mslxdff 0.1.231 → 0.1.233

package/src/providers/cline/chat.js

@@ -1,307 +1,307 @@
-import { joinUrl, sleep } from "../base.js";
-import { appendEvent } from "../../logs.js";
-import { createHash } from "node:crypto";
-import { clineHeaders } from "./headers.js";
-import { createTransport } from "../../transport/index.js";
-import { normalizeProviderId } from "../model-id.js";
-import { recordLimit, recordOutput, exactOutputTokens } from "./usage.js";
-import { createSdkDispatch } from "../../upstream-engine/sdk/dispatch.js";
-
-function genSessionId() { return `sess_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`; }
-
-// 只记哈希标识，不落邮箱 / refreshToken（events.log 可读不可泄）
-function acctId(account) {
-  const seed = String(account?.refreshToken || "");
-  if (!seed) return "slot-unknown";
-  return `acct_${createHash("sha256").update(seed).digest("hex").slice(0, 8)}`;
-}
-// 池子快照：ready/cooling 按「该模型」口径统计（ready=对该模型可用的号数）。
-// isReady 由调用方注入 authPool.accountAvailable(a, model)，避免与 auth 逻辑分叉。
-function poolStat(pool, isReady = () => true) {
-  let cooling = 0;
-  let dead = 0;
-  for (const a of pool || []) {
-    if (a.dead) dead++;
-    else if (!isReady(a)) cooling++;
-  }
-  const total = (pool || []).length;
-  return { total, cooling, dead, ready: total - cooling - dead };
-}
-
-// 429 错误体自带权威模型名（"...Daily free limit reached on model meta/muse-spark-1.3-contributor"）。
-function limitModelOf(bodyText) {
-  // 模型名可含 / . - _ : 字母数字（如 meta/muse-spark-1.3-contributor），
-  // 遇空格/引号/逗号/句号结尾或 "Try again" 前的空白即停。
-  const m = String(bodyText || "").match(/on model\s+([A-Za-z0-9/._:-]+?)(?=[\s"',]|\.\s|$)/i);
-  return m ? m[1] : "";
-}
-
-// stripProviderPrefix 只剥「本供应商」前缀（cline/<裸 id> → 裸 id），非本前缀（含上游自带多段 id
-// 如 meta/vendor/x）原样透传，不误削首段（旧实现按段数剥曾是潜在 bug）。
-function stripProviderPrefix(model, providerId) {
-  const s = String(model || "");
-  const i = s.indexOf("/");
-  if (i <= 0) return s;
-  const low = (v) => normalizeProviderId(String(v || "").toLowerCase());
-  return low(s.slice(0, i)) === low(providerId) ? s.slice(i + 1) : s;
-}
-function unwrapData(obj) {
-  if (obj && obj.data && typeof obj.data === "object") {
-    const d = obj.data;
-    if (d.choices || d.id || d.usage || d.output) return d;
-  }
-  return obj;
-}
-
-async function streamToNonStream(upstream, track = null) {
-  let content = "";
-  let reasoning = "";
-  let finishReason = null;
-  let model = "";
-  let id = "";
-  let usage = null;
-  for await (const ev of upstream.stream()) {
-    if (!ev || ev === "[DONE]") continue;
-    try {
-      const obj = JSON.parse(ev);
-      const normalized = unwrapData(obj);
-      const choice = normalized?.choices?.[0];
-      if (!choice) continue;
-      const delta = choice.delta || {};
-      if (delta.content) content += delta.content;
-      if (delta.reasoning) reasoning += delta.reasoning;
-      if (choice.finish_reason) finishReason = choice.finish_reason;
-      if (normalized.id) id = normalized.id;
-      if (normalized.model) model = normalized.model;
-      if (normalized.usage) usage = normalized.usage;
-    } catch {}
-  }
-  const msg = { role: "assistant", content };
-  if (reasoning) msg.reasoning = reasoning;
-  if (!content && reasoning) { msg.content = reasoning; msg.reasoning_used_as_content = true; }
-  if (track) {
-    const exact = exactOutputTokens(usage);
-    if (exact) track({ tokens: exact, estimated: false });
-    else if (content || reasoning) track({ tokens: Math.ceil((content.length + reasoning.length) / 4), estimated: true });
-  }
-  return {
-    id: id || `gen_${Date.now()}`,
-    object: "chat.completion",
-    created: Math.floor(Date.now() / 1000),
-    model: model || "",
-    choices: [{ index: 0, message: msg, finish_reason: finishReason || "stop", logprobs: null, native_finish_reason: finishReason || "stop" }],
-    usage: usage || { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 },
-  };
-}
-
-export function createChatService({
-  id,
-  baseUrl,
-  chatPath,
-  fetchImpl,
-  dispatcher,
-  authPool,
-  connectTimeoutMs = 30_000,
-} = {}) {
-  const resolvedBase = String(baseUrl).trim().replace(/\/+$/, "");
-  const resolvedChat = chatPath || (String(resolvedBase).includes("/api/v1") ? "/chat/completions" : "/api/v1/chat/completions");
-  const transport = createTransport({ fetchImpl, dispatcher, keepAlive: !!dispatcher, timeoutMs: connectTimeoutMs, retry: {} });
-  const sdk = createSdkDispatch({ id, providerName: id || "cline" });
-
-  async function clineFetch(body, sessionId, allowSdk = false) {
-    const token = await authPool.getAccessToken(body?.model);
-    appendEvent({ type: "cline-account-state", provider: id || "cline", state: "selected", accountId: acctId(authPool.getCurrentAccount()), model: body?.model, pool: poolStat(authPool.getAccounts(), (a) => authPool.accountAvailable(a, body?.model)) });
-    const headers = clineHeaders(sessionId, token);
-    const finalUrl = joinUrl(resolvedBase, resolvedChat);
-    const isStream = body?.stream === true;
-    // SDK 通道（缺省）：仅客户端显式流式；forceStream 聚合与非流式保持原生（避免丢 nonStreamWithContentCheck）。
-    if (allowSdk && sdk.enabled) {
-      const r = await sdk.trySdk({ url: finalUrl, body, headers, fetchImpl });
-      if (r) return r;
-    }
-    return transport.request({ url: finalUrl, headers, body, stream: isStream, timeoutMs: connectTimeoutMs });
-  }
-  // 记账回调：捕获当前账号哈希，成功输出进当前额度周期
-  const track = (model) => ({ tokens, estimated }) => recordOutput({ accountId: acctId(authPool.getCurrentAccount()), model, tokens, estimated }).catch(() => {});
-
-  function isLimitHit(status, bodyText) {
-    if (status === 429) return true;
-    if (status >= 500 && String(bodyText).includes("empty response content")) return true;
-    return false;
-  }
-
-  async function clineFetchWithRetry(body, sessionId, allowSdk = false) {
-    const maxRetries = 4;
-    let lastResp = null;
-    for (let attempt = 0; attempt <= maxRetries; attempt++) {
-      const resp0 = await authPool.enqueue(() => clineFetch(body, sessionId, allowSdk));
-      let resp = resp0;
-      let bodyText = "";
-      if (resp.status !== 200) {
-        try { bodyText = await resp.text(); } catch {}
-        // 错误响应体已在此读走：原生 transport 的 text() 有缓存，而 SDK Response 的 body 一次性，
-        // 重建以便上层/客户端仍能读到错误详情（如 429 的 INFERENCE_CAP_ERROR）。
-        try {
-          resp = new Response(bodyText, { status: resp0.status, statusText: resp0.statusText, headers: new Headers(resp0.headers) });
-        } catch {}
-      }
-      lastResp = resp;
-      const hit = isLimitHit(resp.status, bodyText);
-      if (hit) {
-        const { parseCooldown } = await import("./auth.js");
-        const cooldownMs = parseCooldown(bodyText, resp.status);
-        const cur = authPool.getCurrentAccount();
-        // 免费额度按「账号 × 模型」计：只把该模型挂到这个号上（limits[model]），
-        // 不动账号级 cooldownUntil、不清 accessToken——同号其它模型照常可用。
-        authPool.markLimit(cur, body?.model, cooldownMs);
-        const limitedModel = limitModelOf(bodyText);
-        if (limitedModel && limitedModel !== body?.model) authPool.markLimit(cur, limitedModel, cooldownMs);
-        const cycle = await recordLimit({ accountId: acctId(cur), model: body?.model, reason: resp.status === 429 ? "daily_limit" : "empty_response", status: resp.status, cooldownMs });
-        const pool = authPool.getAccounts();
-        const ready = (a) => authPool.accountAvailable(a, body?.model);
-        const hasOther = pool.some(ready);
-        // 限流是「账号 × 模型」维度：429 只说明当前账号在这个模型上额度用尽，
-        // 换模型（如 muse-spark → deepseek）往往仍可用，故日志必须点名模型。
-        appendEvent({
-          type: "cline-account-state",
-          provider: id || "cline",
-          state: hasOther ? "switch" : "pool-exhausted",
-          accountId: acctId(cur),
-          model: body?.model,
-          limitedModel: limitedModel || undefined,
-          scope: "model",
-          reason: resp.status === 429 ? "daily_limit" : "empty_response",
-          status: resp.status,
-          cooldownMs,
-          pool: poolStat(pool, ready),
-          cycleOutputTokens: cycle?.cycleOutputTokens ?? 0,
-          cycleCount: cycle?.cycleCount ?? 0,
-        });
-        if (!hasOther) return resp;
-        await sleep(500 + Math.floor(Math.random() * 500));
-        continue;
-      }
-      if (resp.ok) return resp;
-      return resp;
-    }
-    return lastResp;
-  }
-
-  async function nonStreamWithContentCheck(body, sessionId, firstResp) {
-    const maxAttempts = 3;
-    let lastData = null;
-    let resp = firstResp;
-    for (let attempt = 0; attempt < maxAttempts; attempt++) {
-      if (!resp) resp = await clineFetchWithRetry(body, sessionId);
-      if (!resp.ok) {
-        const errText = await resp.text().catch(() => "");
-        const hdrs = new Headers(resp.headers);
-        const errBody = JSON.stringify({ error: { message: errText.slice(0, 500), type: "api_error" } });
-        return { error: new Response(errBody, { status: resp.status, headers: hdrs }) };
-      }
-      const ct = resp.headers.get("content-type") || "";
-      let normalized = null;
-      if (ct.includes("text/event-stream")) normalized = await streamToNonStream(resp, track(body?.model));
-      else {
-        const raw = await resp.json().catch(() => null);
-        if (raw) normalized = unwrapData(raw);
-      }
-      if (!normalized) return { error: new Response(JSON.stringify({ error: { message: "upstream returned non-SSE body", type: "api_error" } }), { status: 502 }) };
-      lastData = normalized;
-      const msg = normalized?.choices?.[0]?.message || {};
-      const content = String(msg.content || "").trim();
-      const reasoning = String(msg.reasoning || "").trim();
-      const isFallback = msg.reasoning_used_as_content === true;
-      if (content && !isFallback) return { data: normalized };
-      if (reasoning || isFallback) {
-        const cur = authPool.getCurrentAccount();
-        // 空响应是「模型 × 通道」形态问题（deepseek 非流式特性），同样只挡该模型。
-        authPool.markLimit(cur, body?.model, 30 * 1000);
-        appendEvent({ type: "cline-account-state", provider: id || "cline", state: "cooldown", accountId: acctId(cur), model: body?.model, reason: "empty_response", cooldownMs: 30000, scope: "model", pool: poolStat(authPool.getAccounts(), (a) => authPool.accountAvailable(a, body?.model)) });
-        await sleep(300 + Math.floor(Math.random() * 300));
-        resp = null;
-        continue;
-      }
-      await sleep(300 + Math.floor(Math.random() * 300));
-      resp = null;
-    }
-    return { data: lastData };
-  }
-
-  async function runChat(body, ring, sourceKey) {
-    const model = body?.model || "deepseek/deepseek-v4-flash";
-    const sessionId = genSessionId();
-    const isStream = body?.stream === true;
-    const upstreamModel = stripProviderPrefix(model, id);
-    // token 上限对标 dsh-cline-pass（DEFAULT_MAX_TOKENS=32000）：客户端显式 max 优先，
-    // 缺省给 32000 而非 4096——旧 4096 默认会在长文处 finish=length 拦腰截断（2026-09-23 实测）。
-    // reasoning_effort 缺省不发（dsh 同款：调用方自选 none→max，不替上游做 high 假设；
-    // 高推理+小预算双挤是超大上下文秒回超短答的推手之一）；max_completion_tokens 双写保留
-    //（只发 max_tokens 会被部分通道拒，历史教训）。
-    const tokLimit = body?.max_tokens || body?.max_completion_tokens || 32000;
-    const upstreamBody = {
-      model: upstreamModel,
-      max_tokens: tokLimit,
-      max_completion_tokens: tokLimit,
-      session_id: sessionId,
-      messages: body?.messages || [],
-    };
-    const effort = body?.reasoning_effort || body?.reasoningEffort;
-    if (effort) upstreamBody.reasoning_effort = effort;
-    // 部分免费通道（deepseek 家族，含 cline-free/deepseek-*）原生非流式不可靠
-    //（500 empty response）：这类模型在内部走 stream+聚合成 JSON，对外仍按请求方
-    // stream 标志返回（true 直透，false 聚合）。其它模型完全尊重请求方，不强制。
-    const forceStream = !isStream && String(upstreamModel).toLowerCase().includes("deepseek");
-    if (isStream || forceStream) upstreamBody.stream = true;
-    for (const k of ["temperature", "top_p", "tools", "tool_choice", "stop", "presence_penalty", "frequency_penalty", "response_format", "user", "n", "seed"]) {
-      if (body[k] !== undefined) upstreamBody[k] = body[k];
-    }
-    // Gemini 3.x Flash 经 Vertex 的 maxOutputTokens 上限 65536：客户端（PI-Desktop 默认
-    // 128000）原样透传即 400 invalid argument（网关 SSE 内 detail：
-    // "Unable to submit request because it has a maxOutputTokens va..."）。
-    // 2026-09-25 直连实测：max 128000 必 400，65536/32000/500 全 200；
-    // reasoning_effort=high 单独不触发，仅夹上限。只夹 gemini，其他通道不动。
-    if (/gemini/i.test(String(upstreamModel || "")) && tokLimit > 65536) {
-      upstreamBody.max_tokens = 65536;
-      upstreamBody.max_completion_tokens = 65536;
-    }
-    for (let netAttempt = 0; netAttempt < 3; netAttempt++) {
-      try {
-        const resp = await clineFetchWithRetry(upstreamBody, sessionId, isStream);
-        if (!resp) throw new Error("empty response");
-        if (!resp.ok) return resp;
-        if (isStream) {
-          // 旁路观测：把本次选中的账号哈希与裸 model 挂到透传 Response 上，供 relay 记账用。
-          // 转发链路不读这两个字段；记账失败也不影响返回。见 .agents/notes/implemented/feature/2026-09-24-cline-usage-dualwindow.md
-          try { resp.clineAccountId = acctId(authPool.getCurrentAccount()); resp.clineModel = model; } catch {}
-          return resp;
-        }
-        if (forceStream) {
-          const ret = await nonStreamWithContentCheck(upstreamBody, sessionId, resp);
-          if (ret.error) return ret.error;
-          ret.data.model = model;
-          const hdrs = new Headers({ "Content-Type": "application/json" });
-          return new Response(JSON.stringify(ret.data), { status: 200, headers: hdrs });
-        }
-        const raw = await resp.json().catch(() => null);
-        if (!raw || typeof raw !== "object") {
-          if (!raw) return resp;
-          return new Response(JSON.stringify({ error: { message: "upstream returned non-JSON body", type: "api_error" } }), { status: 502, headers: { "Content-Type": "application/json" } });
-        }
-        const normalized = unwrapData(raw);
-        normalized.model = model;
-        // 非流式终点：正文估算记账（上游有 usage 时 streamToNonStream 路径才拿得到精确值；此处 raw 无流）
-        const m2 = normalized?.choices?.[0]?.message || {};
-        const outLen2 = String(m2.content || "").length + String(m2.reasoning || "").length;
-        if (outLen2 > 0) track(model)({ tokens: Math.ceil(outLen2 / 4), estimated: true });
-        return new Response(JSON.stringify(normalized), { status: 200, headers: { "Content-Type": "application/json" } });
-      } catch (err) {
-        if (netAttempt < 2 && String(err?.message || "").toLowerCase().includes("timed out")) { await sleep(300); continue; }
-        throw err;
-      }
-    }
-    throw new Error("cline chat failed after retries");
-  }
-
-  return { runChat, streamToNonStream, _clineFetch: clineFetch };
-}
+import { joinUrl, sleep } from "../base.js";
+import { appendEvent } from "../../logs.js";
+import { createHash } from "node:crypto";
+import { clineHeaders } from "./headers.js";
+import { createTransport } from "../../transport/index.js";
+import { normalizeProviderId } from "../model-id.js";
+import { recordLimit, recordOutput, exactOutputTokens } from "./usage.js";
+import { createSdkDispatch } from "../../upstream-engine/sdk/dispatch.js";
+
+function genSessionId() { return `sess_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`; }
+
+// 只记哈希标识，不落邮箱 / refreshToken（events.log 可读不可泄）
+function acctId(account) {
+  const seed = String(account?.refreshToken || "");
+  if (!seed) return "slot-unknown";
+  return `acct_${createHash("sha256").update(seed).digest("hex").slice(0, 8)}`;
+}
+// 池子快照：ready/cooling 按「该模型」口径统计（ready=对该模型可用的号数）。
+// isReady 由调用方注入 authPool.accountAvailable(a, model)，避免与 auth 逻辑分叉。
+function poolStat(pool, isReady = () => true) {
+  let cooling = 0;
+  let dead = 0;
+  for (const a of pool || []) {
+    if (a.dead) dead++;
+    else if (!isReady(a)) cooling++;
+  }
+  const total = (pool || []).length;
+  return { total, cooling, dead, ready: total - cooling - dead };
+}
+
+// 429 错误体自带权威模型名（"...Daily free limit reached on model meta/muse-spark-1.3-contributor"）。
+function limitModelOf(bodyText) {
+  // 模型名可含 / . - _ : 字母数字（如 meta/muse-spark-1.3-contributor），
+  // 遇空格/引号/逗号/句号结尾或 "Try again" 前的空白即停。
+  const m = String(bodyText || "").match(/on model\s+([A-Za-z0-9/._:-]+?)(?=[\s"',]|\.\s|$)/i);
+  return m ? m[1] : "";
+}
+
+// stripProviderPrefix 只剥「本供应商」前缀（cline/<裸 id> → 裸 id），非本前缀（含上游自带多段 id
+// 如 meta/vendor/x）原样透传，不误削首段（旧实现按段数剥曾是潜在 bug）。
+function stripProviderPrefix(model, providerId) {
+  const s = String(model || "");
+  const i = s.indexOf("/");
+  if (i <= 0) return s;
+  const low = (v) => normalizeProviderId(String(v || "").toLowerCase());
+  return low(s.slice(0, i)) === low(providerId) ? s.slice(i + 1) : s;
+}
+function unwrapData(obj) {
+  if (obj && obj.data && typeof obj.data === "object") {
+    const d = obj.data;
+    if (d.choices || d.id || d.usage || d.output) return d;
+  }
+  return obj;
+}
+
+async function streamToNonStream(upstream, track = null) {
+  let content = "";
+  let reasoning = "";
+  let finishReason = null;
+  let model = "";
+  let id = "";
+  let usage = null;
+  for await (const ev of upstream.stream()) {
+    if (!ev || ev === "[DONE]") continue;
+    try {
+      const obj = JSON.parse(ev);
+      const normalized = unwrapData(obj);
+      const choice = normalized?.choices?.[0];
+      if (!choice) continue;
+      const delta = choice.delta || {};
+      if (delta.content) content += delta.content;
+      if (delta.reasoning) reasoning += delta.reasoning;
+      if (choice.finish_reason) finishReason = choice.finish_reason;
+      if (normalized.id) id = normalized.id;
+      if (normalized.model) model = normalized.model;
+      if (normalized.usage) usage = normalized.usage;
+    } catch {}
+  }
+  const msg = { role: "assistant", content };
+  if (reasoning) msg.reasoning = reasoning;
+  if (!content && reasoning) { msg.content = reasoning; msg.reasoning_used_as_content = true; }
+  if (track) {
+    const exact = exactOutputTokens(usage);
+    if (exact) track({ tokens: exact, estimated: false });
+    else if (content || reasoning) track({ tokens: Math.ceil((content.length + reasoning.length) / 4), estimated: true });
+  }
+  return {
+    id: id || `gen_${Date.now()}`,
+    object: "chat.completion",
+    created: Math.floor(Date.now() / 1000),
+    model: model || "",
+    choices: [{ index: 0, message: msg, finish_reason: finishReason || "stop", logprobs: null, native_finish_reason: finishReason || "stop" }],
+    usage: usage || { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 },
+  };
+}
+
+export function createChatService({
+  id,
+  baseUrl,
+  chatPath,
+  fetchImpl,
+  dispatcher,
+  authPool,
+  connectTimeoutMs = 30_000,
+} = {}) {
+  const resolvedBase = String(baseUrl).trim().replace(/\/+$/, "");
+  const resolvedChat = chatPath || (String(resolvedBase).includes("/api/v1") ? "/chat/completions" : "/api/v1/chat/completions");
+  const transport = createTransport({ fetchImpl, dispatcher, keepAlive: !!dispatcher, timeoutMs: connectTimeoutMs, retry: {} });
+  const sdk = createSdkDispatch({ id, providerName: id || "cline" });
+
+  async function clineFetch(body, sessionId, allowSdk = false) {
+    const token = await authPool.getAccessToken(body?.model);
+    appendEvent({ type: "cline-account-state", provider: id || "cline", state: "selected", accountId: acctId(authPool.getCurrentAccount()), model: body?.model, pool: poolStat(authPool.getAccounts(), (a) => authPool.accountAvailable(a, body?.model)) });
+    const headers = clineHeaders(sessionId, token);
+    const finalUrl = joinUrl(resolvedBase, resolvedChat);
+    const isStream = body?.stream === true;
+    // SDK 通道（缺省）：仅客户端显式流式；forceStream 聚合与非流式保持原生（避免丢 nonStreamWithContentCheck）。
+    if (allowSdk && sdk.enabled) {
+      const r = await sdk.trySdk({ url: finalUrl, body, headers, fetchImpl });
+      if (r) return r;
+    }
+    return transport.request({ url: finalUrl, headers, body, stream: isStream, timeoutMs: connectTimeoutMs });
+  }
+  // 记账回调：捕获当前账号哈希，成功输出进当前额度周期
+  const track = (model) => ({ tokens, estimated }) => recordOutput({ accountId: acctId(authPool.getCurrentAccount()), model, tokens, estimated }).catch(() => {});
+
+  function isLimitHit(status, bodyText) {
+    if (status === 429) return true;
+    if (status >= 500 && String(bodyText).includes("empty response content")) return true;
+    return false;
+  }
+
+  async function clineFetchWithRetry(body, sessionId, allowSdk = false) {
+    const maxRetries = 4;
+    let lastResp = null;
+    for (let attempt = 0; attempt <= maxRetries; attempt++) {
+      const resp0 = await authPool.enqueue(() => clineFetch(body, sessionId, allowSdk));
+      let resp = resp0;
+      let bodyText = "";
+      if (resp.status !== 200) {
+        try { bodyText = await resp.text(); } catch {}
+        // 错误响应体已在此读走：原生 transport 的 text() 有缓存，而 SDK Response 的 body 一次性，
+        // 重建以便上层/客户端仍能读到错误详情（如 429 的 INFERENCE_CAP_ERROR）。
+        try {
+          resp = new Response(bodyText, { status: resp0.status, statusText: resp0.statusText, headers: new Headers(resp0.headers) });
+        } catch {}
+      }
+      lastResp = resp;
+      const hit = isLimitHit(resp.status, bodyText);
+      if (hit) {
+        const { parseCooldown } = await import("./auth.js");
+        const cooldownMs = parseCooldown(bodyText, resp.status);
+        const cur = authPool.getCurrentAccount();
+        // 免费额度按「账号 × 模型」计：只把该模型挂到这个号上（limits[model]），
+        // 不动账号级 cooldownUntil、不清 accessToken——同号其它模型照常可用。
+        authPool.markLimit(cur, body?.model, cooldownMs);
+        const limitedModel = limitModelOf(bodyText);
+        if (limitedModel && limitedModel !== body?.model) authPool.markLimit(cur, limitedModel, cooldownMs);
+        const cycle = await recordLimit({ accountId: acctId(cur), model: body?.model, reason: resp.status === 429 ? "daily_limit" : "empty_response", status: resp.status, cooldownMs });
+        const pool = authPool.getAccounts();
+        const ready = (a) => authPool.accountAvailable(a, body?.model);
+        const hasOther = pool.some(ready);
+        // 限流是「账号 × 模型」维度：429 只说明当前账号在这个模型上额度用尽，
+        // 换模型（如 muse-spark → deepseek）往往仍可用，故日志必须点名模型。
+        appendEvent({
+          type: "cline-account-state",
+          provider: id || "cline",
+          state: hasOther ? "switch" : "pool-exhausted",
+          accountId: acctId(cur),
+          model: body?.model,
+          limitedModel: limitedModel || undefined,
+          scope: "model",
+          reason: resp.status === 429 ? "daily_limit" : "empty_response",
+          status: resp.status,
+          cooldownMs,
+          pool: poolStat(pool, ready),
+          cycleOutputTokens: cycle?.cycleOutputTokens ?? 0,
+          cycleCount: cycle?.cycleCount ?? 0,
+        });
+        if (!hasOther) return resp;
+        await sleep(500 + Math.floor(Math.random() * 500));
+        continue;
+      }
+      if (resp.ok) return resp;
+      return resp;
+    }
+    return lastResp;
+  }
+
+  async function nonStreamWithContentCheck(body, sessionId, firstResp) {
+    const maxAttempts = 3;
+    let lastData = null;
+    let resp = firstResp;
+    for (let attempt = 0; attempt < maxAttempts; attempt++) {
+      if (!resp) resp = await clineFetchWithRetry(body, sessionId);
+      if (!resp.ok) {
+        const errText = await resp.text().catch(() => "");
+        const hdrs = new Headers(resp.headers);
+        const errBody = JSON.stringify({ error: { message: errText.slice(0, 500), type: "api_error" } });
+        return { error: new Response(errBody, { status: resp.status, headers: hdrs }) };
+      }
+      const ct = resp.headers.get("content-type") || "";
+      let normalized = null;
+      if (ct.includes("text/event-stream")) normalized = await streamToNonStream(resp, track(body?.model));
+      else {
+        const raw = await resp.json().catch(() => null);
+        if (raw) normalized = unwrapData(raw);
+      }
+      if (!normalized) return { error: new Response(JSON.stringify({ error: { message: "upstream returned non-SSE body", type: "api_error" } }), { status: 502 }) };
+      lastData = normalized;
+      const msg = normalized?.choices?.[0]?.message || {};
+      const content = String(msg.content || "").trim();
+      const reasoning = String(msg.reasoning || "").trim();
+      const isFallback = msg.reasoning_used_as_content === true;
+      if (content && !isFallback) return { data: normalized };
+      if (reasoning || isFallback) {
+        const cur = authPool.getCurrentAccount();
+        // 空响应是「模型 × 通道」形态问题（deepseek 非流式特性），同样只挡该模型。
+        authPool.markLimit(cur, body?.model, 30 * 1000);
+        appendEvent({ type: "cline-account-state", provider: id || "cline", state: "cooldown", accountId: acctId(cur), model: body?.model, reason: "empty_response", cooldownMs: 30000, scope: "model", pool: poolStat(authPool.getAccounts(), (a) => authPool.accountAvailable(a, body?.model)) });
+        await sleep(300 + Math.floor(Math.random() * 300));
+        resp = null;
+        continue;
+      }
+      await sleep(300 + Math.floor(Math.random() * 300));
+      resp = null;
+    }
+    return { data: lastData };
+  }
+
+  async function runChat(body, ring, sourceKey) {
+    const model = body?.model || "deepseek/deepseek-v4-flash";
+    const sessionId = genSessionId();
+    const isStream = body?.stream === true;
+    const upstreamModel = stripProviderPrefix(model, id);
+    // token 上限对标 dsh-cline-pass（DEFAULT_MAX_TOKENS=32000）：客户端显式 max 优先，
+    // 缺省给 32000 而非 4096——旧 4096 默认会在长文处 finish=length 拦腰截断（2026-09-23 实测）。
+    // reasoning_effort 缺省不发（dsh 同款：调用方自选 none→max，不替上游做 high 假设；
+    // 高推理+小预算双挤是超大上下文秒回超短答的推手之一）；max_completion_tokens 双写保留
+    //（只发 max_tokens 会被部分通道拒，历史教训）。
+    const tokLimit = body?.max_tokens || body?.max_completion_tokens || 32000;
+    const upstreamBody = {
+      model: upstreamModel,
+      max_tokens: tokLimit,
+      max_completion_tokens: tokLimit,
+      session_id: sessionId,
+      messages: body?.messages || [],
+    };
+    const effort = body?.reasoning_effort || body?.reasoningEffort;
+    if (effort) upstreamBody.reasoning_effort = effort;
+    // 部分免费通道（deepseek 家族，含 cline-free/deepseek-*）原生非流式不可靠
+    //（500 empty response）：这类模型在内部走 stream+聚合成 JSON，对外仍按请求方
+    // stream 标志返回（true 直透，false 聚合）。其它模型完全尊重请求方，不强制。
+    const forceStream = !isStream && String(upstreamModel).toLowerCase().includes("deepseek");
+    if (isStream || forceStream) upstreamBody.stream = true;
+    for (const k of ["temperature", "top_p", "tools", "tool_choice", "stop", "presence_penalty", "frequency_penalty", "response_format", "user", "n", "seed"]) {
+      if (body[k] !== undefined) upstreamBody[k] = body[k];
+    }
+    // Gemini 3.x Flash 经 Vertex 的 maxOutputTokens 上限 65536：客户端（PI-Desktop 默认
+    // 128000）原样透传即 400 invalid argument（网关 SSE 内 detail：
+    // "Unable to submit request because it has a maxOutputTokens va..."）。
+    // 2026-09-25 直连实测：max 128000 必 400，65536/32000/500 全 200；
+    // reasoning_effort=high 单独不触发，仅夹上限。只夹 gemini，其他通道不动。
+    if (/gemini/i.test(String(upstreamModel || "")) && tokLimit > 65536) {
+      upstreamBody.max_tokens = 65536;
+      upstreamBody.max_completion_tokens = 65536;
+    }
+    for (let netAttempt = 0; netAttempt < 3; netAttempt++) {
+      try {
+        const resp = await clineFetchWithRetry(upstreamBody, sessionId, isStream);
+        if (!resp) throw new Error("empty response");
+        if (!resp.ok) return resp;
+        if (isStream) {
+          // 旁路观测：把本次选中的账号哈希与裸 model 挂到透传 Response 上，供 relay 记账用。
+          // 转发链路不读这两个字段；记账失败也不影响返回。见 .agents/notes/implemented/feature/2026-09-24-cline-usage-dualwindow.md
+          try { resp.clineAccountId = acctId(authPool.getCurrentAccount()); resp.clineModel = model; } catch {}
+          return resp;
+        }
+        if (forceStream) {
+          const ret = await nonStreamWithContentCheck(upstreamBody, sessionId, resp);
+          if (ret.error) return ret.error;
+          ret.data.model = model;
+          const hdrs = new Headers({ "Content-Type": "application/json" });
+          return new Response(JSON.stringify(ret.data), { status: 200, headers: hdrs });
+        }
+        const raw = await resp.json().catch(() => null);
+        if (!raw || typeof raw !== "object") {
+          if (!raw) return resp;
+          return new Response(JSON.stringify({ error: { message: "upstream returned non-JSON body", type: "api_error" } }), { status: 502, headers: { "Content-Type": "application/json" } });
+        }
+        const normalized = unwrapData(raw);
+        normalized.model = model;
+        // 非流式终点：正文估算记账（上游有 usage 时 streamToNonStream 路径才拿得到精确值；此处 raw 无流）
+        const m2 = normalized?.choices?.[0]?.message || {};
+        const outLen2 = String(m2.content || "").length + String(m2.reasoning || "").length;
+        if (outLen2 > 0) track(model)({ tokens: Math.ceil(outLen2 / 4), estimated: true });
+        return new Response(JSON.stringify(normalized), { status: 200, headers: { "Content-Type": "application/json" } });
+      } catch (err) {
+        if (netAttempt < 2 && String(err?.message || "").toLowerCase().includes("timed out")) { await sleep(300); continue; }
+        throw err;
+      }
+    }
+    throw new Error("cline chat failed after retries");
+  }
+
+  return { runChat, streamToNonStream, _clineFetch: clineFetch };
+}

package/src/providers/cline.js

@@ -1,2 +1,2 @@
-// Shim — 保持历史导入路径兼容，真实实现下沉至 ./cline/ 深模块
-export * from "./cline/index.js";
+// Shim — 保持历史导入路径兼容，真实实现下沉至 ./cline/ 深模块
+export * from "./cline/index.js";