@orkestrel/ollama 0.0.14 → 0.0.16

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

@@ -1 +0,0 @@
-{"version":3,"file":"index.cjs","names":["#id","#model","#url","#keepAlive","#timeout","#think","#options","#transport","#headers","#format","#fetch","#deltas","#requestHeaders","#body"],"sources":["../../../src/server/constants.ts","../../../src/server/errors.ts","../../../src/server/helpers.ts","../../../src/server/parsers.ts","../../../src/server/OllamaProvider.ts","../../../src/server/factories.ts"],"sourcesContent":["// Ollama constants — the provider's defaults.\n\n/** Names the local Ollama daemon base URL assumed when `OllamaOptions.url` is omitted. */\nexport const DEFAULT_OLLAMA_URL = 'http://localhost:11434'\n\n/**\n * Names 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 *\n * @remarks\n * The name mirrors the Ollama `/api/chat` `keep_alive` field this value is sent as, so\n * the constant, the `OllamaOptions.keepAlive` key, and the wire member read as one term.\n */\nexport const DEFAULT_KEEP_ALIVE = '5m'\n\n/**\n * Names 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 * Names 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. `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.\n\nimport type { OllamaHTTPErrorOptions } from './types.js'\n\n/**\n * Represents an error thrown when the Ollama `/api/chat` HTTP transport fails.\n *\n * @remarks\n * Carries the machine-readable `code` `'HTTP'` and the response `status` (0 when no\n * HTTP response was received at all, for example a `null` body). Thrown by\n * {@link OllamaProvider} at its HTTP failure sites — the non-OK status branch and the\n * null-body branch — so a caller can branch on `error.code` and read `error.status`\n * for the HTTP number instead of parsing the message. Narrow a caught value with\n * {@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\t/**\n\t * Names the machine-readable condition this error reports — `'HTTP'`: an `/api/chat`\n\t * transport, status, or body failure.\n\t */\n\treadonly code = 'HTTP' as const\n\treadonly status: number\n\n\tconstructor(message: string, status: number, options?: OllamaHTTPErrorOptions) {\n\t\tsuper(message, options)\n\t\tthis.name = 'OllamaHTTPError'\n\t\tthis.status = status\n\t}\n}\n\n/**\n * Checks whether a value is an {@link OllamaHTTPError}.\n *\n * @param value - The value to test\n * @returns True if `value` is an `OllamaHTTPError`; false otherwise\n */\nexport function isOllamaHTTPError(value: unknown): value is OllamaHTTPError {\n\treturn value instanceof OllamaHTTPError\n}\n","// The Ollama wire leaves — the request projections and the response extractions\n// `OllamaProvider` composes. Each is a pure, total function of its parameters: a missing\n// or malformed wire field degrades to a sensible default (empty content, no usage, `{}`\n// arguments), never a throw, and no value is reached through `as`.\n\nimport type { Message, ProviderResult, ThinkSplitterInterface } from '@orkestrel/agent'\nimport type { TokenUsage } from '@orkestrel/budget'\nimport type { ToolCall } from '@orkestrel/tool'\nimport type { WireChatRequest } from './types.js'\nimport { isNumber, isRecord, isString, parseJSONAs } from '@orkestrel/contract'\n\n/**\n * Maps conversation turns onto the `/api/chat` wire's minimal message shape.\n *\n * @remarks\n * `tool_calls` is emitted only on a turn that replays them and `images` only on a\n * multimodal turn, so an empty optional never reaches the wire.\n *\n * @param messages - The conversation turns to send\n * @returns The wire `messages` array, one entry per turn, in order\n *\n * @example\n * ```ts\n * mapMessages([{ id: '1', role: 'user', content: 'Say hello.' }])\n * // [{ role: 'user', content: 'Say hello.' }]\n * ```\n */\nexport function mapMessages(messages: readonly Message[]): WireChatRequest['messages'] {\n\treturn messages.map((message) => ({\n\t\trole: message.role,\n\t\tcontent: message.content,\n\t\t...(message.calls !== undefined && message.calls.length > 0\n\t\t\t? {\n\t\t\t\t\ttool_calls: message.calls.map((call) => ({\n\t\t\t\t\t\tfunction: { name: call.name, arguments: call.arguments },\n\t\t\t\t\t})),\n\t\t\t\t}\n\t\t\t: {}),\n\t\t// Forward multimodal image data — Ollama accepts a base64 `images` array on a\n\t\t// message, which a vision-capable model receives alongside the text content.\n\t\t...(message.images !== undefined && message.images.length > 0\n\t\t\t? { images: [...message.images] }\n\t\t\t: {}),\n\t}))\n}\n\n/**\n * Builds a provider result from a turn's content, reasoning, tool calls, and usage.\n *\n * @remarks\n * Only the present optionals are set: no empty `thinking`, no empty `tools`, and no\n * `usage` unless the wire reported one.\n *\n * @param content - The clean assistant content the splitter accumulated\n * @param thinking - The joined reasoning, empty when the turn produced none\n * @param tools - The tool calls collected across the turn\n * @param usage - The token usage, or `undefined` when the wire reported none\n * @returns The result carrying only its populated fields\n *\n * @example\n * ```ts\n * buildResult('ok', '', [], undefined) // { content: 'ok' }\n * ```\n */\nexport function buildResult(\n\tcontent: string,\n\tthinking: string,\n\ttools: readonly ToolCall[],\n\tusage: TokenUsage | undefined,\n): ProviderResult {\n\tconst result: {\n\t\tcontent: string\n\t\tthinking?: string\n\t\ttools?: readonly ToolCall[]\n\t\tusage?: TokenUsage\n\t} = { content }\n\tif (thinking.length > 0) result.thinking = thinking\n\tif (tools.length > 0) result.tools = tools\n\tif (usage !== undefined) result.usage = usage\n\treturn result\n}\n\n/**\n * Extracts the assistant text of one wire record.\n *\n * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line\n * @returns The record's `message.content` when it is a string, else `''`\n *\n * @example\n * ```ts\n * extractContent({ message: { content: 'ok' } }) // 'ok'\n * ```\n */\nexport function extractContent(record: Readonly<Record<string, unknown>>): string {\n\tconst message = Reflect.get(record, 'message')\n\tif (!isRecord(message)) return ''\n\tconst content = Reflect.get(message, 'content')\n\treturn isString(content) ? content : ''\n}\n\n/**\n * Extracts the daemon-side reasoning of one wire record.\n *\n * @remarks\n * `message.thinking` is the `think: true` wire shape. It is read whatever the configured\n * flag says, because a daemon may separate reasoning on its own.\n *\n * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line\n * @returns The record's `message.thinking` when it is a string, else `''`\n *\n * @example\n * ```ts\n * extractThinking({ message: { thinking: 'weighing it' } }) // 'weighing it'\n * ```\n */\nexport function extractThinking(record: Readonly<Record<string, unknown>>): string {\n\tconst message = Reflect.get(record, 'message')\n\tif (!isRecord(message)) return ''\n\tconst thinking = Reflect.get(message, 'thinking')\n\treturn isString(thinking) ? thinking : ''\n}\n\n/**\n * Joins a call's two reasoning carriers into the result's `thinking`.\n *\n * @param splitter - The per-call splitter holding the separated in-content spans\n * @param wired - The accumulated wire-side `message.thinking` text\n * @returns The two carriers separated by a blank line, or whichever one is non-empty\n *\n * @example\n * ```ts\n * joinThinking(createThinkSplitter(), 'from the wire') // 'from the wire'\n * ```\n */\nexport function joinThinking(splitter: ThinkSplitterInterface, wired: string): string {\n\tif (splitter.thinking.length === 0) return wired\n\tif (wired.length === 0) return splitter.thinking\n\treturn `${splitter.thinking}\\n\\n${wired}`\n}\n\n/**\n * Extracts the token usage of one wire record.\n *\n * @remarks\n * Both counts must be numbers, which is true of the non-stream body and the stream's\n * `done: true` line. A delta line carries neither, so it yields `undefined`.\n *\n * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line\n * @returns The `TokenUsage` shape, or `undefined` when either count is absent\n *\n * @example\n * ```ts\n * extractUsage({ prompt_eval_count: 3, eval_count: 4 })\n * // { prompt: 3, completion: 4, total: 7 }\n * ```\n */\nexport function extractUsage(record: Readonly<Record<string, unknown>>): TokenUsage | undefined {\n\tconst prompt = Reflect.get(record, 'prompt_eval_count')\n\tconst completion = Reflect.get(record, 'eval_count')\n\tif (!isNumber(prompt) || !isNumber(completion)) return undefined\n\treturn { prompt, completion, total: prompt + completion }\n}\n\n/**\n * Extracts the tool calls of one wire record's `message.tool_calls`.\n *\n * @remarks\n * Each entry narrows to `{ id, name, arguments }`: the entry and its `function` must be\n * records and `name` a string, else the entry is dropped. An id is minted when the wire\n * omits one.\n *\n * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line\n * @returns The narrowed tool calls, empty when the record carries none\n *\n * @example\n * ```ts\n * extractTools({ message: { tool_calls: [{ function: { name: 'weather' } }] } })\n * // [{ id: '…', name: 'weather', arguments: {} }]\n * ```\n */\nexport function extractTools(record: Readonly<Record<string, unknown>>): readonly ToolCall[] {\n\tconst message = Reflect.get(record, 'message')\n\tif (!isRecord(message)) return []\n\tconst calls = Reflect.get(message, 'tool_calls')\n\tif (!Array.isArray(calls)) return []\n\tconst out: ToolCall[] = []\n\tfor (const entry of calls) {\n\t\tif (!isRecord(entry)) continue\n\t\tconst callable = Reflect.get(entry, 'function')\n\t\tif (!isRecord(callable)) continue\n\t\tconst name = Reflect.get(callable, 'name')\n\t\tif (!isString(name)) continue\n\t\tconst id = Reflect.get(entry, 'id')\n\t\tout.push({\n\t\t\tid: isString(id) ? id : crypto.randomUUID(),\n\t\t\tname,\n\t\t\targuments: extractArguments(Reflect.get(callable, 'arguments')),\n\t\t})\n\t}\n\treturn out\n}\n\n/**\n * Extracts a wire `arguments` value as a record.\n *\n * @remarks\n * Total: an object passes through, a JSON string is parsed when it yields a record, and\n * a malformed string yields `{}` rather than throwing.\n *\n * @param value - The wire's `function.arguments` value, of unknown shape\n * @returns The argument record, or `{}` when the value carries none\n *\n * @example\n * ```ts\n * extractArguments('{\"city\":\"Oslo\"}') // { city: 'Oslo' }\n * ```\n */\nexport function extractArguments(value: unknown): Readonly<Record<string, unknown>> {\n\tif (isRecord(value)) return value\n\tif (isString(value)) return parseJSONAs(value, isRecord) ?? {}\n\treturn {}\n}\n","// The Ollama response coercer — the non-stream `/api/chat` body read off the wire and\n// coerced to a record inside a total guard, never a raw `SyntaxError`.\n\nimport { isRecord, parseJSONAs } from '@orkestrel/contract'\n\n/**\n * Parses a non-stream `/api/chat` response body into a wire record.\n *\n * @remarks\n * Total by construction: an empty body, a body that is not JSON, and a body whose JSON is\n * not an object all yield `undefined`, so a malformed daemon response never escapes as a\n * `SyntaxError`. The call site supplies the empty-record default that reads as empty\n * content and no usage.\n *\n * @param response - The 200-OK `/api/chat` response whose body is read as text\n * @returns The parsed record, or `undefined` when the body is empty or malformed\n *\n * @example\n * ```ts\n * await parseBody(new Response('{\"message\":{\"content\":\"ok\"}}'))\n * // { message: { content: 'ok' } }\n * ```\n */\nexport async function parseBody(\n\tresponse: Response,\n): Promise<Readonly<Record<string, unknown>> | undefined> {\n\treturn parseJSONAs(await response.text(), isRecord)\n}\n","import type {\n\tContextFormat,\n\tMessage,\n\tProviderDelta,\n\tProviderInterface,\n\tProviderResult,\n\tProviderStreamOptions,\n\tThinkSplitterInterface,\n} from '@orkestrel/agent'\nimport type { TokenUsage } from '@orkestrel/budget'\nimport type { ToolCall, ToolDefinition } from '@orkestrel/tool'\nimport type { OllamaOptions, OllamaResponse, WireChatRequest } from './types.js'\nimport { createThinkSplitter, ProviderAbortError } from '@orkestrel/agent'\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'\nimport {\n\tbuildResult,\n\textractContent,\n\textractThinking,\n\textractTools,\n\textractUsage,\n\tjoinThinking,\n\tmapMessages,\n} from './helpers.js'\nimport { parseBody } from './parsers.js'\n\n/**\n * Implements 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 through {@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.** 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 through `extractThinking`) 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.** 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 ⇒ the global `fetch`\n *   and only a JSON content type.\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 name = 'ollama'\n\treadonly #id: string\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:\n\t\t| (() => Readonly<Record<string, string>> | Promise<Readonly<Record<string, string>>>)\n\t\t| undefined\n\treadonly #format: ContextFormat | undefined\n\n\tconstructor(options: OllamaOptions) {\n\t\tthis.#id = crypto.randomUUID()\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 a non-thinking model needs no\n\t\t// configuration and answers immediately. A thinking model whose reasoning is DISPLAYED\n\t\t// separately sets `think: true`, and the daemon then returns it on the\n\t\t// `message.thinking` channel (`extractThinking`) 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: 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 the request goes out over the global fetch carrying 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 through `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` wire param — that one IS sent in `#body`,\n\t\t// but only when a per-call `ProviderStreamOptions.schema` is supplied; the two\n\t\t// merely share a word. Omitted ⇒ undefined ⇒ core's built-in framing.\n\t\tthis.#format = options.format\n\t}\n\n\t/**\n\t * Exposes this instance's identity — a fresh `crypto.randomUUID()` minted at\n\t * construction, satisfying the {@link ProviderInterface.id} contract member. A second\n\t * provider built from identical options carries a distinct id.\n\t *\n\t * @returns The instance's minted identifier\n\t */\n\tget id(): string {\n\t\treturn this.#id\n\t}\n\n\t/**\n\t * Exposes 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 — that one IS sent in `#body`, but only when\n\t * a per-call `ProviderStreamOptions.schema` is supplied; only the word collides.\n\t *\n\t * @returns The configured {@link ContextFormat}, or `undefined` when none\n\t */\n\tget format(): ContextFormat | undefined {\n\t\treturn this.#format\n\t}\n\n\tasync generate(\n\t\tmessages: readonly Message[],\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 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(extractContent(record))\n\t\t\tsplitter.flush()\n\t\t\tconst thinking = joinThinking(splitter, extractThinking(record))\n\t\t\treturn buildResult(splitter.content, thinking, extractTools(record), extractUsage(record))\n\t\t} finally {\n\t\t\ttimeout.clear()\n\t\t}\n\t}\n\n\tasync *stream(\n\t\tmessages: readonly Message[],\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: 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\t// The per-stream accumulators, folded from every `#deltas` return across the live\n\t\t// loop and the post-loop NDJSON tail flush following.\n\t\tlet wired = ''\n\t\tconst calls: ToolCall[] = []\n\t\tlet usage: TokenUsage | undefined\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.\n\t\t\t\tfor (const record of parser.parse(decoder.decode(value, { stream: true }))) {\n\t\t\t\t\tconst increment = yield* this.#deltas(record, splitter, usage)\n\t\t\t\t\twired += increment.thinking\n\t\t\t\t\tcalls.push(...increment.calls)\n\t\t\t\t\tusage = increment.usage\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\tconst increment = yield* this.#deltas(record, splitter, usage)\n\t\t\t\twired += increment.thinking\n\t\t\t\tcalls.push(...increment.calls)\n\t\t\t\tusage = increment.usage\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 { channel: '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 preceding) 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\tbuildResult(splitter.content, joinThinking(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.clear()\n\t\t\ttimeout.clear()\n\t\t}\n\t\treturn buildResult(splitter.content, joinThinking(splitter, wired), calls, usage)\n\t}\n\n\t// Per-record streaming step shared between the live NDJSON loop and the post-loop\n\t// tail flush in `stream()` — a `#` private method (not a free helper) because it is\n\t// the streaming spine that composes the wire leaves and drives the splitter, and\n\t// because its yields are the stream's own. It mutates nothing: it RETURNS the record's\n\t// increments (`thinking` / `calls` / `usage`) and `stream()` folds them, so the\n\t// accumulator's shape is written once, here.\n\t*#deltas(\n\t\trecord: Readonly<Record<string, unknown>>,\n\t\tsplitter: ThinkSplitterInterface,\n\t\tusage: TokenUsage | undefined,\n\t): Generator<\n\t\tProviderDelta,\n\t\t{\n\t\t\treadonly thinking: string\n\t\t\treadonly calls: readonly ToolCall[]\n\t\t\treadonly usage: TokenUsage | undefined\n\t\t}\n\t> {\n\t\tconst delta = splitter.split(extractContent(record))\n\t\tif (delta.length > 0) yield { channel: 'content', text: delta }\n\t\t// The PRIMARY live reasoning channel: each native `message.thinking` wire delta is\n\t\t// surfaced as a tagged `thinking` delta AND returned for the caller's `wired`\n\t\t// accumulation (the two stay in lockstep). The ThinkSplitter's in-content\n\t\t// reclassified spans have no per-delta hook — the final `ProviderResult.thinking`\n\t\t// reconciles them; the native channel (think: true) is what streams live.\n\t\tconst thinking = extractThinking(record)\n\t\tif (thinking.length > 0) yield { channel: 'thinking', text: thinking }\n\t\t// Only the `done` line carries usage, so every other record hands the caller's\n\t\t// current value straight back rather than clearing it.\n\t\treturn {\n\t\t\tthinking,\n\t\t\tcalls: extractTools(record),\n\t\t\tusage: Reflect.get(record, 'done') === true ? extractUsage(record) : usage,\n\t\t}\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 Message[],\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 `#fetch` returns a response.\n\t\t\ttimeout.clear()\n\t\t\tthrow error\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// for example refreshing a token); called inside `#fetch`'s try so a hook rejection\n\t// clears the armed deadline like any other request failure. The hook's result is a\n\t// `Readonly<Record<string, string>>` already — merged through `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` / `format` only when set. The\n\t// wire `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.\n\t// `format` is the wire's structured-output constraint, forwarded verbatim from the per-call\n\t// `ProviderStreamOptions.schema` — unrelated to `OllamaOptions.format` (prompt-context framing).\n\t#body(\n\t\tmessages: readonly Message[],\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: mapMessages(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...(options?.schema !== undefined ? { format: options.schema } : {}),\n\t\t\t...(tools !== undefined && tools.length > 0\n\t\t\t\t? {\n\t\t\t\t\t\ttools: tools.map((tool): NonNullable<WireChatRequest['tools']>[number] => ({\n\t\t\t\t\t\t\ttype: 'function',\n\t\t\t\t\t\t\tfunction: {\n\t\t\t\t\t\t\t\tname: tool.name,\n\t\t\t\t\t\t\t\t...(tool.description === undefined ? {} : { description: tool.description }),\n\t\t\t\t\t\t\t\t...(tool.parameters === undefined ? {} : { parameters: tool.parameters }),\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","import type { ProviderInterface } from '@orkestrel/agent'\nimport type { OllamaOptions } from './types.js'\nimport { OllamaProvider } from './OllamaProvider.js'\n\n/**\n * Creates 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`, and `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 ⇒ the global `fetch` and 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 (beaten by a manager-options or per-item\n * override, beating the managers' built-in framing), declaring how this\n * provider's models prefer context sections framed (for example 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 '@orkestrel/ollama'\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:\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;;;;;;;;;AAUlC,IAAa,qBAAqB;;;;;AAMlC,IAAa,2BAA2B;;;;;;;;;;;;AAaxC,IAAa,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;ACHrC,IAAa,kBAAb,cAAqC,MAAM;;;;;CAK1C,OAAgB;CAChB;CAEA,YAAY,SAAiB,QAAgB,SAAkC;EAC9E,MAAM,SAAS,OAAO;EACtB,KAAK,OAAO;EACZ,KAAK,SAAS;CACf;AACD;;;;;;;AAQA,SAAgB,kBAAkB,OAA0C;CAC3E,OAAO,iBAAiB;AACzB;;;;;;;;;;;;;;;;;;;ACzBA,SAAgB,YAAY,UAA2D;CACtF,OAAO,SAAS,KAAK,aAAa;EACjC,MAAM,QAAQ;EACd,SAAS,QAAQ;EACjB,GAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,MAAM,SAAS,IACvD,EACA,YAAY,QAAQ,MAAM,KAAK,UAAU,EACxC,UAAU;GAAE,MAAM,KAAK;GAAM,WAAW,KAAK;EAAU,EACxD,EAAE,EACH,IACC,CAAC;EAGJ,GAAI,QAAQ,WAAW,KAAA,KAAa,QAAQ,OAAO,SAAS,IACzD,EAAE,QAAQ,CAAC,GAAG,QAAQ,MAAM,EAAE,IAC9B,CAAC;CACL,EAAE;AACH;;;;;;;;;;;;;;;;;;;AAoBA,SAAgB,YACf,SACA,UACA,OACA,OACiB;CACjB,MAAM,SAKF,EAAE,QAAQ;CACd,IAAI,SAAS,SAAS,GAAG,OAAO,WAAW;CAC3C,IAAI,MAAM,SAAS,GAAG,OAAO,QAAQ;CACrC,IAAI,UAAU,KAAA,GAAW,OAAO,QAAQ;CACxC,OAAO;AACR;;;;;;;;;;;;AAaA,SAAgB,eAAe,QAAmD;CACjF,MAAM,UAAU,QAAQ,IAAI,QAAQ,SAAS;CAC7C,IAAI,EAAA,GAAC,oBAAA,SAAA,CAAS,OAAO,GAAG,OAAO;CAC/B,MAAM,UAAU,QAAQ,IAAI,SAAS,SAAS;CAC9C,QAAA,GAAO,oBAAA,SAAA,CAAS,OAAO,IAAI,UAAU;AACtC;;;;;;;;;;;;;;;;AAiBA,SAAgB,gBAAgB,QAAmD;CAClF,MAAM,UAAU,QAAQ,IAAI,QAAQ,SAAS;CAC7C,IAAI,EAAA,GAAC,oBAAA,SAAA,CAAS,OAAO,GAAG,OAAO;CAC/B,MAAM,WAAW,QAAQ,IAAI,SAAS,UAAU;CAChD,QAAA,GAAO,oBAAA,SAAA,CAAS,QAAQ,IAAI,WAAW;AACxC;;;;;;;;;;;;;AAcA,SAAgB,aAAa,UAAkC,OAAuB;CACrF,IAAI,SAAS,SAAS,WAAW,GAAG,OAAO;CAC3C,IAAI,MAAM,WAAW,GAAG,OAAO,SAAS;CACxC,OAAO,GAAG,SAAS,SAAS,MAAM;AACnC;;;;;;;;;;;;;;;;;AAkBA,SAAgB,aAAa,QAAmE;CAC/F,MAAM,SAAS,QAAQ,IAAI,QAAQ,mBAAmB;CACtD,MAAM,aAAa,QAAQ,IAAI,QAAQ,YAAY;CACnD,IAAI,EAAA,GAAC,oBAAA,SAAA,CAAS,MAAM,KAAK,EAAA,GAAC,oBAAA,SAAA,CAAS,UAAU,GAAG,OAAO,KAAA;CACvD,OAAO;EAAE;EAAQ;EAAY,OAAO,SAAS;CAAW;AACzD;;;;;;;;;;;;;;;;;;AAmBA,SAAgB,aAAa,QAAgE;CAC5F,MAAM,UAAU,QAAQ,IAAI,QAAQ,SAAS;CAC7C,IAAI,EAAA,GAAC,oBAAA,SAAA,CAAS,OAAO,GAAG,OAAO,CAAC;CAChC,MAAM,QAAQ,QAAQ,IAAI,SAAS,YAAY;CAC/C,IAAI,CAAC,MAAM,QAAQ,KAAK,GAAG,OAAO,CAAC;CACnC,MAAM,MAAkB,CAAC;CACzB,KAAK,MAAM,SAAS,OAAO;EAC1B,IAAI,EAAA,GAAC,oBAAA,SAAA,CAAS,KAAK,GAAG;EACtB,MAAM,WAAW,QAAQ,IAAI,OAAO,UAAU;EAC9C,IAAI,EAAA,GAAC,oBAAA,SAAA,CAAS,QAAQ,GAAG;EACzB,MAAM,OAAO,QAAQ,IAAI,UAAU,MAAM;EACzC,IAAI,EAAA,GAAC,oBAAA,SAAA,CAAS,IAAI,GAAG;EACrB,MAAM,KAAK,QAAQ,IAAI,OAAO,IAAI;EAClC,IAAI,KAAK;GACR,KAAA,GAAI,oBAAA,SAAA,CAAS,EAAE,IAAI,KAAK,OAAO,WAAW;GAC1C;GACA,WAAW,iBAAiB,QAAQ,IAAI,UAAU,WAAW,CAAC;EAC/D,CAAC;CACF;CACA,OAAO;AACR;;;;;;;;;;;;;;;;AAiBA,SAAgB,iBAAiB,OAAmD;CACnF,KAAA,GAAI,oBAAA,SAAA,CAAS,KAAK,GAAG,OAAO;CAC5B,KAAA,GAAI,oBAAA,SAAA,CAAS,KAAK,GAAG,QAAA,GAAO,oBAAA,YAAA,CAAY,OAAO,oBAAA,QAAQ,KAAK,CAAC;CAC7D,OAAO,CAAC;AACT;;;;;;;;;;;;;;;;;;;;;ACtMA,eAAsB,UACrB,UACyD;CACzD,QAAA,GAAO,oBAAA,YAAA,CAAY,MAAM,SAAS,KAAK,GAAG,oBAAA,QAAQ;AACnD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACqDA,IAAa,iBAAb,MAAyD;CACxD,OAAgB;CAChB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAGA;CAEA,YAAY,SAAwB;EACnC,KAAKA,MAAM,OAAO,WAAW;EAC7B,KAAKC,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;EAOxB,KAAKC,UAAU,QAAQ;CACxB;;;;;;;;CASA,IAAI,KAAa;EAChB,OAAO,KAAKT;CACb;;;;;;;;;;;;;;;;;;CAmBA,IAAI,SAAoC;EACvC,OAAO,KAAKS;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,SAAU,MAAM,UAAU,QAAQ,KAAM,CAAC;GAM/C,MAAM,YAAA,GAAW,iBAAA,oBAAA,CAAoB;GACrC,SAAS,MAAM,eAAe,MAAM,CAAC;GACrC,SAAS,MAAM;GACf,MAAM,WAAW,aAAa,UAAU,gBAAgB,MAAM,CAAC;GAC/D,OAAO,YAAY,SAAS,SAAS,UAAU,aAAa,MAAM,GAAG,aAAa,MAAM,CAAC;EAC1F,UAAU;GACT,QAAQ,MAAM;EACf;CACD;CAEA,OAAO,OACN,UACA,QACA,OACA,SACgD;EAChD,MAAM,EAAE,UAAU,SAAS,aAAa,MAAM,KAAKA,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,GAAS,kBAAA,mBAAA,CAAmB;EAQlC,MAAM,YAAA,GAAW,iBAAA,oBAAA,CAAoB;EAGrC,IAAI,QAAQ;EACZ,MAAM,QAAoB,CAAC;EAC3B,IAAI;EACJ,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,GAAG;KAC3E,MAAM,YAAY,OAAO,KAAKC,QAAQ,QAAQ,UAAU,KAAK;KAC7D,SAAS,UAAU;KACnB,MAAM,KAAK,GAAG,UAAU,KAAK;KAC7B,QAAQ,UAAU;IACnB;GACD;GAIA,MAAM,cAAc,QAAQ,OAAO;GACnC,KAAK,MAAM,UAAU,OAAO,MAAM,YAAY,SAAS,IAAI,GAAG,YAAY,MAAM,IAAI,GAAG;IACtF,MAAM,YAAY,OAAO,KAAKA,QAAQ,QAAQ,UAAU,KAAK;IAC7D,SAAS,UAAU;IACnB,MAAM,KAAK,GAAG,UAAU,KAAK;IAC7B,QAAQ,UAAU;GACnB;GAGA,MAAM,OAAO,SAAS,MAAM;GAC5B,IAAI,KAAK,SAAS,GAAG,MAAM;IAAE,SAAS;IAAW,MAAM;GAAK;EAC7D,SAAS,OAAO;GAGf,IAAI,SAAS,SAAS;IAIrB,SAAS,MAAM;IACf,MAAM,IAAI,iBAAA,mBACT,YAAY,SAAS,SAAS,aAAa,UAAU,KAAK,GAAG,OAAO,KAAK,CAC1E;GACD;GACA,MAAM;EACP,UAAU;GAKT,IAAI;IACH,MAAM,OAAO,OAAO;GACrB,QAAQ,CAER;GACA,OAAO,MAAM;GACb,QAAQ,MAAM;EACf;EACA,OAAO,YAAY,SAAS,SAAS,aAAa,UAAU,KAAK,GAAG,OAAO,KAAK;CACjF;CAQA,CAACA,QACA,QACA,UACA,OAQC;EACD,MAAM,QAAQ,SAAS,MAAM,eAAe,MAAM,CAAC;EACnD,IAAI,MAAM,SAAS,GAAG,MAAM;GAAE,SAAS;GAAW,MAAM;EAAM;EAM9D,MAAM,WAAW,gBAAgB,MAAM;EACvC,IAAI,SAAS,SAAS,GAAG,MAAM;GAAE,SAAS;GAAY,MAAM;EAAS;EAGrE,OAAO;GACN;GACA,OAAO,aAAa,MAAM;GAC1B,OAAO,QAAQ,IAAI,QAAQ,MAAM,MAAM,OAAO,aAAa,MAAM,IAAI;EACtE;CACD;CAIA,MAAMD,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,KAAKU,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;CAUA,MAAMD,kBAAmD;EACxD,MAAM,UAAkC,EAAE,gBAAgB,mBAAmB;EAC7E,IAAI,KAAKJ,aAAa,KAAA,GACrB,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,MAAM,KAAKA,SAAS,CAAC,GAAG,QAAQ,OAAO;EAElF,OAAO;CACR;CAQA,MACC,UACA,QACA,OACA,SACkB;EAClB,OAAO;GACN,OAAO,KAAKP;GACZ,UAAU,YAAY,QAAQ;GAC9B;GACA,YAAY,KAAKE;GACjB,OAAO,SAAS,SAAS,KAAKE;GAC9B,GAAI,KAAKC,aAAa,KAAA,IAAY,EAAE,SAAS,KAAKA,SAAS,IAAI,CAAC;GAChE,GAAI,SAAS,WAAW,KAAA,IAAY,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;GAClE,GAAI,UAAU,KAAA,KAAa,MAAM,SAAS,IACvC,EACA,OAAO,MAAM,KAAK,UAAyD;IAC1E,MAAM;IACN,UAAU;KACT,MAAM,KAAK;KACX,GAAI,KAAK,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,KAAK,YAAY;KAC1E,GAAI,KAAK,eAAe,KAAA,IAAY,CAAC,IAAI,EAAE,YAAY,KAAK,WAAW;IACxE;GACD,EAAE,EACH,IACC,CAAC;EACL;CACD;AACD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC/UA,SAAgB,aAAa,SAA2C;CACvE,OAAO,IAAI,eAAe,OAAO;AAClC"}

package/dist/src/server/index.d.cts

@@ -1,542 +0,0 @@
-import { ContextFormat } from '@orkestrel/agent';
-import { Message } from '@orkestrel/agent';
-import { ProviderDelta } from '@orkestrel/agent';
-import { ProviderInterface } from '@orkestrel/agent';
-import { ProviderResult } from '@orkestrel/agent';
-import { ProviderStreamOptions } from '@orkestrel/agent';
-import { ThinkSplitterInterface } from '@orkestrel/agent';
-import { TimeoutInterface } from '@orkestrel/timeout';
-import { TokenUsage } from '@orkestrel/budget';
-import { ToolCall } from '@orkestrel/tool';
-import { ToolDefinition } from '@orkestrel/tool';
-
-/**
- * Builds a provider result from a turn's content, reasoning, tool calls, and usage.
- *
- * @remarks
- * Only the present optionals are set: no empty `thinking`, no empty `tools`, and no
- * `usage` unless the wire reported one.
- *
- * @param content - The clean assistant content the splitter accumulated
- * @param thinking - The joined reasoning, empty when the turn produced none
- * @param tools - The tool calls collected across the turn
- * @param usage - The token usage, or `undefined` when the wire reported none
- * @returns The result carrying only its populated fields
- *
- * @example
- * ```ts
- * buildResult('ok', '', [], undefined) // { content: 'ok' }
- * ```
- */
-export declare function buildResult(content: string, thinking: string, tools: readonly ToolCall[], usage: TokenUsage | undefined): ProviderResult;
-
-/**
- * Creates 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`, and `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 ⇒ the global `fetch` and only a JSON content type.
- *
- * The optional `format` is the provider's context-framing default — the PROVIDER-DEFAULT
- * level of `AgentContext`'s format cascade (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 (for example 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 '@orkestrel/ollama'
- *
- * 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:
- * ```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>',
- *     },
- *   },
- * })
- * ```
- */
-export declare function createOllama(options: OllamaOptions): ProviderInterface;
-
-/**
- * Names 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.
- *
- * @remarks
- * The name mirrors the Ollama `/api/chat` `keep_alive` field this value is sent as, so
- * the constant, the `OllamaOptions.keepAlive` key, and the wire member read as one term.
- */
-export declare const DEFAULT_KEEP_ALIVE = "5m";
-
-/** Names the local Ollama daemon base URL assumed when `OllamaOptions.url` is omitted. */
-export declare const DEFAULT_OLLAMA_URL = "http://localhost:11434";
-
-/**
- * Names the per-call deadline in milliseconds when `OllamaOptions.timeout` is omitted —
- * generous enough that a cold model load does not trip it.
- */
-export declare const DEFAULT_PROVIDER_TIMEOUT = 120000;
-
-/**
- * Extracts a wire `arguments` value as a record.
- *
- * @remarks
- * Total: an object passes through, a JSON string is parsed when it yields a record, and
- * a malformed string yields `{}` rather than throwing.
- *
- * @param value - The wire's `function.arguments` value, of unknown shape
- * @returns The argument record, or `{}` when the value carries none
- *
- * @example
- * ```ts
- * extractArguments('{"city":"Oslo"}') // { city: 'Oslo' }
- * ```
- */
-export declare function extractArguments(value: unknown): Readonly<Record<string, unknown>>;
-
-/**
- * Extracts the assistant text of one wire record.
- *
- * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line
- * @returns The record's `message.content` when it is a string, else `''`
- *
- * @example
- * ```ts
- * extractContent({ message: { content: 'ok' } }) // 'ok'
- * ```
- */
-export declare function extractContent(record: Readonly<Record<string, unknown>>): string;
-
-/**
- * Extracts the daemon-side reasoning of one wire record.
- *
- * @remarks
- * `message.thinking` is the `think: true` wire shape. It is read whatever the configured
- * flag says, because a daemon may separate reasoning on its own.
- *
- * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line
- * @returns The record's `message.thinking` when it is a string, else `''`
- *
- * @example
- * ```ts
- * extractThinking({ message: { thinking: 'weighing it' } }) // 'weighing it'
- * ```
- */
-export declare function extractThinking(record: Readonly<Record<string, unknown>>): string;
-
-/**
- * Extracts the tool calls of one wire record's `message.tool_calls`.
- *
- * @remarks
- * Each entry narrows to `{ id, name, arguments }`: the entry and its `function` must be
- * records and `name` a string, else the entry is dropped. An id is minted when the wire
- * omits one.
- *
- * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line
- * @returns The narrowed tool calls, empty when the record carries none
- *
- * @example
- * ```ts
- * extractTools({ message: { tool_calls: [{ function: { name: 'weather' } }] } })
- * // [{ id: '…', name: 'weather', arguments: {} }]
- * ```
- */
-export declare function extractTools(record: Readonly<Record<string, unknown>>): readonly ToolCall[];
-
-/**
- * Extracts the token usage of one wire record.
- *
- * @remarks
- * Both counts must be numbers, which is true of the non-stream body and the stream's
- * `done: true` line. A delta line carries neither, so it yields `undefined`.
- *
- * @param record - One parsed `/api/chat` record — a non-stream body or an NDJSON line
- * @returns The `TokenUsage` shape, or `undefined` when either count is absent
- *
- * @example
- * ```ts
- * extractUsage({ prompt_eval_count: 3, eval_count: 4 })
- * // { prompt: 3, completion: 4, total: 7 }
- * ```
- */
-export declare function extractUsage(record: Readonly<Record<string, unknown>>): TokenUsage | undefined;
-
-/**
- * Checks whether a value is an {@link OllamaHTTPError}.
- *
- * @param value - The value to test
- * @returns True if `value` is an `OllamaHTTPError`; false otherwise
- */
-export declare function isOllamaHTTPError(value: unknown): value is OllamaHTTPError;
-
-/**
- * Joins a call's two reasoning carriers into the result's `thinking`.
- *
- * @param splitter - The per-call splitter holding the separated in-content spans
- * @param wired - The accumulated wire-side `message.thinking` text
- * @returns The two carriers separated by a blank line, or whichever one is non-empty
- *
- * @example
- * ```ts
- * joinThinking(createThinkSplitter(), 'from the wire') // 'from the wire'
- * ```
- */
-export declare function joinThinking(splitter: ThinkSplitterInterface, wired: string): string;
-
-/**
- * Maps conversation turns onto the `/api/chat` wire's minimal message shape.
- *
- * @remarks
- * `tool_calls` is emitted only on a turn that replays them and `images` only on a
- * multimodal turn, so an empty optional never reaches the wire.
- *
- * @param messages - The conversation turns to send
- * @returns The wire `messages` array, one entry per turn, in order
- *
- * @example
- * ```ts
- * mapMessages([{ id: '1', role: 'user', content: 'Say hello.' }])
- * // [{ role: 'user', content: 'Say hello.' }]
- * ```
- */
-export declare function mapMessages(messages: readonly Message[]): WireChatRequest['messages'];
-
-/**
- * Names 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. `2048` characters is generous enough to carry a
- * useful diagnostic snippet while staying well short of any practical size
- * concern.
- */
-export declare const MAX_ERROR_BODY_LENGTH = 2048;
-
-/**
- * Represents an error thrown when the Ollama `/api/chat` HTTP transport fails.
- *
- * @remarks
- * Carries the machine-readable `code` `'HTTP'` and the response `status` (0 when no
- * HTTP response was received at all, for example a `null` body). Thrown by
- * {@link OllamaProvider} at its HTTP failure sites — the non-OK status branch and the
- * null-body branch — so a caller can branch on `error.code` and read `error.status`
- * for the HTTP number 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
- * 	}
- * }
- * ```
- */
-export declare class OllamaHTTPError extends Error {
-    /**
-     * Names the machine-readable condition this error reports — `'HTTP'`: an `/api/chat`
-     * transport, status, or body failure.
-     */
-    readonly code: "HTTP";
-    readonly status: number;
-    constructor(message: string, status: number, options?: OllamaHTTPErrorOptions);
-}
-
-/**
- * Represents the options a thrown {@link OllamaHTTPError} accepts beside its message and status —
- * the standard error `cause` link, named so a consumer can reference the shape.
- *
- * @remarks
- * `cause` is the underlying value that produced the HTTP failure: the transport or
- * body-read rejection the provider caught before rethrowing. It is `unknown` because a
- * thrown value is unconstrained. Omitted ⇒ the error carries no cause.
- */
-export declare interface OllamaHTTPErrorOptions {
-    readonly cause?: unknown;
-}
-
-/**
- * Represents the configuration `createOllama` accepts for the local Ollama backend.
- *
- * @remarks
- * Only `model` is required. `url` defaults to the local daemon, `keepAlive` controls
- * how long the model stays resident after a call, `timeout` is the per-call deadline
- * in milliseconds, and `options` is a passthrough bag of sampling parameters
- * (`temperature`, `seed`, and `num_predict`) forwarded verbatim to the wire.
- *
- * The optional `fetch` + `headers` form a **transport seam**: by default the provider
- * talks straight to a local daemon over `globalThis.fetch` with only a JSON content
- * type, but a browser-side runtime can inject a custom transport AND a dynamic header
- * (for example an obfuscated bearer token) so requests route through the developer's OWN
- * server, which validates that header and forwards to the real LLM. Your app never
- * holds a real API key — the real key lives only on the developer's server; the
- * `headers` hook supplies whatever short-lived/obfuscated token that server expects.
- */
-export declare interface OllamaOptions {
-    readonly model: string;
-    /** Sets the daemon base URL; defaults to `'http://localhost:11434'`. */
-    readonly url?: string;
-    /**
-     * Sets how long the model stays resident after a call; defaults to `'5m'`. Mirrors the
-     * Ollama `/api/chat` `keep_alive` field, whose value this key carries verbatim onto
-     * {@link WireChatRequest.keep_alive}.
-     */
-    readonly keepAlive?: string | number;
-    /** Sets the per-call deadline in milliseconds; defaults to `120_000`. */
-    readonly timeout?: number;
-    /**
-     * Carries passthrough sampling parameters (`temperature`, `seed`, and `num_predict`).
-     * Mirrors the Ollama `/api/chat` `options` field, whose value this key carries verbatim
-     * onto {@link WireChatRequest.options}.
-     */
-    readonly options?: Readonly<Record<string, unknown>>;
-    /**
-     * Sets the `/api/chat` `think` wire flag; defaults to `false`. When `true`, a thinking-capable
-     * model (for example `qwen3`) separates its reasoning NATIVELY at the wire — the daemon returns it
-     * on the distinct `message.thinking` channel (surfaced on `ProviderResult.thinking`) rather
-     * than inline in `message.content`. The default is `false`, so a non-thinking model needs no
-     * configuration and answers immediately; the per-call ThinkSplitter
-     * remains the defensive fallback for daemons/models that still inline `<think>` tags either
-     * way. Set it `true` for a thinking model whose reasoning you intend to DISPLAY separately.
-     */
-    readonly think?: boolean;
-    /**
-     * Sets a custom `fetch` implementation for every request; defaults to
-     * `globalThis.fetch`. Lets a runtime inject its own transport (a browser fetch
-     * pointed at the developer's server, an instrumented wrapper) without changing
-     * the wire protocol. Omitted ⇒ the global `fetch`.
-     */
-    readonly fetch?: typeof globalThis.fetch;
-    /**
-     * Sets a dynamic, possibly-async header injector called once per request; its returned
-     * headers are merged into the request on top of the base `Content-Type`. Use it to
-     * attach an authorization header — for example an obfuscated/generated bearer token the
-     * developer's server validates before relaying to the real LLM — so a browser
-     * runtime can authenticate WITHOUT your app ever handling a real API key. Async so a
-     * token can be refreshed/fetched per call. A returned `Content-Type` overrides the
-     * default; other headers add to it. Omitted ⇒ only `Content-Type: application/json`.
-     */
-    readonly headers?: () => Readonly<Record<string, string>> | Promise<Readonly<Record<string, string>>>;
-    /**
-     * Sets the provider's OPTIONAL context-framing default — the PROVIDER-DEFAULT level of
-     * `AgentContext`'s format cascade (beaten by a manager-options or per-item override,
-     * beating the managers' built-in framing). Declares how this provider's models prefer
-     * context sections framed (for example XML group wrappers vs. Markdown headers). Omitted ⇒
-     * the provider is framing-agnostic and core's built-in defaults apply unchanged. NOTE:
-     * this is the prompt-CONTEXT framing consumed by `AgentContext.build()` — it is NOT
-     * Ollama's `/api/chat` `format` wire parameter (structured-output / JSON schema),
-     * which this provider sends only when a call supplies a `schema`; the two are unrelated
-     * despite the shared word.
-     */
-    readonly format?: ContextFormat;
-}
-
-/**
- * Implements 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 through {@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.** 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 through `extractThinking`) 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.** 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 ⇒ the global `fetch`
- *   and only a JSON content type.
- *   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)
- * ```
- */
-export declare class OllamaProvider implements ProviderInterface {
-    #private;
-    readonly name = "ollama";
-    constructor(options: OllamaOptions);
-    /**
-     * Exposes this instance's identity — a fresh `crypto.randomUUID()` minted at
-     * construction, satisfying the {@link ProviderInterface.id} contract member. A second
-     * provider built from identical options carries a distinct id.
-     *
-     * @returns The instance's minted identifier
-     */
-    get id(): string;
-    /**
-     * Exposes 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 — that one IS sent in `#body`, but only when
-     * a per-call `ProviderStreamOptions.schema` is supplied; only the word collides.
-     *
-     * @returns The configured {@link ContextFormat}, or `undefined` when none
-     */
-    get format(): ContextFormat | undefined;
-    generate(messages: readonly Message[], signal: AbortSignal, tools?: readonly ToolDefinition[], options?: ProviderStreamOptions): Promise<ProviderResult>;
-    stream(messages: readonly Message[], signal: AbortSignal, tools?: readonly ToolDefinition[], options?: ProviderStreamOptions): AsyncGenerator<ProviderDelta, ProviderResult>;
-}
-
-/**
- * Represents an open `POST /api/chat` response together with the deadline and the
- * combined signal that bound the request.
- *
- * @remarks
- * The `response` is the open `POST /api/chat` `Response`; `timeout` is the armed
- * {@link TimeoutInterface} the consuming call clears once it finishes reading the body
- * (or that the provider clears on a failed or aborted request); `combined` is the
- * `AbortSignal.any([timeout.signal, callerSignal])` the request was issued under, which
- * the streaming path checks to tell a mid-stream cancel apart from any other error.
- */
-export declare interface OllamaResponse {
-    readonly response: Response;
-    readonly timeout: TimeoutInterface;
-    readonly combined: AbortSignal;
-}
-
-/**
- * Parses a non-stream `/api/chat` response body into a wire record.
- *
- * @remarks
- * Total by construction: an empty body, a body that is not JSON, and a body whose JSON is
- * not an object all yield `undefined`, so a malformed daemon response never escapes as a
- * `SyntaxError`. The call site supplies the empty-record default that reads as empty
- * content and no usage.
- *
- * @param response - The 200-OK `/api/chat` response whose body is read as text
- * @returns The parsed record, or `undefined` when the body is empty or malformed
- *
- * @example
- * ```ts
- * await parseBody(new Response('{"message":{"content":"ok"}}'))
- * // { message: { content: 'ok' } }
- * ```
- */
-export declare function parseBody(response: Response): Promise<Readonly<Record<string, unknown>> | undefined>;
-
-/**
- * Represents the exact `POST /api/chat` request body `OllamaProvider` sends — the internal typed
- * wire contract.
- *
- * @remarks
- * This is the typed wire shape asserted against the official `ollama` client's
- * `ChatRequest` by the compile-time parity test; `src/` never imports `ollama` itself.
- * `messages` mirrors the minimal turn shape `mapMessages` builds (`role` / `content`, plus
- * `tool_calls` only on a turn that replays them and `images` only on a multimodal
- * turn); `options` and `tools` are only present when configured.
- */
-export declare interface WireChatRequest {
-    readonly model: string;
-    readonly messages: ReadonlyArray<{
-        readonly role: string;
-        readonly content: string;
-        readonly tool_calls?: ReadonlyArray<{
-            readonly function: {
-                readonly name: string;
-                readonly arguments: Readonly<Record<string, unknown>>;
-            };
-        }>;
-        readonly images?: readonly string[];
-    }>;
-    readonly stream: boolean;
-    readonly keep_alive: string | number;
-    readonly think: boolean;
-    readonly options?: Readonly<Record<string, unknown>>;
-    readonly tools?: ReadonlyArray<{
-        readonly type: 'function';
-        readonly function: {
-            readonly name: string;
-            readonly description?: string;
-            readonly parameters?: Readonly<Record<string, unknown>>;
-        };
-    }>;
-    /**
-     * Holds the `/api/chat` structured-output constraint — a JSON-Schema object forwarded
-     * verbatim from the per-call `ProviderStreamOptions.schema`. This is NOT
-     * `OllamaOptions.format` (the unrelated prompt-context framing); only present
-     * when a call supplies a `schema`.
-     */
-    readonly format?: Readonly<Record<string, unknown>>;
-}
-
-export { }