Summary

On session initialization, when the primary model (ollama/phi4-mini:latest) is unavailable (Ollama process running but model unresponsive), the main agent skips the configured 2nd fallback (openai/gpt-5.4-mini) and launches directly on the 3rd fallback (anthropic/claude-sonnet-4-6).

Environment

• OpenClaw version: 2026.4.14
• Host: macOS Darwin 24.6.0 (arm64)
• Channel: iMessage
• Agent: main (default)

Agent model config (from openclaw.json)

{
  "id": "main",
  "default": true,
  "model": {
    "primary": "ollama/phi4-mini:latest",
    "fallbacks": [
      "openai/gpt-5.4-mini",
      "anthropic/claude-sonnet-4-6"
    ]
  }
}

Steps to Reproduce

1. Ollama is running (pgrep -x ollama returns a PID) but phi4-mini is unresponsive or slow to load.
2. Start a new OpenClaw session via iMessage (or /new).
3. Session initializes on anthropic/claude-sonnet-4-6 instead of openai/gpt-5.4-mini.

Expected Behavior

Fallback chain should be respected in order:

1. ollama/phi4-mini:latest (primary) — fails
2. openai/gpt-5.4-mini (1st fallback) — should be selected
3. anthropic/claude-sonnet-4-6 (2nd fallback) — only if gpt-mini also fails

Actual Behavior

Session launched on anthropic/claude-sonnet-4-6 (3rd in chain), skipping openai/gpt-5.4-mini entirely.

Runtime header (from session)

model=anthropic/claude-sonnet-4-6 | default_model=ollama/phi4-mini:latest

Additional Notes

• After restarting Ollama (pkill ollama && open -a Ollama), phi4-mini loads correctly.
• The openclaw.json config is correctly structured — this appears to be a fallback selection bug at session init, not a config issue.
• This has a real cost impact: Sonnet is significantly more expensive than gpt-5.4-mini.