@morlay/dsh-llm-openai-compatible 0.0.1

package/README.md

@@ -0,0 +1,134 @@
+# @morlay/dsh-llm-openai-compatible
+
+DeepSeek Harness 的 **OpenAI 兼容 LLM 适配器**插件。与内置 `llm-pi-ai` /
+`llm-deepseek` 不同，本插件的配置 schema 支持 **profile 级默认采样参数**
+（`temperature` / `topP` / `topK` / `presencePenalty` / `frequencyPenalty` /
+`seed`），请求级 `GenerateOptions.temperature` 优先于 profile 默认值。用
+`providers` dict 多路由结构（与 `llm-pi-ai` 一致），用户可将现有
+`llm-pi-ai` 配置近乎无缝迁移。
+
+传输层复用 **[@ai-sdk/openai-compatible](https://www.npmjs.com/package/@ai-sdk/openai-compatible)**
+（`LanguageModelV4.doStream`：wire 序列化与 SSE 解析由 SDK 负责）；本插件负责
+harness 消息 → AI SDK prompt 转换、采样默认合并、stream part → `StreamChunk`
+翻译、错误归一化与凭据策略。
+
+## 安装
+
+```sh
+dsh plugin --profile web add "@morlay/dsh-llm-openai-compatible"
+```
+
+插件默认休眠（空 `providers`）：不声明任何 provider 路由，直到你在
+`settings.yaml` 中配置。
+
+## 配置
+
+`providers` 是 dict：**key 就是 provider 路由键**（选择器与
+`GenerateOptions.provider` 使用），值是 profile。
+
+```yaml
+llm-openai-compatible:
+  providers:
+    ollama:
+      apiKeyEnv: OLLAMA_API_KEY
+      baseURL: https://ollama.com/v1
+      displayName: Ollama Gateway
+      # === 采样默认参数（请求级 temperature 优先）===
+      temperature: 1 # 0..2
+      topP: 0.95 # 0..1 → wire top_p
+      topK: 40 # 正整数 → wire top_k（非标准，仅网关支持时发送）
+      presencePenalty: 0 # -2..2 → wire presence_penalty
+      frequencyPenalty: 0 # -2..2 → wire frequency_penalty
+      seed: 42 # 正整数 → wire seed
+      # === 推理 ===
+      reasoning: high # 部署默认档位（省略 = 提供方默认）
+      # === 模型目录 ===
+      defaultContextWindow: 262144
+      defaultMaxTokens: 32768
+      models:
+        - id: deepseek-v4-flash:0731
+          name: DeepSeek V4 Flash
+          contextWindow: 1000000
+          maxTokens: 65535
+          inputModalities: [text, image]
+          reasoningEfforts:
+            off: # off 空值 = 不发送 reasoning_effort
+            high: high # 档位 → wire reasoning_effort 拼写
+            max: max
+      # === 传输 ===
+      maxRequestImageBytes: 20971520
+      streamIdleTimeoutMs: 300000
+      timeoutMs: 600000 # 整体请求超时；缺省不设
+      retryPolicy:
+        mode: normal
+        maxRetries: 5
+```
+
+### 采样默认值合并规则
+
+| wire 字段                                                   | 取值                                                               | 省略语义                    |
+| ----------------------------------------------------------- | ------------------------------------------------------------------ | --------------------------- |
+| `temperature`                                               | `options.temperature ?? profile.temperature`                       | 都不给 → 不发送，提供方默认 |
+| `max_tokens`                                                | `options.maxTokens ?? model.maxTokens ?? profile.defaultMaxTokens` | 都不给 → 不发送             |
+| `top_p` / `presence_penalty` / `frequency_penalty` / `seed` | `profile` 值                                                       | undefined → 不发送          |
+| `top_k`                                                     | `profile.topK`，经 `providerOptions` 透传进请求体                  | undefined → 不发送          |
+| `reasoning_effort`                                          | 见下                                                               | 解析不出 → 不发送           |
+
+### reasoning 映射（OpenAI 风格）
+
+- 模型声明 `reasoningEfforts`（对象）后，该模型的选器公开 `efforts`（按声明
+  顺序）+ `defaultEffort`（= `profile.reasoning`，须在模型能力内，否则视为无
+  默认值——**描述模型时绝不抛错**）。
+- wire：非 `off` 档位发送 `reasoning_effort: <声明值>`；`off` → 不发送。
+- 请求级 `options.reasoningEffort` 不在模型能力内 → 网络 I/O 前抛
+  `LlmError('UNSUPPORTED_REASONING_EFFORT')`。`profile.reasoning` 配了模型不
+  支持的档位 → 请求执行处失败（同一错误码），配置页面仍可编辑。
+- 模型不声明 `reasoningEfforts`（或 `false`）→ 不公开 reasoning 能力。
+
+### 模型目录
+
+- `models` 缺省 = 服务**空目录**：`listModels` 返回空，未列出的 id 原样透传
+  （`resolveModel` 返回基础信息 + `defaultContextWindow` / `defaultMaxTokens`）。
+- 模型 `maxTokens` 配置后成为该模型的 per-request 默认输出上限。
+- `inputModalities` 缺省 `[text]`；声明含 `image` 的模型接受图片输入
+  （attachments seam，base64 data-URL parts）。
+
+### 凭据
+
+- profile 设置 `apiKeyEnv` 后：`ctx.credentials` 优先，其次
+  `launchEnvironmentOf(ctx)`；解析不到 → `LlmError('MISSING_CREDENTIAL')`。
+- profile 不设置 `apiKeyEnv` → 请求不带 `authorization` 头（无认证端点，
+  如本地 Ollama）。
+
+## 传输
+
+- 端点 = `baseURL` + `/chat/completions`（streaming，`stream_options.include_usage`）。
+- 每个请求携带 `attributionHeaders()` + `x-…-harness-user-id`（+ session-id /
+  compaction 标头），并带 SDK 的 `ai-sdk/openai-compatible` user-agent 后缀。
+- `streamIdleTimeoutMs` 控制流空闲超时（`TIMEOUT`）；`timeoutMs` 控制整体请求
+  超时（缺省不设）。
+- 错误映射：401/403 → `AUTH`、429 → `RATE_LIMIT`、400+上下文 →
+  `CONTEXT_WINDOW_EXCEEDED`、5xx → `SERVER`、配额 → `QUOTA_EXCEEDED`。
+- 用量：`prompt_tokens_details.cached_tokens` 与 DeepSeek 方言的
+  `prompt_cache_hit_tokens` 都被拆出为 `cacheReadTokens`（disjoint 计数）。
+
+## 从 `llm-pi-ai` 迁移
+
+把 `llm-pi-ai.providers.<route>` 的 `api`/`baseURL`/`models`/采样字段平移到
+`llm-openai-compatible.providers.<route>`，`apiKeyEnv` 与 `retryPolicy` 原样
+保留；`reasoningEfforts` 的 `off` 空值语义一致。
+
+## 暂缓能力（YAGNI）
+
+- 不做 `modelOverrides`（providers 里每个路由自己写 `models` 即可）。
+- 不做模型 discovery（端点询问 `GET /models`）；需要时手写 `models`。
+- 不做 OAuth / 非 bearer 认证。
+
+## 构建与验证
+
+```bash
+pnpm install
+pnpm --filter @morlay/dsh-llm-openai-compatible run prepare   # → lib/*.mjs + *.d.mts
+pnpm exec tsc --noEmit
+pnpm exec vitest run
+```

package/cordis.patch.yml

@@ -0,0 +1,24 @@
+- insert:
+    - id: llm-openai-compatible
+      name: "@morlay/dsh-llm-openai-compatible"
+      config:
+        providers:
+          ollama:
+            apiKeyEnv: OLLAMA_API_KEY
+            baseURL: https://ollama.com/v1
+            displayName: Ollama Gateway
+            temperature: 1 # 0..2
+            topP: 0.95
+            reasoning: high
+            defaultContextWindow: 1000000
+            defaultMaxTokens: 65535
+            models:
+              - id: deepseek-v4-flash:0731
+                name: "DeepSeek V4 Flash: 0731"
+                contextWindow: 1000000
+                maxTokens: 65535
+                inputModalities: [text]
+                reasoningEfforts:
+                  off: # off 空值 = 不发送 reasoning_effort
+                  high: high
+                  max: max

package/lib/adapter.d.mts

@@ -0,0 +1,137 @@
+import { GenerateOptions, LlmAdapter, LlmModelInfo, LlmProviderInfo, LlmResolvedModelInfo, ModelModality, ResolvedRetryPolicy, StreamChunk } from "@deepseek-ai/dsh-llm";
+import { CredentialRef } from "@deepseek-ai/dsh-credentials";
+import { LanguageModelV4Usage } from "@ai-sdk/provider";
+import { AttachmentStore } from "@deepseek-ai/dsh-attachment";
+//#region src/adapter.d.ts
+/** Default maximum idle interval while an adapter stream read is outstanding. */
+declare const DEFAULT_STREAM_IDLE_TIMEOUT_MS = 300000;
+/** Default combined request/response context capacity for unconfigured models. */
+declare const DEFAULT_CONTEXT_WINDOW = 262144;
+/** Default per-request output-token cap for unconfigured models. */
+declare const DEFAULT_MAX_TOKENS = 32768;
+/** Default bound on accumulated base64 image payload per request. */
+declare const DEFAULT_MAX_REQUEST_IMAGE_BYTES: number;
+/** Code stamped on the idle-watchdog timeout reason. */
+declare const STREAM_IDLE_TIMEOUT_CODE = "LLM_STREAM_IDLE_TIMEOUT";
+/** Code stamped on the whole-request deadline timeout reason. */
+declare const REQUEST_TIMEOUT_CODE = "LLM_REQUEST_TIMEOUT";
+/** The provider options key the SDK forwards into the request body. */
+declare const PROVIDER_OPTIONS_KEY = "openai-compatible";
+/** Selectable reasoning effort levels for one provider route. */
+type ReasoningEffort = "off" | "low" | "high" | "max";
+/** One validated catalog model of a provider route. */
+interface ResolvedModelProfile {
+  id: string;
+  name?: string;
+  description?: string;
+  contextWindow?: number;
+  maxTokens?: number;
+  inputModalities: readonly ModelModality[];
+  /**
+   * Declared reasoning efforts: key = selectable level, value = wire
+   * `reasoning_effort` spelling. `false` rejects the capability outright;
+   * absent means the model carries no reasoning metadata. A `null` wire
+   * spelling (only legal for `off`) means "omit the field".
+   */
+  reasoningEfforts?: false | Partial<Record<ReasoningEffort, string | null>>;
+}
+/** One validated provider route profile, detached and ready for per-request reads. */
+interface ResolvedProviderProfile {
+  provider: string;
+  displayName: string;
+  /** Credential reference; absence means the route sends no authorization header. */
+  apiKeyEnv?: CredentialRef;
+  /** Required endpoint base; requests hit `${baseURL}/chat/completions`. */
+  baseURL: string;
+  headers?: Readonly<Record<string, string>>;
+  temperature?: number;
+  topP?: number;
+  topK?: number;
+  presencePenalty?: number;
+  frequencyPenalty?: number;
+  seed?: number;
+  /** Deployment default reasoning level; omission keeps the provider default. */
+  reasoning?: ReasoningEffort;
+  models: readonly ResolvedModelProfile[];
+  defaultContextWindow: number;
+  defaultMaxTokens: number;
+  maxRequestImageBytes: number;
+  streamIdleTimeoutMs: number;
+  /** Whole-request deadline in milliseconds; unset arms no overall timer. */
+  timeoutMs?: number;
+  retryPolicy: ResolvedRetryPolicy;
+}
+/** Constructor options for {@link OpenAICompatibleAdapter}: the hooks the plugin owns. */
+interface OpenAICompatibleAdapterOptions {
+  /** Current validated profiles by provider route; called once per operation. */
+  profiles: () => ReadonlyMap<string, ResolvedProviderProfile>;
+  /**
+   * Resolve the credential for one already-resolved profile; called once per
+   * stream call and frozen for that call. `undefined` means the route sends no
+   * authorization header (an unauthenticated endpoint such as local Ollama).
+   */
+  resolveApiKey: (provider: string, profile: ResolvedProviderProfile) => Promise<string | undefined>;
+  /** Resolve the harness anonymous user id for request attribution headers. */
+  resolveUserId: () => string;
+  /** Resolve the optional durable attachment service at request time. */
+  resolveAttachments?: () => AttachmentStore | undefined;
+}
+/** The wire usage shape the SDK converter receives (subset of the OpenAI shape). */
+interface WireUsageLike {
+  prompt_tokens?: number | null | undefined;
+  completion_tokens?: number | null | undefined;
+  prompt_tokens_details?: {
+    cached_tokens?: number | null | undefined;
+  } | null | undefined;
+  /** DeepSeek-dialect cache hits folded into prompt_tokens. */
+  prompt_cache_hit_tokens?: number | null | undefined;
+  completion_tokens_details?: {
+    reasoning_tokens?: number | null | undefined;
+  } | null | undefined;
+}
+/**
+ * Convert provider token accounting into disjoint AI SDK usage. OpenAI's
+ * `prompt_tokens_details.cached_tokens` and the DeepSeek dialect's
+ * `prompt_cache_hit_tokens` both report cache reads folded into
+ * `prompt_tokens`; the harness convention is disjoint counts, so cache reads
+ * are split out regardless of which field the endpoint used.
+ */
+declare function convertUsage(usage: WireUsageLike | null | undefined): LanguageModelV4Usage;
+/**
+ * Map an HTTP status to a stable LlmError code.
+ * @param status - status of a non-2xx provider response.
+ * @param error - parsed provider error body, when available.
+ * @returns the normalized harness error code.
+ */
+declare function httpErrorCode(status: number, error?: {
+  code?: unknown;
+  type?: unknown;
+  message?: unknown;
+}): string;
+/**
+ * Multi-provider adapter. Each operation reads the current profiles, so a
+ * configuration change reaches the next request without a restart; the
+ * underlying SDK provider instance is cached per resolved profile and rebuilt
+ * when the profile object changes.
+ */
+declare class OpenAICompatibleAdapter extends LlmAdapter {
+  private readonly config;
+  private readonly sdkProviders;
+  constructor(config: OpenAICompatibleAdapterOptions);
+  /** The profile for one route, or the not-owned failure. */
+  private profileOf;
+  /** The configured descriptor for one exact route/model pair; unlisted ids pass through. */
+  private modelOf;
+  /** The SDK chat model for one route/model, cached per resolved profile. */
+  private sdkModel;
+  providerInfo(provider: string): LlmProviderInfo;
+  providerRetryPolicy(provider: string): ResolvedRetryPolicy | undefined;
+  listModels(provider: string): Promise<readonly LlmModelInfo[]>;
+  resolveModel(provider: string, model: string, _signal?: AbortSignal): Promise<LlmResolvedModelInfo>;
+  stream(options: GenerateOptions): AsyncGenerator<StreamChunk>;
+  /** Normalize an SDK/transport failure into a harness LlmError. */
+  private normalizeTransportError;
+}
+//#endregion
+export { DEFAULT_CONTEXT_WINDOW, DEFAULT_MAX_REQUEST_IMAGE_BYTES, DEFAULT_MAX_TOKENS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, OpenAICompatibleAdapter, OpenAICompatibleAdapterOptions, PROVIDER_OPTIONS_KEY, REQUEST_TIMEOUT_CODE, ReasoningEffort, ResolvedModelProfile, ResolvedProviderProfile, STREAM_IDLE_TIMEOUT_CODE, convertUsage, httpErrorCode };
+//# sourceMappingURL=adapter.d.mts.map

package/lib/adapter.d.mts.map

@@ -0,0 +1 @@
+{"version":3,"file":"adapter.d.mts","names":[],"sources":["../src/adapter.ts"],"mappings":";;;;;;cA4Ca;;cAEA;;cAEA;;cAEA;;cAEA;;cAEA;;cAEA;;KAGD;;UAGK;EACf;EACA;EACA;EACA;EACA;EACA,0BAA0B;;;;;;;EAO1B,2BAA2B,QAAQ,OAAO;;;UAI3B;EACf;EACA;;EAEA,YAAY;;EAEZ;EACA,UAAU,SAAS;EAEnB;EACA;EACA;EACA;EACA;EACA;;EAEA,YAAY;EAEZ,iBAAiB;EACjB;EACA;EAEA;EACA;;EAEA;EACA,aAAa;;;UAIE;;EAEf,gBAAgB,oBAAoB;;;;;;EAMpC,gBACE,kBACA,SAAS,4BACN;;EAEL;;EAEA,2BAA2B;;;UAInB;EACR;EACA;EACA;IAA0B;;;EAE1B;EACA;IAA8B;;;;;;;;;;iBAUhB,aAAa,OAAO,mCAAmC;;;;;;;iBAqEvD,cACd,gBACA;EAAU;EAAgB;EAAgB;;;;;;;;cA2C/B,gCAAgC;mBAC1B;mBACA;EAEL,YAAA,QAAQ;;UAMZ;;UAWA;;UAQA;EAqBR,aAAa,mBAAmB;EAOhC,oBAAoB,mBAAmB;EAIvC,WAAW,mBAAmB,iBAAiB;EAK/C,aACE,kBACA,eACA,UAAU,cACT,QAAQ;EAeJ,OAAO,SAAS,kBAAkB,eAAe;;UAoHhD"}

package/lib/adapter.mjs

@@ -0,0 +1,295 @@
+import { serializeCallOptions, serializeCallOptionsWithImages } from "./serialize.mjs";
+import { translate } from "./translate.mjs";
+import { CONTEXT_WINDOW_EXCEEDED_CODE, LlmAdapter, LlmError, ProviderRequestId, QUOTA_EXCEEDED_CODE, ReasoningEffortId, attributionHeaders, contentHasImage, isContextWindowExceededError, isQuotaExceededError } from "@deepseek-ai/dsh-llm";
+import { deadline, idleWatchdog, timeoutOf } from "@deepseek-ai/dsh-timeout";
+import { APICallError } from "@ai-sdk/provider";
+import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
+//#region src/adapter.ts
+/**
+* `OpenAICompatibleAdapter`: a multi-provider harness adapter built on
+* `@ai-sdk/openai-compatible`. One instance serves every provider route in
+* the plugin's `providers` dict; profile facts arrive through a thunk resolved
+* once per operation, so the registering plugin owns validation, layering, and
+* credential policy, and a changed profile reaches the next request without a
+* restart. The SDK owns wire serialization and SSE parsing; this adapter owns
+* harness message conversion, sampling-default merging (via
+* `serialize.ts`), chunk translation (via `translate.ts`), and error
+* normalization.
+* @module dsh-llm-openai-compatible/adapter
+*/
+/** Default maximum idle interval while an adapter stream read is outstanding. */
+const DEFAULT_STREAM_IDLE_TIMEOUT_MS = 3e5;
+/** Default combined request/response context capacity for unconfigured models. */
+const DEFAULT_CONTEXT_WINDOW = 262144;
+/** Default per-request output-token cap for unconfigured models. */
+const DEFAULT_MAX_TOKENS = 32768;
+/** Default bound on accumulated base64 image payload per request. */
+const DEFAULT_MAX_REQUEST_IMAGE_BYTES = 20971520;
+/** Code stamped on the idle-watchdog timeout reason. */
+const STREAM_IDLE_TIMEOUT_CODE = "LLM_STREAM_IDLE_TIMEOUT";
+/** Code stamped on the whole-request deadline timeout reason. */
+const REQUEST_TIMEOUT_CODE = "LLM_REQUEST_TIMEOUT";
+/** The provider options key the SDK forwards into the request body. */
+const PROVIDER_OPTIONS_KEY = "openai-compatible";
+/**
+* Convert provider token accounting into disjoint AI SDK usage. OpenAI's
+* `prompt_tokens_details.cached_tokens` and the DeepSeek dialect's
+* `prompt_cache_hit_tokens` both report cache reads folded into
+* `prompt_tokens`; the harness convention is disjoint counts, so cache reads
+* are split out regardless of which field the endpoint used.
+*/
+function convertUsage(usage) {
+	if (usage == null) return {
+		inputTokens: {
+			total: 0,
+			noCache: 0,
+			cacheRead: void 0,
+			cacheWrite: void 0
+		},
+		outputTokens: {
+			total: 0,
+			text: void 0,
+			reasoning: void 0
+		}
+	};
+	const promptTokens = usage.prompt_tokens ?? 0;
+	const completionTokens = usage.completion_tokens ?? 0;
+	const cacheRead = usage.prompt_tokens_details?.cached_tokens ?? usage.prompt_cache_hit_tokens ?? 0;
+	const reasoningTokens = usage.completion_tokens_details?.reasoning_tokens ?? 0;
+	return {
+		inputTokens: {
+			total: promptTokens,
+			noCache: Math.max(0, promptTokens - cacheRead),
+			cacheRead,
+			cacheWrite: void 0
+		},
+		outputTokens: {
+			total: completionTokens,
+			text: Math.max(0, completionTokens - reasoningTokens),
+			reasoning: reasoningTokens
+		}
+	};
+}
+/** The first real blocks of one resolve: display + catalog metadata. */
+function modelInfo(profile, model) {
+	return {
+		provider: profile.provider,
+		id: model.id,
+		name: model.name ?? model.id,
+		...model.description === void 0 ? {} : { description: model.description },
+		inputModalities: [...model.inputModalities]
+	};
+}
+/** The harness reasoning info for one model, honoring its declared efforts. */
+function reasoningInfo(model, defaultEffort) {
+	const declaration = model?.reasoningEfforts;
+	if (declaration === void 0 || declaration === false) return {};
+	return { reasoning: {
+		efforts: Object.entries(declaration).map(([id]) => ({
+			id: ReasoningEffortId(id),
+			name: `${id.charAt(0).toUpperCase()}${id.slice(1)}`
+		})),
+		...defaultEffort !== void 0 && declaration[defaultEffort] !== void 0 ? { defaultEffort: ReasoningEffortId(defaultEffort) } : {}
+	} };
+}
+/**
+* Map an HTTP status to a stable LlmError code.
+* @param status - status of a non-2xx provider response.
+* @param error - parsed provider error body, when available.
+* @returns the normalized harness error code.
+*/
+function httpErrorCode(status, error) {
+	if (status === 401 || status === 403) return "AUTH";
+	if (status === 413) return "INVALID_REQUEST";
+	const detail = [
+		error?.code,
+		error?.type,
+		error?.message
+	].filter(Boolean).join(" ");
+	if (isQuotaExceededError(detail)) return QUOTA_EXCEEDED_CODE;
+	if (status === 429) return "RATE_LIMIT";
+	if (status === 400) {
+		if (isContextWindowExceededError(detail)) return CONTEXT_WINDOW_EXCEEDED_CODE;
+		return "INVALID_REQUEST";
+	}
+	if (status >= 500) return "SERVER";
+	return `HTTP_${status}`;
+}
+/** Parse a provider error body out of an API-call error's JSON response body. */
+function providerErrorBody(error) {
+	if (error.responseBody === void 0) return void 0;
+	try {
+		return JSON.parse(error.responseBody).error;
+	} catch {
+		return;
+	}
+}
+/** Extract a provider-issued request id from response headers when present. */
+function requestId(headers) {
+	if (headers === void 0) return void 0;
+	const value = headers["x-request-id"] ?? headers["x-openai-compatible-request-id"];
+	return value === void 0 || value.length === 0 ? void 0 : ProviderRequestId(value);
+}
+/**
+* Multi-provider adapter. Each operation reads the current profiles, so a
+* configuration change reaches the next request without a restart; the
+* underlying SDK provider instance is cached per resolved profile and rebuilt
+* when the profile object changes.
+*/
+var OpenAICompatibleAdapter = class extends LlmAdapter {
+	config;
+	sdkProviders = /* @__PURE__ */ new Map();
+	constructor(config) {
+		super();
+		this.config = config;
+	}
+	/** The profile for one route, or the not-owned failure. */
+	profileOf(provider) {
+		const profile = this.config.profiles().get(provider);
+		if (profile === void 0) throw new LlmError(`OpenAI-compatible adapter does not own provider "${provider}"`, "NO_ADAPTER");
+		return profile;
+	}
+	/** The configured descriptor for one exact route/model pair; unlisted ids pass through. */
+	modelOf(profile, model) {
+		return profile.models.find((entry) => entry.id === model);
+	}
+	/** The SDK chat model for one route/model, cached per resolved profile. */
+	sdkModel(profile, modelId) {
+		let byModel = this.sdkProviders.get(profile);
+		if (byModel === void 0) {
+			byModel = /* @__PURE__ */ new Map();
+			this.sdkProviders.set(profile, byModel);
+		}
+		let model = byModel.get(modelId);
+		if (model === void 0) {
+			model = createOpenAICompatible({
+				name: PROVIDER_OPTIONS_KEY,
+				baseURL: profile.baseURL,
+				headers: {
+					...profile.headers,
+					...attributionHeaders()
+				},
+				includeUsage: true,
+				convertUsage
+			}).chatModel(modelId);
+			byModel.set(modelId, model);
+		}
+		return model;
+	}
+	providerInfo(provider) {
+		return {
+			id: provider,
+			name: this.config.profiles().get(provider)?.displayName ?? provider
+		};
+	}
+	providerRetryPolicy(provider) {
+		return this.config.profiles().get(provider)?.retryPolicy;
+	}
+	listModels(provider) {
+		const profile = this.profileOf(provider);
+		return Promise.resolve(profile.models.map((model) => modelInfo(profile, model)));
+	}
+	resolveModel(provider, model, _signal) {
+		const profile = this.profileOf(provider);
+		const configured = this.modelOf(profile, model);
+		const contextWindow = configured?.contextWindow ?? profile.defaultContextWindow;
+		const maxTokens = configured?.maxTokens ?? profile.defaultMaxTokens;
+		return Promise.resolve({
+			...configured === void 0 ? {
+				provider,
+				id: model,
+				name: model,
+				inputModalities: ["text"]
+			} : modelInfo(profile, configured),
+			context: { contextWindow },
+			...maxTokens !== void 0 ? { defaultMaxTokens: maxTokens } : {},
+			...reasoningInfo(configured, profile.reasoning)
+		});
+	}
+	async *stream(options) {
+		const profile = this.profileOf(options.provider);
+		const model = this.modelOf(profile, options.model);
+		const hasImages = options.messages.some((message) => contentHasImage(message.content));
+		let attachments;
+		if (hasImages) {
+			if (model?.inputModalities.includes("image") !== true) throw new LlmError(`OpenAI-compatible model "${options.model}" does not accept image input.`, "UNSUPPORTED_CONTENT");
+			attachments = this.config.resolveAttachments?.();
+			if (attachments === void 0) throw new LlmError("OpenAI-compatible image conversion requires the durable attachment service.", "UNSUPPORTED_CONTENT");
+		}
+		const apiKey = await this.config.resolveApiKey(options.provider, profile);
+		const userId = this.config.resolveUserId();
+		const consumer = new AbortController();
+		const upstream = options.signal === void 0 ? consumer.signal : AbortSignal.any([options.signal, consumer.signal]);
+		const overall = profile.timeoutMs === void 0 ? void 0 : deadline(upstream, profile.timeoutMs, REQUEST_TIMEOUT_CODE);
+		const watchdog = idleWatchdog(overall?.signal ?? upstream, profile.streamIdleTimeoutMs, STREAM_IDLE_TIMEOUT_CODE);
+		try {
+			const callOptions = attachments === void 0 ? await serializeCallOptions(options, profile, model) : await serializeCallOptionsWithImages(options, profile, model, {
+				attachments,
+				maxRequestImageBytes: profile.maxRequestImageBytes,
+				signal: watchdog.signal
+			});
+			const sdkModel = this.sdkModel(profile, options.model);
+			let result;
+			try {
+				result = await sdkModel.doStream({
+					...callOptions,
+					abortSignal: watchdog.signal,
+					headers: {
+						...apiKey === void 0 ? {} : { authorization: `Bearer ${apiKey}` },
+						"x-openai-compatible-harness-user-id": String(userId),
+						...options.sessionId !== void 0 ? { "x-openai-compatible-harness-session-id": String(options.sessionId) } : {},
+						...options.purpose === "compaction" ? { "x-openai-compatible-harness-compact": "1" } : {}
+					}
+				});
+			} catch (error) {
+				throw this.normalizeTransportError(error, profile);
+			}
+			const iterator = translate(result.stream)[Symbol.asyncIterator]();
+			let exhausted = false;
+			try {
+				while (true) {
+					const next = await watchdog.next(iterator);
+					if (next.done) {
+						exhausted = true;
+						return;
+					}
+					yield next.value;
+				}
+			} catch (error) {
+				if (timeoutOf(watchdog.signal, "LLM_STREAM_IDLE_TIMEOUT") !== void 0) throw new LlmError(`OpenAI-compatible stream idle timeout after ${profile.streamIdleTimeoutMs}ms`, "TIMEOUT", { cause: error });
+				if (profile.timeoutMs !== void 0 && timeoutOf(watchdog.signal, "LLM_REQUEST_TIMEOUT") !== void 0) throw new LlmError(`OpenAI-compatible request timeout after ${profile.timeoutMs}ms`, "TIMEOUT", { cause: error });
+				if (options.signal?.aborted) throw new LlmError("OpenAI-compatible request aborted by caller", "ABORTED", { cause: error });
+				if (error instanceof LlmError) throw error;
+				throw this.normalizeTransportError(error, profile);
+			} finally {
+				consumer.abort("OpenAI-compatible stream consumer stopped");
+				if (!exhausted) try {
+					await iterator.return(void 0);
+				} catch {}
+			}
+		} finally {
+			watchdog[Symbol.dispose]();
+			overall?.[Symbol.dispose]();
+		}
+	}
+	/** Normalize an SDK/transport failure into a harness LlmError. */
+	normalizeTransportError(error, profile) {
+		if (error instanceof LlmError) return error;
+		if (APICallError.isInstance(error)) {
+			const providerError = providerErrorBody(error);
+			const message = typeof providerError?.message === "string" ? providerError.message : error.message;
+			const id = requestId(error.responseHeaders);
+			return new LlmError(message, httpErrorCode(error.statusCode ?? 0, providerError), {
+				...error.statusCode === void 0 ? {} : { status: error.statusCode },
+				...id === void 0 ? {} : { requestId: id },
+				cause: error
+			});
+		}
+		if (error instanceof Error) return new LlmError(`OpenAI-compatible API request to ${profile.baseURL} failed`, "TRANSPORT", { cause: error });
+		return new LlmError(`OpenAI-compatible API request to ${profile.baseURL} failed`, "TRANSPORT");
+	}
+};
+//#endregion
+export { DEFAULT_CONTEXT_WINDOW, DEFAULT_MAX_REQUEST_IMAGE_BYTES, DEFAULT_MAX_TOKENS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, OpenAICompatibleAdapter, PROVIDER_OPTIONS_KEY, REQUEST_TIMEOUT_CODE, STREAM_IDLE_TIMEOUT_CODE, convertUsage, httpErrorCode };
+
+//# sourceMappingURL=adapter.mjs.map

package/lib/adapter.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"adapter.mjs","names":[],"sources":["../src/adapter.ts"],"sourcesContent":["/**\n * `OpenAICompatibleAdapter`: a multi-provider harness adapter built on\n * `@ai-sdk/openai-compatible`. One instance serves every provider route in\n * the plugin's `providers` dict; profile facts arrive through a thunk resolved\n * once per operation, so the registering plugin owns validation, layering, and\n * credential policy, and a changed profile reaches the next request without a\n * restart. The SDK owns wire serialization and SSE parsing; this adapter owns\n * harness message conversion, sampling-default merging (via\n * `serialize.ts`), chunk translation (via `translate.ts`), and error\n * normalization.\n * @module dsh-llm-openai-compatible/adapter\n */\n\nimport {\n  CONTEXT_WINDOW_EXCEEDED_CODE,\n  QUOTA_EXCEEDED_CODE,\n  LlmAdapter,\n  LlmError,\n  ProviderRequestId,\n  ReasoningEffortId,\n  attributionHeaders,\n  contentHasImage,\n  isContextWindowExceededError,\n  isQuotaExceededError,\n} from \"@deepseek-ai/dsh-llm\";\nimport type {\n  GenerateOptions,\n  LlmModelInfo,\n  LlmProviderInfo,\n  LlmResolvedModelInfo,\n  ModelModality,\n  ResolvedRetryPolicy,\n  StreamChunk,\n} from \"@deepseek-ai/dsh-llm\";\nimport type { AttachmentStore } from \"@deepseek-ai/dsh-attachment\";\nimport type { CredentialRef } from \"@deepseek-ai/dsh-credentials\";\nimport { deadline, idleWatchdog, timeoutOf } from \"@deepseek-ai/dsh-timeout\";\nimport { APICallError } from \"@ai-sdk/provider\";\nimport type { LanguageModelV4, LanguageModelV4Usage } from \"@ai-sdk/provider\";\nimport { createOpenAICompatible } from \"@ai-sdk/openai-compatible\";\nimport { serializeCallOptions, serializeCallOptionsWithImages } from \"./serialize.ts\";\nimport { translate } from \"./translate.ts\";\n\n/** Default maximum idle interval while an adapter stream read is outstanding. */\nexport const DEFAULT_STREAM_IDLE_TIMEOUT_MS = 300_000;\n/** Default combined request/response context capacity for unconfigured models. */\nexport const DEFAULT_CONTEXT_WINDOW = 262_144;\n/** Default per-request output-token cap for unconfigured models. */\nexport const DEFAULT_MAX_TOKENS = 32_768;\n/** Default bound on accumulated base64 image payload per request. */\nexport const DEFAULT_MAX_REQUEST_IMAGE_BYTES = 20 * 1024 * 1024;\n/** Code stamped on the idle-watchdog timeout reason. */\nexport const STREAM_IDLE_TIMEOUT_CODE = \"LLM_STREAM_IDLE_TIMEOUT\";\n/** Code stamped on the whole-request deadline timeout reason. */\nexport const REQUEST_TIMEOUT_CODE = \"LLM_REQUEST_TIMEOUT\";\n/** The provider options key the SDK forwards into the request body. */\nexport const PROVIDER_OPTIONS_KEY = \"openai-compatible\";\n\n/** Selectable reasoning effort levels for one provider route. */\nexport type ReasoningEffort = \"off\" | \"low\" | \"high\" | \"max\";\n\n/** One validated catalog model of a provider route. */\nexport interface ResolvedModelProfile {\n  id: string;\n  name?: string;\n  description?: string;\n  contextWindow?: number;\n  maxTokens?: number;\n  inputModalities: readonly ModelModality[];\n  /**\n   * Declared reasoning efforts: key = selectable level, value = wire\n   * `reasoning_effort` spelling. `false` rejects the capability outright;\n   * absent means the model carries no reasoning metadata. A `null` wire\n   * spelling (only legal for `off`) means \"omit the field\".\n   */\n  reasoningEfforts?: false | Partial<Record<ReasoningEffort, string | null>>;\n}\n\n/** One validated provider route profile, detached and ready for per-request reads. */\nexport interface ResolvedProviderProfile {\n  provider: string;\n  displayName: string;\n  /** Credential reference; absence means the route sends no authorization header. */\n  apiKeyEnv?: CredentialRef;\n  /** Required endpoint base; requests hit `${baseURL}/chat/completions`. */\n  baseURL: string;\n  headers?: Readonly<Record<string, string>>;\n  // === sampling defaults (request-level values win) ===\n  temperature?: number;\n  topP?: number;\n  topK?: number;\n  presencePenalty?: number;\n  frequencyPenalty?: number;\n  seed?: number;\n  /** Deployment default reasoning level; omission keeps the provider default. */\n  reasoning?: ReasoningEffort;\n  // === model catalog ===\n  models: readonly ResolvedModelProfile[];\n  defaultContextWindow: number;\n  defaultMaxTokens: number;\n  // === transport ===\n  maxRequestImageBytes: number;\n  streamIdleTimeoutMs: number;\n  /** Whole-request deadline in milliseconds; unset arms no overall timer. */\n  timeoutMs?: number;\n  retryPolicy: ResolvedRetryPolicy;\n}\n\n/** Constructor options for {@link OpenAICompatibleAdapter}: the hooks the plugin owns. */\nexport interface OpenAICompatibleAdapterOptions {\n  /** Current validated profiles by provider route; called once per operation. */\n  profiles: () => ReadonlyMap<string, ResolvedProviderProfile>;\n  /**\n   * Resolve the credential for one already-resolved profile; called once per\n   * stream call and frozen for that call. `undefined` means the route sends no\n   * authorization header (an unauthenticated endpoint such as local Ollama).\n   */\n  resolveApiKey: (\n    provider: string,\n    profile: ResolvedProviderProfile,\n  ) => Promise<string | undefined>;\n  /** Resolve the harness anonymous user id for request attribution headers. */\n  resolveUserId: () => string;\n  /** Resolve the optional durable attachment service at request time. */\n  resolveAttachments?: () => AttachmentStore | undefined;\n}\n\n/** The wire usage shape the SDK converter receives (subset of the OpenAI shape). */\ninterface WireUsageLike {\n  prompt_tokens?: number | null | undefined;\n  completion_tokens?: number | null | undefined;\n  prompt_tokens_details?: { cached_tokens?: number | null | undefined } | null | undefined;\n  /** DeepSeek-dialect cache hits folded into prompt_tokens. */\n  prompt_cache_hit_tokens?: number | null | undefined;\n  completion_tokens_details?: { reasoning_tokens?: number | null | undefined } | null | undefined;\n}\n\n/**\n * Convert provider token accounting into disjoint AI SDK usage. OpenAI's\n * `prompt_tokens_details.cached_tokens` and the DeepSeek dialect's\n * `prompt_cache_hit_tokens` both report cache reads folded into\n * `prompt_tokens`; the harness convention is disjoint counts, so cache reads\n * are split out regardless of which field the endpoint used.\n */\nexport function convertUsage(usage: WireUsageLike | null | undefined): LanguageModelV4Usage {\n  if (usage == null) {\n    return {\n      inputTokens: { total: 0, noCache: 0, cacheRead: void 0, cacheWrite: void 0 },\n      outputTokens: { total: 0, text: void 0, reasoning: void 0 },\n    };\n  }\n  const promptTokens = usage.prompt_tokens ?? 0;\n  const completionTokens = usage.completion_tokens ?? 0;\n  const cacheRead =\n    usage.prompt_tokens_details?.cached_tokens ?? usage.prompt_cache_hit_tokens ?? 0;\n  const reasoningTokens = usage.completion_tokens_details?.reasoning_tokens ?? 0;\n  return {\n    inputTokens: {\n      total: promptTokens,\n      noCache: Math.max(0, promptTokens - cacheRead),\n      cacheRead,\n      cacheWrite: void 0,\n    },\n    outputTokens: {\n      total: completionTokens,\n      text: Math.max(0, completionTokens - reasoningTokens),\n      reasoning: reasoningTokens,\n    },\n  };\n}\n\n/** The first real blocks of one resolve: display + catalog metadata. */\nfunction modelInfo(profile: ResolvedProviderProfile, model: ResolvedModelProfile): LlmModelInfo {\n  return {\n    provider: profile.provider,\n    id: model.id,\n    name: model.name ?? model.id,\n    ...(model.description === void 0 ? {} : { description: model.description }),\n    inputModalities: [...model.inputModalities],\n  };\n}\n\n/** The harness reasoning info for one model, honoring its declared efforts. */\nfunction reasoningInfo(\n  model: ResolvedModelProfile | undefined,\n  defaultEffort: ReasoningEffort | undefined,\n): Pick<LlmResolvedModelInfo, \"reasoning\"> {\n  const declaration = model?.reasoningEfforts;\n  if (declaration === void 0 || declaration === false) return {};\n  const entries = Object.entries(declaration) as [ReasoningEffort, string | null | undefined][];\n  const efforts = entries.map(([id]) => ({\n    id: ReasoningEffortId(id),\n    name: `${id.charAt(0).toUpperCase()}${id.slice(1)}`,\n  }));\n  return {\n    reasoning: {\n      efforts,\n      // A configured default the model does not declare is silently dropped\n      // here (describing a model must never throw); the request path still\n      // refuses it, which is where a bad deployment default belongs.\n      ...(defaultEffort !== void 0 && declaration[defaultEffort] !== void 0\n        ? { defaultEffort: ReasoningEffortId(defaultEffort) }\n        : {}),\n    },\n  };\n}\n\n/**\n * Map an HTTP status to a stable LlmError code.\n * @param status - status of a non-2xx provider response.\n * @param error - parsed provider error body, when available.\n * @returns the normalized harness error code.\n */\nexport function httpErrorCode(\n  status: number,\n  error?: { code?: unknown; type?: unknown; message?: unknown },\n): string {\n  if (status === 401 || status === 403) return \"AUTH\";\n  if (status === 413) return \"INVALID_REQUEST\";\n  const detail = [error?.code, error?.type, error?.message].filter(Boolean).join(\" \");\n  if (isQuotaExceededError(detail)) return QUOTA_EXCEEDED_CODE;\n  if (status === 429) return \"RATE_LIMIT\";\n  if (status === 400) {\n    if (isContextWindowExceededError(detail)) return CONTEXT_WINDOW_EXCEEDED_CODE;\n    return \"INVALID_REQUEST\";\n  }\n  if (status >= 500) return \"SERVER\";\n  return `HTTP_${status}`;\n}\n\n/** Parse a provider error body out of an API-call error's JSON response body. */\nfunction providerErrorBody(\n  error: APICallError,\n): { code?: unknown; type?: unknown; message?: unknown } | undefined {\n  if (error.responseBody === void 0) return void 0;\n  try {\n    const parsed = JSON.parse(error.responseBody) as {\n      error?: { code?: unknown; type?: unknown; message?: unknown };\n    };\n    return parsed.error;\n  } catch {\n    return void 0;\n  }\n}\n\n/** Extract a provider-issued request id from response headers when present. */\nfunction requestId(headers: Record<string, string> | undefined): ProviderRequestId | undefined {\n  if (headers === void 0) return void 0;\n  const value = headers[\"x-request-id\"] ?? headers[\"x-openai-compatible-request-id\"];\n  return value === void 0 || value.length === 0 ? void 0 : ProviderRequestId(value);\n}\n\n/**\n * Multi-provider adapter. Each operation reads the current profiles, so a\n * configuration change reaches the next request without a restart; the\n * underlying SDK provider instance is cached per resolved profile and rebuilt\n * when the profile object changes.\n */\nexport class OpenAICompatibleAdapter extends LlmAdapter {\n  private readonly config: OpenAICompatibleAdapterOptions;\n  private readonly sdkProviders = new Map<ResolvedProviderProfile, Map<string, LanguageModelV4>>();\n\n  constructor(config: OpenAICompatibleAdapterOptions) {\n    super();\n    this.config = config;\n  }\n\n  /** The profile for one route, or the not-owned failure. */\n  private profileOf(provider: string): ResolvedProviderProfile {\n    const profile = this.config.profiles().get(provider);\n    if (profile === void 0)\n      throw new LlmError(\n        `OpenAI-compatible adapter does not own provider \"${provider}\"`,\n        \"NO_ADAPTER\",\n      );\n    return profile;\n  }\n\n  /** The configured descriptor for one exact route/model pair; unlisted ids pass through. */\n  private modelOf(\n    profile: ResolvedProviderProfile,\n    model: string,\n  ): ResolvedModelProfile | undefined {\n    return profile.models.find((entry) => entry.id === model);\n  }\n\n  /** The SDK chat model for one route/model, cached per resolved profile. */\n  private sdkModel(profile: ResolvedProviderProfile, modelId: string): LanguageModelV4 {\n    let byModel = this.sdkProviders.get(profile);\n    if (byModel === void 0) {\n      byModel = new Map();\n      this.sdkProviders.set(profile, byModel);\n    }\n    let model = byModel.get(modelId);\n    if (model === void 0) {\n      const provider = createOpenAICompatible({\n        name: PROVIDER_OPTIONS_KEY,\n        baseURL: profile.baseURL,\n        headers: { ...profile.headers, ...attributionHeaders() },\n        includeUsage: true,\n        convertUsage,\n      });\n      model = provider.chatModel(modelId);\n      byModel.set(modelId, model);\n    }\n    return model;\n  }\n\n  providerInfo(provider: string): LlmProviderInfo {\n    return {\n      id: provider,\n      name: this.config.profiles().get(provider)?.displayName ?? provider,\n    };\n  }\n\n  providerRetryPolicy(provider: string): ResolvedRetryPolicy | undefined {\n    return this.config.profiles().get(provider)?.retryPolicy;\n  }\n\n  listModels(provider: string): Promise<readonly LlmModelInfo[]> {\n    const profile = this.profileOf(provider);\n    return Promise.resolve(profile.models.map((model) => modelInfo(profile, model)));\n  }\n\n  resolveModel(\n    provider: string,\n    model: string,\n    _signal?: AbortSignal,\n  ): Promise<LlmResolvedModelInfo> {\n    const profile = this.profileOf(provider);\n    const configured = this.modelOf(profile, model);\n    const contextWindow = configured?.contextWindow ?? profile.defaultContextWindow;\n    const maxTokens = configured?.maxTokens ?? profile.defaultMaxTokens;\n    return Promise.resolve({\n      ...(configured === void 0\n        ? { provider, id: model, name: model, inputModalities: [\"text\" as const] }\n        : modelInfo(profile, configured)),\n      context: { contextWindow },\n      ...(maxTokens !== void 0 ? { defaultMaxTokens: maxTokens } : {}),\n      ...reasoningInfo(configured, profile.reasoning),\n    });\n  }\n\n  async *stream(options: GenerateOptions): AsyncGenerator<StreamChunk> {\n    const profile = this.profileOf(options.provider);\n    const model = this.modelOf(profile, options.model);\n    const hasImages = options.messages.some((message) => contentHasImage(message.content));\n    let attachments: AttachmentStore | undefined;\n    if (hasImages) {\n      if (model?.inputModalities.includes(\"image\") !== true) {\n        throw new LlmError(\n          `OpenAI-compatible model \"${options.model}\" does not accept image input.`,\n          \"UNSUPPORTED_CONTENT\",\n        );\n      }\n      attachments = this.config.resolveAttachments?.();\n      if (attachments === void 0)\n        throw new LlmError(\n          \"OpenAI-compatible image conversion requires the durable attachment service.\",\n          \"UNSUPPORTED_CONTENT\",\n        );\n    }\n    const apiKey = await this.config.resolveApiKey(options.provider, profile);\n    const userId = this.config.resolveUserId();\n    const consumer = new AbortController();\n    const upstream =\n      options.signal === void 0\n        ? consumer.signal\n        : AbortSignal.any([options.signal, consumer.signal]);\n    const overall =\n      profile.timeoutMs === void 0\n        ? void 0\n        : deadline(upstream, profile.timeoutMs, REQUEST_TIMEOUT_CODE);\n    const watchdog = idleWatchdog(\n      overall?.signal ?? upstream,\n      profile.streamIdleTimeoutMs,\n      STREAM_IDLE_TIMEOUT_CODE,\n    );\n    try {\n      const callOptions =\n        attachments === void 0\n          ? await serializeCallOptions(options, profile, model)\n          : await serializeCallOptionsWithImages(options, profile, model, {\n              attachments,\n              maxRequestImageBytes: profile.maxRequestImageBytes,\n              signal: watchdog.signal,\n            });\n      const sdkModel = this.sdkModel(profile, options.model);\n      let result;\n      try {\n        result = await sdkModel.doStream({\n          ...callOptions,\n          abortSignal: watchdog.signal,\n          headers: {\n            ...(apiKey === void 0 ? {} : { authorization: `Bearer ${apiKey}` }),\n            \"x-openai-compatible-harness-user-id\": String(userId),\n            ...(options.sessionId !== void 0\n              ? { \"x-openai-compatible-harness-session-id\": String(options.sessionId) }\n              : {}),\n            ...(options.purpose === \"compaction\"\n              ? { \"x-openai-compatible-harness-compact\": \"1\" }\n              : {}),\n          },\n        });\n      } catch (error) {\n        throw this.normalizeTransportError(error, profile);\n      }\n      const iterator = translate(result.stream)[Symbol.asyncIterator]();\n      let exhausted = false;\n      try {\n        while (true) {\n          const next = await watchdog.next(iterator);\n          if (next.done) {\n            exhausted = true;\n            return;\n          }\n          yield next.value;\n        }\n      } catch (error) {\n        if (timeoutOf(watchdog.signal, STREAM_IDLE_TIMEOUT_CODE) !== void 0) {\n          throw new LlmError(\n            `OpenAI-compatible stream idle timeout after ${profile.streamIdleTimeoutMs}ms`,\n            \"TIMEOUT\",\n            { cause: error },\n          );\n        }\n        if (\n          profile.timeoutMs !== void 0 &&\n          timeoutOf(watchdog.signal, REQUEST_TIMEOUT_CODE) !== void 0\n        ) {\n          throw new LlmError(\n            `OpenAI-compatible request timeout after ${profile.timeoutMs}ms`,\n            \"TIMEOUT\",\n            { cause: error },\n          );\n        }\n        if (options.signal?.aborted)\n          throw new LlmError(\"OpenAI-compatible request aborted by caller\", \"ABORTED\", {\n            cause: error,\n          });\n        if (error instanceof LlmError) throw error;\n        throw this.normalizeTransportError(error, profile);\n      } finally {\n        consumer.abort(\"OpenAI-compatible stream consumer stopped\");\n        if (!exhausted) {\n          try {\n            await iterator.return(void 0);\n          } catch {\n            // The transport already aborted; teardown is best-effort.\n          }\n        }\n      }\n    } finally {\n      watchdog[Symbol.dispose]();\n      overall?.[Symbol.dispose]();\n    }\n  }\n\n  /** Normalize an SDK/transport failure into a harness LlmError. */\n  private normalizeTransportError(error: unknown, profile: ResolvedProviderProfile): LlmError {\n    if (error instanceof LlmError) return error;\n    if (APICallError.isInstance(error)) {\n      const providerError = providerErrorBody(error);\n      const message =\n        typeof providerError?.message === \"string\" ? providerError.message : error.message;\n      const id = requestId(error.responseHeaders);\n      return new LlmError(message, httpErrorCode(error.statusCode ?? 0, providerError), {\n        ...(error.statusCode === void 0 ? {} : { status: error.statusCode }),\n        ...(id === void 0 ? {} : { requestId: id }),\n        cause: error,\n      });\n    }\n    if (error instanceof Error) {\n      return new LlmError(\n        `OpenAI-compatible API request to ${profile.baseURL} failed`,\n        \"TRANSPORT\",\n        { cause: error },\n      );\n    }\n    return new LlmError(`OpenAI-compatible API request to ${profile.baseURL} failed`, \"TRANSPORT\");\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AA4CA,MAAa,iCAAiC;;AAE9C,MAAa,yBAAyB;;AAEtC,MAAa,qBAAqB;;AAElC,MAAa,kCAAkC;;AAE/C,MAAa,2BAA2B;;AAExC,MAAa,uBAAuB;;AAEpC,MAAa,uBAAuB;;;;;;;;AAwFpC,SAAgB,aAAa,OAA+D;CAC1F,IAAI,SAAS,MACX,OAAO;EACL,aAAa;GAAE,OAAO;GAAG,SAAS;GAAG,WAAW,KAAK;GAAG,YAAY,KAAK;EAAE;EAC3E,cAAc;GAAE,OAAO;GAAG,MAAM,KAAK;GAAG,WAAW,KAAK;EAAE;CAC5D;CAEF,MAAM,eAAe,MAAM,iBAAiB;CAC5C,MAAM,mBAAmB,MAAM,qBAAqB;CACpD,MAAM,YACJ,MAAM,uBAAuB,iBAAiB,MAAM,2BAA2B;CACjF,MAAM,kBAAkB,MAAM,2BAA2B,oBAAoB;CAC7E,OAAO;EACL,aAAa;GACX,OAAO;GACP,SAAS,KAAK,IAAI,GAAG,eAAe,SAAS;GAC7C;GACA,YAAY,KAAK;EACnB;EACA,cAAc;GACZ,OAAO;GACP,MAAM,KAAK,IAAI,GAAG,mBAAmB,eAAe;GACpD,WAAW;EACb;CACF;AACF;;AAGA,SAAS,UAAU,SAAkC,OAA2C;CAC9F,OAAO;EACL,UAAU,QAAQ;EAClB,IAAI,MAAM;EACV,MAAM,MAAM,QAAQ,MAAM;EAC1B,GAAI,MAAM,gBAAgB,KAAK,IAAI,CAAC,IAAI,EAAE,aAAa,MAAM,YAAY;EACzE,iBAAiB,CAAC,GAAG,MAAM,eAAe;CAC5C;AACF;;AAGA,SAAS,cACP,OACA,eACyC;CACzC,MAAM,cAAc,OAAO;CAC3B,IAAI,gBAAgB,KAAK,KAAK,gBAAgB,OAAO,OAAO,CAAC;CAM7D,OAAO,EACL,WAAW;EACT,SAPY,OAAO,QAAQ,WACT,CAAC,CAAC,KAAK,CAAC,SAAS;GACrC,IAAI,kBAAkB,EAAE;GACxB,MAAM,GAAG,GAAG,OAAO,CAAC,CAAC,CAAC,YAAY,IAAI,GAAG,MAAM,CAAC;EAClD,EAGU;EAIN,GAAI,kBAAkB,KAAK,KAAK,YAAY,mBAAmB,KAAK,IAChE,EAAE,eAAe,kBAAkB,aAAa,EAAE,IAClD,CAAC;CACP,EACF;AACF;;;;;;;AAQA,SAAgB,cACd,QACA,OACQ;CACR,IAAI,WAAW,OAAO,WAAW,KAAK,OAAO;CAC7C,IAAI,WAAW,KAAK,OAAO;CAC3B,MAAM,SAAS;EAAC,OAAO;EAAM,OAAO;EAAM,OAAO;CAAO,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK,GAAG;CAClF,IAAI,qBAAqB,MAAM,GAAG,OAAO;CACzC,IAAI,WAAW,KAAK,OAAO;CAC3B,IAAI,WAAW,KAAK;EAClB,IAAI,6BAA6B,MAAM,GAAG,OAAO;EACjD,OAAO;CACT;CACA,IAAI,UAAU,KAAK,OAAO;CAC1B,OAAO,QAAQ;AACjB;;AAGA,SAAS,kBACP,OACmE;CACnE,IAAI,MAAM,iBAAiB,KAAK,GAAG,OAAO,KAAK;CAC/C,IAAI;EAIF,OAHe,KAAK,MAAM,MAAM,YAGpB,CAAC,CAAC;CAChB,QAAQ;EACN;CACF;AACF;;AAGA,SAAS,UAAU,SAA4E;CAC7F,IAAI,YAAY,KAAK,GAAG,OAAO,KAAK;CACpC,MAAM,QAAQ,QAAQ,mBAAmB,QAAQ;CACjD,OAAO,UAAU,KAAK,KAAK,MAAM,WAAW,IAAI,KAAK,IAAI,kBAAkB,KAAK;AAClF;;;;;;;AAQA,IAAa,0BAAb,cAA6C,WAAW;CACtD;CACA,+BAAgC,IAAI,IAA2D;CAE/F,YAAY,QAAwC;EAClD,MAAM;EACN,KAAK,SAAS;CAChB;;CAGA,UAAkB,UAA2C;EAC3D,MAAM,UAAU,KAAK,OAAO,SAAS,CAAC,CAAC,IAAI,QAAQ;EACnD,IAAI,YAAY,KAAK,GACnB,MAAM,IAAI,SACR,oDAAoD,SAAS,IAC7D,YACF;EACF,OAAO;CACT;;CAGA,QACE,SACA,OACkC;EAClC,OAAO,QAAQ,OAAO,MAAM,UAAU,MAAM,OAAO,KAAK;CAC1D;;CAGA,SAAiB,SAAkC,SAAkC;EACnF,IAAI,UAAU,KAAK,aAAa,IAAI,OAAO;EAC3C,IAAI,YAAY,KAAK,GAAG;GACtB,0BAAU,IAAI,IAAI;GAClB,KAAK,aAAa,IAAI,SAAS,OAAO;EACxC;EACA,IAAI,QAAQ,QAAQ,IAAI,OAAO;EAC/B,IAAI,UAAU,KAAK,GAAG;GAQpB,QAPiB,uBAAuB;IACtC,MAAM;IACN,SAAS,QAAQ;IACjB,SAAS;KAAE,GAAG,QAAQ;KAAS,GAAG,mBAAmB;IAAE;IACvD,cAAc;IACd;GACF,CACe,CAAC,CAAC,UAAU,OAAO;GAClC,QAAQ,IAAI,SAAS,KAAK;EAC5B;EACA,OAAO;CACT;CAEA,aAAa,UAAmC;EAC9C,OAAO;GACL,IAAI;GACJ,MAAM,KAAK,OAAO,SAAS,CAAC,CAAC,IAAI,QAAQ,CAAC,EAAE,eAAe;EAC7D;CACF;CAEA,oBAAoB,UAAmD;EACrE,OAAO,KAAK,OAAO,SAAS,CAAC,CAAC,IAAI,QAAQ,CAAC,EAAE;CAC/C;CAEA,WAAW,UAAoD;EAC7D,MAAM,UAAU,KAAK,UAAU,QAAQ;EACvC,OAAO,QAAQ,QAAQ,QAAQ,OAAO,KAAK,UAAU,UAAU,SAAS,KAAK,CAAC,CAAC;CACjF;CAEA,aACE,UACA,OACA,SAC+B;EAC/B,MAAM,UAAU,KAAK,UAAU,QAAQ;EACvC,MAAM,aAAa,KAAK,QAAQ,SAAS,KAAK;EAC9C,MAAM,gBAAgB,YAAY,iBAAiB,QAAQ;EAC3D,MAAM,YAAY,YAAY,aAAa,QAAQ;EACnD,OAAO,QAAQ,QAAQ;GACrB,GAAI,eAAe,KAAK,IACpB;IAAE;IAAU,IAAI;IAAO,MAAM;IAAO,iBAAiB,CAAC,MAAe;GAAE,IACvE,UAAU,SAAS,UAAU;GACjC,SAAS,EAAE,cAAc;GACzB,GAAI,cAAc,KAAK,IAAI,EAAE,kBAAkB,UAAU,IAAI,CAAC;GAC9D,GAAG,cAAc,YAAY,QAAQ,SAAS;EAChD,CAAC;CACH;CAEA,OAAO,OAAO,SAAuD;EACnE,MAAM,UAAU,KAAK,UAAU,QAAQ,QAAQ;EAC/C,MAAM,QAAQ,KAAK,QAAQ,SAAS,QAAQ,KAAK;EACjD,MAAM,YAAY,QAAQ,SAAS,MAAM,YAAY,gBAAgB,QAAQ,OAAO,CAAC;EACrF,IAAI;EACJ,IAAI,WAAW;GACb,IAAI,OAAO,gBAAgB,SAAS,OAAO,MAAM,MAC/C,MAAM,IAAI,SACR,4BAA4B,QAAQ,MAAM,iCAC1C,qBACF;GAEF,cAAc,KAAK,OAAO,qBAAqB;GAC/C,IAAI,gBAAgB,KAAK,GACvB,MAAM,IAAI,SACR,+EACA,qBACF;EACJ;EACA,MAAM,SAAS,MAAM,KAAK,OAAO,cAAc,QAAQ,UAAU,OAAO;EACxE,MAAM,SAAS,KAAK,OAAO,cAAc;EACzC,MAAM,WAAW,IAAI,gBAAgB;EACrC,MAAM,WACJ,QAAQ,WAAW,KAAK,IACpB,SAAS,SACT,YAAY,IAAI,CAAC,QAAQ,QAAQ,SAAS,MAAM,CAAC;EACvD,MAAM,UACJ,QAAQ,cAAc,KAAK,IACvB,KAAK,IACL,SAAS,UAAU,QAAQ,WAAW,oBAAoB;EAChE,MAAM,WAAW,aACf,SAAS,UAAU,UACnB,QAAQ,qBACR,wBACF;EACA,IAAI;GACF,MAAM,cACJ,gBAAgB,KAAK,IACjB,MAAM,qBAAqB,SAAS,SAAS,KAAK,IAClD,MAAM,+BAA+B,SAAS,SAAS,OAAO;IAC5D;IACA,sBAAsB,QAAQ;IAC9B,QAAQ,SAAS;GACnB,CAAC;GACP,MAAM,WAAW,KAAK,SAAS,SAAS,QAAQ,KAAK;GACrD,IAAI;GACJ,IAAI;IACF,SAAS,MAAM,SAAS,SAAS;KAC/B,GAAG;KACH,aAAa,SAAS;KACtB,SAAS;MACP,GAAI,WAAW,KAAK,IAAI,CAAC,IAAI,EAAE,eAAe,UAAU,SAAS;MACjE,uCAAuC,OAAO,MAAM;MACpD,GAAI,QAAQ,cAAc,KAAK,IAC3B,EAAE,0CAA0C,OAAO,QAAQ,SAAS,EAAE,IACtE,CAAC;MACL,GAAI,QAAQ,YAAY,eACpB,EAAE,uCAAuC,IAAI,IAC7C,CAAC;KACP;IACF,CAAC;GACH,SAAS,OAAO;IACd,MAAM,KAAK,wBAAwB,OAAO,OAAO;GACnD;GACA,MAAM,WAAW,UAAU,OAAO,MAAM,CAAC,CAAC,OAAO,cAAc,CAAC;GAChE,IAAI,YAAY;GAChB,IAAI;IACF,OAAO,MAAM;KACX,MAAM,OAAO,MAAM,SAAS,KAAK,QAAQ;KACzC,IAAI,KAAK,MAAM;MACb,YAAY;MACZ;KACF;KACA,MAAM,KAAK;IACb;GACF,SAAS,OAAO;IACd,IAAI,UAAU,SAAS,QAAA,yBAAgC,MAAM,KAAK,GAChE,MAAM,IAAI,SACR,+CAA+C,QAAQ,oBAAoB,KAC3E,WACA,EAAE,OAAO,MAAM,CACjB;IAEF,IACE,QAAQ,cAAc,KAAK,KAC3B,UAAU,SAAS,QAAA,qBAA4B,MAAM,KAAK,GAE1D,MAAM,IAAI,SACR,2CAA2C,QAAQ,UAAU,KAC7D,WACA,EAAE,OAAO,MAAM,CACjB;IAEF,IAAI,QAAQ,QAAQ,SAClB,MAAM,IAAI,SAAS,+CAA+C,WAAW,EAC3E,OAAO,MACT,CAAC;IACH,IAAI,iBAAiB,UAAU,MAAM;IACrC,MAAM,KAAK,wBAAwB,OAAO,OAAO;GACnD,UAAU;IACR,SAAS,MAAM,2CAA2C;IAC1D,IAAI,CAAC,WACH,IAAI;KACF,MAAM,SAAS,OAAO,KAAK,CAAC;IAC9B,QAAQ,CAER;GAEJ;EACF,UAAU;GACR,SAAS,OAAO,QAAQ,CAAC;GACzB,UAAU,OAAO,QAAQ,CAAC;EAC5B;CACF;;CAGA,wBAAgC,OAAgB,SAA4C;EAC1F,IAAI,iBAAiB,UAAU,OAAO;EACtC,IAAI,aAAa,WAAW,KAAK,GAAG;GAClC,MAAM,gBAAgB,kBAAkB,KAAK;GAC7C,MAAM,UACJ,OAAO,eAAe,YAAY,WAAW,cAAc,UAAU,MAAM;GAC7E,MAAM,KAAK,UAAU,MAAM,eAAe;GAC1C,OAAO,IAAI,SAAS,SAAS,cAAc,MAAM,cAAc,GAAG,aAAa,GAAG;IAChF,GAAI,MAAM,eAAe,KAAK,IAAI,CAAC,IAAI,EAAE,QAAQ,MAAM,WAAW;IAClE,GAAI,OAAO,KAAK,IAAI,CAAC,IAAI,EAAE,WAAW,GAAG;IACzC,OAAO;GACT,CAAC;EACH;EACA,IAAI,iBAAiB,OACnB,OAAO,IAAI,SACT,oCAAoC,QAAQ,QAAQ,UACpD,aACA,EAAE,OAAO,MAAM,CACjB;EAEF,OAAO,IAAI,SAAS,oCAAoC,QAAQ,QAAQ,UAAU,WAAW;CAC/F;AACF"}