Summary

Cron job payload.model field is ignored when the target agent has a model set in agents.list[].model. The agent-level model pin always wins, causing every cron job to cascade through the entire fallback chain and land on the agent's pinned model (usually the most expensive one).

Environment

• OpenClaw version: 2026.3.28
• OS: Ubuntu 22.04 (Linux 5.15.0-171-generic x64)
• Setup: 5 agents (main, builder, orion, bizdev, health) with per-agent model pins

Config

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-opus-4-6",
        "fallbacks": ["anthropic/claude-sonnet-4-6", "openai-codex/gpt-5.4", "google/gemini-3.1-pro", "lmstudio/qwen/qwen3.5-9b"]
      }
    },
    "list": [
      { "id": "builder", "model": "anthropic/claude-opus-4-6" },
      { "id": "orion", "model": "anthropic/claude-sonnet-4-6" }
    ]
  }
}

Cron job example:

{
  "agentId": "builder",
  "payload": {
    "model": "openai-codex/gpt-5.4-mini",
    "message": "Run the daily error scanner..."
  }
}

Expected Behavior

The cron's payload.model field (gpt-5.4-mini) should be used for that session, overriding the agent's default model. This is the whole point of the field — to run cheap/lightweight cron tasks on appropriate models without changing the agent's interactive model.

Actual Behavior

The gateway detects a "live session model switch" and rejects every candidate except the agent's pinned model:

[agent/embedded] live session model switch detected before attempt for <run-id>: openai-codex/gpt-5.4-mini -> anthropic/claude-opus-4-6
[model-fallback/decision] candidate_failed requestedModel=gpt-5.4-mini candidateModel=gpt-5.4-mini reason=unknown errorPreview="Live session model switch requested: anthropic/claude-opus-4-6"
[model-fallback/decision] candidate_failed candidateModel=gpt-5.4 reason=unknown errorPreview="Live session model switch requested: anthropic/claude-opus-4-6"  
[model-fallback/decision] candidate_failed candidateModel=gemini-3.1-pro-preview reason=unknown errorPreview="Live session model switch requested: anthropic/claude-opus-4-6"
[model-fallback/decision] candidate_failed candidateModel=qwen/qwen3.5-9b reason=unknown errorPreview="Live session model switch requested: anthropic/claude-opus-4-6"
[model-fallback/decision] candidate_succeeded candidateModel=claude-opus-4-6 attempt=5 total=5

Every candidate fails with reason=unknown and the same error: "Live session model switch requested: anthropic/claude-opus-4-6". The session always lands on the agent's pinned model.

Impact

• All 15+ enabled cron jobs cascade through 4-5 fallback candidates before landing on the agent's model
• Produces ~60+ warning log lines per cron cycle (every 30 minutes)
• Jobs configured for gpt-5.4-mini ($0.001/run) execute on claude-opus-4-6 ($0.10/run) — 100x cost multiplier
• Operators cannot run cost-tiered cron work without this fix

Suggested Fix

The cron payload.model field should create a per-session model override (similar to /model in interactive sessions) that takes priority over the agent's agents.list[].model pin for the duration of that cron run.

Workaround

Currently no clean workaround. Removing the agent-level model pin would affect interactive sessions. The cron model field is documented but non-functional when agent pins exist.