@yeaft/webchat-agent 0.1.1028 → 0.1.1029

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.1028",
+  "version": "0.1.1029",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",

package/yeaft/debug-trace.js

@@ -449,6 +449,25 @@ export class DebugTrace {
       try { return JSON.parse(s); }
       catch { return null; }
     };
+    const normalizeUsage = (usage, row) => {
+      const inputTokens = Number.isFinite(Number(usage?.inputTokens)) ? Number(usage.inputTokens) : (row.input_tokens || 0);
+      const outputTokens = Number.isFinite(Number(usage?.outputTokens)) ? Number(usage.outputTokens) : (row.output_tokens || 0);
+      const cacheReadTokens = Number.isFinite(Number(usage?.cacheReadTokens)) ? Number(usage.cacheReadTokens) : (row.cache_read_tokens || 0);
+      const cacheWriteTokens = Number.isFinite(Number(usage?.cacheWriteTokens)) ? Number(usage.cacheWriteTokens) : (row.cache_write_tokens || 0);
+      const totalInputTokens = Number.isFinite(Number(usage?.totalInputTokens))
+        ? Number(usage.totalInputTokens)
+        : inputTokens + cacheReadTokens + cacheWriteTokens;
+      return {
+        inputTokens,
+        outputTokens,
+        cacheReadTokens,
+        cacheWriteTokens,
+        totalInputTokens,
+        totalTokens: Number.isFinite(Number(usage?.totalTokens))
+          ? Number(usage.totalTokens)
+          : totalInputTokens + outputTokens,
+      };
+    };
     // Group rows by (turnId, threadId, sessionId, vpId) → frontend Turn
     // record. Each row is also surfaced as a Loop.
     const turnsById = new Map();
@@ -463,11 +482,7 @@ export class DebugTrace {
         messages: parsedMessages,
         response: r.response_text || '',
         toolCalls: parseJsonSafe(r.tool_calls_json) || [],
-        usage: parsedUsage || {
-          inputTokens: r.input_tokens || 0,
-          outputTokens: r.output_tokens || 0,
-          totalTokens: (r.input_tokens || 0) + (r.output_tokens || 0),
-        },
+        usage: normalizeUsage(parsedUsage, r),
         latencyMs: r.latency_ms || 0,
         ttfbMs: r.ttfb_ms || null,
         stopReason: r.stop_reason || null,
@@ -504,10 +519,7 @@ export class DebugTrace {
       // Aggregate per-loop latency / tokens so the Turn header shows the
       // same totals the live `turn_close` event would have stamped.
       t.totalMs += r.latency_ms || 0;
-      const usageTokens = parsedUsage && Number.isFinite(parsedUsage.totalTokens)
-        ? parsedUsage.totalTokens
-        : (r.input_tokens || 0) + (r.output_tokens || 0);
-      t.totalTokens += usageTokens;
+      t.totalTokens += loop.usage.totalTokens || 0;
       if (r.ended_at && (!t.closedAt || r.ended_at > t.closedAt)) t.closedAt = r.ended_at;
       return loop;
     });

package/yeaft/engine.js

@@ -1885,7 +1885,7 @@ export class Engine {
       const toolCalls = [];
       const thinkingBlocks = []; // task-327d: collected from adapter for round-trip
       let stopReason = 'end_turn';
-      const totalUsage = { inputTokens: 0, outputTokens: 0 };
+      const totalUsage = { inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheWriteTokens: 0, cacheInputDeltaTokens: 0 };
       // task-344: capture redacted raw request / raw response for debug panel.
       let rawRequest = null;
       let rawResponse = null;
@@ -2080,13 +2080,22 @@ export class Engine {
               toolCalls.push(event);
               yield event;
               break;
-            case 'usage':
-              totalUsage.inputTokens += event.inputTokens;
-              totalUsage.outputTokens += event.outputTokens;
-              cumulativeInputTokens += event.inputTokens || 0;
-              cumulativeOutputTokens += event.outputTokens || 0;
+            case 'usage': {
+              const inputTokens = event.inputTokens || 0;
+              const outputTokens = event.outputTokens || 0;
+              const cacheReadTokens = event.cacheReadTokens || 0;
+              const cacheWriteTokens = event.cacheWriteTokens || 0;
+              const cacheInputDeltaTokens = event.cacheTokensAreIncludedInInput ? 0 : cacheReadTokens + cacheWriteTokens;
+              totalUsage.inputTokens += inputTokens;
+              totalUsage.outputTokens += outputTokens;
+              totalUsage.cacheReadTokens += cacheReadTokens;
+              totalUsage.cacheWriteTokens += cacheWriteTokens;
+              totalUsage.cacheInputDeltaTokens += cacheInputDeltaTokens;
+              cumulativeInputTokens += inputTokens + cacheInputDeltaTokens;
+              cumulativeOutputTokens += outputTokens;
               yield event;
               break;
+            }
             case 'stop':
               stopReason = event.stopReason;
               yield event;
@@ -2107,6 +2116,8 @@ export class Engine {
           model: currentModel,
           inputTokens: totalUsage.inputTokens,
           outputTokens: totalUsage.outputTokens,
+          cacheReadTokens: totalUsage.cacheReadTokens,
+          cacheWriteTokens: totalUsage.cacheWriteTokens,
           stopReason: 'error',
           latencyMs,
           responseText,
@@ -2119,7 +2130,10 @@ export class Engine {
           usage: {
             inputTokens: totalUsage.inputTokens || 0,
             outputTokens: totalUsage.outputTokens || 0,
-            totalTokens: (totalUsage.inputTokens || 0) + (totalUsage.outputTokens || 0),
+            cacheReadTokens: totalUsage.cacheReadTokens || 0,
+            cacheWriteTokens: totalUsage.cacheWriteTokens || 0,
+            totalInputTokens: (totalUsage.inputTokens || 0) + (totalUsage.cacheInputDeltaTokens || 0),
+            totalTokens: (totalUsage.inputTokens || 0) + (totalUsage.cacheInputDeltaTokens || 0) + (totalUsage.outputTokens || 0),
           },
           ttfbMs,
           rawRequest,
@@ -2127,7 +2141,7 @@ export class Engine {
         });
 
         // Emit `loop` event for error path too (was `debug_turn`).
-        const errLoopInputTokens = totalUsage.inputTokens || 0;
+        const errLoopInputTokens = (totalUsage.inputTokens || 0) + (totalUsage.cacheInputDeltaTokens || 0);
         const errLoopOutputTokens = totalUsage.outputTokens || 0;
         yield {
           type: 'loop',
@@ -2140,8 +2154,11 @@ export class Engine {
           response: responseText || `Error: ${err.message}`,
           toolCalls: toolCalls.map(tc => ({ id: tc.id, name: tc.name, input: tc.input })),
           usage: {
-            inputTokens: errLoopInputTokens,
+            inputTokens: totalUsage.inputTokens || 0,
             outputTokens: errLoopOutputTokens,
+            cacheReadTokens: totalUsage.cacheReadTokens || 0,
+            cacheWriteTokens: totalUsage.cacheWriteTokens || 0,
+            totalInputTokens: errLoopInputTokens,
             totalTokens: errLoopInputTokens + errLoopOutputTokens,
           },
           latencyMs,
@@ -2250,11 +2267,16 @@ export class Engine {
 
       const latencyMs = Date.now() - startTime;
 
+      const turnInputTokens = (totalUsage.inputTokens || 0) + (totalUsage.cacheInputDeltaTokens || 0);
+      const turnOutputTokens = totalUsage.outputTokens || 0;
+
       // Record turn in debug trace
       this.#trace.endTurn(turnId, {
         model: currentModel,
         inputTokens: totalUsage.inputTokens,
         outputTokens: totalUsage.outputTokens,
+        cacheReadTokens: totalUsage.cacheReadTokens,
+        cacheWriteTokens: totalUsage.cacheWriteTokens,
         stopReason,
         latencyMs,
         responseText,
@@ -2268,7 +2290,10 @@ export class Engine {
         usage: {
           inputTokens: totalUsage.inputTokens || 0,
           outputTokens: totalUsage.outputTokens || 0,
-          totalTokens: (totalUsage.inputTokens || 0) + (totalUsage.outputTokens || 0),
+          cacheReadTokens: totalUsage.cacheReadTokens || 0,
+          cacheWriteTokens: totalUsage.cacheWriteTokens || 0,
+          totalInputTokens: turnInputTokens,
+          totalTokens: turnInputTokens + turnOutputTokens,
         },
         ttfbMs,
         rawRequest,
@@ -2284,7 +2309,7 @@ export class Engine {
       // task-331: preserve toolCalls / toolCallId / isError on each message
       // so the panel can render function_call requests and their paired
       // tool_result responses across loops.
-      const loopInputTokens = totalUsage.inputTokens || 0;
+      const loopInputTokens = (totalUsage.inputTokens || 0) + (totalUsage.cacheInputDeltaTokens || 0);
       const loopOutputTokens = totalUsage.outputTokens || 0;
       yield {
         type: 'loop',
@@ -2296,8 +2321,11 @@ export class Engine {
         response: responseText,
         toolCalls: toolCalls.map(tc => ({ id: tc.id, name: tc.name, input: tc.input })),
         usage: {
-          inputTokens: loopInputTokens,
+          inputTokens: totalUsage.inputTokens || 0,
           outputTokens: loopOutputTokens,
+          cacheReadTokens: totalUsage.cacheReadTokens || 0,
+          cacheWriteTokens: totalUsage.cacheWriteTokens || 0,
+          totalInputTokens: loopInputTokens,
           totalTokens: loopInputTokens + loopOutputTokens,
         },
         latencyMs,

package/yeaft/llm/adapter.js

@@ -42,7 +42,7 @@
  * @typedef {{ type: 'thinking_delta', text: string }} ThinkingDeltaEvent
  * @typedef {{ type: 'thinking_block_end', thinking: string, signature: string }} ThinkingBlockEndEvent
  * @typedef {{ type: 'tool_call', id: string, name: string, input: object }} ToolCallEvent
- * @typedef {{ type: 'usage', inputTokens: number, outputTokens: number, cacheReadTokens?: number, cacheWriteTokens?: number }} UsageEvent
+ * @typedef {{ type: 'usage', inputTokens: number, outputTokens: number, cacheReadTokens?: number, cacheWriteTokens?: number, cacheTokensAreIncludedInInput?: boolean }} UsageEvent
  * @typedef {{ type: 'stop', stopReason: 'end_turn' | 'tool_use' | 'max_tokens' }} StopEvent
  * @typedef {{ type: 'error', error: Error, retryable: boolean }} ErrorEvent
  *

package/yeaft/llm/anthropic.js

@@ -508,6 +508,8 @@ export class AnthropicAdapter extends LLMAdapter {
       usage: {
         inputTokens: result.usage?.input_tokens || 0,
         outputTokens: result.usage?.output_tokens || 0,
+        cacheReadTokens: result.usage?.cache_read_input_tokens || 0,
+        cacheWriteTokens: result.usage?.cache_creation_input_tokens || 0,
       },
     };
   }

package/yeaft/llm/openai-responses.js

@@ -435,6 +435,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
               outputTokens: usage.output_tokens || 0,
               cacheReadTokens: usage.input_tokens_details?.cached_tokens || 0,
               cacheWriteTokens: 0,
+              cacheTokensAreIncludedInInput: true,
             };
 
             yield {

package/yeaft/web-bridge.js

@@ -2233,6 +2233,9 @@ function handleEngineEvent(event, hctx) {
         type: 'context_usage',
         inputTokens: event.inputTokens,
         outputTokens: event.outputTokens,
+        cacheReadTokens: event.cacheReadTokens || 0,
+        cacheWriteTokens: event.cacheWriteTokens || 0,
+        totalInputTokens: (event.inputTokens || 0) + (event.cacheReadTokens || 0) + (event.cacheWriteTokens || 0),
       }, envelope);
       break;