mslxdff 0.1.199 → 0.1.203

package/src/anthropic/translate.js

@@ -1,287 +1,287 @@
-/**
- * Anthropic Messages 出站翻译层：chat → Anthropic message / SSE 具名事件（/v1/messages 外壳，ADR-0047）。
- * 纯函数，无网络、无副作用；出站形状依据 code.claude.com/docs/en/llm-gateway-protocol
- * 与参考实现 lamdt1/ms-copilot365-2api/app/formatters/anthropic_sse.py。
- * 分档（ADR-0055，为守本文件 ≤20KB 体积硬门；对外 import 路径全走本文件、调用方零改动）：
- *  · 请求侧 Anthropic → chat 在 ./request-body.js；
- *  · 消息/工具 id 与协议错误体在 ./shared.js；
- *  · token 估算的单一来源在 ../text-tokens.js（上游未报 usage 时由它兜底）。
- *
- * 与 src/responses/translate.js 的三点关键差异：
- *  1) Anthropic 的 content block **只封口不复用**（Responses 侧 ensureTextItem 会复用已开项）；
- *  2) 事件是具名的（`event: <type>` + `data:`），且**没有 `data: [DONE]`**，收场只有
- *     `message_delta`+`message_stop`（成功）或一帧 `error`（失败）——见 grill-decisions.md Q2；
- *  3) 请求侧**保留思考档位**、**必须主动丢** output_config/context_management 与消息块内 thinking
- *     （块内 thinking 我们无法签发 signature，回传必被上游拒）——细则见 ./request-body.js 头（ADR-0047）。
- */
-import { chunkToString } from "../responses/translate.js";
-import { createTokenCounter } from "../text-tokens.js";
-import { newMessageId, newToolUseId } from "./shared.js";
-
-// re-export：拆档不改对外接口，调用方与测试仍从本文件取这些符号
-export { messagesToChatBody, toolResultTextOf } from "./request-body.js";
-export { newMessageId, newToolUseId, anthropicError } from "./shared.js";
-
-// ---------- usage / stop_reason ----------
-
-/**
- * chat 口径 usage → Anthropic 口径。
- * Anthropic 的 `input_tokens` **不含**缓存读命中的部分（两者分开计费），而上游
- * `prompt_tokens` 是含缓存的总输入 —— 故命中数要从 input 里扣掉，否则客户端把同一段
- * 前缀算两遍（读数虚高）。
- * `fallback()` 只在**上游整条未报**（缺字段，或 prompt/output/cache_read/cache_creation 四项全为 0）时才用：
- * 真实模型对非空 prompt 不可能报 0，照抄 0 会让客户端的上下文计度停在原地、永不压实（ADR-0055 决策 3）。
- */
-export function toAnthropicUsage(u = {}, fallback = null) {
-  const num = (v) => (Number.isFinite(Number(v)) ? Number(v) : 0);
-  const src = u && typeof u === "object" ? u : {};
-  const prompt = num(src.prompt_tokens ?? src.input_tokens);
-  const output = num(src.completion_tokens ?? src.output_tokens);
-  const cacheRead = num(src.cache_read_input_tokens ?? src.prompt_tokens_details?.cached_tokens ?? src.input_tokens_details?.cached_tokens);
-  const cacheWrite = num(src.cache_creation_input_tokens ?? src.prompt_tokens_details?.cache_creation_input_tokens);
-  const usage = {
-    input_tokens: Math.max(0, prompt - cacheRead - cacheWrite),
-    output_tokens: output,
-    cache_read_input_tokens: cacheRead,
-    cache_creation_input_tokens: cacheWrite,
-  };
-  // cache 非零也算「报了」：SDK 通道（upstream-engine/sdk/sse.js 的 usageToOpenAI）会输出
-  // `{prompt_tokens:0, …, prompt_tokens_details:{cached_tokens:N}}` 这种形状，若当未报整条覆盖，
-  // 客户端「input + cache_read」会把同一段前缀算两遍（正是上面扣减要防的事，评审 P2）。
-  if (fallback && !(prompt > 0 || output > 0 || cacheRead > 0 || cacheWrite > 0)) return { ...usage, ...(fallback() || {}) };
-  return usage;
-}
-
-function toStopReason(finish, hasToolUse) {
-  if (hasToolUse || finish === "tool_calls" || finish === "function_call") return "tool_use";
-  if (finish === "length") return "max_tokens";
-  return "end_turn";
-}
-
-// ---------- 非流式响应：chat JSON → Anthropic message ----------
-
-function parseToolInput(args) {
-  const raw = String(args || "");
-  if (!raw.trim()) return {};
-  try {
-    const v = JSON.parse(raw);
-    return v && typeof v === "object" && !Array.isArray(v) ? v : { value: v };
-  } catch {
-    return { raw }; // 上游吐出发散 JSON：退化成 raw，不整轮失败
-  }
-}
-
-/**
- * 上游 chat 的 `message.content` 既可能是字符串，也可能是部件数组（`[{type:"text",text:…}, {type:"image_url",…}]`，
- * 多模态上游与部分聚合网关都这么回）。统一抽成正文文本；图片部件在本外壳里无法映射成 Anthropic 的 image 块，只能弃（见 ADR-0047 边界 11）。
- */
-export function textFromChatContent(c) {
-  if (typeof c === "string") return c;
-  if (!Array.isArray(c)) return "";
-  let t = "";
-  for (const p of c) {
-    if (p && typeof p === "object" && typeof p.text === "string") t += p.text;
-  }
-  return t;
-}
-
-export function chatJsonToAnthropic(chatJson = {}, model = "", opts = {}) {
-  const choice = chatJson.choices?.[0] || {};
-  const message = choice.message || {};
-  const content = [];
-  const text = textFromChatContent(message.content);
-  if (text) content.push({ type: "text", text });
-  for (const tc of message.tool_calls || []) {
-    content.push({ type: "tool_use", id: String(tc.id || newToolUseId()), name: String(tc.function?.name || ""), input: parseToolInput(tc.function?.arguments) });
-  }
-  return {
-    id: typeof chatJson.id === "string" && chatJson.id.startsWith("msg_") ? chatJson.id : newMessageId(),
-    type: "message",
-    role: "assistant",
-    model: model || chatJson.model || "",
-    content: content.length ? content : [{ type: "text", text: "" }],
-    stop_reason: toStopReason(choice.finish_reason, content.some((b) => b.type === "tool_use")),
-    stop_sequence: null,
-    usage: toAnthropicUsage(chatJson.usage, () => fallbackUsage(opts, content)),
-  };
-}
-
-// 上游整条未报 usage 时的兜底（ADR-0055 决策 3）：请求侧用注入的 estInput，生成侧按已产出的块估。
-function fallbackUsage(opts = {}, content = []) {
-  const counter = createTokenCounter();
-  for (const b of content) {
-    if (b?.type === "text") counter.add(b.text || "");
-    else if (b?.type === "tool_use") counter.add(JSON.stringify(b.input ?? {})); // 只计参数：与流式 pushTool 同口径（name 从不进 gen）
-  }
-  const n = typeof opts.estInput === "function" ? Number(opts.estInput()) : 0;
-  return { input_tokens: Number.isFinite(n) && n > 0 ? Math.floor(n) : 0, output_tokens: counter.tokens() };
-}
-
-// ---------- 流式响应：chat SSE chunk → Anthropic 具名事件 ----------
-
-// thinking 块必须在 content_block_stop **之前**交一帧 signature_delta，缺它 Claude Code 把整块静默丢弃
-// （routatic/proxy#51 真机复现、PR#133 修）。下面是**身份占位**：不含思考原文、不冒充密码学签名，理由与风险见 ADR-0052。
-const SIG_PREFIX = "mslxdff:thinking:v1:";
-function thinkingSignature(messageId, index) {
-  return Buffer.from(`${SIG_PREFIX}${messageId}:${index}`, "utf8").toString("base64url");
-}
-
-/**
- * @param model 回显给客户端的模型名
- * @param opts  { id?, thinking?: boolean }  thinking=true 才把 reasoning_content 发成 thinking 块
- */
-export function createAnthropicChunkTranslator(model = "", opts = {}) {
-  const id = opts.id || newMessageId();
-  const emitThinking = opts.thinking === true;
-  const created = Math.floor(Date.now() / 1000);
-  const ev = (type, extra = {}) => ({ type, ...extra });
-
-  // 请求侧 token 估算（ADR-0055 决策 3/5）：路由注入 `opts.estInput()`，memoize —— `message_start` 必须在
-  // 首块前就带 input_tokens，全文只能扫一遍；上游后报实测 usage 时估算只留在首帧，不进 message_delta。
-  let estCached = null;
-  const estInput = () => {
-    if (estCached == null) {
-      const n = typeof opts.estInput === "function" ? Number(opts.estInput()) : 0;
-      estCached = Number.isFinite(n) && n > 0 ? Math.floor(n) : 0;
-    }
-    return estCached;
-  };
-  const gen = createTokenCounter(); // 生成侧累计（text + thinking + tool 参数增量）
-
-  let buf = "";
-  let nextIndex = 0;
-  let open = null; // 当前未封口块：{ index, kind: "text"|"thinking"|"tool_use", text? }
-  const tools = new Map(); // 上游 tc.index → { id, name, args, announced, index }
-  let sawToolUse = false;
-  let lastFinish = null;
-  let lastUsage = null;
-  const dbg = { chunks: 0, textChars: 0, toolDeltas: 0, thinkingChars: 0, skippedLines: 0, jsonFails: 0, lateToolDeltas: 0, blocks: 0 };
-
-  function closeOpen(out) {
-    if (!open) return;
-    if (open.kind === "thinking") out.push(ev("content_block_delta", { index: open.index, delta: { type: "signature_delta", signature: thinkingSignature(id, open.index) } }));
-    out.push(ev("content_block_stop", { index: open.index }));
-    open = null;
-  }
-  function openBlock(kind, contentBlock, out) {
-    closeOpen(out);
-    open = { index: nextIndex++, kind, text: "" };
-    dbg.blocks = open.index + 1;
-    out.push(ev("content_block_start", { index: open.index, content_block: contentBlock }));
-    return open;
-  }
-  function pushText(delta) {
-    const out = [];
-    if (!open || open.kind !== "text") openBlock("text", { type: "text", text: "" }, out);
-    open.text += delta;
-    gen.add(delta); // 生成侧只累计字数、末帧一次套公式：逐帧套会把参照实现的 +1 下限乘上帧数
-    out.push(ev("content_block_delta", { index: open.index, delta: { type: "text_delta", text: delta } }));
-    return out;
-  }
-  function pushThinking(delta) {
-    const out = [];
-    if (!open || open.kind !== "thinking") openBlock("thinking", { type: "thinking", thinking: "", signature: "" }, out);
-    open.text += delta;
-    gen.add(delta); // thinking 也按 output 计费（ADR-0055 grill Q6）
-    out.push(ev("content_block_delta", { index: open.index, delta: { type: "thinking_delta", thinking: delta } }));
-    return out;
-  }
-  function announceTool(t, out) {
-    t.announced = true;
-    sawToolUse = true;
-    // 统一走 openBlock：自动封掉当前 text/thinking/上一个工具块，本块成为新的待封 open
-    openBlock("tool_use", { type: "tool_use", id: t.id, name: t.name, input: {} }, out);
-    t.index = open.index;
-    // 增量早于 id/name 到达时先攒着，announce 后补发（事件顺序仍合规）
-    if (t.args) out.push(ev("content_block_delta", { index: t.index, delta: { type: "input_json_delta", partial_json: t.args } }));
-  }
-  function pushTool(tc) {
-    const slot = Number(tc?.index ?? 0);
-    const out = [];
-    let t = tools.get(slot);
-    if (!t) { t = { id: "", name: "", args: "", announced: false, index: null }; tools.set(slot, t); }
-    if (tc?.id) t.id = String(tc.id);
-    if (tc?.function?.name) t.name += String(tc.function.name);
-    const frag = tc?.function?.arguments ? String(tc.function.arguments) : "";
-    if (frag) { dbg.toolDeltas++; t.args += frag; gen.add(frag); }
-    if (!t.announced && t.id && t.name) announceTool(t, out);
-    else if (t.announced && frag && open && open.index === t.index) out.push(ev("content_block_delta", { index: t.index, delta: { type: "input_json_delta", partial_json: frag } }));
-    else if (t.announced && frag) dbg.lateToolDeltas++; // 块已封口后才来的参数增量：发出即违规帧，丢弃但上面已计入 gen（output_tokens 取计费口径：模型确实生成了这些字节）
-    return out;
-  }
-
-  function begin() {
-    return [ev("message_start", { message: { id, type: "message", role: "assistant", model, content: [], stop_reason: null, stop_sequence: null, usage: { input_tokens: estInput(), output_tokens: 0, cache_read_input_tokens: 0, cache_creation_input_tokens: 0 } } })];
-  }
-
-  function push(data) {
-    buf += chunkToString(data);
-    const out = [];
-    let nl;
-    while ((nl = buf.indexOf("\n")) >= 0) {
-      const line = buf.slice(0, nl).trim();
-      buf = buf.slice(nl + 1);
-      // 注释帧/keepalive（`:` 开头）与上游 `[DONE]` 都不进 Anthropic 事件序列（grill Q6）
-      if (!line || line.startsWith(":")) continue;
-      if (!line.startsWith("data:")) { dbg.skippedLines++; continue; }
-      const payload = line.slice(5).trim();
-      if (!payload || payload === "[DONE]") continue;
-      let chunk;
-      try { chunk = JSON.parse(payload); } catch { dbg.jsonFails++; continue; }
-      dbg.chunks++;
-      const choice = chunk.choices?.[0] || {};
-      const delta = choice.delta || {};
-      if (choice.finish_reason) lastFinish = choice.finish_reason;
-      if (chunk.usage) lastUsage = chunk.usage;
-      if (typeof delta.content === "string" && delta.content) { dbg.textChars += delta.content.length; out.push(...pushText(delta.content)); }
-      for (const tc of delta.tool_calls || []) out.push(...pushTool(tc));
-      const rc = typeof delta.reasoning_content === "string" ? delta.reasoning_content : "";
-      if (rc) {
-        dbg.thinkingChars += rc.length;
-        if (emitThinking) out.push(...pushThinking(rc)); // 只发 thinking_delta，无 signature 可发
-        else gen.add(rc); // `MSLXDFF_ANTHROPIC_THINKING=0` 时思考字节仍是真实生成，兜底 output 照计（在此加、不在 pushThinking 内加，避免双计）
-      }
-    }
-    return out;
-  }
-
-  function end({ finish = null, usage = null } = {}) {
-    const out = [];
-    const fr = finish || lastFinish;
-    for (const t of tools.values()) {
-      if (!t.announced && (t.id || t.name || t.args)) announceTool(t, out); // 只有 arguments 无 name 的退化收尾
-    }
-    closeOpen(out);
-    out.push(ev("message_delta", { delta: { stop_reason: toStopReason(fr, sawToolUse), stop_sequence: null }, usage: toAnthropicUsage(usage || lastUsage, () => ({ input_tokens: estInput(), output_tokens: gen.tokens() })) }));
-    out.push(ev("message_stop"));
-    return out;
-  }
-
-  // 上游给聚合 JSON（非 SSE，muse-spark / workbuddy 聚合 / zcode 非流式都是这种）时，
-  // 把整包喂进同一套 pushText/pushTool 状态机 —— 与流式共用封口规则，不出第二套形状。
-  function fromChatJson(chatJson = {}) {
-    const out = [];
-    const choice = chatJson.choices?.[0] || {};
-    const m = choice.message || {};
-    const ct = textFromChatContent(m.content); // 字符串或部件数组都吃，别把数组正文整段丢掉
-    // 聚合上游把思考放在 message.reasoning_content（muse-spark / workbuddy 聚合 / zcode 非流式）：thinking 块须排 text 前。
-    const rt = typeof m.reasoning_content === "string" ? m.reasoning_content : "";
-    if (rt && emitThinking) out.push(...pushThinking(rt));
-    if (ct) out.push(...pushText(ct));
-    (Array.isArray(m.tool_calls) ? m.tool_calls : []).forEach((tc, i) => {
-      out.push(...pushTool({ index: i, id: tc?.id, function: { name: tc?.function?.name, arguments: tc?.function?.arguments } }));
-    });
-    if (choice.finish_reason) lastFinish = choice.finish_reason;
-    if (chatJson.usage) lastUsage = chatJson.usage;
-    return out;
-  }
-
-  function getFinal() {
-    return { finish: lastFinish, usage: lastUsage, sawToolUse };
-  }
-  function stats() {
-    return { ...dbg, openKind: open?.kind || null, toolSlots: tools.size };
-  }
-
-  return { id, created, begin, push, end, fromChatJson, getFinal, stats };
-}
+/**
+ * Anthropic Messages 出站翻译层：chat → Anthropic message / SSE 具名事件（/v1/messages 外壳，ADR-0047）。
+ * 纯函数，无网络、无副作用；出站形状依据 code.claude.com/docs/en/llm-gateway-protocol
+ * 与参考实现 lamdt1/ms-copilot365-2api/app/formatters/anthropic_sse.py。
+ * 分档（ADR-0055，为守本文件 ≤20KB 体积硬门；对外 import 路径全走本文件、调用方零改动）：
+ *  · 请求侧 Anthropic → chat 在 ./request-body.js；
+ *  · 消息/工具 id 与协议错误体在 ./shared.js；
+ *  · token 估算的单一来源在 ../text-tokens.js（上游未报 usage 时由它兜底）。
+ *
+ * 与 src/responses/translate.js 的三点关键差异：
+ *  1) Anthropic 的 content block **只封口不复用**（Responses 侧 ensureTextItem 会复用已开项）；
+ *  2) 事件是具名的（`event: <type>` + `data:`），且**没有 `data: [DONE]`**，收场只有
+ *     `message_delta`+`message_stop`（成功）或一帧 `error`（失败）——见 grill-decisions.md Q2；
+ *  3) 请求侧**保留思考档位**、**必须主动丢** output_config/context_management 与消息块内 thinking
+ *     （块内 thinking 我们无法签发 signature，回传必被上游拒）——细则见 ./request-body.js 头（ADR-0047）。
+ */
+import { chunkToString } from "../responses/translate.js";
+import { createTokenCounter } from "../text-tokens.js";
+import { newMessageId, newToolUseId } from "./shared.js";
+
+// re-export：拆档不改对外接口，调用方与测试仍从本文件取这些符号
+export { messagesToChatBody, toolResultTextOf } from "./request-body.js";
+export { newMessageId, newToolUseId, anthropicError } from "./shared.js";
+
+// ---------- usage / stop_reason ----------
+
+/**
+ * chat 口径 usage → Anthropic 口径。
+ * Anthropic 的 `input_tokens` **不含**缓存读命中的部分（两者分开计费），而上游
+ * `prompt_tokens` 是含缓存的总输入 —— 故命中数要从 input 里扣掉，否则客户端把同一段
+ * 前缀算两遍（读数虚高）。
+ * `fallback()` 只在**上游整条未报**（缺字段，或 prompt/output/cache_read/cache_creation 四项全为 0）时才用：
+ * 真实模型对非空 prompt 不可能报 0，照抄 0 会让客户端的上下文计度停在原地、永不压实（ADR-0055 决策 3）。
+ */
+export function toAnthropicUsage(u = {}, fallback = null) {
+  const num = (v) => (Number.isFinite(Number(v)) ? Number(v) : 0);
+  const src = u && typeof u === "object" ? u : {};
+  const prompt = num(src.prompt_tokens ?? src.input_tokens);
+  const output = num(src.completion_tokens ?? src.output_tokens);
+  const cacheRead = num(src.cache_read_input_tokens ?? src.prompt_tokens_details?.cached_tokens ?? src.input_tokens_details?.cached_tokens);
+  const cacheWrite = num(src.cache_creation_input_tokens ?? src.prompt_tokens_details?.cache_creation_input_tokens);
+  const usage = {
+    input_tokens: Math.max(0, prompt - cacheRead - cacheWrite),
+    output_tokens: output,
+    cache_read_input_tokens: cacheRead,
+    cache_creation_input_tokens: cacheWrite,
+  };
+  // cache 非零也算「报了」：SDK 通道（upstream-engine/sdk/sse.js 的 usageToOpenAI）会输出
+  // `{prompt_tokens:0, …, prompt_tokens_details:{cached_tokens:N}}` 这种形状，若当未报整条覆盖，
+  // 客户端「input + cache_read」会把同一段前缀算两遍（正是上面扣减要防的事，评审 P2）。
+  if (fallback && !(prompt > 0 || output > 0 || cacheRead > 0 || cacheWrite > 0)) return { ...usage, ...(fallback() || {}) };
+  return usage;
+}
+
+function toStopReason(finish, hasToolUse) {
+  if (hasToolUse || finish === "tool_calls" || finish === "function_call") return "tool_use";
+  if (finish === "length") return "max_tokens";
+  return "end_turn";
+}
+
+// ---------- 非流式响应：chat JSON → Anthropic message ----------
+
+function parseToolInput(args) {
+  const raw = String(args || "");
+  if (!raw.trim()) return {};
+  try {
+    const v = JSON.parse(raw);
+    return v && typeof v === "object" && !Array.isArray(v) ? v : { value: v };
+  } catch {
+    return { raw }; // 上游吐出发散 JSON：退化成 raw，不整轮失败
+  }
+}
+
+/**
+ * 上游 chat 的 `message.content` 既可能是字符串，也可能是部件数组（`[{type:"text",text:…}, {type:"image_url",…}]`，
+ * 多模态上游与部分聚合网关都这么回）。统一抽成正文文本；图片部件在本外壳里无法映射成 Anthropic 的 image 块，只能弃（见 ADR-0047 边界 11）。
+ */
+export function textFromChatContent(c) {
+  if (typeof c === "string") return c;
+  if (!Array.isArray(c)) return "";
+  let t = "";
+  for (const p of c) {
+    if (p && typeof p === "object" && typeof p.text === "string") t += p.text;
+  }
+  return t;
+}
+
+export function chatJsonToAnthropic(chatJson = {}, model = "", opts = {}) {
+  const choice = chatJson.choices?.[0] || {};
+  const message = choice.message || {};
+  const content = [];
+  const text = textFromChatContent(message.content);
+  if (text) content.push({ type: "text", text });
+  for (const tc of message.tool_calls || []) {
+    content.push({ type: "tool_use", id: String(tc.id || newToolUseId()), name: String(tc.function?.name || ""), input: parseToolInput(tc.function?.arguments) });
+  }
+  return {
+    id: typeof chatJson.id === "string" && chatJson.id.startsWith("msg_") ? chatJson.id : newMessageId(),
+    type: "message",
+    role: "assistant",
+    model: model || chatJson.model || "",
+    content: content.length ? content : [{ type: "text", text: "" }],
+    stop_reason: toStopReason(choice.finish_reason, content.some((b) => b.type === "tool_use")),
+    stop_sequence: null,
+    usage: toAnthropicUsage(chatJson.usage, () => fallbackUsage(opts, content)),
+  };
+}
+
+// 上游整条未报 usage 时的兜底（ADR-0055 决策 3）：请求侧用注入的 estInput，生成侧按已产出的块估。
+function fallbackUsage(opts = {}, content = []) {
+  const counter = createTokenCounter();
+  for (const b of content) {
+    if (b?.type === "text") counter.add(b.text || "");
+    else if (b?.type === "tool_use") counter.add(JSON.stringify(b.input ?? {})); // 只计参数：与流式 pushTool 同口径（name 从不进 gen）
+  }
+  const n = typeof opts.estInput === "function" ? Number(opts.estInput()) : 0;
+  return { input_tokens: Number.isFinite(n) && n > 0 ? Math.floor(n) : 0, output_tokens: counter.tokens() };
+}
+
+// ---------- 流式响应：chat SSE chunk → Anthropic 具名事件 ----------
+
+// thinking 块必须在 content_block_stop **之前**交一帧 signature_delta，缺它 Claude Code 把整块静默丢弃
+// （routatic/proxy#51 真机复现、PR#133 修）。下面是**身份占位**：不含思考原文、不冒充密码学签名，理由与风险见 ADR-0052。
+const SIG_PREFIX = "mslxdff:thinking:v1:";
+function thinkingSignature(messageId, index) {
+  return Buffer.from(`${SIG_PREFIX}${messageId}:${index}`, "utf8").toString("base64url");
+}
+
+/**
+ * @param model 回显给客户端的模型名
+ * @param opts  { id?, thinking?: boolean }  thinking=true 才把 reasoning_content 发成 thinking 块
+ */
+export function createAnthropicChunkTranslator(model = "", opts = {}) {
+  const id = opts.id || newMessageId();
+  const emitThinking = opts.thinking === true;
+  const created = Math.floor(Date.now() / 1000);
+  const ev = (type, extra = {}) => ({ type, ...extra });
+
+  // 请求侧 token 估算（ADR-0055 决策 3/5）：路由注入 `opts.estInput()`，memoize —— `message_start` 必须在
+  // 首块前就带 input_tokens，全文只能扫一遍；上游后报实测 usage 时估算只留在首帧，不进 message_delta。
+  let estCached = null;
+  const estInput = () => {
+    if (estCached == null) {
+      const n = typeof opts.estInput === "function" ? Number(opts.estInput()) : 0;
+      estCached = Number.isFinite(n) && n > 0 ? Math.floor(n) : 0;
+    }
+    return estCached;
+  };
+  const gen = createTokenCounter(); // 生成侧累计（text + thinking + tool 参数增量）
+
+  let buf = "";
+  let nextIndex = 0;
+  let open = null; // 当前未封口块：{ index, kind: "text"|"thinking"|"tool_use", text? }
+  const tools = new Map(); // 上游 tc.index → { id, name, args, announced, index }
+  let sawToolUse = false;
+  let lastFinish = null;
+  let lastUsage = null;
+  const dbg = { chunks: 0, textChars: 0, toolDeltas: 0, thinkingChars: 0, skippedLines: 0, jsonFails: 0, lateToolDeltas: 0, blocks: 0 };
+
+  function closeOpen(out) {
+    if (!open) return;
+    if (open.kind === "thinking") out.push(ev("content_block_delta", { index: open.index, delta: { type: "signature_delta", signature: thinkingSignature(id, open.index) } }));
+    out.push(ev("content_block_stop", { index: open.index }));
+    open = null;
+  }
+  function openBlock(kind, contentBlock, out) {
+    closeOpen(out);
+    open = { index: nextIndex++, kind, text: "" };
+    dbg.blocks = open.index + 1;
+    out.push(ev("content_block_start", { index: open.index, content_block: contentBlock }));
+    return open;
+  }
+  function pushText(delta) {
+    const out = [];
+    if (!open || open.kind !== "text") openBlock("text", { type: "text", text: "" }, out);
+    open.text += delta;
+    gen.add(delta); // 生成侧只累计字数、末帧一次套公式：逐帧套会把参照实现的 +1 下限乘上帧数
+    out.push(ev("content_block_delta", { index: open.index, delta: { type: "text_delta", text: delta } }));
+    return out;
+  }
+  function pushThinking(delta) {
+    const out = [];
+    if (!open || open.kind !== "thinking") openBlock("thinking", { type: "thinking", thinking: "", signature: "" }, out);
+    open.text += delta;
+    gen.add(delta); // thinking 也按 output 计费（ADR-0055 grill Q6）
+    out.push(ev("content_block_delta", { index: open.index, delta: { type: "thinking_delta", thinking: delta } }));
+    return out;
+  }
+  function announceTool(t, out) {
+    t.announced = true;
+    sawToolUse = true;
+    // 统一走 openBlock：自动封掉当前 text/thinking/上一个工具块，本块成为新的待封 open
+    openBlock("tool_use", { type: "tool_use", id: t.id, name: t.name, input: {} }, out);
+    t.index = open.index;
+    // 增量早于 id/name 到达时先攒着，announce 后补发（事件顺序仍合规）
+    if (t.args) out.push(ev("content_block_delta", { index: t.index, delta: { type: "input_json_delta", partial_json: t.args } }));
+  }
+  function pushTool(tc) {
+    const slot = Number(tc?.index ?? 0);
+    const out = [];
+    let t = tools.get(slot);
+    if (!t) { t = { id: "", name: "", args: "", announced: false, index: null }; tools.set(slot, t); }
+    if (tc?.id) t.id = String(tc.id);
+    if (tc?.function?.name) t.name += String(tc.function.name);
+    const frag = tc?.function?.arguments ? String(tc.function.arguments) : "";
+    if (frag) { dbg.toolDeltas++; t.args += frag; gen.add(frag); }
+    if (!t.announced && t.id && t.name) announceTool(t, out);
+    else if (t.announced && frag && open && open.index === t.index) out.push(ev("content_block_delta", { index: t.index, delta: { type: "input_json_delta", partial_json: frag } }));
+    else if (t.announced && frag) dbg.lateToolDeltas++; // 块已封口后才来的参数增量：发出即违规帧，丢弃但上面已计入 gen（output_tokens 取计费口径：模型确实生成了这些字节）
+    return out;
+  }
+
+  function begin() {
+    return [ev("message_start", { message: { id, type: "message", role: "assistant", model, content: [], stop_reason: null, stop_sequence: null, usage: { input_tokens: estInput(), output_tokens: 0, cache_read_input_tokens: 0, cache_creation_input_tokens: 0 } } })];
+  }
+
+  function push(data) {
+    buf += chunkToString(data);
+    const out = [];
+    let nl;
+    while ((nl = buf.indexOf("\n")) >= 0) {
+      const line = buf.slice(0, nl).trim();
+      buf = buf.slice(nl + 1);
+      // 注释帧/keepalive（`:` 开头）与上游 `[DONE]` 都不进 Anthropic 事件序列（grill Q6）
+      if (!line || line.startsWith(":")) continue;
+      if (!line.startsWith("data:")) { dbg.skippedLines++; continue; }
+      const payload = line.slice(5).trim();
+      if (!payload || payload === "[DONE]") continue;
+      let chunk;
+      try { chunk = JSON.parse(payload); } catch { dbg.jsonFails++; continue; }
+      dbg.chunks++;
+      const choice = chunk.choices?.[0] || {};
+      const delta = choice.delta || {};
+      if (choice.finish_reason) lastFinish = choice.finish_reason;
+      if (chunk.usage) lastUsage = chunk.usage;
+      if (typeof delta.content === "string" && delta.content) { dbg.textChars += delta.content.length; out.push(...pushText(delta.content)); }
+      for (const tc of delta.tool_calls || []) out.push(...pushTool(tc));
+      const rc = typeof delta.reasoning_content === "string" ? delta.reasoning_content : "";
+      if (rc) {
+        dbg.thinkingChars += rc.length;
+        if (emitThinking) out.push(...pushThinking(rc)); // 只发 thinking_delta，无 signature 可发
+        else gen.add(rc); // `MSLXDFF_ANTHROPIC_THINKING=0` 时思考字节仍是真实生成，兜底 output 照计（在此加、不在 pushThinking 内加，避免双计）
+      }
+    }
+    return out;
+  }
+
+  function end({ finish = null, usage = null } = {}) {
+    const out = [];
+    const fr = finish || lastFinish;
+    for (const t of tools.values()) {
+      if (!t.announced && (t.id || t.name || t.args)) announceTool(t, out); // 只有 arguments 无 name 的退化收尾
+    }
+    closeOpen(out);
+    out.push(ev("message_delta", { delta: { stop_reason: toStopReason(fr, sawToolUse), stop_sequence: null }, usage: toAnthropicUsage(usage || lastUsage, () => ({ input_tokens: estInput(), output_tokens: gen.tokens() })) }));
+    out.push(ev("message_stop"));
+    return out;
+  }
+
+  // 上游给聚合 JSON（非 SSE，muse-spark / workbuddy 聚合 / zcode 非流式都是这种）时，
+  // 把整包喂进同一套 pushText/pushTool 状态机 —— 与流式共用封口规则，不出第二套形状。
+  function fromChatJson(chatJson = {}) {
+    const out = [];
+    const choice = chatJson.choices?.[0] || {};
+    const m = choice.message || {};
+    const ct = textFromChatContent(m.content); // 字符串或部件数组都吃，别把数组正文整段丢掉
+    // 聚合上游把思考放在 message.reasoning_content（muse-spark / workbuddy 聚合 / zcode 非流式）：thinking 块须排 text 前。
+    const rt = typeof m.reasoning_content === "string" ? m.reasoning_content : "";
+    if (rt && emitThinking) out.push(...pushThinking(rt));
+    if (ct) out.push(...pushText(ct));
+    (Array.isArray(m.tool_calls) ? m.tool_calls : []).forEach((tc, i) => {
+      out.push(...pushTool({ index: i, id: tc?.id, function: { name: tc?.function?.name, arguments: tc?.function?.arguments } }));
+    });
+    if (choice.finish_reason) lastFinish = choice.finish_reason;
+    if (chatJson.usage) lastUsage = chatJson.usage;
+    return out;
+  }
+
+  function getFinal() {
+    return { finish: lastFinish, usage: lastUsage, sawToolUse };
+  }
+  function stats() {
+    return { ...dbg, openKind: open?.kind || null, toolSlots: tools.size };
+  }
+
+  return { id, created, begin, push, end, fromChatJson, getFinal, stats };
+}