how model overrides are applied:

1. Session creation uses the agent's model.primary (e.g., anthropic/claude-opus-4-6)
2. Model override from payload.model (e.g., zai/glm-5.1) triggers resolveLiveSessionModelSelection() which detects the difference
3. LiveSessionModelSwitchError is thrown in auth-profiles-B5ypC5S-.js at the session runner level
4. The runner in agent-runner.runtime-C-sR1PRP.js does handle this error gracefully for interactive sessions (it catches it, updates followupRun.run.provider/model, and continues)
5. However, the cron scheduler treats it as fatal — the run is marked status: "error" and no retry occurs

Key Code Paths

Error generation (auth-profiles-B5ypC5S-.js):

// Before each attempt, checks if model selection changed
const nextSelection = resolvePersistedLiveSelection();
if (hasDifferentLiveSessionModelSelection(resolveCurrentLiveSelection(), nextSelection)) {
    throw new LiveSessionModelSwitchError(nextSelection);
}

Interactive runner handles it (agent-runner.runtime-C-sR1PRP.js):

if (err instanceof LiveSessionModelSwitchError) {
    params.followupRun.run.provider = err.provider;
    params.followupRun.run.model = err.model;
    params.followupRun.run.authProfileId = err.authProfileId;
    // continues...
}

But isExpectedNonErrorLaneFailure() marks it as expected non-error, suggesting the cron lane may not have the same retry logic.

Reproduction Steps

1. Configure an agent with model.primary: "anthropic/claude-opus-4-6" (or any provider A model)
2. Create a cron job with payload.model: "zai/glm-5.1" (or any provider B model)
3. Trigger the cron manually or wait for schedule
4. Observe: run fails immediately with LiveSessionModelSwitchError

Minimal Config

{
  "agents": {
    "list": [{
      "id": "my-agent",
      "model": {
        "primary": "anthropic/claude-opus-4-6",
        "fallbacks": ["zai/glm-5-turbo"]
      }
    }]
  }
}

{
  "name": "test-cron",
  "schedule": { "kind": "every", "everyMs": 600000 },
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "model": "zai/glm-5.1",
    "message": "Hello world"
  }
}

Expected Behavior

The cron job's payload.model should be used as the session's initial model, not trigger a model switch after session creation.

Actual Behavior

LiveSessionModelSwitchError: Live session model switch requested: anthropic/claude-opus-4-6

Duration: ~2-6 seconds, no actual LLM call is made.

Workaround

Set the agent's model.primary to the same provider used in cron payloads. This defeats the purpose of per-cron model selection.

Impact

• All cron jobs on agents with cross-provider primary models are broken
• Affects qa-tester, supply-chain-monitor, clawbbo-sync — 41+ consecutive failures on one job alone
• The model field in cron payload is documented as functional but silently fails

Suggested Fix

Either:

1. Apply payload.model before session creation (ideal — prevents the switch entirely)
2. Add the same retry/switch logic to the cron lane that the interactive runner already has
3. Clear isExpectedNonErrorLaneFailure handling so the error is retried as a transient failure