@orkestrel/ollama 0.0.1

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Orkestrel
+
+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 +1,66 @@
+# @orkestrel/ollama
+
+A typed local-LLM provider for the `@orkestrel` line — a `ProviderInterface`
+implementation over a local Ollama daemon's `POST /api/chat`, with NDJSON
+streaming, tool calls, thinking, and usage accounting, built on pure
+web-standard `fetch` / `ReadableStream` (no Ollama SDK dependency).
+
+## Install
+
+```sh
+npm install @orkestrel/ollama
+```
+
+## Requirements
+
+- Node.js >= 24
+- A running Ollama daemon (default `http://localhost:11434`) with a pulled
+  model — required at runtime by any consumer, and by this repo's live-only
+  `src:server` test suite
+- ESM + CJS (dual-format build)
+
+## Usage
+
+```ts
+import { createAbort } from '@orkestrel/abort'
+import { createOllama } from '@orkestrel/ollama'
+
+const provider = createOllama({ model: 'qwen3.5:2b-q4_K_M' })
+const abort = createAbort()
+const messages = [{ id: '1', role: 'user', content: 'Reply with exactly: ok' }] as const
+
+const result = await provider.generate(messages, abort.signal)
+result.content // 'ok'
+result.usage // { prompt, completion, total }
+```
+
+`generate` resolves the assembled `ProviderResult` in one call. For live
+output, `stream` is the same call streamed — it yields `ProviderDelta`s
+(`content` for answer text, `thinking` for live reasoning) and returns the
+assembled result when the stream completes:
+
+```ts
+const generator = provider.stream(messages, abort.signal)
+let step = await generator.next()
+while (!step.done) {
+	if (step.value.type === 'content') process.stdout.write(step.value.text)
+	step = await generator.next()
+}
+const streamed = step.value // the assembled ProviderResult
+```
+
+## Guide
+
+For the full surface — `createOllama`, `OllamaProvider`, `OllamaOptions`,
+tool calls, thinking, and the context-framing default — see
+[`guides/src/ollama.md`](guides/src/ollama.md).
+
+## Package
+
+Published as a single server surface per the `exports` field in
+`package.json` — one `.` entry backed by a dual ESM + CommonJS build of
+`src/server`.
+
+## License
+
+MIT © [Orkestrel](https://github.com/orkestrel) — see [LICENSE](./LICENSE).

package/dist/src/server/index.cjs

@@ -0,0 +1,459 @@
+Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
+let _orkestrel_agent = require("@orkestrel/agent");
+let _orkestrel_contract = require("@orkestrel/contract");
+let _orkestrel_ndjson = require("@orkestrel/ndjson");
+let _orkestrel_timeout = require("@orkestrel/timeout");
+//#region src/server/constants.ts
+/** The local Ollama daemon base URL assumed when `OllamaOptions.url` is omitted. */
+var DEFAULT_OLLAMA_URL = "http://localhost:11434";
+/**
+* How long the model stays resident after a call when `OllamaOptions.keepAlive` is
+* omitted — Ollama's own `keep_alive` default, expressed as a duration string.
+*/
+var DEFAULT_KEEP_ALIVE = "5m";
+/**
+* The per-call deadline in milliseconds when `OllamaOptions.timeout` is omitted —
+* generous enough that a cold model load does not trip it.
+*/
+var DEFAULT_PROVIDER_TIMEOUT = 12e4;
+/**
+* The cap, in characters, on how much of a non-OK response body is
+* incorporated into a thrown {@link OllamaHTTPError}'s message.
+*
+* @remarks
+* Bounds the excerpt so a defensive proxy or a misbehaving daemon handing
+* back an unbounded response body cannot inflate the thrown error's message
+* without limit (§14). `2048` characters is generous enough to carry a
+* useful diagnostic snippet while staying well short of any practical size
+* concern.
+*/
+var MAX_ERROR_BODY_LENGTH = 2048;
+//#endregion
+//#region src/server/errors.ts
+/**
+* An error thrown when the Ollama `/api/chat` HTTP transport fails.
+*
+* @remarks
+* Carries the response `status` (0 when no HTTP response was received at all,
+* e.g. a `null` body). Thrown by {@link OllamaProvider} at its two HTTP
+* failure sites — the non-OK status branch and the null-body branch — so a
+* caller can branch on `error.status` instead of parsing the message. Narrow
+* a caught value with {@link isOllamaHTTPError}.
+*
+* @example
+* ```ts
+* try {
+* 	await provider.generate(messages, signal)
+* } catch (error) {
+* 	if (isOllamaHTTPError(error) && error.status === 404) {
+* 		// the configured model isn't pulled
+* 	}
+* }
+* ```
+*/
+var OllamaHTTPError = class extends Error {
+	status;
+	constructor(message, status, options) {
+		super(message, options);
+		this.name = "OllamaHTTPError";
+		this.status = status;
+	}
+};
+/**
+* Whether a value is an {@link OllamaHTTPError}.
+*
+* @param value - The value to test
+* @returns `true` when `value` is an `OllamaHTTPError`
+*/
+function isOllamaHTTPError(value) {
+	return value instanceof OllamaHTTPError;
+}
+//#endregion
+//#region src/server/OllamaProvider.ts
+/**
+* The local Ollama inference boundary — a {@link ProviderInterface} over Ollama's
+* `POST /api/chat`, both non-streaming (`generate`) and streaming NDJSON (`stream`).
+*
+* @remarks
+* - **Wire protocol.** Posts `{ model, messages, stream, keep_alive, think }` plus
+*   passthrough sampling `options` and mapped function `tools`. The `think` flag is
+*   CONFIGURABLE via {@link OllamaOptions.think} (default `false`). Non-stream parses
+*   one JSON body; stream consumes NDJSON (one JSON object per `\n`-terminated line) —
+*   deltas carry `message.content`, the final `done: true` line carries the token usage.
+* - **Think separation (H4).** The wire `think` flag is configurable
+*   ({@link OllamaOptions.think}, default `false`). With `think: true` a thinking model's
+*   daemon separates reasoning NATIVELY — returning it on the distinct `message.thinking`
+*   channel (read here via `#thinking`) instead of inline in `message.content`. EITHER
+*   way the per-call {@link ThinkSplitterInterface} is the defensive guarantee: a daemon
+*   may ignore `think: false` for a thinking model and inline `<think>` tags, so every
+*   content delta routes through the splitter, only CLEAN content is yielded / assembled,
+*   and the separated reasoning (plus any daemon-side `message.thinking` deltas) lands on
+*   `ProviderResult.thinking`, never in the conversation.
+* - **Boundary narrowing (§14).** Every wire value arrives as `unknown` and is
+*   narrowed through guards (`isRecord` / `isString` / `isNumber`) — never `as`. A
+*   missing / malformed field degrades to a sensible default (empty content, no
+*   usage, `{}` arguments), never a throw.
+* - **Bounded.** Each call arms a {@link Timeout} for `OllamaOptions.timeout` and
+*   passes `AbortSignal.any([timeout.signal, signal])` to `fetch`, so the caller's
+*   signal AND the deadline both cancel the request. The timeout is always cleared —
+*   in `#fetch` if the request fails/aborts, otherwise in the consuming call's `finally`.
+* - **Abort recovers partial.** A `stream` cancelled mid-flight throws a
+*   `ProviderAbortError` carrying the partial result assembled so far; pairing the
+*   `TextDecoder({ stream: true })` with the {@link NDJSONParser} parser keeps multi-byte
+*   UTF-8 splits and partial lines honest.
+* - **Event-free.** A pure functional boundary — no Emitter, no events.
+* - **Transport seam.** {@link OllamaOptions.fetch} swaps the transport (default
+*   `globalThis.fetch`) and {@link OllamaOptions.headers} is a per-request, possibly
+*   async header injector merged over the base `Content-Type` — so a browser runtime
+*   can route through the developer's own server with an obfuscated bearer token,
+*   without this library ever handling a real API key. Both omitted ⇒ today's behaviour.
+*   Orthogonal to the deadline: the hook is awaited inside `#fetch`'s try, so a hook
+*   rejection clears the armed timer like any other request failure.
+*
+* @example
+* ```ts
+* const provider = new OllamaProvider({ model: 'qwen3.5:2b-q4_K_M' })
+* const result = await provider.generate(messages, abort.signal)
+* ```
+*/
+var OllamaProvider = class {
+	id = crypto.randomUUID();
+	name = "ollama";
+	#model;
+	#url;
+	#keepAlive;
+	#timeout;
+	#think;
+	#options;
+	#transport;
+	#headers;
+	#format;
+	constructor(options) {
+		this.#model = options.model;
+		this.#url = options.url ?? "http://localhost:11434";
+		this.#keepAlive = options.keepAlive ?? "5m";
+		this.#timeout = options.timeout ?? 12e4;
+		this.#think = options.think ?? false;
+		this.#options = options.options;
+		this.#transport = options.fetch ?? globalThis.fetch.bind(globalThis);
+		this.#headers = options.headers;
+		this.#format = options.format;
+	}
+	/**
+	* The provider's context-framing default — the PROVIDER-DEFAULT level of
+	* {@link import('@orkestrel/agent').AgentContextInterface.build}'s format cascade (it BEATS
+	* the managers' built-in framing, is BEATEN by a manager-options or per-item override).
+	* Satisfies the OPTIONAL {@link ProviderInterface.format} contract member: `undefined`
+	* when {@link OllamaOptions.format} was omitted (the framing-agnostic default ⇒ core's
+	* built-in framing applies unchanged), else the exact configured framing the Agent
+	* threads into `build()`.
+	*
+	* @remarks
+	* EXPOSE-ONLY — read by the Agent loop and consumed by core's cascade; it is NEVER sent
+	* on the `/api/chat` wire (it is absent from `#body` / the request). This is NOT Ollama's
+	* structured-output `format` wire parameter, which this provider does not send; only the
+	* word collides.
+	*
+	* @returns The configured {@link ContextFormatInterface}, or `undefined` when none
+	*/
+	get format() {
+		return this.#format;
+	}
+	async generate(messages, signal, tools, options) {
+		const { response, timeout } = await this.#fetch(messages, false, signal, tools, options);
+		try {
+			const record = await this.#parseBody(response);
+			const splitter = (0, _orkestrel_agent.createThinkSplitter)();
+			splitter.split(this.#content(record));
+			splitter.flush();
+			const thinking = this.#thought(splitter, this.#thinking(record));
+			return this.#result(splitter.content, thinking, this.#tools(record), this.#usage(record));
+		} finally {
+			timeout.clear();
+		}
+	}
+	async *stream(messages, signal, tools, options) {
+		const { response, timeout, combined } = await this.#fetch(messages, true, signal, tools, options);
+		const body = response.body;
+		if (body === null) {
+			timeout.clear();
+			throw new OllamaHTTPError("Ollama API error: no response body", 0);
+		}
+		const reader = body.getReader();
+		const decoder = new TextDecoder();
+		const parser = (0, _orkestrel_ndjson.createNDJSONParser)();
+		const splitter = (0, _orkestrel_agent.createThinkSplitter)();
+		let wired = "";
+		const calls = [];
+		let usage;
+		const contentOf = (record) => this.#content(record);
+		const thinkingOf = (record) => this.#thinking(record);
+		const toolsOf = (record) => this.#tools(record);
+		const usageOf = (record) => this.#usage(record);
+		function* handle(record) {
+			const delta = splitter.split(contentOf(record));
+			if (delta.length > 0) yield {
+				type: "content",
+				text: delta
+			};
+			const thinking = thinkingOf(record);
+			if (thinking.length > 0) yield {
+				type: "thinking",
+				text: thinking
+			};
+			wired += thinking;
+			calls.push(...toolsOf(record));
+			if (Reflect.get(record, "done") === true) usage = usageOf(record);
+		}
+		try {
+			for (;;) {
+				const { value, done } = await reader.read();
+				if (done) break;
+				for (const record of parser.parse(decoder.decode(value, { stream: true }))) yield* handle(record);
+			}
+			const decoderTail = decoder.decode();
+			for (const record of parser.parse(decoderTail.length > 0 ? `${decoderTail}\n` : "\n")) yield* handle(record);
+			const tail = splitter.flush();
+			if (tail.length > 0) yield {
+				type: "content",
+				text: tail
+			};
+		} catch (error) {
+			if (combined.aborted) {
+				splitter.flush();
+				throw new _orkestrel_agent.ProviderAbortError(this.#result(splitter.content, this.#thought(splitter, wired), calls, usage));
+			}
+			throw error;
+		} finally {
+			try {
+				await reader.cancel();
+			} catch {}
+			parser.reset();
+			timeout.clear();
+		}
+		return this.#result(splitter.content, this.#thought(splitter, wired), calls, usage);
+	}
+	async #fetch(messages, stream, signal, tools, options) {
+		const timeout = new _orkestrel_timeout.Timeout({ ms: this.#timeout });
+		timeout.start();
+		const combined = AbortSignal.any([timeout.signal, signal]);
+		try {
+			const response = await this.#transport(`${this.#url}/api/chat`, {
+				method: "POST",
+				headers: await this.#requestHeaders(),
+				body: JSON.stringify(this.#body(messages, stream, tools, options)),
+				signal: combined
+			});
+			if (!response.ok) {
+				let detail;
+				try {
+					const text = await response.text();
+					detail = text.length > 2048 ? text.slice(0, MAX_ERROR_BODY_LENGTH) : text;
+				} catch (cause) {
+					throw new OllamaHTTPError(`Ollama API error: ${response.status} - (error body unavailable)`, response.status, { cause });
+				}
+				throw new OllamaHTTPError(`Ollama API error: ${response.status} - ${detail}`, response.status);
+			}
+			return {
+				response,
+				timeout,
+				combined
+			};
+		} catch (error) {
+			timeout.clear();
+			throw error;
+		}
+	}
+	async #parseBody(response) {
+		const text = await response.text();
+		if (text.length === 0) return {};
+		try {
+			const data = JSON.parse(text);
+			return (0, _orkestrel_contract.isRecord)(data) ? data : {};
+		} catch {
+			return {};
+		}
+	}
+	async #requestHeaders() {
+		const headers = { "Content-Type": "application/json" };
+		if (this.#headers !== void 0) for (const [key, value] of Object.entries(await this.#headers())) headers[key] = value;
+		return headers;
+	}
+	#body(messages, stream, tools, options) {
+		return {
+			model: this.#model,
+			messages: this.#plain(messages),
+			stream,
+			keep_alive: this.#keepAlive,
+			think: options?.think ?? this.#think,
+			...this.#options !== void 0 ? { options: this.#options } : {},
+			...tools !== void 0 && tools.length > 0 ? { tools: tools.map((tool) => ({
+				type: "function",
+				function: {
+					name: tool.name,
+					description: tool.description,
+					parameters: tool.parameters
+				}
+			})) } : {}
+		};
+	}
+	#plain(messages) {
+		return messages.map((message) => ({
+			role: message.role,
+			content: message.content,
+			...message.calls !== void 0 && message.calls.length > 0 ? { tool_calls: message.calls.map((call) => ({ function: {
+				name: call.name,
+				arguments: call.arguments
+			} })) } : {},
+			...message.images !== void 0 && message.images.length > 0 ? { images: [...message.images] } : {}
+		}));
+	}
+	#result(content, thinking, tools, usage) {
+		const result = { content };
+		if (thinking.length > 0) result.thinking = thinking;
+		if (tools.length > 0) result.tools = tools;
+		if (usage !== void 0) result.usage = usage;
+		return result;
+	}
+	#content(record) {
+		const message = Reflect.get(record, "message");
+		if (!(0, _orkestrel_contract.isRecord)(message)) return "";
+		const content = Reflect.get(message, "content");
+		return (0, _orkestrel_contract.isString)(content) ? content : "";
+	}
+	#thinking(record) {
+		const message = Reflect.get(record, "message");
+		if (!(0, _orkestrel_contract.isRecord)(message)) return "";
+		const thinking = Reflect.get(message, "thinking");
+		return (0, _orkestrel_contract.isString)(thinking) ? thinking : "";
+	}
+	#thought(splitter, wired) {
+		if (splitter.thinking.length === 0) return wired;
+		if (wired.length === 0) return splitter.thinking;
+		return `${splitter.thinking}\n\n${wired}`;
+	}
+	#usage(record) {
+		const prompt = Reflect.get(record, "prompt_eval_count");
+		const completion = Reflect.get(record, "eval_count");
+		if (!(0, _orkestrel_contract.isNumber)(prompt) || !(0, _orkestrel_contract.isNumber)(completion)) return void 0;
+		return {
+			prompt,
+			completion,
+			total: prompt + completion
+		};
+	}
+	#tools(record) {
+		const message = Reflect.get(record, "message");
+		if (!(0, _orkestrel_contract.isRecord)(message)) return [];
+		const calls = Reflect.get(message, "tool_calls");
+		if (!Array.isArray(calls)) return [];
+		const out = [];
+		for (const entry of calls) {
+			if (!(0, _orkestrel_contract.isRecord)(entry)) continue;
+			const callable = Reflect.get(entry, "function");
+			if (!(0, _orkestrel_contract.isRecord)(callable)) continue;
+			const name = Reflect.get(callable, "name");
+			if (!(0, _orkestrel_contract.isString)(name)) continue;
+			const id = Reflect.get(entry, "id");
+			out.push({
+				id: (0, _orkestrel_contract.isString)(id) ? id : crypto.randomUUID(),
+				name,
+				arguments: this.#arguments(Reflect.get(callable, "arguments"))
+			});
+		}
+		return out;
+	}
+	#arguments(value) {
+		if ((0, _orkestrel_contract.isRecord)(value)) return value;
+		if ((0, _orkestrel_contract.isString)(value)) try {
+			const parsed = JSON.parse(value);
+			if ((0, _orkestrel_contract.isRecord)(parsed)) return parsed;
+		} catch {
+			return {};
+		}
+		return {};
+	}
+};
+//#endregion
+//#region src/server/factories.ts
+/**
+* Create a local Ollama inference provider — a {@link ProviderInterface} over the
+* daemon's `POST /api/chat`, supporting non-streaming `generate` and streaming
+* `stream`.
+*
+* @remarks
+* Only `model` is required; `url` defaults to the local daemon, `keepAlive` to `'5m'`,
+* `timeout` to `120_000`ms, and `options` is forwarded verbatim as sampling
+* parameters (`temperature` / `seed` / `num_predict` / …). Both calls take an
+* `AbortSignal` to bound the request; a `stream` cancelled mid-flight throws a
+* `ProviderAbortError` carrying the partial result.
+*
+* The optional `fetch` + `headers` form a transport seam (see {@link OllamaOptions}):
+* point `url` at your own server, inject a custom `fetch`, and have `headers` attach a
+* generated/obfuscated bearer token your server validates — so a browser runtime
+* reaches the LLM through your middleware WITHOUT this library ever handling the real API
+* key. Both omitted ⇒ today's behaviour (the global `fetch`, only a JSON content type).
+*
+* The optional `format` is the provider's context-framing default — the PROVIDER-DEFAULT
+* level of `AgentContext`'s format cascade (see [agents.md]; beaten by a manager-options
+* or per-item override, beating the managers' built-in framing), declaring how this
+* provider's models prefer context sections framed (e.g. XML group wrappers vs. Markdown
+* headers). It is EXPOSED on the provider for the Agent's `build()` and is NOT Ollama's
+* `/api/chat` `format` wire parameter (structured output) — the two are unrelated despite
+* the shared word. Omitted ⇒ the provider is framing-agnostic (core's built-in defaults).
+*
+* @param options - `model` (required), and optional `url` / `keepAlive` / `timeout` /
+*   `options` / `fetch` / `headers` / `format` (see {@link OllamaOptions})
+* @returns A working {@link ProviderInterface} backed by Ollama
+*
+* @example
+* ```ts
+* import { createAbort } from '@orkestrel/abort'
+* import { createOllama } from '@src/server'
+*
+* const provider = createOllama({ model: 'qwen3.5:2b-q4_K_M' })
+* const abort = createAbort()
+* const result = await provider.generate(messages, abort.signal)
+* ```
+*
+* @example
+* Route through your own server with an obfuscated token (deployment scenario S2):
+* ```ts
+* const provider = createOllama({
+*   model: 'qwen3.5:2b-q4_K_M',
+*   url: 'https://my-app.example.com/llm', // your server, not the daemon
+*   fetch: myFetch, // optional custom transport
+*   headers: () => ({ authorization: `Bearer ${myToken}` }), // your server validates this
+* })
+* ```
+*
+* @example
+* Declare a context-framing default — wrap the instructions section in an XML group (the
+* provider-default level of `AgentContext`'s cascade; NOT the wire `format`):
+* ```ts
+* const provider = createOllama({
+*   model: 'qwen3.5:2b-q4_K_M',
+*   format: {
+*     instructions: {
+*       open: '<instructions>',
+*       render: (i) => `<instruction>${i.content}</instruction>`,
+*       close: '</instructions>',
+*     },
+*   },
+* })
+* ```
+*/
+function createOllama(options) {
+	return new OllamaProvider(options);
+}
+//#endregion
+exports.DEFAULT_KEEP_ALIVE = DEFAULT_KEEP_ALIVE;
+exports.DEFAULT_OLLAMA_URL = DEFAULT_OLLAMA_URL;
+exports.DEFAULT_PROVIDER_TIMEOUT = DEFAULT_PROVIDER_TIMEOUT;
+exports.MAX_ERROR_BODY_LENGTH = MAX_ERROR_BODY_LENGTH;
+exports.OllamaHTTPError = OllamaHTTPError;
+exports.OllamaProvider = OllamaProvider;
+exports.createOllama = createOllama;
+exports.isOllamaHTTPError = isOllamaHTTPError;
+
+//# sourceMappingURL=index.cjs.map

package/dist/src/server/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.cjs","names":["#model","#url","#keepAlive","#timeout","#think","#options","#transport","#headers","#format","#fetch","#parseBody","#content","#thought","#thinking","#result","#tools","#usage","#requestHeaders","#body","#plain","#arguments"],"sources":["../../../src/server/constants.ts","../../../src/server/errors.ts","../../../src/server/OllamaProvider.ts","../../../src/server/factories.ts"],"sourcesContent":["// Ollama constants — the provider's defaults (AGENTS §5).\n\n/** The local Ollama daemon base URL assumed when `OllamaOptions.url` is omitted. */\nexport const DEFAULT_OLLAMA_URL = 'http://localhost:11434'\n\n/**\n * How long the model stays resident after a call when `OllamaOptions.keepAlive` is\n * omitted — Ollama's own `keep_alive` default, expressed as a duration string.\n */\nexport const DEFAULT_KEEP_ALIVE = '5m'\n\n/**\n * The per-call deadline in milliseconds when `OllamaOptions.timeout` is omitted —\n * generous enough that a cold model load does not trip it.\n */\nexport const DEFAULT_PROVIDER_TIMEOUT = 120_000\n\n/**\n * The cap, in characters, on how much of a non-OK response body is\n * incorporated into a thrown {@link OllamaHTTPError}'s message.\n *\n * @remarks\n * Bounds the excerpt so a defensive proxy or a misbehaving daemon handing\n * back an unbounded response body cannot inflate the thrown error's message\n * without limit (§14). `2048` characters is generous enough to carry a\n * useful diagnostic snippet while staying well short of any practical size\n * concern.\n */\nexport const MAX_ERROR_BODY_LENGTH = 2048\n","// Errors for the Ollama provider. A single `OllamaHTTPError` carries the\n// `/api/chat` HTTP status at the boundary — non-OK responses and a missing\n// response body both throw it — so a `catch` can branch on `error.status`\n// rather than parsing a message (AGENTS §12).\n\n/**\n * An error thrown when the Ollama `/api/chat` HTTP transport fails.\n *\n * @remarks\n * Carries the response `status` (0 when no HTTP response was received at all,\n * e.g. a `null` body). Thrown by {@link OllamaProvider} at its two HTTP\n * failure sites — the non-OK status branch and the null-body branch — so a\n * caller can branch on `error.status` instead of parsing the message. Narrow\n * a caught value with {@link isOllamaHTTPError}.\n *\n * @example\n * ```ts\n * try {\n * \tawait provider.generate(messages, signal)\n * } catch (error) {\n * \tif (isOllamaHTTPError(error) && error.status === 404) {\n * \t\t// the configured model isn't pulled\n * \t}\n * }\n * ```\n */\nexport class OllamaHTTPError extends Error {\n\treadonly status: number\n\n\tconstructor(message: string, status: number, options?: { readonly cause?: unknown }) {\n\t\tsuper(message, options)\n\t\tthis.name = 'OllamaHTTPError'\n\t\tthis.status = status\n\t}\n}\n\n/**\n * Whether a value is an {@link OllamaHTTPError}.\n *\n * @param value - The value to test\n * @returns `true` when `value` is an `OllamaHTTPError`\n */\nexport function isOllamaHTTPError(value: unknown): value is OllamaHTTPError {\n\treturn value instanceof OllamaHTTPError\n}\n","import type {\n\tContextFormatInterface,\n\tMessageInterface,\n\tProviderDelta,\n\tProviderInterface,\n\tProviderResult,\n\tProviderStreamOptions,\n\tThinkSplitterInterface,\n\tToolCall,\n\tToolDefinition,\n} from '@orkestrel/agent'\nimport type { TokenUsage } from '@orkestrel/budget'\nimport type { OllamaOptions, OllamaResponse, WireChatRequest } from './types.js'\nimport { createThinkSplitter, ProviderAbortError } from '@orkestrel/agent'\nimport { isNumber, isRecord, isString } from '@orkestrel/contract'\nimport { createNDJSONParser } from '@orkestrel/ndjson'\nimport { Timeout } from '@orkestrel/timeout'\nimport {\n\tDEFAULT_KEEP_ALIVE,\n\tDEFAULT_OLLAMA_URL,\n\tDEFAULT_PROVIDER_TIMEOUT,\n\tMAX_ERROR_BODY_LENGTH,\n} from './constants.js'\nimport { OllamaHTTPError } from './errors.js'\n\n/**\n * The local Ollama inference boundary — a {@link ProviderInterface} over Ollama's\n * `POST /api/chat`, both non-streaming (`generate`) and streaming NDJSON (`stream`).\n *\n * @remarks\n * - **Wire protocol.** Posts `{ model, messages, stream, keep_alive, think }` plus\n *   passthrough sampling `options` and mapped function `tools`. The `think` flag is\n *   CONFIGURABLE via {@link OllamaOptions.think} (default `false`). Non-stream parses\n *   one JSON body; stream consumes NDJSON (one JSON object per `\\n`-terminated line) —\n *   deltas carry `message.content`, the final `done: true` line carries the token usage.\n * - **Think separation (H4).** The wire `think` flag is configurable\n *   ({@link OllamaOptions.think}, default `false`). With `think: true` a thinking model's\n *   daemon separates reasoning NATIVELY — returning it on the distinct `message.thinking`\n *   channel (read here via `#thinking`) instead of inline in `message.content`. EITHER\n *   way the per-call {@link ThinkSplitterInterface} is the defensive guarantee: a daemon\n *   may ignore `think: false` for a thinking model and inline `<think>` tags, so every\n *   content delta routes through the splitter, only CLEAN content is yielded / assembled,\n *   and the separated reasoning (plus any daemon-side `message.thinking` deltas) lands on\n *   `ProviderResult.thinking`, never in the conversation.\n * - **Boundary narrowing (§14).** Every wire value arrives as `unknown` and is\n *   narrowed through guards (`isRecord` / `isString` / `isNumber`) — never `as`. A\n *   missing / malformed field degrades to a sensible default (empty content, no\n *   usage, `{}` arguments), never a throw.\n * - **Bounded.** Each call arms a {@link Timeout} for `OllamaOptions.timeout` and\n *   passes `AbortSignal.any([timeout.signal, signal])` to `fetch`, so the caller's\n *   signal AND the deadline both cancel the request. The timeout is always cleared —\n *   in `#fetch` if the request fails/aborts, otherwise in the consuming call's `finally`.\n * - **Abort recovers partial.** A `stream` cancelled mid-flight throws a\n *   `ProviderAbortError` carrying the partial result assembled so far; pairing the\n *   `TextDecoder({ stream: true })` with the {@link NDJSONParser} parser keeps multi-byte\n *   UTF-8 splits and partial lines honest.\n * - **Event-free.** A pure functional boundary — no Emitter, no events.\n * - **Transport seam.** {@link OllamaOptions.fetch} swaps the transport (default\n *   `globalThis.fetch`) and {@link OllamaOptions.headers} is a per-request, possibly\n *   async header injector merged over the base `Content-Type` — so a browser runtime\n *   can route through the developer's own server with an obfuscated bearer token,\n *   without this library ever handling a real API key. Both omitted ⇒ today's behaviour.\n *   Orthogonal to the deadline: the hook is awaited inside `#fetch`'s try, so a hook\n *   rejection clears the armed timer like any other request failure.\n *\n * @example\n * ```ts\n * const provider = new OllamaProvider({ model: 'qwen3.5:2b-q4_K_M' })\n * const result = await provider.generate(messages, abort.signal)\n * ```\n */\nexport class OllamaProvider implements ProviderInterface {\n\treadonly id = crypto.randomUUID()\n\treadonly name = 'ollama'\n\treadonly #model: string\n\treadonly #url: string\n\treadonly #keepAlive: string | number\n\treadonly #timeout: number\n\treadonly #think: boolean\n\treadonly #options: Readonly<Record<string, unknown>> | undefined\n\treadonly #transport: typeof globalThis.fetch\n\treadonly #headers: (() => Record<string, string> | Promise<Record<string, string>>) | undefined\n\treadonly #format: ContextFormatInterface | undefined\n\n\tconstructor(options: OllamaOptions) {\n\t\tthis.#model = options.model\n\t\tthis.#url = options.url ?? DEFAULT_OLLAMA_URL\n\t\tthis.#keepAlive = options.keepAlive ?? DEFAULT_KEEP_ALIVE\n\t\tthis.#timeout = options.timeout ?? DEFAULT_PROVIDER_TIMEOUT\n\t\t// The `/api/chat` `think` wire flag — DEFAULT `false` so the general-purpose provider\n\t\t// stays backward-compatible and immediate for non-thinking models. A thinking model whose\n\t\t// reasoning is DISPLAYED separately sets `think: true`, and the daemon then returns it on\n\t\t// the `message.thinking` channel (`#thinking`) rather than inline in `message.content`.\n\t\tthis.#think = options.think ?? false\n\t\tthis.#options = options.options\n\t\t// The transport seam (§21): a custom fetch (defaulting to the global, BOUND to its\n\t\t// `globalThis` receiver — invoking a bare reference through a field loses the `window`\n\t\t// receiver and browsers throw `Illegal invocation`; node's fetch is receiver-agnostic,\n\t\t// so only a browser runtime ever saw it) and a dynamic header injector — both omitted\n\t\t// by default, so today's behaviour is byte-identical (the global fetch, only the JSON\n\t\t// content type). The injected transport is `#transport` (the request METHOD already\n\t\t// owns the `#fetch` name).\n\t\tthis.#transport = options.fetch ?? globalThis.fetch.bind(globalThis)\n\t\tthis.#headers = options.headers\n\t\t// The context-framing default (the provider-DEFAULT level of AgentContext's format\n\t\t// cascade) — EXPOSE-ONLY: read by the Agent via `build(this.#provider.format)` and\n\t\t// consumed by core's cascade, it NEVER enters `#body` / the `/api/chat` wire. It is\n\t\t// NOT Ollama's structured-output `format` param (which this provider doesn't send) —\n\t\t// only the shared word collides. Omitted ⇒ undefined ⇒ core's built-in framing.\n\t\tthis.#format = options.format\n\t}\n\n\t/**\n\t * The provider's context-framing default — the PROVIDER-DEFAULT level of\n\t * {@link import('@orkestrel/agent').AgentContextInterface.build}'s format cascade (it BEATS\n\t * the managers' built-in framing, is BEATEN by a manager-options or per-item override).\n\t * Satisfies the OPTIONAL {@link ProviderInterface.format} contract member: `undefined`\n\t * when {@link OllamaOptions.format} was omitted (the framing-agnostic default ⇒ core's\n\t * built-in framing applies unchanged), else the exact configured framing the Agent\n\t * threads into `build()`.\n\t *\n\t * @remarks\n\t * EXPOSE-ONLY — read by the Agent loop and consumed by core's cascade; it is NEVER sent\n\t * on the `/api/chat` wire (it is absent from `#body` / the request). This is NOT Ollama's\n\t * structured-output `format` wire parameter, which this provider does not send; only the\n\t * word collides.\n\t *\n\t * @returns The configured {@link ContextFormatInterface}, or `undefined` when none\n\t */\n\tget format(): ContextFormatInterface | undefined {\n\t\treturn this.#format\n\t}\n\n\tasync generate(\n\t\tmessages: readonly MessageInterface[],\n\t\tsignal: AbortSignal,\n\t\ttools?: readonly ToolDefinition[],\n\t\toptions?: ProviderStreamOptions,\n\t): Promise<ProviderResult> {\n\t\tconst { response, timeout } = await this.#fetch(messages, false, signal, tools, options)\n\t\ttry {\n\t\t\tconst record = await this.#parseBody(response)\n\t\t\t// The one-body call routes through the SAME splitter as the stream (the daemon may\n\t\t\t// ignore `think: false` — the splitter is the guarantee): the assembled content is\n\t\t\t// CLEAN (the splitter's authoritative `content`, which also covers the qwen3\n\t\t\t// template's IMPLICIT leading open), the separated spans + any wire-side\n\t\t\t// `message.thinking` land on `thinking`.\n\t\t\tconst splitter = createThinkSplitter()\n\t\t\tsplitter.split(this.#content(record))\n\t\t\tsplitter.flush()\n\t\t\tconst thinking = this.#thought(splitter, this.#thinking(record))\n\t\t\treturn this.#result(splitter.content, thinking, this.#tools(record), this.#usage(record))\n\t\t} finally {\n\t\t\ttimeout.clear()\n\t\t}\n\t}\n\n\tasync *stream(\n\t\tmessages: readonly MessageInterface[],\n\t\tsignal: AbortSignal,\n\t\ttools?: readonly ToolDefinition[],\n\t\toptions?: ProviderStreamOptions,\n\t): AsyncGenerator<ProviderDelta, ProviderResult> {\n\t\tconst { response, timeout, combined } = await this.#fetch(\n\t\t\tmessages,\n\t\t\ttrue,\n\t\t\tsignal,\n\t\t\ttools,\n\t\t\toptions,\n\t\t)\n\t\tconst body = response.body\n\t\tif (body === null) {\n\t\t\ttimeout.clear()\n\t\t\tthrow new OllamaHTTPError('Ollama API error: no response body', 0)\n\t\t}\n\t\tconst reader = body.getReader()\n\t\tconst decoder = new TextDecoder()\n\t\tconst parser = createNDJSONParser()\n\t\t// The per-call think separator (H4): every wire content delta routes through it, so\n\t\t// only CLEAN content is yielded / assembled even when the daemon ignores `think: false`\n\t\t// for a thinking model; daemon-side `message.thinking` deltas accumulate beside it.\n\t\t// The ASSEMBLED content is the splitter's authoritative `content` — across the qwen3\n\t\t// template's IMPLICIT leading open (a bare `</think>` with the open pre-seeded into the\n\t\t// prompt scaffold) the splitter RECLASSIFIES the already-yielded prefix into `thinking`,\n\t\t// so the result stays clean even though those deltas could not be recalled.\n\t\tconst splitter = createThinkSplitter()\n\t\tlet wired = ''\n\t\tconst calls: ToolCall[] = []\n\t\tlet usage: TokenUsage | undefined\n\t\t// Per-record handling shared between the live loop and the post-loop NDJSON\n\t\t// tail flush below — captured over `this` so a nested `function*` (not an\n\t\t// arrow, which cannot be a generator) needs no rebinding.\n\t\tconst contentOf = (record: Record<string, unknown>): string => this.#content(record)\n\t\tconst thinkingOf = (record: Record<string, unknown>): string => this.#thinking(record)\n\t\tconst toolsOf = (record: Record<string, unknown>): readonly ToolCall[] => this.#tools(record)\n\t\tconst usageOf = (record: Record<string, unknown>): TokenUsage | undefined => this.#usage(record)\n\t\tfunction* handle(record: Record<string, unknown>): Generator<ProviderDelta> {\n\t\t\tconst delta = splitter.split(contentOf(record))\n\t\t\tif (delta.length > 0) yield { type: 'content', text: delta }\n\t\t\t// The PRIMARY live reasoning channel: each native `message.thinking` wire delta is\n\t\t\t// surfaced as a tagged `thinking` delta AND accumulated into `wired` for the assembled\n\t\t\t// result (the two stay in lockstep). The ThinkSplitter's in-content reclassified spans\n\t\t\t// have no per-delta hook — the final `ProviderResult.thinking` reconciles them; the\n\t\t\t// native channel (think: true) is what streams live.\n\t\t\tconst thinking = thinkingOf(record)\n\t\t\tif (thinking.length > 0) yield { type: 'thinking', text: thinking }\n\t\t\twired += thinking\n\t\t\tcalls.push(...toolsOf(record))\n\t\t\tif (Reflect.get(record, 'done') === true) usage = usageOf(record)\n\t\t}\n\t\ttry {\n\t\t\tfor (;;) {\n\t\t\t\tconst { value, done } = await reader.read()\n\t\t\t\tif (done) break\n\t\t\t\t// Pair the streaming decoder with the line parser: the decoder handles\n\t\t\t\t// partial multi-byte CHARS, the parser handles partial LINES (§14).\n\t\t\t\tfor (const record of parser.parse(decoder.decode(value, { stream: true }))) {\n\t\t\t\t\tyield* handle(record)\n\t\t\t\t}\n\t\t\t}\n\t\t\t// Flush the decoder's held partial multi-byte tail and feed it (plus a\n\t\t\t// terminating `\\n`) through the parser, so a non-conformant proxy's final\n\t\t\t// unterminated `done` line is recovered instead of silently dropped.\n\t\t\tconst decoderTail = decoder.decode()\n\t\t\tfor (const record of parser.parse(decoderTail.length > 0 ? `${decoderTail}\\n` : '\\n')) {\n\t\t\t\tyield* handle(record)\n\t\t\t}\n\t\t\t// Stream end: a held partial tag that never completed was real content — it is the\n\t\t\t// final delta (the splitter folds it into its `content` too).\n\t\t\tconst tail = splitter.flush()\n\t\t\tif (tail.length > 0) yield { type: 'content', text: tail }\n\t\t} catch (error) {\n\t\t\t// A mid-stream cancel (the caller's signal or the deadline) surfaces the\n\t\t\t// partial so the loop can recover what streamed; anything else propagates.\n\t\t\tif (combined.aborted) {\n\t\t\t\t// Flush the splitter's held partial tail first (mirrors the\n\t\t\t\t// normal-completion assembly above) so the recovered partial includes\n\t\t\t\t// any clean content that never crossed a tag boundary.\n\t\t\t\tsplitter.flush()\n\t\t\t\tthrow new ProviderAbortError(\n\t\t\t\t\tthis.#result(splitter.content, this.#thought(splitter, wired), calls, usage),\n\t\t\t\t)\n\t\t\t}\n\t\t\tthrow error\n\t\t} finally {\n\t\t\t// Cancel (not merely release) the reader on early return so the\n\t\t\t// underlying HTTP connection is freed; a normal-done or already-errored\n\t\t\t// reader tolerates the redundant cancel as a no-op. `cancel()` also\n\t\t\t// releases the lock — never call `releaseLock()` afterward.\n\t\t\ttry {\n\t\t\t\tawait reader.cancel()\n\t\t\t} catch {\n\t\t\t\t// Never mask the primary error/result with a cancel failure.\n\t\t\t}\n\t\t\tparser.reset()\n\t\t\ttimeout.clear()\n\t\t}\n\t\treturn this.#result(splitter.content, this.#thought(splitter, wired), calls, usage)\n\t}\n\n\t// Arm the deadline, POST `/api/chat`, and hand back the response + the handles\n\t// that bound it. On a non-OK status, clear the deadline and throw with the body.\n\tasync #fetch(\n\t\tmessages: readonly MessageInterface[],\n\t\tstream: boolean,\n\t\tsignal: AbortSignal,\n\t\ttools?: readonly ToolDefinition[],\n\t\toptions?: ProviderStreamOptions,\n\t): Promise<OllamaResponse> {\n\t\tconst timeout = new Timeout({ ms: this.#timeout })\n\t\ttimeout.start()\n\t\tconst combined = AbortSignal.any([timeout.signal, signal])\n\t\ttry {\n\t\t\tconst response = await this.#transport(`${this.#url}/api/chat`, {\n\t\t\t\tmethod: 'POST',\n\t\t\t\theaders: await this.#requestHeaders(),\n\t\t\t\tbody: JSON.stringify(this.#body(messages, stream, tools, options)),\n\t\t\t\tsignal: combined,\n\t\t\t})\n\t\t\tif (!response.ok) {\n\t\t\t\t// Bound the incorporated body: a defensive proxy or daemon could hand\n\t\t\t\t// back an unbounded response — read defensively so a body-read\n\t\t\t\t// failure still throws with the status, never a masked/unbounded read.\n\t\t\t\tlet detail: string\n\t\t\t\ttry {\n\t\t\t\t\tconst text = await response.text()\n\t\t\t\t\tdetail = text.length > MAX_ERROR_BODY_LENGTH ? text.slice(0, MAX_ERROR_BODY_LENGTH) : text\n\t\t\t\t} catch (cause) {\n\t\t\t\t\tthrow new OllamaHTTPError(\n\t\t\t\t\t\t`Ollama API error: ${response.status} - (error body unavailable)`,\n\t\t\t\t\t\tresponse.status,\n\t\t\t\t\t\t{ cause },\n\t\t\t\t\t)\n\t\t\t\t}\n\t\t\t\tthrow new OllamaHTTPError(\n\t\t\t\t\t`Ollama API error: ${response.status} - ${detail}`,\n\t\t\t\t\tresponse.status,\n\t\t\t\t)\n\t\t\t}\n\t\t\treturn { response, timeout, combined }\n\t\t} catch (error) {\n\t\t\t// `fetch` rejected (pre-aborted signal / unreachable / network) or the status\n\t\t\t// was non-OK — clear the deadline so the armed timer can't outlive the failed\n\t\t\t// call. The caller's `finally` only takes ownership once we return a response.\n\t\t\ttimeout.clear()\n\t\t\tthrow error\n\t\t}\n\t}\n\n\t// The non-stream `/api/chat` body — read the 200-OK response as text and parse it\n\t// inside a total guard (§14): a malformed or empty body degrades to `{}` (empty\n\t// content, no usage), never a raw `SyntaxError` escaping to the caller.\n\tasync #parseBody(response: Response): Promise<Record<string, unknown>> {\n\t\tconst text = await response.text()\n\t\tif (text.length === 0) return {}\n\t\ttry {\n\t\t\tconst data: unknown = JSON.parse(text)\n\t\t\treturn isRecord(data) ? data : {}\n\t\t} catch {\n\t\t\treturn {}\n\t\t}\n\t}\n\n\t// The request headers — the base JSON content type, plus the dynamic `headers`\n\t// hook's result merged ON TOP when configured (so a dev can attach an obfuscated\n\t// bearer the server validates). Merge order: `Content-Type` is seeded first, then\n\t// the hook's entries overlay it — so the hook ADDS auth headers but only clobbers\n\t// `Content-Type` if the dev explicitly returns one. Awaited (the hook may be async,\n\t// e.g. refreshing a token); called inside `#fetch`'s try so a hook rejection clears\n\t// the armed deadline like any other request failure. §14: the hook's result is a\n\t// `Record<string, string>` already — merged via `Object.entries`, no `as`.\n\tasync #requestHeaders(): Promise<Record<string, string>> {\n\t\tconst headers: Record<string, string> = { 'Content-Type': 'application/json' }\n\t\tif (this.#headers !== undefined) {\n\t\t\tfor (const [key, value] of Object.entries(await this.#headers())) headers[key] = value\n\t\t}\n\t\treturn headers\n\t}\n\n\t// The `/api/chat` request body — conditional `options` / `tools` only when set. The wire\n\t// `think` flag honours a PER-CALL override (`options.think`) over the constructor default\n\t// (`#think`), so a caller can flip reasoning on / off for one turn without reconfiguring the\n\t// provider; no per-call option ⇒ the constructed default, byte-for-byte the prior behaviour.\n\t#body(\n\t\tmessages: readonly MessageInterface[],\n\t\tstream: boolean,\n\t\ttools?: readonly ToolDefinition[],\n\t\toptions?: ProviderStreamOptions,\n\t): WireChatRequest {\n\t\treturn {\n\t\t\tmodel: this.#model,\n\t\t\tmessages: this.#plain(messages),\n\t\t\tstream,\n\t\t\tkeep_alive: this.#keepAlive,\n\t\t\tthink: options?.think ?? this.#think,\n\t\t\t...(this.#options !== undefined ? { options: this.#options } : {}),\n\t\t\t...(tools !== undefined && tools.length > 0\n\t\t\t\t? {\n\t\t\t\t\t\ttools: tools.map(\n\t\t\t\t\t\t\t(\n\t\t\t\t\t\t\t\ttool,\n\t\t\t\t\t\t\t): {\n\t\t\t\t\t\t\t\ttype: 'function'\n\t\t\t\t\t\t\t\tfunction: {\n\t\t\t\t\t\t\t\t\tname: string\n\t\t\t\t\t\t\t\t\tdescription?: string\n\t\t\t\t\t\t\t\t\tparameters?: Readonly<Record<string, unknown>>\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t} => ({\n\t\t\t\t\t\t\t\ttype: 'function',\n\t\t\t\t\t\t\t\tfunction: {\n\t\t\t\t\t\t\t\t\tname: tool.name,\n\t\t\t\t\t\t\t\t\tdescription: tool.description,\n\t\t\t\t\t\t\t\t\tparameters: tool.parameters,\n\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t}),\n\t\t\t\t\t\t),\n\t\t\t\t\t}\n\t\t\t\t: {}),\n\t\t}\n\t}\n\n\t// Map messages to the wire's minimal turn shape — `tool_calls` only on a turn\n\t// that replays them, `images` only on a multimodal turn (omit empty optionals).\n\t#plain(messages: readonly MessageInterface[]): WireChatRequest['messages'] {\n\t\treturn messages.map((message) => ({\n\t\t\trole: message.role,\n\t\t\tcontent: message.content,\n\t\t\t...(message.calls !== undefined && message.calls.length > 0\n\t\t\t\t? {\n\t\t\t\t\t\ttool_calls: message.calls.map((call) => ({\n\t\t\t\t\t\t\tfunction: { name: call.name, arguments: call.arguments },\n\t\t\t\t\t\t})),\n\t\t\t\t\t}\n\t\t\t\t: {}),\n\t\t\t// Forward multimodal image data — Ollama accepts a base64 `images` array on a\n\t\t\t// message, which a vision-capable model receives alongside the text content.\n\t\t\t...(message.images !== undefined && message.images.length > 0\n\t\t\t\t? { images: [...message.images] }\n\t\t\t\t: {}),\n\t\t}))\n\t}\n\n\t// Assemble a ProviderResult including only the present optionals — no empty\n\t// `thinking` / `tools`, no `usage` unless the wire reported it.\n\t#result(\n\t\tcontent: string,\n\t\tthinking: string,\n\t\ttools: readonly ToolCall[],\n\t\tusage: TokenUsage | undefined,\n\t): ProviderResult {\n\t\tconst result: {\n\t\t\tcontent: string\n\t\t\tthinking?: string\n\t\t\ttools?: readonly ToolCall[]\n\t\t\tusage?: TokenUsage\n\t\t} = { content }\n\t\tif (thinking.length > 0) result.thinking = thinking\n\t\tif (tools.length > 0) result.tools = tools\n\t\tif (usage !== undefined) result.usage = usage\n\t\treturn result\n\t}\n\n\t// The assistant text of one wire record — `message.content` when a string, else\n\t// `''` (a delta line, a tool-only turn, or a malformed shape).\n\t#content(record: Record<string, unknown>): string {\n\t\tconst message = Reflect.get(record, 'message')\n\t\tif (!isRecord(message)) return ''\n\t\tconst content = Reflect.get(message, 'content')\n\t\treturn isString(content) ? content : ''\n\t}\n\n\t// The daemon-side reasoning of one wire record — `message.thinking` when a string\n\t// (the `think: true` wire shape — surfaced when the configured `think` flag is on, and\n\t// handled defensively regardless since the daemon may vary), else `''`.\n\t#thinking(record: Record<string, unknown>): string {\n\t\tconst message = Reflect.get(record, 'message')\n\t\tif (!isRecord(message)) return ''\n\t\tconst thinking = Reflect.get(message, 'thinking')\n\t\treturn isString(thinking) ? thinking : ''\n\t}\n\n\t// Join a call's two reasoning carriers — the splitter's separated in-content spans and\n\t// the accumulated wire-side `message.thinking` — blank-line separated when both exist.\n\t#thought(splitter: ThinkSplitterInterface, wired: string): string {\n\t\tif (splitter.thinking.length === 0) return wired\n\t\tif (wired.length === 0) return splitter.thinking\n\t\treturn `${splitter.thinking}\\n\\n${wired}`\n\t}\n\n\t// Token usage from a wire record — only when BOTH counts are numbers (`done`\n\t// line / non-stream body); a delta line carries neither, so it yields undefined.\n\t#usage(record: Record<string, unknown>): TokenUsage | undefined {\n\t\tconst prompt = Reflect.get(record, 'prompt_eval_count')\n\t\tconst completion = Reflect.get(record, 'eval_count')\n\t\tif (!isNumber(prompt) || !isNumber(completion)) return undefined\n\t\treturn { prompt, completion, total: prompt + completion }\n\t}\n\n\t// Tool calls from a wire record's `message.tool_calls` — each entry narrowed to\n\t// `{ id, name, arguments }`, minting an id when the wire omits one and coercing a\n\t// JSON-string `arguments` to a record (defaulting to `{}`); §14, no `as`.\n\t#tools(record: Record<string, unknown>): readonly ToolCall[] {\n\t\tconst message = Reflect.get(record, 'message')\n\t\tif (!isRecord(message)) return []\n\t\tconst calls = Reflect.get(message, 'tool_calls')\n\t\tif (!Array.isArray(calls)) return []\n\t\tconst out: ToolCall[] = []\n\t\tfor (const entry of calls) {\n\t\t\tif (!isRecord(entry)) continue\n\t\t\tconst callable = Reflect.get(entry, 'function')\n\t\t\tif (!isRecord(callable)) continue\n\t\t\tconst name = Reflect.get(callable, 'name')\n\t\t\tif (!isString(name)) continue\n\t\t\tconst id = Reflect.get(entry, 'id')\n\t\t\tout.push({\n\t\t\t\tid: isString(id) ? id : crypto.randomUUID(),\n\t\t\t\tname,\n\t\t\t\targuments: this.#arguments(Reflect.get(callable, 'arguments')),\n\t\t\t})\n\t\t}\n\t\treturn out\n\t}\n\n\t// Narrow a wire `arguments` value to a record — an object as-is, a JSON string\n\t// parsed (when it yields a record), otherwise `{}`. Total: a bad string never throws.\n\t#arguments(value: unknown): Readonly<Record<string, unknown>> {\n\t\tif (isRecord(value)) return value\n\t\tif (isString(value)) {\n\t\t\ttry {\n\t\t\t\tconst parsed: unknown = JSON.parse(value)\n\t\t\t\tif (isRecord(parsed)) return parsed\n\t\t\t} catch {\n\t\t\t\treturn {}\n\t\t\t}\n\t\t}\n\t\treturn {}\n\t}\n}\n","import type { ProviderInterface } from '@orkestrel/agent'\nimport type { OllamaOptions } from './types.js'\nimport { OllamaProvider } from './OllamaProvider.js'\n\n/**\n * Create a local Ollama inference provider — a {@link ProviderInterface} over the\n * daemon's `POST /api/chat`, supporting non-streaming `generate` and streaming\n * `stream`.\n *\n * @remarks\n * Only `model` is required; `url` defaults to the local daemon, `keepAlive` to `'5m'`,\n * `timeout` to `120_000`ms, and `options` is forwarded verbatim as sampling\n * parameters (`temperature` / `seed` / `num_predict` / …). Both calls take an\n * `AbortSignal` to bound the request; a `stream` cancelled mid-flight throws a\n * `ProviderAbortError` carrying the partial result.\n *\n * The optional `fetch` + `headers` form a transport seam (see {@link OllamaOptions}):\n * point `url` at your own server, inject a custom `fetch`, and have `headers` attach a\n * generated/obfuscated bearer token your server validates — so a browser runtime\n * reaches the LLM through your middleware WITHOUT this library ever handling the real API\n * key. Both omitted ⇒ today's behaviour (the global `fetch`, only a JSON content type).\n *\n * The optional `format` is the provider's context-framing default — the PROVIDER-DEFAULT\n * level of `AgentContext`'s format cascade (see [agents.md]; beaten by a manager-options\n * or per-item override, beating the managers' built-in framing), declaring how this\n * provider's models prefer context sections framed (e.g. XML group wrappers vs. Markdown\n * headers). It is EXPOSED on the provider for the Agent's `build()` and is NOT Ollama's\n * `/api/chat` `format` wire parameter (structured output) — the two are unrelated despite\n * the shared word. Omitted ⇒ the provider is framing-agnostic (core's built-in defaults).\n *\n * @param options - `model` (required), and optional `url` / `keepAlive` / `timeout` /\n *   `options` / `fetch` / `headers` / `format` (see {@link OllamaOptions})\n * @returns A working {@link ProviderInterface} backed by Ollama\n *\n * @example\n * ```ts\n * import { createAbort } from '@orkestrel/abort'\n * import { createOllama } from '@src/server'\n *\n * const provider = createOllama({ model: 'qwen3.5:2b-q4_K_M' })\n * const abort = createAbort()\n * const result = await provider.generate(messages, abort.signal)\n * ```\n *\n * @example\n * Route through your own server with an obfuscated token (deployment scenario S2):\n * ```ts\n * const provider = createOllama({\n *   model: 'qwen3.5:2b-q4_K_M',\n *   url: 'https://my-app.example.com/llm', // your server, not the daemon\n *   fetch: myFetch, // optional custom transport\n *   headers: () => ({ authorization: `Bearer ${myToken}` }), // your server validates this\n * })\n * ```\n *\n * @example\n * Declare a context-framing default — wrap the instructions section in an XML group (the\n * provider-default level of `AgentContext`'s cascade; NOT the wire `format`):\n * ```ts\n * const provider = createOllama({\n *   model: 'qwen3.5:2b-q4_K_M',\n *   format: {\n *     instructions: {\n *       open: '<instructions>',\n *       render: (i) => `<instruction>${i.content}</instruction>`,\n *       close: '</instructions>',\n *     },\n *   },\n * })\n * ```\n */\nexport function createOllama(options: OllamaOptions): ProviderInterface {\n\treturn new OllamaProvider(options)\n}\n"],"mappings":";;;;;;;AAGA,IAAa,qBAAqB;;;;;AAMlC,IAAa,qBAAqB;;;;;AAMlC,IAAa,2BAA2B;;;;;;;;;;;;AAaxC,IAAa,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;ACFrC,IAAa,kBAAb,cAAqC,MAAM;CAC1C;CAEA,YAAY,SAAiB,QAAgB,SAAwC;EACpF,MAAM,SAAS,OAAO;EACtB,KAAK,OAAO;EACZ,KAAK,SAAS;CACf;AACD;;;;;;;AAQA,SAAgB,kBAAkB,OAA0C;CAC3E,OAAO,iBAAiB;AACzB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC2BA,IAAa,iBAAb,MAAyD;CACxD,KAAc,OAAO,WAAW;CAChC,OAAgB;CAChB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA,YAAY,SAAwB;EACnC,KAAKA,SAAS,QAAQ;EACtB,KAAKC,OAAO,QAAQ,OAAA;EACpB,KAAKC,aAAa,QAAQ,aAAA;EAC1B,KAAKC,WAAW,QAAQ,WAAA;EAKxB,KAAKC,SAAS,QAAQ,SAAS;EAC/B,KAAKC,WAAW,QAAQ;EAQxB,KAAKC,aAAa,QAAQ,SAAS,WAAW,MAAM,KAAK,UAAU;EACnE,KAAKC,WAAW,QAAQ;EAMxB,KAAKC,UAAU,QAAQ;CACxB;;;;;;;;;;;;;;;;;;CAmBA,IAAI,SAA6C;EAChD,OAAO,KAAKA;CACb;CAEA,MAAM,SACL,UACA,QACA,OACA,SAC0B;EAC1B,MAAM,EAAE,UAAU,YAAY,MAAM,KAAKC,OAAO,UAAU,OAAO,QAAQ,OAAO,OAAO;EACvF,IAAI;GACH,MAAM,SAAS,MAAM,KAAKC,WAAW,QAAQ;GAM7C,MAAM,YAAA,GAAA,iBAAA,oBAAA,CAA+B;GACrC,SAAS,MAAM,KAAKC,SAAS,MAAM,CAAC;GACpC,SAAS,MAAM;GACf,MAAM,WAAW,KAAKC,SAAS,UAAU,KAAKC,UAAU,MAAM,CAAC;GAC/D,OAAO,KAAKC,QAAQ,SAAS,SAAS,UAAU,KAAKC,OAAO,MAAM,GAAG,KAAKC,OAAO,MAAM,CAAC;EACzF,UAAU;GACT,QAAQ,MAAM;EACf;CACD;CAEA,OAAO,OACN,UACA,QACA,OACA,SACgD;EAChD,MAAM,EAAE,UAAU,SAAS,aAAa,MAAM,KAAKP,OAClD,UACA,MACA,QACA,OACA,OACD;EACA,MAAM,OAAO,SAAS;EACtB,IAAI,SAAS,MAAM;GAClB,QAAQ,MAAM;GACd,MAAM,IAAI,gBAAgB,sCAAsC,CAAC;EAClE;EACA,MAAM,SAAS,KAAK,UAAU;EAC9B,MAAM,UAAU,IAAI,YAAY;EAChC,MAAM,UAAA,GAAA,kBAAA,mBAAA,CAA4B;EAQlC,MAAM,YAAA,GAAA,iBAAA,oBAAA,CAA+B;EACrC,IAAI,QAAQ;EACZ,MAAM,QAAoB,CAAC;EAC3B,IAAI;EAIJ,MAAM,aAAa,WAA4C,KAAKE,SAAS,MAAM;EACnF,MAAM,cAAc,WAA4C,KAAKE,UAAU,MAAM;EACrF,MAAM,WAAW,WAAyD,KAAKE,OAAO,MAAM;EAC5F,MAAM,WAAW,WAA4D,KAAKC,OAAO,MAAM;EAC/F,UAAU,OAAO,QAA2D;GAC3E,MAAM,QAAQ,SAAS,MAAM,UAAU,MAAM,CAAC;GAC9C,IAAI,MAAM,SAAS,GAAG,MAAM;IAAE,MAAM;IAAW,MAAM;GAAM;GAM3D,MAAM,WAAW,WAAW,MAAM;GAClC,IAAI,SAAS,SAAS,GAAG,MAAM;IAAE,MAAM;IAAY,MAAM;GAAS;GAClE,SAAS;GACT,MAAM,KAAK,GAAG,QAAQ,MAAM,CAAC;GAC7B,IAAI,QAAQ,IAAI,QAAQ,MAAM,MAAM,MAAM,QAAQ,QAAQ,MAAM;EACjE;EACA,IAAI;GACH,SAAS;IACR,MAAM,EAAE,OAAO,SAAS,MAAM,OAAO,KAAK;IAC1C,IAAI,MAAM;IAGV,KAAK,MAAM,UAAU,OAAO,MAAM,QAAQ,OAAO,OAAO,EAAE,QAAQ,KAAK,CAAC,CAAC,GACxE,OAAO,OAAO,MAAM;GAEtB;GAIA,MAAM,cAAc,QAAQ,OAAO;GACnC,KAAK,MAAM,UAAU,OAAO,MAAM,YAAY,SAAS,IAAI,GAAG,YAAY,MAAM,IAAI,GACnF,OAAO,OAAO,MAAM;GAIrB,MAAM,OAAO,SAAS,MAAM;GAC5B,IAAI,KAAK,SAAS,GAAG,MAAM;IAAE,MAAM;IAAW,MAAM;GAAK;EAC1D,SAAS,OAAO;GAGf,IAAI,SAAS,SAAS;IAIrB,SAAS,MAAM;IACf,MAAM,IAAI,iBAAA,mBACT,KAAKF,QAAQ,SAAS,SAAS,KAAKF,SAAS,UAAU,KAAK,GAAG,OAAO,KAAK,CAC5E;GACD;GACA,MAAM;EACP,UAAU;GAKT,IAAI;IACH,MAAM,OAAO,OAAO;GACrB,QAAQ,CAER;GACA,OAAO,MAAM;GACb,QAAQ,MAAM;EACf;EACA,OAAO,KAAKE,QAAQ,SAAS,SAAS,KAAKF,SAAS,UAAU,KAAK,GAAG,OAAO,KAAK;CACnF;CAIA,MAAMH,OACL,UACA,QACA,QACA,OACA,SAC0B;EAC1B,MAAM,UAAU,IAAI,mBAAA,QAAQ,EAAE,IAAI,KAAKN,SAAS,CAAC;EACjD,QAAQ,MAAM;EACd,MAAM,WAAW,YAAY,IAAI,CAAC,QAAQ,QAAQ,MAAM,CAAC;EACzD,IAAI;GACH,MAAM,WAAW,MAAM,KAAKG,WAAW,GAAG,KAAKL,KAAK,YAAY;IAC/D,QAAQ;IACR,SAAS,MAAM,KAAKgB,gBAAgB;IACpC,MAAM,KAAK,UAAU,KAAKC,MAAM,UAAU,QAAQ,OAAO,OAAO,CAAC;IACjE,QAAQ;GACT,CAAC;GACD,IAAI,CAAC,SAAS,IAAI;IAIjB,IAAI;IACJ,IAAI;KACH,MAAM,OAAO,MAAM,SAAS,KAAK;KACjC,SAAS,KAAK,SAAA,OAAiC,KAAK,MAAM,GAAG,qBAAqB,IAAI;IACvF,SAAS,OAAO;KACf,MAAM,IAAI,gBACT,qBAAqB,SAAS,OAAO,8BACrC,SAAS,QACT,EAAE,MAAM,CACT;IACD;IACA,MAAM,IAAI,gBACT,qBAAqB,SAAS,OAAO,KAAK,UAC1C,SAAS,MACV;GACD;GACA,OAAO;IAAE;IAAU;IAAS;GAAS;EACtC,SAAS,OAAO;GAIf,QAAQ,MAAM;GACd,MAAM;EACP;CACD;CAKA,MAAMR,WAAW,UAAsD;EACtE,MAAM,OAAO,MAAM,SAAS,KAAK;EACjC,IAAI,KAAK,WAAW,GAAG,OAAO,CAAC;EAC/B,IAAI;GACH,MAAM,OAAgB,KAAK,MAAM,IAAI;GACrC,QAAA,GAAA,oBAAA,SAAA,CAAgB,IAAI,IAAI,OAAO,CAAC;EACjC,QAAQ;GACP,OAAO,CAAC;EACT;CACD;CAUA,MAAMO,kBAAmD;EACxD,MAAM,UAAkC,EAAE,gBAAgB,mBAAmB;EAC7E,IAAI,KAAKV,aAAa,KAAA,GACrB,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,MAAM,KAAKA,SAAS,CAAC,GAAG,QAAQ,OAAO;EAElF,OAAO;CACR;CAMA,MACC,UACA,QACA,OACA,SACkB;EAClB,OAAO;GACN,OAAO,KAAKP;GACZ,UAAU,KAAKmB,OAAO,QAAQ;GAC9B;GACA,YAAY,KAAKjB;GACjB,OAAO,SAAS,SAAS,KAAKE;GAC9B,GAAI,KAAKC,aAAa,KAAA,IAAY,EAAE,SAAS,KAAKA,SAAS,IAAI,CAAC;GAChE,GAAI,UAAU,KAAA,KAAa,MAAM,SAAS,IACvC,EACA,OAAO,MAAM,KAEX,UAQK;IACL,MAAM;IACN,UAAU;KACT,MAAM,KAAK;KACX,aAAa,KAAK;KAClB,YAAY,KAAK;IAClB;GACD,EACD,EACD,IACC,CAAC;EACL;CACD;CAIA,OAAO,UAAoE;EAC1E,OAAO,SAAS,KAAK,aAAa;GACjC,MAAM,QAAQ;GACd,SAAS,QAAQ;GACjB,GAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,MAAM,SAAS,IACvD,EACA,YAAY,QAAQ,MAAM,KAAK,UAAU,EACxC,UAAU;IAAE,MAAM,KAAK;IAAM,WAAW,KAAK;GAAU,EACxD,EAAE,EACH,IACC,CAAC;GAGJ,GAAI,QAAQ,WAAW,KAAA,KAAa,QAAQ,OAAO,SAAS,IACzD,EAAE,QAAQ,CAAC,GAAG,QAAQ,MAAM,EAAE,IAC9B,CAAC;EACL,EAAE;CACH;CAIA,QACC,SACA,UACA,OACA,OACiB;EACjB,MAAM,SAKF,EAAE,QAAQ;EACd,IAAI,SAAS,SAAS,GAAG,OAAO,WAAW;EAC3C,IAAI,MAAM,SAAS,GAAG,OAAO,QAAQ;EACrC,IAAI,UAAU,KAAA,GAAW,OAAO,QAAQ;EACxC,OAAO;CACR;CAIA,SAAS,QAAyC;EACjD,MAAM,UAAU,QAAQ,IAAI,QAAQ,SAAS;EAC7C,IAAI,EAAA,GAAA,oBAAA,SAAA,CAAU,OAAO,GAAG,OAAO;EAC/B,MAAM,UAAU,QAAQ,IAAI,SAAS,SAAS;EAC9C,QAAA,GAAA,oBAAA,SAAA,CAAgB,OAAO,IAAI,UAAU;CACtC;CAKA,UAAU,QAAyC;EAClD,MAAM,UAAU,QAAQ,IAAI,QAAQ,SAAS;EAC7C,IAAI,EAAA,GAAA,oBAAA,SAAA,CAAU,OAAO,GAAG,OAAO;EAC/B,MAAM,WAAW,QAAQ,IAAI,SAAS,UAAU;EAChD,QAAA,GAAA,oBAAA,SAAA,CAAgB,QAAQ,IAAI,WAAW;CACxC;CAIA,SAAS,UAAkC,OAAuB;EACjE,IAAI,SAAS,SAAS,WAAW,GAAG,OAAO;EAC3C,IAAI,MAAM,WAAW,GAAG,OAAO,SAAS;EACxC,OAAO,GAAG,SAAS,SAAS,MAAM;CACnC;CAIA,OAAO,QAAyD;EAC/D,MAAM,SAAS,QAAQ,IAAI,QAAQ,mBAAmB;EACtD,MAAM,aAAa,QAAQ,IAAI,QAAQ,YAAY;EACnD,IAAI,EAAA,GAAA,oBAAA,SAAA,CAAU,MAAM,KAAK,EAAA,GAAA,oBAAA,SAAA,CAAU,UAAU,GAAG,OAAO,KAAA;EACvD,OAAO;GAAE;GAAQ;GAAY,OAAO,SAAS;EAAW;CACzD;CAKA,OAAO,QAAsD;EAC5D,MAAM,UAAU,QAAQ,IAAI,QAAQ,SAAS;EAC7C,IAAI,EAAA,GAAA,oBAAA,SAAA,CAAU,OAAO,GAAG,OAAO,CAAC;EAChC,MAAM,QAAQ,QAAQ,IAAI,SAAS,YAAY;EAC/C,IAAI,CAAC,MAAM,QAAQ,KAAK,GAAG,OAAO,CAAC;EACnC,MAAM,MAAkB,CAAC;EACzB,KAAK,MAAM,SAAS,OAAO;GAC1B,IAAI,EAAA,GAAA,oBAAA,SAAA,CAAU,KAAK,GAAG;GACtB,MAAM,WAAW,QAAQ,IAAI,OAAO,UAAU;GAC9C,IAAI,EAAA,GAAA,oBAAA,SAAA,CAAU,QAAQ,GAAG;GACzB,MAAM,OAAO,QAAQ,IAAI,UAAU,MAAM;GACzC,IAAI,EAAA,GAAA,oBAAA,SAAA,CAAU,IAAI,GAAG;GACrB,MAAM,KAAK,QAAQ,IAAI,OAAO,IAAI;GAClC,IAAI,KAAK;IACR,KAAA,GAAA,oBAAA,SAAA,CAAa,EAAE,IAAI,KAAK,OAAO,WAAW;IAC1C;IACA,WAAW,KAAKe,WAAW,QAAQ,IAAI,UAAU,WAAW,CAAC;GAC9D,CAAC;EACF;EACA,OAAO;CACR;CAIA,WAAW,OAAmD;EAC7D,KAAA,GAAA,oBAAA,SAAA,CAAa,KAAK,GAAG,OAAO;EAC5B,KAAA,GAAA,oBAAA,SAAA,CAAa,KAAK,GACjB,IAAI;GACH,MAAM,SAAkB,KAAK,MAAM,KAAK;GACxC,KAAA,GAAA,oBAAA,SAAA,CAAa,MAAM,GAAG,OAAO;EAC9B,QAAQ;GACP,OAAO,CAAC;EACT;EAED,OAAO,CAAC;CACT;AACD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC3aA,SAAgB,aAAa,SAA2C;CACvE,OAAO,IAAI,eAAe,OAAO;AAClC"}