@prismer/runtime 2.2.63 → 2.2.64

package/dist/cli.cjs

@@ -2101,16 +2101,20 @@ function thinkingText(message) {
 }
 function mapUsage(usage) {
   if (!usage) return void 0;
+  const measured = (value) => typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
   return {
-    inputTokens: usage.input,
-    cachedInputTokens: usage.cacheRead,
-    outputTokens: usage.output,
-    totalCostUsd: usage.cost.total,
-    contextWindowUsedTokens: usage.totalTokens
+    inputTokens: measured(usage.input),
+    cachedInputTokens: measured(usage.cacheRead),
+    cacheWriteTokens: measured(usage.cacheWrite),
+    outputTokens: measured(usage.output),
+    totalCostUsd: measured(usage.cost?.total),
+    contextWindowUsedTokens: measured(usage.totalTokens)
   };
 }
 function extractAssistantUsage(message) {
-  return message.role === "assistant" ? mapUsage(message.usage) : void 0;
+  if (message.role !== "assistant") return void 0;
+  if ((message.stopReason === "error" || message.stopReason === "aborted") && ![message.usage?.input, message.usage?.output, message.usage?.cacheRead, message.usage?.cacheWrite].some((value) => typeof value === "number" && value > 0)) return void 0;
+  return mapUsage(message.usage);
 }
 function modelDefinition(model) {
   return {
@@ -2686,7 +2690,15 @@ var init_agent = __esm({
               };
             })
           },
-          streamFn: options.models.streamSimple.bind(options.models),
+          streamFn: (...args) => {
+            const run = this.activeRun;
+            if (run) {
+              run.callCount = (run.callCount ?? 0) + 1;
+              run.modelCall = { sequence: run.callCount, startedAt: Date.now(), clock: performance.now(), firstTextMs: null };
+              this.emitObservation({ type: "model_request_started", provider: this.provider, startedAt: run.modelCall.startedAt, turnId: run.turnId });
+            }
+            return options.models.streamSimple(...args);
+          },
           sessionId: this.id,
           convertToLlm: (messages) => messages.filter(isProviderMessage),
           toolExecution: "parallel",
@@ -2698,7 +2710,7 @@ var init_agent = __esm({
           // 「模型请求了一个不存在的工具」不会产生 tool_started——事件面只记真实执行。
           beforeToolCall: async (context, signal) => {
             if (sink) {
-              this.toolStartedAt.set(context.toolCall.id, Date.now());
+              this.toolStartedAt.set(context.toolCall.id, performance.now());
               emitToolEvent(sink, {
                 kind: "tool_started",
                 name: context.toolCall.name,
@@ -2723,7 +2735,7 @@ var init_agent = __esm({
                 name,
                 isError: true,
                 resultSummary: reason,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
               return { block: true, reason };
             }
@@ -2738,7 +2750,7 @@ var init_agent = __esm({
                 name: context.toolCall.name,
                 resultSummary: summarizeToolResult(context.result, secrets),
                 isError: context.isError,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
             }
             return void 0;
@@ -2854,6 +2866,14 @@ var init_agent = __esm({
       emit(event) {
         for (const subscriber of this.subscribers) subscriber(event);
       }
+      emitObservation(event) {
+        for (const subscriber of this.subscribers) {
+          try {
+            void Promise.resolve(subscriber(event)).catch(() => void 0);
+          } catch {
+          }
+        }
+      }
       recordTimeline(item) {
         this.activeRun?.timeline.push(item);
         this.emit({
@@ -2873,6 +2893,9 @@ var init_agent = __esm({
           case "message_update": {
             const sub = event.assistantMessageEvent;
             if (sub.type === "text_delta" || sub.type === "thinking_delta") {
+              if (sub.type === "text_delta" && sub.delta && run?.modelCall && run.modelCall.firstTextMs === null) {
+                run.modelCall.firstTextMs = performance.now() - run.modelCall.clock;
+              }
               this.emit({
                 type: "text_delta",
                 provider: this.provider,
@@ -2882,7 +2905,7 @@ var init_agent = __esm({
               });
               break;
             }
-            this.handleMessageEvent(event.message);
+            this.handleMessageEvent(event.message, false);
             break;
           }
           case "message_end":
@@ -2920,7 +2943,7 @@ var init_agent = __esm({
             break;
         }
       }
-      handleMessageEvent(message) {
+      handleMessageEvent(message, settled = true) {
         const run = this.activeRun;
         if (!run) return;
         if (message.role !== "assistant") return;
@@ -2933,8 +2956,34 @@ var init_agent = __esm({
           run.finalText = text2;
           this.recordTimeline({ type: "assistant_message", text: text2 });
         }
-        const usage = extractAssistantUsage(message);
-        if (usage) run.usage = usage;
+        const usage = settled ? extractAssistantUsage(message) : void 0;
+        if (settled && !usage) run.usage = {};
+        if (usage) {
+          const previous = run.usage;
+          run.usage = { ...usage };
+          for (const field2 of ["inputTokens", "outputTokens", "cachedInputTokens", "cacheWriteTokens", "totalCostUsd"]) {
+            if (previous?.[field2] !== void 0 && usage[field2] !== void 0) {
+              run.usage[field2] = previous[field2] + usage[field2];
+            } else if (previous) {
+              delete run.usage[field2];
+            }
+          }
+        }
+        if (settled && run.modelCall) {
+          const call = run.modelCall;
+          delete run.modelCall;
+          this.emitObservation({ type: "model_call_observed", provider: this.provider, turnId: run.turnId, call: {
+            sequence: call.sequence,
+            source: "pi-engine",
+            provider: message.provider,
+            model: message.responseModel ?? message.model,
+            startedAt: call.startedAt,
+            durationMs: performance.now() - call.clock,
+            firstTextMs: call.firstTextMs,
+            status: message.stopReason === "error" ? "failed" : message.stopReason === "aborted" ? "canceled" : "completed",
+            usage: usage ?? null
+          } });
+        }
         run.servedModel = message.responseModel ?? message.model;
         run.servedProvider = message.provider;
       }
@@ -16003,7 +16052,7 @@ var require_package = __commonJS({
   "package.json"(exports2, module2) {
     module2.exports = {
       name: "@prismer/runtime",
-      version: "2.2.63",
+      version: "2.2.64",
       description: "Prismer Cloud daemon runtime \u2014 TS-only adapter host for hosted IM agents",
       type: "module",
       main: "dist/index.js",
@@ -78652,6 +78701,21 @@ function classifyProviderError(message) {
   return /ENOTFOUND|ECONNREFUSED|ETIMEDOUT|EAI_AGAIN|fetch failed|socket hang up|network/i.test(message) ? "provider_unreachable" : "internal";
 }
 async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
+  const started = performance.now();
+  const observation = {
+    version: 1,
+    source: "runtime",
+    sessionInitMs: null,
+    totalMs: 0,
+    modelCalls: [],
+    droppedModelCalls: 0
+  };
+  const result = await runTurnExecution(envelope2, egress, deps, onToolEvent, observation);
+  observation.totalMs = performance.now() - started;
+  result.observation = observation;
+  return result;
+}
+async function runTurnExecution(envelope2, egress, deps, onToolEvent, observation) {
   const createSession = deps.createSession ?? defaultCreateSession;
   const version = runtimeVersion();
   const env = egressSessionEnv(envelope2, egress);
@@ -78659,7 +78723,9 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
   let session = null;
   let t8 = null;
   let timedOut = false;
-  const t7 = Date.now();
+  let t7 = null;
+  let sawRequest = false;
+  const initStarted = performance.now();
   try {
     let tenantComponents;
     if (envelope2.tenantComponents?.length) {
@@ -78683,11 +78749,20 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
       ...onToolEvent ? { onToolEvent } : {}
     });
     session.subscribe((event) => {
+      if (event.type === "model_request_started" && !sawRequest && typeof event.startedAt === "number") {
+        t7 = event.startedAt;
+        sawRequest = true;
+      }
+      if (event.type === "model_call_observed" && event.call) {
+        if (observation.modelCalls.length < 128) observation.modelCalls.push(event.call);
+        else observation.droppedModelCalls += 1;
+      }
       if (timedOut || t8 !== null) return;
       if (event.type === "text_delta" && event.deltaKind === "text" && event.delta) {
         t8 = Date.now();
       }
     });
+    observation.sessionInitMs = performance.now() - initStarted;
     const timer = setTimeout(() => {
       timedOut = true;
       void session?.interrupt().catch(() => void 0);
@@ -79383,7 +79458,7 @@ async function readResponseError(res) {
 
 // src/cli/index.ts
 init_ui();
-var VERSION3 = "2.2.63";
+var VERSION3 = "2.2.64";
 function buildProgram() {
   const program = new import_commander32.Command("prismer").description("Prismer Runtime host and local control CLI (TS-only).").version(VERSION3).addHelpText(
     "after",

package/dist/cli.js

@@ -2113,16 +2113,20 @@ function thinkingText(message) {
 }
 function mapUsage(usage) {
   if (!usage) return void 0;
+  const measured = (value) => typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
   return {
-    inputTokens: usage.input,
-    cachedInputTokens: usage.cacheRead,
-    outputTokens: usage.output,
-    totalCostUsd: usage.cost.total,
-    contextWindowUsedTokens: usage.totalTokens
+    inputTokens: measured(usage.input),
+    cachedInputTokens: measured(usage.cacheRead),
+    cacheWriteTokens: measured(usage.cacheWrite),
+    outputTokens: measured(usage.output),
+    totalCostUsd: measured(usage.cost?.total),
+    contextWindowUsedTokens: measured(usage.totalTokens)
   };
 }
 function extractAssistantUsage(message) {
-  return message.role === "assistant" ? mapUsage(message.usage) : void 0;
+  if (message.role !== "assistant") return void 0;
+  if ((message.stopReason === "error" || message.stopReason === "aborted") && ![message.usage?.input, message.usage?.output, message.usage?.cacheRead, message.usage?.cacheWrite].some((value) => typeof value === "number" && value > 0)) return void 0;
+  return mapUsage(message.usage);
 }
 function modelDefinition(model) {
   return {
@@ -2683,7 +2687,15 @@ var init_agent = __esm({
               };
             })
           },
-          streamFn: options.models.streamSimple.bind(options.models),
+          streamFn: (...args) => {
+            const run = this.activeRun;
+            if (run) {
+              run.callCount = (run.callCount ?? 0) + 1;
+              run.modelCall = { sequence: run.callCount, startedAt: Date.now(), clock: performance.now(), firstTextMs: null };
+              this.emitObservation({ type: "model_request_started", provider: this.provider, startedAt: run.modelCall.startedAt, turnId: run.turnId });
+            }
+            return options.models.streamSimple(...args);
+          },
           sessionId: this.id,
           convertToLlm: (messages) => messages.filter(isProviderMessage),
           toolExecution: "parallel",
@@ -2695,7 +2707,7 @@ var init_agent = __esm({
           // 「模型请求了一个不存在的工具」不会产生 tool_started——事件面只记真实执行。
           beforeToolCall: async (context, signal) => {
             if (sink) {
-              this.toolStartedAt.set(context.toolCall.id, Date.now());
+              this.toolStartedAt.set(context.toolCall.id, performance.now());
               emitToolEvent(sink, {
                 kind: "tool_started",
                 name: context.toolCall.name,
@@ -2720,7 +2732,7 @@ var init_agent = __esm({
                 name,
                 isError: true,
                 resultSummary: reason,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
               return { block: true, reason };
             }
@@ -2735,7 +2747,7 @@ var init_agent = __esm({
                 name: context.toolCall.name,
                 resultSummary: summarizeToolResult(context.result, secrets),
                 isError: context.isError,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
             }
             return void 0;
@@ -2851,6 +2863,14 @@ var init_agent = __esm({
       emit(event) {
         for (const subscriber of this.subscribers) subscriber(event);
       }
+      emitObservation(event) {
+        for (const subscriber of this.subscribers) {
+          try {
+            void Promise.resolve(subscriber(event)).catch(() => void 0);
+          } catch {
+          }
+        }
+      }
       recordTimeline(item) {
         this.activeRun?.timeline.push(item);
         this.emit({
@@ -2870,6 +2890,9 @@ var init_agent = __esm({
           case "message_update": {
             const sub = event.assistantMessageEvent;
             if (sub.type === "text_delta" || sub.type === "thinking_delta") {
+              if (sub.type === "text_delta" && sub.delta && run?.modelCall && run.modelCall.firstTextMs === null) {
+                run.modelCall.firstTextMs = performance.now() - run.modelCall.clock;
+              }
               this.emit({
                 type: "text_delta",
                 provider: this.provider,
@@ -2879,7 +2902,7 @@ var init_agent = __esm({
               });
               break;
             }
-            this.handleMessageEvent(event.message);
+            this.handleMessageEvent(event.message, false);
             break;
           }
           case "message_end":
@@ -2917,7 +2940,7 @@ var init_agent = __esm({
             break;
         }
       }
-      handleMessageEvent(message) {
+      handleMessageEvent(message, settled = true) {
         const run = this.activeRun;
         if (!run) return;
         if (message.role !== "assistant") return;
@@ -2930,8 +2953,34 @@ var init_agent = __esm({
           run.finalText = text2;
           this.recordTimeline({ type: "assistant_message", text: text2 });
         }
-        const usage = extractAssistantUsage(message);
-        if (usage) run.usage = usage;
+        const usage = settled ? extractAssistantUsage(message) : void 0;
+        if (settled && !usage) run.usage = {};
+        if (usage) {
+          const previous = run.usage;
+          run.usage = { ...usage };
+          for (const field2 of ["inputTokens", "outputTokens", "cachedInputTokens", "cacheWriteTokens", "totalCostUsd"]) {
+            if (previous?.[field2] !== void 0 && usage[field2] !== void 0) {
+              run.usage[field2] = previous[field2] + usage[field2];
+            } else if (previous) {
+              delete run.usage[field2];
+            }
+          }
+        }
+        if (settled && run.modelCall) {
+          const call = run.modelCall;
+          delete run.modelCall;
+          this.emitObservation({ type: "model_call_observed", provider: this.provider, turnId: run.turnId, call: {
+            sequence: call.sequence,
+            source: "pi-engine",
+            provider: message.provider,
+            model: message.responseModel ?? message.model,
+            startedAt: call.startedAt,
+            durationMs: performance.now() - call.clock,
+            firstTextMs: call.firstTextMs,
+            status: message.stopReason === "error" ? "failed" : message.stopReason === "aborted" ? "canceled" : "completed",
+            usage: usage ?? null
+          } });
+        }
         run.servedModel = message.responseModel ?? message.model;
         run.servedProvider = message.provider;
       }
@@ -16034,7 +16083,7 @@ var require_package = __commonJS({
   "package.json"(exports, module) {
     module.exports = {
       name: "@prismer/runtime",
-      version: "2.2.63",
+      version: "2.2.64",
       description: "Prismer Cloud daemon runtime \u2014 TS-only adapter host for hosted IM agents",
       type: "module",
       main: "dist/index.js",
@@ -78696,6 +78745,21 @@ function classifyProviderError(message) {
   return /ENOTFOUND|ECONNREFUSED|ETIMEDOUT|EAI_AGAIN|fetch failed|socket hang up|network/i.test(message) ? "provider_unreachable" : "internal";
 }
 async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
+  const started = performance.now();
+  const observation = {
+    version: 1,
+    source: "runtime",
+    sessionInitMs: null,
+    totalMs: 0,
+    modelCalls: [],
+    droppedModelCalls: 0
+  };
+  const result = await runTurnExecution(envelope2, egress, deps, onToolEvent, observation);
+  observation.totalMs = performance.now() - started;
+  result.observation = observation;
+  return result;
+}
+async function runTurnExecution(envelope2, egress, deps, onToolEvent, observation) {
   const createSession = deps.createSession ?? defaultCreateSession;
   const version = runtimeVersion();
   const env = egressSessionEnv(envelope2, egress);
@@ -78703,7 +78767,9 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
   let session = null;
   let t8 = null;
   let timedOut = false;
-  const t7 = Date.now();
+  let t7 = null;
+  let sawRequest = false;
+  const initStarted = performance.now();
   try {
     let tenantComponents;
     if (envelope2.tenantComponents?.length) {
@@ -78727,11 +78793,20 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
       ...onToolEvent ? { onToolEvent } : {}
     });
     session.subscribe((event) => {
+      if (event.type === "model_request_started" && !sawRequest && typeof event.startedAt === "number") {
+        t7 = event.startedAt;
+        sawRequest = true;
+      }
+      if (event.type === "model_call_observed" && event.call) {
+        if (observation.modelCalls.length < 128) observation.modelCalls.push(event.call);
+        else observation.droppedModelCalls += 1;
+      }
       if (timedOut || t8 !== null) return;
       if (event.type === "text_delta" && event.deltaKind === "text" && event.delta) {
         t8 = Date.now();
       }
     });
+    observation.sessionInitMs = performance.now() - initStarted;
     const timer = setTimeout(() => {
       timedOut = true;
       void session?.interrupt().catch(() => void 0);
@@ -79427,7 +79502,7 @@ async function readResponseError(res) {
 
 // src/cli/index.ts
 init_ui();
-var VERSION3 = "2.2.63";
+var VERSION3 = "2.2.64";
 function buildProgram() {
   const program = new Command32("prismer").description("Prismer Runtime host and local control CLI (TS-only).").version(VERSION3).addHelpText(
     "after",

package/dist/index.cjs

@@ -15707,16 +15707,20 @@ function thinkingText(message) {
 }
 function mapUsage(usage) {
   if (!usage) return void 0;
+  const measured = (value) => typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
   return {
-    inputTokens: usage.input,
-    cachedInputTokens: usage.cacheRead,
-    outputTokens: usage.output,
-    totalCostUsd: usage.cost.total,
-    contextWindowUsedTokens: usage.totalTokens
+    inputTokens: measured(usage.input),
+    cachedInputTokens: measured(usage.cacheRead),
+    cacheWriteTokens: measured(usage.cacheWrite),
+    outputTokens: measured(usage.output),
+    totalCostUsd: measured(usage.cost?.total),
+    contextWindowUsedTokens: measured(usage.totalTokens)
   };
 }
 function extractAssistantUsage(message) {
-  return message.role === "assistant" ? mapUsage(message.usage) : void 0;
+  if (message.role !== "assistant") return void 0;
+  if ((message.stopReason === "error" || message.stopReason === "aborted") && ![message.usage?.input, message.usage?.output, message.usage?.cacheRead, message.usage?.cacheWrite].some((value) => typeof value === "number" && value > 0)) return void 0;
+  return mapUsage(message.usage);
 }
 function modelDefinition(model) {
   return {
@@ -16292,7 +16296,15 @@ var init_agent = __esm({
               };
             })
           },
-          streamFn: options.models.streamSimple.bind(options.models),
+          streamFn: (...args) => {
+            const run = this.activeRun;
+            if (run) {
+              run.callCount = (run.callCount ?? 0) + 1;
+              run.modelCall = { sequence: run.callCount, startedAt: Date.now(), clock: performance.now(), firstTextMs: null };
+              this.emitObservation({ type: "model_request_started", provider: this.provider, startedAt: run.modelCall.startedAt, turnId: run.turnId });
+            }
+            return options.models.streamSimple(...args);
+          },
           sessionId: this.id,
           convertToLlm: (messages) => messages.filter(isProviderMessage),
           toolExecution: "parallel",
@@ -16304,7 +16316,7 @@ var init_agent = __esm({
           // 「模型请求了一个不存在的工具」不会产生 tool_started——事件面只记真实执行。
           beforeToolCall: async (context, signal) => {
             if (sink) {
-              this.toolStartedAt.set(context.toolCall.id, Date.now());
+              this.toolStartedAt.set(context.toolCall.id, performance.now());
               emitToolEvent(sink, {
                 kind: "tool_started",
                 name: context.toolCall.name,
@@ -16329,7 +16341,7 @@ var init_agent = __esm({
                 name,
                 isError: true,
                 resultSummary: reason,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
               return { block: true, reason };
             }
@@ -16344,7 +16356,7 @@ var init_agent = __esm({
                 name: context.toolCall.name,
                 resultSummary: summarizeToolResult(context.result, secrets),
                 isError: context.isError,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
             }
             return void 0;
@@ -16460,6 +16472,14 @@ var init_agent = __esm({
       emit(event) {
         for (const subscriber of this.subscribers) subscriber(event);
       }
+      emitObservation(event) {
+        for (const subscriber of this.subscribers) {
+          try {
+            void Promise.resolve(subscriber(event)).catch(() => void 0);
+          } catch {
+          }
+        }
+      }
       recordTimeline(item) {
         this.activeRun?.timeline.push(item);
         this.emit({
@@ -16479,6 +16499,9 @@ var init_agent = __esm({
           case "message_update": {
             const sub = event.assistantMessageEvent;
             if (sub.type === "text_delta" || sub.type === "thinking_delta") {
+              if (sub.type === "text_delta" && sub.delta && run?.modelCall && run.modelCall.firstTextMs === null) {
+                run.modelCall.firstTextMs = performance.now() - run.modelCall.clock;
+              }
               this.emit({
                 type: "text_delta",
                 provider: this.provider,
@@ -16488,7 +16511,7 @@ var init_agent = __esm({
               });
               break;
             }
-            this.handleMessageEvent(event.message);
+            this.handleMessageEvent(event.message, false);
             break;
           }
           case "message_end":
@@ -16526,7 +16549,7 @@ var init_agent = __esm({
             break;
         }
       }
-      handleMessageEvent(message) {
+      handleMessageEvent(message, settled = true) {
         const run = this.activeRun;
         if (!run) return;
         if (message.role !== "assistant") return;
@@ -16539,8 +16562,34 @@ var init_agent = __esm({
           run.finalText = text2;
           this.recordTimeline({ type: "assistant_message", text: text2 });
         }
-        const usage = extractAssistantUsage(message);
-        if (usage) run.usage = usage;
+        const usage = settled ? extractAssistantUsage(message) : void 0;
+        if (settled && !usage) run.usage = {};
+        if (usage) {
+          const previous = run.usage;
+          run.usage = { ...usage };
+          for (const field2 of ["inputTokens", "outputTokens", "cachedInputTokens", "cacheWriteTokens", "totalCostUsd"]) {
+            if (previous?.[field2] !== void 0 && usage[field2] !== void 0) {
+              run.usage[field2] = previous[field2] + usage[field2];
+            } else if (previous) {
+              delete run.usage[field2];
+            }
+          }
+        }
+        if (settled && run.modelCall) {
+          const call = run.modelCall;
+          delete run.modelCall;
+          this.emitObservation({ type: "model_call_observed", provider: this.provider, turnId: run.turnId, call: {
+            sequence: call.sequence,
+            source: "pi-engine",
+            provider: message.provider,
+            model: message.responseModel ?? message.model,
+            startedAt: call.startedAt,
+            durationMs: performance.now() - call.clock,
+            firstTextMs: call.firstTextMs,
+            status: message.stopReason === "error" ? "failed" : message.stopReason === "aborted" ? "canceled" : "completed",
+            usage: usage ?? null
+          } });
+        }
         run.servedModel = message.responseModel ?? message.model;
         run.servedProvider = message.provider;
       }
@@ -17065,7 +17114,7 @@ var require_package = __commonJS({
   "package.json"(exports2, module2) {
     module2.exports = {
       name: "@prismer/runtime",
-      version: "2.2.63",
+      version: "2.2.64",
       description: "Prismer Cloud daemon runtime \u2014 TS-only adapter host for hosted IM agents",
       type: "module",
       main: "dist/index.js",
@@ -79256,6 +79305,21 @@ function classifyProviderError(message) {
   return /ENOTFOUND|ECONNREFUSED|ETIMEDOUT|EAI_AGAIN|fetch failed|socket hang up|network/i.test(message) ? "provider_unreachable" : "internal";
 }
 async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
+  const started = performance.now();
+  const observation = {
+    version: 1,
+    source: "runtime",
+    sessionInitMs: null,
+    totalMs: 0,
+    modelCalls: [],
+    droppedModelCalls: 0
+  };
+  const result = await runTurnExecution(envelope2, egress, deps, onToolEvent, observation);
+  observation.totalMs = performance.now() - started;
+  result.observation = observation;
+  return result;
+}
+async function runTurnExecution(envelope2, egress, deps, onToolEvent, observation) {
   const createSession = deps.createSession ?? defaultCreateSession;
   const version = runtimeVersion();
   const env = egressSessionEnv(envelope2, egress);
@@ -79263,7 +79327,9 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
   let session = null;
   let t8 = null;
   let timedOut = false;
-  const t7 = Date.now();
+  let t7 = null;
+  let sawRequest = false;
+  const initStarted = performance.now();
   try {
     let tenantComponents;
     if (envelope2.tenantComponents?.length) {
@@ -79287,11 +79353,20 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
       ...onToolEvent ? { onToolEvent } : {}
     });
     session.subscribe((event) => {
+      if (event.type === "model_request_started" && !sawRequest && typeof event.startedAt === "number") {
+        t7 = event.startedAt;
+        sawRequest = true;
+      }
+      if (event.type === "model_call_observed" && event.call) {
+        if (observation.modelCalls.length < 128) observation.modelCalls.push(event.call);
+        else observation.droppedModelCalls += 1;
+      }
       if (timedOut || t8 !== null) return;
       if (event.type === "text_delta" && event.deltaKind === "text" && event.delta) {
         t8 = Date.now();
       }
     });
+    observation.sessionInitMs = performance.now() - initStarted;
     const timer = setTimeout(() => {
       timedOut = true;
       void session?.interrupt().catch(() => void 0);
@@ -79987,7 +80062,7 @@ async function readResponseError(res) {
 
 // src/cli/index.ts
 init_ui();
-var VERSION3 = "2.2.63";
+var VERSION3 = "2.2.64";
 function buildProgram() {
   const program = new import_commander32.Command("prismer").description("Prismer Runtime host and local control CLI (TS-only).").version(VERSION3).addHelpText(
     "after",

package/dist/index.d.cts

@@ -1875,11 +1875,25 @@ interface AgentRunOptions {
 interface AgentUsage {
     inputTokens?: number;
     cachedInputTokens?: number;
+    /** PI cache creation tokens, separate from uncached input and cache reads. */
+    cacheWriteTokens?: number;
     outputTokens?: number;
     totalCostUsd?: number;
     contextWindowMaxTokens?: number;
     contextWindowUsedTokens?: number;
 }
+/** Measurements at the PI engine boundary, not a provider billing receipt. */
+interface ModelCallObservation {
+    sequence: number;
+    source: 'pi-engine';
+    provider: string;
+    model: string;
+    startedAt: number;
+    durationMs: number;
+    firstTextMs: number | null;
+    status: 'completed' | 'failed' | 'canceled';
+    usage: AgentUsage | null;
+}
 declare const TOOL_CALL_ICON_NAMES: readonly ["wrench", "square_terminal", "eye", "pencil", "search", "bot", "sparkles", "brain", "mic_vocal"];
 type ToolCallIconName = (typeof TOOL_CALL_ICON_NAMES)[number];
 type ToolCallDetail = {
@@ -2032,6 +2046,16 @@ type AgentTimelineItem = {
     message: string;
 } | CompactionTimelineItem;
 type AgentStreamEvent = {
+    type: 'model_request_started';
+    provider: AgentProvider;
+    startedAt: number;
+    turnId?: string;
+} | {
+    type: 'model_call_observed';
+    provider: AgentProvider;
+    call: ModelCallObservation;
+    turnId?: string;
+} | {
     type: "thread_started";
     sessionId: string;
     provider: AgentProvider;
@@ -7260,6 +7284,7 @@ declare class PiAgentCoreSession implements AgentSession {
     close(): Promise<void>;
     listCommands(): Promise<AgentSlashCommand[]>;
     private emit;
+    private emitObservation;
     private recordTimeline;
     private handlePiEvent;
     private handleMessageEvent;

package/dist/index.d.ts

@@ -1875,11 +1875,25 @@ interface AgentRunOptions {
 interface AgentUsage {
     inputTokens?: number;
     cachedInputTokens?: number;
+    /** PI cache creation tokens, separate from uncached input and cache reads. */
+    cacheWriteTokens?: number;
     outputTokens?: number;
     totalCostUsd?: number;
     contextWindowMaxTokens?: number;
     contextWindowUsedTokens?: number;
 }
+/** Measurements at the PI engine boundary, not a provider billing receipt. */
+interface ModelCallObservation {
+    sequence: number;
+    source: 'pi-engine';
+    provider: string;
+    model: string;
+    startedAt: number;
+    durationMs: number;
+    firstTextMs: number | null;
+    status: 'completed' | 'failed' | 'canceled';
+    usage: AgentUsage | null;
+}
 declare const TOOL_CALL_ICON_NAMES: readonly ["wrench", "square_terminal", "eye", "pencil", "search", "bot", "sparkles", "brain", "mic_vocal"];
 type ToolCallIconName = (typeof TOOL_CALL_ICON_NAMES)[number];
 type ToolCallDetail = {
@@ -2032,6 +2046,16 @@ type AgentTimelineItem = {
     message: string;
 } | CompactionTimelineItem;
 type AgentStreamEvent = {
+    type: 'model_request_started';
+    provider: AgentProvider;
+    startedAt: number;
+    turnId?: string;
+} | {
+    type: 'model_call_observed';
+    provider: AgentProvider;
+    call: ModelCallObservation;
+    turnId?: string;
+} | {
     type: "thread_started";
     sessionId: string;
     provider: AgentProvider;
@@ -7260,6 +7284,7 @@ declare class PiAgentCoreSession implements AgentSession {
     close(): Promise<void>;
     listCommands(): Promise<AgentSlashCommand[]>;
     private emit;
+    private emitObservation;
     private recordTimeline;
     private handlePiEvent;
     private handleMessageEvent;

package/dist/index.js

@@ -15752,16 +15752,20 @@ function thinkingText(message) {
 }
 function mapUsage(usage) {
   if (!usage) return void 0;
+  const measured = (value) => typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
   return {
-    inputTokens: usage.input,
-    cachedInputTokens: usage.cacheRead,
-    outputTokens: usage.output,
-    totalCostUsd: usage.cost.total,
-    contextWindowUsedTokens: usage.totalTokens
+    inputTokens: measured(usage.input),
+    cachedInputTokens: measured(usage.cacheRead),
+    cacheWriteTokens: measured(usage.cacheWrite),
+    outputTokens: measured(usage.output),
+    totalCostUsd: measured(usage.cost?.total),
+    contextWindowUsedTokens: measured(usage.totalTokens)
   };
 }
 function extractAssistantUsage(message) {
-  return message.role === "assistant" ? mapUsage(message.usage) : void 0;
+  if (message.role !== "assistant") return void 0;
+  if ((message.stopReason === "error" || message.stopReason === "aborted") && ![message.usage?.input, message.usage?.output, message.usage?.cacheRead, message.usage?.cacheWrite].some((value) => typeof value === "number" && value > 0)) return void 0;
+  return mapUsage(message.usage);
 }
 function modelDefinition(model) {
   return {
@@ -16322,7 +16326,15 @@ var init_agent = __esm({
               };
             })
           },
-          streamFn: options.models.streamSimple.bind(options.models),
+          streamFn: (...args) => {
+            const run = this.activeRun;
+            if (run) {
+              run.callCount = (run.callCount ?? 0) + 1;
+              run.modelCall = { sequence: run.callCount, startedAt: Date.now(), clock: performance.now(), firstTextMs: null };
+              this.emitObservation({ type: "model_request_started", provider: this.provider, startedAt: run.modelCall.startedAt, turnId: run.turnId });
+            }
+            return options.models.streamSimple(...args);
+          },
           sessionId: this.id,
           convertToLlm: (messages) => messages.filter(isProviderMessage),
           toolExecution: "parallel",
@@ -16334,7 +16346,7 @@ var init_agent = __esm({
           // 「模型请求了一个不存在的工具」不会产生 tool_started——事件面只记真实执行。
           beforeToolCall: async (context, signal) => {
             if (sink) {
-              this.toolStartedAt.set(context.toolCall.id, Date.now());
+              this.toolStartedAt.set(context.toolCall.id, performance.now());
               emitToolEvent(sink, {
                 kind: "tool_started",
                 name: context.toolCall.name,
@@ -16359,7 +16371,7 @@ var init_agent = __esm({
                 name,
                 isError: true,
                 resultSummary: reason,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
               return { block: true, reason };
             }
@@ -16374,7 +16386,7 @@ var init_agent = __esm({
                 name: context.toolCall.name,
                 resultSummary: summarizeToolResult(context.result, secrets),
                 isError: context.isError,
-                ...startedAt !== void 0 ? { durationMs: Date.now() - startedAt } : {}
+                ...startedAt !== void 0 ? { durationMs: performance.now() - startedAt } : {}
               });
             }
             return void 0;
@@ -16490,6 +16502,14 @@ var init_agent = __esm({
       emit(event) {
         for (const subscriber of this.subscribers) subscriber(event);
       }
+      emitObservation(event) {
+        for (const subscriber of this.subscribers) {
+          try {
+            void Promise.resolve(subscriber(event)).catch(() => void 0);
+          } catch {
+          }
+        }
+      }
       recordTimeline(item) {
         this.activeRun?.timeline.push(item);
         this.emit({
@@ -16509,6 +16529,9 @@ var init_agent = __esm({
           case "message_update": {
             const sub = event.assistantMessageEvent;
             if (sub.type === "text_delta" || sub.type === "thinking_delta") {
+              if (sub.type === "text_delta" && sub.delta && run?.modelCall && run.modelCall.firstTextMs === null) {
+                run.modelCall.firstTextMs = performance.now() - run.modelCall.clock;
+              }
               this.emit({
                 type: "text_delta",
                 provider: this.provider,
@@ -16518,7 +16541,7 @@ var init_agent = __esm({
               });
               break;
             }
-            this.handleMessageEvent(event.message);
+            this.handleMessageEvent(event.message, false);
             break;
           }
           case "message_end":
@@ -16556,7 +16579,7 @@ var init_agent = __esm({
             break;
         }
       }
-      handleMessageEvent(message) {
+      handleMessageEvent(message, settled = true) {
         const run = this.activeRun;
         if (!run) return;
         if (message.role !== "assistant") return;
@@ -16569,8 +16592,34 @@ var init_agent = __esm({
           run.finalText = text2;
           this.recordTimeline({ type: "assistant_message", text: text2 });
         }
-        const usage = extractAssistantUsage(message);
-        if (usage) run.usage = usage;
+        const usage = settled ? extractAssistantUsage(message) : void 0;
+        if (settled && !usage) run.usage = {};
+        if (usage) {
+          const previous = run.usage;
+          run.usage = { ...usage };
+          for (const field2 of ["inputTokens", "outputTokens", "cachedInputTokens", "cacheWriteTokens", "totalCostUsd"]) {
+            if (previous?.[field2] !== void 0 && usage[field2] !== void 0) {
+              run.usage[field2] = previous[field2] + usage[field2];
+            } else if (previous) {
+              delete run.usage[field2];
+            }
+          }
+        }
+        if (settled && run.modelCall) {
+          const call = run.modelCall;
+          delete run.modelCall;
+          this.emitObservation({ type: "model_call_observed", provider: this.provider, turnId: run.turnId, call: {
+            sequence: call.sequence,
+            source: "pi-engine",
+            provider: message.provider,
+            model: message.responseModel ?? message.model,
+            startedAt: call.startedAt,
+            durationMs: performance.now() - call.clock,
+            firstTextMs: call.firstTextMs,
+            status: message.stopReason === "error" ? "failed" : message.stopReason === "aborted" ? "canceled" : "completed",
+            usage: usage ?? null
+          } });
+        }
         run.servedModel = message.responseModel ?? message.model;
         run.servedProvider = message.provider;
       }
@@ -17094,7 +17143,7 @@ var require_package = __commonJS({
   "package.json"(exports, module) {
     module.exports = {
       name: "@prismer/runtime",
-      version: "2.2.63",
+      version: "2.2.64",
       description: "Prismer Cloud daemon runtime \u2014 TS-only adapter host for hosted IM agents",
       type: "module",
       main: "dist/index.js",
@@ -79189,6 +79238,21 @@ function classifyProviderError(message) {
   return /ENOTFOUND|ECONNREFUSED|ETIMEDOUT|EAI_AGAIN|fetch failed|socket hang up|network/i.test(message) ? "provider_unreachable" : "internal";
 }
 async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
+  const started = performance.now();
+  const observation = {
+    version: 1,
+    source: "runtime",
+    sessionInitMs: null,
+    totalMs: 0,
+    modelCalls: [],
+    droppedModelCalls: 0
+  };
+  const result = await runTurnExecution(envelope2, egress, deps, onToolEvent, observation);
+  observation.totalMs = performance.now() - started;
+  result.observation = observation;
+  return result;
+}
+async function runTurnExecution(envelope2, egress, deps, onToolEvent, observation) {
   const createSession = deps.createSession ?? defaultCreateSession;
   const version = runtimeVersion();
   const env = egressSessionEnv(envelope2, egress);
@@ -79196,7 +79260,9 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
   let session = null;
   let t8 = null;
   let timedOut = false;
-  const t7 = Date.now();
+  let t7 = null;
+  let sawRequest = false;
+  const initStarted = performance.now();
   try {
     let tenantComponents;
     if (envelope2.tenantComponents?.length) {
@@ -79220,11 +79286,20 @@ async function runTurnOnce(envelope2, egress, deps = {}, onToolEvent) {
       ...onToolEvent ? { onToolEvent } : {}
     });
     session.subscribe((event) => {
+      if (event.type === "model_request_started" && !sawRequest && typeof event.startedAt === "number") {
+        t7 = event.startedAt;
+        sawRequest = true;
+      }
+      if (event.type === "model_call_observed" && event.call) {
+        if (observation.modelCalls.length < 128) observation.modelCalls.push(event.call);
+        else observation.droppedModelCalls += 1;
+      }
       if (timedOut || t8 !== null) return;
       if (event.type === "text_delta" && event.deltaKind === "text" && event.delta) {
         t8 = Date.now();
       }
     });
+    observation.sessionInitMs = performance.now() - initStarted;
     const timer = setTimeout(() => {
       timedOut = true;
       void session?.interrupt().catch(() => void 0);
@@ -79920,7 +79995,7 @@ async function readResponseError(res) {
 
 // src/cli/index.ts
 init_ui();
-var VERSION3 = "2.2.63";
+var VERSION3 = "2.2.64";
 function buildProgram() {
   const program = new Command32("prismer").description("Prismer Runtime host and local control CLI (TS-only).").version(VERSION3).addHelpText(
     "after",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@prismer/runtime",
-  "version": "2.2.63",
+  "version": "2.2.64",
   "description": "Prismer Cloud daemon runtime — TS-only adapter host for hosted IM agents",
   "type": "module",
   "main": "dist/index.js",