Bug: LiveSessionModelSwitchError crashes all isolated cron sessions

Version

OpenClaw 2026.3.28 (f9b1079, latest on npm)

Description

All cron jobs with sessionTarget: "isolated" crash immediately with LiveSessionModelSwitchError. The error occurs regardless of model configuration, payload settings, or agent setup.

Error Pattern

LiveSessionModelSwitchError: Live session model switch requested: zai/glm-4.7

The "requested" model in the error message always matches the agent's configured primary model — no actual model switch is happening.

Affected Jobs

All jobs with sessionTarget: "isolated" fail. Jobs using other session targets (e.g., main with kind: "systemEvent") work fine.

What We Tested (none fixed it)

1. Removed model from job payload → still crashes
2. Set model in payload to match agent primary exactly → still crashes
3. Removed lightContext, thinking, fallbacks from payload → still crashes
4. Gateway restart (clean, single instance) → still crashes
5. Deleted all stale session files for affected agents → still crashes
6. Reset job state (removed cached sessionId, reset consecutiveErrors) → still crashes
7. Verified single gateway process (found and killed duplicate) → still crashes
8. All 6 model APIs respond correctly when tested directly
9. All agent API keys are valid (unique per-agent keys)
10. Provider configs are complete in each agent's models.json

Source Code Analysis

Examined the bundled source in dist/auth-profiles-B5ypC5S-.js:

// These are the SAME function — both read from LIVE_MODEL_SELECTIONS Map
function resolveCurrentLiveSelection(sessionKey) {
    const current = LIVE_MODEL_SELECTIONS.get(sessionKey);
    return current ?? void 0;
}
function resolvePersistedLiveSelection(sessionKey) {
    return resolveCurrentLiveSelection(sessionKey);
}

The check in agent-runner.runtime:

const nextSelection = resolvePersistedLiveSelection();
if (hasDifferentLiveSessionModelSelection(resolveCurrentLiveSelection(), nextSelection)) {
    throw new LiveSessionModelSwitchError(nextSelection);
}

Since both functions are identical and called without arguments, both should return undefined → hasDifferentLiveSessionModelSelection(undefined, undefined) returns false → error should never trigger. Yet it does.

This suggests either:

• An async operation modifies LIVE_MODEL_SELECTIONS between the two calls
• Or the resolveModelSelectionForSession() call earlier in the run sets a selection that conflicts with itself
• Or the functions are being called WITH different session keys in the actual runtime (the minified code may behave differently than the source suggests)

Reproduction

1. Create a cron job with sessionTarget: "isolated" and payload.kind: "agentTurn"
2. Manually trigger it: via the cron API/tool
3. Observe immediate crash with LiveSessionModelSwitchError

Environment

• OS: WSL2 (Linux 6.6.87.2-microsoft-standard-WSL2, x64)
• Node: v22.22.1
• Agents: 6 configured agents (main/dora, scout, forge, archivist, watchman, rainmaker) with per-agent API keys and models
• Global default model: zai/glm-5-turbo
• Provider: ZAI (first-party, 6 models registered)

Workaround

Disable all isolated cron jobs and run them manually from an interactive session when needed.