Summary

When creating an isolated cron job with an explicit model override like --model "ollama/gemma3:12b", the cron job metadata correctly shows the overridden model, but the actual run still executes with the agent’s default model (in this case gpt-5.4 / openai-codex).

This makes it impossible to reliably move lightweight status/report cron jobs to a local model, even though the CLI and payload schema indicate model override support.

Expected behavior

For an isolated cron with:

• payload.kind = "agentTurn"
• sessionTarget = "isolated"
• payload.model = "ollama/gemma3:12b"

the actual run should execute with:

• model: ollama/gemma3:12b

Actual behavior

The cron job metadata shows:

• model: ollama/gemma3:12b

but the actual run history shows:

• model: gpt-5.4
• provider: openai-codex

So the override appears to be accepted at config/metadata level, but ignored at execution time.

Reproduction steps

1. Create a new isolated cron job for an agentTurn payload:

openclaw cron add \
  --name "coding-agent-status-report-20m-local" \
  --agent coding \
  --every 20m \
  --session isolated \
  --message "..." \
  --light-context \
  --announce \
  --channel telegram \
  --to 8261811945 \
  --timeout-seconds 600 \
  --model "ollama/gemma3:12b"

2. Verify the cron metadata shows the override:

• model: ollama/gemma3:12b

3. Trigger it manually:

openclaw cron run <job-id>

4. Inspect run history:

openclaw cron runs --id <job-id> --limit 1 --expect-final

5. Observe the run result shows execution with:

• model: gpt-5.4
• provider: openai-codex

instead of the requested local model.

Additional verification

The issue is not with Ollama itself.

Direct local model test works:

ollama run gemma3:12b "用中文简短回复：本地模型联通测试成功。只输出这一句。"

This returns successfully, confirming:

• Ollama is reachable
• gemma3:12b is installed
• local inference works

So the problem is specifically in the OpenClaw cron → isolated agentTurn execution path.

Observed scope

Affected setup:

• isolated cron jobs
• payload kind: agentTurn
• explicit --model override
• actual execution still uses agent default model

This was reproduced even after:

• deleting the old cron jobs entirely
• recreating new cron jobs from scratch
• verifying the new cron metadata had the correct local model

Likely issue

Possible bug/limitation in the execution path where:

• cron payload stores the model
• but the isolated agentTurn run resolves to the agent default model instead of honoring payload.model

Why this matters

This prevents routing lightweight cron/report jobs to local models, which is important for:

• cost control
• reducing unnecessary cloud model usage
• using strong cloud models only for actual coding/execution tasks
• using local models for frequent summary/status jobs

Example affected use case

Two high-frequency status cron jobs intended to use local model:

• coding status report
• customer-support-bot status report

Both were recreated fresh with:

• model: ollama/gemma3:12b

but actual execution still used:

• gpt-5.4

Requested fix

Ensure that for:

• sessionTarget = isolated
• payload.kind = agentTurn
• payload.model is present

the runtime actually executes with payload.model instead of silently falling back to the agent default model.