pi2web 0.2.6 → 0.2.12

package/dist/sotto-stt.js

@@ -0,0 +1,217 @@
+import { WebSocket } from 'ws';
+export const DEFAULT_SOTTO_ENDPOINT = 'http://127.0.0.1:8766';
+const invalid = (message, code = 'INVALID_SOTTO_SETTINGS') => Object.assign(new Error(message), { code });
+/** No arbitrary URLs: the Sotto Web API is local-only, and this server is the authenticated proxy. */
+export function validateSottoEndpoint(value) {
+    if (typeof value !== 'string')
+        throw invalid('Sotto endpoint must be a loopback HTTP URL');
+    let url;
+    try {
+        url = new URL(value);
+    }
+    catch {
+        throw invalid('Sotto endpoint must be a valid URL');
+    }
+    if (url.protocol !== 'http:' ||
+        url.hostname !== '127.0.0.1' ||
+        !url.port ||
+        url.username ||
+        url.password ||
+        (url.pathname !== '/' && url.pathname !== '') ||
+        url.search ||
+        url.hash)
+        throw invalid('Sotto endpoint must be http://127.0.0.1:<port> with no path or credentials');
+    return url.origin;
+}
+/** An authenticated browser socket relays only audio and start/stop, never a Sotto control API. */
+function proxySottoStream(client, url) {
+    const upstream = new WebSocket(url, { handshakeTimeout: 5000, maxPayload: 65536 });
+    let started = false, finished = false, closed = false;
+    let timer;
+    const close = () => {
+        closed = true;
+        clearTimeout(timer);
+        if (upstream.readyState === WebSocket.OPEN)
+            upstream.close();
+        else if (upstream.readyState === WebSocket.CONNECTING)
+            upstream.terminate();
+    };
+    const fail = (message, code = 1011) => {
+        if (closed)
+            return;
+        if (client.readyState === WebSocket.OPEN) {
+            client.send(JSON.stringify({ type: 'error', message }));
+            client.close(code);
+        }
+        close();
+    };
+    // Silence still contains PCM. Only a stalled capture/connection times out, not a long dictation.
+    const touch = () => {
+        clearTimeout(timer);
+        timer = setTimeout(() => fail('Audio stream stalled: no microphone frames received for 30 seconds'), 30000);
+    };
+    touch();
+    client.on('close', close);
+    client.on('error', close);
+    upstream.on('error', () => fail('Sotto streaming service is unavailable'));
+    upstream.on('close', code => {
+        if (closed)
+            return;
+        if (!finished || code !== 1000)
+            fail('Sotto streaming connection closed unexpectedly');
+        else {
+            clearTimeout(timer);
+            if (client.readyState === WebSocket.OPEN)
+                client.close(1000);
+        }
+    });
+    upstream.on('message', (data, binary) => {
+        if (closed || binary || client.readyState !== WebSocket.OPEN)
+            return;
+        if (client.bufferedAmount > 1_000_000) {
+            fail('Transcription receiver is too slow', 1009);
+            return;
+        }
+        client.send(data.toString());
+    });
+    client.on('message', (data, binary) => {
+        if (closed || finished || client.readyState !== WebSocket.OPEN)
+            return;
+        const length = data.byteLength;
+        if (upstream.readyState !== WebSocket.OPEN) {
+            fail('Sotto streaming service is not ready');
+            return;
+        }
+        if (!binary) {
+            try {
+                const value = JSON.parse(data.toString());
+                if (!started &&
+                    value.type === 'start' &&
+                    value.format === 'float32' &&
+                    Number.isInteger(value.sample_rate) &&
+                    value.sample_rate >= 16000 &&
+                    value.sample_rate <= 96000) {
+                    started = true;
+                    upstream.send(JSON.stringify({ type: 'start', format: 'float32', sample_rate: value.sample_rate }));
+                    touch();
+                }
+                else if (started && value.type === 'stop') {
+                    finished = true;
+                    upstream.send(JSON.stringify({ type: 'stop' }));
+                }
+                else
+                    fail('Invalid audio protocol', 1008);
+            }
+            catch {
+                fail('Invalid audio protocol', 1008);
+            }
+        }
+        else if (started &&
+            length > 0 &&
+            length % 4 === 0 &&
+            length <= 32768 &&
+            upstream.bufferedAmount < 1_000_000) {
+            upstream.send(data, { binary: true });
+            touch();
+        }
+        else
+            fail('Invalid audio frame or Sotto cannot keep up with microphone audio', 1009);
+    });
+    return fail;
+}
+export class SottoSttService {
+    metadata;
+    fetcher;
+    streams = new Set();
+    settings = { enabled: false, endpoint: DEFAULT_SOTTO_ENDPOINT };
+    /** Bumped by every update, so a slow enable probe cannot overwrite a newer choice. */
+    updates = 0;
+    constructor(metadata, fetcher = fetch) {
+        this.metadata = metadata;
+        this.fetcher = fetcher;
+    }
+    load() {
+        const saved = this.metadata.getSetting('sottoStt');
+        if (saved && typeof saved.enabled === 'boolean') {
+            try {
+                this.settings = { enabled: saved.enabled, endpoint: validateSottoEndpoint(saved.endpoint) };
+            }
+            catch {
+                /* Corrupt or obsolete settings must not enable an untrusted destination. */
+            }
+        }
+    }
+    getSettings() {
+        return { ...this.settings };
+    }
+    streamUrl() {
+        if (!this.settings.enabled)
+            throw invalid('Sotto transcription is disabled', 'SOTTO_DISABLED');
+        const port = Number(new URL(this.settings.endpoint).port) + 2;
+        if (port > 65535)
+            throw invalid('Invalid Sotto streaming port');
+        return `ws://127.0.0.1:${port}/asr`;
+    }
+    stream(client) {
+        const stop = proxySottoStream(client, this.streamUrl());
+        this.streams.add(stop);
+        client.once('close', () => this.streams.delete(stop));
+    }
+    async request(endpoint, path, init, timeoutMs) {
+        const response = await this.fetcher(`${endpoint}${path}`, {
+            ...init,
+            redirect: 'error',
+            signal: AbortSignal.timeout(timeoutMs)
+        }).catch(error => {
+            throw invalid(`Cannot reach local Sotto: ${error.message}`, 'SOTTO_UNAVAILABLE');
+        });
+        if (!response.ok) {
+            const message = (await response.json().catch(() => ({})));
+            throw invalid(message.error || `Sotto returned HTTP ${response.status}`, 'SOTTO_UNAVAILABLE');
+        }
+        return response;
+    }
+    async update(input) {
+        if (!input || typeof input !== 'object')
+            throw invalid('enabled and endpoint are required');
+        const { enabled, endpoint } = input;
+        if (typeof enabled !== 'boolean')
+            throw invalid('enabled must be a boolean');
+        const next = { enabled, endpoint: validateSottoEndpoint(endpoint) };
+        const update = ++this.updates;
+        if (enabled) {
+            const capabilities = await (await this.request(next.endpoint, '/api/capabilities', { method: 'GET' }, 4000)).json();
+            if (capabilities?.stt !== true ||
+                capabilities?.format !== 'pcm-f32le' ||
+                capabilities?.sample_rate !== 16000 ||
+                capabilities?.stream_port !== Number(new URL(next.endpoint).port) + 2)
+                throw invalid('Sotto does not support live PCM transcription', 'SOTTO_UNAVAILABLE');
+            if (update !== this.updates)
+                throw invalid('Sotto settings were changed by a newer request', 'SOTTO_SETTINGS_SUPERSEDED');
+        }
+        this.metadata.setSetting('sottoStt', next);
+        const changed = !next.enabled || next.endpoint !== this.settings.endpoint;
+        this.settings = next;
+        if (changed) {
+            for (const stop of this.streams)
+                stop('Sotto 语音输入已关闭或服务地址已更改');
+            this.streams.clear();
+        }
+        return this.getSettings();
+    }
+    async transcribe(audio) {
+        if (!this.settings.enabled)
+            throw invalid('Sotto transcription is disabled', 'SOTTO_DISABLED');
+        if (!audio.length || audio.length % 4 || audio.length > 16000 * 4 * 60)
+            throw invalid('Expected at most 60 seconds of 16 kHz float32 PCM');
+        const response = await this.request(this.settings.endpoint, '/api/transcribe', {
+            method: 'POST',
+            headers: { 'content-type': 'application/octet-stream' },
+            body: new Uint8Array(audio)
+        }, 120000);
+        const body = (await response.json());
+        if (typeof body.text !== 'string')
+            throw invalid('Invalid response from Sotto', 'SOTTO_UNAVAILABLE');
+        return { text: body.text };
+    }
+}

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": "0.2.6",
+  "version": "0.2.12",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "pi2web",
-      "version": "0.2.6",
+      "version": "0.2.12",
       "license": "MIT",
       "dependencies": {
         "@earendil-works/chord": "1.1.0",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi2web",
-  "version": "0.2.6",
+  "version": "0.2.12",
   "description": "Secure remote web client server for Pi coding agent",
   "type": "module",
   "author": "sxwxs",

package/packages/protocol/openapi.yaml

@@ -21,11 +21,32 @@ paths:
     post: {operationId: createAgent, requestBody: {$ref: '#/components/requestBodies/Json'}, responses: {'201': {$ref: '#/components/responses/Success'}}}
   /agents/{id}/timeline:
     get: {operationId: agentTimeline, description: 'Older history as display items (see timelineItem in websocket.schema.json): about `limit` items before item `before`, never splitting a tool call from its result. Data {sessionId, items, total, start, hasMore}. Viewers get the newest page from the WebSocket focus snapshot. 404 when `before` is not on the current branch.', parameters: [{$ref: '#/components/parameters/Id'}, {name: limit, in: query, schema: {type: integer, minimum: 1, maximum: 500, default: 50}}, {name: before, in: query, schema: {type: string}}], responses: {'200': {$ref: '#/components/responses/Success'}}}
+  /agents/{id}/performance:
+    get: {operationId: agentPerformance, description: 'Current branch wall-clock model and tool spans recorded by recent runs; older sessions return only available tool durations. Data {sessionId, runs, legacyTools}. Parallel tools overlap.', parameters: [{$ref: '#/components/parameters/Id'}], responses: {'200': {$ref: '#/components/responses/Success'}}}
   /agents/{id}/sessions/{sessionId}/items/{itemId}:
     get: {operationId: agentTimelineItem, description: 'Full body of one timeline item, data {id, body}. Items never change, so it is served with Cache-Control private, max-age=31536000, immutable. 404 when the Agent is no longer on that session or branch.', parameters: [{$ref: '#/components/parameters/Id'}, {name: sessionId, in: path, required: true, schema: {type: string}}, {name: itemId, in: path, required: true, schema: {type: string}}], responses: {'200': {$ref: '#/components/responses/Success'}}}
+  /agents/{id}/system-prompt:
+    get: {operationId: getAgentSystemPrompt, parameters: [{$ref: '#/components/parameters/Id'}], responses: {'200': {$ref: '#/components/responses/Success'}}}
+    post: {operationId: changeAgentSystemPrompt, description: 'Only before the first conversation message; body {templateId?: string}, empty ID selects the Pi default.', parameters: [{$ref: '#/components/parameters/Id'}], requestBody: {$ref: '#/components/requestBodies/Json'}, responses: {'200': {$ref: '#/components/responses/Success'}}}
+  /sotto-stt:
+    get: {operationId: getSottoSttSettings, description: 'Backend-wide Sotto transcription setting; disabled by default, endpoint defaults to http://127.0.0.1:8766.', responses: {'200': {$ref: '#/components/responses/Success'}}}
+    put: {operationId: setSottoSttSettings, description: 'Body {enabled: boolean, endpoint: http://127.0.0.1:<port>}; probes local Sotto capabilities before enabling. Requires full access token.', requestBody: {$ref: '#/components/requestBodies/Json'}, responses: {'200': {$ref: '#/components/responses/Success'}}}
+  /transcribe:
+    post: {operationId: transcribeDraft, description: 'Backend-wide authenticated transcription, not tied to an Agent. Audio format follows the configured Sotto or STT service; returns {data:{text}}.', responses: {'200': {$ref: '#/components/responses/Success'}}}
+  /transcribe/ws:
+    get: {operationId: streamSottoTranscription, description: 'Authenticated backend-wide Sotto transcription WebSocket; start/PCM/stop protocol matches the Agent route.', responses: {'101': {description: 'WebSocket upgrade'}}}
+  /agents/{id}/transcribe:
+    post: {operationId: transcribeAgentDraft, description: 'When Sotto is enabled, body is 16 kHz mono float32 little-endian PCM (application/octet-stream; max 60 s); otherwise uses configured OpenAI-compatible STT audio. Returns {data:{text}} without submitting to the Agent.', parameters: [{$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};
   }