Bug: The /model Command Changes the Label But Not the Actual Model in OpenClaw TUI

A frustrating bug has surfaced in the OpenClaw TUI: when you use the /model command to switch models mid-session, the footer updates to show your new selection—but the agent continues using the old model. You think you're talking to GPT 5.3, but you're actually still chatting with MiniMax-M2.5.

The Problem

The /model command is supposed to let you switch models on the fly. It's convenient for testing different models on the same problem or switching to a more capable model for a complex task. When you run /model kimi-coding/k2p5, here's what happens:

1. The sessions.patch API call succeeds
2. The footer/status bar updates to show kimi-coding/k2p5
3. You feel satisfied that you've switched
4. The agent responds... and identifies itself as MiniMax-M2.5

Wait, what?

What's Actually Happening

The issue is that sessions.patch updates the session metadata but doesn't propagate to the runtime. The model change is cosmetic—it affects what the UI displays but not what the inference engine actually uses.

One user on GitHub demonstrated this with screenshots: the TUI footer clearly shows kimi-coding/k2p5 while the agent's response in the same session says "I'm MiniMax-M2.5." Same conversation, seconds apart, completely out of sync.

This isn't a config issue. This isn't user error. The model switching pipeline is fundamentally broken.

Why This Matters

If you're debugging an agent's behavior and think you've switched models, you could spend hours chasing phantom issues. You might conclude "Claude is bad at this task" when you've actually been testing on Gemini the whole time.

For users who rely on /model for:

• A/B testing responses across models
• Switching to cheaper models for simple tasks
• Escalating to more capable models for hard problems

...this bug makes the feature completely unreliable.

Current Workarounds

Until this is fixed, your options are limited:

1. Start a new session after changing models (the nuclear option)
2. Edit your config file and restart the gateway (heavy-handed but reliable)
3. Use per-agent model config instead of /model switching

None of these are great. The whole point of /model is quick, mid-conversation switching without restarting anything.

The Fix

The runtime needs to respect session model changes, not just record them. When sessions.patch updates the model, that change must propagate to wherever the actual inference call is made.

This is being tracked as #29572. If this bug affects your workflow, add a thumbs-up to help prioritize it.

Takeaway

If you've been using /model and getting inconsistent results, now you know why. The label changes, the model doesn't. Until the fix lands, don't trust the footer—verify your model by asking the agent directly.