nolo-cli 0.54.0-alpha.2 → 0.61.0-alpha.2

package/{chunk-INXBNYCG.js → chunk-7A3RRUCQ.js}

@@ -1,6 +1,6 @@
 import {
   sanitizeForOutbound
-} from "./chunk-T27IEBNQ.js";
+} from "./chunk-2QDDKFG5.js";
 import {
   accumulateToolCallDelta,
   createThinkParserState,
@@ -15,15 +15,17 @@ import {
   processThinkChunk
 } from "./chunk-N7J63Q6C.js";
 import {
+  collectStablePromptPrefix,
   convertMessagesToResponsesInput,
   extractTextFromResponseOutput,
   extractToolCallsFromResponseOutput,
   getUsageRequestOptions,
   hasImageParts,
   shouldStripReasoningContentForOutbound,
+  stablePromptCacheKey,
   toOpenAiCompatibleMessages,
   toResponsesTools
-} from "./chunk-6BYAXNH5.js";
+} from "./chunk-XTJTN3RW.js";
 import {
   buildProviderExecutionPlan,
   canUsePlatformChatProvider,
@@ -33,13 +35,15 @@ import {
   resolvePlatformAuthToken,
   resolvePlatformServerUrl,
   resolveProviderTransportDecision
-} from "./chunk-SP3DPVBH.js";
+} from "./chunk-RDBARW6T.js";
 import {
-  clipCompactText
-} from "./chunk-H4RYESI6.js";
+  clipCompactText,
+  isMeaningfulChatCompletionOutput,
+  isMeaningfulResponsesOutput
+} from "./chunk-Z4QSJILV.js";
 import {
   NOLO_CLIENT_VERSION_HEADER
-} from "./chunk-WL3UOGYO.js";
+} from "./chunk-22BTGNSP.js";
 import {
   isRecord
 } from "./chunk-KUG52IAP.js";
@@ -239,6 +243,11 @@ function finalizeResponsesToolCalls(calls) {
 }
 
 // packages/agent-runtime/processChatCompletionDelta.ts
+function observeFirstMeaningfulChatCompletionDelta(state) {
+  if (!state.onMeaningfulDelta || state.sawMeaningfulDelta) return;
+  state.sawMeaningfulDelta = true;
+  state.onMeaningfulDelta();
+}
 function extractChatCompletionStreamError(parsed) {
   const raw = parsed?.error;
   if (!raw) return null;
@@ -281,6 +290,9 @@ function applyChatCompletionDelta(parsed, state) {
   }
   const delta = parsed?.choices?.[0]?.delta;
   if (!delta || typeof delta !== "object") return false;
+  if (isMeaningfulChatCompletionOutput(delta)) {
+    observeFirstMeaningfulChatCompletionDelta(state);
+  }
   const reasoningChunk = typeof delta.reasoning_content === "string" ? delta.reasoning_content : typeof delta.reasoning === "string" ? delta.reasoning : "";
   if (reasoningChunk) {
     state.reasoning += reasoningChunk;
@@ -353,6 +365,16 @@ async function* readSseFrames(response) {
       boundary = findSseFrameBoundary(buffer);
     }
   }
+  const remaining = decoder.decode();
+  if (remaining) {
+    buffer += remaining;
+    let boundary = findSseFrameBoundary(buffer);
+    while (boundary) {
+      yield buffer.slice(0, boundary.index);
+      buffer = buffer.slice(boundary.index + boundary.separatorLength);
+      boundary = findSseFrameBoundary(buffer);
+    }
+  }
   if (buffer.trim()) yield buffer;
 }
 function findSseFrameBoundary(buffer) {
@@ -378,6 +400,16 @@ async function* streamSseDataValues(response, parse) {
       if (parsed) yield parsed;
     }
   }
+  const remaining = decoder.decode();
+  if (remaining) {
+    buffer += remaining;
+    const lines = buffer.split("\n");
+    buffer = lines.pop() ?? "";
+    for (const line of lines) {
+      const parsed = parse(line);
+      if (parsed) yield parsed;
+    }
+  }
   if (buffer.trim()) {
     const parsed = parse(buffer);
     if (parsed) yield parsed;
@@ -392,6 +424,45 @@ async function readSseDataValues(response, parse) {
 }
 
 // packages/agent-runtime/platformChatProvider.ts
+var PlatformStreamInterruptedError = class extends Error {
+  code = "PLATFORM_STREAM_INTERRUPTED";
+  name = "PlatformStreamInterrupted";
+  phase;
+  httpStatus;
+  frameCount;
+  textChars;
+  sawDone;
+  sawUsage;
+  sawFinishReason;
+  elapsedMs;
+  constructor(args) {
+    super(
+      args.message ?? `Platform LLM stream interrupted (${args.phase}, frames=${args.frameCount ?? 0}, chars=${args.textChars ?? 0})`,
+      args.cause !== void 0 ? { cause: args.cause } : void 0
+    );
+    this.name = "PlatformStreamInterrupted";
+    this.phase = args.phase;
+    this.httpStatus = args.httpStatus ?? (args.phase === "streaming" ? 200 : 0);
+    this.frameCount = args.frameCount ?? 0;
+    this.textChars = args.textChars ?? 0;
+    this.sawDone = args.sawDone ?? false;
+    this.sawUsage = args.sawUsage ?? false;
+    this.sawFinishReason = args.sawFinishReason ?? false;
+    this.elapsedMs = args.elapsedMs ?? 0;
+    if (args.cause !== void 0 && this.cause === void 0) {
+      this.cause = args.cause;
+    }
+  }
+};
+function isAbortOrTimeoutError(error) {
+  if (typeof DOMException !== "undefined" && error instanceof DOMException) {
+    return error.name === "AbortError" || error.name === "TimeoutError";
+  }
+  if (error instanceof Error) {
+    return error.name === "AbortError" || error.name === "TimeoutError";
+  }
+  return false;
+}
 var CHAT_PROXY_PATH = "/api/v1/chat";
 function shouldDisableThinking(providerConfig) {
   return providerConfig.provider.toLowerCase() === "mimo" || /xiaomimimo\.com/i.test(providerConfig.endpoint);
@@ -477,7 +548,8 @@ function buildPlatformChatCompletionRequest(args) {
     },
     stream: args.stream ?? false,
     ...args.stream ? getUsageRequestOptions(args.providerConfig.usageProvider ?? args.providerConfig.provider, {
-      api: usesResponsesApi ? "responses" : "chat-completions"
+      api: usesResponsesApi ? "responses" : "chat-completions",
+      endpoint: usesResponsesApi ? null : args.providerConfig.endpoint
     }) : {},
     ...requestOptions,
     ...args.tools && args.tools.length > 0 ? {
@@ -485,6 +557,23 @@ function buildPlatformChatCompletionRequest(args) {
       ...usesResponsesApi ? {} : { tool_choice: "auto" }
     } : {},
     ...shouldDisableThinking(args.providerConfig) ? { thinking: { type: "disabled" } } : {},
+    // 缓存路由亲和提示：托管上游是多副本共享池，前缀缓存是节点本地的，
+    // 请求落到另一副本就整条 miss（实测占该线全部 miss token 的 52%，
+    // 依据见 promptCacheKey.ts）。键只取跨轮稳定的部分（model + system 稳定
+    // 前缀 + 工具定义），历史增长不参与，否则键每轮都变等于没有键。
+    // 已实测 RunInfra 接受该字段且不破坏既有命中；效果待上线后按
+    // promptCacheKey.ts 末尾的口径复测。
+    prompt_cache_key: stablePromptCacheKey(
+      [
+        args.providerConfig.model,
+        // 只从消息里取：本 builder 只拿得到 providerConfig，没有 agentConfig，
+        // agent prompt 早已由 localLoop 的 buildMessages 拼进 system 稳定前缀，
+        // 所以 sanitizedMessages 里的 system 就是完整取材，无需 fallback。
+        collectStablePromptPrefix(sanitizedMessages),
+        args.tools ?? []
+      ],
+      "nolo-chat"
+    ),
     url: args.providerConfig.endpoint,
     provider: args.providerConfig.provider,
     agentKey: args.providerConfig.agentKey,
@@ -500,6 +589,18 @@ function buildPlatformChatCompletionRequest(args) {
       headers: {
         "Content-Type": "application/json",
         Authorization: `Bearer ${args.providerConfig.authToken}`,
+        ...args.requestId ? { "x-request-id": args.requestId } : {},
+        // 流式请求显式放弃压缩协商。实测记录（2026-09-01，glm-5-3-flash 经
+        // nolo.chat）：某个时间窗内 SSE 响应带 content-encoding: gzip 返回，
+        // 整条流被攒成 1 个 chunk，首字节 = 总生成时长（5.8/10.6/16.3s）；
+        // 同一分钟交替发起的 identity 请求为 123-138 chunks、首字节
+        // 1.2-3.6s。即压缩中间件一旦介入，流式的首个输出收益归零。
+        // 该窗口之后线上不再压缩 SSE（同一探针 6/6 未压缩），具体是哪一跳
+        // 压的没有定论；边缘侧已把流式路径排除出 encode
+        // （configureCaddyProxy.sh），这里是不依赖任何中间跳行为的客户端兜底
+        // ——用户侧本地代理同样可能压。SSE 正文只有几十 KB，放弃压缩的带宽
+        // 代价可忽略。
+        ...args.stream ? { "Accept-Encoding": "identity" } : {},
         // 客户端版本闸门第 2 层：与鉴权头同处一处统一注入，避免某条命令路径
         // 漏发。server 据此拒绝「该客户端认不了 wire 的托管模型」；不发这个头
         // 的旧客户端在 server 侧 fail-open（已知局限）。
@@ -653,15 +754,24 @@ function processPlatformChatSseEvent(event, state) {
       if (dsml.content) {
         state.content += dsml.content;
         state.onTextDelta?.(dsml.content);
+        if (isMeaningfulResponsesOutput(parsed)) {
+          observeFirstMeaningfulChatCompletionDelta(state);
+        }
       }
       continue;
     }
     if (eventType === "response.reasoning_text.delta" && typeof parsed?.delta === "string") {
       state.reasoning += parsed.delta;
       state.onReasoningDelta?.(parsed.delta);
+      if (isMeaningfulResponsesOutput(parsed)) {
+        observeFirstMeaningfulChatCompletionDelta(state);
+      }
       continue;
     }
     if (eventType === "response.output_item.added" || eventType === "response.output_item.done" || eventType === "response.function_call_arguments.delta" || eventType === "response.function_call_arguments.done") {
+      if (isMeaningfulResponsesOutput(parsed)) {
+        observeFirstMeaningfulChatCompletionDelta(state);
+      }
       applyResponsesToolEvent(state.responsesToolCalls, parsed);
       continue;
     }
@@ -671,6 +781,8 @@ function processPlatformChatSseEvent(event, state) {
   }
 }
 async function readPlatformChatSseCompletion(args) {
+  const startedAtMs = Date.now();
+  let frameCount = 0;
   const state = {
     content: "",
     reasoning: "",
@@ -679,6 +791,7 @@ async function readPlatformChatSseCompletion(args) {
     usesResponsesApi: args.usesResponsesApi,
     onTextDelta: args.onTextDelta,
     onReasoningDelta: args.onReasoningDelta,
+    ...args.onMeaningfulDelta ? { onMeaningfulDelta: args.onMeaningfulDelta } : {},
     completedResponsesPayload: void 0,
     responsesToolCalls: createResponsesToolAccumulator(),
     dsmlToolCallState: createDsmlParserState(),
@@ -689,8 +802,35 @@ async function readPlatformChatSseCompletion(args) {
     finishReason: void 0,
     sawDone: false
   };
-  for await (const frame of readSseFrames(args.response)) {
-    processPlatformChatSseEvent(frame, state);
+  try {
+    for await (const frame of readSseFrames(args.response)) {
+      frameCount += 1;
+      processPlatformChatSseEvent(frame, state);
+    }
+  } catch (cause) {
+    if (cause instanceof PlatformStreamInterruptedError) {
+      throw cause;
+    }
+    if (isAbortOrTimeoutError(cause)) {
+      throw cause;
+    }
+    const elapsedMs = Date.now() - startedAtMs;
+    const textChars = state.content.length;
+    const sawDone = state.sawDone;
+    const sawUsage = Boolean(state.usage);
+    const sawFinishReason = Boolean(state.finishReason);
+    throw new PlatformStreamInterruptedError({
+      message: cause instanceof Error ? cause.message : "Platform chat SSE stream interrupted",
+      phase: "streaming",
+      httpStatus: 200,
+      frameCount,
+      textChars,
+      sawDone,
+      sawUsage,
+      sawFinishReason,
+      elapsedMs,
+      cause
+    });
   }
   flushChatCompletionStream(state);
   if (state.usesResponsesApi) {
@@ -706,6 +846,8 @@ async function readPlatformChatSseCompletion(args) {
     if (dsml.content) state.content += dsml.content;
   }
   throwIfChatCompletionStreamFailed(state);
+  const streamComplete = state.sawDone || Boolean(state.usage) || Boolean(state.finishReason);
+  const streamTruncated = !streamComplete;
   if (state.usesResponsesApi) {
     const response = state.completedResponsesPayload;
     const completedToolCalls = response ? extractToolCallsFromResponseOutput(response) : [];
@@ -721,13 +863,13 @@ async function readPlatformChatSseCompletion(args) {
     );
     const tool_calls2 = completedToolCalls.length > 0 ? completedToolCalls : normalizedDsmlCalls.length > 0 ? normalizedDsmlCalls : completionsToolCalls.length > 0 ? completionsToolCalls : finalizeResponsesToolCalls(state.responsesToolCalls);
     const finalUsage2 = mergeBillingIntoUsage(state.usage, state.billing);
-    const streamComplete2 = state.sawDone || !!state.usage || !!state.finishReason;
     return {
       content: normalizedFinal.content || state.content,
       ...state.reasoning ? { reasoning_content: state.reasoning } : {},
       ...tool_calls2.length > 0 ? { tool_calls: tool_calls2 } : {},
       ...finalUsage2 ? { usage: finalUsage2 } : {},
-      ...streamComplete2 ? { stream_complete: true } : {},
+      ...streamComplete ? { stream_complete: true } : {},
+      ...streamTruncated ? { streamTruncated: true } : {},
       // finish_reason 供 executePlatformChatCompletion 穿透给 localLoop 的
       // 空轮/截断判定；此前 reader 从不回传，流式路径恒缺。
       ...state.finishReason ? { finish_reason: state.finishReason } : {}
@@ -735,13 +877,13 @@ async function readPlatformChatSseCompletion(args) {
   }
   const tool_calls = finalizeAccumulatedToolCalls(state.accumulatedToolCalls);
   const finalUsage = mergeBillingIntoUsage(state.usage, state.billing);
-  const streamComplete = state.sawDone || !!state.usage || !!state.finishReason;
   return {
     content: state.content,
     ...state.reasoning ? { reasoning_content: state.reasoning } : {},
     ...tool_calls.length > 0 ? { tool_calls } : {},
     ...finalUsage ? { usage: finalUsage } : {},
     ...streamComplete ? { stream_complete: true } : {},
+    ...streamTruncated ? { streamTruncated: true } : {},
     ...state.finishReason ? { finish_reason: state.finishReason } : {}
   };
 }
@@ -785,6 +927,7 @@ export {
   createResponsesToolAccumulator,
   applyResponsesToolEvent,
   finalizeResponsesToolCalls,
+  PlatformStreamInterruptedError,
   resolvePlatformChatProviderConfig,
   buildPlatformChatCompletionRequest,
   parsePlatformChatCompletionData,