mslxdff 0.1.188 → 0.1.197

package/src/transport/sse.js

@@ -1,93 +1,93 @@
-/**
- * SSE 行协议深模块 — 纯函数 + 流式
- * 负责 `data: / : / event: / [DONE]` 的解析与粘包 remain 处理
- */
-
-export function parseSseChunk(buf) {
-  const events = [];
-  let remain = String(buf || "");
-  // 按 \n\n 分帧，\r\n 也兼容
-  // 保留最后一帧若不以 \n\n 结尾则为 remain
-  const normalized = remain.replace(/\r\n/g, "\n");
-  const parts = normalized.split("\n\n");
-  // 最后一部分若原串不以 \n\n 结尾则为 remain
-  const endsWithDelim = normalized.endsWith("\n\n") || normalized.endsWith("\n\r\n");
-  // parts 最后一项在非结尾时为未完成帧
-  const complete = endsWithDelim ? parts : parts.slice(0, -1);
-  remain = endsWithDelim ? "" : parts[parts.length - 1] || "";
-  for (const frame of complete) {
-    if (!frame) continue; // 空帧
-    const lines = frame.split("\n");
-    const dataLines = [];
-    for (const line of lines) {
-      if (!line) continue;
-      if (line.startsWith(":")) continue; // 注释/keepalive
-      if (line.startsWith("event:")) continue; // 事件类型忽略
-      if (line.startsWith("data:")) {
-        let v = line.slice(5);
-        if (v.startsWith(" ")) v = v.slice(1);
-        dataLines.push(v);
-      }
-    }
-    if (dataLines.length === 0) continue;
-    // 多行 data 按 \n 拼接（SSE 规范）
-    const payload = dataLines.join("\n");
-    // 空 data: 仍产生事件，由上层过滤；此处保留以便测试可见
-    events.push(payload);
-  }
-  return { events, remain };
-}
-
-export function createSseParser() {
-  let buf = "";
-  return {
-    push(chunk) {
-      buf += String(chunk || "");
-      const { events, remain } = parseSseChunk(buf);
-      buf = remain;
-      return events;
-    },
-    flush() {
-      // 未形成完整帧的不发（避免半包误触发）
-      if (!buf.trim()) { buf = ""; return []; }
-      // 若残留包含完整 data: 行但缺结尾，不强行发
-      return [];
-    },
-    getRemain() { return buf; },
-  };
-}
-
-export async function* streamFromResponse(res) {
-  if (!res || !res.body) return;
-  const parser = createSseParser();
-  const reader = res.body.getReader ? res.body.getReader() : null;
-  const decoder = new TextDecoder();
-  if (reader) {
-    while (true) {
-      const { done, value } = await reader.read();
-      if (done) break;
-      const text = decoder.decode(value, { stream: true });
-      const evs = parser.push(text);
-      for (const e of evs) {
-        if (e === "[DONE]") return;
-        if (e === "") continue;
-        yield e;
-      }
-    }
-    // flush 残留
-    const rest = parser.flush();
-    for (const e of rest) {
-      if (e === "[DONE]") return;
-      if (e) yield e;
-    }
-  } else if (typeof res.text === "function") {
-    // 回退：一次性 text
-    const txt = await res.text();
-    const evs = parser.push(txt);
-    for (const e of evs) {
-      if (e === "[DONE]") return;
-      if (e === "") continue;
-      yield e;
-    }
-  }
-}
+/**
+ * SSE 行协议深模块 — 纯函数 + 流式
+ * 负责 `data: / : / event: / [DONE]` 的解析与粘包 remain 处理
+ */
+
+export function parseSseChunk(buf) {
+  const events = [];
+  let remain = String(buf || "");
+  // 按 \n\n 分帧，\r\n 也兼容
+  // 保留最后一帧若不以 \n\n 结尾则为 remain
+  const normalized = remain.replace(/\r\n/g, "\n");
+  const parts = normalized.split("\n\n");
+  // 最后一部分若原串不以 \n\n 结尾则为 remain
+  const endsWithDelim = normalized.endsWith("\n\n") || normalized.endsWith("\n\r\n");
+  // parts 最后一项在非结尾时为未完成帧
+  const complete = endsWithDelim ? parts : parts.slice(0, -1);
+  remain = endsWithDelim ? "" : parts[parts.length - 1] || "";
+  for (const frame of complete) {
+    if (!frame) continue; // 空帧
+    const lines = frame.split("\n");
+    const dataLines = [];
+    for (const line of lines) {
+      if (!line) continue;
+      if (line.startsWith(":")) continue; // 注释/keepalive
+      if (line.startsWith("event:")) continue; // 事件类型忽略
+      if (line.startsWith("data:")) {
+        let v = line.slice(5);
+        if (v.startsWith(" ")) v = v.slice(1);
+        dataLines.push(v);
+      }
+    }
+    if (dataLines.length === 0) continue;
+    // 多行 data 按 \n 拼接（SSE 规范）
+    const payload = dataLines.join("\n");
+    // 空 data: 仍产生事件，由上层过滤；此处保留以便测试可见
+    events.push(payload);
+  }
+  return { events, remain };
+}
+
+export function createSseParser() {
+  let buf = "";
+  return {
+    push(chunk) {
+      buf += String(chunk || "");
+      const { events, remain } = parseSseChunk(buf);
+      buf = remain;
+      return events;
+    },
+    flush() {
+      // 未形成完整帧的不发（避免半包误触发）
+      if (!buf.trim()) { buf = ""; return []; }
+      // 若残留包含完整 data: 行但缺结尾，不强行发
+      return [];
+    },
+    getRemain() { return buf; },
+  };
+}
+
+export async function* streamFromResponse(res) {
+  if (!res || !res.body) return;
+  const parser = createSseParser();
+  const reader = res.body.getReader ? res.body.getReader() : null;
+  const decoder = new TextDecoder();
+  if (reader) {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      const text = decoder.decode(value, { stream: true });
+      const evs = parser.push(text);
+      for (const e of evs) {
+        if (e === "[DONE]") return;
+        if (e === "") continue;
+        yield e;
+      }
+    }
+    // flush 残留
+    const rest = parser.flush();
+    for (const e of rest) {
+      if (e === "[DONE]") return;
+      if (e) yield e;
+    }
+  } else if (typeof res.text === "function") {
+    // 回退：一次性 text
+    const txt = await res.text();
+    const evs = parser.push(txt);
+    for (const e of evs) {
+      if (e === "[DONE]") return;
+      if (e === "") continue;
+      yield e;
+    }
+  }
+}

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

@@ -38,6 +38,8 @@ export function createSseSerializer(captured = null) {
   let nextIndex = 0;
   const toolIndex = new Map();
   const toolDeltaIds = new Set();
+  // 「当前正在流式输入的工具」的 index：无 id 的 tool-input-delta 靠它归位（见该 case 的注释）
+  let lastToolIdx = null;
   // 加密思考往返：reasoning-start 的 providerMetadata 带 itemId + 加密态，随首帧透出
   let reasoningMeta = null;
   let reasoningFrameSent = false;
@@ -104,18 +106,30 @@ export function createSseSerializer(captured = null) {
         const idx = nextIndex++;
         toolIndex.set(part.id, idx);
         toolDeltaIds.add(part.id);
+        lastToolIdx = idx;   // 后续无 id 的 delta 归到这个工具
         return ensureRole() + frame({
           tool_calls: [{ index: idx, id: part.id, type: "function", function: { name: part.toolName || "", arguments: "" } }],
         });
       }
       case "tool-input-delta": {
-        const idx = toolIndex.has(part.id) ? toolIndex.get(part.id) : 0;
+        // 未知 id **绝不能回退到 index 0**：那会把这段参数写进另一个工具（实测：callA 占 index 0 时，
+        // 一个 id 丢失的 delta 让 funcA 的参数变成 `{"hijack":true}`），客户端会拿着被污染的参数去调用
+        // 错误的函数。改为按 `tool-call` 分支的口径分配新 index —— 宁可多一个空工具槽，也不污染已有工具。
+        // 无 id 的 delta 属于「当前正在流式输入的那个工具」：复用上一次用过的 index。
+        // 每个都分配新 index 会把一段参数拆成多个工具（实测：无 id 的 `{"a":` 与 `1}` 分别落到
+        // index 1 和 2，两个都是非法 JSON）—— 比旧代码的「都落 index 0」更糟。
+        let idx;
+        if (part.id && toolIndex.has(part.id)) idx = toolIndex.get(part.id);
+        else if (!part.id && lastToolIdx != null) idx = lastToolIdx;   // 无 id：沿用当前工具
+        else { idx = nextIndex++; if (part.id) toolIndex.set(part.id, idx); }
+        lastToolIdx = idx;
         return frame({ tool_calls: [{ index: idx, function: { arguments: String(part.delta ?? "") } }] });
       }
       case "tool-call": {
         const id = part.toolCallId;
         const idx = toolIndex.has(id) ? toolIndex.get(id) : nextIndex++;
         toolIndex.set(id, idx);
+        lastToolIdx = idx;   // 同 tool-input-start：后续无 id 的 delta 归到这个工具
         if (toolDeltaIds.has(id)) return null;
         const args = typeof part.input === "string" ? part.input : JSON.stringify(part.input ?? {});
         return ensureRole() + frame({ tool_calls: [{ index: idx, id, type: "function", function: { name: part.toolName || "", arguments: args } }] });

package/src/upstream-probe/display.js

@@ -1,52 +1,52 @@
-// ADR-0015 展示渲染：-group list 行缀 + -status 汇总（纯函数，输入 routes 数据）
-// 四态文案：无数据→引导；有数据→摘要；坏数据→标 offline
-
-function bestDisplay(entry) {
-  if (!entry) return null;
-  const best = String(entry.best || "direct");
-  const ms = best === "direct" ? (entry.direct?.ttfbMs ?? entry.direct?.totalMs) : (entry.via?.[best.slice(4)]?.ttfbMs ?? entry.via?.[best.slice(4)]?.totalMs);
-  const tag = best === "direct" ? "direct" : `via ${best.slice(4)}`;
-  if (entry.direct?.ok === false && !best.startsWith("via:")) return `${tag} (${entry.direct?.label || "offline"})`;
-  return ms != null ? `${tag} ${ms}ms` : tag;
-}
-
-// -group list 静态成员行尾缀：该成员作为 best 出现时的供应商与相对直连收益
-export function groupRowSuffix(peerUrl, routes) {
-  if (!peerUrl || !routes || typeof routes !== "object") return null;
-  const label = (() => {
-    const raw = String(peerUrl);
-    if (raw.includes("://")) {
-      try { const u = new URL(raw); return `${u.hostname}${u.port ? `:${u.port}` : ""}`; } catch { return raw.slice(-16); }
-    }
-    return raw;
-  })();
-  const wins = [];
-  for (const [key, entry] of Object.entries(routes)) {
-    if (!key.startsWith("provider:")) continue;
-    const prov = key.slice(9);
-    if (entry?.best === `via:${label}`) {
-      const d = entry.deltaMs;
-      wins.push(d != null ? `${prov} ${d > 0 ? "+" : ""}${d}ms` : prov);
-    }
-  }
-  if (!wins.length) return null;
-  return `  via-routes: ${wins.join(", ")}`;
-}
-
-// -status 汇总（多行字符串）：探针条目摘要 ≤5 条；空数据给引导
-export function statusViaSummary({ routes, meta, probeMs }) {
-  const list = Object.entries(routes || {}).filter(([k]) => k.startsWith("provider:"));
-  if (!list.length) {
-    return `via-routes: 无探针数据（后台探针 ${probeMs === 0 ? "已关闭" : "自动运行中"}，加入组并配置 key/token 类供应商后自动生成）`;
-  }
-  const at = meta?.at ? new Date(meta.at) : null;
-  const atStr = at && !Number.isNaN(at.getTime()) ? at.toISOString().slice(11, 19) + "Z" : "";
-  const lines = [`via-routes: ${list.length} 条（探针自动${atStr ? `，${atStr} 刷新` : ""}）`];
-  for (const [key, entry] of list.slice(0, 5)) {
-    const prov = key.slice(9);
-    const b = bestDisplay(entry);
-    if (b) lines.push(`  ${prov} → ${b}${entry.deltaMs != null ? ` (${entry.deltaMs > 0 ? "+" : ""}${entry.deltaMs}ms vs direct)` : ""}`);
-  }
-  if (list.length > 5) lines.push(`  … 其余 ${list.length - 5} 条见 via-routes.json`);
-  return lines.join("\n");
-}
+// ADR-0015 展示渲染：-group list 行缀 + -status 汇总（纯函数，输入 routes 数据）
+// 四态文案：无数据→引导；有数据→摘要；坏数据→标 offline
+
+function bestDisplay(entry) {
+  if (!entry) return null;
+  const best = String(entry.best || "direct");
+  const ms = best === "direct" ? (entry.direct?.ttfbMs ?? entry.direct?.totalMs) : (entry.via?.[best.slice(4)]?.ttfbMs ?? entry.via?.[best.slice(4)]?.totalMs);
+  const tag = best === "direct" ? "direct" : `via ${best.slice(4)}`;
+  if (entry.direct?.ok === false && !best.startsWith("via:")) return `${tag} (${entry.direct?.label || "offline"})`;
+  return ms != null ? `${tag} ${ms}ms` : tag;
+}
+
+// -group list 静态成员行尾缀：该成员作为 best 出现时的供应商与相对直连收益
+export function groupRowSuffix(peerUrl, routes) {
+  if (!peerUrl || !routes || typeof routes !== "object") return null;
+  const label = (() => {
+    const raw = String(peerUrl);
+    if (raw.includes("://")) {
+      try { const u = new URL(raw); return `${u.hostname}${u.port ? `:${u.port}` : ""}`; } catch { return raw.slice(-16); }
+    }
+    return raw;
+  })();
+  const wins = [];
+  for (const [key, entry] of Object.entries(routes)) {
+    if (!key.startsWith("provider:")) continue;
+    const prov = key.slice(9);
+    if (entry?.best === `via:${label}`) {
+      const d = entry.deltaMs;
+      wins.push(d != null ? `${prov} ${d > 0 ? "+" : ""}${d}ms` : prov);
+    }
+  }
+  if (!wins.length) return null;
+  return `  via-routes: ${wins.join(", ")}`;
+}
+
+// -status 汇总（多行字符串）：探针条目摘要 ≤5 条；空数据给引导
+export function statusViaSummary({ routes, meta, probeMs }) {
+  const list = Object.entries(routes || {}).filter(([k]) => k.startsWith("provider:"));
+  if (!list.length) {
+    return `via-routes: 无探针数据（后台探针 ${probeMs === 0 ? "已关闭" : "自动运行中"}，加入组并配置 key/token 类供应商后自动生成）`;
+  }
+  const at = meta?.at ? new Date(meta.at) : null;
+  const atStr = at && !Number.isNaN(at.getTime()) ? at.toISOString().slice(11, 19) + "Z" : "";
+  const lines = [`via-routes: ${list.length} 条（探针自动${atStr ? `，${atStr} 刷新` : ""}）`];
+  for (const [key, entry] of list.slice(0, 5)) {
+    const prov = key.slice(9);
+    const b = bestDisplay(entry);
+    if (b) lines.push(`  ${prov} → ${b}${entry.deltaMs != null ? ` (${entry.deltaMs > 0 ? "+" : ""}${entry.deltaMs}ms vs direct)` : ""}`);
+  }
+  if (list.length > 5) lines.push(`  … 其余 ${list.length - 5} 条见 via-routes.json`);
+  return lines.join("\n");
+}

package/src/upstream-probe/probe.js

@@ -1,49 +1,49 @@
-// ADR-0015 探针薄层：direct GET 探针 + 组员 /v1/relay 代发探针
-// 只测 GET <modelsPath>，不烧 token；组长侧计时。
-import { errMsg } from "../cli/util.js";
-import { compatFetch, timeoutSignal } from "../compat.js";
-
-export const PROBE_TIMEOUT_MS = 5000;
-
-export async function directProbe({ baseUrl, modelsPath = "/models", key = "", timeoutMs = PROBE_TIMEOUT_MS, fetchImpl = compatFetch, clock = Date.now, sampleMs } = {}) {
-  const base = String(baseUrl || "").replace(/\/+$/, "");
-  if (!base) return { ok: false, label: "配置错误", error: "missing baseUrl", ttfbMs: null };
-  const p = String(modelsPath || "/models");
-  const url = `${base}${p.startsWith("/") ? p : `/${p}`}`;
-  const t0 = sampleMs ? sampleMs() : clock();
-  try {
-    const headers = { Accept: "application/json" };
-    if (key) headers.Authorization = `Bearer ${key}`;
-    const res = await fetchImpl(url, { method: "GET", headers, signal: timeoutSignal(timeoutMs) });
-    const ms = sampleMs ? sampleMs() - t0 : clock() - t0;
-    const status = typeof res.status === "number" ? res.status : (res.ok ? 200 : 0);
-    if (status >= 400) return { ok: false, label: `HTTP ${status}`, ttfbMs: null };
-    return { ok: true, ttfbMs: Math.max(0, Math.round(ms)) };
-  } catch (e) {
-    return { ok: false, label: errMsg(e).slice(0, 60), ttfbMs: null };
-  }
-}
-
-// 组员代发：POST <peer>/v1/relay { targetUrl, method:"GET", headers } → 组员 fetch 上游后原样回
-export async function relayViaProbe({ peerUrl, peerToken = "", targetUrl, authHeader = "", timeoutMs = PROBE_TIMEOUT_MS, fetchImpl = compatFetch, clock = Date.now, sampleMs } = {}) {
-  const peer = String(peerUrl || "").replace(/\/+$/, "");
-  if (!peer || !targetUrl) return { ok: false, label: "配置错误", ttfbMs: null };
-  const t0 = sampleMs ? sampleMs() : clock();
-  try {
-    const headers = { "Content-Type": "application/json", Accept: "application/json" };
-    if (peerToken) headers.Authorization = `Bearer ${peerToken}`;
-    const res = await fetchImpl(`${peer}/v1/relay`, {
-      method: "POST",
-      headers,
-      body: JSON.stringify({ targetUrl, method: "GET", headers: authHeader ? { Authorization: authHeader } : {} }),
-      signal: timeoutSignal(timeoutMs),
-    });
-    const ms = sampleMs ? sampleMs() - t0 : clock() - t0;
-    const relayStatus = res.headers?.get?.("x-mslxdff-relay-status");
-    const status = relayStatus != null ? Number(relayStatus) : (res.status ?? 0);
-    if (status >= 400) return { ok: false, label: `HTTP ${status}`, ttfbMs: null };
-    return { ok: true, ttfbMs: Math.max(0, Math.round(ms)) };
-  } catch (e) {
-    return { ok: false, label: errMsg(e).slice(0, 60), ttfbMs: null };
-  }
-}
+// ADR-0015 探针薄层：direct GET 探针 + 组员 /v1/relay 代发探针
+// 只测 GET <modelsPath>，不烧 token；组长侧计时。
+import { errMsg } from "../cli/util.js";
+import { compatFetch, timeoutSignal } from "../compat.js";
+
+export const PROBE_TIMEOUT_MS = 5000;
+
+export async function directProbe({ baseUrl, modelsPath = "/models", key = "", timeoutMs = PROBE_TIMEOUT_MS, fetchImpl = compatFetch, clock = Date.now, sampleMs } = {}) {
+  const base = String(baseUrl || "").replace(/\/+$/, "");
+  if (!base) return { ok: false, label: "配置错误", error: "missing baseUrl", ttfbMs: null };
+  const p = String(modelsPath || "/models");
+  const url = `${base}${p.startsWith("/") ? p : `/${p}`}`;
+  const t0 = sampleMs ? sampleMs() : clock();
+  try {
+    const headers = { Accept: "application/json" };
+    if (key) headers.Authorization = `Bearer ${key}`;
+    const res = await fetchImpl(url, { method: "GET", headers, signal: timeoutSignal(timeoutMs) });
+    const ms = sampleMs ? sampleMs() - t0 : clock() - t0;
+    const status = typeof res.status === "number" ? res.status : (res.ok ? 200 : 0);
+    if (status >= 400) return { ok: false, label: `HTTP ${status}`, ttfbMs: null };
+    return { ok: true, ttfbMs: Math.max(0, Math.round(ms)) };
+  } catch (e) {
+    return { ok: false, label: errMsg(e).slice(0, 60), ttfbMs: null };
+  }
+}
+
+// 组员代发：POST <peer>/v1/relay { targetUrl, method:"GET", headers } → 组员 fetch 上游后原样回
+export async function relayViaProbe({ peerUrl, peerToken = "", targetUrl, authHeader = "", timeoutMs = PROBE_TIMEOUT_MS, fetchImpl = compatFetch, clock = Date.now, sampleMs } = {}) {
+  const peer = String(peerUrl || "").replace(/\/+$/, "");
+  if (!peer || !targetUrl) return { ok: false, label: "配置错误", ttfbMs: null };
+  const t0 = sampleMs ? sampleMs() : clock();
+  try {
+    const headers = { "Content-Type": "application/json", Accept: "application/json" };
+    if (peerToken) headers.Authorization = `Bearer ${peerToken}`;
+    const res = await fetchImpl(`${peer}/v1/relay`, {
+      method: "POST",
+      headers,
+      body: JSON.stringify({ targetUrl, method: "GET", headers: authHeader ? { Authorization: authHeader } : {} }),
+      signal: timeoutSignal(timeoutMs),
+    });
+    const ms = sampleMs ? sampleMs() - t0 : clock() - t0;
+    const relayStatus = res.headers?.get?.("x-mslxdff-relay-status");
+    const status = relayStatus != null ? Number(relayStatus) : (res.status ?? 0);
+    if (status >= 400) return { ok: false, label: `HTTP ${status}`, ttfbMs: null };
+    return { ok: true, ttfbMs: Math.max(0, Math.round(ms)) };
+  } catch (e) {
+    return { ok: false, label: errMsg(e).slice(0, 60), ttfbMs: null };
+  }
+}