fix(chat): keep one global model choice across conversations

docs/dialog-improvement-api.md

@@ -357,7 +357,7 @@ Query 参数：
 
 | 参数 | 类型 | 必填 | 说明 |
 | --- | --- | --- | --- |
-| `projectKey` | string | 是 | 项目注册表中的标识 |
+| `projectKey` | string | 否 | 仅兼容旧客户端；模型目录为全局配置，不按项目过滤或扫描项目 |
 | `query` | string | 否 | 按 provider、model、displayName 检索 |
 | `provider` | string | 否 | 过滤 provider |
 | `includeAuto` | boolean | 否 | 是否在 Router 可用时返回 auto |
@@ -388,6 +388,7 @@ type ModelCatalogItem = {
 };
 
 type ModelsResponse = {
+  defaultSelection: { mode: "model"; provider: string; model: string };
   items: ModelCatalogItem[];
   router: {
     enabled: boolean;
@@ -408,6 +409,10 @@ type ModelsResponse = {
 - 未返回的能力表示该模型不支持对应参数。
 - Router 开启且支持 auto 时，接口可返回 `{ provider: "router", model: "auto" }` 虚拟条目。
 
+Web Composer 使用同一浏览器、同一站点的全局模型偏好：首次使用 `defaultSelection`，用户手动选择后保存完整模型及参数，跨项目、会话、刷新和标签页复用。系统默认值变化只影响未手动选择的用户；不可用的已选模型保留并提示重新选择。旧项目和会话草稿不参与全局偏好恢复，避免无法确定时间顺序时任意继承旧选择。
+
+Composer 不再请求会话模型 GET/PUT 来恢复或保存选择。每条消息仍携带提交时的 `modelSelection` 快照，运行中和排队消息不受后续选择影响。下列会话 API 为兼容已有客户端保留，不决定 Web Composer 的全局选择。模型目录由页面内共享缓存复用，配置重载或 WebSocket 重连后失效。
+
 ### 6.2 查询会话模型设置
 
 `GET /api/sessions/model?sessionKey=<sessionKey>&projectKey=<projectKey>`
@@ -500,6 +505,7 @@ Gateway WebSocket 方法：`submit_turn`
 
 ```ts
 type SessionModelOverride = {
+  mode: "model";
   provider: string;
   model: string;
   reasoning?: number;
@@ -522,21 +528,22 @@ type SubmitTurnRequest = {
   mode?: "default" | "plan" | "bypassPermissions";
   basePermissionMode?: "default" | "plan" | "bypassPermissions";
   modelOverride?: SessionModelOverride;
+  modelSelection?: { mode: "auto" } | SessionModelOverride;
   runId?: string;
 };
 ```
 
-`modelOverride` 仅覆盖当前 turn，不修改第 6.3 节保存的会话模型设置。未传 `modelOverride` 时使用保存的会话设置；会话未设置时，Router 开启则使用 auto，否则使用系统默认模型。
+Web Composer 始终提交 `modelSelection` 快照，可明确选择具体模型或 Auto，并随已接收输入记录。`modelOverride` 为已有客户端保留，只覆盖当前 turn；两者不能同时传入。仅当两者都未传时，才使用兼容的会话设置；会话未设置时，Router 开启则使用 auto，否则使用系统默认模型。
 
 服务端校验：
 
-- `modelOverride.provider/model` 必须存在且可用。
+- 显式选择的 `provider/model` 必须存在且可用；Auto 必须有可用 Router。
 - reasoning、temperature、speed 必须满足模型 capabilities；speed 使用 `0..1` 的统一数值语义。
 - `uploadedAttachments` 必须属于同一 `projectKey`、状态为 completed 且未过期。
 - `mode` 和 `basePermissionMode` 必须属于声明枚举。
 - 校验失败时不得启动模型调用。
 
-模型选择优先级：本轮 `modelOverride` > 已保存的会话模型 > Router auto > 系统默认模型。
+模型选择优先级：本轮 `modelSelection` 或 `modelOverride`（互斥）> 兼容会话设置 > Router auto > 系统默认模型。Web Composer 始终发送快照，所以不依赖会话回退。
 
 响应为 Gateway 事件流，新增事件：
 

docs/trd-dialog-improvement.md

@@ -213,7 +213,9 @@ type UploadedAttachmentRef = {
 
 ### 9.1 模型目录
 
-`GET /api/models?projectKey=&query=&provider=&includeAuto=`
+`GET /api/models?query=&provider=&includeAuto=`
+
+模型目录直接读取全局配置，不枚举项目或会话；旧客户端传入的 `projectKey` 仅为兼容保留。返回 `defaultSelection` 明确指定系统默认模型，目录顺序不影响选择。
 
 返回 provider、model、displayName、available 以及 reasoning（推理强度）、temperature 和可选 speed 的能力声明。对话框统一使用 0..1 的数值语义；每个模型可通过能力声明限制可用范围、步长或枚举值，后端负责把 0..1 值映射为 Provider 所需参数。temperature 和 speed 统一范围为 0..1。官方 OpenAI / Anthropic 模型默认声明 speed；自定义模型需显式 `supportsSpeed: true`，且 Google Provider 当前不支持该字段。目录将 speed 暴露为枚举 `0`（标准）与 `1`（快速）。
 
@@ -227,6 +229,7 @@ type UploadedAttachmentRef = {
 
 ```ts
 type SessionModelOverride = {
+  mode: "model";
   provider: string;
   model: string;
   reasoning?: number;
@@ -235,11 +238,16 @@ type SessionModelOverride = {
 };
 type GatewaySubmitTurnInput = ExistingGatewaySubmitTurnInput & {
   modelOverride?: SessionModelOverride;
+  modelSelection?: { mode: "auto" } | SessionModelOverride;
   uploadedAttachments?: UploadedAttachmentRef[];
 };
 ```
 
-### 9.3 会话模型状态
+### 9.3 Web 全局偏好与兼容会话模型状态
+
+Web Composer 只保存同一浏览器、同一站点内最近一次手动选择的模型及参数，不区分项目或会话，并同步其他标签页。未手动选择时使用系统默认模型；手动选择后，配置默认值变化、会话切换、提交确认和运行结果都不改写偏好。已选模型不可用时提示重新选择，不自动切换。
+
+Composer 不再调用下列会话模型读写接口。它在准备附件之前固定每条提交的 `modelSelection`，队列、编辑重发和历史执行记录继续保留各自的快照。会话 API 保留给已有客户端，其保存值不参与 Web 全局选择恢复。
 
 新增会话模型读写接口：
 
@@ -249,7 +257,7 @@ type GatewaySubmitTurnInput = ExistingGatewaySubmitTurnInput & {
 
 保存值写入 session metadata，会话恢复后继续生效。`mode=auto` 仅在 Router 开启时允许；清除设置后，Router 开启则回到 auto，Router 关闭则回到 `agent.model`。
 
-`submit_turn.modelOverride` 只覆盖本轮，不修改会话保存值。模型解析顺序：本轮 `modelOverride` > 会话保存模型 > Router auto/路由决策 > `agent.model` 默认模型。
+`submit_turn.modelSelection` 为 Web 提交时固定的模型快照，随已接收输入记录；`modelOverride` 只覆盖本轮，不修改会话保存值。两者互斥，均优先于兼容会话设置。只有两者都未传时才使用会话保存模型 > Router auto/路由决策 > `agent.model` 默认模型；Web Composer 不依赖这条回退路径。
 
 `provider/model` 不存在或不可用返回 `INVALID_MODEL_OVERRIDE`；reasoning、temperature 或 speed 不满足模型能力返回 `UNSUPPORTED_MODEL_PARAMETER`。未声明支持的参数不发送给 Provider。speed 必须在 canonical request 入口通过 `0..1` 校验，再由支持 speed 的 Provider adapter 映射为原生字段；Google Provider 不声明或接收 speed。
 

src/agent/protocol/input.ts

@@ -35,4 +35,6 @@ export type AgentSubmitOptions = {
    */
   syntheticMessages?: import("../../model/index.js").CanonicalMessage[];
   modelOverride?: AgentModelOverride;
+  /** Submitted model snapshot, recorded for replay and legacy session clients. */
+  modelSelection?: NonNullable<import("../../session/transcript/TranscriptEntry.js").SessionMetadataValue["modelSelection"]>;
 };

src/agent/session/AgentSession.ts

@@ -95,6 +95,7 @@ export class AgentSession {
       permissionRules: submitOptions.permissionRules,
       syntheticMessages: submitOptions.syntheticMessages,
       modelOverride: submitOptions.modelOverride,
+      modelSelection: submitOptions.modelSelection,
       abortSignal: this.state.abortController.signal,
       openSteerMailbox: () => this.steerMailbox.start(turnId),
       drainSteerMessages: () => this.steerMailbox.drain(turnId),

src/agent/turn/TurnRunner.ts

@@ -36,6 +36,7 @@ export type TurnRunnerOptions = {
   /** Synthetic messages appended after user input; stored with metadata.synthetic flag. */
   syntheticMessages?: CanonicalMessage[];
   modelOverride?: AgentModelOverride;
+  modelSelection?: NonNullable<SessionMetadataValue["modelSelection"]>;
   openSteerMailbox?: () => void;
   drainSteerMessages?: () => AgentSteerMessage[];
   drainOrCloseSteerMailbox?: () => { messages: AgentSteerMessage[]; closed: boolean };
@@ -488,12 +489,15 @@ export class TurnRunner {
 
     const snapshot = metadataStore.getSnapshot();
     const prompt = allHumanText(acceptedMessages);
-    if (!prompt) return;
+    if (!prompt && !options.modelSelection) return;
 
-    const boundedPrompt = prompt.slice(0, SESSION_LISTING_PROMPT_MAX_CHARS);
+    const boundedPrompt = prompt?.slice(0, SESSION_LISTING_PROMPT_MAX_CHARS);
     await metadataStore.record(options.turnId, {
-      ...(snapshot.firstPrompt ? {} : { firstPrompt: boundedPrompt }),
-      lastPrompt: boundedPrompt,
+      ...(boundedPrompt ? {
+        ...(snapshot.firstPrompt ? {} : { firstPrompt: boundedPrompt }),
+        lastPrompt: boundedPrompt,
+      } : {}),
+      ...(options.modelSelection ? { modelSelection: { ...options.modelSelection } } : {}),
       updatedAt: this.now().toISOString(),
     }).catch(() => {});
   }
@@ -505,6 +509,8 @@ function isVisibleFailureStatus(status: AgentStatusMessageInput): boolean {
 
 function acceptedInputMetadata(options: TurnRunnerOptions): Record<string, unknown> | undefined {
   const metadata: Record<string, unknown> = {};
+  // Save alongside input so a crash before the metadata snapshot cannot lose the choice.
+  if (options.modelSelection) metadata.modelSelection = { ...options.modelSelection };
   if (options.permissionMode) {
     metadata.permissionMode = options.permissionMode;
   }

src/cli/createLocalGateway.ts

@@ -382,8 +382,7 @@ export function createLocalGateway(options: CreateLocalGatewayOptions = {}): Cre
       return listCommands({ ...input, projectKey }, pilotHome);
     },
     async modelCatalogList(input) {
-      const projectKey = await dialogProjects.resolveProjectKey(input.projectKey);
-      return listModelCatalog({ ...input, projectKey }, env);
+      return listModelCatalog(input, env);
     },
     async sessionModelGet(input) {
       const projectKey = await dialogProjects.resolveProjectKey(input.projectKey);
@@ -415,13 +414,22 @@ export function createLocalGateway(options: CreateLocalGatewayOptions = {}): Cre
     },
     async resolveTurnModelSelection(input) {
       const projectKey = await dialogProjects.resolveProjectKey(input.projectKey ?? fallbackProjectRoot);
+      if (input.modelSelection !== undefined && input.modelOverride !== undefined) {
+        throw new DialogGatewayError("INVALID_MODEL_OVERRIDE", "Specify modelSelection or modelOverride, not both.");
+      }
+      if (input.modelSelection !== undefined) {
+        validateModelSelection(projectKey, input.modelSelection, env);
+        return input.modelSelection.mode === "model"
+          ? { selection: input.modelSelection, source: "turn" as const }
+          : { source: "router" as const };
+      }
       if (input.modelOverride) {
         validateExplicitModelSelection(projectKey, input.modelOverride, env);
         return { selection: input.modelOverride, source: "turn" as const };
       }
       const saved = await readSavedModel(projectKey, input.sessionKey);
+      if (saved) validateModelSelection(projectKey, saved, env);
       if (saved?.mode === "model") {
-        validateExplicitModelSelection(projectKey, saved, env);
         return { selection: saved, source: "session" as const };
       }
       const config = loadPilotConfig({ projectRoot: projectKey, env }).config;

src/gateway/client/InProcessGateway.ts

@@ -554,10 +554,13 @@ export class InProcessGateway implements Gateway {
         }));
         const modelSelection = this.options.resolveTurnModelSelection
           ? await this.options.resolveTurnModelSelection(input)
-          : input.modelOverride
-            ? { selection: input.modelOverride, source: "turn" as const }
-            : { source: "default" as const };
+          : input.modelSelection?.mode === "auto"
+            ? { source: "router" as const }
+            : input.modelSelection?.mode === "model" || input.modelOverride
+              ? { selection: input.modelSelection?.mode === "model" ? input.modelSelection : input.modelOverride, source: "turn" as const }
+              : { source: "default" as const };
         let lastEmittedModel: string | undefined;
+        let actualRequestModel: string | undefined;
         if (modelSelection.selection) {
           const event: GatewayEvent = {
             type: "model_selection_changed",
@@ -577,6 +580,7 @@ export class InProcessGateway implements Gateway {
           agentInput,
           {
             turnId: runId,
+            modelSelection: input.modelSelection,
             maxTurns: input.maxTurns,
             runMode,
             permissionMode,
@@ -620,6 +624,7 @@ export class InProcessGateway implements Gateway {
           if (event.type === "input_accepted") {
             await this.commitAcceptedTurnReplacement(input.sessionKey, runId);
           }
+          if (event.type === "model_event" && event.event.type === "request_started") actualRequestModel = event.event.model;
           if (event.type === "model_event" && event.event.type === "request_started"
             && lastEmittedModel !== `${event.event.provider}\0${event.event.model}`) {
             const selectionEvent: GatewayEvent = {
@@ -634,6 +639,10 @@ export class InProcessGateway implements Gateway {
             lastEmittedModel = `${event.event.provider}\0${event.event.model}`;
           }
           for (const gatewayEvent of mapAgentEvent(event, runId)) {
+            if (gatewayEvent.type === "assistant_text_delta" && actualRequestModel) gatewayEvent.model = actualRequestModel;
+            if (gatewayEvent.type === "input_accepted" && input.modelSelection) {
+              gatewayEvent.modelSelection = { ...input.modelSelection };
+            }
             if (gatewayEvent.type === "context_budget") {
               this.recordGatewayStatusMessage({
                 sessionKey: input.sessionKey,

src/gateway/dialog/modelCatalog.ts

@@ -16,8 +16,7 @@ const REASONING_VALUES = new Map<number, ThinkingMode>([
 ]);
 
 export function listModelCatalog(input: ModelCatalogListInput, env: NodeJS.ProcessEnv = process.env): ModelCatalogListResult {
-  if (!input.projectKey?.trim()) throw new DialogGatewayError("PROJECT_NOT_FOUND", "projectKey is required.");
-  const config = loadPilotConfig({ projectRoot: input.projectKey, env }).config;
+  const config = loadPilotConfig({ env }).config;
   const query = input.query?.trim().toLocaleLowerCase() ?? "";
   const items: ModelCatalogItem[] = [];
   for (const [providerId, provider] of Object.entries(config.model.providers)) {
@@ -51,11 +50,15 @@ export function listModelCatalog(input: ModelCatalogListInput, env: NodeJS.Proce
     && (!query || "router auto".includes(query))) {
     items.unshift({ id: "router/auto", provider: "router", model: "auto", displayName: "Auto", available: true, capabilities: {} });
   }
-  return { items, router: { enabled: routerEnabled, autoAvailable: routerEnabled } };
+  return {
+    items,
+    defaultSelection: { mode: "model", provider: config.agent.model.provider, model: config.agent.model.model },
+    router: { enabled: routerEnabled, autoAvailable: routerEnabled },
+  };
 }
 
 export function validateModelSelection(projectKey: string, selection: SessionModelSelection, env: NodeJS.ProcessEnv = process.env): void {
-  if (selection.mode === "auto") {
+  if (selection?.mode === "auto") {
     if (!listModelCatalog({ projectKey }, env).router.autoAvailable) {
       throw new DialogGatewayError("ROUTER_AUTO_UNAVAILABLE", "Router auto is not available for this project.");
     }

src/gateway/protocol/types.ts

@@ -100,6 +100,8 @@ export type GatewaySubmitTurnInput = {
   uploadedAttachments?: UploadedAttachmentRef[];
   /** A one-turn model override. Persisted session preferences are managed separately. */
   modelOverride?: ExplicitModelSelection;
+  /** Submitted choice: used for this turn and recorded with accepted input; never updates the Web global preference. */
+  modelSelection?: SessionModelSelection;
   runMode?: AgentRunMode;
   mode?: GatewayMode;
   /** The user's actual permission preference before plan-mode override. */
@@ -172,7 +174,7 @@ type GatewayTurnScopedEventMetadata = {
 
 export type GatewayEvent = GatewayTurnScopedEventMetadata & (
   | { type: "turn_started"; runId: string }
-  | { type: "input_accepted"; runId: string }
+  | { type: "input_accepted"; runId: string; modelSelection?: SessionModelSelection }
   | { type: "steer_applied"; itemId: string; message: CanonicalMessage }
   | { type: "steer_unapplied"; itemId: string; reason: "turn_ended" }
   | { type: "model_request_started"; model?: string; provider?: string }
@@ -185,7 +187,7 @@ export type GatewayEvent = GatewayTurnScopedEventMetadata & (
       temperature?: number;
       speed?: number;
     }
-  | { type: "assistant_text_delta"; text: string }
+  | { type: "assistant_text_delta"; text: string; model?: string }
   | { type: "assistant_attachment"; attachment: GatewayOutboundAttachment }
   | { type: "file_artifacts"; artifacts: import("../../session/artifacts/FileArtifact.js").FileArtifact[] }
   | { type: "assistant_thinking_delta"; text: string }
@@ -467,13 +469,15 @@ export type ModelCatalogItem = {
 };
 
 export type ModelCatalogListInput = {
-  projectKey: string;
+  /** Accepted for compatibility; the model catalog is global. */
+  projectKey?: string;
   query?: string;
   provider?: string;
   includeAuto?: boolean;
 };
 
 export type ModelCatalogListResult = {
+  defaultSelection: ExplicitModelSelection;
   items: ModelCatalogItem[];
   router: { enabled: boolean; autoAvailable: boolean };
 };

src/model/protocol/canonical.ts

@@ -134,6 +134,8 @@ export type CanonicalContentBlock =
   | CanonicalMediaReferenceBlock;
 
 export type CanonicalMessageMetadata = {
+  /** Actual model that generated this assistant message. */
+  model?: string;
   /** True for messages injected by the system (e.g. JSON self-correct prompts). */
   synthetic?: boolean;
   /** Synthetic prompt that should be consumed by the next assistant response only. */

src/model/streaming/assembleModelMessage.ts

@@ -18,6 +18,7 @@ import {
 export type ModelMessageAssemblerState = {
   content: CanonicalContentBlock[];
   textBuffer: string;
+  model?: string;
   thinkingBuffer: string;
   thinkingReasoningContentBuffer: string;
   thinkingSignature?: string;
@@ -67,6 +68,8 @@ export function applyModelEventToAssembler(
 ): void {
   switch (event.type) {
     case "request_started":
+      state.model = event.model;
+      return;
     case "message_start":
     case "tool_call_start":
     case "tool_call_delta":
@@ -151,6 +154,7 @@ export function assembleAssistantMessage(state: ModelMessageAssemblerState): Ass
     message: {
       role: "assistant",
       content: [...state.content],
+      ...(state.model ? { metadata: { model: state.model } } : {}),
     },
     finishReason: state.finishReason ?? (state.error ? "error" : "unknown"),
     hasMessageEnd: state.hasMessageEnd,

src/session/transcript/TranscriptReplay.ts

@@ -60,6 +60,12 @@ export function replayTranscriptEntries(entries: AgentTranscriptEntry[]): AgentT
 
     switch (entry.type) {
       case "accepted_input":
+        if (entry.metadata?.modelSelection) {
+          const choice = entry.metadata.modelSelection as SessionMetadataValue["modelSelection"];
+          if (choice?.mode === "auto" || (choice?.mode === "model" && typeof choice.provider === "string" && typeof choice.model === "string")) {
+            metadata = mergeMetadata(metadata, { modelSelection: { ...choice } });
+          }
+        }
         if (!beforeBoundary) {
           messages.push(...cloneMessages(entry.messages));
           events.push({