openclaw - 💡(How to fix) Fix Cron isolated sessions fail with LiveSessionModelSwitchError when payload model differs from agent default [1 comments, 2 participants]

Bug Description

Cron jobs with sessionTarget: "isolated" that specify a model different from the agent's default model fail immediately with LiveSessionModelSwitchError. This is a regression introduced in 2026.3.28 — all prior versions allowed cron jobs to freely override the model.

Steps to Reproduce

1. Configure an agent with a default model, e.g. openai-codex/gpt-5.3-codex for agent codex
2. Create a cron job with sessionTarget: "isolated", agentId: "codex", and a different model in the payload, e.g. model: "openai-codex/gpt-5.2"
3. Run the cron job (scheduled or manual trigger)

Expected Behavior

The cron job runs using the model specified in the payload (gpt-5.2), as it did in all versions prior to 2026.3.28.

Actual Behavior

The job fails within ~200ms with:

LiveSessionModelSwitchError: Live session model switch requested: openai-codex/gpt-5.3-codex

The error occurs because the run loop's live model switch detection (resolveLiveSessionModelSelection in auth-profiles-*.js) compares the currently running model (from the cron payload) against the agent's default model (resolved via resolveDefaultModelForAgent). Since the newly created isolated session has no providerOverride/modelOverride in the session store, it falls back to the agent default — which differs from the cron-specified model. This triggers a false positive "model switch detected" and throws LiveSessionModelSwitchError.

In the cron execution path, this error is not caught and retried the way it is in the interactive agent-runner loop (where the catch block at ~line 522 of agent-runner.runtime-*.js does continue with the new model). Instead, it propagates as an unhandled error and the job is marked as failed.

Root Cause Analysis

In auth-profiles-*.js around line 161503, resolveLiveSessionModelSelection():

const defaultModelRef = agentId 
  ? resolveDefaultModelForAgent({ cfg, agentId }) 
  : { provider, model };
const entry = loadSessionStore(...)[sessionKey];
const provider = entry?.providerOverride?.trim() || defaultModelRef.provider;
const model = entry?.modelOverride?.trim() || defaultModelRef.model;

For a fresh cron isolated session, entry has no overrides, so it resolves to the agent default model rather than the cron payload model. The main run loop then detects a mismatch between the running model and this "persisted" selection, and throws.

The fix should likely seed the session store entry with the cron payload's model override before the run loop begins, or skip the live model switch check for isolated/ephemeral sessions that have an explicit model specified in the run parameters.

Impact

• All cron jobs using sessionTarget: "isolated" with a model override different from the agent's default are broken
• Cross-provider overrides (e.g. agent default anthropic/claude-opus-4-6, cron model openai-codex/gpt-5.2) and same-provider overrides (e.g. gpt-5.3-codex vs gpt-5.2) are both affected
• Jobs that happen to use the same model as the agent default still work (e.g. zombie-process-watchdog with gpt-5.3-codex on the codex agent)

Environment

• OpenClaw version: 2026.3.28
• Node.js: v25.8.1
• Platform: macOS arm64

Workaround

Remove the model field from the cron payload or set it to match the agent's default model. This defeats the purpose of model overrides for cost optimization on lightweight tasks.