dsh-neartoken 0.17.0 → 0.17.1

package/README.md

@@ -228,6 +228,8 @@ Gemini 系走原生 surface 而不是 openai 兼容层：同模型同提示词
 
 - 四套 wire（chat / anthropic / genai / responses）各自解析 SSE；**图片按各协议发送** —— chat 用 `image_url` part、anthropic 用 `source.base64`、genai 用 `inlineData`、responses 用 `input_image`。发送前先按 `imageMaxEdge` / `imageMaxBytes` 重编码（见「图片压缩」），所以线上字节与媒体类型未必等于附件里的原图。**工具结果里的图片也发**（`read_image` 的结果就是一段文本信封 + 一张图）：anthropic 把图塞回同一个 `tool_result` 的 content 块数组；genai 在同一 user turn 的 parts 里补 `inlineData`；chat 的 tool 消息只能放字符串、responses 的 `function_call_output` 同理，所以图跟在**整段连续工具结果之后**的一条 user 消息 / 条目里（并行调用时不会把 tool 消息切断）。字节由 `attachments.readImage` 在请求前内联；没有 attachments 服务或读不出来会**让请求失败**，不再静默降级成纯文本（曾经就是这样：贴的截图模型根本收不到）。会话里被 image-offload 标记的图片（`offloaded: true`）不发字节；插件也不替它生成占位文本，所以那一次图片对模型完全不可见。文件块（非图片附件）仍不发；`/model` 弹窗仍是内置平铺列表；箭头键盘导航未接。
 - **genai 线的并行工具调用必须收在同一个 user turn 里**：Gemini 校验「一个 functionCall turn ↔ 紧随其后那条 user turn 里等量的 functionResponse parts」，每条工具结果各开一条 user turn 会被上游拒成 400 `Please ensure that the number of function response parts is equal to the number of function call parts of the function call turn`。它是**确定性**错误：历史不变就必然复现，重试、换分组都无用，只能新开会话（0.16.1 及以前踩中过一次：一轮发了 6 个并行 `read`，整轮就此作废）。0.16.2 起把整段连续的工具结果折进同一个 user turn；Gemini ≥ 3 另给 `functionCall` / `functionResponse` 带上 `id`（同一个函数被并行调多次时才分得清谁答谁），`gemma*` 与更早的 Gemini 不发该字段，只按函数名匹配。
+- **anthropic 线的工具结果同样必须合批**：messages API 校验「一条 assistant turn 的 `tool_use` ↔ 紧随其后那条 user turn 的 `tool_result`」一一对应。并行调用时若每条结果各开一条 user 消息，网关报 400 `Invalid message sequence: tool_use and tool_result blocks must be correctly paired and ordered`（0.17.1 以前 anthropic 线就是这个形状 —— genai 线 0.16.2 修了同类问题，anthropic 线漏了）。现在整段连续结果折进紧邻 assistant 的那一条 user 消息；结果没记录（中断、fork、压缩裁剪）的调用补一句占位结果；对不上任何调用的结果丢弃；批内混进来的 user 文本移到结果之后；相邻同角色轮次合并、`tool_result` 固定排在 user 消息首位。这条是**请求体形状**的错，与历史记录无关：升级后此前必挂的会话下一轮就能继续。
+- **请求体按需 brotli 压缩**：每次请求都带全量工具定义，body 常在几十到几百 KB，带截图时上 MB。上游（New API 的 `DecompressRequestMiddleware`）在解析前解压 `content-encoding` 为 gzip / br / zstd 的 body，实测三条 wire（chat / messages / responses）全通；deflate 不支持，且不带该头就按明文解析。所以 **≥16 KB 的 body 以 brotli quality 5 发送**：实测 57 KB 的工具密集请求压到 468 B、2 MB 真实源码压到 418 KB（19%）。四条聊天 wire 与出图的三条通道（`images` / `genai` / `chat`）共用同一个出口 `postJson`，阈值与回落只有一份 —— 带参考图的出图请求是另一个 base64 大户。选型依据：同 CPU 下 brotli 比 gzip 小 8–14%；zstd level 3 与 gzip 等大但快 5 倍，只在 CPU 吃紧时更优；而 brotli 的**运行时默认 quality 是 11**，2 MB 上要 3.7 秒，必须显式钉 5。低于阈值走明文 —— 那么小的上行不值一次压缩加解压。图片主导的请求收益有天花板（base64 编码的已压缩图只能压到 75%，那 25% 只是还原 base64 膨胀），真正的杠杆仍是 `imageMaxEdge` / `imageMaxBytes`。若上游某天不认该编码（400 且 body 解析失败），同一请求立刻以明文重发一次：被拒的请求从未计费，代价是一个往返而不是整次调用。
 - 四套 wire 的 SSE 解析都有「帧样本 → 期望 chunk 序列」回归测试（`src/host/wire-parse.spec.ts`），帧样本取自生产网关的真实响应。网关在内容帧上发 `"usage": null`，只有末帧带真值：任何一处把 `usage` 当对象直接读都会让整条流中途崩掉（曾导致 claude 模型全不可用）。
 - **每个请求都带会话亲和键**：chat 与 responses 发 `prompt_cache_key`，anthropic 发 `metadata.user_id`，取值都是 DSH 的 `GenerateOptions.sessionId`。站点就是靠这个值挑上游渠道的（New API 的「codex cli trace」规则：`^gpt-.*$` + `/v1/responses`，键取自 body 的 `prompt_cache_key`，取不到就把整条规则跳过），而提示词缓存按**上游账号**存 —— 不带键时每个请求随机换账号，缓存永远不中。实测同一 6k 前缀：不带键连发 4 次 `cached_tokens` 全 0；带键后第 2 次起 5888/6011 ≈ 98%。`/v1/chat/completions` 上目前没有对应规则，chat 线即使带键也只是偶发命中，要彻底解决需要站点加一条匹配该路径的亲和规则；anthropic 线的 `metadata.user_id` 同理（规则只作用于 `/v1/messages`，这一条已经匹配）。
 - **anthropic 线显式打缓存断点**：Anthropic 只在 `cache_control` 断点处写缓存，没有断点就整段 prompt 按新输入计价（亲和键只保证落在同一上游账号，不产生缓存）。每个请求打 4 个断点：`tools` 末项、`system` 末块、最后两条消息各自末块 —— harness 每轮全量重放历史，尾部双断点保证下一轮几乎全命中，每轮只写新增片段。实测（Claude Max 超惠分组，~6.3k token 前缀）：无断点连发两次 `cache_read` 全 0；打点后第 2 次起 `in=4, cr=6316`，三轮增长历史 `cr` 累积到 10811、`in` 恒为个位数。断点也覆盖 `tool_result` 结尾的轮次（DSH 会话的常态形态）。

package/lib/client.js

@@ -401,23 +401,23 @@ window.__ModuleLoader__.load({
 			document.head.appendChild(tag);
 		}
 		var group_tags_module_css_default = {
-			"toneCyan": "FotuUq_toneCyan",
-			"toneOrange": "FotuUq_toneOrange",
+			"toneIndigo": "FotuUq_toneIndigo",
+			"toneLime": "FotuUq_toneLime",
 			"toneAmber": "FotuUq_toneAmber",
+			"ratioPricey": "FotuUq_ratioPricey",
+			"toneCyan": "FotuUq_toneCyan",
+			"toneTeal": "FotuUq_toneTeal",
+			"tonePink": "FotuUq_tonePink",
+			"tag": "FotuUq_tag",
+			"toneFuchsia": "FotuUq_toneFuchsia",
+			"ratioBase": "FotuUq_ratioBase",
 			"toneRose": "FotuUq_toneRose",
-			"ratioCheap": "FotuUq_ratioCheap",
-			"toneLime": "FotuUq_toneLime",
+			"toneViolet": "FotuUq_toneViolet",
 			"ratioChip": "FotuUq_ratioChip",
 			"toneEmerald": "FotuUq_toneEmerald",
-			"toneViolet": "FotuUq_toneViolet",
-			"toneIndigo": "FotuUq_toneIndigo",
-			"toneTeal": "FotuUq_toneTeal",
-			"ratioPricey": "FotuUq_ratioPricey",
-			"ratioBase": "FotuUq_ratioBase",
 			"toneSky": "FotuUq_toneSky",
-			"tag": "FotuUq_tag",
-			"toneFuchsia": "FotuUq_toneFuchsia",
-			"tonePink": "FotuUq_tonePink"
+			"ratioCheap": "FotuUq_ratioCheap",
+			"toneOrange": "FotuUq_toneOrange"
 		};
 		//#endregion
 		//#region src/client/group-tags.tsx
@@ -472,39 +472,39 @@ window.__ModuleLoader__.load({
 			document.head.appendChild(tag);
 		}
 		var ModelGroupSelect_module_css_default = {
-			"root": "_2pt-bW_root",
-			"warning": "_2pt-bW_warning",
-			"optionTrailing": "_2pt-bW_optionTrailing",
-			"cellLabel": "_2pt-bW_cellLabel",
-			"cellValue": "_2pt-bW_cellValue",
-			"optionCopy": "_2pt-bW_optionCopy",
 			"groups": "_2pt-bW_groups",
-			"groupName": "_2pt-bW_groupName",
-			"searchRow": "_2pt-bW_searchRow",
-			"status": "_2pt-bW_status",
-			"triggerIcon": "_2pt-bW_triggerIcon",
+			"groupOptionLabel": "_2pt-bW_groupOptionLabel",
 			"menu": "_2pt-bW_menu",
+			"empty": "_2pt-bW_empty",
+			"optionCopy": "_2pt-bW_optionCopy",
+			"root": "_2pt-bW_root",
+			"cellLabel": "_2pt-bW_cellLabel",
 			"searchClear": "_2pt-bW_searchClear",
-			"cellChevron": "_2pt-bW_cellChevron",
-			"trigger": "_2pt-bW_trigger",
-			"groupInline": "_2pt-bW_groupInline",
-			"optionActive": "_2pt-bW_optionActive",
-			"chevron": "_2pt-bW_chevron",
-			"modelName": "_2pt-bW_modelName",
-			"triggerEffort": "_2pt-bW_triggerEffort",
-			"triggerLabel": "_2pt-bW_triggerLabel",
+			"toast": "_2pt-bW_toast",
 			"retry": "_2pt-bW_retry",
+			"groupName": "_2pt-bW_groupName",
+			"trigger": "_2pt-bW_trigger",
+			"modelOption": "_2pt-bW_modelOption",
+			"triggerIcon": "_2pt-bW_triggerIcon",
 			"error": "_2pt-bW_error",
+			"triggerEffort": "_2pt-bW_triggerEffort",
+			"modelName": "_2pt-bW_modelName",
 			"variantCount": "_2pt-bW_variantCount",
+			"option": "_2pt-bW_option",
+			"cellValue": "_2pt-bW_cellValue",
+			"warning": "_2pt-bW_warning",
+			"optionTrailing": "_2pt-bW_optionTrailing",
 			"selected": "_2pt-bW_selected",
-			"check": "_2pt-bW_check",
+			"cellChevron": "_2pt-bW_cellChevron",
+			"triggerLabel": "_2pt-bW_triggerLabel",
 			"search": "_2pt-bW_search",
-			"toast": "_2pt-bW_toast",
-			"groupOptionLabel": "_2pt-bW_groupOptionLabel",
+			"chevron": "_2pt-bW_chevron",
+			"optionActive": "_2pt-bW_optionActive",
+			"groupInline": "_2pt-bW_groupInline",
+			"check": "_2pt-bW_check",
+			"searchRow": "_2pt-bW_searchRow",
 			"cell": "_2pt-bW_cell",
-			"option": "_2pt-bW_option",
-			"modelOption": "_2pt-bW_modelOption",
-			"empty": "_2pt-bW_empty"
+			"status": "_2pt-bW_status"
 		};
 		//#endregion
 		//#region src/client/ModelGroupSelect.tsx
@@ -1070,77 +1070,77 @@ window.__ModuleLoader__.load({
 			document.head.appendChild(tag);
 		}
 		var NearTokenSection_module_css_default = {
-			"badgeOn": "_02HI2q_badgeOn",
+			"metaLabel": "_02HI2q_metaLabel",
+			"codeRow": "_02HI2q_codeRow",
+			"tipTitle": "_02HI2q_tipTitle",
+			"name": "_02HI2q_name",
+			"linkLine": "_02HI2q_linkLine",
+			"metaEditor": "_02HI2q_metaEditor",
+			"fact": "_02HI2q_fact",
+			"chevronOpen": "_02HI2q_chevronOpen",
+			"valueRow": "_02HI2q_valueRow",
+			"form": "_02HI2q_form",
+			"link": "_02HI2q_link",
+			"count": "_02HI2q_count",
+			"code": "_02HI2q_code",
+			"metaTitle": "_02HI2q_metaTitle",
+			"metaIntro": "_02HI2q_metaIntro",
+			"countPicked": "_02HI2q_countPicked",
+			"linkRow": "_02HI2q_linkRow",
+			"rowName": "_02HI2q_rowName",
+			"list": "_02HI2q_list",
+			"voucherEmpty": "_02HI2q_voucherEmpty",
+			"title": "_02HI2q_title",
+			"metaField": "_02HI2q_metaField",
 			"factValue": "_02HI2q_factValue",
+			"cardToggle": "_02HI2q_cardToggle",
+			"topupBar": "_02HI2q_topupBar",
+			"label": "_02HI2q_label",
+			"metaBtn": "_02HI2q_metaBtn",
+			"search": "_02HI2q_search",
+			"voucherAmount": "_02HI2q_voucherAmount",
+			"topupInput": "_02HI2q_topupInput",
 			"notice": "_02HI2q_notice",
 			"tipRow": "_02HI2q_tipRow",
-			"cardToggle": "_02HI2q_cardToggle",
-			"tipGroup": "_02HI2q_tipGroup",
-			"count": "_02HI2q_count",
-			"voucherGroup": "_02HI2q_voucherGroup",
-			"select": "_02HI2q_select",
+			"autosave": "_02HI2q_autosave",
 			"metaHint": "_02HI2q_metaHint",
-			"list": "_02HI2q_list",
-			"form": "_02HI2q_form",
-			"voucherEmpty": "_02HI2q_voucherEmpty",
-			"topup": "_02HI2q_topup",
+			"chevron": "_02HI2q_chevron",
+			"invite": "_02HI2q_invite",
 			"facts": "_02HI2q_facts",
-			"fact": "_02HI2q_fact",
 			"badge": "_02HI2q_badge",
-			"codeRow": "_02HI2q_codeRow",
-			"metaActions": "_02HI2q_metaActions",
-			"buttonPrimary": "_02HI2q_buttonPrimary",
-			"linkRow": "_02HI2q_linkRow",
-			"linkLine": "_02HI2q_linkLine",
-			"card": "_02HI2q_card",
-			"rowName": "_02HI2q_rowName",
-			"factLabel": "_02HI2q_factLabel",
-			"metaCheck": "_02HI2q_metaCheck",
+			"rowMeta": "_02HI2q_rowMeta",
+			"select": "_02HI2q_select",
+			"voucherGroup": "_02HI2q_voucherGroup",
+			"hint": "_02HI2q_hint",
+			"rowRatio": "_02HI2q_rowRatio",
+			"tipGroup": "_02HI2q_tipGroup",
+			"row": "_02HI2q_row",
+			"toolbar": "_02HI2q_toolbar",
+			"metaBtnOn": "_02HI2q_metaBtnOn",
+			"topupChips": "_02HI2q_topupChips",
+			"cardHead": "_02HI2q_cardHead",
+			"voucherList": "_02HI2q_voucherList",
+			"topup": "_02HI2q_topup",
 			"intro": "_02HI2q_intro",
-			"label": "_02HI2q_label",
+			"metaInput": "_02HI2q_metaInput",
+			"badgeOn": "_02HI2q_badgeOn",
 			"handle": "_02HI2q_handle",
-			"tipGroupName": "_02HI2q_tipGroupName",
-			"search": "_02HI2q_search",
-			"voucherList": "_02HI2q_voucherList",
-			"chevron": "_02HI2q_chevron",
-			"chevronOpen": "_02HI2q_chevronOpen",
-			"topupBar": "_02HI2q_topupBar",
-			"metaBtnOn": "_02HI2q_metaBtnOn",
-			"row": "_02HI2q_row",
-			"voucherAmount": "_02HI2q_voucherAmount",
-			"topupChipActive": "_02HI2q_topupChipActive",
+			"refresh": "_02HI2q_refresh",
+			"factLabel": "_02HI2q_factLabel",
+			"countDot": "_02HI2q_countDot",
+			"buttonPrimary": "_02HI2q_buttonPrimary",
+			"tip": "_02HI2q_tip",
+			"card": "_02HI2q_card",
 			"button": "_02HI2q_button",
-			"hint": "_02HI2q_hint",
-			"autosave": "_02HI2q_autosave",
-			"metaEditor": "_02HI2q_metaEditor",
-			"metaIntro": "_02HI2q_metaIntro",
-			"link": "_02HI2q_link",
 			"section": "_02HI2q_section",
-			"topupChips": "_02HI2q_topupChips",
-			"title": "_02HI2q_title",
+			"voucherFact": "_02HI2q_voucherFact",
 			"topupChip": "_02HI2q_topupChip",
-			"rowMeta": "_02HI2q_rowMeta",
-			"name": "_02HI2q_name",
-			"valueRow": "_02HI2q_valueRow",
-			"toolbar": "_02HI2q_toolbar",
+			"tipGroupName": "_02HI2q_tipGroupName",
+			"topupChipActive": "_02HI2q_topupChipActive",
 			"actions": "_02HI2q_actions",
-			"tip": "_02HI2q_tip",
-			"code": "_02HI2q_code",
+			"metaCheck": "_02HI2q_metaCheck",
 			"voucherRow": "_02HI2q_voucherRow",
-			"tipTitle": "_02HI2q_tipTitle",
-			"metaTitle": "_02HI2q_metaTitle",
-			"metaField": "_02HI2q_metaField",
-			"refresh": "_02HI2q_refresh",
-			"countPicked": "_02HI2q_countPicked",
-			"rowRatio": "_02HI2q_rowRatio",
-			"metaInput": "_02HI2q_metaInput",
-			"invite": "_02HI2q_invite",
-			"topupInput": "_02HI2q_topupInput",
-			"metaLabel": "_02HI2q_metaLabel",
-			"voucherFact": "_02HI2q_voucherFact",
-			"metaBtn": "_02HI2q_metaBtn",
-			"cardHead": "_02HI2q_cardHead",
-			"countDot": "_02HI2q_countDot"
+			"metaActions": "_02HI2q_metaActions"
 		};
 		//#endregion
 		//#region src/client/NearTokenSection.tsx

package/lib/index.js

@@ -3,6 +3,8 @@ import { credentialRef, isCredentialRefName } from "@deepseek-ai/dsh-credentials
 import { HarnessError, LlmAdapter, LlmError, ProviderRequestId, attributionHeaders, errorChain, isContextWindowExceededError, isQuotaExceededError } from "@deepseek-ai/dsh-llm";
 import { readFile, rename, unlink, writeFile } from "node:fs/promises";
 import { join } from "node:path";
+import { brotliCompress, constants } from "node:zlib";
+import { promisify } from "node:util";
 import { AttachmentId } from "@deepseek-ai/dsh-attachment";
 import { defineTool } from "@deepseek-ai/dsh-tools";
 import { brandString } from "@deepseek-ai/dsh-brand";
@@ -970,6 +972,77 @@ var NearTokenAccount = class {
 		return info;
 	}
 };
+/** Brotli quality to encode at: past this a smaller body costs real CPU. */
+const QUALITY = 5;
+/** The header pair that announces a brotli body. */
+const BROTLI_HEADERS = { "content-encoding": "br" };
+const compress = promisify(brotliCompress);
+/** One body in the clear: the shape every bail-out below returns. */
+const CLEAR = {
+	headers: {},
+	compressed: false
+};
+/**
+* POST one JSON body, compressed when that pays for itself.
+*
+* The one place a call leaves this plugin, so the threshold, the encoding, and
+* the fallback below cannot drift between the chat wires and the picture
+* channel.
+* @param send - the HTTP carrier, so a test or a caller can substitute one.
+* @param url - absolute endpoint.
+* @param request - headers, body, and cancellation.
+* @returns the response, compressed or not; never a response the caller has to discard.
+*/
+async function postJson(send, url, request) {
+	const encoded = await encodeBody(request.json);
+	const post = (payload, encoding) => send(url, {
+		method: "POST",
+		headers: {
+			...request.headers,
+			...encoding
+		},
+		body: payload,
+		signal: request.signal
+	});
+	const response = await post(encoded.body, encoded.headers);
+	if (response.status !== 400 || !encoded.compressed) return response;
+	await response.body?.cancel().catch(() => {});
+	return post(request.json, {});
+}
+/**
+* Compress one request body, when that pays for itself.
+*
+* Compression is an optimization on top of a request that must go out either
+* way, so every failure here — a runtime without brotli, a body that does not
+* shrink, an encoder error — sends the body in the clear rather than failing
+* the call.
+* @param json - the serialized request body.
+* @returns the bytes to send, the headers they need, and whether they are packed.
+*/
+async function encodeBody(json) {
+	const bytes = Buffer.byteLength(json, "utf8");
+	if (bytes < 16384) return {
+		body: json,
+		...CLEAR
+	};
+	try {
+		const packed = await compress(Buffer.from(json, "utf8"), { params: { [constants.BROTLI_PARAM_QUALITY]: QUALITY } });
+		if (packed.byteLength >= bytes) return {
+			body: json,
+			...CLEAR
+		};
+		return {
+			body: new Uint8Array(packed),
+			headers: BROTLI_HEADERS,
+			compressed: true
+		};
+	} catch {
+		return {
+			body: json,
+			...CLEAR
+		};
+	}
+}
 //#endregion
 //#region src/host/images.ts
 /** Model names that generate pictures on this kind of site, in or out of the catalog's endpoint list. */
@@ -1228,15 +1301,14 @@ var NearTokenImageClient = class {
 	async attempt(request) {
 		const pictures = request.images ?? [];
 		const path = request.channel === "images" ? "/v1/images/generations" : request.channel === "genai" ? `/v1beta/models/${request.model}:generateContent` : "/v1/chat/completions";
-		const response = await request.send(`${this.deps.baseURL()}${path}`, {
-			method: "POST",
+		const response = await postJson(request.send, `${this.deps.baseURL()}${path}`, {
 			headers: {
 				...attributionHeaders(),
 				"content-type": "application/json",
 				...request.channel === "genai" ? { "x-goog-api-key": request.key } : { authorization: `Bearer ${request.key}` },
 				...request.route?.groupRef === void 0 ? {} : { "x-neartoken-group": request.route.groupRef }
 			},
-			body: JSON.stringify(request.channel === "images" ? {
+			json: JSON.stringify(request.channel === "images" ? {
 				model: request.model,
 				prompt: request.prompt,
 				n: 1,
@@ -2397,40 +2469,95 @@ function serializeMessage$3(message, images = /* @__PURE__ */ new Map()) {
 				content: blocks
 			};
 		}
-		case "tool": {
-			const text = textOf$3(content);
-			const pictures = imagesIn(content, images);
-			if (pictures.length === 0) return {
-				role: "user",
-				content: [{
-					type: "tool_result",
-					tool_use_id: message.toolCallId,
-					content: text
-				}]
-			};
-			const blocks = text === "" ? [] : [{
-				type: "text",
-				text
-			}];
-			for (const image of pictures) blocks.push({
-				type: "image",
-				source: {
-					type: "base64",
-					media_type: image.mediaType,
-					data: image.data
-				}
+		default: return;
+	}
+}
+/**
+* What a tool call whose result is missing reads as.
+*
+* Every call the harness sent is answered, but a replayed log, a fork seed, or a
+* compacted history can lack the result. The messages API refuses a request whose
+* `tool_use` has no answer, so the call is answered here instead of dropped.
+*/
+const MISSING_TOOL_RESULT = "No result was recorded for this tool call.";
+/**
+* The block one tool message contributes to the turn that answers its call.
+*
+* A tool result is itself a block list, so the pictures a tool handed back
+* (read_image does exactly that) travel inside the result rather than being
+* flattened into its text.
+* @param message - the harness tool message.
+* @param images - the request's inlined images.
+* @returns the `tool_result` block.
+*/
+function toolResultBlock(message, images = /* @__PURE__ */ new Map()) {
+	const content = message.content;
+	const text = textOf$3(content);
+	const pictures = imagesIn(content, images);
+	if (pictures.length === 0) return {
+		type: "tool_result",
+		tool_use_id: message.toolCallId,
+		content: text
+	};
+	const blocks = text === "" ? [] : [{
+		type: "text",
+		text
+	}];
+	for (const image of pictures) blocks.push({
+		type: "image",
+		source: {
+			type: "base64",
+			media_type: image.mediaType,
+			data: image.data
+		}
+	});
+	return {
+		type: "tool_result",
+		tool_use_id: message.toolCallId,
+		content: blocks
+	};
+}
+/**
+* The call ids one assistant turn asks the next turn to answer.
+* @param turn - the serialized assistant turn.
+* @returns the ids, in block order.
+*/
+function toolUseIds(turn) {
+	const ids = [];
+	for (const block of turn.content) if (block.type === "tool_use") ids.push(String(block.id));
+	return ids;
+}
+/**
+* Fold adjacent same-role turns into one.
+*
+* The messages API reads the conversation as one turn per role, and a gateway in
+* front of it may not merge a repeated role itself. Inside a user turn the
+* `tool_result` blocks come first: a result that follows text is refused.
+* @param turns - the settled turns, in send order.
+* @returns the merged messages, ready to send.
+*/
+function mergeAdjacent(turns) {
+	const merged = [];
+	for (const turn of turns) {
+		const last = merged.at(-1);
+		if (last === void 0 || last.role !== turn.role) {
+			merged.push({
+				role: turn.role,
+				content: [...turn.content]
 			});
-			return {
-				role: "user",
-				content: [{
-					type: "tool_result",
-					tool_use_id: message.toolCallId,
-					content: blocks
-				}]
-			};
+			continue;
 		}
-		default: return;
+		if (last.role === "assistant") {
+			last.content.push(...turn.content);
+			continue;
+		}
+		last.content.unshift(...turn.content.filter((block) => block.type === "tool_result"));
+		last.content.push(...turn.content.filter((block) => block.type !== "tool_result"));
 	}
+	return merged.map((turn) => ({
+		role: turn.role,
+		content: turn.content
+	}));
 }
 /**
 * The breakpoint that marks a cacheable prefix end.
@@ -2472,6 +2599,13 @@ function claudeEffort(effort) {
 /**
 * Serialize one harness request into an Anthropic messages body.
 *
+* The turn list is settled into the pairing the messages API checks: every
+* `tool_use` of an assistant turn is answered by a `tool_result` in the single
+* user turn that follows it, calls the harness never recorded a result for are
+* answered with a placeholder, results are collected into one turn rather than
+* one turn each (a parallel batch of calls and a split batch of results are
+* refused as unpaired), and adjacent same-role turns are folded into one.
+*
 * Reasoning is requested as `thinking.type = "adaptive"` with the effort in
 * `output_config`, which every Claude generation behind this gateway accepts.
 * The older `thinking.type = "enabled"` budget form is not portable: models of
@@ -2485,9 +2619,35 @@ function claudeEffort(effort) {
 * @returns the ready-to-JSON body object.
 */
 function serializeRequest$3(options, model, images = /* @__PURE__ */ new Map()) {
-	const messages = [];
+	const turns = [];
+	const results = [];
+	let open = [];
+	let deferred = [];
+	const settle = () => {
+		if (results.length === 0 && open.length === 0) return;
+		for (const id of open) results.push({
+			type: "tool_result",
+			tool_use_id: id,
+			content: MISSING_TOOL_RESULT
+		});
+		for (const turn of deferred) results.push(...turn.content);
+		turns.push({
+			role: "user",
+			content: results.splice(0, results.length)
+		});
+		open = [];
+		deferred = [];
+	};
 	let system = options.system;
 	for (const message of options.messages) {
+		if (message.role === "tool") {
+			const callId = String(message.toolCallId);
+			if (!open.includes(callId)) continue;
+			results.push(toolResultBlock(message, images));
+			open = open.filter((pending) => pending !== callId);
+			if (open.length === 0) settle();
+			continue;
+		}
 		if (message.role === "system" && system === void 0) {
 			const text = textOf$3(message.content);
 			if (text !== "") {
@@ -2495,9 +2655,22 @@ function serializeRequest$3(options, model, images = /* @__PURE__ */ new Map())
 				continue;
 			}
 		}
-		const wire = serializeMessage$3(message, images);
-		if (wire !== void 0) messages.push(wire);
+		const turn = serializeMessage$3(message, images);
+		if (turn === void 0) continue;
+		if (turn.role === "assistant") {
+			settle();
+			turns.push(turn);
+			open = toolUseIds(turn);
+			continue;
+		}
+		if (open.length > 0) {
+			deferred.push(turn);
+			continue;
+		}
+		turns.push(turn);
 	}
+	settle();
+	const messages = mergeAdjacent(turns);
 	const body = {
 		model,
 		messages,
@@ -4290,13 +4463,12 @@ var NearTokenAdapter = class extends LlmAdapter {
 		const images = await inlineImages(this.deps.attachments(), options.messages, options.signal, this.deps.imagePolicy());
 		let response;
 		try {
-			response = await fetch(wire.url(this.deps.baseURL(), model), {
-				method: "POST",
+			response = await postJson(fetch, wire.url(this.deps.baseURL(), model), {
 				headers: {
 					...attributionHeaders(),
 					...wire.headers(key, groupRef)
 				},
-				body: JSON.stringify(wire.serialize(options, model, images)),
+				json: JSON.stringify(wire.serialize(options, model, images)),
 				signal: options.signal
 			});
 		} catch (error) {

package/lib/types/host/anthropic-wire.d.ts

@@ -3,6 +3,13 @@ import type { WireImages } from './wire-images.ts';
 /**
  * Serialize one harness request into an Anthropic messages body.
  *
+ * The turn list is settled into the pairing the messages API checks: every
+ * `tool_use` of an assistant turn is answered by a `tool_result` in the single
+ * user turn that follows it, calls the harness never recorded a result for are
+ * answered with a placeholder, results are collected into one turn rather than
+ * one turn each (a parallel batch of calls and a split batch of results are
+ * refused as unpaired), and adjacent same-role turns are folded into one.
+ *
  * Reasoning is requested as `thinking.type = "adaptive"` with the effort in
  * `output_config`, which every Claude generation behind this gateway accepts.
  * The older `thinking.type = "enabled"` budget form is not portable: models of

package/lib/types/host/request-body.d.ts

@@ -0,0 +1,48 @@
+/**
+ * Smallest body, in bytes, that is worth a compression round trip.
+ *
+ * Below this the saved upload is a few milliseconds even on a slow link, which
+ * the fixed cost of encoding and decoding eats.
+ */
+export declare const MIN_COMPRESS_BYTES: number;
+/** One request body ready to send, with the headers its encoding requires. */
+export interface EncodedBody {
+    /** The bytes to send: the JSON text itself, or its brotli form. */
+    readonly body: Uint8Array<ArrayBuffer> | string;
+    /** Encoding headers; empty when the body travels in the clear. */
+    readonly headers: Readonly<Record<string, string>>;
+    /** Whether {@link EncodedBody.body} is compressed, so a caller can retry in the clear. */
+    readonly compressed: boolean;
+}
+/** One JSON POST the encoding layer shapes: where it goes, what it sends, how it authenticates. */
+export interface JsonPost {
+    /** Headers the call needs regardless of encoding. */
+    readonly headers: Readonly<Record<string, string>>;
+    /** The serialized JSON body. */
+    readonly json: string;
+    /** Cancellation passthrough. */
+    readonly signal?: AbortSignal | undefined;
+}
+/**
+ * POST one JSON body, compressed when that pays for itself.
+ *
+ * The one place a call leaves this plugin, so the threshold, the encoding, and
+ * the fallback below cannot drift between the chat wires and the picture
+ * channel.
+ * @param send - the HTTP carrier, so a test or a caller can substitute one.
+ * @param url - absolute endpoint.
+ * @param request - headers, body, and cancellation.
+ * @returns the response, compressed or not; never a response the caller has to discard.
+ */
+export declare function postJson(send: typeof fetch, url: string, request: JsonPost): Promise<Response>;
+/**
+ * Compress one request body, when that pays for itself.
+ *
+ * Compression is an optimization on top of a request that must go out either
+ * way, so every failure here — a runtime without brotli, a body that does not
+ * shrink, an encoder error — sends the body in the clear rather than failing
+ * the call.
+ * @param json - the serialized request body.
+ * @returns the bytes to send, the headers they need, and whether they are packed.
+ */
+export declare function encodeBody(json: string): Promise<EncodedBody>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-neartoken",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "publishConfig": {
     "access": "public"
   },