@graphorin/provider 0.5.0

package/CHANGELOG.md

@@ -0,0 +1,6 @@
+# @graphorin/provider
+
+## 0.1.0
+
+Initial release. See the root `CHANGELOG.md` and the corresponding
+changeset for the full list of features.

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Oleksiy Stepurenko
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,105 @@
+# `@graphorin/provider`
+
+> Vendor-neutral LLM provider layer for the
+> [**Graphorin**](https://github.com/o-stepper/graphorin) framework.
+
+The package owns four moving parts:
+
+1. **`createProvider(...)`** — wraps any adapter in a stable `Provider`
+   shape with sensitivity, capability, and reasoning-retention defaults.
+2. **Adapters** — `vercelAdapter` (default cloud path; wraps the Vercel
+   AI SDK), `ollamaAdapter` (direct Ollama HTTP), `llamaCppServerAdapter`
+   (the upstream `llama-server` binary from llama.cpp), and
+   `openAICompatibleAdapter` (LMStudio / LocalAI / vLLM / Together-style
+   self-host endpoints).
+3. **Middleware** — `composeProviderMiddleware([...])` enforces a
+   canonical order at startup and throws `MiddlewareOrderingError` on
+   violation. Built-ins: `withTracing`, `withRetry`, `withRateLimit`,
+   `withCostLimit`, `withCostTracking`, `withFallback`, and
+   `withRedaction` (mandatory in production).
+4. **Token counting** — pluggable `TokenCounter` dispatcher. Default
+   `JsTiktokenCounter` for OpenAI-compatible models; per-vendor native
+   counters for Anthropic, Google, and Bedrock; heuristic fallback for
+   unknown providers with a one-time WARN.
+
+## Installation
+
+```bash
+pnpm add @graphorin/provider
+# Optional peer for the cloud adapter:
+pnpm add ai
+# Optional peer for the default token counter:
+pnpm add js-tiktoken
+```
+
+For the in-process llama.cpp companion adapter, install the separate
+package:
+
+```bash
+pnpm add @graphorin/provider-llamacpp-node node-llama-cpp
+```
+
+## Quick start
+
+```ts
+import { composeProviderMiddleware, createProvider } from '@graphorin/provider';
+import { vercelAdapter } from '@graphorin/provider/adapters/vercel';
+import {
+  withCostLimit,
+  withFallback,
+  withRateLimit,
+  withRedaction,
+  withRetry,
+  withTracing,
+} from '@graphorin/provider/middleware';
+import { BUILT_IN_PATTERNS } from '@graphorin/observability/redaction/patterns';
+import { openai } from '@ai-sdk/openai';
+
+const provider = createProvider(vercelAdapter(openai('gpt-4o')));
+
+const safeProvider = composeProviderMiddleware([
+  withTracing(),
+  withRetry({ maxRetries: 3 }),
+  withRateLimit({ requestsPerMinute: 60 }),
+  withCostLimit({ maxPerSession: 1.0 }),
+  withFallback([fallbackProvider]),
+  withRedaction({ patterns: BUILT_IN_PATTERNS }), // INNERMOST
+])(provider);
+```
+
+## Local-first stack
+
+```ts
+import { ollamaAdapter } from '@graphorin/provider/adapters/ollama';
+
+// Auto-classified as 'loopback' — no warning, no first-run prompt.
+const local = createProvider(
+  ollamaAdapter({ model: 'llama3.1:8b', baseUrl: 'http://127.0.0.1:11434' }),
+);
+```
+
+The same `LocalProviderTrust` classifier (`'loopback' | 'private' |
+'public-tls' | 'public-cleartext'`) drives the trust auto-detection,
+the sensitivity-tier defaults, and the `withRedaction` policy table
+for every `baseUrl`-driven adapter — `ollamaAdapter`,
+`llamaCppServerAdapter`, and `openAICompatibleAdapter`. The classifier
+lives at `@graphorin/provider/trust`. Public-cleartext URLs refuse to
+start with `LocalProviderInsecureTransportError`.
+
+## Cost-tier vocabulary
+
+```ts
+import type { ModelHint } from '@graphorin/core';
+import { classifyModelTier } from '@graphorin/provider/model-tier';
+
+classifyModelTier(provider); // ⇒ 'fast' | 'balanced' | 'smart' | undefined
+```
+
+The classifier is consumed by the agent runtime (Phase 12) to validate
+operator-supplied per-tier mappings and to surface tier-not-mapped
+recommendations.
+
+## Project metadata
+
+- **Project Graphorin** · v0.5.0 · MIT License · © 2026 Oleksiy Stepurenko
+- Repository: <https://github.com/o-stepper/graphorin>

package/dist/adapters/index.d.ts

@@ -0,0 +1,4 @@
+import { LlamaCppServerAdapterOptions, llamaCppServerAdapter } from "./llamacpp-server.js";
+import { OllamaAdapterOptions, ollamaAdapter } from "./ollama.js";
+import { OpenAICompatibleAdapterOptions, openAICompatibleAdapter } from "./openai-compatible.js";
+import { VercelAdapterOptions, vercelAdapter } from "./vercel.js";

package/dist/adapters/index.js

@@ -0,0 +1,6 @@
+import { llamaCppServerAdapter } from "./llamacpp-server.js";
+import { ollamaAdapter } from "./ollama.js";
+import { openAICompatibleAdapter } from "./openai-compatible.js";
+import { vercelAdapter } from "./vercel.js";
+
+export {  };

package/dist/adapters/llamacpp-server.d.ts

@@ -0,0 +1,69 @@
+import * as _graphorin_core0 from "@graphorin/core";
+import { Provider, Sensitivity } from "@graphorin/core";
+
+//#region src/adapters/llamacpp-server.d.ts
+
+/**
+ * Default port used by the upstream `llama-server` binary.
+ *
+ * @stable
+ */
+declare const DEFAULT_LLAMACPP_SERVER_BASE_URL = "http://127.0.0.1:8080";
+/**
+ * Options accepted by {@link llamaCppServerAdapter}.
+ *
+ * @stable
+ */
+interface LlamaCppServerAdapterOptions {
+  /** GGUF model identifier exposed by the running server (e.g. `'qwen2.5:7b-instruct-q4_k_m'`). */
+  readonly model: string;
+  /** Base URL of the running `llama-server` process. Defaults to `http://127.0.0.1:8080`. */
+  readonly baseUrl?: string;
+  /** Optional bearer-auth API key (`--api-key` flag on the server). */
+  readonly apiKey?: string;
+  /** Extra headers merged on top of `content-type` + `accept` defaults. */
+  readonly headers?: Readonly<Record<string, string>>;
+  /** Custom `fetch` implementation; useful for tests. */
+  readonly fetchImpl?: typeof fetch;
+  /**
+   * Time-to-response budget per request (PS-24). Default
+   * `DEFAULT_REQUEST_TIMEOUT_MS` (120s); `0` disables.
+   */
+  readonly timeoutMs?: number;
+  /** Capability overrides merged on top of the adapter defaults. */
+  readonly capabilities?: Partial<_graphorin_core0.ProviderCapabilities>;
+  /**
+   * Acknowledge the risk of running over plaintext HTTP against a
+   * public host. Without this flag the adapter throws
+   * {@link import('../errors/errors.js').LocalProviderInsecureTransportError}.
+   */
+  readonly allowInsecureTransport?: boolean;
+  /** Override for the default `acceptsSensitivity` value. */
+  readonly acceptsSensitivity?: ReadonlyArray<Sensitivity>;
+  /** Provider name attached to spans / log lines. */
+  readonly name?: string;
+  /** Optional log sink. Tests pass a fixture sink to silence the console. */
+  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;
+}
+/**
+ * Build a Graphorin {@link Provider} backed by the upstream
+ * `llama-server` binary. The factory does not start the binary —
+ * operators launch it themselves with the desired model + GPU flags
+ * and pass the URL here.
+ *
+ * @example
+ * ```ts
+ * const local = createProvider(
+ *   llamaCppServerAdapter({
+ *     model: 'qwen2.5:7b-instruct-q4_k_m',
+ *     baseUrl: 'http://127.0.0.1:8080',
+ *   }),
+ * );
+ * ```
+ *
+ * @stable
+ */
+declare function llamaCppServerAdapter(options: LlamaCppServerAdapterOptions): Provider;
+//#endregion
+export { DEFAULT_LLAMACPP_SERVER_BASE_URL, LlamaCppServerAdapterOptions, llamaCppServerAdapter };
+//# sourceMappingURL=llamacpp-server.d.ts.map

package/dist/adapters/llamacpp-server.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"llamacpp-server.d.ts","names":[],"sources":["../../src/adapters/llamacpp-server.ts"],"sourcesContent":[],"mappings":";;;;;AAgFA;;;;;cAzDa,gCAAA;;;;;;UAOI,4BAAA;;;;;;;;qBAQI,SAAS;;8BAEA;;;;;;;0BAOJ,QAPS,gBAAA,CAOyB,oBAAA;;;;;;;;gCAQ5B,cAAc;;;;;;;;;;;;;;;;;;;;;;;;iBAyB9B,qBAAA,UAA+B,+BAA+B"}

package/dist/adapters/llamacpp-server.js

@@ -0,0 +1,51 @@
+import { buildOpenAIShapedProvider } from "../internal/openai-shaped.js";
+
+//#region src/adapters/llamacpp-server.ts
+/**
+* Default port used by the upstream `llama-server` binary.
+*
+* @stable
+*/
+const DEFAULT_LLAMACPP_SERVER_BASE_URL = "http://127.0.0.1:8080";
+/**
+* Build a Graphorin {@link Provider} backed by the upstream
+* `llama-server` binary. The factory does not start the binary —
+* operators launch it themselves with the desired model + GPU flags
+* and pass the URL here.
+*
+* @example
+* ```ts
+* const local = createProvider(
+*   llamaCppServerAdapter({
+*     model: 'qwen2.5:7b-instruct-q4_k_m',
+*     baseUrl: 'http://127.0.0.1:8080',
+*   }),
+* );
+* ```
+*
+* @stable
+*/
+function llamaCppServerAdapter(options) {
+	const baseUrl = options.baseUrl ?? DEFAULT_LLAMACPP_SERVER_BASE_URL;
+	return buildOpenAIShapedProvider({
+		providerName: options.name ?? `llamacpp-server-${options.model}`,
+		model: options.model,
+		baseUrl,
+		...options.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
+		...options.headers !== void 0 ? { headers: options.headers } : {},
+		...options.fetchImpl !== void 0 ? { fetchImpl: options.fetchImpl } : {},
+		...options.allowInsecureTransport !== void 0 ? { allowInsecureTransport: options.allowInsecureTransport } : {},
+		...options.acceptsSensitivity !== void 0 ? { acceptsSensitivity: options.acceptsSensitivity } : {},
+		...options.logger !== void 0 ? { logger: options.logger } : {},
+		...options.timeoutMs !== void 0 ? { timeoutMs: options.timeoutMs } : {},
+		capabilities: {
+			multimodal: false,
+			parallelToolCalls: false,
+			...options.capabilities
+		}
+	}).provider;
+}
+
+//#endregion
+export { DEFAULT_LLAMACPP_SERVER_BASE_URL, llamaCppServerAdapter };
+//# sourceMappingURL=llamacpp-server.js.map

package/dist/adapters/llamacpp-server.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"llamacpp-server.js","names":[],"sources":["../../src/adapters/llamacpp-server.ts"],"sourcesContent":["/**\n * Direct adapter for the upstream `llama-server` binary from the\n * llama.cpp project. The binary speaks the OpenAI-compatible REST\n * contract end-to-end (`POST /v1/chat/completions`, `POST /v1/completions`,\n * `POST /v1/embeddings`); streaming is via `text/event-stream` chunks\n * terminated by `data: [DONE]` exactly as the upstream OpenAI shape.\n *\n * The adapter shares a single `LocalProviderTrust` classifier with\n * `ollamaAdapter` and `openAICompatibleAdapter` — one classifier, one\n * policy table, one error type.\n *\n * @packageDocumentation\n */\n\nimport type { Provider, Sensitivity } from '@graphorin/core';\n\nimport { buildOpenAIShapedProvider } from '../internal/openai-shaped.js';\n\n/**\n * Default port used by the upstream `llama-server` binary.\n *\n * @stable\n */\nexport const DEFAULT_LLAMACPP_SERVER_BASE_URL = 'http://127.0.0.1:8080';\n\n/**\n * Options accepted by {@link llamaCppServerAdapter}.\n *\n * @stable\n */\nexport interface LlamaCppServerAdapterOptions {\n  /** GGUF model identifier exposed by the running server (e.g. `'qwen2.5:7b-instruct-q4_k_m'`). */\n  readonly model: string;\n  /** Base URL of the running `llama-server` process. Defaults to `http://127.0.0.1:8080`. */\n  readonly baseUrl?: string;\n  /** Optional bearer-auth API key (`--api-key` flag on the server). */\n  readonly apiKey?: string;\n  /** Extra headers merged on top of `content-type` + `accept` defaults. */\n  readonly headers?: Readonly<Record<string, string>>;\n  /** Custom `fetch` implementation; useful for tests. */\n  readonly fetchImpl?: typeof fetch;\n  /**\n   * Time-to-response budget per request (PS-24). Default\n   * `DEFAULT_REQUEST_TIMEOUT_MS` (120s); `0` disables.\n   */\n  readonly timeoutMs?: number;\n  /** Capability overrides merged on top of the adapter defaults. */\n  readonly capabilities?: Partial<import('@graphorin/core').ProviderCapabilities>;\n  /**\n   * Acknowledge the risk of running over plaintext HTTP against a\n   * public host. Without this flag the adapter throws\n   * {@link import('../errors/errors.js').LocalProviderInsecureTransportError}.\n   */\n  readonly allowInsecureTransport?: boolean;\n  /** Override for the default `acceptsSensitivity` value. */\n  readonly acceptsSensitivity?: ReadonlyArray<Sensitivity>;\n  /** Provider name attached to spans / log lines. */\n  readonly name?: string;\n  /** Optional log sink. Tests pass a fixture sink to silence the console. */\n  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;\n}\n\n/**\n * Build a Graphorin {@link Provider} backed by the upstream\n * `llama-server` binary. The factory does not start the binary —\n * operators launch it themselves with the desired model + GPU flags\n * and pass the URL here.\n *\n * @example\n * ```ts\n * const local = createProvider(\n *   llamaCppServerAdapter({\n *     model: 'qwen2.5:7b-instruct-q4_k_m',\n *     baseUrl: 'http://127.0.0.1:8080',\n *   }),\n * );\n * ```\n *\n * @stable\n */\nexport function llamaCppServerAdapter(options: LlamaCppServerAdapterOptions): Provider {\n  const baseUrl = options.baseUrl ?? DEFAULT_LLAMACPP_SERVER_BASE_URL;\n  const providerName = options.name ?? `llamacpp-server-${options.model}`;\n  const built = buildOpenAIShapedProvider({\n    providerName,\n    model: options.model,\n    baseUrl,\n    ...(options.apiKey !== undefined ? { apiKey: options.apiKey } : {}),\n    ...(options.headers !== undefined ? { headers: options.headers } : {}),\n    ...(options.fetchImpl !== undefined ? { fetchImpl: options.fetchImpl } : {}),\n    ...(options.allowInsecureTransport !== undefined\n      ? { allowInsecureTransport: options.allowInsecureTransport }\n      : {}),\n    ...(options.acceptsSensitivity !== undefined\n      ? { acceptsSensitivity: options.acceptsSensitivity }\n      : {}),\n    ...(options.logger !== undefined ? { logger: options.logger } : {}),\n    ...(options.timeoutMs !== undefined ? { timeoutMs: options.timeoutMs } : {}),\n    capabilities: { multimodal: false, parallelToolCalls: false, ...options.capabilities },\n  });\n  return built.provider;\n}\n"],"mappings":";;;;;;;;AAuBA,MAAa,mCAAmC;;;;;;;;;;;;;;;;;;;AAyDhD,SAAgB,sBAAsB,SAAiD;CACrF,MAAM,UAAU,QAAQ,WAAW;AAmBnC,QAjBc,0BAA0B;EACtC,cAFmB,QAAQ,QAAQ,mBAAmB,QAAQ;EAG9D,OAAO,QAAQ;EACf;EACA,GAAI,QAAQ,WAAW,SAAY,EAAE,QAAQ,QAAQ,QAAQ,GAAG,EAAE;EAClE,GAAI,QAAQ,YAAY,SAAY,EAAE,SAAS,QAAQ,SAAS,GAAG,EAAE;EACrE,GAAI,QAAQ,cAAc,SAAY,EAAE,WAAW,QAAQ,WAAW,GAAG,EAAE;EAC3E,GAAI,QAAQ,2BAA2B,SACnC,EAAE,wBAAwB,QAAQ,wBAAwB,GAC1D,EAAE;EACN,GAAI,QAAQ,uBAAuB,SAC/B,EAAE,oBAAoB,QAAQ,oBAAoB,GAClD,EAAE;EACN,GAAI,QAAQ,WAAW,SAAY,EAAE,QAAQ,QAAQ,QAAQ,GAAG,EAAE;EAClE,GAAI,QAAQ,cAAc,SAAY,EAAE,WAAW,QAAQ,WAAW,GAAG,EAAE;EAC3E,cAAc;GAAE,YAAY;GAAO,mBAAmB;GAAO,GAAG,QAAQ;GAAc;EACvF,CAAC,CACW"}

package/dist/adapters/ollama.d.ts

@@ -0,0 +1,43 @@
+import { Provider, ProviderCapabilities, Sensitivity } from "@graphorin/core";
+
+//#region src/adapters/ollama.d.ts
+
+/**
+ * Default Ollama base URL.
+ *
+ * @stable
+ */
+declare const DEFAULT_OLLAMA_BASE_URL = "http://127.0.0.1:11434";
+/**
+ * Options accepted by {@link ollamaAdapter}.
+ *
+ * @stable
+ */
+interface OllamaAdapterOptions {
+  readonly model: string;
+  readonly baseUrl?: string;
+  readonly chatPath?: string;
+  readonly headers?: Readonly<Record<string, string>>;
+  readonly fetchImpl?: typeof fetch;
+  /**
+   * Time-to-response budget per request (PS-24). Default
+   * `DEFAULT_REQUEST_TIMEOUT_MS` (120s); `0` disables.
+   */
+  readonly timeoutMs?: number;
+  readonly allowInsecureTransport?: boolean;
+  readonly acceptsSensitivity?: ReadonlyArray<Sensitivity>;
+  readonly capabilities?: Partial<ProviderCapabilities>;
+  readonly name?: string;
+  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;
+}
+/**
+ * Build a Graphorin {@link Provider} backed by Ollama's native HTTP
+ * API. The adapter is fail-safe by default: public-cleartext URLs
+ * refuse to start with `LocalProviderInsecureTransportError`.
+ *
+ * @stable
+ */
+declare function ollamaAdapter(options: OllamaAdapterOptions): Provider;
+//#endregion
+export { DEFAULT_OLLAMA_BASE_URL, OllamaAdapterOptions, ollamaAdapter };
+//# sourceMappingURL=ollama.d.ts.map

package/dist/adapters/ollama.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"ollama.d.ts","names":[],"sources":["../../src/adapters/ollama.ts"],"sourcesContent":[],"mappings":";;;;;AAoFA;;;;cA5Ca,uBAAA;;;;;;UAOI,oBAAA;;;;qBAII,SAAS;8BACA;;;;;;;gCAOE,cAAc;0BACpB,QAAQ;;;;;;;;;;;iBAwBlB,aAAA,UAAuB,uBAAuB"}

package/dist/adapters/ollama.js

@@ -0,0 +1,246 @@
+import { LocalProviderInsecureTransportError, ProviderStreamParseError } from "../errors/errors.js";
+import { applyReasoningPolicy } from "../reasoning/apply-policy.js";
+import { resolveReasoningRetention } from "../reasoning/retention.js";
+import { classifyLocalProvider } from "../trust/classify-local-provider.js";
+import { callJsonHttp, makeStreamStartEvent, toOllamaChatMessages } from "../internal/http.js";
+import { parseNdJsonStream } from "../internal/sse.js";
+import { stripTrailingSlashes } from "../internal/url-utils.js";
+import { randomUUID } from "node:crypto";
+
+//#region src/adapters/ollama.ts
+/**
+* Direct adapter for the Ollama HTTP API. The adapter speaks the
+* native Ollama streaming JSON protocol (`POST /api/chat` returning
+* newline-delimited JSON objects). For operators who prefer the
+* OpenAI-compatible variant exposed by recent Ollama releases, the
+* generic {@link openAICompatibleAdapter} is the better choice — both
+* adapters share the same {@link LocalProviderTrust} classifier and
+* {@link LocalProviderInsecureTransportError} startup behaviour.
+*
+* @packageDocumentation
+*/
+/**
+* Default Ollama base URL.
+*
+* @stable
+*/
+const DEFAULT_OLLAMA_BASE_URL = "http://127.0.0.1:11434";
+const DEFAULT_CAPABILITIES = {
+	streaming: true,
+	toolCalling: true,
+	parallelToolCalls: false,
+	multimodal: false,
+	structuredOutput: true,
+	reasoning: false,
+	contextWindow: 8192,
+	maxOutput: 4096,
+	reasoningContract: "optional"
+};
+/**
+* Build a Graphorin {@link Provider} backed by Ollama's native HTTP
+* API. The adapter is fail-safe by default: public-cleartext URLs
+* refuse to start with `LocalProviderInsecureTransportError`.
+*
+* @stable
+*/
+function ollamaAdapter(options) {
+	const baseUrl = options.baseUrl ?? DEFAULT_OLLAMA_BASE_URL;
+	const classification = classifyLocalProvider(baseUrl);
+	if (classification.trust === "public-cleartext" && options.allowInsecureTransport !== true) throw new LocalProviderInsecureTransportError(baseUrl);
+	emitTrustWarning(options.logger ?? defaultLogger, classification, baseUrl);
+	const acceptsSensitivity = options.acceptsSensitivity ?? classification.acceptsSensitivity;
+	const providerName = options.name ?? `ollama-${options.model}`;
+	const capabilities = {
+		...DEFAULT_CAPABILITIES,
+		...options.capabilities
+	};
+	const chatPath = options.chatPath ?? "/api/chat";
+	const url = `${stripTrailingSlashes(baseUrl)}${chatPath}`;
+	return {
+		name: providerName,
+		modelId: options.model,
+		capabilities,
+		acceptsSensitivity,
+		stream(req) {
+			return streamOllama(options, providerName, url, applyOllamaPreflight(req, capabilities));
+		},
+		async generate(req) {
+			return generateOllama(options, providerName, url, applyOllamaPreflight(req, capabilities));
+		}
+	};
+}
+function applyOllamaPreflight(req, capabilities) {
+	const retention = resolveReasoningRetention({
+		...req.reasoningRetention !== void 0 ? { requested: req.reasoningRetention } : {},
+		...capabilities.reasoningContract !== void 0 ? { contract: capabilities.reasoningContract } : {}
+	});
+	if (retention === "pass-through-all") return req;
+	const filtered = applyReasoningPolicy({
+		messages: req.messages,
+		retention
+	});
+	if (filtered === req.messages) return req;
+	return {
+		...req,
+		messages: filtered
+	};
+}
+async function* streamOllama(options, providerName, url, req) {
+	const body = buildBody(options.model, req, true, options.capabilities?.structuredOutput ?? true);
+	const resp = await callJsonHttp({
+		providerName,
+		url,
+		headers: buildHeaders(options),
+		body,
+		...req.signal !== void 0 ? { signal: req.signal } : {},
+		...options.fetchImpl !== void 0 ? { fetchImpl: options.fetchImpl } : {},
+		...options.timeoutMs !== void 0 ? { timeoutMs: options.timeoutMs } : {}
+	});
+	yield makeStreamStartEvent({
+		providerName,
+		modelId: options.model
+	});
+	let usage = {
+		promptTokens: 0,
+		completionTokens: 0,
+		totalTokens: 0
+	};
+	let finishReason = "stop";
+	for await (const line of parseNdJsonStream(resp.body, req.signal !== void 0 ? { signal: req.signal } : {})) {
+		if (req.signal?.aborted) {
+			finishReason = "aborted";
+			break;
+		}
+		let chunk;
+		try {
+			chunk = JSON.parse(line);
+		} catch (cause) {
+			throw new ProviderStreamParseError(providerName, `failed to parse ndjson line: ${cause.message}`, cause);
+		}
+		if (chunk.message?.content !== void 0 && chunk.message.content.length > 0) yield {
+			type: "text-delta",
+			delta: chunk.message.content
+		};
+		if (Array.isArray(chunk.message?.tool_calls)) for (const tc of chunk.message.tool_calls) {
+			const toolCallId = tc.id ?? `call_${randomUUID()}`;
+			const toolName = tc.function?.name ?? "";
+			if (toolName.length === 0) continue;
+			yield {
+				type: "tool-call-start",
+				toolCallId,
+				toolName
+			};
+			yield {
+				type: "tool-call-end",
+				toolCallId,
+				finalArgs: tc.function?.arguments ?? {}
+			};
+		}
+		if (chunk.done === true) {
+			finishReason = mapFinishReason(chunk.done_reason);
+			usage = mapUsage(chunk);
+			break;
+		}
+	}
+	yield {
+		type: "finish",
+		finishReason,
+		usage
+	};
+}
+async function generateOllama(options, providerName, url, req) {
+	const body = buildBody(options.model, req, false, options.capabilities?.structuredOutput ?? true);
+	const resp = await callJsonHttp({
+		providerName,
+		url,
+		headers: buildHeaders(options),
+		body,
+		...req.signal !== void 0 ? { signal: req.signal } : {},
+		...options.fetchImpl !== void 0 ? { fetchImpl: options.fetchImpl } : {},
+		...options.timeoutMs !== void 0 ? { timeoutMs: options.timeoutMs } : {}
+	});
+	let json;
+	try {
+		json = await resp.json();
+	} catch (cause) {
+		throw new ProviderStreamParseError(providerName, "response body was not valid JSON", cause);
+	}
+	return {
+		usage: mapUsage(json),
+		finishReason: mapFinishReason(json.done_reason),
+		...typeof json.message?.content === "string" && json.message.content.length > 0 ? { text: json.message.content } : {},
+		...Array.isArray(json.message?.tool_calls) && json.message.tool_calls.length > 0 ? { toolCalls: json.message.tool_calls.map((tc) => ({
+			toolCallId: tc.id ?? `call_${randomUUID()}`,
+			toolName: tc.function?.name ?? "",
+			args: tc.function?.arguments ?? {}
+		})) } : {}
+	};
+}
+function buildBody(model, req, stream, structuredOutput) {
+	const body = {
+		model,
+		messages: toOllamaChatMessages(req.systemMessage !== void 0 ? [{
+			role: "system",
+			content: req.systemMessage
+		}, ...req.messages] : req.messages),
+		stream
+	};
+	if (req.temperature !== void 0 || req.maxTokens !== void 0) {
+		const optionsBlock = {};
+		if (req.temperature !== void 0) optionsBlock.temperature = req.temperature;
+		if (req.maxTokens !== void 0) optionsBlock.num_predict = req.maxTokens;
+		body.options = optionsBlock;
+	}
+	if (req.tools !== void 0 && req.tools.length > 0) body.tools = req.tools.map((t) => ({
+		type: "function",
+		function: {
+			name: t.name,
+			...t.description !== void 0 ? { description: t.description } : {},
+			parameters: t.inputSchema
+		}
+	}));
+	if (structuredOutput && req.outputType?.kind === "structured") body.format = req.outputType.jsonSchema ?? "json";
+	if (req.providerOptions !== void 0) Object.assign(body, req.providerOptions);
+	return body;
+}
+function buildHeaders(options) {
+	return {
+		"content-type": "application/json",
+		accept: "application/json",
+		...options.headers
+	};
+}
+function mapFinishReason(value) {
+	switch (value) {
+		case "stop":
+		case "length": return value;
+		case "tool_calls": return "tool-calls";
+		default: return "stop";
+	}
+}
+function mapUsage(chunk) {
+	const promptTokens = chunk.prompt_eval_count ?? 0;
+	const completionTokens = chunk.eval_count ?? 0;
+	return {
+		promptTokens,
+		completionTokens,
+		totalTokens: promptTokens + completionTokens
+	};
+}
+function emitTrustWarning(log, classification, baseUrl) {
+	if (classification.trust === "public-cleartext") log("warn", `[ollama] allowInsecureTransport=true accepted for ${baseUrl}`, { baseUrl });
+	else if (classification.trust === "public-tls") log("warn", `[ollama] public-TLS endpoint; treating as cloud-tier`, { baseUrl });
+	else if (classification.trust === "private") log("warn", `[ollama] private-network endpoint detected (${classification.reason})`, {
+		baseUrl,
+		acceptsSensitivity: classification.acceptsSensitivity
+	});
+}
+function defaultLogger(level, message, meta) {
+	const fn = level === "warn" ? console.warn : console.info;
+	if (meta !== void 0) fn(`[graphorin/provider] ${message}`, meta);
+	else fn(`[graphorin/provider] ${message}`);
+}
+
+//#endregion
+export { DEFAULT_OLLAMA_BASE_URL, ollamaAdapter };
+//# sourceMappingURL=ollama.js.map

package/dist/adapters/ollama.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ollama.js","names":["DEFAULT_CAPABILITIES: ProviderCapabilities","capabilities: ProviderCapabilities","usage: Usage","finishReason: FinishReason","chunk: OllamaChatChunk","json: OllamaChatChunk","body: Record<string, unknown>","optionsBlock: Record<string, unknown>"],"sources":["../../src/adapters/ollama.ts"],"sourcesContent":["/**\n * Direct adapter for the Ollama HTTP API. The adapter speaks the\n * native Ollama streaming JSON protocol (`POST /api/chat` returning\n * newline-delimited JSON objects). For operators who prefer the\n * OpenAI-compatible variant exposed by recent Ollama releases, the\n * generic {@link openAICompatibleAdapter} is the better choice — both\n * adapters share the same {@link LocalProviderTrust} classifier and\n * {@link LocalProviderInsecureTransportError} startup behaviour.\n *\n * @packageDocumentation\n */\n\nimport { randomUUID } from 'node:crypto';\nimport type {\n  FinishReason,\n  Provider,\n  ProviderCapabilities,\n  ProviderEvent,\n  ProviderRequest,\n  ProviderResponse,\n  Sensitivity,\n  Usage,\n} from '@graphorin/core';\n\nimport { LocalProviderInsecureTransportError, ProviderStreamParseError } from '../errors/errors.js';\nimport { callJsonHttp, makeStreamStartEvent, toOllamaChatMessages } from '../internal/http.js';\nimport { parseNdJsonStream } from '../internal/sse.js';\nimport { stripTrailingSlashes } from '../internal/url-utils.js';\nimport { applyReasoningPolicy } from '../reasoning/apply-policy.js';\nimport { resolveReasoningRetention } from '../reasoning/retention.js';\nimport {\n  classifyLocalProvider,\n  type LocalProviderClassification,\n} from '../trust/classify-local-provider.js';\n\n/**\n * Default Ollama base URL.\n *\n * @stable\n */\nexport const DEFAULT_OLLAMA_BASE_URL = 'http://127.0.0.1:11434';\n\n/**\n * Options accepted by {@link ollamaAdapter}.\n *\n * @stable\n */\nexport interface OllamaAdapterOptions {\n  readonly model: string;\n  readonly baseUrl?: string;\n  readonly chatPath?: string;\n  readonly headers?: Readonly<Record<string, string>>;\n  readonly fetchImpl?: typeof fetch;\n  /**\n   * Time-to-response budget per request (PS-24). Default\n   * `DEFAULT_REQUEST_TIMEOUT_MS` (120s); `0` disables.\n   */\n  readonly timeoutMs?: number;\n  readonly allowInsecureTransport?: boolean;\n  readonly acceptsSensitivity?: ReadonlyArray<Sensitivity>;\n  readonly capabilities?: Partial<ProviderCapabilities>;\n  readonly name?: string;\n  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;\n}\n\nconst DEFAULT_CAPABILITIES: ProviderCapabilities = {\n  streaming: true,\n  toolCalling: true,\n  parallelToolCalls: false,\n  multimodal: false,\n  structuredOutput: true,\n  reasoning: false,\n  contextWindow: 8_192,\n  maxOutput: 4_096,\n  reasoningContract: 'optional',\n};\n\n/**\n * Build a Graphorin {@link Provider} backed by Ollama's native HTTP\n * API. The adapter is fail-safe by default: public-cleartext URLs\n * refuse to start with `LocalProviderInsecureTransportError`.\n *\n * @stable\n */\nexport function ollamaAdapter(options: OllamaAdapterOptions): Provider {\n  const baseUrl = options.baseUrl ?? DEFAULT_OLLAMA_BASE_URL;\n  const classification = classifyLocalProvider(baseUrl);\n  if (classification.trust === 'public-cleartext' && options.allowInsecureTransport !== true) {\n    throw new LocalProviderInsecureTransportError(baseUrl);\n  }\n  const log = options.logger ?? defaultLogger;\n  emitTrustWarning(log, classification, baseUrl);\n  const acceptsSensitivity = options.acceptsSensitivity ?? classification.acceptsSensitivity;\n  const providerName = options.name ?? `ollama-${options.model}`;\n  const capabilities: ProviderCapabilities = {\n    ...DEFAULT_CAPABILITIES,\n    ...options.capabilities,\n  };\n  const chatPath = options.chatPath ?? '/api/chat';\n  const url = `${stripTrailingSlashes(baseUrl)}${chatPath}`;\n  return {\n    name: providerName,\n    modelId: options.model,\n    capabilities,\n    acceptsSensitivity,\n    stream(req) {\n      return streamOllama(options, providerName, url, applyOllamaPreflight(req, capabilities));\n    },\n    async generate(req) {\n      return generateOllama(options, providerName, url, applyOllamaPreflight(req, capabilities));\n    },\n  };\n}\n\nfunction applyOllamaPreflight(\n  req: ProviderRequest,\n  capabilities: ProviderCapabilities,\n): ProviderRequest {\n  const retention = resolveReasoningRetention({\n    ...(req.reasoningRetention !== undefined ? { requested: req.reasoningRetention } : {}),\n    ...(capabilities.reasoningContract !== undefined\n      ? { contract: capabilities.reasoningContract }\n      : {}),\n  });\n  if (retention === 'pass-through-all') return req;\n  const filtered = applyReasoningPolicy({ messages: req.messages, retention });\n  if (filtered === req.messages) return req;\n  return { ...req, messages: filtered };\n}\n\nasync function* streamOllama(\n  options: OllamaAdapterOptions,\n  providerName: string,\n  url: string,\n  req: ProviderRequest,\n): AsyncIterable<ProviderEvent> {\n  const body = buildBody(options.model, req, true, options.capabilities?.structuredOutput ?? true);\n  const resp = await callJsonHttp({\n    providerName,\n    url,\n    headers: buildHeaders(options),\n    body,\n    ...(req.signal !== undefined ? { signal: req.signal } : {}),\n    ...(options.fetchImpl !== undefined ? { fetchImpl: options.fetchImpl } : {}),\n    ...(options.timeoutMs !== undefined ? { timeoutMs: options.timeoutMs } : {}),\n  });\n  yield makeStreamStartEvent({ providerName, modelId: options.model });\n  let usage: Usage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 };\n  let finishReason: FinishReason = 'stop';\n  for await (const line of parseNdJsonStream(\n    resp.body,\n    req.signal !== undefined ? { signal: req.signal } : {},\n  )) {\n    if (req.signal?.aborted) {\n      finishReason = 'aborted'; // PS-12: honest abort reason, not 'stop'\n      break;\n    }\n    let chunk: OllamaChatChunk;\n    try {\n      chunk = JSON.parse(line) as OllamaChatChunk;\n    } catch (cause) {\n      throw new ProviderStreamParseError(\n        providerName,\n        `failed to parse ndjson line: ${(cause as Error).message}`,\n        cause,\n      );\n    }\n    if (chunk.message?.content !== undefined && chunk.message.content.length > 0) {\n      yield { type: 'text-delta', delta: chunk.message.content };\n    }\n    if (Array.isArray(chunk.message?.tool_calls)) {\n      for (const tc of chunk.message.tool_calls) {\n        const toolCallId = tc.id ?? `call_${randomUUID()}`;\n        const toolName = tc.function?.name ?? '';\n        if (toolName.length === 0) continue;\n        yield { type: 'tool-call-start', toolCallId, toolName };\n        yield {\n          type: 'tool-call-end',\n          toolCallId,\n          finalArgs: tc.function?.arguments ?? {},\n        };\n      }\n    }\n    if (chunk.done === true) {\n      finishReason = mapFinishReason(chunk.done_reason);\n      usage = mapUsage(chunk);\n      break;\n    }\n  }\n  yield { type: 'finish', finishReason, usage };\n}\n\nasync function generateOllama(\n  options: OllamaAdapterOptions,\n  providerName: string,\n  url: string,\n  req: ProviderRequest,\n): Promise<ProviderResponse> {\n  const body = buildBody(options.model, req, false, options.capabilities?.structuredOutput ?? true);\n  const resp = await callJsonHttp({\n    providerName,\n    url,\n    headers: buildHeaders(options),\n    body,\n    ...(req.signal !== undefined ? { signal: req.signal } : {}),\n    ...(options.fetchImpl !== undefined ? { fetchImpl: options.fetchImpl } : {}),\n    ...(options.timeoutMs !== undefined ? { timeoutMs: options.timeoutMs } : {}),\n  });\n  let json: OllamaChatChunk;\n  try {\n    json = (await resp.json()) as OllamaChatChunk;\n  } catch (cause) {\n    throw new ProviderStreamParseError(providerName, 'response body was not valid JSON', cause);\n  }\n  return {\n    usage: mapUsage(json),\n    finishReason: mapFinishReason(json.done_reason),\n    ...(typeof json.message?.content === 'string' && json.message.content.length > 0\n      ? { text: json.message.content }\n      : {}),\n    ...(Array.isArray(json.message?.tool_calls) && json.message.tool_calls.length > 0\n      ? {\n          toolCalls: json.message.tool_calls.map((tc) => ({\n            toolCallId: tc.id ?? `call_${randomUUID()}`,\n            toolName: tc.function?.name ?? '',\n            args: tc.function?.arguments ?? {},\n          })),\n        }\n      : {}),\n  };\n}\n\nfunction buildBody(\n  model: string,\n  req: ProviderRequest,\n  stream: boolean,\n  structuredOutput: boolean,\n): Record<string, unknown> {\n  const messages =\n    req.systemMessage !== undefined\n      ? [{ role: 'system' as const, content: req.systemMessage }, ...req.messages]\n      : req.messages;\n  const body: Record<string, unknown> = {\n    model,\n    messages: toOllamaChatMessages(messages),\n    stream,\n  };\n  if (req.temperature !== undefined || req.maxTokens !== undefined) {\n    const optionsBlock: Record<string, unknown> = {};\n    if (req.temperature !== undefined) optionsBlock.temperature = req.temperature;\n    if (req.maxTokens !== undefined) optionsBlock.num_predict = req.maxTokens;\n    body.options = optionsBlock;\n  }\n  if (req.tools !== undefined && req.tools.length > 0) {\n    body.tools = req.tools.map((t) => ({\n      type: 'function',\n      function: {\n        name: t.name,\n        ...(t.description !== undefined ? { description: t.description } : {}),\n        parameters: t.inputSchema,\n      },\n    }));\n  }\n  // PS-24: Ollama's native structured output — `format` takes a JSON\n  // schema object (or 'json' for schema-less JSON mode).\n  if (structuredOutput && req.outputType?.kind === 'structured') {\n    body.format = req.outputType.jsonSchema ?? 'json';\n  }\n  if (req.providerOptions !== undefined) {\n    Object.assign(body, req.providerOptions);\n  }\n  return body;\n}\n\nfunction buildHeaders(options: OllamaAdapterOptions): Record<string, string> {\n  return {\n    'content-type': 'application/json',\n    accept: 'application/json',\n    ...options.headers,\n  };\n}\n\nfunction mapFinishReason(value: string | null | undefined): FinishReason {\n  switch (value) {\n    case 'stop':\n    case 'length':\n      return value;\n    case 'tool_calls':\n      return 'tool-calls';\n    default:\n      return 'stop';\n  }\n}\n\nfunction mapUsage(chunk: OllamaChatChunk): Usage {\n  const promptTokens = chunk.prompt_eval_count ?? 0;\n  const completionTokens = chunk.eval_count ?? 0;\n  return {\n    promptTokens,\n    completionTokens,\n    totalTokens: promptTokens + completionTokens,\n  };\n}\n\nfunction emitTrustWarning(\n  log: (level: 'warn' | 'info', message: string, meta?: object) => void,\n  classification: LocalProviderClassification,\n  baseUrl: string,\n): void {\n  if (classification.trust === 'public-cleartext') {\n    log('warn', `[ollama] allowInsecureTransport=true accepted for ${baseUrl}`, { baseUrl });\n  } else if (classification.trust === 'public-tls') {\n    log('warn', `[ollama] public-TLS endpoint; treating as cloud-tier`, { baseUrl });\n  } else if (classification.trust === 'private') {\n    log('warn', `[ollama] private-network endpoint detected (${classification.reason})`, {\n      baseUrl,\n      acceptsSensitivity: classification.acceptsSensitivity,\n    });\n  }\n}\n\nfunction defaultLogger(level: 'warn' | 'info', message: string, meta?: object): void {\n  const fn = level === 'warn' ? console.warn : console.info;\n  if (meta !== undefined) {\n    fn(`[graphorin/provider] ${message}`, meta);\n  } else {\n    fn(`[graphorin/provider] ${message}`);\n  }\n}\n\ninterface OllamaChatChunk {\n  readonly model?: string;\n  readonly created_at?: string;\n  readonly message?: {\n    readonly role?: string;\n    readonly content?: string;\n    readonly tool_calls?: ReadonlyArray<{\n      readonly id?: string;\n      readonly function?: { readonly name?: string; readonly arguments?: unknown };\n    }>;\n  };\n  readonly done?: boolean;\n  readonly done_reason?: string;\n  readonly prompt_eval_count?: number;\n  readonly eval_count?: number;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AAwCA,MAAa,0BAA0B;AAyBvC,MAAMA,uBAA6C;CACjD,WAAW;CACX,aAAa;CACb,mBAAmB;CACnB,YAAY;CACZ,kBAAkB;CAClB,WAAW;CACX,eAAe;CACf,WAAW;CACX,mBAAmB;CACpB;;;;;;;;AASD,SAAgB,cAAc,SAAyC;CACrE,MAAM,UAAU,QAAQ,WAAW;CACnC,MAAM,iBAAiB,sBAAsB,QAAQ;AACrD,KAAI,eAAe,UAAU,sBAAsB,QAAQ,2BAA2B,KACpF,OAAM,IAAI,oCAAoC,QAAQ;AAGxD,kBADY,QAAQ,UAAU,eACR,gBAAgB,QAAQ;CAC9C,MAAM,qBAAqB,QAAQ,sBAAsB,eAAe;CACxE,MAAM,eAAe,QAAQ,QAAQ,UAAU,QAAQ;CACvD,MAAMC,eAAqC;EACzC,GAAG;EACH,GAAG,QAAQ;EACZ;CACD,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,MAAM,GAAG,qBAAqB,QAAQ,GAAG;AAC/C,QAAO;EACL,MAAM;EACN,SAAS,QAAQ;EACjB;EACA;EACA,OAAO,KAAK;AACV,UAAO,aAAa,SAAS,cAAc,KAAK,qBAAqB,KAAK,aAAa,CAAC;;EAE1F,MAAM,SAAS,KAAK;AAClB,UAAO,eAAe,SAAS,cAAc,KAAK,qBAAqB,KAAK,aAAa,CAAC;;EAE7F;;AAGH,SAAS,qBACP,KACA,cACiB;CACjB,MAAM,YAAY,0BAA0B;EAC1C,GAAI,IAAI,uBAAuB,SAAY,EAAE,WAAW,IAAI,oBAAoB,GAAG,EAAE;EACrF,GAAI,aAAa,sBAAsB,SACnC,EAAE,UAAU,aAAa,mBAAmB,GAC5C,EAAE;EACP,CAAC;AACF,KAAI,cAAc,mBAAoB,QAAO;CAC7C,MAAM,WAAW,qBAAqB;EAAE,UAAU,IAAI;EAAU;EAAW,CAAC;AAC5E,KAAI,aAAa,IAAI,SAAU,QAAO;AACtC,QAAO;EAAE,GAAG;EAAK,UAAU;EAAU;;AAGvC,gBAAgB,aACd,SACA,cACA,KACA,KAC8B;CAC9B,MAAM,OAAO,UAAU,QAAQ,OAAO,KAAK,MAAM,QAAQ,cAAc,oBAAoB,KAAK;CAChG,MAAM,OAAO,MAAM,aAAa;EAC9B;EACA;EACA,SAAS,aAAa,QAAQ;EAC9B;EACA,GAAI,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE;EAC1D,GAAI,QAAQ,cAAc,SAAY,EAAE,WAAW,QAAQ,WAAW,GAAG,EAAE;EAC3E,GAAI,QAAQ,cAAc,SAAY,EAAE,WAAW,QAAQ,WAAW,GAAG,EAAE;EAC5E,CAAC;AACF,OAAM,qBAAqB;EAAE;EAAc,SAAS,QAAQ;EAAO,CAAC;CACpE,IAAIC,QAAe;EAAE,cAAc;EAAG,kBAAkB;EAAG,aAAa;EAAG;CAC3E,IAAIC,eAA6B;AACjC,YAAW,MAAM,QAAQ,kBACvB,KAAK,MACL,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE,CACvD,EAAE;AACD,MAAI,IAAI,QAAQ,SAAS;AACvB,kBAAe;AACf;;EAEF,IAAIC;AACJ,MAAI;AACF,WAAQ,KAAK,MAAM,KAAK;WACjB,OAAO;AACd,SAAM,IAAI,yBACR,cACA,gCAAiC,MAAgB,WACjD,MACD;;AAEH,MAAI,MAAM,SAAS,YAAY,UAAa,MAAM,QAAQ,QAAQ,SAAS,EACzE,OAAM;GAAE,MAAM;GAAc,OAAO,MAAM,QAAQ;GAAS;AAE5D,MAAI,MAAM,QAAQ,MAAM,SAAS,WAAW,CAC1C,MAAK,MAAM,MAAM,MAAM,QAAQ,YAAY;GACzC,MAAM,aAAa,GAAG,MAAM,QAAQ,YAAY;GAChD,MAAM,WAAW,GAAG,UAAU,QAAQ;AACtC,OAAI,SAAS,WAAW,EAAG;AAC3B,SAAM;IAAE,MAAM;IAAmB;IAAY;IAAU;AACvD,SAAM;IACJ,MAAM;IACN;IACA,WAAW,GAAG,UAAU,aAAa,EAAE;IACxC;;AAGL,MAAI,MAAM,SAAS,MAAM;AACvB,kBAAe,gBAAgB,MAAM,YAAY;AACjD,WAAQ,SAAS,MAAM;AACvB;;;AAGJ,OAAM;EAAE,MAAM;EAAU;EAAc;EAAO;;AAG/C,eAAe,eACb,SACA,cACA,KACA,KAC2B;CAC3B,MAAM,OAAO,UAAU,QAAQ,OAAO,KAAK,OAAO,QAAQ,cAAc,oBAAoB,KAAK;CACjG,MAAM,OAAO,MAAM,aAAa;EAC9B;EACA;EACA,SAAS,aAAa,QAAQ;EAC9B;EACA,GAAI,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE;EAC1D,GAAI,QAAQ,cAAc,SAAY,EAAE,WAAW,QAAQ,WAAW,GAAG,EAAE;EAC3E,GAAI,QAAQ,cAAc,SAAY,EAAE,WAAW,QAAQ,WAAW,GAAG,EAAE;EAC5E,CAAC;CACF,IAAIC;AACJ,KAAI;AACF,SAAQ,MAAM,KAAK,MAAM;UAClB,OAAO;AACd,QAAM,IAAI,yBAAyB,cAAc,oCAAoC,MAAM;;AAE7F,QAAO;EACL,OAAO,SAAS,KAAK;EACrB,cAAc,gBAAgB,KAAK,YAAY;EAC/C,GAAI,OAAO,KAAK,SAAS,YAAY,YAAY,KAAK,QAAQ,QAAQ,SAAS,IAC3E,EAAE,MAAM,KAAK,QAAQ,SAAS,GAC9B,EAAE;EACN,GAAI,MAAM,QAAQ,KAAK,SAAS,WAAW,IAAI,KAAK,QAAQ,WAAW,SAAS,IAC5E,EACE,WAAW,KAAK,QAAQ,WAAW,KAAK,QAAQ;GAC9C,YAAY,GAAG,MAAM,QAAQ,YAAY;GACzC,UAAU,GAAG,UAAU,QAAQ;GAC/B,MAAM,GAAG,UAAU,aAAa,EAAE;GACnC,EAAE,EACJ,GACD,EAAE;EACP;;AAGH,SAAS,UACP,OACA,KACA,QACA,kBACyB;CAKzB,MAAMC,OAAgC;EACpC;EACA,UAAU,qBALV,IAAI,kBAAkB,SAClB,CAAC;GAAE,MAAM;GAAmB,SAAS,IAAI;GAAe,EAAE,GAAG,IAAI,SAAS,GAC1E,IAAI,SAGgC;EACxC;EACD;AACD,KAAI,IAAI,gBAAgB,UAAa,IAAI,cAAc,QAAW;EAChE,MAAMC,eAAwC,EAAE;AAChD,MAAI,IAAI,gBAAgB,OAAW,cAAa,cAAc,IAAI;AAClE,MAAI,IAAI,cAAc,OAAW,cAAa,cAAc,IAAI;AAChE,OAAK,UAAU;;AAEjB,KAAI,IAAI,UAAU,UAAa,IAAI,MAAM,SAAS,EAChD,MAAK,QAAQ,IAAI,MAAM,KAAK,OAAO;EACjC,MAAM;EACN,UAAU;GACR,MAAM,EAAE;GACR,GAAI,EAAE,gBAAgB,SAAY,EAAE,aAAa,EAAE,aAAa,GAAG,EAAE;GACrE,YAAY,EAAE;GACf;EACF,EAAE;AAIL,KAAI,oBAAoB,IAAI,YAAY,SAAS,aAC/C,MAAK,SAAS,IAAI,WAAW,cAAc;AAE7C,KAAI,IAAI,oBAAoB,OAC1B,QAAO,OAAO,MAAM,IAAI,gBAAgB;AAE1C,QAAO;;AAGT,SAAS,aAAa,SAAuD;AAC3E,QAAO;EACL,gBAAgB;EAChB,QAAQ;EACR,GAAG,QAAQ;EACZ;;AAGH,SAAS,gBAAgB,OAAgD;AACvE,SAAQ,OAAR;EACE,KAAK;EACL,KAAK,SACH,QAAO;EACT,KAAK,aACH,QAAO;EACT,QACE,QAAO;;;AAIb,SAAS,SAAS,OAA+B;CAC/C,MAAM,eAAe,MAAM,qBAAqB;CAChD,MAAM,mBAAmB,MAAM,cAAc;AAC7C,QAAO;EACL;EACA;EACA,aAAa,eAAe;EAC7B;;AAGH,SAAS,iBACP,KACA,gBACA,SACM;AACN,KAAI,eAAe,UAAU,mBAC3B,KAAI,QAAQ,qDAAqD,WAAW,EAAE,SAAS,CAAC;UAC/E,eAAe,UAAU,aAClC,KAAI,QAAQ,wDAAwD,EAAE,SAAS,CAAC;UACvE,eAAe,UAAU,UAClC,KAAI,QAAQ,+CAA+C,eAAe,OAAO,IAAI;EACnF;EACA,oBAAoB,eAAe;EACpC,CAAC;;AAIN,SAAS,cAAc,OAAwB,SAAiB,MAAqB;CACnF,MAAM,KAAK,UAAU,SAAS,QAAQ,OAAO,QAAQ;AACrD,KAAI,SAAS,OACX,IAAG,wBAAwB,WAAW,KAAK;KAE3C,IAAG,wBAAwB,UAAU"}

package/dist/adapters/openai-compatible.d.ts

@@ -0,0 +1,48 @@
+import { Provider, Sensitivity } from "@graphorin/core";
+
+//#region src/adapters/openai-compatible.d.ts
+
+/**
+ * Options accepted by {@link openAICompatibleAdapter}.
+ *
+ * @stable
+ */
+interface OpenAICompatibleAdapterOptions {
+  /** Model identifier sent in the request body's `model` field. */
+  readonly model: string;
+  /**
+   * Base URL of the OpenAI-compatible server. The classifier inspects
+   * the protocol + host to assign a `LocalProviderTrust` value.
+   */
+  readonly baseUrl: string;
+  /** Optional REST path override. Defaults to `/v1/chat/completions`. */
+  readonly chatPath?: string;
+  /** Optional bearer-auth API key. */
+  readonly apiKey?: string;
+  /** Extra headers merged on top of `content-type` + `accept` defaults. */
+  readonly headers?: Readonly<Record<string, string>>;
+  /** Custom `fetch` implementation; useful for tests. */
+  readonly fetchImpl?: typeof fetch;
+  /**
+   * Acknowledge the risk of running over plaintext HTTP against a
+   * public host.
+   */
+  readonly allowInsecureTransport?: boolean;
+  /** Override for the default `acceptsSensitivity` value. */
+  readonly acceptsSensitivity?: ReadonlyArray<Sensitivity>;
+  /** Provider name attached to spans / log lines. */
+  readonly name?: string;
+  /** Optional log sink. */
+  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;
+}
+/**
+ * Build a Graphorin {@link Provider} backed by an OpenAI-compatible
+ * HTTP server. The same code path serves LMStudio, LocalAI, vLLM, and
+ * any other compatible self-host endpoint.
+ *
+ * @stable
+ */
+declare function openAICompatibleAdapter(options: OpenAICompatibleAdapterOptions): Provider;
+//#endregion
+export { OpenAICompatibleAdapterOptions, openAICompatibleAdapter };
+//# sourceMappingURL=openai-compatible.d.ts.map