openclaw - 💡(How to fix) Fix Cannot switch models via /models when active provider is rate-limited

Bug

When the active provider (e.g. Anthropic/Claude) hits its rate limit or quota cap, the /models slash command becomes unusable. The user cannot switch to an alternative provider (e.g. OpenAI GPT, OpenRouter) because the /models command appears to route through the currently active (and failing) provider rather than through the control layer.

This makes it impossible to fall back to an alternative model when the primary provider's quota is exhausted.

Steps to Reproduce

1. Use OpenClaw with Claude (Anthropic) as the active model
2. Exhaust Claude quota / hit rate limit
3. Attempt to use /models to switch to a different provider (e.g. OpenAI GPT)
4. The command fails or is unresponsive because it depends on the rate-limited provider

Expected Behavior

/models (and related model-switching commands) should work regardless of the active provider's rate limit status. Model switching is a control-plane operation and should not depend on the active AI provider being available.

Actual Behavior

When the active provider is rate-limited, /models fails, leaving the user stuck on the exhausted provider with no way to switch to an available alternative.

Related Issues

• #43447 — Kimi/Moonshot rate limit causes UI lockout and prevents model switching (same root cause, different provider). That issue's "Proposed Fix #3: Emergency Model Override" describes the needed fix: route /model commands through the control layer, not the active AI provider.
• #61096 (closed) — /model override not persisting when provider is rate-limited; fixed in v2026.4.10, but the fix addressed persistence of the override, not the fundamental issue of the command being blocked when the provider is down.
• #6966 — Dynamic model switching based on rate limits (automatic downgrade; related but different — this issue is about manual switching being blocked).

Proposed Fix

Route /models and all model-switching commands through the control/session layer directly, bypassing the active AI provider entirely. Model selection is a session-management operation that should never touch the inference API.