Background subagent notification silently reverts manually-selected model to config default

[jessie-amaranthine]

Description

When "OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true" is set and I manually select a non-default model in the picker, the model silently reverts to the agent's configured default (per "opencode.json") the moment the first background subagent notification is injected into the session.

The system prompt does not refresh after the revert; the agent continues reporting the old model as its identity because it has no way to detect the switch. Per session data from Tokscale (a CLI token-tracking dashboard), I was charged on both models sequentially -- the user-selected model handles messages before the notification; the config default handles messages after.

This does NOT happen with foreground subagent dispatch (multiple parallel task calls without background: true). It only triggers observably when background notifications are injected and I've manually selected a non-default model. When I stay on the config default model, any revert would be unobservable because the fallback target matches the active model.

Plugins

None

OpenCode version

OpenCode v1.18.4

Steps to reproduce

1. Set "OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true" in environment variables
2. In opencode.json, set a primary agent's model to Model A (e.g., opencode-go/deepseek-v4-pro)
3. Open a fresh session, manually select a different Model B (e.g., opencode-go/minimax-m3)
4. Send a message that causes the agent to dispatch subagents with "background: true"
5. Observe: the model switches from Model B back to Model A as soon as the first background notification arrives
6. The agent's system prompt still reports Model B (exported Markdown conversation + Tokscale confirms the responding model is actually Model A)

Screenshot and/or share link

No response

Operating System

Windows 10

Terminal

OpenChamber v1.16.3

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Model selection silently reverts after answering a question #34207: Model selection silently reverts after answering a question — describes the same silent model revert during an active session event (question injection), which may share the same root cause as the background subagent notification trigger reported here.

[jessie-amaranthine]

I checked #34207 -- related symptom but different trigger. That issue involves the user manually switching models mid-stream while the agent is composing a question, then the revert happening on question-answer. My case requires no mid-stream switching: the model is selected before the session starts, and the revert is triggered specifically by background subagent notification injection ("background: true" + "OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true"). Foreground dispatch does not trigger it. May share a root cause in the prompt injection path, but the reproduction and trigger conditions are distinct.

[jessie-amaranthine]

Update: confirmed in OpenCode CLI (not OpenChamber-specific).

Reproduced the same behavior in the OpenCode TUI using /model to switch to "opencode-go/kimi-k3", then sent my test prompt telling the model to dispatch 4 background subagents. The model reverted to my config default ("alibaba-token-plan/qwen3.8-max-preview") on the first background notification injection. Session export shows the assistant header switching from "Kimi K3" to "Qwen3.8 Max Preview" between the task launch message and the first notification response.

This rules out OpenChamber's UI layer as the cause. Also confirms it's model-agnostic (tested with both "opencode-go/minimax-m3" and "opencode-go/kimi-k3").