Agent Profile model always set to auto after I set to a custom model

[liuxuan30]

Pre-submit Checks

• I have searched Warp bugs and there are no duplicates
• I have searched Warp known issues page and my issue is not there
• I have an issue with AI and have included the debugging ID (Optional, but helps expedite the AI quality fix). Debugging ID instructions
• I have technical issue and have included the logs (optional, but helps expedite the bug fix). Log instructions

Describe the bug

My credits are used out, so I bought Z.ai coding plan lite.

I setup the urls and keys in warp, and set profile base model and full terminal use to this custom GLM5.2 model.

The model will be set back to auto no matter what I do.

Even creating a new profile will have such issue. It's annoying

To reproduce

1. use out warp builder plan credits
2. got z.ai coding plan lite.
3. setup endpoint for z.ai GLM-5.2 model
4. set agent profile base model and full terminal use to this GLM 5.2 model.
5. try start a new agent session, it will be the z.ai model.
6. take a coffee, watch tiktok for a few moment
7. start a new session or just check profile
8. the base model will be changed to auto(genius) for no idea why

Expected behavior

the model keeps the same, no matter what we do.

Screenshots, videos, and logs

No response

Operating system (OS)

Windows

Operating system and version

Windows 11 26200.8655

Shell Version

warp + builtin windows11 powershell (I didn't install any other shell)

Current Warp version

v0.2026.06.24.09.19.stable_02

Regression

No, this bug or issue has existed throughout my experience using Warp

Recent working Warp date

Yesterday

Additional context

No response

Does this block you from using Warp daily?

Yes, this issue prevents me from using Warp daily.

Is this an issue only in Warp?

Yes, I confirmed that this only happens in Warp, not other terminals.

Warp Internal (ignore): linear-label:b9d78064-c89e-4973-b153-5178a31ee54e

None

[oz-for-oss]

@liuxuan30

Here's what I found while triaging this issue:

Custom endpoint model selections should persist in Agent Profiles; reverting to Auto/Genius after a delay is not expected. This looks like an Agent/Profile settings persistence issue rather than normal credit-fallback behavior.

I have a few follow-up questions before I can narrow this down:

1. Could you attach a short screen recording showing the profile after saving the GLM model and then after it reverts to Auto/Genius?
2. Does the profile revert if Warp is left idle after saving the custom model, or only after starting/running a new agent session?

Reply in-thread with those details and the triage workflow will automatically re-evaluate the issue and update the diagnosis, labels, and next steps.

Maintainer details

I concluded that agent Profile custom model selections on Windows appear to revert to Auto/Genius after a delay, likely when the selected custom endpoint model no longer resolves during profile/model reconciliation..

Triage summary

This appears to be an Agent Profile model persistence bug for custom inference endpoints on Windows. The report says a Z.ai GLM-5.2 custom endpoint can be selected for both the Base model and Full terminal use model, works for an initial agent session, then later the saved profile shows Auto/Genius again.

Reproducibility

Medium. The report includes clear setup steps, OS, Warp version, and expected vs. actual behavior, but it does not include a screenshot/recording or logs showing the moment the profile value changes. The delayed timing also leaves open whether the reset is triggered by idle background refresh, starting a new agent session, custom endpoint/key state changes, or profile cloud sync.

Likely root cause

Agent Profile model preferences are stored as LLM IDs/config keys in AIExecutionProfile and edited through app/src/ai/execution_profiles/editor/mod.rs. Custom endpoint models are synthesized from ApiKeyManager.custom_endpoints in app/src/ai/llms.rs and crates/ai/src/api_keys.rs. During model/key/workspace refreshes, LLMPreferences::reconcile_disabled_model_preferences in app/src/ai/llms.rs clears stored profile model fields when the selected ID no longer resolves as usable; clearing the field makes the UI fall back to the default Auto/Genius model.

There is existing test coverage intended to preserve custom endpoint selections through a normal model-catalog refresh, so the remaining ambiguity is what makes this reporter's Z.ai custom model unresolvable later (for example a key/custom endpoint reload, entitlement state, or cloud profile update).

Duplicate check

I checked all open repository issues plus issues closed in the last 7 days. #11564 is related because it reports an AI profile model persistence problem, but it is a restart persistence issue for a different model path and is the only close candidate. Other matches concern BYOK gating, SSH quota errors, or custom endpoint request failures rather than Agent Profile values reverting to Auto. Because fewer than two likely duplicate issues match this exact symptom, I am not marking this as a duplicate.

Question reasoning

1. Could you attach a short screen recording showing the profile after saving the GLM model and then after it reverts to Auto/Genius? — This confirms the exact UI surface and whether both Base model and Full terminal use model are changing or only one field is affected.
2. Does the profile revert if Warp is left idle after saving the custom model, or only after starting/running a new agent session? — This distinguishes a background model/key/profile refresh from a reset caused by an agent request or quota/fallback path.

This is my automated analysis and may be incorrect. A maintainer will verify the details.

Powered by Oz

[liuxuan30]

1.Could you attach a short screen recording showing the profile after saving the GLM model and then after it reverts to Auto/Genius?

I cannot attach a recording; the length cannot be quantified, so I don't know when it will trigger or not. A shortest reproducing is I change to custom model, wait for like one minute, after directly openning the agent profile page, the model is set from GLM5.2 to auto while I'm looking at it. So there must a logic that will scan regularly and revert back to auto under certain circumstances.

2.Does the profile revert if Warp is left idle after saving the custom model, or only after starting/running a new agent session?

Both, either idle or staring a new agent.

Another fact is that, after saving the custom model and exit warp and reopen, the custom model will persist. It's just revert back while warp is opened.

My wild guess is the lack of credits triggering a banner or something to recharge credits might cause the auto revert?

[oz-for-oss]

@liuxuan30

Here's what I found while triaging this issue:

Thanks for the follow-up. The added detail that the custom model survives restart initially but reverts while Warp remains open makes this look like a live refresh/reconciliation bug rather than a simple save failure.

Maintainer details

I concluded that custom endpoint model selections in Agent Profiles appear to revert to Auto while Warp is running, likely through a live model/profile reconciliation path rather than a simple save failure..

Triage summary

The reporter sees a custom Z.ai GLM 5.2 endpoint selected for both the Agent Profile base model and full terminal use model, but within roughly a minute while Warp remains open the profile switches back to Auto/Genius. The reporter clarified that this happens both while idle and after starting a new agent session, while quitting and reopening initially preserves the custom model.

Reproducibility

medium — the report includes OS/version, Warp version, custom endpoint provider/model, affected fields, and timing, but reproducing likely requires a Windows custom Z.ai endpoint setup plus the reporter's out-of-Warp-credits account state.

Likely root cause

The current client stores profile model choices as optional LLMIds on AIExecutionProfile; when a stored model is cleared to None, the UI displays the default Auto model. Custom endpoint models are local synthetic LLMInfos built from ApiKeyManager.custom_endpoints. LLMPreferences::reconcile_disabled_model_preferences and sanitize_disabled_custom_model_preferences can clear base/coding/CLI custom selections when the selected ID is not currently usable/known or custom inference is considered disabled. Those reconciliation paths run on model catalog refresh, BYOK/custom endpoint key changes, and workspace/team changes, which fits a while-open revert and the reporter's observation that restart initially reloads the saved custom selection.

The implementation should check whether a periodic server/workspace/model refresh, API-key refresh, or cloud-object profile sync temporarily marks the custom endpoint model unavailable or overwrites the profile value, especially for out-of-credits users using custom inference.

Relevant files:

• app/src/ai/llms.rs
• app/src/ai/execution_profiles/profiles.rs
• app/src/ai/execution_profiles/editor/mod.rs
• app/src/settings_view/ai_page.rs
• crates/ai/src/api_keys.rs
• app/src/workspaces/user_workspaces.rs

Duplicate check

I scanned all open issues via gh api --paginate excluding #13136. The closest related issues were #11564 (default profile model persistence across restart), #11568 (BYOK model selection plan gating), and #12251 (custom provider support/persistence), but none matches this while-open custom endpoint model revert closely enough to mark as a duplicate.

This is my automated analysis and may be incorrect. A maintainer will verify the details.

Powered by Oz

[liuxuan30]

yep, I did a experiment, just open the profile and change to the z.ai custom model, leave the profile page open, and after a while, it was reverted back to auto. so defintely something is scanning.