@springbrand/agent-runtime 0.2.0-alpha.72 → 0.3.0-alpha.1

package/src/lib/mcp.ts

@@ -1,5 +1,5 @@
 import type { RetryOptions } from "agents";
-import { withRuntimeLoadTimeout } from "../kernel/runtime-load";
+import { withRuntimeLoadTimeout } from "../config/runtime-load";
 
 const MCP_RETRY: RetryOptions = {
   maxAttempts: 3,

package/src/{pi/runtime-adapter/models.ts → models/index.ts}

@@ -48,9 +48,9 @@ import type {
   RuntimeModelInputModality,
   RuntimeModelProtocol,
   RuntimeProviderPort,
-} from "../../kernel/bindings";
-import { modelInputMessages } from "../message/model-input";
-import { openRouterMessagesApi } from "./openrouter-messages";
+} from "../config/ports";
+import { modelInputMessages } from "../conversation/message/model-input";
+import { openRouterMessagesApi } from "./openrouter/messages";
 
 const CATALOGS = {
   "openai-chat": openaiProvider().getModels(),
@@ -60,8 +60,6 @@ const CATALOGS = {
   "openai-codex-responses": openaiCodexProvider().getModels(),
 } satisfies Record<RuntimeModelProtocol, readonly Model<Api>[]>;
 
-const PROVIDER_MAX_RETRIES = 2;
-
 const MESSAGES_HOSTED_INPUT_MODALITIES = ["image", "file"] as const;
 const NO_HOSTED_INPUT_MODALITIES: readonly RuntimeModelInputModality[] = [];
 
@@ -83,7 +81,7 @@ function adapterHostedInputModalities(
 // 结构上不可能收敛**,表现为前端 think 转到恢复预算耗尽为止。
 //
 // 调到 240s:比实测静默期留约 7 倍余量。代价是连接真死时单次要等更久,
-// 所以 runtime.ts 同时把 stall 的恢复次数单独收窄(见 CHAT_STALL_MAX_ATTEMPTS)。
+// Harness owns bounded provider retries; this timeout only detects a stream that stopped making progress.
 export const MODEL_STREAM_STALL_TIMEOUT_MS = 240_000;
 export const MODEL_STREAM_STALL_MESSAGE =
   `Chat stream stalled: no activity for ${MODEL_STREAM_STALL_TIMEOUT_MS}ms; the turn was aborted by the stall watchdog.`;
@@ -113,6 +111,12 @@ export type PiGenerationLifecycleEvent =
       readonly generationId: string;
       readonly timestamp: number;
       readonly durationMs: number;
+      readonly responseHeadersMs?: number;
+      readonly firstRawEventMs?: number;
+      readonly firstMeaningfulEventMs?: number;
+      readonly firstTextEventMs?: number;
+      readonly sourceWaitMs?: number;
+      readonly forwardingOverheadMs?: number;
       readonly outcome: "success" | "error" | "cancelled";
       readonly stopReason?: string;
       readonly errorName?: string;
@@ -302,8 +306,12 @@ async function* stopStalledModelStream(
   let stalled = false;
   let stallError: ChatStreamStalledError | undefined;
   let idleWaitMs = 0;
+  let sourceWaitMs = 0;
+  let forwardingOverheadMs = 0;
+  let yieldedAt: number | undefined;
   let rawEventCount = 0;
   let meaningfulEventCount = 0;
+  let firstTextSeen = false;
   let lastRawEventAt: number | undefined;
   let lastRawEventType: AssistantMessageEvent["type"] | undefined;
   const streamedContent = new Set<string>();
@@ -344,6 +352,10 @@ async function* stopStalledModelStream(
   });
   try {
     while (true) {
+      if (yieldedAt !== undefined) {
+        forwardingOverheadMs += Math.max(0, Date.now() - yieldedAt);
+        yieldedAt = undefined;
+      }
       const waitStartedAt = Date.now();
       const nextPromise = iterator
         ? iterator.next()
@@ -375,6 +387,7 @@ async function* stopStalledModelStream(
         timer = undefined;
       }
       if (stalled) throw stallError;
+      sourceWaitMs += Math.max(0, Date.now() - waitStartedAt);
       if (next.done) break;
       const event = next.value;
       rawEventCount += 1;
@@ -395,6 +408,8 @@ async function* stopStalledModelStream(
           output: event.type === "done" ? event.message : event.error,
           rawEventCount,
           meaningfulEventCount,
+          sourceWaitMs,
+          forwardingOverheadMs,
         });
         yield event;
         return;
@@ -410,6 +425,15 @@ async function* stopStalledModelStream(
         lastMeaningfulActivityType = progress;
         idleWaitMs = 0;
       }
+      if (
+        !firstTextSeen &&
+        (event.type === "text_delta" || event.type === "text_end") &&
+        (event.type === "text_delta" ? event.delta : event.content).trim()
+      ) {
+        firstTextSeen = true;
+        probe("first_text_event", { eventType: event.type });
+      }
+      yieldedAt = Date.now();
       yield event;
     }
   } finally {
@@ -431,7 +455,7 @@ function trimTrailingSlash(value: string): string {
 /**
  * 算出一个已解析模型真正会被请求的 URL。
  *
- * {@link withProviderRetry} 在每次派发模型请求前调用它写观测日志；诊断"这个 Agent 到底调了哪个 LLM"时也可以直接复用。
+ * {@link withProviderBoundary} 在每次派发模型请求前调用它写观测日志；诊断"这个 Agent 到底调了哪个 LLM"时也可以直接复用。
  *
  * 各协议的路径由底层 SDK 决定，这里必须与之逐条对齐：OpenAI 兼容 SDK 用 `baseURL + /chat/completions`，
  * Anthropic SDK 用 `baseURL + /v1/messages`（纯字符串拼接，不会去重 `/v1`），Google 的 baseUrl 已含版本段，
@@ -458,18 +482,15 @@ export function modelRequestUrl(model: Model<Api>): string {
 /**
  * 给模型请求补上可中断的空闲终止边界。
  *
- * Runtime Turn 和 SubAgent 在把 `Models.streamSimple` 交给 Pi 前调用；显式传入的重试次数优先。
- *
- * 调用方可以覆盖默认的 2 次 Provider 重试；主 Turn 传 0，由 Submission
- * 统一持有恢复预算。连续 {@link MODEL_STREAM_STALL_TIMEOUT_MS} 毫秒没有可展示进展时中止 provider。
+ * Runtime Turn 和 SubAgent 在把 `Models.streamSimple` 交给 Harness 前调用。
+ * 本边界只负责载荷转换、观测和空闲超时；重试次数由 Harness 传入，这里不提供第二个默认值。
  *
  * 可选 observer 接收配对的 generation started/finished 事实及实际模型输入输出；响应头仅保留
  * 白名单 request id，永不携带 key 或任意响应 header。正文由 telemetry 隐私模式统一脱敏。
  * observer 的异常被隔离，不能改变模型执行结果。
  */
-export function withProviderRetry(
+export function withProviderBoundary(
   streamFn: StreamFn,
-  defaultMaxRetries = PROVIDER_MAX_RETRIES,
   defaultSessionId?: string,
   onGeneration?: PiGenerationLifecycleObserver,
 ): StreamFn {
@@ -503,6 +524,12 @@ export function withProviderRetry(
       let responseCount = 0;
       let upstreamRequestId: string | undefined;
       let responseStatus: number | undefined;
+      let responseHeadersMs: number | undefined;
+      let firstRawEventMs: number | undefined;
+      let firstMeaningfulEventMs: number | undefined;
+      let firstTextEventMs: number | undefined;
+      let sourceWaitMs: number | undefined;
+      let forwardingOverheadMs: number | undefined;
       let finished = false;
       const finish = (
         outcome: "success" | "error" | "cancelled",
@@ -515,11 +542,18 @@ export function withProviderRetry(
       ) => {
         if (finished) return;
         finished = true;
+        const finishedAt = Date.now();
         notify({
           type: "finished",
           generationId: requestId,
-          timestamp: Date.now(),
-          durationMs: Math.max(0, Date.now() - startedAt),
+          timestamp: finishedAt,
+          durationMs: Math.max(0, finishedAt - startedAt),
+          responseHeadersMs,
+          firstRawEventMs,
+          firstMeaningfulEventMs,
+          firstTextEventMs,
+          sourceWaitMs,
+          forwardingOverheadMs,
           outcome,
           ...details,
           ...(upstreamRequestId ? { upstreamRequestId } : {}),
@@ -527,7 +561,9 @@ export function withProviderRetry(
         });
       };
       const probe = (phase: string, details: Record<string, unknown> = {}) => {
+        const elapsedMs = Math.max(0, Date.now() - startedAt);
         if (phase === "response_headers") {
+          responseHeadersMs ??= elapsedMs;
           if (typeof details.upstreamRequestId === "string") {
             upstreamRequestId = details.upstreamRequestId;
           }
@@ -536,6 +572,17 @@ export function withProviderRetry(
           }
           return;
         }
+        if (phase === "first_raw_event") firstRawEventMs ??= elapsedMs;
+        if (phase === "first_meaningful_event") {
+          firstMeaningfulEventMs ??= elapsedMs;
+        }
+        if (phase === "first_text_event") firstTextEventMs ??= elapsedMs;
+        if (typeof details.sourceWaitMs === "number") {
+          sourceWaitMs = details.sourceWaitMs;
+        }
+        if (typeof details.forwardingOverheadMs === "number") {
+          forwardingOverheadMs = details.forwardingOverheadMs;
+        }
         if (phase === "done") {
           finish("success", {
             ...(typeof details.stopReason === "string"
@@ -579,7 +626,7 @@ export function withProviderRetry(
           signal: options?.signal
             ? AbortSignal.any([options.signal, watchdog.signal])
             : watchdog.signal,
-          maxRetries: options?.maxRetries ?? defaultMaxRetries,
+          maxRetries: options?.maxRetries,
           sessionId,
           onPayload: async (payload, activeModel) =>
             pdfPayload(
@@ -743,7 +790,6 @@ function configuredModel(
       ? {
           compat: {
             ...anthropicMessagesCompat,
-            // OpenRouter rejects Anthropic's mid-conversation effort messages and block binding.
             supportsMidConvoEffort: false,
             supportsEagerToolInputStreaming: false,
             supportsStrictTools:
@@ -789,7 +835,7 @@ export function resolvePiModel(
 /**
  * 用当前部署端点完整重建一个已有的 Pi 模型集合。
  *
- * PiRuntimeAdapter 激活新运行快照时调用它；需要复用 MutableModels 的调用方也可以调用，并继续使用返回的同一实例。
+ * HarnessRuntimeConfig 激活新运行快照时调用它；需要复用 MutableModels 的调用方也可以调用，并继续使用返回的同一实例。
  * 该函数在全部新 provider 校验通过后才清空并替换集合，所以失败时旧集合保持可用。
  *
  * Pi 当前用 provider.id 在 Map 中注册和路由请求，因此每个部署端点都必须同时注册唯一 provider、对应模型、认证解析器和协议实现。