@rexvane/piweb 0.3.21 → 0.3.23

package/src/i18n.tsx

@@ -38,13 +38,42 @@ const zh = {
 	capacityFields: "容量",
 	contextWindowLabel: "上下文窗口",
 	maxTokensLabel: "最大输出",
-	capacityHint: "留空用默认 · 支持 256K / 2M 写法",
+	capacityHint: "留空按 pi 缺省（128K / 16K）· 支持 256K / 2M 写法",
+	// 模型能力（视觉）与配置体检
+	visionLabel: "视觉",
+	visionHint: "该模型能否接收图片。未声明时 pi 按纯文本处理：你贴的截图会被换成一行占位文字（提示语还会说成「模型不支持图像」）",
+	visionUnset: "未声明",
+	visionInputText: "文本",
+	visionInputTextHint: "pi 总是接受文本输入；贴图能不能发由「图片」决定",
+	visionInputImage: "图片",
+	visionInputImageHint: "勾上表示该模型能收图（写进 models.json 的 input）；不勾则按纯文本处理",
+	visionOn: "有视觉",
+	visionOff: "纯文本",
+	imagePreview: "点击放大预览",
+	imageViewer: "图片预览",
+	imageClearAll: "清空图片",
+	imagePrev: "上一张",
+	imageNext: "下一张",
+	chipExpandMore: "展开 {n}",
+	chipCollapse: "收起",
+	modelAuditTitle: "{n} 个模型没写能力字段",
+	modelAuditDesc: "没写的字段，pi 会按上下文 {ctx}、输出上限 {out}、纯文本处理。端点没报的值不会替你猜：照模型的真实规格填上（支持 256K / 1M 写法），保存时只补空着的格子，已写的值不改。",
+	modelAuditColModel: "模型",
+	modelAuditColContext: "上下文",
+	modelAuditColOutput: "输出上限",
+	modelAuditColVision: "视觉",
+	modelAuditVisionQuick: "视觉一起设：",
+	modelAuditVisionOn: "都有视觉",
+	modelAuditVisionOff: "都是纯文本",
+	modelAuditSave: "保存 {n} 个模型",
+	modelAuditSaveIdle: "填好后保存",
 	capacityInvalid: "第 {n} 行：容量需为数字（可用 K/M）",
 	discoverEmpty: "远端没有返回可识别的模型",
 	candidateTitle: "发现 {count} 个模型，勾选要添加的",
 	candidateAll: "全选",
 	candidatePicked: "已选 {n} 个",
 	candidateAdopt: "添加 {n} 个",
+	candidateInCatalog: "已在目录中",
 	runningTool: "Running {name}…",
 	thinkingLive: "Thinking",
 	stepRunning: "Running",
@@ -127,6 +156,10 @@ const zh = {
 	gitNoChanges: "工作区干净，没有变更",
 	gitStatusUnknown: "Git 状态读取失败，无法确认是否有变更",
 	gitRecentCommits: "最近提交",
+	gitStagingArea: "暂存区",
+	gitWorkingTree: "工作区",
+	gitGithubAccount: "GitHub 账号",
+	gitGithubUnknown: "未识别到 GitHub 账号",
 	toolNoOutput: "无输出",
 	toolFailed: "失败",
 	thinkingClamped: "当前模型不支持「{requested}」，已改为「{level}」",
@@ -285,6 +318,11 @@ const zh = {
 	extensionStatus: "扩展状态",
 	contextUsed: "上下文已用",
 	contextUnknown: "上下文用量未知",
+	contextEstimated: "上下文估算",
+	contextMeasured: "上下文用量",
+	contextBackground: "Main Conversation",
+	contextApiInput: "最近请求输入（API）",
+	contextChars: "字符",
 	contextSystem: "系统提示词",
 	contextMessages: "对话消息",
 	contextTools: "工具",
@@ -326,6 +364,20 @@ const zh = {
 	contextColdHint: "会话还没启动，System Prompt 和工具定义要启动后才算得出",
 	contextCountNow: "现在统计",
 	contextCounting: "统计中…",
+	contextWindowEdit: "修改",
+	contextWindowBuiltin: "pi 内置目录：{value}",
+	contextWindowUseBuiltin: "填入",
+	contextWindowScope: "改的是这个模型的全局配置（models.json），pi 终端和其它会话一起生效",
+	contextWindowOverBuiltin: "比内置目录大：超过模型实际上限时，请求会被上游拒绝（400）",
+	contextWindowNoBuiltin: "填模型实际支持的上限；填大了，请求会被上游拒绝（400）",
+	contextWindowInvalid: "填正整数，可带 K / M（如 200K）",
+	// @ 文件引用（输入 @ 弹候选）
+	mentionFiles: "@ 引用文件 / 文件夹",
+	mentionHint: "↑↓ 选择 · Enter 引用 · Esc 关闭",
+	mentionNoMatch: "没有匹配的文件",
+	mentionTruncated: "候选已截断",
+	mentionDiskSource: "目录遍历",
+	mentionDiskSourceHint: "当前工作区不是 git 仓库（或没装 git）：候选来自目录遍历，会跳过 node_modules、构建产物等目录；列为 git 仓库后候选会自动按 .gitignore 过滤",
 	// 统计条
 	turns: "轮次",
 	toolCalls: "工具调用",
@@ -416,6 +468,21 @@ const zh = {
 	drawerPanel: "右侧抽屉",
 	drawerSections: "抽屉入口",
 	drawerPrompts: "提示词",
+	sideChat: "旁聊",
+	sideChatModel: "旁聊模型",
+	sideChatFollowMain: "跟随主会话",
+	sideChatNew: "新旁聊",
+	sideChatHistory: "旁聊记录",
+	sideChatMessages: "旁聊消息",
+	sideChatQuestion: "旁聊提问",
+	sideChatPlaceholder: "读到哪里有疑问，就在这里问…",
+	sideChatEmpty: "边读边问，主会话停在原处",
+	sideChatHint: "参考主会话，问答独立。记录仅保存在此浏览器。",
+	sideChatPickSession: "打开一个主会话后，就可以旁聊",
+	sideChatAnswering: "正在回答…",
+	sideChatContextTruncated: "背景较长，本次仅参考了部分主会话或旁聊记录。",
+	sideChatLengthLimited: "已达到本次输出上限，可以继续追问。",
+	sideChatStorageFailed: "浏览器未能保存旁聊记录，请先保留需要的内容。",
 	setPlugins: "插件",
 	setRemote: "远程连接",
 	remoteBuiltin: "内置插件",
@@ -507,10 +574,14 @@ const zh = {
 	allowPrivateDiscovery: "允许本机/私网",
 	allowPrivateHint: "模型地址指向本机或私有网络（如 Ollama、LM Studio、one-api 本地网关）时勾选，本次探测放行；发消息不受此限制",
 	discoveringModels: "获取中...",
+	autoDiscovering: "正在自动识别模型…",
+	autoDiscovered: "已自动识别 {n} 个模型（端点报了容量与视觉的已一并填好）",
+	autoDiscoverFailed: "自动识别失败，可点「获取可用模型」重试",
 	usingDefaultModels: "正在使用 Pi SDK 默认模型 · {count} 个",
 	customModelsRequired: "至少添加一个可发送的模型 ID",
 	modelCatalogEmptyBuiltin: "未添加覆盖模型；模型选择器会继续实时使用 Pi SDK 的官方模型目录。",
 	modelCatalogEmptyCustom: "模型选择器中将不显示任何模型；请至少添加一个模型 ID。",
+	catalogOverrideHint: "与内置目录同 ID 的行只覆盖填写的字段，推理、价格、思考档位等沿用内置值。",
 	addModel: "添加模型",
 	modelId: "模型 ID",
 	createProvider: "创建提供方",
@@ -668,6 +739,9 @@ const zh = {
 	viewerNotInSnapshot: "此文件不在快照里（超过 1MB 或被排除），显示磁盘上的当前内容",
 	viewerLoadFailed: "加载失败",
 	viewerEmpty: "从左侧树里选择一个文件",
+	viewerOpenFolder: "打开文件夹预览",
+	viewerDirHint: "文件夹：共 {n} 项（已按 .gitignore 过滤）",
+	viewerDirMore: "加载更多",
 	// 流式输出
 	streamStyleTitle: "流式输出样式",
 	streamStyleDesc: "回答逐字出现的方式；改完立刻生效，正在流出的回答也一并换",
@@ -723,13 +797,42 @@ const en: Dict = {
 	capacityFields: "Capacities",
 	contextWindowLabel: "Context window",
 	maxTokensLabel: "Max output",
-	capacityHint: "blank = default · 256K / 2M accepted",
+	capacityHint: "blank = pi default (128K / 16K) · 256K / 2M accepted",
+	// model capability (vision) and config checkup
+	visionLabel: "Vision",
+	visionHint: "Whether this model accepts images. Left unset, pi treats it as text-only and replaces pasted screenshots with a placeholder line.",
+	visionUnset: "Not declared",
+	visionInputText: "Text",
+	visionInputTextHint: "pi always accepts text; whether images can be sent is decided by the Image box",
+	visionInputImage: "Image",
+	visionInputImageHint: "Checked means this model accepts images (written into models.json input); unchecked is treated as text-only",
+	visionOn: "Vision",
+	visionOff: "Text only",
+	imagePreview: "Click to preview",
+	imageViewer: "Image preview",
+	imageClearAll: "Clear images",
+	imagePrev: "Previous image",
+	imageNext: "Next image",
+	chipExpandMore: "Show {n} more",
+	chipCollapse: "Collapse",
+	modelAuditTitle: "{n} models don't declare their capabilities",
+	modelAuditDesc: "For fields left out, pi assumes a {ctx} context window, {out} max output and text-only input. Values the endpoint didn't report are never guessed: fill in the model's real limits (256K / 1M accepted). Saving only fills empty cells and never changes values already set.",
+	modelAuditColModel: "Model",
+	modelAuditColContext: "Context",
+	modelAuditColOutput: "Max output",
+	modelAuditColVision: "Vision",
+	modelAuditVisionQuick: "Set vision for all:",
+	modelAuditVisionOn: "All vision",
+	modelAuditVisionOff: "All text only",
+	modelAuditSave: "Save {n} models",
+	modelAuditSaveIdle: "Fill in, then save",
 	capacityInvalid: "Row {n}: capacity must be a number (K/M ok)",
 	discoverEmpty: "No recognizable models returned",
 	candidateTitle: "Found {count} models — pick ones to add",
 	candidateAll: "Select all",
 	candidatePicked: "{n} selected",
 	candidateAdopt: "Add {n}",
+	candidateInCatalog: "Already in catalog",
 	runningTool: "Running {name}…",
 	thinkingLive: "Thinking",
 	stepRunning: "Running",
@@ -812,6 +915,10 @@ const en: Dict = {
 	gitNoChanges: "Working tree clean",
 	gitStatusUnknown: "Git status unavailable; changes could not be determined",
 	gitRecentCommits: "Recent commits",
+	gitStagingArea: "Staging area",
+	gitWorkingTree: "Working tree",
+	gitGithubAccount: "GitHub account",
+	gitGithubUnknown: "No GitHub account detected",
 	toolNoOutput: "no output",
 	toolFailed: "failed",
 	thinkingClamped: "This model does not support \"{requested}\"; using \"{level}\" instead",
@@ -968,6 +1075,11 @@ const en: Dict = {
 	extensionStatus: "Extension status",
 	contextUsed: "context used",
 	contextUnknown: "Context usage unknown",
+	contextEstimated: "estimated context usage",
+	contextMeasured: "context usage",
+	contextBackground: "Main Conversation",
+	contextApiInput: "Last request input (API)",
+	contextChars: "chars",
 	contextSystem: "System prompt",
 	contextMessages: "Messages",
 	contextTools: "Tools",
@@ -1009,6 +1121,20 @@ const en: Dict = {
 	contextColdHint: "Session not started yet — System Prompt and tool definitions are counted once it starts",
 	contextCountNow: "Count now",
 	contextCounting: "Counting…",
+	contextWindowEdit: "Edit",
+	contextWindowBuiltin: "pi built-in catalog: {value}",
+	contextWindowUseBuiltin: "Use",
+	contextWindowScope: "Changes this model's global config (models.json); the pi terminal and other sessions use it too",
+	contextWindowOverBuiltin: "Larger than the built-in catalog: requests past the model's real limit are rejected upstream (400)",
+	contextWindowNoBuiltin: "Enter the model's real limit; a larger value gets requests rejected upstream (400)",
+	contextWindowInvalid: "Enter a positive integer; K / M allowed (e.g. 200K)",
+	// @ file references (type @ to open the picker)
+	mentionFiles: "@ reference files / folders",
+	mentionHint: "↑↓ select · Enter reference · Esc close",
+	mentionNoMatch: "No matching files",
+	mentionTruncated: "list truncated",
+	mentionDiskSource: "directory scan",
+	mentionDiskSourceHint: "This workspace is not a git repository (or git is missing): candidates come from a directory scan that skips node_modules and build output; in a git repository they follow .gitignore automatically",
 	turns: "turns",
 	toolCalls: "tool calls",
 	cost: "cost",
@@ -1094,6 +1220,21 @@ const en: Dict = {
 	drawerPanel: "Drawer",
 	drawerSections: "Drawer sections",
 	drawerPrompts: "Prompts",
+	sideChat: "Side chat",
+	sideChatModel: "Side-chat model",
+	sideChatFollowMain: "Follow main chat",
+	sideChatNew: "New side chat",
+	sideChatHistory: "Side-chat history",
+	sideChatMessages: "Side-chat messages",
+	sideChatQuestion: "Side-chat question",
+	sideChatPlaceholder: "Ask about anything you're reading…",
+	sideChatEmpty: "Ask as you read. Keep your place.",
+	sideChatHint: "Uses the main chat as background. Separate answers, saved only in this browser.",
+	sideChatPickSession: "Open a main conversation to start a side chat",
+	sideChatAnswering: "Answering…",
+	sideChatContextTruncated: "The background is long; this answer uses part of the main or side-chat history.",
+	sideChatLengthLimited: "This answer reached its output limit. You can ask it to continue.",
+	sideChatStorageFailed: "Your browser could not save the side chat. Keep a copy of anything you need.",
 	setPlugins: "Plugins",
 	setRemote: "Remote access",
 	remoteBuiltin: "Built-in plugin",
@@ -1185,10 +1326,14 @@ const en: Dict = {
 	allowPrivateDiscovery: "Allow local/private",
 	allowPrivateHint: "Check this when the model address points to loopback or a private network (Ollama, LM Studio, one-api gateways) to allow this probe; chat requests are not affected",
 	discoveringModels: "Fetching...",
+	autoDiscovering: "Detecting models…",
+	autoDiscovered: "Detected {n} models (capacity and vision filled in where the endpoint reports them)",
+	autoDiscoverFailed: "Automatic detection failed — use the button to retry",
 	usingDefaultModels: "Using Pi SDK default models · {count}",
 	customModelsRequired: "Add at least one model ID that can be sent",
 	modelCatalogEmptyBuiltin: "No model overrides. The model selector will continue using Pi SDK's live official catalog.",
 	modelCatalogEmptyCustom: "No models will appear in the selector. Add at least one model ID.",
+	catalogOverrideHint: "Rows matching a built-in model ID only override the fields you fill in; reasoning, pricing and thinking levels stay built-in.",
 	addModel: "Add model",
 	modelId: "Model ID",
 	createProvider: "Create provider",
@@ -1345,6 +1490,9 @@ const en: Dict = {
 	viewerNotInSnapshot: "Not in the snapshot (over 1MB or excluded); showing the current content from disk",
 	viewerLoadFailed: "Failed to load",
 	viewerEmpty: "Pick a file from the tree on the left",
+	viewerOpenFolder: "Preview folder",
+	viewerDirHint: "Folder: {n} items (ignored files hidden)",
+	viewerDirMore: "Load more",
 	// Streaming text
 	streamStyleTitle: "Streaming text",
 	streamStyleDesc: "How the answer appears while it arrives; applies to the stream already running",

package/src/lib/agent-manager.ts

@@ -25,17 +25,18 @@ import {
 	SessionManager,
 	createAgentSession,
 	getModelRuntime,
+	refreshModelRuntime,
 	openSessionManager,
 } from "./pi";
 import { createExtensionUiBridge, type ExtensionUiBridge } from "./extension-ui";
 import { createGrowthTracker, GROWTH_TRACKER_VERSION, type GrowthTracker } from "./growth-tracker";
 import { userTurnsFromEntries } from "./growth-turns";
+import { projectedContextUsage } from "./session/metrics";
 import { TrajLedger, buildTrajectoryFromEntries, toTrajTokens } from "./trajectory";
 import { startArchivedSessionRetention } from "./session-retention";
 import { sanitizeToolOutput } from "./text-sanitize";
 import { BoundaryError } from "./path-security";
-import { estimateTokensOf } from "./process-format";
-import { getPiSettings } from "./pi-settings";
+import { newSessionThinkingLevel } from "./thinking-level";
 import { getGrowthInitConsent, setGrowthInitConsent } from "./workspace-store";
 import { clampThinkingLevel, getSupportedThinkingLevels } from "@earendil-works/pi-ai";
 import type {
@@ -67,6 +68,8 @@ interface Managed {
 	viewerGraceUntil?: number;
 	/** 本轮已排队推送上下文分项（publishContextBreakdown 合并同一轮的多次触发） */
 	breakdownQueued?: boolean;
+	/** 首次运行后 pi 才会把系统分段持久化，届时用它把分类从差分换成投影口径，只换一次 */
+	breakdownProjected?: boolean;
 	buffer: { seq: number; json: string }[];
 	seq: number;
 	/**
@@ -349,36 +352,58 @@ function reapplyDesiredLevel(m: Managed, session: AgentSession): void {
 	if (target !== session.thinkingLevel) session.setThinkingLevel(target as never);
 }
 
+/**
+ * 新会话页选定模型后的思考档位：与界面显示的生效值同一规则（每模型覆盖 → 设置默认 → medium，再按模型能力钳制）。
+ * SDK 切换模型时会沿用会话原来的档位，新会话不能沿用：那个档位是 SDK 按它自己的初始模型定的。
+ */
+function applyNewSessionThinking(m: Managed, session: AgentSession): void {
+	const model = session.model;
+	if (!model) return;
+	const settings = getSettingsManager(m.cwd);
+	const level = newSessionThinkingLevel({
+		supported: session.getAvailableThinkingLevels() as unknown as string[],
+		modelOverride: settings.getModelThinkingLevel(model.provider, model.id),
+		defaultLevel: settings.getDefaultThinkingLevel(),
+	});
+	if (level !== session.thinkingLevel) session.setThinkingLevel(level as never);
+}
+
 function publishTraj(m: Managed): void {
 	for (const entry of m.ledger.drainDirty()) {
 		publish(m, { type: "traj", entry, ts: Date.now() });
 	}
 }
 
-function timingStats(m: Managed): Pick<WebStats, "llmMs" | "toolMs" | "ttftMs" | "ttftSteps" | "decodeMs" | "decodeTokens"> {
+function timingStats(m: Managed): Pick<WebStats, "llmMs" | "toolMs" | "ttftMs" | "ttftSteps" | "decodeMs" | "decodeTokens" | "llmTimedSteps" | "toolTimedCalls" | "decodeSteps"> {
 	let llmMs = 0;
 	let toolMs = 0;
 	let ttftMs = 0;
 	let ttftSteps = 0;
 	let decodeMs = 0;
 	let decodeTokens = 0;
+	let llmTimedSteps = 0;
+	let toolTimedCalls = 0;
+	let decodeSteps = 0;
 	for (const entry of m.ledger.entries) {
 		if (entry.kind === "tool") {
+			if (entry.timing?.durationMs !== undefined) toolTimedCalls++;
 			toolMs += Math.max(0, entry.timing?.durationMs ?? 0);
 			continue;
 		}
 		if (entry.kind !== "message") continue;
+		if (entry.timing?.durationMs !== undefined) llmTimedSteps++;
 		llmMs += Math.max(0, entry.timing?.durationMs ?? 0);
 		if (entry.timing?.ttftMs !== undefined) {
 			ttftMs += Math.max(0, entry.timing.ttftMs);
 			ttftSteps += 1;
 		}
-		if (entry.timing?.decodeMs !== undefined && entry.tokens?.output !== undefined) {
-			decodeMs += Math.max(0, entry.timing.decodeMs);
+		if ((entry.timing?.decodeMs ?? 0) > 0 && entry.tokens?.output !== undefined) {
+			decodeSteps++;
+			decodeMs += Math.max(0, entry.timing?.decodeMs ?? 0);
 			decodeTokens += Math.max(0, entry.tokens.output);
 		}
 	}
-	return { llmMs, toolMs, ttftMs, ttftSteps, decodeMs, decodeTokens };
+	return { llmMs, toolMs, ttftMs, ttftSteps, decodeMs, decodeTokens, llmTimedSteps, toolTimedCalls, decodeSteps };
 }
 
 function sessionStats(m: Managed): WebStats | null {
@@ -415,7 +440,7 @@ function coldSessionStats(m: Managed): WebStats {
 		cost += Number(usage.cost?.total ?? 0);
 	};
 	for (const entry of m.sm.getEntries() as any[]) {
-		if ((entry.type === "branch_summary" || entry.type === "compaction") && entry.usage) addUsage(entry.usage);
+		if ((entry.type === "usage" || entry.type === "branch_summary" || entry.type === "compaction") && entry.usage) addUsage(entry.usage);
 		if (entry.type !== "message" || !entry.message) continue;
 		totalMessages += 1;
 		const message = entry.message;
@@ -497,7 +522,14 @@ function translate(m: Managed, evt: AgentSessionEvent): void {
 			// the browser receives the durable entry ID without reordering later tool events.
 			deferPublish(m, () => {
 				const msg = projectMessage(m, raw);
-				if (msg.role === "assistant") msg.endedAt = now;
+				if (msg.role === "assistant") {
+					msg.endedAt = now;
+					// 首次回复结束后 pi 已把系统分段写进会话文件，此时分类能读到投影（比 systemPrompt 差分精确）
+					if (!m.breakdownProjected) {
+						m.breakdownProjected = true;
+						publishContextBreakdown(m);
+					}
+				}
 				const usage = usageEvent(m);
 				return usage
 					? [{ type: "message", message: msg, phase: "end", ts: now }, usage]
@@ -597,8 +629,8 @@ function translate(m: Managed, evt: AgentSessionEvent): void {
 			else publish(m, { type: "model", thinkingLevel: (evt as any).level, ts: now });
 			break;
 		case "auto_retry_start":
-			// 重试是流程内通知：发一条结构化 retry 事件（界面聚合成「重试 n/max」一行），
-			// 不再走 error 通道，免得每次尝试都多出一条报错。
+			// 重试是流程内通知:发一条结构化 retry 事件(界面聚合成「重试 n/max」一行),
+			// 不再走 error 通道,免得每次尝试都多出一条报错。
 			publish(m, {
 				type: "retry",
 				attempt: Number((evt as any).attempt) || 0,
@@ -607,6 +639,13 @@ function translate(m: Managed, evt: AgentSessionEvent): void {
 				ts: now,
 			});
 			break;
+		case "auto_retry_end":
+			// 重试收尾:SDK 成功/最终失败/被中止都会发这条,而重试成功后的这轮对话
+			// agent_settled 之前界面上还留着「重试 1/5 terminated」那一行(用户看到的
+			// 就是「已经恢复了还显示错误」)。这里显式发一条清除事件,重试区间一结束
+			// 就收掉这行;真失败时紧随其后的 assistant error 消息会照常出错误块。
+			publish(m, { type: "retry_end", attempt: Number((evt as any).attempt) || 0, ts: now });
+			break;
 		default:
 			break;
 	}
@@ -742,13 +781,9 @@ const SKILLS_SECTION_START = "The following skills provide specialized instructi
 const SKILLS_SECTION_END = "</available_skills>";
 
 /**
- * 上下文占用的服务端估算分项。systemPrompt 是整段拼接（正文 + memory + skills + 工具列表），
- * 所以 Memory / Skills 从资源加载器与 systemPrompt 的资源段差分得出：
- * SDK 的拼接顺序固定为 正文 → context files → skills → 工具列表，按内容定位切块。
- *
- * 冷会话（只浏览历史、不冷启动 agent）拿不到组装后的 system prompt 和工具定义，这三项为 0、标 cold；
- * Memory / Skills / Compacted / Buffer 不依赖 agent，照常从发现缓存、会话文件和设置里算，
- * 不能因为没启动就整块归零（那样弹窗只剩消息几类）。
+ * 当前上下文的字符分类，优先读 SDK 投影中的系统分段、工具 schema 与摘要。
+ * 旧客户端仍可读取字符/4 的估算字段；新客户端仅将字符数作为分类数据。
+ * 没有持久化系统消息且尚未启动的会话，资源来自发现缓存并标记 cold。
  */
 async function buildContextBreakdown(m: Managed, resources: ContextResource[]): Promise<ContextBreakdown> {
 	const session = m.session;
@@ -758,64 +793,56 @@ async function buildContextBreakdown(m: Managed, resources: ContextResource[]):
 	};
 	// Memory：AGENTS.md 等注入文件（contextResources 的 project 部分）
 	const memoryText = resources.filter((r) => r.source === "project").map((r) => r.content).join("\n\n");
-	breakdown.memory = estimateTokensOf(memoryText);
-	if (session) {
-		try {
-			const prompt = session.systemPrompt ?? "";
-			// Skills：按 SDK 的 formatSkillsForPrompt 段落定位。该段落没有 Markdown 标题（以
-			// "The following skills provide specialized instructions…" 开头、</available_skills> 结尾），
-			// 以前按 "## Skills" 找永远找不到，技能段一直是 0。
-			try {
-				if (session.resourceLoader.getSkills().skills.length && prompt) {
-					const start = prompt.indexOf(SKILLS_SECTION_START);
-					const end = prompt.indexOf(SKILLS_SECTION_END);
-					if (start >= 0 && end > start) breakdown.skills = estimateTokensOf(prompt.slice(start, end + SKILLS_SECTION_END.length));
-				}
-			} catch {
-				/* loader 不可用则跳过 */
+	// 字符分项只读取当前投影中的内容。历史压缩摘要、未激活工具不混入当前统计。
+	const characters = { systemPrompt: 0, systemTools: 0, customTools: 0, memory: 0, skills: 0, compacted: 0 };
+	try {
+		const messages = m.sm.buildSessionProjection().messages as any[];
+		const system = messages.findLast((message) => message.role === "system");
+		if (system) {
+			const contentChars = typeof system.content === "string" ? system.content.length : textOf(system.content).length;
+			characters.memory = system.sections?.project_context?.length ?? 0;
+			characters.skills = system.sections?.skills?.length ?? 0;
+			characters.systemPrompt = contentChars + Object.entries(system.sections ?? {})
+				.filter(([key]) => key !== "project_context" && key !== "skills")
+				.reduce((sum, [, value]) => sum + (typeof value === "string" ? value.length : 0), 0);
+			const builtinNames = new Set(session
+				? session.getAllTools().filter((tool: any) => tool.sourceInfo?.source === "builtin").map((tool) => tool.name)
+				: ["read", "bash", "edit", "write", "grep", "find", "ls"]);
+			for (const tool of system.toolsAdded ?? []) {
+				characters[builtinNames.has(tool.name) ? "systemTools" : "customTools"] += JSON.stringify(tool).length;
 			}
-			// System Prompt 正文 = 全文减去资源段与 skills 段的近似差分，下限 0
-			breakdown.systemPrompt = Math.max(0, estimateTokensOf(prompt) - breakdown.memory - breakdown.skills);
-			// 工具：builtin 与扩展注册的分开统计（schema + description + guidelines 序列化估算）
-			for (const tool of session.getAllTools()) {
-				const size = estimateTokensOf(JSON.stringify({
-					name: tool.name,
-					description: tool.description,
-					parameters: tool.parameters,
-					promptGuidelines: tool.promptGuidelines,
-				}));
-				if ((tool as { sourceInfo?: { source?: string } }).sourceInfo?.source === "builtin") breakdown.systemTools += size;
-				else breakdown.customTools += size;
+			delete breakdown.cold;
+		} else if (session) {
+			const prompt = session.systemPrompt ?? "";
+			characters.memory = memoryText.length;
+			const start = prompt.indexOf(SKILLS_SECTION_START);
+			const end = prompt.indexOf(SKILLS_SECTION_END);
+			characters.skills = start >= 0 && end > start ? prompt.slice(start, end + SKILLS_SECTION_END.length).length : 0;
+			characters.systemPrompt = Math.max(0, prompt.length - characters.memory - characters.skills);
+			const active = new Set(session.getActiveToolNames());
+			for (const tool of session.getAllTools().filter((tool) => active.has(tool.name))) {
+				const size = JSON.stringify({ name: tool.name, description: tool.description, parameters: tool.parameters }).length;
+				characters[(tool as any).sourceInfo?.source === "builtin" ? "systemTools" : "customTools"] += size;
 			}
-		} catch {
-			/* systemPrompt / getAllTools 不可用则保持 0 */
-		}
-	} else {
-		// 冷会话：技能段用发现缓存里的清单按 SDK 同一个格式化函数估算（启动后改用实际 prompt 里的那段）
-		try {
+		} else {
+			// 初次启动前只能读到待注入的资源，界面保留 cold 提示。
+			characters.memory = memoryText.length;
 			const loader = peekResourceLoader(m.cwd) ?? getResourceLoader(m.cwd);
-			breakdown.skills = estimateTokensOf(formatSkillsForPrompt(loader.getSkills().skills).trim());
-		} catch {
-			/* 发现缓存还没好则跳过 */
+			characters.skills = formatSkillsForPrompt(loader.getSkills().skills).trim().length;
 		}
-	}
-	// 最新一次压缩的摘要（会话文件里就有，冷热都能算）
-	try {
-		const entries = m.sm.getEntries() as unknown as { type?: string; summary?: string }[];
-		for (let i = entries.length - 1; i >= 0; i -= 1) {
-			const entry = entries[i];
-			if (entry?.type === "compaction" && typeof entry.summary === "string") {
-				breakdown.compacted = estimateTokensOf(entry.summary);
-				break;
-			}
+		for (const message of messages) {
+			if (message.role === "compactionSummary" || message.role === "branchSummary") characters.compacted += message.summary?.length ?? 0;
 		}
+		breakdown.characters = characters;
+		// 保留旧客户端的估算字段，当前客户端直接展示实际字符数。
+		for (const key of Object.keys(characters) as Array<keyof typeof characters>) breakdown[key] = Math.ceil(characters[key] / 4);
 	} catch {
-		/* entries 不可用则跳过 */
+		/* 投影不可用时不给字符分项，不能从旧 token 估算反推字符数。 */
 	}
-	// auto-compact 预留 = compaction.reserveTokens（启用时）
+	// auto-compact 预留读本工作区、当前模型实际使用的设置。
 	try {
-		const settings = await getPiSettings();
-		if (settings.compaction.enabled) breakdown.autoCompactBuffer = settings.compaction.reserveTokens;
+		const compaction = getSettingsManager(m.cwd).getCompactionSettings(session?.model);
+		if (compaction.enabled) breakdown.autoCompactBuffer = compaction.reserveTokens;
 	} catch {
 		/* 设置读取失败则保持 0 */
 	}
@@ -1208,11 +1235,7 @@ async function reloadManagedResources(m: Managed): Promise<boolean> {
 	m.resourceReloadPending = false;
 	try {
 		await m.session.reload();
-		// SDK 的 reload 不会换当前模型对象：models.json 改了 baseUrl / contextWindow 后，
-		// 会话手里还是旧对象。按 provider/id 从（已就地刷新的）共享 runtime 取一次新的。
-		const current = m.session.model;
-		const fresh = current && m.session.modelRuntime.getModel(current.provider, current.id);
-		if (fresh && fresh !== current) m.session.agent.state.model = fresh;
+		swapCurrentModelObject(m.session);
 		return true;
 	} catch (err) {
 		publish(m, { type: "error", message: `重载扩展失败：${String((err as Error)?.message ?? err)}`, ts: Date.now() });
@@ -1222,6 +1245,32 @@ async function reloadManagedResources(m: Managed): Promise<boolean> {
 	}
 }
 
+/**
+ * SDK 的 reload 不会换当前模型对象：models.json 改了 baseUrl / contextWindow 后，会话手里还是旧对象。
+ * 按 provider/id 从（已就地刷新的）共享 runtime 取一次新的。
+ */
+function swapCurrentModelObject(session: AgentSession): void {
+	const current = session.model;
+	const fresh = current && session.modelRuntime.getModel(current.provider, current.id);
+	if (fresh && fresh !== current) session.agent.state.model = fresh;
+}
+
+/**
+ * 共享目录刷新后，给正在用这个模型的会话换上新模型对象（上下文窗口等字段跟着变），再推一次用量让圆环当场更新。
+ * 运行中的会话也换：SDK 在扩展注册 provider 时同样会中途替换 agent.state.model，已发出的请求不受影响。
+ */
+export function refreshSessionModels(provider: string, id: string): number {
+	let n = 0;
+	for (const m of sessions.values()) {
+		const current = m.session?.model;
+		if (!m.session || current?.provider !== provider || current.id !== id) continue;
+		swapCurrentModelObject(m.session);
+		publishUsage(m);
+		n += 1;
+	}
+	return n;
+}
+
 export async function reloadSessionsForCwd(cwd?: string): Promise<number> {
 	const key = (p: string) => (process.platform === "win32" ? p.replace(/\\/g, "/").toLowerCase() : p.replace(/\\/g, "/"));
 	let n = 0;
@@ -1325,42 +1374,13 @@ export async function buildSnapshot(m: Managed): Promise<WebSnapshot> {
 		const usage = session.getContextUsage() ?? null;
 		contextUsage = usage ? { tokens: usage.tokens, contextWindow: usage.contextWindow, percent: usage.percent } : null;
 	} else if (model) {
-		// 冷会话没有 SDK 实例：用最后一轮 assistant 用量（input+cache+output ≈ 该轮后的
-		// 上下文规模）除以模型窗口估算，避免上下文计量永远显示 0% / —
+		// 冷会话也使用当前分支投影，包含最后一次 API 用量之后的用户消息和工具输出。
 		try {
 			const rt = await getModelRuntime();
 			const modelMeta = (rt.getModels(model.provider) as any[]).find((x) => String(x.id) === model.id);
-			const window = Number(modelMeta?.contextWindow ?? 0);
-			let lastTurnTokens = 0;
-			// 与 SDK getContextUsage 同一条规则：最近一次压缩之后还没有新的助手回复时，上下文大小未知（不能拿压缩前的用量充数）
-			const entries = m.sm.getEntries() as Array<{ type?: string; message?: { role?: string; stopReason?: string } }>;
-			let compactedSinceLastAssistant = false;
-			for (let i = entries.length - 1; i >= 0; i--) {
-				const e = entries[i];
-				if (e.type === "compaction" || e.type === "branch_summary") {
-					compactedSinceLastAssistant = true;
-					break;
-				}
-				if (e.type === "message" && e.message?.role === "assistant" && e.message.stopReason !== "error" && e.message.stopReason !== "aborted") break;
-			}
-			for (let i = messages.length - 1; i >= 0 && !compactedSinceLastAssistant; i--) {
-				const msg = messages[i];
-				if (msg.role === "assistant" && msg.usage) {
-					lastTurnTokens =
-						Number(msg.usage.input ?? 0) +
-						Number(msg.usage.cacheRead ?? 0) +
-						Number(msg.usage.cacheWrite ?? 0) +
-						Number(msg.usage.output ?? 0);
-					break;
-				}
-			}
-			if (window > 0 && lastTurnTokens > 0) {
-				contextUsage = { tokens: lastTurnTokens, contextWindow: window, percent: (lastTurnTokens / window) * 100 };
-			} else if (window > 0) {
-				contextUsage = { tokens: null, contextWindow: window, percent: null };
-			}
+			contextUsage = projectedContextUsage(m.sm.buildSessionProjection(), m.sm.getBranch(), Number(modelMeta?.contextWindow ?? 0));
 		} catch {
-			/* 模型目录不可用则维持 null（UI 显示 0%） */
+			/* 模型或投影不可用时用量未知。 */
 		}
 	}
 
@@ -1600,6 +1620,8 @@ export async function execute(m: Managed, cmd: AgentCommand): Promise<CommandRes
 				if (!model) return { ok: false, error: `unknown model ${cmd.provider}/${cmd.modelId}` };
 				await session.setModel(model, { persist: false });
 				reapplyDesiredLevel(m, session);
+				// 新会话页选的模型（fresh）：思考档位按新会话的规则定；之后显式选的档位仍会覆盖
+				if (cmd.fresh === true && !m.desiredThinkingLevel) applyNewSessionThinking(m, session);
 				publishModelState(m, session);
 				return { ok: true, data: { thinkingLevel: session.thinkingLevel, thinkingLevels: session.getAvailableThinkingLevels() } };
 			}
@@ -1712,8 +1734,17 @@ export async function execute(m: Managed, cmd: AgentCommand): Promise<CommandRes
 			}
 			case "reload": {
 				if (isManagedBusy(m)) return { ok: false, error: "session is busy" };
+				// /reload：先就地刷新共享的模型目录（models.json / auth.json 可能是手改的）；其它会话手里的
+				// 模型对象跟着换新，忙着的等本轮结束再换；本会话重载资源后再换一次
+				await refreshModelRuntime();
+				for (const other of sessions.values()) {
+					if (other === m || !other.session) continue;
+					if (isManagedBusy(other)) other.resourceReloadPending = true;
+					else swapCurrentModelObject(other.session);
+				}
 				const session = await ensureSession(m);
 				await session.reload();
+				swapCurrentModelObject(session);
 				return { ok: true };
 			}
 			case "fork": {