UI model picker silently resets to agent's default after first message in new session

Summary

In a new session, after the user manually selects a model via /model and sends the first message, the model picker silently resets back to the agent's default model after the response. The second message — if sent without re-selecting — uses the agent default, not the previously chosen model. Once the user manually selects a model for the second time, the picker sticks and subsequent messages persist correctly.

So the first manual selection is effectively a "one-shot" that silently expires, while any subsequent selection is persistent. The asymmetry is confusing, and the reset is invisible unless the user notices the status bar label has changed.

Repro

Configure an agent with an explicit model:

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "build": { "model": "anthropic/claude-opus-4-7" }
  }
}

Open opencode on a fresh session (homepage).
/model → switch to openai/gpt-5.4.
Send message 1: test
- Response: GPT-5.4 ✓
- Status bar after response: Claude Opus 4.7 — silently reverted
Send message 2 (no manual switch): test
- Response: Claude Opus 4.7 ✗ — agent default, not the earlier choice
/model → switch back to openai/gpt-5.4.
Send message 3: test
- Response: GPT-5.4 ✓
Send message 4 (no manual switch): test
- Response: GPT-5.4 ✓ — now persists correctly

Reproduced multiple times; the asymmetric behavior is deterministic. Same pattern reproduces with models routed through custom OpenAI-compatible providers.

Backend fallback is fine; issue is in the frontend state

From packages/opencode/src/session/prompt.ts:932 (createUserMessage):

const model = input.model ?? ag.model ?? (yield* lastModel(input.sessionID))

This three-way fallback is reasonable. The observable behavior above implies the UI stops sending input.model on message 2 but resumes sending it from message 3 onward — so the backend correctly falls through to ag.model exactly once. In other words the picker's selectedModel state is being reset exactly once, after the first message of a new session completes.

I didn't pinpoint the frontend line, but the four-step repro makes the behavior deterministic and isolates it to frontend picker state management.

Expected

The picker should persist the user's manual selection across all subsequent messages in the session, including the one immediately after the first response — same as messages 3+ already do.

Impact

Users unknowingly send their second message to the agent's default model instead of their chosen one
The reset is not surfaced via toast or warning
Particularly impactful for users on custom OpenAI-compatible providers where provider-level cost and capability differences matter — silent routing to the wrong backend happens with zero indication

This issue might be related to existing issues about model selection persistence. Please check:

[FEATURE]: Interactively selected model resets to agent default when switching between Plan and Build modes (TUI) #21351: Model selected via /model resets to agent default when switching between Plan and Build modes — similar symptom of session-level model selection not sticking
Subagent with specific model results in model change in build and plan mode #6636: Subagent with specific model causes main agent's model to change — related model state management issue
Only current selected agent's model persists when resuming a session (other agent(s) revert to default) #23369: Model reverts to default when resuming a session — another variant of model persistence not working as expected

These are not exact duplicates (the specific trigger here — first message of a new session — appears distinct), but the underlying cause may be shared.

Thanks bot — cross-checked all three and I think the likely shared root cause sits in how the frontend tracks per-agent selected model state, and specifically how it decides whether to send input.model on each submission.

All four issues (this one plus #21351, #6636, #23369) are different triggers of the same underlying asymmetry around packages/opencode/src/session/prompt.ts:932:

const model = input.model ?? ag.model ?? (yield* lastModel(input.sessionID))

The backend fallback order is fine in isolation. What varies across these issues is whether the frontend correctly sends input.model on a given submission:

UI model picker silently resets to agent's default after first message in new session #23666 (this issue): UI drops input.model on message 2 of a new session, exactly once
[FEATURE]: Interactively selected model resets to agent default when switching between Plan and Build modes (TUI) #21351: UI drops input.model on Tab-switch between Plan/Build, so each mode re-resolves to the new agent's ag.model
Only current selected agent's model persists when resuming a session (other agent(s) revert to default) #23369: Session resume only restores one agent's selected model; other agents re-resolve to ag.model on switch
Subagent with specific model results in model change in build and plan mode #6636 goes in the opposite direction (subagent's model bleeding into main agent). Without reading that code path I can't say if it's the same state map, but it's clearly in the same "per-agent selected model state" problem area — possibly the same map written with the wrong agent key

Net: these look like four symptoms of one missing invariant — "each agent in a session has its own persistent selectedModel state, independent of ag.model, surviving first-message completion, mode switches, subagent dispatch, and session resume". Fixing that one invariant should close this whole cluster.

Happy to keep these linked; feel free to triage this as a cluster under a single model-selection-persistence umbrella if that fits the roadmap better.

I can take this. The root cause is in packages/opencode/src/cli/cmd/tui/context/local.tsx — the createEffect that syncs agent model fires on every reactive sync.data.agent refresh (new object references), not just on actual agent switches. After message 1 completes, sync refreshes → effect fires → overwrites user selection with agent config model. Will submit a fix shortly.

OpenCode model-switch bug investigation:

Confirmed the bug you described — it matches issues UI model picker silently resets to agent's default after first message in new session #23666 and TUI footer shows stale default provider/model after switching via /models #23741 on the opencode GitHub
Root cause identified by contributor kagera-agent: the createEffect in local.tsx that syncs agent model fires on every reactive refresh (after message completion), overwriting the user's manual selection
Related issues: [FEATURE]: Interactively selected model resets to agent default when switching between Plan and Build modes (TUI) #21351 (Plan/Build switch resets model), Only current selected agent's model persists when resuming a session (other agent(s) revert to default) #23369 (session resume loses model), Subagent with specific model results in model change in build and plan mode #6636 (subagent model bleeds into main agent)
All appear to share one root cause: per-agent selectedModel state not persisting correctly across mode switches, message completions, and session resume

Still not fixed

Reproduction steps below...

This is so annoying.

Steps to reproduce:

Launch opencode with an alias 'minimax'
opencode -m minimax-coding-plan/MiniMax-M2.7

This is what appears on screen:

3. Send a message, and this is what appears on screen: 4. Notice that the model is routing to the wrong provider (since my limit of Github Copilot has ran out). Its clearly using/routing the wrong model. It should be Minimax, but its trying to use GPT-5.4 (since its my default model in OpenCode):
To get this session back to normal, i now have to do the following workaround:

Press ESC on the keyboard.

type /model, switch to minimax

copy the previous message, paste it again, now it proceeds (see below)

This is a major glitch.

Proposed solution:
#23741 (comment)

I second this as it is PEAK annoying. Please let us know how we can help you with getting this fixed. Thanks!

Would also love to see this prioritized as it has tripped me up a few times when I start a conversation with GPT 5.5 selected to override my default MiniMax M2.7. Then, after the initial turn, the model appears to be automatically reverted to the default (MiniMax M2.7).

(Previously, I have used Codex CLI for GPT interactions and Opencode for interactions with self-hosted models and ones from the Chinese providers, so this was not a big problem for me. But, Codex CLI has become so buggy lately that I am migrating completely to Opencode. Need my overrides in Opencode to stick.)

This is the only glitch that made me move away from OpenCode.

Navigation Menu

Repository navigation

UI model picker silently resets to agent's default after first message in new session #23666

Summary

Repro

Backend fallback is fine; issue is in the frontend state

Expected

Impact

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

UI model picker silently resets to agent's default after first message in new session #23666

Description

Summary

Repro

Backend fallback is fine; issue is in the frontend state

Expected

Impact

Activity

github-actions commented on Apr 21, 2026

XIYBHK commented on Apr 21, 2026

kagura-chen commented on Apr 21, 2026

coygeek commented on Apr 25, 2026

coygeek commented on Apr 27, 2026

coygeek commented on Apr 27, 2026

coygeek commented on Apr 27, 2026

Fohlen commented on Apr 30, 2026

emcd commented on May 6, 2026

coygeek commented on May 6, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions