Manually selected model in UI is overridden by agent config model via runtime-fallback

[alferd0806]

Bug Description

When a user manually selects a model in the OpenCode UI (using the model selector dropdown) and sends a message, the model is immediately overridden back to the OMO agent-configured model (from oh-my-openagent.json → agents.sisyphus.model). The manually selected model is used for the message, but the session reverts to the agent config model.

Root Cause

The bug involves two interacting code paths in the runtime-fallback handler:

1. handleSessionCreated unconditionally overwrites session state

In runtime-fallback/event-handler.ts, handleSessionCreated (around line 111688) always overwrites sessionStates[sessionID] with a new state derived from the session.created event info, without checking if the user has already made a manual model selection.

// Simplified logic:
const sessionModel = sessionInfo.model;            // e.g., "B" (user selected)
const preferredModel = resolvePreferredSessionModel(...); // "A" (agent config)
const fallbackIndex = preferredModel !== model
  ? getFallbackModelsForSession(...).indexOf(model) : -1;

const state = createFallbackState(
  fallbackIndex >= 0 && preferredModel ? preferredModel : model
);
// If model B is in the fallback list:
if (fallbackIndex >= 0) {
  state.currentModel = model;      // B
  state.originalModel = preferredModel; // A
}
sessionStates.set(sessionID, state);

Since the user-selected model B is in the agent's fallback_models list, the handler creates a state where:

• originalModel = A (agent config model)
• currentModel = B (user-selected, but treated as "fallback")

2. createChatMessageHandler2 interprets the state as an active fallback

In runtime-fallback/chat-message-handler.ts, the chat.message handler (around line 111305) processes each message and at lines 111348-111364:

const activeModel = state.currentModel;  // B
if (activeModel === state.originalModel) return;  // B !== A → continues

// ❌ Applies "fallback override" — actually reverts to agent config!
output.message.model = {
  providerID: parts[0],
  modelID: parts.slice(1).join("/")
};  // Forces model back to A

The handler interprets currentModel !== originalModel as "a fallback is active" and overrides output.message.model with the agent config model — even when the user manually selected the current model.

Event timing scenario

1. User selects model B in UI, sends message
2. session.created fires → handleSessionCreated
   → state = { originalModel: A, currentModel: B }
   → B is treated as "fallback" because it's in fallback_models list
3. chat.message fires → runtime-fallback handler
   → input.model = B, state.currentModel = B
   → requestedModel === state.currentModel → no "manual change" detected
   → activeModel (B) !== originalModel (A) → treats as fallback override
   → output.message.model = A ← BUG!

Expected Behavior

When a user manually selects a model in the UI:

1. The session state should respect the user's choice, not treat it as a fallback
2. handleSessionCreated should not overwrite session state that was set by a user's manual model selection
3. createChatMessageHandler2 should distinguish between "fallback due to error" and "user manually selected a different model"

Suggested Fix

There are two approaches:

Approach A: Fix handleSessionCreated to not overwrite on manual selection

In the session.created event handler, check if the user already has a pending model selection before overwriting. If chat.message already processed the message and detected a manual model change (reset state), session.created should not overwrite that.

Approach B: Fix createChatMessageHandler2 to respect user selection

When activeModel !== originalModel:

1. Check if input.model contains the user's UI selection
2. If the "fallback" model matches the user's UI choice, it is NOT a fallback — it is a user preference
3. Only apply the model override when the model change was due to an API error, not a user action

A minimal fix would be in createChatMessageHandler2: add a check that if input.model matches state.currentModel and differs from state.originalModel, the user intentionally chose this model — do not override.

// Additional check before applying fallback override:
if (input.model) {
  const requestedModel = `${input.model.providerID}/${input.model.modelID}`;
  if (requestedModel === state.currentModel && requestedModel !== state.originalModel) {
    // User intentionally selected this model — not a fallback
    return;
  }
}

Additional Context

• OMO version: 4.15.0
• The bug occurs regardless of which models are involved, as long as the UI-selected model is in the agent's fallback_models list
• The bug does NOT occur for agents without explicit model config (because hasExplicitAgentModelOverride returns false, causing a different code path that does not interact with session state)
• Discovered while investigating issue Runtime-fallback doesn't trigger on "Free usage exceeded" error — missing patterns in classifier #5815 (runtime-fallback not triggering on Free usage exceeded)

[alferd0806]

补充：切 Tab 也会触发模型覆盖

新现象

桌面端切换 Tab 时，目标 Tab 的模型选择会自动被覆盖为该 Tab 最后一轮对话使用的模型。

调用链分析

根因

和原 issue 同一个 root cause —— 中的 在非首条消息且 为空时，无条件从 缓存中取值并覆盖 。

为什么要修

用户希望：

1. 手动选的模型不应被自动覆盖（原 issue）
2. 切 Tab 时应保留该 Tab 当前 UI 上选择的模型，而不是回退到历史消息的模型

可能的修复方向

• 不应简单地返回 ，它应该区分「用户手动选择的模型」和「历史消息记录中的模型」
• 或者在 中不覆盖已由用户手动设置的模型

[alferd0806]

补充：切 Tab 也会触发模型覆盖

新现象

桌面端切换 Tab 时，目标 Tab 的模型选择会自动被覆盖为该 Tab 最后一轮对话使用的模型。

调用链分析

Tab Switch -> Session Restore
  -> handleMessageUpdatedSessionState()  [dist/index.js:149499]
      -> 从 session 的最后一条消息提取 providerID + modelID
      -> noteSessionModel(= setLastKnownModel) + setSessionModel()
      -> sessionModels.set(sessionID, {providerID, modelID})
  |
  -> createChatMessageHandler3()  [dist/index.js:148675]
      -> getStoredMainSessionModel()  [dist/index.js:148623]
          -> isFirstMessage?        -> NO（非首条）
          -> subagent?              -> NO
          -> input.model?           -> NO（restore 不带 model）
          -> hasExplicitAgentOverride? -> NO
          -> return getSessionModel(sessionID)  <- 返回刚记录的 model
      -> output.message.model = getSessionModel()  -> 被覆盖

根因

和原 issue 同一个 root cause —— 在 createChatMessageHandler3 中，getStoredMainSessionModel 在非首条消息且 input.model 为空时，无条件从 sessionModels 缓存中取值并覆盖 output.message.model。

为什么和原 issue 表现不同

• 原 issue：首条消息被 handleSessionCreated + runtime-fallback 覆盖
• 切 Tab：非首条消息被 handleMessageUpdatedSessionState + getStoredMainSessionModel 覆盖

两者的共同点是 output.message.model 都被插件无条件覆写。

用户期望

1. 手动选的模型不应被自动覆盖
2. 切 Tab 后应保留该 Tab 当前 UI 上选择的模型，而不是回退到历史消息记录的模型

文件位置

• dist/index.js 中 getStoredMainSessionModel（约 line 148623）
• dist/index.js 中 createChatMessageHandler3（约 line 148675）
• dist/index.js 中 handleMessageUpdatedSessionState（约 line 149499）
• dist/index.js 中 recordSessionModel（约 line 148641）