mslxdff 0.1.112 → 0.1.114

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mslxdff",
-  "version": "0.1.112",
+  "version": "0.1.114",
   "description": "测试项目，请勿使用。",
   "type": "module",
   "bin": {

package/src/chat-pipeline/serial-trial.js

@@ -49,6 +49,8 @@ export async function runSerialTrial(ctx, deps = {}) {
   for (let idx = 0; idx < order.length; idx++) {
     const model = order[idx];
     handlerCtx.model = model;
+    handlerCtx.orderLen = order.length;
+    handlerCtx.idx = idx;
     evt("model-try", { reqId, model, idx, remaining: order.length - idx });
     if (plugins?.length) {
       const bt = await runHook(plugins, "model:beforeTry", { reqId, requested, model, idx, hops });

package/src/routes/chat/relay-pipeline.js

@@ -3,6 +3,13 @@ import { recordModelStats } from "../../state.js";
 import { normalizeFullId } from "../../providers/model-id.js";
 import { computeMetrics, extractUsageFromJson, extractUsageFromSseText } from "../../metrics.js";
 
+// 唯一/最后候选没有 failover 去向：首块闸门退化为纯"防连接泄漏"，放宽避免误杀慢模型
+// （参考 opencode：zen 通道不设超时；openai responses 硬编码 300s headerTimeout）
+const LAST_CANDIDATE_TIMEOUT_MS = (() => {
+  const n = Number(process.env.MSLXDFF_LAST_CANDIDATE_TIMEOUT_MS);
+  return Number.isInteger(n) && n >= 0 ? n : 120_000;
+})();
+
 /**
  * RelayPipeline 深模块
  * 把 5 个 handler 各自的 fallback→relay→scoring→事件 6段流水收敛为单一真相。
@@ -72,9 +79,16 @@ export function createRelayPipeline({
     }
     _evt("relay-start", { reqId, model: actual, via, isStream: Boolean(body?.stream), fallback });
 
-    // 3. relay
+    // 3. relay（唯一/最后候选：无 failover 去向 → 闸门放宽到防泄漏级别）
+    const orderLen = handlerCtx?.orderLen;
+    const curIdx = handlerCtx?.idx;
+    const isLastCandidate =
+      Number.isInteger(orderLen) && orderLen > 0 &&
+      (orderLen === 1 || (Number.isInteger(curIdx) && curIdx >= orderLen - 1));
+    const streamTimeoutMs = isLastCandidate ? LAST_CANDIDATE_TIMEOUT_MS : C.STREAM_TIMEOUT_MS;
     const out = await _relay(res, upRes, body, {
       fallback,
+      streamTimeoutMs,
       onFirstChunk: (delta) => {
         try { markFn(`ttf-${actual}`); } catch {}
         _evt("relay-first-chunk", { reqId, model: actual, ttfMs: delta, via });
@@ -99,12 +113,12 @@ export function createRelayPipeline({
     });
 
     // 5a. 首块超时未写字节 → 回退
-    if (out.status === C.STREAM_TIMEOUT_MS) {
-      if (auto) try { await auto.recordError(actual, { status: 502, slow: true, note: `stream timeout ${C.STREAM_TIMEOUT_MS}ms` }); } catch {}
-      try { _logError(actual, 502, `stream timeout ${C.STREAM_TIMEOUT_MS}ms`); } catch {}
+    if (streamTimeoutMs > 0 && out.status === streamTimeoutMs) {
+      if (auto) try { await auto.recordError(actual, { status: 502, slow: true, note: `stream timeout ${streamTimeoutMs}ms` }); } catch {}
+      try { _logError(actual, 502, `stream timeout ${streamTimeoutMs}ms`); } catch {}
       _evt("upstream-error", { reqId, model: actual, status: 502, message: "stream timeout", timing: null });
       _evt("fallback", { reqId, from: actual, to: null, reason: "stream timeout" });
-      return { handled: false, upRes: null, lastErr: { model: actual, upstream: null, status: 502, message: `stream timed out after ${C.STREAM_TIMEOUT_MS}ms` } };
+      return { handled: false, upRes: null, lastErr: { model: actual, upstream: null, status: 502, message: `stream timed out after ${streamTimeoutMs}ms` } };
     }
 
     // 5b. 中断（stall 超时 / max 流时长）

package/src/routes/stream.js

@@ -2,6 +2,23 @@ import { performance } from "node:perf_hooks";
 import { applyFallbackHeaders, enrichNonStreamJson, enrichSseChunkText } from "./fallback.js";
 import { json } from "./helpers.js";
 
+// SDK 通道（TextEncoder）产出 Uint8Array，legacy 通道为 Buffer；
+// 统一转文本，避免 [DONE]/finish_reason/usage/chars 统计在 SDK 路径下静默失效。
+function chunkText(chunk) {
+  if (typeof chunk === "string") return chunk;
+  if (Buffer.isBuffer(chunk)) return chunk.toString("utf8");
+  if (chunk instanceof Uint8Array) return Buffer.from(chunk).toString("utf8");
+  return "";
+}
+
+// body.cancel() 可能返回非 Promise（自定义/AI SDK 流）——同步异常与 rejection 双路径都要吞掉
+function cancelBody(body) {
+  try {
+    const p = typeof body?.cancel === "function" ? body.cancel() : null;
+    if (p && typeof p.catch === "function") p.catch(() => {});
+  } catch { /* ignore */ }
+}
+
 export const SLOW_TOTAL_MS = (() => {
   const n = Number(process.env.MSLXDFF_SLOW_TOTAL_MS);
   return Number.isInteger(n) && n > 0 ? n : 20_000;
@@ -9,7 +26,8 @@ export const SLOW_TOTAL_MS = (() => {
 
 export const STREAM_TIMEOUT_MS = (() => {
   const n = Number(process.env.MSLXDFF_STREAM_TIMEOUT_MS);
-  return Number.isInteger(n) && n > 0 ? n : 25_000;
+  // 0 = 显式关闭首块超时（慢思考模型专用）；未设/非法值 → 默认 25s
+  return Number.isInteger(n) && n >= 0 ? n : 25_000;
 })();
 
 export const STALL_TIMEOUT_MS = (() => {
@@ -42,6 +60,8 @@ export async function relay(res, upRes, body, { onFirstChunk, onDownstreamAbort,
     if (reason) res.setHeader("x-mslxdff-workbuddy-reason", reason);
     const allow = upRes.headers.get("x-mslxdff-allowlist");
     if (allow) res.setHeader("x-mslxdff-allowlist", allow);
+    const engine = upRes.headers.get("x-mslxdff-upstream-engine");
+    if (engine) res.setHeader("x-mslxdff-upstream-engine", engine);
   } catch {}
   if (fallback) applyFallbackHeaders(res, fallback);
 
@@ -64,6 +84,7 @@ export async function relay(res, upRes, body, { onFirstChunk, onDownstreamAbort,
     downstreamClosed: false,
     usage: null,
     chars: 0,
+    recoveries: 0,
   };
   let prevChunkAt = t0;
   const onClose = () => {
@@ -95,20 +116,22 @@ export async function relay(res, upRes, body, { onFirstChunk, onDownstreamAbort,
           ? setTimeout(() => {
               stalled = true;
               detail.exitReason = "stall";
-              if (typeof upRes.body.cancel === "function") upRes.body.cancel().catch(() => {});
+              cancelBody(upRes.body);
             }, STALL_TIMEOUT_MS)
           : null;
       };
-      let firstTimer = setTimeout(() => {
-        timedOut = true;
-        detail.exitReason = "first-timeout";
-        if (typeof upRes.body.cancel === "function") upRes.body.cancel().catch(() => {});
-      }, streamTimeoutMs);
+      let firstTimer = streamTimeoutMs > 0
+        ? setTimeout(() => {
+            timedOut = true;
+            detail.exitReason = "first-timeout";
+            cancelBody(upRes.body);
+          }, streamTimeoutMs)
+        : null;
       const maxTimer = MAX_STREAM_MS
         ? setTimeout(() => {
             tooLong = true;
             detail.exitReason = "max";
-            if (typeof upRes.body.cancel === "function") upRes.body.cancel().catch(() => {});
+            cancelBody(upRes.body);
           }, MAX_STREAM_MS)
         : null;
       try {
@@ -124,7 +147,7 @@ export async function relay(res, upRes, body, { onFirstChunk, onDownstreamAbort,
           if (gap > SCORE_STALL_MS) detail.stallHits += 1;
           prevChunkAt = now;
           try {
-            const txt = Buffer.isBuffer(chunk) ? chunk.toString("utf8") : typeof chunk === "string" ? chunk : "";
+            const txt = chunkText(chunk);
             if (txt.includes("[DONE]")) detail.sawDone = true;
             const m = txt.match(/"finish_reason"\s*:\s*"([^"]+)"/);
             if (m) detail.sawFinishReason = m[1];
@@ -157,7 +180,7 @@ export async function relay(res, upRes, body, { onFirstChunk, onDownstreamAbort,
             } else {
               // 非 usage 的普通 delta 也累 chars
               try {
-                const txt2 = Buffer.isBuffer(chunk) ? chunk.toString("utf8") : String(chunk);
+                const txt2 = chunkText(chunk);
                 const ms = txt2.match(/"content"\s*:\s*"([^"]*)"/g);
                 if (ms) for (const mm of ms) {
                   const c = JSON.parse(`{${mm}}`);
@@ -166,6 +189,15 @@ export async function relay(res, upRes, body, { onFirstChunk, onDownstreamAbort,
               } catch {}
             }
           } catch { /* ignore */ }
+          // 首块/空闲超时后上游仍吐出了数据 → 只是慢，不是死：撤销超时判定，照常转发
+          //（cancel 是协作式的，缓冲数据仍会到达；丢掉已到达的数据是纯损失）
+          if (timedOut || stalled) {
+            timedOut = false;
+            stalled = false;
+            detail.recoveries = (detail.recoveries || 0) + 1;
+            if (detail.exitReason === "first-timeout" || detail.exitReason === "stall") detail.exitReason = null;
+            if (firstTimer) { clearTimeout(firstTimer); firstTimer = null; }
+          }
           if (timedOut || stalled || tooLong) break;
           if (first) {
             first = false;
@@ -205,7 +237,7 @@ export async function relay(res, upRes, body, { onFirstChunk, onDownstreamAbort,
       }
       if (timedOut && !wroteAny) {
         res.removeListener("close", onClose);
-        return { status: STREAM_TIMEOUT_MS, ttfMs: null, totalMs: Math.round(performance.now() - t0), aborted: true, interrupted: false, detail };
+        return { status: streamTimeoutMs, ttfMs: null, totalMs: Math.round(performance.now() - t0), aborted: true, interrupted: false, detail };
       }
       if ((stalled || tooLong) && wroteAny) {
         interrupted = true;

package/src/upstream-engine/sdk/sse.js

@@ -2,12 +2,34 @@
 // 纯函数序列化器：push(part) 返回该 part 产生的 SSE 文本（可能为空串/组合帧），end() 返回 [DONE]。
 // 见 .scratch/workbuddy-sdk-channel/SPEC.md。
 
+// 数值兼容：扁平 number / { total } 嵌套
+function pickTokens(v) {
+  if (typeof v === "number") return v;
+  if (v && typeof v === "object" && typeof v.total === "number") return v.total;
+  return undefined;
+}
+function num(v) {
+  const n = Number(v);
+  return Number.isFinite(n) ? n : undefined;
+}
+
 export function usageToOpenAI(u) {
   if (!u || typeof u !== "object") return undefined;
-  if (u.raw && typeof u.raw === "object") return u.raw;
-  const prompt = u.inputTokens?.total ?? 0;
-  const completion = u.outputTokens?.total ?? 0;
-  return { prompt_tokens: prompt, completion_tokens: completion, total_tokens: prompt + completion };
+  const raw = u.raw && typeof u.raw === "object" ? u.raw : null;
+  // raw 已是 OpenAI chat 口径 → 直接用（最保真）
+  if (raw && (raw.prompt_tokens != null || raw.completion_tokens != null || raw.total_tokens != null)) return raw;
+  // 否则从 raw（Responses 口径 input_tokens/output_tokens）或 V2 扁平 / V3 嵌套标准字段归一
+  const src = raw || u;
+  const prompt = num(src.input_tokens) ?? pickTokens(src.inputTokens) ?? pickTokens(u.inputTokens) ?? 0;
+  const completion = num(src.output_tokens) ?? pickTokens(src.outputTokens) ?? pickTokens(u.outputTokens) ?? 0;
+  const total = num(src.total_tokens) ?? num(src.totalTokens) ?? pickTokens(u.totalTokens) ?? (prompt + completion);
+  const out = { prompt_tokens: prompt, completion_tokens: completion, total_tokens: total };
+  // Responses 口径的 details 映射到 OpenAI chat details（有则带上）
+  const cached = num(src.input_tokens_details?.cached_tokens ?? src.prompt_tokens_details?.cached_tokens);
+  if (cached != null) out.prompt_tokens_details = { cached_tokens: cached };
+  const reasoningTokens = num(src.output_tokens_details?.reasoning_tokens ?? src.completion_tokens_details?.reasoning_tokens);
+  if (reasoningTokens != null) out.completion_tokens_details = { reasoning_tokens: reasoningTokens };
+  return out;
 }
 
 export function createSseSerializer() {
@@ -72,7 +94,10 @@ export function createSseSerializer() {
         return ensureRole() + frame({ tool_calls: [{ index: idx, id, type: "function", function: { name: part.toolName || "", arguments: args } }] });
       }
       case "finish": {
-        const reason = part.finishReason?.raw ?? part.finishReason?.unified ?? "stop";
+        // V2 spec 的 finishReason 是字符串；V3 spec 是 { unified, raw }
+        const fr = part.finishReason;
+        const raw = typeof fr === "string" ? fr : (fr?.raw ?? fr?.unified);
+        const reason = raw === "tool-calls" ? "tool_calls" : raw === "content-filter" ? "content_filter" : (raw ?? "stop");
         return frame({}, { finishReason: reason, usage: usageToOpenAI(part.usage) });
       }
       case "error": {