feat: per-model max output tokens and a conversation-time thinking level backed by agent settings (#28)

Co-authored-by: Alice <alice@prismshadow.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

packages/web/src/lib/omni/stream-model.ts

@@ -241,6 +241,13 @@ export interface StreamModel {
   items: ChatItem[];
   /** A nested sub-session model (produces no stats row; its stats count toward the parent). */
   nested: boolean;
+  /**
+   * The session's thinking level, captured from the main session's `session_meta` on history
+   * replay ("default" when the Agent config leaves it unset; null until a session_meta has been
+   * seen). llmConfig is assembled once per session, so this is fixed for the session's lifetime —
+   * shown read-only in the input area next to the locked model.
+   */
+  thinkingLevel: string | null;
   stats: TaskStatsTracker;
   /** The currently open text/thinking fragment (opened by start, closed by stop). */
   openText: AssistantTextItem | null;
@@ -318,6 +325,7 @@ function newModel(nested: boolean, localDecisions: Set<string>): StreamModel {
   return {
     items: [],
     nested,
+    thinkingLevel: null,
     stats: createTaskStatsTracker(),
     openText: null,
     openThinking: null,
@@ -392,7 +400,13 @@ export function pushMessage(
     advanceLastTs(model, msg.timestamp);
     return;
   }
-  // session_meta (main session): not rendered.
+  // session_meta (main session): not rendered as an item, but its thinking level is captured
+  // for the input area's read-only display (fixed per session: llmConfig is assembled once at
+  // session creation, so a mid-conversation Agent-config change doesn't affect this session).
+  if (msg.type === "session_meta") {
+    const level = (msg.payload as { thinking_level?: unknown }).thinking_level;
+    if (typeof level === "string" && level) model.thinkingLevel = level;
+  }
 }
 
 /** ISO timestamp → milliseconds (returns undefined if invalid). */

packages/web/src/lib/strings-en.ts

@@ -312,6 +312,10 @@ export const en: Strings = {
     contextWindow: "Context window",
     contextWindowUnit: "tokens",
     contextWindowHint: "Leave empty if unknown",
+    maxTokens: "Max output tokens",
+    maxTokensHint: "Leave empty to inherit the agent setting",
+    maxTokensCapHint: "Caps output tokens per request; lower it for small-context local models",
+    maxTokensInvalid: "Must be a positive integer",
     clientTypeLocked: (t: string): string => `Protocol: ${t} (kept as configured; not editable)`,
     vision: "Supports image input (vision / multimodal)",
     visionHint:
@@ -480,6 +484,14 @@ export const en: Strings = {
     newSessionMenu: "New chat",
     chooseAgent: "Choose agent",
     chooseModel: "Choose model",
+    thinkingLevel: "Thinking level",
+    thinkingLevelNames: {
+      none: "None",
+      low: "Low",
+      medium: "Medium",
+      high: "High",
+      xhigh: "Extreme High",
+    },
     workspaceUseThis: "Use this dir",
     workspaceUp: "Parent dir",
     workspaceNoSubdirs: "No subdirectories",

packages/web/src/lib/strings.ts

@@ -287,6 +287,10 @@ export const zh = {
     contextWindow: "上下文窗口",
     contextWindowUnit: "tokens",
     contextWindowHint: "留空表示未知",
+    maxTokens: "最大输出长度（Token）",
+    maxTokensHint: "留空沿用 Agent 设置",
+    maxTokensCapHint: "限制单次请求的输出 Token 数；小上下文的本地模型建议调低",
+    maxTokensInvalid: "必须为正整数",
     clientTypeLocked: (t: string): string => `协议：${t}（沿用原配置，不可修改）`,
     vision: "支持图片输入（视觉/多模态）",
     visionHint:
@@ -458,6 +462,15 @@ export const zh = {
     newSessionMenu: "新建对话",
     chooseAgent: "选择 Agent",
     chooseModel: "选择模型",
+    thinkingLevel: "思考等级",
+    /** 会话前拾取器的档位短名（评审要求：只写短名、不带说明、无“缺省”项）。 */
+    thinkingLevelNames: {
+      none: "无",
+      low: "低",
+      medium: "中",
+      high: "高",
+      xhigh: "极高",
+    } as Readonly<Record<string, string>>,
     workspaceUseThis: "使用此目录",
     workspaceUp: "上级目录",
     workspaceNoSubdirs: "无子目录",