pi2web 0.2.8 → 0.2.12

package/README.md

@@ -223,8 +223,9 @@ Web 端能力：
 
 - Workspace 添加、目录浏览、文本文件分页及二进制提示。
 - Workspace 目录右键在交互式 Terminal 中打开；Terminal 与 Agent 显示在同一列表，支持刷新重连、清屏和关闭。
-- 创建/恢复 Session、Agent 列表、停止与 abort；Agent 右键 Archive。
-- Prompt、steer、follow-up 和流式对话。
+- 新建 Agent 时始终创建新 Session；已有 Agent 可在刷新或重启后继续使用。通过 Session 文件恢复对话暂仅支持 API，不在新建弹窗中提供；Agent 列表支持停止与 abort，右键 Archive。
+- Prompt、steer、follow-up 和流式对话。HTTP / WebSocket 发送命令在 Pi 校验并接收消息后即确认，不等待整个回合结束；执行进度继续通过 WebSocket 同步。
+- 手机切后台或网络中断导致发送确认丢失时，显示「送达状态未知」并保留原文供核对，不自动重发或放回输入框。提示在同一页面内切换会话、重连后仍保留，可手动关闭；刷新页面不保留。明确被服务端拒绝的消息仍恢复为草稿。
 - 对话区只保留用户输入、Agent 输出、Thinking、Extension 对话框和错误；Tool call、Session bash 输出、Retry、Summarization retry 等过程事件自动折叠成一张"活动"卡片，卡片标题实时显示最新一项的名称和时间，展开后逐条查看，再点击单条查看内容。
 - 同一次 Tool call 的 start / update / end 合并为一张卡片（参数 + 输出 + 结果），结束后显示 Pi 记录的执行耗时。
 - Tool 自己发起的调用（codemode 脚本、MCP 批量调用）带 `parentToolCallId`，会嵌套显示在发起它的那张卡片里。
@@ -362,6 +363,7 @@ Web 看板在 `/collab.html`（首页顶部"协作"入口）：会话列表与
 - `GET /api/v1/agents/:id/mcp` 返回该 Agent cwd 下的 MCP 服务器配置（全局 `mcp.json` 与受信任项目的 `.pi/mcp.json`），`POST` 用 `{name, enabled}` 启用或停用。只有项目被 Pi 显式信任时才会写入项目级覆盖，否则写全局文件。修改在 Agent 下次加载时生效。
 - `GET /api/v1/agents/:id/commands` 返回该 Agent cwd 下可用的 prompt template 和 skill（`{name, description, argumentHint, source}`）。`prompt` 接口本身会展开 `/命令`、skill 和 prompt template。
 - `POST /api/v1/agents/:id/bash` 在 Agent Session 内执行命令（`{command, excludeFromContext}`），输出通过 `bash_execution_update` 事件流式推送并记录到 Session；`POST /api/v1/agents/:id/bash-abort` 取消。
+- `POST /api/v1/agents/:id/pause` 安全暂停：对话不再前进——正在执行的工具跑完、结果写入 Session 后，不再产生新的模型回复或工具调用（Agent 循环的下一轮，以及自动重试、溢出压缩后的重试、排队消息、扩展 `agent_before_settle` 要求的续跑都会被拦下；正在等待的重试退避会被立即取消）。Session 以工具结果（或用户消息）结尾，重启后也不丢失，之后发送新消息即从这里继续；暂停时排队的消息留在队列里，随下一次 prompt 送达。不拦截不推进对话的整理性模型调用（压缩摘要、提示缓存保温；保温仅在 `cacheWarming: idle` 时于暂停后继续，最长 30 分钟）。进度通过 `safe_pause` 事件推送：`requested`、`reached`（已拦下一次模型请求，正在收尾）、`paused`、`cancelled`、`finished`（未到暂停点运行就自行结束）；活动状态带 `pauseRequested {since, reached?}`。被暂停运行的 `agent_settled` 带 `paused: true`，邮件通知按“已中断”处理、语音不播报。`DELETE` 同一路径撤回暂停：已到暂停点、运行仍在收尾时（`reached` 之后、`paused` 之前）返回 `PAUSE_ALREADY_REACHED`；没有待生效的暂停时（包括运行已结束）为无操作并返回成功。撤回是否赶上以 `safe_pause` 事件（`cancelled` 或 `paused`）为准。
 - WebSocket `/api/v1/ws`（协议 v2，不做事件回放）：`subscribe_all` 先下发每个 Agent 的当前状态，再推送所有 Agent 的概览事件（状态、活动、队列、对话框、通知）；`focus` 选定正在查看的 Agent，只为它推送流式文本、时间线条目变化和用户 Shell 输出；`watch` 订阅已展开的运行中工具的输出；`subscribe_collab` 推送协作事件。
 - 会话内容通过 HTTP 时间线读取：`GET /api/v1/agents/:id/timeline?limit&before` 返回条目摘要（正文较小或需要阅读的条目内联），`GET /api/v1/agents/:id/sessions/:sessionId/items/:itemId` 按需返回完整正文；已完成条目永不改变，响应带 `immutable` 缓存头，浏览器可永久缓存。
 

package/dist/activity.js

@@ -22,6 +22,7 @@ const num = (value) => (typeof value === 'number' && Number.isFinite(value) ? va
 export class ActivityTracker {
     now;
     state;
+    lastMetricsPublish = 0;
     constructor(now = Date.now) {
         this.now = now;
         this.state = { run: null, phase: { kind: 'idle', since: now() }, bash: [], dialogs: [] };
@@ -32,7 +33,7 @@ export class ActivityTracker {
     /** Forget everything, e.g. when the backend is unloaded. Returns whether anything changed. */
     reset() {
         const s = this.state;
-        if (!s.run && s.phase.kind === 'idle' && !s.bash.length && !s.dialogs.length)
+        if (!s.run && s.phase.kind === 'idle' && !s.bash.length && !s.dialogs.length && !s.pauseRequested)
             return false;
         this.state = { run: null, phase: { kind: 'idle', since: this.now() }, bash: [], dialogs: [] };
         return true;
@@ -51,26 +52,52 @@ export class ActivityTracker {
                 const startedRun = !s.run;
                 if (startedRun)
                     s.run = { startedAt: this.now() };
+                if (startedRun)
+                    s.metrics = { turnStartedAt: this.now(), chars: 0, estimatedTokens: 0 };
                 return this.setPhase({ kind: 'waiting_llm' }) || startedRun;
             }
             case 'turn_start':
-                return this.setPhase({ kind: 'waiting_llm' });
+                s.metrics = { turnStartedAt: this.now(), chars: 0, estimatedTokens: 0 };
+                this.lastMetricsPublish = 0;
+                this.setPhase({ kind: 'waiting_llm' });
+                // Even when the last tool already moved us to waiting_llm, clients need the cleared metrics.
+                return true;
             case 'message_update': {
                 const update = e.assistantMessageEvent, type = str(update?.type);
+                const at = this.now();
+                const firstResponse = !!s.metrics &&
+                    s.metrics.firstResponseAt === undefined &&
+                    (type === 'thinking_delta' || type === 'text_delta' || type === 'toolcall_delta');
+                if (firstResponse)
+                    s.metrics.firstResponseAt = at;
+                let changed, lateName = false;
                 if (type === 'thinking_start' || type === 'thinking_delta')
-                    return this.setPhase({ kind: 'thinking' });
-                if (type === 'text_start' || type === 'text_delta')
-                    return this.setPhase({ kind: 'generating' });
-                if (type === 'toolcall_start' || type === 'toolcall_delta') {
+                    changed = this.setPhase({ kind: 'thinking' });
+                else if (type === 'text_start' || type === 'text_delta')
+                    changed = this.setPhase({ kind: 'generating' });
+                else if (type === 'toolcall_start' || type === 'toolcall_delta') {
                     const toolName = str(update.partial?.content?.[update.contentIndex]?.name);
                     // The name may only arrive with a later delta: fill it in without restarting the timer.
                     if (s.phase.kind === 'preparing_tool' && !s.phase.toolName && toolName) {
                         s.phase.toolName = toolName;
-                        return true;
+                        changed = lateName = true;
                     }
-                    return this.setPhase({ kind: 'preparing_tool', ...(toolName ? { toolName } : {}) });
+                    else
+                        changed = this.setPhase({ kind: 'preparing_tool', ...(toolName ? { toolName } : {}) });
                 }
-                return false;
+                else
+                    return false;
+                // Throughput is per stream: each content block, e.g. each of two consecutive `write` calls,
+                // and each phase, so thinking does not dilute the text rate.
+                // A late tool name only labels the stream already running; it does not start a new one.
+                const cleared = ((changed && !lateName) || type.endsWith('_start')) && this.clearStream();
+                const counted = this.countDelta(update);
+                // Publish a stream's first count at once; later ones at most once a second.
+                if (counted && (counted === 'first' || at - this.lastMetricsPublish >= 1000)) {
+                    this.lastMetricsPublish = at;
+                    return true;
+                }
+                return changed || cleared || firstResponse;
             }
             case 'tool_execution_start': {
                 // Tools called by other tools run inside their caller's time; show only top-level calls.
@@ -143,6 +170,23 @@ export class ActivityTracker {
                 s.dialogs.push({ requestId, title: str(e.title) ?? str(e.message), since: this.now() });
                 return true;
             }
+            case 'safe_pause':
+                if (e.state === 'requested') {
+                    if (s.pauseRequested)
+                        return false;
+                    s.pauseRequested = { since: this.now() };
+                    return true;
+                }
+                if (e.state === 'reached') {
+                    if (s.pauseRequested?.reached)
+                        return false;
+                    s.pauseRequested = { since: s.pauseRequested?.since ?? this.now(), reached: true };
+                    return true;
+                }
+                if (!s.pauseRequested)
+                    return false;
+                delete s.pauseRequested;
+                return true;
             case 'extension_ui_response': {
                 const before = s.dialogs.length;
                 s.dialogs = s.dialogs.filter(d => d.requestId !== e.requestId);
@@ -152,9 +196,39 @@ export class ActivityTracker {
                 return false;
         }
     }
+    /** Adds a thinking, text or tool call delta to the current stream's estimate. Returns whether it counted. */
+    /** Restarts the current stream's counters. Returns whether there was anything to clear. */
+    clearStream() {
+        const metrics = this.state.metrics;
+        if (!metrics || (metrics.streamStartedAt === undefined && !metrics.chars))
+            return false;
+        delete metrics.streamStartedAt;
+        metrics.chars = 0;
+        metrics.estimatedTokens = 0;
+        return true;
+    }
+    /**
+     * Adds a thinking, text or tool call delta to the current stream's estimate. Returns `'first'` for
+     * the stream's first counted delta, `true` for later ones and `false` if nothing was counted.
+     */
+    countDelta(update) {
+        const metrics = this.state.metrics, delta = update?.delta;
+        if (!metrics || !String(update?.type).endsWith('_delta') || typeof delta !== 'string' || !delta)
+            return false;
+        const first = metrics.streamStartedAt === undefined;
+        if (first)
+            metrics.streamStartedAt = this.now();
+        // Approximate Latin/code text at 4 chars/token, CJK at 1.5 chars/token.
+        const cjk = delta.match(/[\u3400-\u9fff\u3040-\u30ff\uac00-\ud7af]/gu)?.length ?? 0;
+        const chars = [...delta].length;
+        metrics.chars += chars;
+        metrics.estimatedTokens += (chars - cjk) / 4 + cjk / 1.5;
+        return first ? 'first' : true;
+    }
     endRun(settled = false) {
         const hadRun = !!this.state.run;
         this.state.run = null;
+        delete this.state.metrics;
         // Compaction may continue after the run has ended; agent_settled closes it.
         if (this.state.phase.kind === 'compacting' && !settled)
             return hadRun;

package/dist/agents.js

@@ -57,6 +57,7 @@ export class MockBackend {
     entries = [];
     timer;
     queue = { steering: [], followUp: [] };
+    pauseRequested = false;
     constructor(agentId, cwd, sessionId = randomUUID(), options = {}) {
         this.options = options;
         this.state = { agentId, sessionId, cwd, status: 'idle' };
@@ -82,11 +83,12 @@ export class MockBackend {
             this.timer = setTimeout(r, 0);
         });
     }
-    async prompt(message) {
+    async prompt(message, onAccepted) {
         if (this.state.status === 'streaming')
             throw Object.assign(new Error('Agent is busy'), { code: 'AGENT_BUSY' });
         this.state.status = 'streaming';
         this.emit({ type: 'agent_start', message });
+        onAccepted?.();
         await this.runTurn();
         const user = { role: 'user', content: message };
         this.emit({ type: 'message_end', message: user });
@@ -98,6 +100,25 @@ export class MockBackend {
         this.append(assistant);
         this.state.status = 'idle';
         this.emit({ type: 'agent_end' });
+        // A mock run is a single final answer, so a pending pause never finds a safe point.
+        if (this.pauseRequested) {
+            this.pauseRequested = false;
+            this.emit({ type: 'safe_pause', state: 'finished' });
+        }
+    }
+    async requestPause() {
+        if (this.state.status !== 'streaming')
+            throw Object.assign(new Error('Agent is not running'), { code: 'AGENT_NOT_RUNNING' });
+        if (this.pauseRequested)
+            return;
+        this.pauseRequested = true;
+        this.emit({ type: 'safe_pause', state: 'requested' });
+    }
+    async cancelPause() {
+        if (!this.pauseRequested)
+            return;
+        this.pauseRequested = false;
+        this.emit({ type: 'safe_pause', state: 'cancelled' });
     }
     async steer(message) {
         const queued = this.state.status === 'streaming';
@@ -146,6 +167,10 @@ export class MockBackend {
                 clearTimeout(this.timer);
             this.state.status = 'idle';
             this.emit({ type: 'agent_end', aborted: true });
+            if (this.pauseRequested) {
+                this.pauseRequested = false;
+                this.emit({ type: 'safe_pause', state: 'finished' });
+            }
         }
     }
     async getState() {
@@ -471,7 +496,29 @@ export class AgentManager {
         })();
         return entry.loading;
     }
-    async command(id, kind, message = '') {
+    /** Transport acknowledgement: an accepted prompt keeps running under the normal write admission. */
+    submitCommand(id, kind, message = '') {
+        return new Promise((resolve, reject) => {
+            let accepted = false;
+            void this.command(id, kind, message, () => {
+                accepted = true;
+                resolve();
+            }).then(resolve, error => {
+                if (!accepted)
+                    return reject(error);
+                // The sender already has its acknowledgement. Report execution errors through the existing
+                // event stream rather than turning a delivered prompt into a failed send.
+                const entry = this.agents.get(id);
+                if (entry)
+                    this.publish(entry, {
+                        type: 'extension_ui_notify',
+                        notificationType: 'error',
+                        message: error instanceof Error ? error.message : String(error)
+                    });
+            });
+        });
+    }
+    async command(id, kind, message = '', onAccepted) {
         if (kind === 'abort') {
             const entry = this.get(id);
             if (!entry.backend)
@@ -489,7 +536,9 @@ export class AgentManager {
             if (effective === 'prompt')
                 entry.promptPending = true;
             try {
-                return await backend[effective === 'follow-up' ? 'followUp' : effective](message);
+                return effective === 'prompt'
+                    ? await backend.prompt(message, onAccepted)
+                    : await backend[effective === 'follow-up' ? 'followUp' : 'steer'](message);
             }
             finally {
                 if (effective === 'prompt' && entry.backend === backend)
@@ -734,6 +783,16 @@ export class AgentManager {
     async runBash(id, command, excludeFromContext = false) {
         return this.writeSession(id, backend => backend.runBash(command, excludeFromContext));
     }
+    /** Stop the running turn at its next safe point: after its tools, before the next model request. */
+    async requestPause(id) {
+        const entry = this.get(id);
+        if (!entry.backend)
+            throw Object.assign(new Error('Agent is not running'), { code: 'AGENT_NOT_RUNNING' });
+        return entry.backend.requestPause();
+    }
+    async cancelPause(id) {
+        return this.get(id).backend?.cancelPause();
+    }
     async abortBash(id) {
         const entry = this.get(id);
         if (entry.loading)

package/dist/http/agent-routes.js

@@ -155,6 +155,14 @@ export async function handleAgentRoutes(ctx, req, res, u, parts) {
             ctx.metadata.scheduleAgent(ctx.agents.get(id).record);
             return send(200, { data: result });
         }
+        // Safe pause: POST asks the running turn to stop before its next model request; DELETE withdraws that.
+        if (parts[4] === 'pause' && (req.method === 'POST' || req.method === 'DELETE')) {
+            if (req.method === 'POST')
+                await ctx.agents.requestPause(id);
+            else
+                await ctx.agents.cancelPause(id);
+            return send(200, { data: { success: true } });
+        }
         if (req.method === 'POST' && parts[4] === 'bash-abort') {
             await ctx.agents.abortBash(id);
             return send(200, { data: { aborted: true } });
@@ -210,7 +218,7 @@ export async function handleAgentRoutes(ctx, req, res, u, parts) {
                 ctx.voice?.recordUserPrompt(id, String(b.message ?? ''));
                 ctx.sessionNamer?.recordUserPrompt(id, String(b.message ?? ''));
             }
-            const disposition = await ctx.agents.command(id, parts[4], b.message ?? '');
+            const disposition = await ctx.agents.submitCommand(id, parts[4], b.message ?? '');
             const record = ctx.agents.get(id).record;
             ctx.metadata.scheduleAgent(record);
             if (parts[4] === 'prompt')

package/dist/mail-notifier.js

@@ -76,8 +76,13 @@ export class MailNotifier {
         this.lastAssistant.delete(agentId);
         if (!this.settings.enabled)
             return;
-        // Pi 0.99+ reports whether the run was cancelled; a cancelled run is not a completed task.
-        this.queue.push({ agentId, finalOutput, context, aborted: event.aborted === true });
+        // Pi 0.99+ reports whether the run was cancelled; a cancelled or safely paused run is not a completed task.
+        this.queue.push({
+            agentId,
+            finalOutput,
+            context,
+            aborted: event.aborted === true || event.paused === true
+        });
         if (!this.timer)
             this.scheduleFlush();
     }

package/dist/safe-pause.js

@@ -0,0 +1,100 @@
+const error = (message, code) => Object.assign(new Error(message), { code });
+/**
+ * Stops a Pi session before the conversation advances: no further model response or tool call
+ * once the pause is reached. Tools that are running finish and their results are recorded first,
+ * so the transcript ends with tool results (or the user's message) and nothing is lost.
+ *
+ * A Pi run asks the model again through exactly two entries, and each one is gated:
+ * - Gate A, the agent loop's `finishTurn`: the next turn inside the loop.
+ * - Gate B, `agent.continue()`: every continuation the session's run controller starts after the
+ *   loop ends, i.e. auto-retry, overflow-compaction retry, queued messages and extension
+ *   `agent_before_settle` continuations.
+ * Neither gate depends on event timing. Housekeeping model calls that do not advance the
+ * conversation (compaction summaries, prompt-cache warming) are not gated.
+ */
+export class SafePause {
+    session;
+    emit;
+    state = 'none';
+    /** Set for the agent_settled of a run that ended at the safe point. */
+    settledPaused = false;
+    constructor(session, emit) {
+        this.session = session;
+        this.emit = emit;
+        const agent = session.agent;
+        // Gate A. AgentSession installed its own finishTurn (extension turn_end boundaries); run it first.
+        const finishTurn = agent.finishTurn;
+        agent.finishTurn = async (turn, signal) => {
+            const decision = (await finishTurn?.(turn, signal)) ?? undefined;
+            if (this.state === 'none' || decision?.action === 'end')
+                return decision;
+            // An error or aborted turn ends the loop anyway; a retry has to pass gate B.
+            if (turn.message.stopReason === 'error' || turn.message.stopReason === 'aborted')
+                return decision;
+            // A final answer with nothing queued and no continuation request ends the loop by itself.
+            const wouldContinue = turn.toolResults.length > 0 || decision?.action === 'continue' || agent.hasQueuedMessages();
+            if (!wouldContinue)
+                return decision;
+            this.reach();
+            return { action: 'end' };
+        };
+        // Gate B. The run controller calls agent.continue() only after the previous loop has fully
+        // ended, so no request is in flight and no tool is running here. Instead of continuing, ask
+        // the session to stop its run: abort() first sets the controller's stop flag synchronously,
+        // which ends the run before another request. Queued messages stay queued for the next prompt.
+        const resume = agent.continue.bind(agent);
+        agent.continue = async () => {
+            if (this.state === 'none')
+                return resume();
+            this.reach();
+            // abort() waits for the run to settle, which happens only after this call returns.
+            void session.abort().catch(() => { });
+        };
+        session.subscribe(event => {
+            if (event.type !== 'agent_settled') {
+                this.settledPaused = false;
+                return;
+            }
+            const state = this.state;
+            this.state = 'none';
+            this.settledPaused = state === 'reached';
+            if (state === 'reached')
+                this.emit('paused');
+            else if (state === 'requested')
+                this.emit('finished');
+        });
+    }
+    /** Whether the agent_settled being delivered ends a run that stopped at the safe point. */
+    get pausedRunSettled() {
+        return this.settledPaused;
+    }
+    async request() {
+        if (!this.session.isStreaming)
+            throw error('Agent is not running', 'AGENT_NOT_RUNNING');
+        if (this.state !== 'none')
+            return;
+        this.state = 'requested';
+        this.emit('requested');
+        // During a retry backoff the next step is the retried model request. isRetrying means Pi has
+        // created the retry's abort controller, so cancelling it now takes effect; the run then winds
+        // down through the controller, and any continuation it still attempts meets gate B.
+        if (this.session.isRetrying) {
+            this.reach();
+            this.session.abortRetry();
+        }
+    }
+    async cancel() {
+        if (this.state === 'reached')
+            throw error('The run has already stopped at the safe point', 'PAUSE_ALREADY_REACHED');
+        if (this.state !== 'requested')
+            return;
+        this.state = 'none';
+        this.emit('cancelled');
+    }
+    reach() {
+        if (this.state === 'reached')
+            return;
+        this.state = 'reached';
+        this.emit('reached');
+    }
+}

package/dist/sdk-backend.js

@@ -2,6 +2,7 @@ import { createAgentSession, DefaultResourceLoader, getAgentDir, SessionManager
 import { EventEmitter } from 'node:events';
 import { randomUUID } from 'node:crypto';
 import { PERFORMANCE_ENTRY, SessionPerformanceTracker } from './session-performance.js';
+import { SafePause } from './safe-pause.js';
 import { createCollabExtension } from './collab/pi-extension.js';
 /** Older Pi builds resolved steer()/followUp() with undefined; treat anything but an explicit `handled` as queued. */
 const normalizeDisposition = (value) => value === 'handled' ? 'handled' : 'queued';
@@ -12,6 +13,7 @@ export class SdkBackend {
     cwd;
     disposed = false;
     events = new EventEmitter();
+    safePause;
     pendingUi = new Map();
     constructor(session, agentId, cwd) {
         this.session = session;
@@ -19,6 +21,14 @@ export class SdkBackend {
         this.cwd = cwd;
         const performance = new SessionPerformanceTracker(run => this.session.sessionManager.appendCustomEntry(PERFORMANCE_ENTRY, run));
         this.session.subscribe(event => performance.apply(event));
+        this.safePause = new SafePause(session, state => this.events.emit('event', { type: 'safe_pause', state }));
+    }
+    /** Ask the running turn to stop before its next model request (see SafePause). */
+    requestPause() {
+        return this.safePause.request();
+    }
+    cancelPause() {
+        return this.safePause.cancel();
     }
     static async create(agentId, cwd, sessionFile, profile = 'default', collabBridge, options = {}) {
         const sessionManager = sessionFile ? SessionManager.open(sessionFile) : SessionManager.create(cwd);
@@ -56,14 +66,18 @@ export class SdkBackend {
             session.dispose();
             throw Object.assign(new Error(modelFallbackMessage), { code: 'PI_MODEL_UNAVAILABLE' });
         }
+        return SdkBackend.fromSession(session, agentId, cwd);
+    }
+    /** Wrap an existing Pi session; tests use it to drive a real session with a scripted model. */
+    static async fromSession(session, agentId, cwd) {
         const backend = new SdkBackend(session, agentId, cwd);
         await session.bindExtensions({ uiContext: backend.extensionUi() });
         return backend;
     }
     // Slash commands, skills and prompt templates are expanded by prompt() itself (expandPromptTemplates defaults to true),
     // so `/name args` typed in the web composer behaves exactly like the Pi TUI.
-    prompt(message) {
-        return this.session.prompt(message);
+    prompt(message, onAccepted) {
+        return this.session.prompt(message, { preflightResult: onAccepted });
     }
     // Since pi 0.99 steer()/followUp() report whether the message reached the running turn (`handled`)
     // or waits for it (`queued`); the web composer shows that difference instead of guessing.
@@ -267,7 +281,10 @@ export class SdkBackend {
             throw Object.assign(new Error('Extension UI request not found'), { code: 'EXTENSION_REQUEST_NOT_FOUND' });
     }
     subscribe(listener) {
-        const fromSession = (event) => listener(event);
+        // A paused run settles like a finished one; notifications must not report it as completed.
+        const fromSession = (event) => listener(event.type === 'agent_settled' && this.safePause.pausedRunSettled
+            ? { ...event, paused: true }
+            : event);
         const unsubscribe = this.session.subscribe(fromSession);
         this.events.on('event', listener);
         return () => {

package/dist/session-performance.js

@@ -43,6 +43,8 @@ export class SessionPerformanceTracker {
                 break;
             case 'message_end':
                 if (e.message?.role === 'assistant') {
+                    if (this.model && Number.isFinite(e.message.usage?.output) && e.message.usage.output >= 0)
+                        this.model.outputTokens = e.message.usage.output;
                     this.finish(this.model, at);
                     this.model = null;
                 }

package/dist/voice.js

@@ -241,8 +241,8 @@ export class VoiceManager {
             const finalOutput = this.lastAssistant.get(agentId), userPrompt = this.lastPrompt.get(agentId) ?? '';
             this.lastAssistant.delete(agentId);
             this.lastPrompt.delete(agentId);
-            // A run the user cancelled has no result worth speaking; Pi 0.99+ reports that as `aborted`.
-            if (finalOutput && event.aborted !== true)
+            // A run the user cancelled or safely paused has no result worth speaking (`aborted` / `paused`).
+            if (finalOutput && event.aborted !== true && event.paused !== true)
                 void this.announce(agentId, { userPrompt, finalOutput, ...context });
         }
     }

package/dist/ws/agent-socket.js

@@ -10,6 +10,7 @@ const OVERVIEW_EVENTS = new Set([
     'agent_settled',
     'activity_changed',
     'queue_update',
+    'safe_pause',
     'session_info_changed',
     'agent_unloaded',
     'agent_loaded',
@@ -225,7 +226,7 @@ export function handleAgentSocket(ctx, ws) {
                         ctx.voice?.recordUserPrompt(m.agentId, String(m.message ?? ''));
                         ctx.sessionNamer?.recordUserPrompt(m.agentId, String(m.message ?? ''));
                     }
-                    const disposition = await ctx.agents.command(m.agentId, m.type, m.message);
+                    const disposition = await ctx.agents.submitCommand(m.agentId, m.type, m.message);
                     send({ type: 'command_result', requestId: m.requestId, success: true, disposition });
                     return;
                 }

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "pi2web",
-  "version": "0.2.8",
+  "version": "0.2.12",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "pi2web",
-      "version": "0.2.8",
+      "version": "0.2.12",
       "license": "MIT",
       "dependencies": {
         "@earendil-works/chord": "1.1.0",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi2web",
-  "version": "0.2.8",
+  "version": "0.2.12",
   "description": "Secure remote web client server for Pi coding agent",
   "type": "module",
   "author": "sxwxs",

package/packages/protocol/openapi.yaml

@@ -39,11 +39,14 @@ paths:
     post: {operationId: transcribeAgentDraft, description: 'When Sotto is enabled, body is 16 kHz mono float32 little-endian PCM (application/octet-stream; max 60 s); otherwise uses configured OpenAI-compatible STT audio. Returns {data:{text}} without submitting to the Agent.', parameters: [{$ref: '#/components/parameters/Id'}], responses: {'200': {$ref: '#/components/responses/Success'}}}
   /agents/{id}/transcribe/ws:
     get: {operationId: streamSottoDraft, description: 'Authenticated WebSocket: start {type:start,format:float32,sample_rate}, binary mono PCM chunks, stop {type:stop}; forwards Sotto ready/partial/final/error messages. Only when Sotto is enabled; never submits to Agent.', parameters: [{$ref: '#/components/parameters/Id'}], responses: {'101': {description: 'WebSocket upgrade'}}}
+  /agents/{id}/pause:
+    post: {operationId: requestSafePause, description: 'Safe pause: the conversation stops advancing. Running tools finish and their results are recorded, then no further model response or tool call is produced: the next agent-loop turn, auto-retry, overflow-compaction retry, queued messages and extension agent_before_settle continuations are all withheld, and a waiting retry backoff is cancelled. Queued input stays queued for the next prompt. Housekeeping model calls (compaction summaries, prompt-cache warming) are not withheld. Progress is published as safe_pause events (requested, reached, paused, cancelled, finished) and as activity.pauseRequested; the agent_settled of a paused run carries paused: true. Rejects with AGENT_NOT_RUNNING while idle.', parameters: [{$ref: '#/components/parameters/Id'}], responses: {'200': {$ref: '#/components/responses/Success'}}}
+    delete: {operationId: cancelSafePause, description: 'Withdraw a pending safe pause. Rejects with PAUSE_ALREADY_REACHED while the run is winding down at the safe point (after reached, before paused). Without a pending pause, including after the run has settled, it is a no-op that succeeds. Whether the withdrawal took effect is reported by the safe_pause event (cancelled or paused).', parameters: [{$ref: '#/components/parameters/Id'}], responses: {'200': {$ref: '#/components/responses/Success'}}}
   /agents/{id}/queue/move:
     post: {operationId: moveQueuedMessage, description: 'Move one unconsumed message to the other queue during a run. Body {from: steering|followUp, index: integer, expected: string}; rejects stale entries.', parameters: [{$ref: '#/components/parameters/Id'}], requestBody: {$ref: '#/components/requestBodies/Json'}, responses: {'200': {$ref: '#/components/responses/Success'}}}
   /agents/{id}/{action}:
     get: {operationId: getAgentResource, parameters: [{$ref: '#/components/parameters/Id'}, {name: action, in: path, required: true, schema: {enum: [state,capabilities,session]}}], responses: {'200': {$ref: '#/components/responses/Success'}}}
-    post: {operationId: agentCommand, parameters: [{$ref: '#/components/parameters/Id'}, {name: action, in: path, required: true, schema: {enum: [prompt,steer,follow-up,abort,compact,model,thinking,session-name,navigate,fork,extension-response]}}], requestBody: {$ref: '#/components/requestBodies/Json'}, responses: {'200': {$ref: '#/components/responses/Success'}}}
+    post: {operationId: agentCommand, description: 'For prompt/steer/follow-up, success acknowledges acceptance by Pi, not completion of the run. Preflight rejections are HTTP errors; subsequent execution updates and errors use the WebSocket event stream. A lost response does not imply the message was rejected; clients must not automatically resend it.', parameters: [{$ref: '#/components/parameters/Id'}, {name: action, in: path, required: true, schema: {enum: [prompt,steer,follow-up,abort,compact,model,thinking,session-name,navigate,fork,extension-response]}}], requestBody: {$ref: '#/components/requestBodies/Json'}, responses: {'200': {$ref: '#/components/responses/Success'}}}
   /draft:
     post: {operationId: draftText, description: 'Insert text into the currently displayed Web UI input box on this backend, without sending it to an Agent. Body {text, mode?: append|replace, source?}; explicit agent/slot targets are rejected. Pending text is kept in memory until a page acknowledges it.', requestBody: {$ref: '#/components/requestBodies/Json'}, responses: {'200': {$ref: '#/components/responses/Success'}}}
 components:

package/packages/protocol/websocket.schema.json

@@ -19,9 +19,9 @@
     "focus": {"type":"object","additionalProperties":false,"required":["type","agentId"],"properties":{"type":{"const":"focus"},"agentId":{"type":["string","null"]},"limit":{"type":"integer","minimum":1,"maximum":500,"description":"Size of the snapshot's page, 50 by default."}},"description":"View one Agent in detail (null: none). Loads it if needed and answers with focus_state, or focus_error; replaces the previous focus and its watches."},
     "watch": {"type":"object","additionalProperties":false,"required":["type","toolCallId"],"properties":{"type":{"enum":["watch","unwatch"]},"toolCallId":{"type":"string"}},"description":"Start or stop tool_output for a running tool of the focused Agent, e.g. while its card is expanded."},
     "command": {"type":"object","additionalProperties":false,"required":["type","requestId","agentId"],"properties":{"type":{"enum":["prompt","steer","follow-up","abort","clear-queue"]},"requestId":{"type":"string"},"agentId":{"type":"string"},"message":{"type":"string","maxLength":2097152}}},
-    "commandResult": {"type":"object","required":["type","success"],"properties":{"type":{"const":"command_result"},"requestId":{"type":"string"},"success":{"type":"boolean"},"error":{"type":"object"},"disposition":{"enum":["handled","queued"],"description":"For steer and follow-up: whether the running turn took the message or it waits in the queue."},"queue":{"$ref":"#/$defs/queue"}}},
+    "commandResult": {"type":"object","description":"For prompt/steer/follow-up, success acknowledges acceptance, not completion. A lost acknowledgement must not trigger an automatic resend.","required":["type","success"],"properties":{"type":{"const":"command_result"},"requestId":{"type":"string"},"success":{"type":"boolean"},"error":{"type":"object"},"disposition":{"enum":["handled","queued"],"description":"For steer and follow-up: whether the running turn took the message or it waits in the queue."},"queue":{"$ref":"#/$defs/queue"}}},
     "queue": {"type":"object","required":["steering","followUp"],"description":"Messages waiting for the running turn.","properties":{"steering":{"type":"array","items":{"type":"string"}},"followUp":{"type":"array","items":{"type":"string"}}}},
-    "agentEvent": {"type":"object","required":["type","agentId","timestamp","status","event"],"properties":{"type":{"const":"agent_event"},"agentId":{"type":"string"},"timestamp":{"type":"integer","minimum":0,"description":"Epoch milliseconds."},"status":{"$ref":"#/$defs/agentStatus","description":"Agent status after this event."},"event":{"type":"object","required":["type"],"description":"Overview events of every Agent: agent_loaded (its backend started on demand), agent_start, agent_end {willRetry, aborted}, agent_settled, activity_changed {activity}, queue_update, session_info_changed, agent_unloaded, auto_retry_*, summarization_retry_*, extension_ui_*. For the focused Agent also bash_execution_start/update/end (output of user shell commands)."}}},
+    "agentEvent": {"type":"object","required":["type","agentId","timestamp","status","event"],"properties":{"type":{"const":"agent_event"},"agentId":{"type":"string"},"timestamp":{"type":"integer","minimum":0,"description":"Epoch milliseconds."},"status":{"$ref":"#/$defs/agentStatus","description":"Agent status after this event."},"event":{"type":"object","required":["type"],"description":"Overview events of every Agent: agent_loaded (its backend started on demand), agent_start, agent_end {willRetry, aborted}, agent_settled, activity_changed {activity}, queue_update, safe_pause {state: requested|reached|paused|cancelled|finished}, session_info_changed, agent_unloaded, auto_retry_*, summarization_retry_*, extension_ui_*. For the focused Agent also bash_execution_start/update/end (output of user shell commands)."}}},
     "focusState": {"type":"object","required":["type","agentId","sessionId","page","live"],"description":"Snapshot of the focused Agent. Every detail message after it is newer.","properties":{"type":{"const":"focus_state"},"agentId":{"type":"string"},"sessionId":{"type":"string"},"page":{"type":"object","required":["items","total","start","hasMore"],"properties":{"items":{"type":"array","items":{"$ref":"#/$defs/timelineItem"}},"total":{"type":"integer"},"start":{"type":"integer"},"hasMore":{"type":"boolean"}}},"live":{"oneOf":[{"type":"null"},{"type":"object","required":["text","thinking"],"properties":{"text":{"type":"string"},"thinking":{"type":"string"}}}],"description":"Text streamed so far for the message being generated."},"dialogRequests":{"type":"array","description":"Pending dialogs, to rebuild their controls in the transcript."}}},
     "focusError": {"type":"object","required":["type","agentId","error"],"properties":{"type":{"const":"focus_error"},"agentId":{"type":"string"},"error":{"type":"object"}}},
     "delta": {"type":"object","required":["type","agentId","kind","delta"],"properties":{"type":{"const":"delta"},"agentId":{"type":"string"},"kind":{"enum":["text","thinking"]},"delta":{"type":"string"}},"description":"More streamed text; adjacent tokens are coalesced."},
@@ -63,7 +63,8 @@
           }
         },
         "bash":{"type":"array","items":{"type":"object","required":["id","since"],"properties":{"id":{"type":"string"},"command":{"type":"string"},"since":{"type":"integer"}}}},
-        "dialogs":{"type":"array","items":{"type":"object","required":["requestId","since"],"properties":{"requestId":{"type":"string"},"title":{"type":"string"},"since":{"type":"integer"}}}}
+        "dialogs":{"type":"array","items":{"type":"object","required":["requestId","since"],"properties":{"requestId":{"type":"string"},"title":{"type":"string"},"since":{"type":"integer"}}}},
+        "pauseRequested":{"type":"object","required":["since"],"description":"A safe pause is pending: the run stops before its next model request. reached: a request was withheld and the run is winding down; it can no longer be withdrawn.","properties":{"since":{"type":"integer"},"reached":{"const":true}}}
       }
     },
     "agentState": {

package/web/activity-view.js

@@ -13,6 +13,8 @@
     compacting: ['🗜️', '压缩上下文'],
     stopping: ['⏹️', '正在停止']
   };
+  /** Streaming phases that report an estimated throughput, and how their output count reads. */
+  const STREAM_COUNT = {thinking: '已思考', generating: '已输出', preparing_tool: '参数'};
 
   const isBusy = activity =>
     !!activity &&
@@ -21,6 +23,10 @@
       !!activity.bash?.length ||
       !!activity.dialogs?.length);
 
+  function formatDuration(ms) {
+    return ms < 1000 ? `${Math.round(ms)} ms` : `${(ms / 1000).toFixed(1)} s`;
+  }
+
   function formatElapsed(ms) {
     const total = Math.max(0, Math.floor(ms / 1000)),
       h = Math.floor(total / 3600),
@@ -40,6 +46,20 @@
     const phase = activity.phase;
     if (phase && phase.kind !== 'idle') {
       const [icon, label] = PHASE_TEXT[phase.kind] || ['•', phase.kind];
+      const metrics = activity.metrics;
+      const ttft =
+        metrics?.firstResponseAt === undefined
+          ? ''
+          : ` · TTFT ${formatDuration(metrics.firstResponseAt - metrics.turnStartedAt)}`;
+      // `textStartedAt`: backends before per-phase throughput only measured the text phase.
+      const streamStartedAt = metrics?.streamStartedAt ?? metrics?.textStartedAt;
+      const streaming =
+        STREAM_COUNT[phase.kind] && (metrics?.streamStartedAt !== undefined || phase.kind === 'generating');
+      const output = streaming && metrics?.chars ? ` · ${STREAM_COUNT[phase.kind]} ${metrics.chars} 字` : '';
+      const speed =
+        streaming && streamStartedAt !== undefined
+          ? {speed: {tokens: metrics.estimatedTokens, since: streamStartedAt}}
+          : {};
       if (phase.kind === 'running_tools')
         for (const tool of phase.tools || [])
           parts.push({
@@ -50,8 +70,9 @@
       else if (phase.kind === 'preparing_tool')
         parts.push({
           icon,
-          text: `${label}${phase.toolName ? `：${phase.toolName}` : ''}`,
-          since: phase.since
+          text: `${label}${phase.toolName ? `：${phase.toolName}` : ''}${ttft}${output}`,
+          since: phase.since,
+          ...speed
         });
       else if (phase.kind === 'retrying')
         parts.push({
@@ -59,8 +80,14 @@
           text: `${label}${phase.attempt ? ` ${phase.attempt}/${phase.maxAttempts || '?'}` : ''}${phase.error ? ` · ${phase.error}` : ''}`,
           since: phase.since
         });
-      else parts.push({icon, text: label, since: phase.since});
+      else parts.push({icon, text: `${label}${ttft}${output}`, since: phase.since, ...speed});
     }
+    if (activity.pauseRequested)
+      parts.push({
+        icon: '⏸️',
+        text: activity.pauseRequested.reached ? '已到暂停点，正在收尾' : '将在下次调用模型前暂停',
+        since: activity.pauseRequested.since
+      });
     for (const bash of activity.bash || [])
       parts.push({icon: '$', text: bash.command || 'bash', since: bash.since});
     return parts;
@@ -86,7 +113,17 @@
       text.className = 'activity-text';
       text.textContent = `${part.icon} ${part.text}`;
       text.title = part.text;
-      node.append(text, elapsed(backendId, part.since));
+      node.append(text);
+      if (part.speed) {
+        const speed = document.createElement('span');
+        speed.className = 'activity-speed';
+        speed.dataset.tokens = String(part.speed.tokens);
+        speed.dataset.since = String(part.speed.since);
+        speed.dataset.backend = backendId || '';
+        speed.textContent = speedText(part.speed.tokens, serverNow(backendId) - part.speed.since);
+        node.append(speed);
+      }
+      node.append(elapsed(backendId, part.since));
       return node;
     }
     /** The full activity of the selected Agent: every part plus the run's total time. */
@@ -110,9 +147,17 @@
       if (parts.length > 1) node.append(` +${parts.length - 1}`);
       return node;
     }
+    function speedText(tokens, ms) {
+      return ms < 1000 ? ' · 估算中' : ` · 约 ${((tokens * 1000) / ms).toFixed(1)} token/s`;
+    }
     function tick() {
       for (const el of document.querySelectorAll('.elapsed[data-since]'))
         el.textContent = formatElapsed(serverNow(el.dataset.backend) - Number(el.dataset.since));
+      for (const el of document.querySelectorAll('.activity-speed[data-since]'))
+        el.textContent = speedText(
+          Number(el.dataset.tokens),
+          serverNow(el.dataset.backend) - Number(el.dataset.since)
+        );
     }
     return {bar, line, tick};
   }

package/web/app.js

@@ -69,6 +69,8 @@
     voiceEnabled: false,
     voiceSttEnabled: false,
     voicePlaybackEnabled: localStorage.rpVoicePlayback === 'true',
+    // Per browser: 'enter' sends on Enter (Shift+Enter for a new line), 'ctrl-enter' on Ctrl/⌘+Enter.
+    enterSendMode: localStorage.rpEnterSendMode === 'enter' ? 'enter' : 'ctrl-enter',
     voiceAudio: {
       context: null,
       nextTime: 0,
@@ -87,6 +89,8 @@
     mediaBackendId: null,
     /** Messages that failed after the user switched away; restored when that Agent is selected again. */
     failedSends: new Map(),
+    /** Lost acknowledgements are notices, not drafts to send again. Kept across focus snapshots. */
+    uncertainSends: new Map(),
     mailBackendId: null,
     mailRequest: 0,
     sottoBackendId: null,
@@ -1300,9 +1304,16 @@
       nodes = agent ? activityView.bar(state.activities.get(agent.agentId), agent.backendId) : [];
     bar.replaceChildren(...nodes);
     bar.hidden = !nodes.length;
+    // A pending safe pause can be withdrawn with the same button until a model request was withheld.
+    const pause = agent ? state.activities.get(agent.agentId)?.pauseRequested : null,
+      button = $('safePause');
+    button.textContent = pause?.reached ? '暂停中…' : pause ? '取消暂停' : '安全暂停';
+    button.disabled = !!pause?.reached;
   }
   function setAgentActivity(agentId, activity) {
-    if (window.ActivityView.isBusy(activity)) state.activities.set(agentId, activity);
+    // A pause request outlives the run (agent_end → agent_settled); its safe_pause event clears it.
+    if (window.ActivityView.isBusy(activity) || activity?.pauseRequested)
+      state.activities.set(agentId, activity);
     else state.activities.delete(agentId);
     sessionKeyboard.setActivity(agentId, activity);
     refreshToolCards(agentId);
@@ -2030,6 +2041,7 @@
       box.scrollTop = box.scrollHeight;
     } else box.innerHTML = '<div class="empty">正在加载…</div>';
     if (!sendFocus(true) && !cached) box.innerHTML = '<div class="empty">等待连接 backend…</div>';
+    for (const text of state.uncertainSends.get(agent.agentId) || []) showUncertainSend(agent.agentId, text);
   }
   /** Ask for a new snapshot, e.g. when the branch changed. */
   function refocus() {
@@ -2046,6 +2058,8 @@
     timeline.ready = true;
     applyLive(message.live);
     restoreExtensionRequests(message.agentId, {dialogRequests: message.dialogRequests || []});
+    for (const text of state.uncertainSends.get(message.agentId) || [])
+      showUncertainSend(message.agentId, text);
     saveTimelineCache();
     box.scrollTop = box.scrollHeight;
   }
@@ -2814,6 +2828,11 @@
       setAgentActivity(message.agentId, ev.activity);
       return;
     }
+    if (ev.type === 'safe_pause') {
+      if (selected && ev.state === 'paused') toast('已安全暂停：工具结果已保存，发送消息即可继续');
+      else if (selected && ev.state === 'finished') toast('运行已结束，未到达暂停点');
+      return;
+    }
     if (ev.type === 'agent_unloaded') {
       setAgentStatus(message.agentId, 'unloaded');
       setAgentQueue(message.agentId, null);
@@ -3321,18 +3340,18 @@
     if (!workspace) return toast('请先选择 Workspace');
     const cwd = $('agentCwd').value.trim() || '.';
     const preferenceKey = `rpLastTemplate:${workspace.backendId}`;
-    let sessions,
-      templates = [];
+    let templates = [],
+      modelCatalog;
     try {
-      [sessions, templates] = await Promise.all([
-        requestBackend(
-          workspace.backendId,
-          `/api/v1/sessions?workspaceId=${encodeURIComponent(workspace.id)}&path=${encodeURIComponent(cwd)}`
-        ),
+      [templates, modelCatalog] = await Promise.all([
         // Older backends have no templates endpoint; they simply offer Pi's default prompt.
         requestBackend(workspace.backendId, '/api/v1/agent-templates')
           .then(data => data.templates || [])
-          .catch(() => [])
+          .catch(() => []),
+        requestBackend(
+          workspace.backendId,
+          `/api/v1/agents/models?workspaceId=${encodeURIComponent(workspace.id)}&relativeCwd=${encodeURIComponent(cwd)}`
+        ).catch(() => null)
       ]);
     } catch (error) {
       return toast(error.message);
@@ -3341,19 +3360,71 @@
       '新建 Agent',
       body => {
         const c = field(body, '工作路径', cwd);
-        const l = document.createElement('label');
-        l.textContent = 'Session';
-        const s = document.createElement('select');
-        s.innerHTML =
-          '<option value="">新 Session</option>' +
-          sessions
-            .map(
-              x =>
-                `<option value="${esc(x.path)}">${esc(x.name || x.sessionName || x.id || x.path)} · ${esc(x.modified || '')}</option>`
-            )
-            .join('');
-        l.append(s);
-        body.append(l);
+        const modelLabel = document.createElement('label');
+        modelLabel.textContent = '模型';
+        const modelSelect = document.createElement('select');
+        const renderModels = (catalog, selected = modelSelect.value) => {
+          modelSelect.replaceChildren();
+          const fallback = document.createElement('option');
+          fallback.value = '';
+          fallback.textContent = 'Pi 默认模型';
+          modelSelect.append(fallback);
+          for (const model of catalog?.models || []) {
+            const option = document.createElement('option');
+            option.value = JSON.stringify([model.provider, model.id]);
+            option.textContent = `${model.provider} / ${model.name || model.id}`;
+            modelSelect.append(option);
+          }
+          if (selected && [...modelSelect.options].some(option => option.value === selected))
+            modelSelect.value = selected;
+          else if (catalog?.model) {
+            const current = JSON.stringify([catalog.model.provider, catalog.model.id]);
+            if ([...modelSelect.options].some(option => option.value === current))
+              modelSelect.value = current;
+          }
+          if (selected && modelSelect.value !== selected)
+            toast('所选模型在新工作路径不可用，已改用当前默认模型；请确认选择。');
+        };
+        renderModels(modelCatalog);
+        modelLabel.append(modelSelect);
+        body.append(modelLabel);
+        let modelRequest = 0;
+        const submit = $('modalOk');
+        $('modal').addEventListener(
+          'close',
+          () => {
+            ++modelRequest;
+            submit.disabled = false;
+          },
+          {once: true}
+        );
+        c.addEventListener('change', async () => {
+          const request = ++modelRequest;
+          const selected = modelSelect.value;
+          submit.disabled = true;
+          modelSelect.disabled = true;
+          try {
+            const catalog = await requestBackend(
+              workspace.backendId,
+              `/api/v1/agents/models?workspaceId=${encodeURIComponent(workspace.id)}&relativeCwd=${encodeURIComponent(c.value.trim() || '.')}`
+            );
+            if (request === modelRequest) {
+              modelCatalog = catalog;
+              renderModels(catalog, selected);
+            }
+          } catch (error) {
+            if (request === modelRequest) {
+              modelCatalog = null;
+              renderModels(null, '');
+              toast(`读取模型失败：${error.message}；模型选择已重置，请确认后创建。`);
+            }
+          } finally {
+            if (request === modelRequest) {
+              modelSelect.disabled = false;
+              submit.disabled = false;
+            }
+          }
+        });
         const t = templateSelect(body, templates);
         const preferred = globalThis.localStorage?.getItem(preferenceKey);
         if (preferred && [...t.options].some(option => option.value === preferred)) t.value = preferred;
@@ -3371,8 +3442,8 @@
         t.onchange = () => preview.reset();
         return () => ({
           cwd: c.value.trim() || '.',
-          sessionFile: s.value || undefined,
-          templateId: t.value || undefined
+          templateId: t.value || undefined,
+          model: modelSelect.value ? JSON.parse(modelSelect.value) : null
         });
       },
       '创建'
@@ -3382,12 +3453,12 @@
       const created = await postBackend(workspace.backendId, '/api/v1/agents', {
         workspaceId: workspace.id,
         relativeCwd: result.cwd,
-        sessionFile: result.sessionFile,
-        templateId: result.templateId
+        templateId: result.templateId,
+        ...(result.model ? {model: {provider: result.model[0], modelId: result.model[1]}} : {})
       });
       localStorage.setItem(preferenceKey, result.templateId || '');
       const agent = stampNewItem(workspace.backendId, created);
-      const binding = result.sessionFile ? null : autoBindNewSession(agent);
+      const binding = autoBindNewSession(agent);
       await refreshAgents();
       await selectAgent(agent);
       if (binding) toast(`新 Session 已自动绑定到按键 ${binding.slot + 1}`);
@@ -3775,7 +3846,7 @@
         const info = document.createElement('p');
         info.className = 'muted';
         info.textContent =
-          '横轴为每轮运行的实际时间；同一时间段的工具调用可能并行。模型耗时包含首响应等待与输出；网络、排队和其他开销可能落在空白时间。';
+          '横轴为每轮运行的实际时间；同一时间段的工具调用可能并行。模型耗时包含首响应等待与输出；网络、排队和其他开销可能落在空白时间。实际输出 token 可能包含思考 token。';
         body.append(info);
         if (!runs.length && !legacyTools.length) {
           body.append(el('p', {textContent: '当前分支没有可分析的耗时记录；新运行结束后会记录精确时间。'}));
@@ -3810,6 +3881,10 @@
               span.kind === 'model' && span.firstResponse !== undefined
                 ? ` · 首响应 ${performanceTime(span.firstResponse - span.start)} · 响应 ${performanceTime(span.end - span.firstResponse)}`
                 : '';
+            const tokens =
+              span.kind === 'model' && span.outputTokens !== undefined
+                ? ` · 输出 ${span.outputTokens} token${span.firstResponse !== undefined && span.end > span.firstResponse ? ` · 均速 ${((span.outputTokens * 1000) / (span.end - span.firstResponse)).toFixed(1)} token/s` : ''}`
+                : '';
             const label =
               span.kind === 'model'
                 ? '模型调用'
@@ -3821,7 +3896,7 @@
             const row = el('div', {className: 'performance-row'});
             const text = el('span', {
               className: 'performance-label',
-              textContent: `${label} · +${performanceTime(span.start - run.start)} · ${performanceTime(length)}${first}${span.incomplete ? ' · 未完成' : ''}`
+              textContent: `${label} · +${performanceTime(span.start - run.start)} · ${performanceTime(length)}${first}${tokens}${span.incomplete ? ' · 未完成' : ''}`
             });
             const track = el('div', {className: 'performance-track'});
             const bar = el('i', {className: `performance-${span.kind}`});
@@ -4126,6 +4201,18 @@
     }
   };
   $('workspaces').onchange = () => void selectWorkspace();
+  function applyEnterSendMode() {
+    $('enterSendMode').value = state.enterSendMode;
+    $('input').placeholder = `输入 @ 引用文件，/ 使用命令，! 执行 Shell；${
+      state.enterSendMode === 'enter' ? 'Enter 发送，Shift+Enter 换行' : 'Ctrl/⌘+Enter 发送'
+    }`;
+  }
+  applyEnterSendMode();
+  $('enterSendMode').onchange = () => {
+    state.enterSendMode = $('enterSendMode').value === 'enter' ? 'enter' : 'ctrl-enter';
+    localStorage.rpEnterSendMode = state.enterSendMode;
+    applyEnterSendMode();
+  };
   $('agentPageSize').value = String(state.agentPageSize);
   $('agentPageSize').onchange = () => {
     state.agentPageSize = Number($('agentPageSize').value) || 10;
@@ -4203,6 +4290,13 @@
     const id = state.agent.agentId;
     Promise.all([postAgent(id, '/abort'), postAgent(id, '/bash-abort')]).catch(e => toast(e.message));
   };
+  $('safePause').onclick = () => {
+    if (!state.agent) return;
+    const id = state.agent.agentId,
+      pause = state.activities.get(id)?.pauseRequested;
+    if (pause?.reached) return;
+    requestAgent(id, '/pause', {method: pause ? 'DELETE' : 'POST'}).catch(e => toast(e.message));
+  };
   $('stop').onclick = async () => {
     if (state.agent && confirm('停止这个 Agent？之后再次使用时会按需启动。')) {
       await requestAgent(state.agent.agentId, '', {method: 'DELETE'});
@@ -4248,6 +4342,25 @@
   function prependInput(text) {
     $('input').value = text + ($('input').value ? `\n${$('input').value}` : '');
   }
+  function showUncertainSend(agentId, text) {
+    const card = addCard(
+      '送达状态未知',
+      `连接中断，以下消息可能已送达。请先核对会话和队列，不要直接重复发送。\n\n${text}`,
+      'error',
+      true
+    );
+    const dismiss = document.createElement('button');
+    dismiss.type = 'button';
+    dismiss.textContent = '已核对，关闭提示';
+    dismiss.onclick = () => {
+      const pending = state.uncertainSends.get(agentId) || [],
+        index = pending.indexOf(text);
+      if (index >= 0) pending.splice(index, 1);
+      if (!pending.length) state.uncertainSends.delete(agentId);
+      removeCard(card);
+    };
+    card.details.append(dismiss);
+  }
   $('prompt').onsubmit = async event => {
     event.preventDefault();
     if (!state.agent) return toast('请先创建或选择 Agent');
@@ -4285,6 +4398,17 @@
       if (mode !== 'prompt' && sent?.disposition)
         toast(sent.disposition === 'queued' ? '已排队，等待当前回合' : '已送达当前回合');
     } catch (e) {
+      // A lost response (including a gateway timeout) says nothing about whether Pi accepted the
+      // input. Keep the text for review, but never turn it back into a sendable draft automatically.
+      if (!e.status || e.status === 408 || e.status >= 500) {
+        const pending = state.uncertainSends.get(agentId) || [];
+        pending.push(text);
+        state.uncertainSends.set(agentId, pending);
+        if (state.selectedKind === 'agent' && state.agent?.agentId === agentId)
+          showUncertainSend(agentId, text);
+        else toast('消息送达状态未知；切回该会话后请核对会话和队列');
+        return;
+      }
       // The composer is shared: only restore into it while the failed Agent is still selected.
       if (state.agent?.agentId === agentId) {
         prependInput(text);
@@ -4297,6 +4421,15 @@
       }
     }
   };
+  /**
+   * Whether a keydown in the prompt sends it. Ctrl/⌘+Enter always sends; in 'enter' mode a plain Enter
+   * does too. Enter that confirms an IME composition never sends.
+   */
+  function isSendKey(event, mode) {
+    if (event.key !== 'Enter' || event.isComposing || event.keyCode === 229) return false;
+    if (event.ctrlKey || event.metaKey) return true;
+    return mode === 'enter' && !event.shiftKey && !event.altKey;
+  }
   $('input').onkeydown = event => {
     if (event.altKey && event.shiftKey && event.key.toLowerCase() === 'v' && !event.isComposing) {
       event.preventDefault();
@@ -4337,7 +4470,8 @@
         return;
       }
     }
-    if ((event.ctrlKey || event.metaKey) && event.key === 'Enter') {
+    if (isSendKey(event, state.enterSendMode)) {
+      event.preventDefault();
       $('prompt').requestSubmit();
     }
   };

package/web/index.html

@@ -50,6 +50,7 @@
       <button id="systemPrompt" class="agent-action" title="查看此 Agent 的 System Prompt 与模板">Prompt</button>
       <button id="switchPrompt" class="agent-action" title="仅新会话首次对话前可切换 System Prompt">切换 Prompt</button>
       <button id="compact" class="agent-action">Compact</button>
+      <button id="safePause" class="agent-action" title="当前工具执行完后、下一次调用模型前暂停；之后发送消息即可继续">安全暂停</button>
       <button id="stop" class="agent-action">停止</button>
       <button id="abort" class="agent-action danger-action">Abort</button>
     </div>
@@ -76,7 +77,7 @@
   </article>
 </main>
 <div id="contextMenu" class="context-menu" hidden><button id="menuStartAgent">启动 Agent</button><button id="menuOpenTerminal">在 Terminal 中打开</button><button id="menuSetCwd">设为 Agent 工作路径</button><button id="menuMention">在输入框中引用</button></div>
-<dialog id="configDialog"><form method="dialog" id="configForm"><h2>Remote Pi 配置</h2><section class="config-section"><h3>Backend 连接</h3><div id="backendList"></div><div class="config-actions"><button id="connect" type="button">添加 Backend</button><button id="disconnectAll" type="button" hidden>断开全部</button></div><small id="configConnectionStatus" class="muted"></small></section><section class="config-section"><h3>六键 Session 键盘</h3><div class="config-actions"><button id="keyboardConnect" type="button">连接键盘</button><button id="keyboardDisconnect" type="button" hidden>断开键盘</button></div><small id="keyboardStatus" class="muted">未连接 · 需要 Chrome/Edge 和 HTTPS 或 localhost</small><div id="keyboardSlots" class="keyboard-slots" aria-label="键盘按键绑定"></div></section><section class="config-section"><h3>浏览器通知与语音</h3><div class="config-actions"><button id="notifications" type="button" title="启用所有 Agent 的完成通知">启用通知</button><button id="voicePlayback" type="button" title="启用 Agent 语音摘要播放" hidden>启用语音</button></div><small id="browserNotificationStatus" class="muted"></small></section><section class="config-section"><h3>Sotto 语音输入</h3><p class="muted">语音输入始终使用提供当前页面的本机 Backend 配置；本机未启用识别时不可用。音频不会发送给远端 Backend，识别出的文字仍可用于远端会话。</p><label class="toggle-row"><input id="sottoEnabled" type="checkbox">启用 Sotto 语音识别（默认关闭）</label><label>Endpoint（pi2web 服务所在机器的本机地址）<input id="sottoEndpoint" type="url" value="http://127.0.0.1:8766" placeholder="http://127.0.0.1:8766"></label><small id="sottoSettingsStatus" class="muted">使用本机 Sotto 已加载的 ASR 模型；未开启时不显示输入框旁的麦克风。</small><div class="config-actions"><button id="saveSotto" type="button">保存语音输入设置</button></div></section><section class="config-section"><h3>System Prompt 与 Agent 模板</h3><p class="muted">查看、修改 Pi 的全局 System Prompt（<code>~/.pi/agent/SYSTEM.md</code>、<code>APPEND_SYSTEM.md</code>），并管理新建 Agent 时可选择的模板（<code>~/.pi/pi2web/agent-templates/*.md</code>）。作用于当前 Workspace 所在的 Backend。</p><div class="config-actions"><button id="openSystemPrompts" type="button">管理 System Prompt…</button></div></section><section class="config-section"><h3>邮件通知</h3><p id="mailNotificationAvailability" class="muted"></p><label class="toggle-row"><input id="mailNotificationEnabled" type="checkbox">启动邮件通知</label><label>邮件通知聚合时间（秒）<input id="mailAggregationDelay" type="number" min="0" max="86400" step="1" value="0"><small class="muted">首个任务完成后等待这段时间，期间完成的其他任务合并到同一封邮件；0 表示立即发送。</small></label><label class="toggle-row"><input id="mailIncludeResponse" type="checkbox" checked>包含 Agent 回复内容</label><label class="toggle-row"><input id="mailIncludeSessionDetails" type="checkbox" checked>包含 Session 名称和路径</label><label class="toggle-row"><input id="mailCollabEscalations" type="checkbox" checked>协作会话需要人工裁定或停滞时立即发信</label><small class="muted">如果回复内容和 Session 信息都不包含，邮件只会说明有任务完成。</small></section><div class="dialog-actions"><button id="configClose" type="button">关闭</button><button id="configSave" type="button">保存邮件设置</button></div></form></dialog>
+<dialog id="configDialog"><form method="dialog" id="configForm"><h2>Remote Pi 配置</h2><section class="config-section"><h3>Backend 连接</h3><div id="backendList"></div><div class="config-actions"><button id="connect" type="button">添加 Backend</button><button id="disconnectAll" type="button" hidden>断开全部</button></div><small id="configConnectionStatus" class="muted"></small></section><section class="config-section"><h3>六键 Session 键盘</h3><div class="config-actions"><button id="keyboardConnect" type="button">连接键盘</button><button id="keyboardDisconnect" type="button" hidden>断开键盘</button></div><small id="keyboardStatus" class="muted">未连接 · 需要 Chrome/Edge 和 HTTPS 或 localhost</small><div id="keyboardSlots" class="keyboard-slots" aria-label="键盘按键绑定"></div></section><section class="config-section"><h3>浏览器通知与语音</h3><div class="config-actions"><button id="notifications" type="button" title="启用所有 Agent 的完成通知">启用通知</button><button id="voicePlayback" type="button" title="启用 Agent 语音摘要播放" hidden>启用语音</button></div><small id="browserNotificationStatus" class="muted"></small></section><section class="config-section"><h3>输入</h3><label>发送快捷键 <select id="enterSendMode" aria-label="发送快捷键"><option value="ctrl-enter">Ctrl/⌘+Enter 发送，Enter 换行</option><option value="enter">Enter 发送，Shift+Enter 换行</option></select></label><small class="muted">仅保存在本浏览器，立即生效。输入法选词时的 Enter 不会发送。</small></section><section class="config-section"><h3>Sotto 语音输入</h3><p class="muted">语音输入始终使用提供当前页面的本机 Backend 配置；本机未启用识别时不可用。音频不会发送给远端 Backend，识别出的文字仍可用于远端会话。</p><label class="toggle-row"><input id="sottoEnabled" type="checkbox">启用 Sotto 语音识别（默认关闭）</label><label>Endpoint（pi2web 服务所在机器的本机地址）<input id="sottoEndpoint" type="url" value="http://127.0.0.1:8766" placeholder="http://127.0.0.1:8766"></label><small id="sottoSettingsStatus" class="muted">使用本机 Sotto 已加载的 ASR 模型；未开启时不显示输入框旁的麦克风。</small><div class="config-actions"><button id="saveSotto" type="button">保存语音输入设置</button></div></section><section class="config-section"><h3>System Prompt 与 Agent 模板</h3><p class="muted">查看、修改 Pi 的全局 System Prompt（<code>~/.pi/agent/SYSTEM.md</code>、<code>APPEND_SYSTEM.md</code>），并管理新建 Agent 时可选择的模板（<code>~/.pi/pi2web/agent-templates/*.md</code>）。作用于当前 Workspace 所在的 Backend。</p><div class="config-actions"><button id="openSystemPrompts" type="button">管理 System Prompt…</button></div></section><section class="config-section"><h3>邮件通知</h3><p id="mailNotificationAvailability" class="muted"></p><label class="toggle-row"><input id="mailNotificationEnabled" type="checkbox">启动邮件通知</label><label>邮件通知聚合时间（秒）<input id="mailAggregationDelay" type="number" min="0" max="86400" step="1" value="0"><small class="muted">首个任务完成后等待这段时间，期间完成的其他任务合并到同一封邮件；0 表示立即发送。</small></label><label class="toggle-row"><input id="mailIncludeResponse" type="checkbox" checked>包含 Agent 回复内容</label><label class="toggle-row"><input id="mailIncludeSessionDetails" type="checkbox" checked>包含 Session 名称和路径</label><label class="toggle-row"><input id="mailCollabEscalations" type="checkbox" checked>协作会话需要人工裁定或停滞时立即发信</label><small class="muted">如果回复内容和 Session 信息都不包含，邮件只会说明有任务完成。</small></section><div class="dialog-actions"><button id="configClose" type="button">关闭</button><button id="configSave" type="button">保存邮件设置</button></div></form></dialog>
 <dialog id="pairDialog"><form method="dialog" id="pairForm"><h2>连接 Backend</h2><p class="muted">可以添加多个 backend；连接成功后可选择是否将配对码保存到浏览器本地存储。请仅在可信设备上保存。</p><label>短名称<input id="pairName" placeholder="例如 home / lab，显示在 Session 标签上"></label><label>服务地址<input id="pairBase" required></label><label>配对码<input id="pairToken" type="password" autocomplete="off" placeholder="服务端使用 --no-auth 时留空" autofocus></label><div class="dialog-actions"><button id="pairCancel" type="button">取消</button><button id="pairSubmit" type="submit" value="default">连接</button></div></form></dialog>
 <div id="agentContextMenu" class="context-menu" hidden><button id="menuBindKeyboard">绑定键盘按键…</button><button id="menuKeyboardColor">更换按键颜色…</button><button id="menuUnbindKeyboard">取消按键绑定</button><button id="menuArchiveAgent">Archive</button><button id="menuCloseTerminal" hidden>关闭 Terminal</button></div>
 <dialog id="modal"><form method="dialog"><h2 id="modalTitle"></h2><div id="modalBody"></div><div class="dialog-actions"><button id="modalCancel" type="button">取消</button><button id="modalOk" type="submit" value="default">确定</button></div></form></dialog>

package/web/styles.css

@@ -100,6 +100,7 @@
 #activityBar[hidden],.agent-activity[hidden]{display:none}
 .activity-chip{display:inline-flex;align-items:center;gap:6px;min-width:0;max-width:100%}
 .activity-text{overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+.activity-speed{flex:0 0 auto;white-space:nowrap;font-variant-numeric:tabular-nums;color:var(--blue)}
 .elapsed{flex:0 0 auto;font-variant-numeric:tabular-nums;color:var(--yellow)}
 .activity-run{margin-left:auto;color:var(--muted);white-space:nowrap}
 .agent-activity{display:flex;align-items:center;gap:6px;margin-top:5px;font-size:11px;color:#c9d1de}