[Bug] Cron isolated session model parameter ignored, falls back to default agent model

Bug Type

Regression / Configuration Issue

Summary

When using sessionTarget: "isolated" with payload.kind: "agentTurn" in cron jobs, the payload.model parameter is ignored at runtime. The job always uses agents.defaults.model.primary (e.g., claude-sonnet-4-5) instead of the specified model (e.g., claude-haiku-4).

This prevents cost optimization for cron jobs, as cheaper models cannot be selected per-job.

Expected Behavior

When a cron job is configured with:

{
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "...",
    "model": "anthropic/claude-haiku-4",
    "timeoutSeconds": 300
  }
}

The isolated session should use Claude Haiku 4 as specified.

Actual Behavior

The isolated session uses the default agent model (anthropic/claude-sonnet-4-5) instead, ignoring the payload.model parameter entirely.

Evidence from cron run logs:

{
  "status": "ok",
  "model": "claude-sonnet-4-5",  // <-- Wrong model!
  "provider": "anthropic",
  "usage": {
    "input_tokens": 16,
    "output_tokens": 493
  }
}

Steps to Reproduce

1. Ensure target model is enabled

Add anthropic/claude-haiku-4 to agent's enabled models:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5"
      },
      "models": {
        "anthropic/claude-sonnet-4-5": {},
        "anthropic/claude-haiku-4": {}
      }
    }
  }
}

2. Create isolated cron job with model override

openclaw cron add \
  --name "Test Model Override" \
  --every 1h \
  --session isolated \
  --message "Run a simple test script and report results" \
  --model "anthropic/claude-haiku-4" \
  --timeout-seconds 300 \
  --announce \
  --channel discord \
  --to <channel-id>

3. Run job manually

openclaw cron run <job-id>

4. Check run logs

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

Observe: "model": "claude-sonnet-4-5" instead of "model": "claude-haiku-4"

Configuration Used

OpenClaw Version: 2026.4.2 (d74a122)

Job Configuration:

{
  "id": "730648b9-7854-4b2b-b02b-b13b6fa63c57",
  "name": "Discord Security Real-Time Monitor",
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "Run enhanced Discord security monitoring...",
    "model": "anthropic/claude-haiku-4",
    "timeoutSeconds": 300
  },
  "delivery": {
    "mode": "announce",
    "channel": "discord",
    "to": "1485243052418138172"
  }
}

Agent Config:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5"
      },
      "models": {
        "anthropic/claude-sonnet-4-5": {},
        "anthropic/claude-haiku-4": {}
      }
    }
  }
}

Test Results

Multiple test runs with different model specifications:

Conclusion: Model parameter is completely ignored for isolated cron sessions.

Impact & Severity

Affected: Users attempting cost optimization via cheaper models in cron jobs

Severity: Medium (functional workaround exists, but blocks cost optimization)

Frequency: Every isolated cron job with custom model parameter

Consequence:

• Cannot use budget models (Haiku, GPT-4o-mini, etc.) for cron automation
• Forces expensive models (Sonnet 4-5) for simple tasks
• Blocks 90%+ cost reduction for automated jobs

Cost impact example:

• Sonnet 4-5: $3 input / $15 output per 1M tokens
• Haiku 4: $0.25 input / $1.25 output per 1M tokens
• Savings blocked: ~92% per run

Environment

• OS: Linux 6.12.57+deb13-amd64 (Debian 13, Docker/Umbrel)
• Node.js: v22.22.2
• OpenClaw: 2026.4.2 (d74a122)
• Install method: Umbrel app (npm global equivalent)
• Agent: main (default)
• Provider: Anthropic (API key configured)

Workarounds Attempted

❌ Model name format variations

• anthropic/claude-haiku-4 → Ignored
• claude-haiku-4 → Ignored
• Result: Format doesn't matter, parameter is ignored

❌ Free local models (Ollama)

• ollama/qwen2.5:3b → Timeout (separate issue)
• ollama/qwen2.5:14b → Not tested (likely same timeout)
• Result: Works but too slow (600s+ timeout needed)

✅ Revert to main session (systemEvent)

• sessionTarget: "main", payload.kind: "systemEvent"
• Result: Works but no model override capability
• Cost optimization: Frequency reduction only (75% savings)

Suggested Fix

The payload.model parameter should be respected for isolated cron sessions, similar to how sessions_spawn with runtime: "subagent" respects the model parameter.

Possible implementation:

1. Pass payload.model to isolated session spawn
2. Override session model before agent turn execution
3. Validate model is in enabled models list
4. Fall back to default only if model unavailable

Related Issues

• Potentially related to sessions_spawn model handling
• May share code path with subagent model override logic

Additional Context

Use case: Security monitoring cron jobs

• Simple script execution + log checking
• No complex reasoning required
• Perfect candidate for budget models
• Current forced use of Sonnet 4-5 is wasteful

Desired outcome:

• Enable per-job model selection
• Allow cost optimization for automated tasks
• Maintain reliable execution

Reporter: User via OpenClaw agent
Date: 2026-04-04
Priority: Medium (blocks cost optimization feature)