Summary

Model selection, reasoning effort, speed, and permission mode in the AI panel are not persisted. They live only in component state and reset back to the provider-reported default every time the panel re-initializes (e.g. on app return from background, on session disconnect, on sessionState === "ended" | "expired").

On iOS this is very visible: pick a non-default model (anything other than the GPT-5.5 default), send a message, leave the app, return — and the selector has snapped back to GPT-5.5.

Repro

1. Open the app, connect a CLI session.
2. In the AI panel, change model to something other than the backend's default (e.g. for Codex pick a non-default model), set reasoning to medium, set permission mode to full-access.
3. Send a message — it sends with the chosen options as expected.
4. Background the app for a moment and reopen (or disconnect/reconnect the session).
5. Reopen the panel: model has reverted to the provider default; reasoning/speed/permission have reverted to their initial constants.

Expected

• Once the user has picked model / reasoning / speed / permission, those choices stick — both within an ongoing chat and across app launches — until the user changes them again.
• Falling back to the provider default should only happen the first time a user touches the backend, or when the previously-selected value is no longer in the provider's supported list.

Where it happens

All state is plain useState (no persistence):

app/plugins/core/ai/Panel.tsx ~ L2715–L2729

const [selectedModelByBackend, setSelectedModelByBackend] = useState<Record<AiBackend, string>>({ opencode: "", codex: "" });
const [codexReasoningEffort, setCodexReasoningEffort] = useState(..."medium");
const [codexSpeed, setCodexSpeed] = useState("default");
const [codexPermissionMode, setCodexPermissionMode] = useState(..."default");

The init effect re-fetches providers and unconditionally overwrites the selection to the backend default, at ~ L3308–L3345:

setSelectedModelByBackend((prev) => ({
  ...prev,
  [backend]: models.length > 0 ? (defaultModelId || models[0].id) : "",
}));

isInitialized is reset on disconnect / session end at ~ L3395–L3400, which causes init to run again and clobber the selection on return.

Only AI_DETAILED_VIEW_STORAGE_KEY is persisted via AsyncStorage today (Panel.tsx L75, L3556, L3568). Nothing else.

Proposed fix (minimal)

Persist the four pieces of selection state via @react-native-async-storage/async-storage (already imported) using keys like:

• ai-selected-model:<backend>
• ai-selected-agent:<backend>
• ai-codex-reasoning
• ai-codex-speed
• ai-codex-permission

Then:

1. Hydrate them on mount before/alongside provider fetch.
2. Write on every change.
3. In the provider-fetch block (L3342), prefer the persisted value if it still exists in models, else fall back to defaultModelId || models[0].id.
4. The existing validators at L2900–L2913 already snap reasoning/speed back to a valid value if a persisted value is no longer supported, so no extra work needed there.

This is a small, contained change inside Panel.tsx. Happy to send a PR if a maintainer can confirm direction.

Scope / non-goals

• Not adding a settings pane in this issue (that's a follow-up — app/settings/ai.tsx currently only exposes aiFontSize).
• Not changing the picker UI.
• Not changing defaults for first-time users.

Environment

• Reproduced on iPhone app build (latest from App Store as of 2026-05-18).
• Source inspected at lunel-dev/lunel@main.

Related

• Adjacent (not duplicate): Clearer “reasoning” selection and plan mode in Codex #26 — about clearer reasoning level / plan-mode UI.
• Adjacent (not blocking): PR Add Pi coding agent support and model search in model picker #2 — touches Panel.tsx to add Pi backend + in-picker search; does not persist selection.