@yeaft/webchat-agent 1.0.141 → 1.0.143

package/context.js

@@ -1,3 +1,5 @@
+import { randomUUID } from 'crypto';
+
 // 共享上下文对象 - 所有模块通过 import 访问
 // 由 index.js 在启动时初始化
 
@@ -22,6 +24,7 @@ export default {
   // Slash command 描述映射: { commandName: description } — 从 plugin commands/*.md 提取
   slashCommandDescriptions: {},
   agentMetrics: {
+    metricEpoch: randomUUID(),
     chatTurns: 0,
     yeaftTurns: 0,
     sessionsCreated: 0,

package/metrics.js

@@ -10,6 +10,9 @@ function ensureMetrics() {
     ctx.agentMetrics = {};
   }
   const metrics = ctx.agentMetrics;
+  if (typeof metrics.metricEpoch !== 'string' || !metrics.metricEpoch) {
+    metrics.metricEpoch = `legacy-${process.pid}-${Date.now()}`;
+  }
   metrics.chatTurns = finiteNumber(metrics.chatTurns);
   metrics.yeaftTurns = finiteNumber(metrics.yeaftTurns);
   metrics.sessionsCreated = finiteNumber(metrics.sessionsCreated);
@@ -78,6 +81,7 @@ export function recordAgentTokenUsage(usage = {}) {
 export function snapshotAgentMetrics() {
   const metrics = ensureMetrics();
   return {
+    metricEpoch: metrics.metricEpoch,
     chatTurns: metrics.chatTurns,
     yeaftTurns: metrics.yeaftTurns,
     totalTurns: metrics.chatTurns + metrics.yeaftTurns,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "1.0.141",
+  "version": "1.0.143",
   "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/openai-responses.js

@@ -572,6 +572,9 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
       usage: {
         inputTokens: usage.input_tokens || 0,
         outputTokens: usage.output_tokens || 0,
+        cacheReadTokens: usage.input_tokens_details?.cached_tokens || 0,
+        cacheWriteTokens: 0,
+        cacheTokensAreIncludedInInput: true,
       },
     };
   }

package/yeaft/llm/usage-accounting.js

@@ -0,0 +1,112 @@
+import { LLMAdapter } from './adapter.js';
+
+function tokenCount(value) {
+  const number = Number(value);
+  return Number.isFinite(number) && number > 0 ? number : 0;
+}
+
+/**
+ * Normalize one provider usage payload without double-counting cached input.
+ * OpenAI includes cached tokens in input_tokens; Anthropic reports cache input
+ * separately. Adapters expose that distinction through
+ * `cacheTokensAreIncludedInInput`.
+ */
+export function normalizeTokenUsage(usage = {}) {
+  const inputTokens = tokenCount(usage.inputTokens ?? usage.input_tokens);
+  const outputTokens = tokenCount(usage.outputTokens ?? usage.output_tokens);
+  const cacheReadTokens = tokenCount(
+    usage.cacheReadTokens ?? usage.cache_read_tokens ?? usage.cache_read_input_tokens
+  );
+  const cacheWriteTokens = tokenCount(
+    usage.cacheWriteTokens ?? usage.cache_write_tokens ?? usage.cache_creation_input_tokens
+  );
+  const explicitTotal = tokenCount(usage.totalTokens ?? usage.total_tokens);
+  const cacheInputTokens = usage.cacheTokensAreIncludedInInput === true
+    ? 0
+    : cacheReadTokens + cacheWriteTokens;
+
+  return {
+    inputTokens,
+    outputTokens,
+    cacheReadTokens,
+    cacheWriteTokens,
+    totalTokens: explicitTotal || inputTokens + outputTokens + cacheInputTokens,
+  };
+}
+
+function addUsage(total, usage) {
+  const normalized = normalizeTokenUsage(usage);
+  total.inputTokens += normalized.inputTokens;
+  total.outputTokens += normalized.outputTokens;
+  total.cacheReadTokens += normalized.cacheReadTokens;
+  total.cacheWriteTokens += normalized.cacheWriteTokens;
+  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.
+ *
+ * Parent VP engines, sub-agent engines, Dream, compact, reflection, AMS, and
+ * classifiers all reuse this adapter, so none need their own accounting hook.
+ */
+export class UsageAccountingAdapter extends LLMAdapter {
+  #adapter;
+  #onUsage;
+
+  constructor(adapter, onUsage) {
+    super(adapter?.config || {});
+    this.#adapter = adapter;
+    this.#onUsage = onUsage;
+  }
+
+  #report(usage) {
+    if (!hasUsage(usage)) return;
+    try {
+      this.#onUsage(usage);
+    } catch (error) {
+      console.warn(`[llm-usage] accounting callback failed: ${error?.message || error}`);
+    }
+  }
+
+  async *stream(params) {
+    const total = normalizeTokenUsage();
+    try {
+      for await (const event of this.#adapter.stream(params)) {
+        if (event?.type === 'usage') addUsage(total, event);
+        yield event;
+      }
+    } finally {
+      this.#report(total);
+    }
+  }
+
+  async call(params) {
+    const result = await this.#adapter.call(params);
+    this.#report(normalizeTokenUsage(result?.usage || {}));
+    return result;
+  }
+
+  refreshProviders(providers) {
+    return this.#adapter.refreshProviders?.(providers);
+  }
+
+  getProviderForModel(modelId) {
+    return this.#adapter.getProviderForModel?.(modelId) || null;
+  }
+
+  listAvailableModels() {
+    return this.#adapter.listAvailableModels?.() || [];
+  }
+}
+
+export function withUsageAccounting(adapter, onUsage) {
+  return typeof onUsage === 'function'
+    ? new UsageAccountingAdapter(adapter, onUsage)
+    : adapter;
+}

package/yeaft/session.js

@@ -17,6 +17,8 @@ import { initYeaftDir, DEFAULT_YEAFT_DIR, isWritable } from './init.js';
 import { loadConfig, loadMCPConfig } from './config.js';
 import { createTrace } from './debug-trace.js';
 import { createLLMAdapter } from './llm/adapter.js';
+import { withUsageAccounting } from './llm/usage-accounting.js';
+import { recordAgentTokenUsage } from '../metrics.js';
 import { ConversationStore, setDefaultRecentTurnsLimit } from './conversation/persist.js';
 import { SkillManager, createSkillManager } from './skills.js';
 import { MCPManager } from './mcp.js';
@@ -262,7 +264,10 @@ export async function loadSession(options = {}) {
   });
 
   // ─── 4. Create LLM adapter ────────────────────────────
-  const adapter = await createLLMAdapter(config);
+  const adapter = withUsageAccounting(
+    await createLLMAdapter(config),
+    recordAgentTokenUsage
+  );
 
   // ─── 5. Create stores ──────────────────────────────────
   const conversationStore = new ConversationStore(yeaftDir);

package/yeaft/web-bridge.js

@@ -72,7 +72,7 @@ import { buildMcpFlattenedTools } from './tools/mcp-tools.js';
 import { getAgentRegistry, agentBelongsToScope } from './tools/agent.js';
 import { isPromptableAgentStatus } from './sub-agent/status.js';
 import { perfNowMs, recordAgentPerfTrace } from './perf-trace.js';
-import { recordAgentSessionCreated, recordAgentTokenUsage, recordAgentTurn } from '../metrics.js';
+import { recordAgentSessionCreated, recordAgentTurn } from '../metrics.js';
 
 const LEGACY_SKILL_COMMAND_PREFIX = 'skill:';
 const YEAFT_SKILL_COMMAND_PREFIX = 'yeaft-skills:';
@@ -2212,9 +2212,6 @@ function mergedStatusForProjectRuntime(runtime) {
 
 /** Send a Yeaft Session metadata event over the legacy-compatible envelope. */
 function sendSessionEvent(event, { sessionId, chatId, vpId, turnId, threadId, perfTraceId } = {}) {
-  if (event?.type === 'loop') {
-    recordAgentTokenUsage(event.usage || {});
-  }
   sendToServer({
     type: 'yeaft_output',
     conversationId: yeaftConversationId,