@agent-proxy/agy 0.16.0 → 0.16.1

package/dist/index.js

@@ -1,4 +1,4 @@
-import { bindHostMethods, withHostServices } from "@agent-proxy/plugin-api";
+import { bindHostMethods, getHostServices, withHostServices } from "@agent-proxy/plugin-api";
 import path from "node:path";
 import { isIP } from "node:net";
 import { atomicWrite } from "@agent-proxy/plugin-api";
@@ -247,9 +247,15 @@ function buildAgyAdapter(config, dependencies = {}) {
                 const translated = translateAgyRequest(input, resolvedModel, to);
                 abort.signal.throwIfAborted();
                 const upstream = await transport.streamInfer(credential, translated.wrapper, abort.signal);
+                // 流式响应在 forward 返回后由客户端 pull 驱动，下列回调可能在调用上下文之外
+                // 触发（Bun 的流调度不继承 withHostServices 的 AsyncLocalStorage）；日志等
+                // 宿主服务调用必须重入当前宿主上下文，否则 logExchange 抛错被 safely 吞掉、
+                // 请求日志静默丢失。
+                const host = getHostServices();
+                const inHostContext = (callback) => withHostServices(host, callback);
                 const response = await createAgyClientResponse(to, upstream, {
                     model, tools: translated.tools, stream, signal: abort.signal,
-                    sanitizeError: (error) => {
+                    sanitizeError: (error) => inHostContext(() => {
                         // 上游流内错误的原始对象只进本机网关日志（截断），供排查；客户端只见预定义分类。
                         if (error?.code !== "response_cancelled") {
                             logGatewayError(sink?.processLog, {
@@ -260,17 +266,17 @@ function buildAgyAdapter(config, dependencies = {}) {
                             });
                         }
                         return safeAgyUpstreamError(0, `${String(error.code ?? "")} ${String(error.message ?? "")}`);
-                    },
-                    abort: () => { abort.abort(); cleanup(); },
-                    onChunk: (chunk) => { if (sink)
-                        chunks += chunk; },
-                    onComplete: (payload) => {
+                    }),
+                    abort: () => inHostContext(() => { abort.abort(); cleanup(); }),
+                    onChunk: (chunk) => inHostContext(() => { if (sink)
+                        chunks += chunk; }),
+                    onComplete: (payload) => inHostContext(() => {
                         cleanup();
                         const stripped = translated.dropped.length ? `\n--- stripped built-in tools: ${translated.dropped.join(", ")} ---\n` : "";
                         if (!mapResult)
                             log(stream ? 200 : payload.status === "failed" ? 502 : 200, stream ? `${chunks}${stripped}\n--- response.${String(payload.status)} ---\n${JSON.stringify(payload)}`
                                 : `${JSON.stringify(payload)}${stripped}`, new Headers({ "content-type": stream ? "text/event-stream" : "application/json" }));
-                    },
+                    }),
                 });
                 if (!mapResult)
                     return response;

package/dist/request.js

@@ -455,6 +455,9 @@ function normalizeSourceItems(body, to) {
         const calls = [];
         const results = [];
         const parts = [];
+        // 带响应侧 agn1: 信封的 thinking 块回放为 reasoning 项，由共用路径把签名挂回
+        // 下一个 functionCall（上游强制要求 thoughtSignature，缺失直接 400）。
+        const signatures = [];
         for (const block of raw.content) {
             if (!record(block))
                 fail("messages 包含无效内容块");
@@ -476,7 +479,14 @@ function normalizeSourceItems(body, to) {
                     parts.push({ type: "input_image", image_url: url });
                 }
                 else {
-                    parts.push({ type: "input_text", text: `[thinking] ${typeof block.thinking === "string" ? block.thinking : ""}` });
+                    const signature = typeof block.signature === "string" ? block.signature : "";
+                    if (signature.startsWith("agn1:")) {
+                        // 签名 thinking 块只回放 agn1 信封，思考文本不回放（thought 属于模型私有态）。
+                        signatures.push({ type: "reasoning", encrypted_content: signature });
+                    }
+                    else {
+                        parts.push({ type: "input_text", text: `[thinking] ${typeof block.thinking === "string" ? block.thinking : ""}` });
+                    }
                 }
             }
         }
@@ -484,6 +494,8 @@ function normalizeSourceItems(body, to) {
             items.push(...results);
         else if (parts.length > 0)
             items.push({ type: "message", role: raw.role, content: parts });
+        if (signatures.length > 0)
+            items.push(...signatures);
         if (calls.length > 0)
             items.push(...calls);
     }

package/dist/response.js

@@ -93,9 +93,37 @@ function finishDisposition(finishReason) {
         return { status: "failed", failure: "Antigravity 上游报告函数调用格式无效，请检查工具声明后重试" };
     return { status: "failed", failure: `Antigravity 上游返回不支持的 finishReason：${finishReason}` };
 }
+/* ====================== 四种客户端协议的独立渲染层 ====================== */
+/**
+ * 渲染事件迭代器 → pull 型增量 SSE 响应：客户端每拉一次消费一帧上游事件，
+ * 背压沿链路回传；取消时同步取消上游流，让 streamGenerateContent 一并中止。
+ */
+function incrementalSseResponse(chunks, upstream) {
+    return new Response(new ReadableStream({
+        async pull(controller) {
+            try {
+                const next = await chunks.next();
+                if (next.done) {
+                    controller.close();
+                    return;
+                }
+                controller.enqueue(encoder.encode(next.value));
+            }
+            catch (error) {
+                controller.error(error);
+            }
+        },
+        async cancel() {
+            // 先取消上游以解除挂起的读取，再终结渲染迭代器释放读取锁。
+            await upstream.cancel().catch(() => undefined);
+            await chunks.return(undefined).catch(() => undefined);
+        },
+    }, { highWaterMark: 0 }), { headers: sseHeaders() });
+}
 /**
  * AGY 自产 Responses SSE → Chat Completions chunks（含 [DONE]）。本渲染器属于
- * agy 插件包（不依赖桥接包）：解析自身产出的受信事件流并重组为客户端协议外壳。
+ * agy 插件包（不依赖桥接包）：解析自身产出的受信事件流并重组为客户端协议外壳，
+ * 逐帧增量下发（非流式客户端由渲染器聚合）。
  */
 async function renderChatClient(responsesSse, state) {
     const id = `chatcmpl-${crypto.randomUUID().replaceAll("-", "")}`;
@@ -105,48 +133,62 @@ async function renderChatClient(responsesSse, state) {
     let model = state.model;
     const frames = [];
     const push = (delta) => {
-        frames.push(`data: ${JSON.stringify({ id, object: "chat.completion.chunk", created, model, choices: [{ index: 0, delta, finish_reason: null }] })}\n\n`);
+        const chunk = `data: ${JSON.stringify({ id, object: "chat.completion.chunk", created, model, choices: [{ index: 0, delta, finish_reason: null }] })}\n\n`;
+        frames.push(chunk);
+        return chunk;
     };
     let contentStarted = false;
-    for await (const frame of iterateOwnSse(responsesSse)) {
-        if (frame.type === "response.output_text.delta" && typeof frame.delta === "string") {
-            contentStarted = true;
-            push({ content: frame.delta, role: contentStarted && !frames.length ? "assistant" : undefined });
-        }
-        else if (frame.type === "response.reasoning_summary_text.delta" && typeof frame.delta === "string") {
-            push({ reasoning_content: frame.delta });
-        }
-        else if (frame.type === "response.refusal.delta" && typeof frame.delta === "string") {
-            push({ refusal: frame.delta });
-        }
-        else if (frame.type === "response.output_item.added" && record(frame.item, "item").type === "function_call") {
-            const item = record(frame.item, "item");
-            push({ role: "assistant", tool_calls: [{ index: Number(frame.output_index ?? 0), id: item.call_id ?? "", type: "function", function: { name: item.name ?? "", arguments: "" } }] });
-        }
-        else if (frame.type === "response.function_call_arguments.delta" && typeof frame.delta === "string") {
-            push({ tool_calls: [{ index: Number(frame.output_index ?? 0), function: { arguments: frame.delta } }] });
-        }
-        else if (frame.type === "response.completed" || frame.type === "response.incomplete" || frame.type === "response.failed") {
-            const response = record(frame.response, "response");
-            if (typeof response.model === "string")
-                model = state.model;
-            const output = Array.isArray(response.output) ? response.output : [];
-            finish = response.status === "incomplete" ? "length" : output.some((entry) => record(entry, "output").type === "function_call") ? "tool_calls" : "stop";
-            if (record(response.usage ?? {}, "usage") && response.usage !== null) {
-                const value = record(response.usage, "usage");
-                usage = value;
+    async function* iterate() {
+        for await (const frame of iterateOwnSse(responsesSse)) {
+            const from = frames.length;
+            if (frame.type === "response.output_text.delta" && typeof frame.delta === "string") {
+                contentStarted = true;
+                push({ content: frame.delta, role: contentStarted && !frames.length ? "assistant" : undefined });
             }
-        }
-    }
-    frames.push(`data: ${JSON.stringify({ id, object: "chat.completion.chunk", created, model, choices: [{ index: 0, delta: {}, finish_reason: finish }] })}\n\n`);
-    if (usage) {
-        frames.push(`data: ${JSON.stringify({ id, object: "chat.completion.chunk", created, model, choices: [], usage })}\n\n`);
+            else if (frame.type === "response.reasoning_summary_text.delta" && typeof frame.delta === "string") {
+                push({ reasoning_content: frame.delta });
+            }
+            else if (frame.type === "response.refusal.delta" && typeof frame.delta === "string") {
+                push({ refusal: frame.delta });
+            }
+            else if (frame.type === "response.output_item.added" && record(frame.item, "item").type === "function_call") {
+                const item = record(frame.item, "item");
+                push({ role: "assistant", tool_calls: [{ index: Number(frame.output_index ?? 0), id: item.call_id ?? "", type: "function", function: { name: item.name ?? "", arguments: "" } }] });
+            }
+            else if (frame.type === "response.function_call_arguments.delta" && typeof frame.delta === "string") {
+                push({ tool_calls: [{ index: Number(frame.output_index ?? 0), function: { arguments: frame.delta } }] });
+            }
+            else if (frame.type === "response.completed" || frame.type === "response.incomplete" || frame.type === "response.failed") {
+                const response = record(frame.response, "response");
+                if (typeof response.model === "string")
+                    model = state.model;
+                const output = Array.isArray(response.output) ? response.output : [];
+                finish = response.status === "incomplete" ? "length" : output.some((entry) => record(entry, "output").type === "function_call") ? "tool_calls" : "stop";
+                if (record(response.usage ?? {}, "usage") && response.usage !== null) {
+                    const value = record(response.usage, "usage");
+                    usage = value;
+                }
+            }
+            if (frames.length === from)
+                continue;
+            const chunk = frames.slice(from).join("");
+            safely(() => state.onChunk?.(chunk));
+            yield chunk;
+        }
+        const tail = [
+            `data: ${JSON.stringify({ id, object: "chat.completion.chunk", created, model, choices: [{ index: 0, delta: {}, finish_reason: finish }] })}\n\n`,
+            ...(usage ? [`data: ${JSON.stringify({ id, object: "chat.completion.chunk", created, model, choices: [], usage })}\n\n`] : []),
+            "data: [DONE]\n\n",
+        ];
+        frames.push(...tail);
+        const text = tail.join("");
+        safely(() => state.onChunk?.(text));
+        yield text;
     }
-    frames.push("data: [DONE]\n\n");
-    const text = frames.join("");
-    safely(() => state.onChunk?.(text));
     if (!state.stream) {
         // 非流式：聚合为完整 completion JSON。
+        for await (const chunk of iterate())
+            void chunk;
         let content = "";
         const toolCalls = new Map();
         let reasoning = "";
@@ -192,174 +234,212 @@ async function renderChatClient(responsesSse, state) {
         };
         return Response.json(completion);
     }
-    return new Response(strBytes(text), { headers: sseHeaders() });
+    return incrementalSseResponse(iterate(), responsesSse);
 }
-/** AGY 自产 Responses SSE → Gemini generateContent SSE / 完整 JSON（零桥接依赖）。 */
+/** AGY 自产 Responses SSE → Gemini generateContent SSE / 完整 JSON（零桥接依赖，文本增量下发）。 */
 async function renderGeminiClient(responsesSse, state) {
     let finish = "STOP";
     let model = state.model;
     const parts = [];
     const toolArgs = new Map();
     const toolNames = new Map();
-    let toolCalled = false;
     let usage;
-    for await (const frame of iterateOwnSse(responsesSse)) {
-        if (frame.type === "response.output_text.delta" && typeof frame.delta === "string") {
-            parts.push({ text: frame.delta });
-        }
-        else if (frame.type === "response.reasoning_summary_text.delta" && typeof frame.delta === "string") {
-            parts.push({ text: frame.delta, thought: true });
-        }
-        else if (frame.type === "response.output_item.added" && record(frame.item, "item").type === "function_call") {
-            const item = record(frame.item, "item");
-            const callId = typeof item.call_id === "string" ? item.call_id : "";
-            toolArgs.set(callId, "");
-            toolNames.set(callId, typeof item.name === "string" ? item.name : "");
-        }
-        else if (frame.type === "response.function_call_arguments.delta" && typeof frame.delta === "string") {
-            const callId = [...toolArgs.keys()].at(-1) ?? "";
-            toolArgs.set(callId, (toolArgs.get(callId) ?? "") + frame.delta);
-        }
-        else if (frame.type === "response.completed" || frame.type === "response.incomplete" || frame.type === "response.failed") {
-            const response = record(frame.response, "response");
-            if (typeof response.model === "string")
-                model = state.model;
-            const output = Array.isArray(response.output) ? response.output : [];
-            finish = response.status === "incomplete" ? "MAX_TOKENS" : "STOP";
-            if (response.status === "failed")
-                finish = "OTHER";
-            if (record(response.usage ?? {}, "usage") && response.usage !== null)
-                usage = record(response.usage, "usage");
-            void output;
-        }
-    }
-    // functionCall parts 依到达顺序附加在正文之后（Gemini 的候选内容是单序列）。
-    for (const [callId, argsText] of toolArgs) {
-        let args = {};
-        try {
-            args = record(JSON.parse(argsText || "{}"), "args");
-        }
-        catch {
-            args = {};
-        }
-        parts.push({ functionCall: { name: toolNames.get(callId) ?? "", args } });
-        toolCalled = true;
-    }
-    void toolCalled;
-    const inputTokens = typeof usage?.input_tokens === "number" ? usage.input_tokens : 0;
-    const outputTokens = typeof usage?.output_tokens === "number" ? usage.output_tokens : 0;
-    const usageMetadata = usage
-        ? {
+    const usageMetadata = () => {
+        if (!usage)
+            return undefined;
+        const inputTokens = typeof usage.input_tokens === "number" ? usage.input_tokens : 0;
+        const outputTokens = typeof usage.output_tokens === "number" ? usage.output_tokens : 0;
+        return {
             promptTokenCount: inputTokens,
             candidatesTokenCount: outputTokens,
             totalTokenCount: typeof usage.total_tokens === "number" ? usage.total_tokens : inputTokens + outputTokens,
+        };
+    };
+    async function* iterate() {
+        const frameFor = (part) => `data: ${JSON.stringify({ candidates: [{ content: { parts: [part], role: "model" } }], modelVersion: model })}\n\n`;
+        for await (const frame of iterateOwnSse(responsesSse)) {
+            if (frame.type === "response.output_text.delta" && typeof frame.delta === "string") {
+                const part = { text: frame.delta };
+                parts.push(part);
+                const text = frameFor(part);
+                safely(() => state.onChunk?.(text));
+                yield text;
+            }
+            else if (frame.type === "response.reasoning_summary_text.delta" && typeof frame.delta === "string") {
+                const part = { text: frame.delta, thought: true };
+                parts.push(part);
+                const text = frameFor(part);
+                safely(() => state.onChunk?.(text));
+                yield text;
+            }
+            else if (frame.type === "response.output_item.added" && record(frame.item, "item").type === "function_call") {
+                const item = record(frame.item, "item");
+                const callId = typeof item.call_id === "string" ? item.call_id : "";
+                toolArgs.set(callId, "");
+                toolNames.set(callId, typeof item.name === "string" ? item.name : "");
+            }
+            else if (frame.type === "response.function_call_arguments.delta" && typeof frame.delta === "string") {
+                const callId = [...toolArgs.keys()].at(-1) ?? "";
+                toolArgs.set(callId, (toolArgs.get(callId) ?? "") + frame.delta);
+            }
+            else if (frame.type === "response.completed" || frame.type === "response.incomplete" || frame.type === "response.failed") {
+                const response = record(frame.response, "response");
+                if (typeof response.model === "string")
+                    model = state.model;
+                finish = response.status === "incomplete" ? "MAX_TOKENS" : "STOP";
+                if (response.status === "failed")
+                    finish = "OTHER";
+                if (record(response.usage ?? {}, "usage") && response.usage !== null)
+                    usage = record(response.usage, "usage");
+            }
         }
-        : undefined;
-    if (state.stream) {
-        const frames = [];
-        for (const part of parts) {
-            const chunk = { candidates: [{ content: { parts: [part], role: "model" } }], modelVersion: model };
-            frames.push(`data: ${JSON.stringify(chunk)}\n\n`);
-        }
-        const terminal = { candidates: [{ content: { role: "model" }, finishReason: finish }], modelVersion: model, ...(usageMetadata ? { usageMetadata } : {}) };
-        frames.push(`data: ${JSON.stringify(terminal)}\n\n`);
-        const text = frames.join("");
+        // functionCall parts 依到达顺序附加在正文之后（Gemini 的候选内容是单序列）。
+        for (const [callId, argsText] of toolArgs) {
+            let args = {};
+            try {
+                args = record(JSON.parse(argsText || "{}"), "args");
+            }
+            catch {
+                args = {};
+            }
+            const part = { functionCall: { name: toolNames.get(callId) ?? "", args } };
+            parts.push(part);
+            const text = frameFor(part);
+            safely(() => state.onChunk?.(text));
+            yield text;
+        }
+        const terminal = { candidates: [{ content: { role: "model" }, finishReason: finish }], modelVersion: model, ...(usageMetadata() ? { usageMetadata: usageMetadata() } : {}) };
+        const text = `data: ${JSON.stringify(terminal)}\n\n`;
         safely(() => state.onChunk?.(text));
-        return new Response(strBytes(text), { headers: sseHeaders() });
+        yield text;
     }
+    if (state.stream)
+        return incrementalSseResponse(iterate(), responsesSse);
+    // 非流式：聚合为完整 generateContent JSON。
+    for await (const chunk of iterate())
+        void chunk;
     const completion = {
         candidates: [{ content: { ...(parts.length > 0 ? { parts } : { parts: [{ text: "" }] }), role: "model" }, finishReason: finish }],
         modelVersion: model,
-        ...(usageMetadata ? { usageMetadata } : {}),
+        ...(usageMetadata() ? { usageMetadata: usageMetadata() } : {}),
     };
     safely(() => state.onChunk?.(JSON.stringify(completion)));
     return Response.json(completion);
 }
-/** AGY 自产 Responses SSE → Anthropic Messages 事件（message_start/块生命周期/终态）。 */
+/** AGY 自产 Responses SSE → Anthropic Messages 事件（message_start/块生命周期/终态），逐帧增量下发。 */
 async function renderAnthropicClient(responsesSse, state) {
     const id = `msg_${crypto.randomUUID().replaceAll("-", "")}`;
     let model = state.model;
     const frames = [];
+    const emit = (event, payload) => {
+        frames.push(`event: ${event}\ndata: ${JSON.stringify(payload)}\n\n`);
+    };
     let started = false;
     let blockIndex = -1;
     let blockKind;
     let textValue = "";
     let thinkingValue = "";
-    const openTool = () => {
+    const openTool = (item) => {
         blockIndex += 1;
         blockKind = "tool";
-        frames.push(`event: content_block_start\ndata: ${JSON.stringify({ type: "content_block_start", index: blockIndex, content_block: { type: "tool_use", id: "", name: "", input: {} } })}\n\n`);
+        emit("content_block_start", { type: "content_block_start", index: blockIndex, content_block: { type: "tool_use", id: item.call_id ?? "", name: item.name ?? "", input: {} } });
+    };
+    const openThinking = () => {
+        blockIndex += 1;
+        blockKind = "thinking";
+        thinkingValue = "";
+        emit("content_block_start", { type: "content_block_start", index: blockIndex, content_block: { type: "thinking", thinking: "", signature: "" } });
+    };
+    const openText = () => {
+        blockIndex += 1;
+        blockKind = "text";
+        textValue = "";
+        emit("content_block_start", { type: "content_block_start", index: blockIndex, content_block: { type: "text", text: "" } });
     };
     const closeBlock = () => {
         if (blockKind === undefined)
             return;
-        frames.push(`event: content_block_stop\ndata: ${JSON.stringify({ type: "content_block_stop", index: blockIndex })}\n\n`);
+        emit("content_block_stop", { type: "content_block_stop", index: blockIndex });
         blockKind = undefined;
     };
     let usage = {};
     let stopReason = "end_turn";
-    for await (const frame of iterateOwnSse(responsesSse)) {
-        if (frame.type === "response.created" || frame.type === "response.in_progress") {
-            if (started)
-                continue;
-            started = true;
-            const response = record(frame.response, "response");
-            if (typeof response.model === "string")
-                model = state.model;
-            frames.push(`event: message_start\ndata: ${JSON.stringify({ type: "message_start", message: { id, type: "message", role: "assistant", model, content: [], stop_reason: null, stop_sequence: null, usage: { input_tokens: 0, output_tokens: 0 } } })}\n\n`);
-            frames.push('event: ping\ndata: {"type": "ping"}\n\n');
-        }
-        else if (frame.type === "response.output_text.delta" && typeof frame.delta === "string") {
-            if (blockKind !== "text") {
+    async function* iterate() {
+        for await (const frame of iterateOwnSse(responsesSse)) {
+            const from = frames.length;
+            if (frame.type === "response.created" || frame.type === "response.in_progress") {
+                if (!started) {
+                    started = true;
+                    const response = record(frame.response, "response");
+                    if (typeof response.model === "string")
+                        model = state.model;
+                    emit("message_start", { type: "message_start", message: { id, type: "message", role: "assistant", model, content: [], stop_reason: null, stop_sequence: null, usage: { input_tokens: 0, output_tokens: 0 } } });
+                    emit("ping", { type: "ping" });
+                }
+            }
+            else if (frame.type === "response.output_text.delta" && typeof frame.delta === "string") {
+                if (blockKind !== "text") {
+                    closeBlock();
+                    openText();
+                }
+                textValue += frame.delta;
+                emit("content_block_delta", { type: "content_block_delta", index: blockIndex, delta: { type: "text_delta", text: frame.delta } });
+            }
+            else if (frame.type === "response.reasoning_summary_text.delta" && typeof frame.delta === "string") {
+                if (blockKind !== "thinking") {
+                    closeBlock();
+                    openThinking();
+                }
+                thinkingValue += frame.delta;
+                emit("content_block_delta", { type: "content_block_delta", index: blockIndex, delta: { type: "thinking_delta", thinking: frame.delta } });
+            }
+            else if (frame.type === "response.output_item.added") {
+                const item = record(frame.item, "item");
+                if (item.type === "function_call") {
+                    closeBlock();
+                    openTool(item);
+                }
+                else if (item.type === "reasoning" && typeof item.encrypted_content === "string" && item.encrypted_content.startsWith("agn1:")) {
+                    // thoughtSignature 信封项：渲染为带 signature 的独立 thinking 块，客户端
+                    // 下一轮回传后由请求侧还原挂回 functionCall（上游缺失直接 400）。
+                    closeBlock();
+                    openThinking();
+                    emit("content_block_delta", { type: "content_block_delta", index: blockIndex, delta: { type: "signature_delta", signature: item.encrypted_content } });
+                    closeBlock();
+                }
+            }
+            else if (frame.type === "response.function_call_arguments.delta" && typeof frame.delta === "string") {
+                emit("content_block_delta", { type: "content_block_delta", index: blockIndex, delta: { type: "input_json_delta", partial_json: frame.delta } });
+            }
+            else if (frame.type === "response.output_item.done" && record(frame.item, "item").type === "function_call") {
                 closeBlock();
-                blockIndex += 1;
-                blockKind = "text";
-                textValue = "";
-                frames.push(`event: content_block_start\ndata: ${JSON.stringify({ type: "content_block_start", index: blockIndex, content_block: { type: "text", text: "" } })}\n\n`);
             }
-            textValue += frame.delta;
-            frames.push(`event: content_block_delta\ndata: ${JSON.stringify({ type: "content_block_delta", index: blockIndex, delta: { type: "text_delta", text: frame.delta } })}\n\n`);
-        }
-        else if (frame.type === "response.reasoning_summary_text.delta" && typeof frame.delta === "string") {
-            if (blockKind !== "thinking") {
+            else if (frame.type === "response.completed" || frame.type === "response.incomplete" || frame.type === "response.failed") {
+                const response = record(frame.response, "response");
                 closeBlock();
-                blockIndex += 1;
-                blockKind = "thinking";
-                thinkingValue = "";
-                frames.push(`event: content_block_start\ndata: ${JSON.stringify({ type: "content_block_start", index: blockIndex, content_block: { type: "thinking", thinking: "" } })}\n\n`);
+                if (record(response.usage ?? {}, "usage") && response.usage !== null) {
+                    const value = record(response.usage, "usage");
+                    usage = { input_tokens: typeof value.input_tokens === "number" ? value.input_tokens : 0, output_tokens: typeof value.output_tokens === "number" ? value.output_tokens : 0 };
+                }
+                const output = Array.isArray(response.output) ? response.output : [];
+                stopReason = response.status === "incomplete" ? "max_tokens" : output.some((entry) => record(entry, "output").type === "function_call") ? "tool_use" : "end_turn";
             }
-            thinkingValue += frame.delta;
-            frames.push(`event: content_block_delta\ndata: ${JSON.stringify({ type: "content_block_delta", index: blockIndex, delta: { type: "thinking_delta", thinking: frame.delta } })}\n\n`);
-        }
-        else if (frame.type === "response.output_item.added" && record(frame.item, "item").type === "function_call") {
-            closeBlock();
-            openTool();
-            const item = record(frame.item, "item");
-            frames.push(`event: content_block_start\ndata: ${JSON.stringify({ type: "content_block_start", index: blockIndex, content_block: { type: "tool_use", id: item.call_id ?? "", name: item.name ?? "", input: {} } })}\n\n`);
-        }
-        else if (frame.type === "response.function_call_arguments.delta" && typeof frame.delta === "string") {
-            frames.push(`event: content_block_delta\ndata: ${JSON.stringify({ type: "content_block_delta", index: blockIndex, delta: { type: "input_json_delta", partial_json: frame.delta } })}\n\n`);
-        }
-        else if (frame.type === "response.output_item.done" && record(frame.item, "item").type === "function_call") {
-            closeBlock();
-        }
-        else if (frame.type === "response.completed" || frame.type === "response.incomplete" || frame.type === "response.failed") {
-            const response = record(frame.response, "response");
-            closeBlock();
-            if (record(response.usage ?? {}, "usage") && response.usage !== null) {
-                const value = record(response.usage, "usage");
-                usage = { input_tokens: typeof value.input_tokens === "number" ? value.input_tokens : 0, output_tokens: typeof value.output_tokens === "number" ? value.output_tokens : 0 };
-            }
-            const output = Array.isArray(response.output) ? response.output : [];
-            stopReason = response.status === "incomplete" ? "max_tokens" : output.some((entry) => record(entry, "output").type === "function_call") ? "tool_use" : "end_turn";
-        }
+            if (frames.length === from)
+                continue;
+            const chunk = frames.slice(from).join("");
+            safely(() => state.onChunk?.(chunk));
+            yield chunk;
+        }
+        emit("message_delta", { type: "message_delta", delta: { stop_reason: stopReason, stop_sequence: null }, usage: { output_tokens: typeof usage.output_tokens === "number" ? usage.output_tokens : 0 } });
+        emit("message_stop", { type: "message_stop" });
+        const tail = frames.slice(-2).join("");
+        safely(() => state.onChunk?.(tail));
+        yield tail;
     }
-    frames.push(`event: message_delta\ndata: ${JSON.stringify({ type: "message_delta", delta: { stop_reason: stopReason, stop_sequence: null }, usage: { output_tokens: typeof usage.output_tokens === "number" ? usage.output_tokens : 0 } })}\n\n`);
-    frames.push('event: message_stop\ndata: {"type": "message_stop"}\n\n');
-    const text = frames.join("");
-    safely(() => state.onChunk?.(text));
     if (!state.stream) {
         // 非流式：聚合为完整 Message JSON。
+        for await (const chunk of iterate())
+            void chunk;
         const content = [];
         let current;
         for (const frame of frames) {
@@ -388,6 +468,8 @@ async function renderAnthropicClient(responsesSse, state) {
                 target.text = `${target.text}${delta.text}`;
             else if (delta.type === "thinking_delta" && typeof delta.thinking === "string")
                 target.thinking = `${target.thinking}${delta.thinking}`;
+            else if (delta.type === "signature_delta" && typeof delta.signature === "string")
+                target.signature = `${target.signature}${delta.signature}`;
             else if (delta.type === "input_json_delta" && typeof delta.partial_json === "string") {
                 // 参数按块整段缓存后一次性解析（上游 functionCall 本就整帧到达）。
                 try {
@@ -408,7 +490,7 @@ async function renderAnthropicClient(responsesSse, state) {
         };
         return Response.json(message);
     }
-    return new Response(strBytes(text), { headers: sseHeaders() });
+    return incrementalSseResponse(iterate(), responsesSse);
 }
 /** 解析 AGY 自产的 Responses SSE（受信输入：严格 JSON，忽略 event 行）。 */
 async function* iterateOwnSse(body) {
@@ -443,12 +525,6 @@ async function* iterateOwnSse(body) {
         catch { /* 读取已由取消收尾。 */ }
     }
 }
-function strBytes(text) {
-    const bytes = encoder.encode(text);
-    const buffer = new ArrayBuffer(bytes.byteLength);
-    new Uint8Array(buffer).set(bytes);
-    return buffer;
-}
 function sseHeaders() {
     return new Headers({ "content-type": "text/event-stream; charset=utf-8", "cache-control": "no-cache" });
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agent-proxy/agy",
-  "version": "0.16.0",
+  "version": "0.16.1",
   "description": "Agent Proxy agy 插件；保留基线转换行为。",
   "type": "module",
   "exports": {
@@ -40,5 +40,5 @@
   "dependencies": {
     "@agent-proxy/plugin-api": "0.16.0"
   },
-  "gitHead": "062121f545d3391665789b2e8baae842029262be"
+  "gitHead": "559eed932ecf1112d90c17f50a7908411574031a"
 }