openclaw - 💡(How to fix) Fix Wizard: un-selecting the primary model leaves provider models: [] while keeping primary pointing at it [2 comments, 2 participants]

Summary

In openclaw configure (local mode), the model-selection step lets you un-check a model that is currently set as agents.defaults.model.primary and then confirm. The wizard writes the resulting (empty) selection to agents/main/agent/models.json as "models": [] for that provider, but does not clear or update agents.defaults.model.primary — which still points at the un-selected model under that now-empty provider.

The result: models.json has "openai-codex": { "models": [] } while openclaw.json still has "primary": "openai-codex/gpt-5.5".

Reproduction

1. Fresh OAuth-authenticated install with provider openai-codex and model gpt-5.5 selected as primary.
2. Run openclaw configure.
3. At the model-selection step, un-select the currently-primary gpt-5.5.
4. Confirm and exit the wizard.
5. Inspect:
   • agents/main/agent/models.json → openai-codex.models is []
   • openclaw.json → agents.defaults.model.primary is still openai-codex/gpt-5.5

Effect

Every subsequent agent run spends ~15s in the model-resolution startup phase (presumably falling through to API discovery against https://chatgpt.com/backend-api/codex since the local catalog is empty). Combined with auth (~7s) and dispatch (~6s), total startup is ~29s before streaming begins. During this window the Node event loop stalls (P99 delay 11–25s, utilization 1.0), producing:

• [fetch-timeout] fetch timeout reached; aborting operation
• [telegram] sendChatAction failed: Network request for 'sendChatAction' failed!
• [diagnostic] liveness warning: reasons=event_loop_delay,event_loop_utilization,cpu

Suggested fix

The wizard should either:

• Refuse to confirm if the un-selected set leaves primary/fallbacks references dangling, or
• Auto-update primary/fallbacks to a still-selected model (or clear them), or
• Refuse to write models: [] at all when the provider has model references elsewhere in config.

Workaround

Re-add the model definitions to the affected provider in agents/main/agent/models.json (can be copied from a sibling provider entry). Restart the gateway.

Environment

• OpenClaw version: 2026.4.29
• Platform: Raspberry Pi 5, ARM64, Debian 12