@yeaft/webchat-agent 1.0.151 → 1.0.152

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "1.0.151",
+  "version": "1.0.152",
   "description": "Remote worker agent for Yeaft Web Code Agent — connects the native Yeaft engine, CLI providers, and workbench tools",
   "main": "index.js",
   "type": "module",

package/yeaft/llm/anthropic.js

@@ -240,7 +240,7 @@ export class AnthropicAdapter extends LLMAdapter {
    * @param {{ model: string, system: string, messages: import('./adapter.js').UnifiedMessage[], tools?: import('./adapter.js').UnifiedToolDef[], maxTokens?: number, effort?: 'low'|'medium'|'high'|'xhigh'|'max', effortSource?: 'user'|'auto', effortContext?: object, signal?: AbortSignal, onRawExchange?: ({rawRequest, rawResponse}) => void }} params
    * @returns {AsyncGenerator<import('./adapter.js').StreamEvent>}
    */
-  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, effortContext, signal, onRawExchange }) {
+  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, effortContext, signal, onRawExchange, onRequestStart }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -272,6 +272,7 @@ export class AnthropicAdapter extends LLMAdapter {
 
     let response;
     try {
+      onRequestStart?.();
       response = await fetch(url, {
         method: 'POST',
         headers,
@@ -328,6 +329,7 @@ export class AnthropicAdapter extends LLMAdapter {
     const responseStatus = response.status;
     let sawStop = false;
     let sawMessageStart = false;
+    let cumulativeOutputTokens = 0;
 
     try {
       while (true) {
@@ -454,12 +456,17 @@ export class AnthropicAdapter extends LLMAdapter {
                 stopReason: this.#mapStopReason(stopReason),
               };
             }
-            // Usage from message_delta
+            // Anthropic message_delta usage is cumulative across the response.
+            // Expose only the newly consumed output tokens so shared accounting
+            // can safely add events from message_start and multiple deltas.
             if (event.usage) {
+              const nextOutputTokens = Math.max(0, Number(event.usage.output_tokens) || 0);
+              const outputTokens = Math.max(0, nextOutputTokens - cumulativeOutputTokens);
+              cumulativeOutputTokens = Math.max(cumulativeOutputTokens, nextOutputTokens);
               yield {
                 type: 'usage',
                 inputTokens: 0, // Only in message_start
-                outputTokens: event.usage.output_tokens || 0,
+                outputTokens,
               };
             }
           } else if (type === 'message_stop') {
@@ -468,10 +475,14 @@ export class AnthropicAdapter extends LLMAdapter {
             sawMessageStart = true;
             // Usage from message_start
             if (event.message?.usage) {
+              cumulativeOutputTokens = Math.max(
+                cumulativeOutputTokens,
+                Math.max(0, Number(event.message.usage.output_tokens) || 0),
+              );
               yield {
                 type: 'usage',
                 inputTokens: event.message.usage.input_tokens || 0,
-                outputTokens: event.message.usage.output_tokens || 0,
+                outputTokens: cumulativeOutputTokens,
                 cacheReadTokens: event.message.usage.cache_read_input_tokens || 0,
                 cacheWriteTokens: event.message.usage.cache_creation_input_tokens || 0,
               };
@@ -518,7 +529,7 @@ export class AnthropicAdapter extends LLMAdapter {
    * models silently drop the param. max_tokens auto-widens to budget+1024
    * when needed.
    */
-  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, effortContext, signal }) {
+  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, effortContext, signal, onRequestStart }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -536,6 +547,7 @@ export class AnthropicAdapter extends LLMAdapter {
 
     let response;
     try {
+      onRequestStart?.();
       response = await fetch(`${this.#baseUrl}/v1/messages`, {
         method: 'POST',
         headers: this.#headers(),

package/yeaft/llm/openai-responses.js

@@ -257,7 +257,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
    * `api-key` headers are auto-redacted (see `redactRawRequest` in
    * `adapter.js`); request-body fields are caller-controlled.
    */
-  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, extraBody, signal, onRawExchange }) {
+  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, extraBody, signal, onRawExchange, onRequestStart }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -297,6 +297,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
 
     let response;
     try {
+      onRequestStart?.();
       response = await fetch(url, {
         method: 'POST',
         headers,
@@ -516,7 +517,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
    * expose them, mirror the stream() instrumentation. Parity with
    * anthropic.js's `call()`.
    */
-  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, extraBody, signal }) {
+  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, extraBody, signal, onRequestStart }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -540,6 +541,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
 
     let response;
     try {
+      onRequestStart?.();
       response = await fetch(`${this.#baseUrl}/responses`, {
         method: 'POST',
         headers: {

package/yeaft/llm/usage-accounting.js

@@ -43,14 +43,10 @@ function addUsage(total, usage) {
   total.totalTokens += normalized.totalTokens;
 }
 
-function hasUsage(usage) {
-  return Object.values(usage).some(value => value > 0);
-}
-
 /**
  * Wrap the shared adapter at the provider-call boundary. Every stream request
- * reports once after its event stream finishes (or aborts after returning
- * usage), and every non-streaming side call reports once after success.
+ * reports once after its event stream finishes or aborts, and every
+ * non-streaming side call reports once whether it succeeds or fails.
  *
  * Parent VP engines, sub-agent engines, Dream, compact, reflection, AMS, and
  * classifiers all reuse this adapter, so none need their own accounting hook.
@@ -58,15 +54,35 @@ function hasUsage(usage) {
 export class UsageAccountingAdapter extends LLMAdapter {
   #adapter;
   #onUsage;
+  #onRequest;
 
-  constructor(adapter, onUsage) {
+  constructor(adapter, onUsage, onRequest = null) {
     super(adapter?.config || {});
     this.#adapter = adapter;
     this.#onUsage = onUsage;
+    this.#onRequest = onRequest;
+  }
+
+  #requestParams(params) {
+    if (typeof this.#onRequest !== 'function') return params;
+    const upstream = params?.onRequestStart;
+    return {
+      ...params,
+      onRequestStart: () => {
+        try {
+          upstream?.();
+        } finally {
+          try {
+            this.#onRequest();
+          } catch (error) {
+            console.warn(`[llm-usage] request callback failed: ${error?.message || error}`);
+          }
+        }
+      },
+    };
   }
 
   #report(usage) {
-    if (!hasUsage(usage)) return;
     try {
       this.#onUsage(usage);
     } catch (error) {
@@ -77,7 +93,7 @@ export class UsageAccountingAdapter extends LLMAdapter {
   async *stream(params) {
     const total = normalizeTokenUsage();
     try {
-      for await (const event of this.#adapter.stream(params)) {
+      for await (const event of this.#adapter.stream(this.#requestParams(params))) {
         if (event?.type === 'usage') addUsage(total, event);
         yield event;
       }
@@ -87,9 +103,14 @@ export class UsageAccountingAdapter extends LLMAdapter {
   }
 
   async call(params) {
-    const result = await this.#adapter.call(params);
-    this.#report(normalizeTokenUsage(result?.usage || {}));
-    return result;
+    let usage = normalizeTokenUsage();
+    try {
+      const result = await this.#adapter.call(this.#requestParams(params));
+      usage = normalizeTokenUsage(result?.usage || {});
+      return result;
+    } finally {
+      this.#report(usage);
+    }
   }
 
   refreshProviders(providers) {
@@ -105,8 +126,8 @@ export class UsageAccountingAdapter extends LLMAdapter {
   }
 }
 
-export function withUsageAccounting(adapter, onUsage) {
+export function withUsageAccounting(adapter, onUsage, onRequest = null) {
   return typeof onUsage === 'function'
-    ? new UsageAccountingAdapter(adapter, onUsage)
+    ? new UsageAccountingAdapter(adapter, onUsage, onRequest)
     : adapter;
 }

package/yeaft/work-center/controller.js

@@ -82,6 +82,12 @@ function normalizeTerminalResult(result, action) {
       : null,
     loopCount: Math.max(0, Number(result.loopCount) || 0),
     toolCount: Math.max(0, Number(result.toolCount) || 0),
+    llmRequestCount: Math.max(0, Number(result.llmRequestCount) || 0),
+    inputTokens: Math.max(0, Number(result.inputTokens) || 0),
+    outputTokens: Math.max(0, Number(result.outputTokens) || 0),
+    cacheReadTokens: Math.max(0, Number(result.cacheReadTokens) || 0),
+    cacheWriteTokens: Math.max(0, Number(result.cacheWriteTokens) || 0),
+    totalTokens: Math.max(0, Number(result.totalTokens) || 0),
     acceptanceChecks: Array.isArray(result.acceptanceChecks) ? result.acceptanceChecks : [],
     checkpoint: result.checkpoint && typeof result.checkpoint === 'object'
       ? result.checkpoint : null,

package/yeaft/work-center/projection.js

@@ -7,22 +7,45 @@ function count(value) {
   return Math.max(0, Number(value) || 0);
 }
 
+function emptyExecutionStats() {
+  return {
+    llmRequestCount: 0,
+    loopCount: 0,
+    toolCount: 0,
+    inputTokens: 0,
+    outputTokens: 0,
+    cacheReadTokens: 0,
+    cacheWriteTokens: 0,
+    totalTokens: 0,
+  };
+}
+
+function executionStats(value) {
+  const stats = emptyExecutionStats();
+  for (const key of Object.keys(stats)) stats[key] = count(value?.[key]);
+  return stats;
+}
+
+function sumExecutionStats(values) {
+  return values.reduce((total, value) => {
+    const next = executionStats(value);
+    for (const key of Object.keys(total)) total[key] += next[key];
+    return total;
+  }, emptyExecutionStats());
+}
+
 function actionExecution(action, runs) {
   const matchingRuns = Array.isArray(runs)
     ? runs.filter(run => run?.actionId === action?.id)
     : [];
   if (matchingRuns.length === 0) {
     return {
-      loopCount: count(action?.loopCount),
-      toolCount: count(action?.toolCount),
+      ...executionStats(action),
       response: '',
       progressRevision: 0,
     };
   }
-  const stats = matchingRuns.reduce((total, run) => ({
-    loopCount: total.loopCount + count(run.loopCount),
-    toolCount: total.toolCount + count(run.toolCount),
-  }), { loopCount: 0, toolCount: 0 });
+  const stats = sumExecutionStats(matchingRuns);
   const latest = [...matchingRuns].sort((left, right) => (
     count(right.progressRevision) - count(left.progressRevision)
       || count(right.startedAt) - count(left.startedAt)
@@ -65,6 +88,7 @@ function projectAction(action, runs) {
     assignmentPolicy: projectAssignmentPolicy(action.assignmentPolicy),
     requiredRole: action.requiredRole || '',
     status: action.status,
+    executionStats: executionStats(execution),
     loopCount: execution.loopCount,
     toolCount: execution.toolCount,
     response: execution.response,
@@ -79,6 +103,7 @@ function projectActionStats(detail) {
     return {
       id: projected.id,
       status: projected.status,
+      executionStats: projected.executionStats,
       loopCount: projected.loopCount,
       toolCount: projected.toolCount,
       response: projected.response,
@@ -97,7 +122,7 @@ function waitingReason(detail) {
 
 /**
  * Authenticated browser detail DTO. Raw execution records stay Agent-local;
- * the browser receives only Action status/counts plus the explicit user-facing response.
+ * the browser receives only aggregate execution stats plus the explicit user-facing response.
  */
 export function projectWorkItemDetail(detail) {
   if (!detail) return null;
@@ -112,6 +137,7 @@ export function projectWorkItemDetail(detail) {
     planningMode: detail.workflowSnapshot?.planningMode || 'static',
     status: detail.status,
     currentActionId: detail.currentActionId || null,
+    executionStats: sumExecutionStats(Array.isArray(detail.runs) ? detail.runs : []),
     reuseMemory: detail.reuseMemory !== false,
     waitingReason: waitingReason(detail),
     origin: detail.origin?.sessionId ? { sessionId: detail.origin.sessionId } : null,
@@ -142,6 +168,7 @@ export function projectWorkItemSummary(detail) {
       status: detail.status,
       currentActionId: detail.currentActionId || null,
       currentAction: null,
+      executionStats: executionStats(detail.executionStats),
       origin: detail.origin?.sessionId ? { sessionId: detail.origin.sessionId } : null,
       linkedSessionIds: Array.isArray(detail.linkedSessionIds) ? detail.linkedSessionIds : [],
       attachmentCount: Array.isArray(detail.attachments) ? detail.attachments.length : 0,
@@ -160,6 +187,7 @@ export function projectWorkItemSummary(detail) {
     planningMode: detail.workflowSnapshot?.planningMode || 'static',
     status: detail.status,
     currentActionId: detail.currentActionId || null,
+    executionStats: sumExecutionStats(Array.isArray(detail.runs) ? detail.runs : []),
     currentAction: projectedAction ? {
       id: projectedAction.id,
       type: projectedAction.type,

package/yeaft/work-center/runner.js

@@ -12,6 +12,7 @@ import { formatPickedForInjection } from '../sessions/pre-flow.js';
 import { existsSync, lstatSync, realpathSync } from 'node:fs';
 import path from 'node:path';
 import { buildWorkItemAttachmentContext } from './attachments.js';
+import { withUsageAccounting } from '../llm/usage-accounting.js';
 import {
   appendCheckpointToolEvent,
   renderActionResumeBlock,
@@ -499,34 +500,24 @@ export class WorkItemRunner {
     );
     if (!snapshotsWritten) throw new Error('Work Center Run lost its lease before execution');
 
-    const engine = new Engine({
-      adapter: runtime.adapter,
-      trace: new NullTrace(),
-      config,
-      conversationStore: null,
-      memoryIndex: null,
-      amsRegistry: null,
-      toolRegistry,
-      skillManager: null,
-      mcpManager: null,
-      // WorkItem execution has its own durable Run/Event record. Do not let
-      // the Engine create Session exec logs, archives, or shared tool stats.
-      yeaftDir: null,
-      toolStats: null,
-      taskManager: null,
-      vpId: vp.id,
-    });
-
     let text = '';
     let loopCount = 0;
     let toolCount = 0;
     let checkpoint = null;
     const toolInputs = new Map();
+    const usageStats = {
+      llmRequestCount: 0,
+      inputTokens: 0,
+      outputTokens: 0,
+      cacheReadTokens: 0,
+      cacheWriteTokens: 0,
+      totalTokens: 0,
+    };
     let lastProgressAt = 0;
+    const executionStats = () => ({ loopCount, toolCount, ...usageStats });
     const currentProgress = () => ({
       response: publicWorkItemResponse(text),
-      loopCount,
-      toolCount,
+      ...executionStats(),
       checkpoint,
     });
     const reportProgress = (force = false) => {
@@ -537,6 +528,34 @@ export class WorkItemRunner {
       return onProgress(currentProgress());
     };
     if (typeof registerProgressReader === 'function') registerProgressReader(currentProgress);
+    const adapter = withUsageAccounting(runtime.adapter, usage => {
+      usageStats.inputTokens += usage.inputTokens;
+      usageStats.outputTokens += usage.outputTokens;
+      usageStats.cacheReadTokens += usage.cacheReadTokens;
+      usageStats.cacheWriteTokens += usage.cacheWriteTokens;
+      usageStats.totalTokens += usage.totalTokens;
+      reportProgress(true);
+    }, () => {
+      usageStats.llmRequestCount += 1;
+      reportProgress(true);
+    });
+    const engine = new Engine({
+      adapter,
+      trace: new NullTrace(),
+      config,
+      conversationStore: null,
+      memoryIndex: null,
+      amsRegistry: null,
+      toolRegistry,
+      skillManager: null,
+      mcpManager: null,
+      // WorkItem execution has its own durable Run/Event record. Do not let
+      // the Engine create Session exec logs, archives, or shared tool stats.
+      yeaftDir: null,
+      toolStats: null,
+      taskManager: null,
+      vpId: vp.id,
+    });
     try {
       const prompt = `${executionAction.instruction}${resumeBlock}${attachmentContext.promptBlock}${memoryBlock}${completionContract(executionAction, workItem)}`;
       const promptParts = attachmentContext.promptParts.length > 0
@@ -567,15 +586,10 @@ export class WorkItemRunner {
           });
         }
         if (event?.type === 'text_delta' && typeof event.text === 'string') text += event.text;
-        reportProgress();
+        reportProgress(event?.type === 'loop');
       }
     } catch (error) {
-      error.workItemExecutionStats = {
-        response: publicWorkItemResponse(text),
-        loopCount,
-        toolCount,
-        checkpoint,
-      };
+      error.workItemExecutionStats = currentProgress();
       throw error;
     } finally {
       try { engine.abort?.('work_item_run_finished'); } catch {}
@@ -585,8 +599,7 @@ export class WorkItemRunner {
     return {
       ...parseStructuredResult(text, executionAction.type),
       response,
-      loopCount,
-      toolCount,
+      ...executionStats(),
       checkpoint,
     };
   }

package/yeaft/work-center/store.js

@@ -48,6 +48,16 @@ function mapWorkItem(row) {
     origin: parseJson(row.origin, null),
     linkedSessionIds: parseJson(row.linked_session_ids, []),
     attachments: parseJson(row.attachments, []),
+    executionStats: {
+      llmRequestCount: Math.max(0, Number(row.usage_llm_request_count) || 0),
+      loopCount: Math.max(0, Number(row.usage_loop_count) || 0),
+      toolCount: Math.max(0, Number(row.usage_tool_count) || 0),
+      inputTokens: Math.max(0, Number(row.usage_input_tokens) || 0),
+      outputTokens: Math.max(0, Number(row.usage_output_tokens) || 0),
+      cacheReadTokens: Math.max(0, Number(row.usage_cache_read_tokens) || 0),
+      cacheWriteTokens: Math.max(0, Number(row.usage_cache_write_tokens) || 0),
+      totalTokens: Math.max(0, Number(row.usage_total_tokens) || 0),
+    },
     createdAt: row.created_at,
     updatedAt: row.updated_at,
   };
@@ -102,6 +112,12 @@ function mapRun(row) {
     contractPatch: parseJson(row.contract_patch, null),
     loopCount: Math.max(0, Number(row.loop_count) || 0),
     toolCount: Math.max(0, Number(row.tool_count) || 0),
+    llmRequestCount: Math.max(0, Number(row.llm_request_count) || 0),
+    inputTokens: Math.max(0, Number(row.input_tokens) || 0),
+    outputTokens: Math.max(0, Number(row.output_tokens) || 0),
+    cacheReadTokens: Math.max(0, Number(row.cache_read_tokens) || 0),
+    cacheWriteTokens: Math.max(0, Number(row.cache_write_tokens) || 0),
+    totalTokens: Math.max(0, Number(row.total_tokens) || 0),
     progressRevision: Math.max(0, Number(row.progress_revision) || 0),
     checkpoint: normalizeActionCheckpoint(parseJson(row.checkpoint, null)),
   };
@@ -220,6 +236,12 @@ export class WorkItemStore {
         response TEXT NOT NULL DEFAULT '',
         loop_count INTEGER NOT NULL DEFAULT 0,
         tool_count INTEGER NOT NULL DEFAULT 0,
+        llm_request_count INTEGER NOT NULL DEFAULT 0,
+        input_tokens INTEGER NOT NULL DEFAULT 0,
+        output_tokens INTEGER NOT NULL DEFAULT 0,
+        cache_read_tokens INTEGER NOT NULL DEFAULT 0,
+        cache_write_tokens INTEGER NOT NULL DEFAULT 0,
+        total_tokens INTEGER NOT NULL DEFAULT 0,
         progress_revision INTEGER NOT NULL DEFAULT 0,
         checkpoint TEXT
       );
@@ -274,6 +296,18 @@ export class WorkItemStore {
     if (!hasColumn(this.db, 'runs', 'tool_count')) {
       this.db.exec('ALTER TABLE runs ADD COLUMN tool_count INTEGER NOT NULL DEFAULT 0');
     }
+    for (const column of [
+      'llm_request_count',
+      'input_tokens',
+      'output_tokens',
+      'cache_read_tokens',
+      'cache_write_tokens',
+      'total_tokens',
+    ]) {
+      if (!hasColumn(this.db, 'runs', column)) {
+        this.db.exec(`ALTER TABLE runs ADD COLUMN ${column} INTEGER NOT NULL DEFAULT 0`);
+      }
+    }
     if (!hasColumn(this.db, 'runs', 'response')) {
       this.db.exec("ALTER TABLE runs ADD COLUMN response TEXT NOT NULL DEFAULT ''");
     }
@@ -437,22 +471,32 @@ export class WorkItemStore {
     const where = [];
     const values = [];
     if (typeof filters.status === 'string' && filters.status) {
-      where.push('status = ?');
+      where.push('w.status = ?');
       values.push(filters.status);
     }
     if (typeof filters.sessionId === 'string' && filters.sessionId.trim()) {
       const sessionId = filters.sessionId.trim();
-      where.push('(instr(origin, ?) > 0 OR instr(linked_session_ids, ?) > 0)');
+      where.push('(instr(w.origin, ?) > 0 OR instr(w.linked_session_ids, ?) > 0)');
       values.push(`\"sessionId\":${JSON.stringify(sessionId)}`, JSON.stringify(sessionId));
     }
     if (typeof filters.search === 'string' && filters.search.trim()) {
-      where.push('(title LIKE ? OR goal LIKE ?)');
+      where.push('(w.title LIKE ? OR w.goal LIKE ?)');
       const query = `%${filters.search.trim()}%`;
       values.push(query, query);
     }
     const limit = Math.min(Math.max(Number(filters.limit) || 100, 1), 500);
-    const sql = `SELECT * FROM work_items ${where.length ? `WHERE ${where.join(' AND ')}` : ''}
-      ORDER BY updated_at DESC LIMIT ?`;
+    const sql = `SELECT w.*,
+        COALESCE(SUM(r.llm_request_count), 0) AS usage_llm_request_count,
+        COALESCE(SUM(r.loop_count), 0) AS usage_loop_count,
+        COALESCE(SUM(r.tool_count), 0) AS usage_tool_count,
+        COALESCE(SUM(r.input_tokens), 0) AS usage_input_tokens,
+        COALESCE(SUM(r.output_tokens), 0) AS usage_output_tokens,
+        COALESCE(SUM(r.cache_read_tokens), 0) AS usage_cache_read_tokens,
+        COALESCE(SUM(r.cache_write_tokens), 0) AS usage_cache_write_tokens,
+        COALESCE(SUM(r.total_tokens), 0) AS usage_total_tokens
+      FROM work_items w LEFT JOIN runs r ON r.work_item_id = w.id
+      ${where.length ? `WHERE ${where.join(' AND ')}` : ''}
+      GROUP BY w.id ORDER BY w.updated_at DESC LIMIT ?`;
     return this.db.prepare(sql).all(...values, limit).map(mapWorkItem);
   }
 
@@ -762,7 +806,8 @@ export class WorkItemStore {
       const hasFinalProgress = finalProgress && typeof finalProgress === 'object';
       const runChanged = hasFinalProgress
         ? this.db.prepare(`UPDATE runs SET status = 'interrupted', ended_at = ?, error = ?,
-          response = ?, loop_count = ?, tool_count = ?, checkpoint = ?,
+          response = ?, loop_count = ?, tool_count = ?, llm_request_count = ?, input_tokens = ?,
+          output_tokens = ?, cache_read_tokens = ?, cache_write_tokens = ?, total_tokens = ?, checkpoint = ?,
           progress_revision = progress_revision + 1
           WHERE id = ? AND owner_boot_id = ? AND lease_epoch = ? AND status = 'running'`).run(
           now,
@@ -770,6 +815,12 @@ export class WorkItemStore {
           normalizeRunResponse(finalProgress.response),
           Math.max(0, Number(finalProgress.loopCount) || 0),
           Math.max(0, Number(finalProgress.toolCount) || 0),
+          Math.max(0, Number(finalProgress.llmRequestCount) || 0),
+          Math.max(0, Number(finalProgress.inputTokens) || 0),
+          Math.max(0, Number(finalProgress.outputTokens) || 0),
+          Math.max(0, Number(finalProgress.cacheReadTokens) || 0),
+          Math.max(0, Number(finalProgress.cacheWriteTokens) || 0),
+          Math.max(0, Number(finalProgress.totalTokens) || 0),
           stringify(normalizeActionCheckpoint(finalProgress.checkpoint)),
           runId,
           ownerBootId,
@@ -835,12 +886,19 @@ export class WorkItemStore {
       const active = this.#activeRunRow(runId, ownerBootId, leaseEpoch, true);
       if (!active) return null;
       const checkpoint = normalizeActionCheckpoint(progress.checkpoint);
-      const result = this.db.prepare(`UPDATE runs SET response = ?, loop_count = ?, tool_count = ?, checkpoint = ?,
-        progress_revision = progress_revision + 1 WHERE id = ? AND owner_boot_id = ?
-        AND lease_epoch = ? AND status = 'running'`).run(
+      const result = this.db.prepare(`UPDATE runs SET response = ?, loop_count = ?, tool_count = ?,
+        llm_request_count = ?, input_tokens = ?, output_tokens = ?, cache_read_tokens = ?,
+        cache_write_tokens = ?, total_tokens = ?, checkpoint = ?, progress_revision = progress_revision + 1
+        WHERE id = ? AND owner_boot_id = ? AND lease_epoch = ? AND status = 'running'`).run(
         normalizeRunResponse(progress.response),
         Math.max(0, Number(progress.loopCount) || 0),
         Math.max(0, Number(progress.toolCount) || 0),
+        Math.max(0, Number(progress.llmRequestCount) || 0),
+        Math.max(0, Number(progress.inputTokens) || 0),
+        Math.max(0, Number(progress.outputTokens) || 0),
+        Math.max(0, Number(progress.cacheReadTokens) || 0),
+        Math.max(0, Number(progress.cacheWriteTokens) || 0),
+        Math.max(0, Number(progress.totalTokens) || 0),
         stringify(checkpoint),
         runId,
         ownerBootId,
@@ -886,7 +944,9 @@ export class WorkItemStore {
       const now = this.now();
       this.db.prepare(`UPDATE runs SET status = ?, ended_at = ?, response = ?, summary = ?, evidence = ?,
         waiting_reason = ?, error = ?, review_decision = ?, contract_patch = ?, checkpoint = ?,
-        loop_count = ?, tool_count = ?, progress_revision = progress_revision + 1 WHERE id = ?`).run(
+        loop_count = ?, tool_count = ?, llm_request_count = ?, input_tokens = ?, output_tokens = ?,
+        cache_read_tokens = ?, cache_write_tokens = ?, total_tokens = ?,
+        progress_revision = progress_revision + 1 WHERE id = ?`).run(
         result.outcome,
         now,
         normalizeRunResponse(result.response),
@@ -899,6 +959,12 @@ export class WorkItemStore {
         stringify(normalizeActionCheckpoint(result.checkpoint)),
         Math.max(0, Number(result.loopCount) || 0),
         Math.max(0, Number(result.toolCount) || 0),
+        Math.max(0, Number(result.llmRequestCount) || 0),
+        Math.max(0, Number(result.inputTokens) || 0),
+        Math.max(0, Number(result.outputTokens) || 0),
+        Math.max(0, Number(result.cacheReadTokens) || 0),
+        Math.max(0, Number(result.cacheWriteTokens) || 0),
+        Math.max(0, Number(result.totalTokens) || 0),
         runId,
       );
       this.onTransitionStep?.('after_run_update');

package/yeaft/work-center/watcher.js

@@ -36,6 +36,8 @@ export class WorkItemWatcher {
     if (this.timer) clearInterval(this.timer);
     this.timer = null;
     const active = Array.from(this.activeRuns.values());
+    for (const entry of active) entry.abortController.abort('watcher_stopped');
+    await Promise.allSettled(active.map(entry => entry.promise));
     const failures = [];
     for (const entry of active) {
       try {
@@ -50,11 +52,8 @@ export class WorkItemWatcher {
       } catch (error) {
         entry.interrupted = false;
         failures.push(error);
-      } finally {
-        entry.abortController.abort('watcher_stopped');
       }
     }
-    await Promise.allSettled(active.map(entry => entry.promise));
     if (failures.length > 0) {
       throw new AggregateError(failures, 'Could not persist one or more Work Center interruptions');
     }
@@ -140,6 +139,12 @@ export class WorkItemWatcher {
         error: err?.message || String(err),
         loopCount: err?.workItemExecutionStats?.loopCount || 0,
         toolCount: err?.workItemExecutionStats?.toolCount || 0,
+        llmRequestCount: err?.workItemExecutionStats?.llmRequestCount || 0,
+        inputTokens: err?.workItemExecutionStats?.inputTokens || 0,
+        outputTokens: err?.workItemExecutionStats?.outputTokens || 0,
+        cacheReadTokens: err?.workItemExecutionStats?.cacheReadTokens || 0,
+        cacheWriteTokens: err?.workItemExecutionStats?.cacheWriteTokens || 0,
+        totalTokens: err?.workItemExecutionStats?.totalTokens || 0,
         checkpoint: err?.workItemExecutionStats?.checkpoint || null,
       };
     }