agentfootprint 9.72.0 → 9.74.0

package/dist/adapters/llm/FoundryLocalProvider.js

@@ -0,0 +1,992 @@
+"use strict";
+/**
+ * FoundryLocalProvider — on-device models over Foundry Local's
+ * OpenAI-compatible `/v1/chat/completions`.
+ *
+ * Pattern: Adapter (GoF) + Ports-and-Adapters (Cockburn 2005).
+ * Role:    Outer ring — translates `LLMRequest`/`LLMResponse` to/from the
+ *          wire Foundry Local serves on localhost. Knows nothing about
+ *          agents, recorders, or compositions.
+ * Emits:   N/A.
+ *
+ * ─── Why this exists ─────────────────────────────────────────────────
+ *
+ * The adapter ladder is `mock()` → a local model → a paid API, and
+ * `ollama()` is the proof that the middle rung is worth owning: zero
+ * dependencies, honest refusals, real token counts. Foundry Local is
+ * Microsoft's runtime for the same rung — ONNX under the hood, models
+ * pulled with `foundry model run <alias>`, no key, no account — and a
+ * Windows or macOS machine that has it installed deserves the same
+ * one-import experience. So this file owns that wire the way
+ * `OllamaProvider.ts` owns Ollama's:
+ *
+ *   • ZERO dependencies — one `fetch` POST and SSE. The official
+ *     `foundry-local-sdk` is NOT imported; its manager is accepted
+ *     duck-typed (see {@link FoundryLocalProviderOptions.manager}) so a
+ *     consumer who already uses it can hand over the discovered URL
+ *     without this package gaining a dependency.
+ *   • HONEST REFUSALS — a typed {@link FoundryLocalUnavailableError}
+ *     that names the endpoint it tried and the command to run. The
+ *     service's port is DYNAMIC per start, which makes "nothing is
+ *     answering" the most likely first failure — so that message
+ *     carries the discovery command, not just the start command. A
+ *     failure the service reports IN BAND — an `error` frame on an
+ *     already-200 stream, the out-of-memory a laptop runtime really does
+ *     hit — is RAISED the same way, never handed over as a shorter
+ *     answer that reads like a clean stop.
+ *   • REAL TOKEN COUNTS while streaming — `stream_options:
+ *     { include_usage: true }` is always sent. This is OUR wire, a
+ *     documented Foundry Local surface, not an arbitrary
+ *     OpenAI-compatible server — so the caution that made
+ *     `openai({ baseURL })` withhold the field (and silently zero every
+ *     local token count until 9.73.0) does not apply here.
+ *
+ * ─── Wire realities this file owns ───────────────────────────────────
+ *
+ * • THE PORT IS DYNAMIC. Every `foundry server start` may pick a new
+ *   port; the docs' own REST example shows `http://localhost:5272` and
+ *   that is the default here, but the truthful discovery is
+ *   `foundry server status` (or the SDK manager's `.urls`). Note the
+ *   CLI group was RENAMED from `foundry service` to `foundry server` —
+ *   every message in this file uses the NEW spelling.
+ * • ALIASES vs VARIANT IDS. The catalog speaks in aliases
+ *   (`qwen2.5-0.5b`) that fan out to hardware variants
+ *   (`qwen2.5-0.5b-instruct-generic-cpu:1`), but REST chat calls take
+ *   the FULL variant id. This adapter resolves an alias through
+ *   `GET /foundry/list` — first matching variant wins, because the
+ *   list's order IS the service's priority order — and caches the
+ *   answer per provider instance, HIT OR MISS: exactly one catalog
+ *   attempt per name, so an alias the catalog never answers for cannot
+ *   re-ask before every call. A fresh provider is the retry. A name that
+ *   already carries a variant's execution-provider suffix
+ *   (`-cpu`/`-gpu`/`-npu`, optional `:version`) is used as-is with no
+ *   catalog round-trip.
+ * • NO API KEY EXISTS. The docs' own samples pass placeholders. This
+ *   adapter sends no `Authorization` header at all — there is nothing
+ *   to put in one, and an invented value would only end up in somebody's
+ *   proxy log.
+ *
+ * ─── Ceilings (stated, not worked around) ────────────────────────────
+ *
+ * • NO FORCED TOOL CHOICE. `tool_choice` support is UNDOCUMENTED on
+ *   this wire, so `carriesForcedToolChoice` is `false` and an agent
+ *   using `.outputSchema(parser, { strategy: 'tool-forced' })` refuses
+ *   at run start, naming this provider. Claiming an undocumented field
+ *   works would turn a guarantee into a suggestion.
+ * • TOOL CALLING IS MODEL-DEPENDENT. `/foundry/list` reports
+ *   `supportsToolCalling` per variant, but this adapter does not
+ *   preflight-refuse on it — a wrong refusal is worse than a weak
+ *   answer, the same stance `ollama()` takes on `/api/show`. Pick a
+ *   tool-capable variant.
+ * • NO MULTI-MODAL. `LLMMessage.content` is a string. Same ceiling as
+ *   every other adapter here.
+ * • NO PROMPT CACHING — resolves to the NoOp cache strategy.
+ * • NO STRUCTURED THINKING. The wire has no thinking field; a reasoning
+ *   model's `<think>` tags ride the answer text untouched.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.FoundryLocalProvider = exports.foundryLocal = exports.FoundryLocalUnavailableError = void 0;
+/** Cap for any text the wire supplies before it becomes one of our messages. */
+const ERROR_TEXT_CAP = 200;
+// ─── Errors ─────────────────────────────────────────────────────────
+/**
+ * The two failures an on-device runtime actually has, told in words that
+ * contain the fix.
+ *
+ * Both are things the person at the keyboard can resolve in one command,
+ * which is exactly why they get a type instead of a wrapped
+ * `ECONNREFUSED` or a bare `404`. `reason` is the discriminator; the
+ * message already reads as instructions — and because Foundry Local's
+ * port changes per start, the unreachable message teaches the discovery
+ * command (`foundry server status`) alongside the start command.
+ */
+class FoundryLocalUnavailableError extends Error {
+    name = 'FoundryLocalUnavailableError';
+    /** Which of the two situations this is. */
+    reason;
+    /** The endpoint that was tried — the thing to check or change. */
+    endpoint;
+    /** The model asked for. Absent when the service never answered at all. */
+    model;
+    /** Models this machine DOES have cached, when the service could tell us. */
+    availableModels;
+    constructor(init) {
+        super(buildUnavailableMessage(init));
+        this.reason = init.reason;
+        this.endpoint = init.endpoint;
+        if (init.model !== undefined)
+            this.model = init.model;
+        if (init.availableModels !== undefined)
+            this.availableModels = init.availableModels;
+        if (init.cause !== undefined)
+            this.cause = init.cause;
+    }
+}
+exports.FoundryLocalUnavailableError = FoundryLocalUnavailableError;
+function buildUnavailableMessage(init) {
+    if (init.reason === 'service-unreachable') {
+        return (`foundryLocal: nothing is answering at ${init.endpoint}. ` +
+            'Start it with `foundry server start` (Foundry Local, install: ' +
+            'https://learn.microsoft.com/azure/foundry-local). ' +
+            'The port is dynamic — `foundry server status` prints the live URL; ' +
+            'pin one with `foundry server start --port <p>`. ' +
+            "Running elsewhere? Pass foundryLocal('<model>', { endpoint }) or set FOUNDRY_LOCAL_ENDPOINT.");
+    }
+    const model = init.model ?? '(unnamed)';
+    const have = init.availableModels && init.availableModels.length > 0
+        ? ` Models on this machine: ${init.availableModels.join(', ')}.`
+        : '';
+    // A 404 that did not speak the dialect is as likely a wrong route as a
+    // missing model. Say both rather than one confidently.
+    const route = init.routeUnconfirmed
+        ? ` That 404 named no model error, so ${init.endpoint} may not be Foundry Local's chat route at all — \`foundry server status\` prints the live URL.`
+        : '';
+    return (`foundryLocal: model '${model}' is not available on the service at ${init.endpoint}. ` +
+        `Run: foundry model run ${model} — it downloads the model if needed (aliases work too).${have}${route}`);
+}
+/**
+ * The docs' own REST example port. Only a default, never a promise —
+ * see {@link FoundryLocalProviderOptions.endpoint} for the truth about
+ * dynamic ports.
+ */
+const DEFAULT_ENDPOINT = 'http://localhost:5272';
+/** The alias the Foundry Local docs use in their own REST walkthrough. */
+const DEFAULT_MODEL = 'qwen2.5-0.5b';
+const DEFAULT_TIMEOUT_MS = 10_000;
+/** Request `model` values that mean "whatever this provider was configured with". */
+const MODEL_SHORTHANDS = new Set(['foundry-local']);
+/**
+ * Which roles this wire carries inside `messages`.
+ *
+ * The OpenAI dialect takes the system prompt as a message like any other
+ * (no separate top-level `system` field), so all three roles survive the
+ * trip.
+ */
+const CARRIES_IN_MESSAGES = Object.freeze(['system', 'user', 'assistant']);
+function foundryLocal(modelOrOptions, maybeOptions) {
+    const options = typeof modelOrOptions === 'string' ? maybeOptions ?? {} : modelOrOptions ?? {};
+    const positionalModel = typeof modelOrOptions === 'string' ? modelOrOptions : undefined;
+    const endpoint = resolveEndpoint(options);
+    const defaultModel = positionalModel ?? options.defaultModel ?? DEFAULT_MODEL;
+    const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+    const fetchImpl = options._fetch ?? ((...args) => fetch(...args));
+    const chatUrl = `${endpoint}/v1/chat/completions`;
+    const cfg = {
+        defaultModel,
+        ...(options.defaultMaxTokens !== undefined && { defaultMaxTokens: options.defaultMaxTokens }),
+    };
+    // Tool-call ids are synthesized per provider instance — the dialect says
+    // every call carries one, small local models sometimes disagree, and the
+    // whole tool round-trip in this library is keyed by id.
+    let toolCallSeq = 0;
+    const nextToolCallId = () => `foundry-call-${++toolCallSeq}`;
+    // Alias → variant-id resolutions, cached per provider instance — HIT OR
+    // MISS. The catalog does not change under a running service often enough
+    // to be worth re-asking on every call, and the cache is what keeps an
+    // alias-configured agent at one catalog fetch per process. Caching the
+    // FALLBACK too is what keeps that promise true when the catalog has no
+    // answer for the alias.
+    const resolutionCache = new Map();
+    /**
+     * REST chat calls take the FULL variant id, so an alias must be
+     * resolved first:
+     *
+     *   • A name already shaped like a variant id (it ends in the
+     *     execution-provider suffix `-cpu`/`-gpu`/`-npu`, optionally
+     *     `:version`) is used AS-IS — no catalog round-trip. That suffix is
+     *     how Foundry Local itself distinguishes a variant from an alias.
+     *   • Anything else is treated as an alias: `GET /foundry/list`, first
+     *     variant whose alias matches wins (the list's order is the
+     *     service's priority order), and the answer is cached.
+     *   • A silent or alias-less catalog resolves to the name UNCHANGED —
+     *     the chat call's own 404 then reports honestly, with the model
+     *     name the caller actually wrote. A failed lookup must never
+     *     replace the error that was coming anyway. That fallback is CACHED
+     *     exactly like a success: ONE catalog attempt per name per provider,
+     *     so an alias the service has no answer for cannot put a
+     *     `/foundry/list` round-trip — and, against a host that accepts but
+     *     never answers, a whole `timeoutMs` — in front of every single
+     *     call. A model pulled with `foundry model run` after the fact is
+     *     picked up by a FRESH provider; that is the retry.
+     */
+    const resolveModel = async (requested) => {
+        const named = MODEL_SHORTHANDS.has(requested) ? cfg.defaultModel : requested;
+        if (looksLikeVariantId(named))
+            return named;
+        const cached = resolutionCache.get(named);
+        if (cached !== undefined)
+            return cached;
+        const catalog = await fetchJsonBounded(fetchImpl, `${endpoint}/foundry/list`, timeoutMs);
+        // The fallback is cached like an answer, so the catalog is asked once
+        // per name whatever it says. A fresh provider re-asks.
+        const answer = firstVariantForAlias(catalog, named) ?? named;
+        resolutionCache.set(named, answer);
+        return answer;
+    };
+    const post = async (body, req) => {
+        // No Authorization header: no key exists on this wire, and nothing is
+        // sent where a secret could be invented, logged, or leaked.
+        const response = await fetchUntilHeaders(fetchImpl, chatUrl, {
+            method: 'POST',
+            headers: { 'content-type': 'application/json' },
+            body: JSON.stringify(body),
+        }, { timeoutMs, endpoint, ...(req.signal && { signal: req.signal }) });
+        if (!response.ok) {
+            throw await describeFailure(response, body.model, endpoint, fetchImpl, timeoutMs);
+        }
+        return response;
+    };
+    const provider = {
+        name: 'foundry-local',
+        carriesInMessages: CARRIES_IN_MESSAGES,
+        // `tool_choice` support is undocumented on this wire. Absence would mean
+        // the same thing; saying it out loud documents that this was checked
+        // rather than forgotten. See the header's ceilings.
+        carriesForcedToolChoice: false,
+        async complete(req) {
+            // A signal that fired BEFORE the call stops it here, before any socket
+            // opens — the catalog lookup included. An already-aborted signal never
+            // dispatches another 'abort' event, so a listener alone cannot see it.
+            throwIfAborted(req.signal);
+            const model = await resolveModel(req.model);
+            const body = buildBody(req, cfg, model, false);
+            const response = await post(body, req);
+            // `fetch` resolves on HEADERS, so reading the body is the stretch where
+            // a caller abort would otherwise be inert — honor it for that too.
+            const json = (await untilAborted(response.json(), req.signal));
+            return fromFoundryResponse(json, nextToolCallId);
+        },
+        async *stream(req) {
+            throwIfAborted(req.signal);
+            const model = await resolveModel(req.model);
+            const body = buildBody(req, cfg, model, true);
+            const response = await post(body, req);
+            if (!response.body)
+                throw new Error('[foundry-local] response has no body');
+            const textParts = [];
+            // The dialect streams tool calls as DELTAS — id and name on the first
+            // fragment, argument JSON split across the rest — assembled by index.
+            const toolCallsByIndex = new Map();
+            let finishReason = null;
+            let usage;
+            let lastId = '';
+            let tokenIndex = 0;
+            // The signal rides INTO the parser: it is what stops a generation the
+            // caller no longer wants, and what cancels the body when it does.
+            for await (const parsed of parseSse(response.body, req.signal)) {
+                const frame = parsed.data;
+                // An already-200 stream can still FAIL mid-generation, and this
+                // dialect says so IN BAND. Such a frame carries no `choices`, so the
+                // guard below would drop it and the terminal chunk would report a
+                // truncated answer as `stopReason: 'stop'` with nobody told. Raise
+                // instead: a failed generation is a failure, never a shorter answer.
+                const failure = frameFailureText(parsed);
+                if (failure) {
+                    throw providerError(`[foundry-local] the stream failed mid-generation — ${failure}`);
+                }
+                // Usage FIRST, and OUTSIDE the choice guard. With
+                // `stream_options.include_usage` (set in buildBody) the token counts
+                // ride a FINAL chunk whose `choices` array is EMPTY — the exact bug
+                // class 9.73.0 fixed in the OpenAI adapter: a `continue` on a missing
+                // choice threw away the only usage the stream ever reports, and every
+                // streamed local call read zero tokens everywhere usage is consumed.
+                if (frame.id)
+                    lastId = frame.id;
+                if (frame.usage)
+                    usage = frame.usage;
+                const choice = frame.choices?.[0];
+                if (!choice)
+                    continue;
+                if (choice.finish_reason)
+                    finishReason = choice.finish_reason;
+                const delta = choice.delta;
+                if (!delta)
+                    continue;
+                if (delta.content) {
+                    textParts.push(delta.content);
+                    yield { tokenIndex, content: delta.content, done: false };
+                    tokenIndex++;
+                }
+                if (delta.tool_calls) {
+                    for (const tcDelta of delta.tool_calls) {
+                        const idx = tcDelta.index ?? 0;
+                        const existing = toolCallsByIndex.get(idx) ?? { id: '', name: '', argsJson: '' };
+                        if (tcDelta.id)
+                            existing.id = tcDelta.id;
+                        if (tcDelta.function?.name)
+                            existing.name = tcDelta.function.name;
+                        if (tcDelta.function?.arguments)
+                            existing.argsJson += tcDelta.function.arguments;
+                        toolCallsByIndex.set(idx, existing);
+                    }
+                }
+            }
+            // Sorted by INDEX, not by first-seen order: a wire is free to open
+            // index 1 before index 0, and a consumer reading `toolCalls[0]` as
+            // "the first tool the model asked for" would then name the wrong one.
+            const toolCalls = Array.from(toolCallsByIndex.entries())
+                .sort(([a], [b]) => a - b)
+                .map(([, tc]) => ({
+                id: tc.id.length > 0 ? tc.id : nextToolCallId(),
+                name: tc.name,
+                args: coerceArgs(tc.argsJson),
+            }));
+            const authoritative = {
+                content: textParts.join(''),
+                toolCalls,
+                usage: {
+                    input: usage?.prompt_tokens ?? 0,
+                    output: usage?.completion_tokens ?? 0,
+                },
+                stopReason: normalizeStopReason(finishReason ?? 'stop', toolCalls.length > 0),
+                ...(lastId && { providerRef: lastId }),
+            };
+            yield { tokenIndex, content: '', done: true, response: authoritative };
+        },
+    };
+    return provider;
+}
+exports.foundryLocal = foundryLocal;
+/**
+ * Class form for consumers who prefer `new FoundryLocalProvider(...)`.
+ */
+class FoundryLocalProvider {
+    name = 'foundry-local';
+    carriesInMessages = CARRIES_IN_MESSAGES;
+    carriesForcedToolChoice = false;
+    inner;
+    constructor(model, options) {
+        this.inner =
+            typeof model === 'string'
+                ? foundryLocal(model, options)
+                : foundryLocal(model);
+    }
+    // `hooks` is FORWARDED, not dropped — see LLMCallHooks in adapters/types.ts.
+    complete(req, hooks) {
+        return this.inner.complete(req, hooks);
+    }
+    stream(req, hooks) {
+        if (!this.inner.stream)
+            throw new Error('stream() unavailable');
+        return this.inner.stream(req, hooks);
+    }
+}
+exports.FoundryLocalProvider = FoundryLocalProvider;
+// ─── Internals ──────────────────────────────────────────────────────
+/**
+ * Resolve where the service lives. Most specific wins: an explicit
+ * `endpoint`, then a duck-typed SDK manager's discovered URL, then the
+ * two env spellings, then the docs' example port.
+ *
+ * A `/v1` suffix is trimmed rather than rejected — someone copying the
+ * chat URL out of `foundry server status` output means the same
+ * machine — and a bare `host:port` gets `http://`.
+ */
+function resolveEndpoint(options) {
+    const env = typeof process !== 'undefined' ? process.env : undefined;
+    const raw = firstConfigured(options.endpoint, options.manager?.urls?.[0], env?.FOUNDRY_LOCAL_ENDPOINT, env?.FOUNDRY_LOCAL_BASE_URL) ?? DEFAULT_ENDPOINT;
+    const withScheme = /^https?:\/\//i.test(raw) ? raw : `http://${raw}`;
+    return withScheme.replace(/\/+$/, '').replace(/\/v1$/i, '');
+}
+/**
+ * First candidate that actually says something.
+ *
+ * A blank value is treated as ABSENT rather than as an endpoint. `??` would
+ * accept `''`, and `'' → 'http://' → strip trailing slashes` leaves the URL
+ * `http:` — a refusal naming nothing the reader can check or fix. An env key
+ * present with an empty value is a config that forgot to fill it in, so the
+ * next candidate (and ultimately the default) gets its turn.
+ */
+function firstConfigured(...candidates) {
+    for (const candidate of candidates) {
+        if (typeof candidate === 'string' && candidate.trim().length > 0)
+            return candidate.trim();
+    }
+    return undefined;
+}
+/**
+ * Does this name already carry a variant's shape?
+ *
+ * Full Foundry Local variant ids name their execution provider as the
+ * terminal segment — `-cpu`, `-gpu` or `-npu`, optionally followed by a
+ * `:version` — e.g. `qwen2.5-0.5b-instruct-generic-cpu:1`. Aliases never
+ * do. This structural fact is what lets a fully-qualified id skip the
+ * catalog round-trip entirely.
+ */
+function looksLikeVariantId(model) {
+    const bare = model.replace(/:\d+$/, '');
+    return /-(cpu|gpu|npu)$/i.test(bare);
+}
+/**
+ * First catalog variant whose alias matches — `/foundry/list` order is
+ * the service's priority order, so first wins. Tolerates both a bare
+ * array and a `{ models: [...] }` wrapper, and both id spellings.
+ */
+function firstVariantForAlias(catalog, alias) {
+    const entries = Array.isArray(catalog)
+        ? catalog
+        : catalog?.models;
+    if (!Array.isArray(entries))
+        return undefined;
+    for (const raw of entries) {
+        if (typeof raw !== 'object' || raw === null)
+            continue;
+        const entry = raw;
+        if (entry.alias !== alias)
+            continue;
+        const id = entry.name ?? entry.id;
+        if (typeof id === 'string' && id.length > 0)
+            return id;
+    }
+    return undefined;
+}
+function buildBody(req, cfg, model, stream) {
+    const body = {
+        model,
+        messages: toFoundryMessages(req.messages, req.systemPrompt),
+        stream,
+    };
+    // Always ask for usage on a stream. This wire is a documented Foundry
+    // Local surface, not an arbitrary OpenAI-compatible server, so the
+    // reject-unknown-field caution that cost `openai({ baseURL })` its token
+    // counts (fixed in 9.73.0) has no purchase here — and a streamed call
+    // that reports zero tokens silently disarms `.compaction()` and budgets.
+    if (stream)
+        body.stream_options = { include_usage: true };
+    if (req.tools && req.tools.length > 0)
+        body.tools = req.tools.map(toFoundryTool);
+    const maxTokens = req.maxTokens ?? cfg.defaultMaxTokens;
+    if (maxTokens !== undefined)
+        body.max_tokens = maxTokens;
+    if (req.temperature !== undefined)
+        body.temperature = req.temperature;
+    if (req.stop && req.stop.length > 0)
+        body.stop = [...req.stop];
+    // `req.toolChoice` is intentionally NOT translated: support for it is
+    // undocumented on this wire, `carriesForcedToolChoice` says so, and the
+    // agent refuses before it ever reaches here.
+    return body;
+}
+/**
+ * messages → wire messages.
+ *
+ * Roles map 1:1. The system prompt is prepended as an ordinary `system`
+ * message (this dialect has no separate system field). Assistant turns
+ * carry their `tool_calls` back — arguments re-serialized to the JSON
+ * STRING the dialect expects — and tool results carry `tool_call_id`,
+ * which is how this wire correlates a result with the call that asked
+ * for it. A role the port does not define is dropped, not forwarded
+ * blindly.
+ */
+function toFoundryMessages(messages, systemPrompt) {
+    const result = [];
+    if (systemPrompt)
+        result.push({ role: 'system', content: systemPrompt });
+    for (const m of messages) {
+        if (m.role === 'system' || m.role === 'user') {
+            result.push({ role: m.role, content: m.content });
+            continue;
+        }
+        if (m.role === 'assistant') {
+            const msg = { role: 'assistant', content: m.content };
+            if (m.toolCalls && m.toolCalls.length > 0) {
+                msg.tool_calls = m.toolCalls.map((tc) => ({
+                    id: tc.id,
+                    type: 'function',
+                    function: { name: tc.name, arguments: JSON.stringify(tc.args) },
+                }));
+            }
+            result.push(msg);
+            continue;
+        }
+        if (m.role === 'tool') {
+            result.push({
+                role: 'tool',
+                content: m.content,
+                ...(m.toolCallId && { tool_call_id: m.toolCallId }),
+            });
+            continue;
+        }
+    }
+    return result;
+}
+function toFoundryTool(schema) {
+    return {
+        type: 'function',
+        function: {
+            name: schema.name,
+            description: schema.description,
+            parameters: { ...schema.inputSchema },
+        },
+    };
+}
+/**
+ * Wire response → the port's shape.
+ *
+ * @throws a `FoundryLocalProviderError` when the 200 body is really a
+ *   failure (`{"error": ...}`). Reading it as a response would yield empty
+ *   content with `stopReason: 'stop'` and zero tokens — a failed call
+ *   dressed as a successful one.
+ */
+function fromFoundryResponse(response, nextToolCallId) {
+    if (response.error !== undefined && response.error !== null) {
+        const detail = extractErrorPayload(response.error) || 'the service named no reason';
+        throw providerError(`[foundry-local] the service answered 200 with an error — ${detail}`);
+    }
+    const choice = response.choices?.[0];
+    const message = choice?.message;
+    const wireToolCalls = message?.tool_calls ?? [];
+    const toolCalls = wireToolCalls.map((tc) => toLLMToolCall(tc, nextToolCallId));
+    return {
+        content: message?.content ?? '',
+        toolCalls,
+        usage: {
+            input: response.usage?.prompt_tokens ?? 0,
+            output: response.usage?.completion_tokens ?? 0,
+        },
+        stopReason: normalizeStopReason(choice?.finish_reason ?? 'stop', toolCalls.length > 0),
+        ...(response.id && { providerRef: response.id }),
+    };
+}
+/**
+ * One wire tool call → the port's shape.
+ *
+ * Two wire facts handled here:
+ *   • `id` should always be present on this dialect, but small local
+ *     models have been seen to omit it — and the whole tool round-trip
+ *     in this library is keyed by id. So an id is SYNTHESIZED when
+ *     missing, unique per provider instance.
+ *   • `arguments` is a JSON STRING per the dialect (Ollama sends an
+ *     object); the object form is still tolerated in case a proxy in the
+ *     middle reshaped it.
+ */
+function toLLMToolCall(tc, nextToolCallId) {
+    return {
+        id: tc.id && tc.id.length > 0 ? tc.id : nextToolCallId(),
+        name: tc.function?.name ?? '',
+        args: coerceArgs(tc.function?.arguments),
+    };
+}
+function coerceArgs(args) {
+    if (args === undefined || args === null)
+        return {};
+    if (typeof args === 'string') {
+        if (args.length === 0)
+            return {};
+        try {
+            const parsed = JSON.parse(args);
+            return typeof parsed === 'object' && parsed !== null
+                ? parsed
+                : {};
+        }
+        catch {
+            // Malformed args are rare but observed on small local models. Surface
+            // empty rather than crash — the tool-call event still fires, so the
+            // problem is visible in the trace.
+            return {};
+        }
+    }
+    return { ...args };
+}
+/**
+ * `finish_reason` → the port's stop vocabulary.
+ *
+ * `tool_calls` is the dialect's own word for a tool-ending turn, but a
+ * local model has been seen to report plain `stop` with tool calls
+ * attached — so the presence of tool calls also decides, the same
+ * correction the Ollama adapter makes.
+ */
+function normalizeStopReason(raw, hasToolCalls) {
+    if (hasToolCalls && (raw === 'stop' || raw === ''))
+        return 'tool_use';
+    switch (raw) {
+        case 'stop':
+            return 'stop';
+        case 'tool_calls':
+            return 'tool_use';
+        case 'length':
+            return 'max_tokens';
+        case 'content_filter':
+            return 'content_filter';
+        default:
+            return raw;
+    }
+}
+/**
+ * Parse an SSE body — `data: {...}` lines, one JSON payload each, each
+ * carrying its frame's `event:` name when it had one.
+ *
+ * The same discipline as the Ollama adapter's NDJSON parser, adapted to
+ * SSE: a cross-read buffer so a frame split anywhere (even mid-byte
+ * sequence — TextDecoder streams) reassembles; a malformed line is
+ * SKIPPED, never fatal; comments and every other field are ignored; and
+ * the `[DONE]` sentinel ends the stream, so anything a broken server
+ * writes after it never reaches a consumer.
+ *
+ * Two things here that a plain SSE reader would not do, both because the
+ * other end is a model on THIS machine: the `event:` name survives (it is
+ * one of the two ways this dialect spells a mid-generation failure), and
+ * the caller's `signal` both interrupts the read and CANCELS the body —
+ * a generation nobody is reading still occupies the GPU.
+ */
+async function* parseSse(body, signal) {
+    const reader = body.getReader();
+    const decoder = new TextDecoder();
+    let buf = '';
+    let eventName;
+    try {
+        for (;;) {
+            // The read is RACED against the caller's signal. `fetch` resolved on
+            // headers, so this loop is the whole rest of the call — a signal that
+            // only reached the headers would be a cancellation that cancels nothing.
+            const { value, done } = await untilAborted(reader.read(), signal);
+            if (done)
+                break;
+            buf += decoder.decode(value, { stream: true });
+            let idx;
+            while ((idx = buf.indexOf('\n')) >= 0) {
+                const line = buf.slice(0, idx).trim(); // trim eats the \r of a \r\n wire
+                buf = buf.slice(idx + 1);
+                if (line.length === 0) {
+                    eventName = undefined; // the blank line ends a frame
+                    continue;
+                }
+                if (line.startsWith('event:')) {
+                    eventName = line.slice('event:'.length).trim();
+                    continue;
+                }
+                if (!line.startsWith('data:'))
+                    continue;
+                const payload = line.slice('data:'.length).trim();
+                if (payload === '[DONE]')
+                    return;
+                let data;
+                try {
+                    data = JSON.parse(payload);
+                }
+                catch {
+                    /* skip malformed line */
+                    eventName = undefined;
+                    continue;
+                }
+                yield { ...(eventName !== undefined && { event: eventName }), data };
+                eventName = undefined;
+            }
+        }
+        // A final data line with no trailing newline still counts.
+        const tail = buf.trim();
+        if (tail.startsWith('data:')) {
+            const payload = tail.slice('data:'.length).trim();
+            if (payload !== '[DONE]') {
+                let data;
+                try {
+                    data = JSON.parse(payload);
+                }
+                catch {
+                    return; // skip a malformed tail — there is nothing after it anyway
+                }
+                yield { ...(eventName !== undefined && { event: eventName }), data };
+            }
+        }
+    }
+    finally {
+        // CANCEL, not merely release the lock. `[DONE]`, a consumer that breaks
+        // out of the loop, an error frame, a caller abort — every one of them
+        // leaves an open body, and an open body means the on-device model keeps
+        // generating for nobody and the socket stays up.
+        try {
+            await reader.cancel();
+        }
+        catch {
+            /* already closed or errored — nothing left to close */
+        }
+        try {
+            reader.releaseLock();
+        }
+        catch {
+            /* a lock the runtime already dropped */
+        }
+    }
+}
+/**
+ * POST, but never hang waiting for a service that is not there.
+ *
+ * The timer bounds the wait for RESPONSE HEADERS and is cleared the moment
+ * they arrive, so a model that takes four minutes to write a long answer is
+ * unaffected — `fetch` resolves on headers, and the body streams afterwards.
+ *
+ * The deadline is a RACE, not just an `AbortSignal`. Aborting is the polite
+ * request — it releases the socket and is what a real `fetch` acts on — but a
+ * promise that never settles is exactly the failure this guards against, and
+ * "never hangs" cannot be a promise the caller's `fetch` implementation gets
+ * to break on our behalf. So the timeout wins on its own.
+ *
+ * A caller's own `AbortSignal` is forwarded and, if IT is what fired, the
+ * abort is re-thrown as an abort rather than blamed on the service —
+ * including a signal that was ALREADY aborted when the call was made, which
+ * never dispatches an event for a listener to hear.
+ */
+async function fetchUntilHeaders(fetchImpl, url, init, opts) {
+    // An already-aborted signal never dispatches another 'abort' event, so the
+    // bridge below could not fire and the request would go out on a turn the
+    // caller has cancelled. Refuse before the socket opens, and refuse with the
+    // caller's own reason: an abort is their call, never a service fault.
+    throwIfAborted(opts.signal);
+    const controller = new AbortController();
+    let timedOut = false;
+    let timer;
+    const onCallerAbort = () => controller.abort();
+    opts.signal?.addEventListener('abort', onCallerAbort, { once: true });
+    const deadline = new Promise((_resolve, reject) => {
+        timer = setTimeout(() => {
+            timedOut = true;
+            controller.abort();
+            reject(new FoundryLocalUnavailableError({
+                reason: 'service-unreachable',
+                endpoint: opts.endpoint,
+            }));
+        }, opts.timeoutMs);
+    });
+    try {
+        return await Promise.race([fetchImpl(url, { ...init, signal: controller.signal }), deadline]);
+    }
+    catch (err) {
+        if (err instanceof FoundryLocalUnavailableError)
+            throw err; // the deadline fired
+        if (opts.signal?.aborted && !timedOut)
+            throw err; // the caller's call, not ours
+        // Timed out, or connection refused / DNS failure / TLS failure — either
+        // way, nothing answered at that endpoint. The original is preserved as
+        // `cause` for anyone who wants it — just never in the message.
+        throw new FoundryLocalUnavailableError({
+            reason: 'service-unreachable',
+            endpoint: opts.endpoint,
+            cause: err,
+        });
+    }
+    finally {
+        if (timer !== undefined)
+            clearTimeout(timer);
+        opts.signal?.removeEventListener('abort', onCallerAbort);
+        // When the fetch wins the race, the deadline promise may still reject with
+        // nobody listening. Swallow it deliberately: the answer already arrived, so
+        // an unhandled-rejection warning here would be noise about a non-event.
+        deadline.catch(() => undefined);
+    }
+}
+/**
+ * Best-effort GET-and-parse, bounded the same way the main POST is.
+ *
+ * Used for the two side lookups (`/foundry/list`, `/openai/models`),
+ * where the answer improves an outcome but its absence must never worsen
+ * one. Any failure — non-2xx, bad JSON, a fetch that hangs past the
+ * deadline — resolves to `undefined`; the race (not just the abort)
+ * keeps the "never hangs" promise even against a fetch implementation
+ * that ignores its signal.
+ */
+async function fetchJsonBounded(fetchImpl, url, timeoutMs) {
+    const controller = new AbortController();
+    let timer;
+    const deadline = new Promise((resolve) => {
+        timer = setTimeout(() => {
+            controller.abort();
+            resolve(undefined);
+        }, timeoutMs);
+    });
+    const attempt = (async () => {
+        const response = await fetchImpl(url, { signal: controller.signal });
+        if (!response.ok)
+            return undefined;
+        return (await response.json());
+    })();
+    try {
+        return await Promise.race([attempt, deadline]);
+    }
+    catch {
+        // Best-effort lookup. Its failure must never replace the error the
+        // caller is on the way to reporting.
+        return undefined;
+    }
+    finally {
+        if (timer !== undefined)
+            clearTimeout(timer);
+        // If the deadline won, the losing attempt may still reject later with
+        // nobody listening — swallow that deliberately.
+        attempt.catch(() => undefined);
+    }
+}
+/**
+ * Turn a non-2xx into the most actionable error available.
+ *
+ * A 404 from chat means the model is not on this service. Before saying
+ * so we ask `/openai/models` — a cheap local call — so the message can
+ * also name what IS cached here, which is usually enough to spot a typo.
+ * If that lookup fails too, the `foundry model run` instruction stands
+ * on its own.
+ *
+ * A 404 whose body does not speak this dialect's `{"error": ...}` may not
+ * be a model 404 at all — the endpoint may point at something that is not
+ * Foundry Local's chat route. The refusal still names the model command
+ * (guessing the other way would be just as confident and just as wrong),
+ * but it also names the endpoint as a suspect instead of pretending to know.
+ */
+async function describeFailure(response, model, endpoint, fetchImpl, timeoutMs) {
+    const bodyText = await safeText(response);
+    if (response.status === 404) {
+        const availableModels = await listCachedModels(fetchImpl, endpoint, timeoutMs);
+        return new FoundryLocalUnavailableError({
+            reason: 'model-not-available',
+            endpoint,
+            model,
+            ...(availableModels && { availableModels }),
+            ...(speaksDialectError(bodyText) ? {} : { routeUnconfirmed: true }),
+        });
+    }
+    const detail = extractErrorText(bodyText);
+    const said = detail ? ` — ${detail}` : '';
+    return providerError(`[foundry-local] ${response.status} ${response.statusText}${said}`.trim(), response.status);
+}
+/** `GET /openai/models` — a bare JSON array of cached model names. */
+async function listCachedModels(fetchImpl, endpoint, timeoutMs) {
+    const json = await fetchJsonBounded(fetchImpl, `${endpoint}/openai/models`, timeoutMs);
+    if (!Array.isArray(json))
+        return undefined;
+    const names = json.filter((n) => typeof n === 'string' && n.length > 0);
+    return names.length > 0 ? names : undefined;
+}
+async function safeText(response) {
+    try {
+        return await response.text();
+    }
+    catch {
+        return '';
+    }
+}
+/**
+ * Errors are `{"error": {"message": "..."}}` per the OpenAI dialect; a
+ * bare `{"error": "..."}` string and raw text are tolerated. Every path
+ * is capped at 200 chars — the wire's words help diagnose, but a server
+ * echoing something enormous (or poisoned) must not become the message.
+ */
+function extractErrorText(bodyText) {
+    if (!bodyText)
+        return '';
+    try {
+        const parsed = JSON.parse(bodyText);
+        const detail = extractErrorPayload(parsed.error);
+        if (detail)
+            return detail;
+    }
+    catch {
+        /* not JSON — fall through */
+    }
+    return bodyText.slice(0, ERROR_TEXT_CAP);
+}
+/**
+ * Did this body speak the dialect's `{"error": ...}`?
+ *
+ * The one fact {@link describeFailure} needs to tell a model 404 from a 404
+ * that is really "this route is not ours".
+ */
+function speaksDialectError(bodyText) {
+    if (!bodyText)
+        return false;
+    try {
+        const parsed = JSON.parse(bodyText);
+        return parsed.error !== undefined && parsed.error !== null;
+    }
+    catch {
+        return false;
+    }
+}
+/**
+ * The words out of an `error` payload — `{"message": "..."}` per the
+ * dialect, a bare string tolerated — capped like every other piece of wire
+ * text this file repeats. Empty when there is nothing usable to say, so a
+ * caller can tell "no error" from "an error that named no reason".
+ */
+function extractErrorPayload(err) {
+    if (typeof err === 'string')
+        return err.slice(0, ERROR_TEXT_CAP);
+    if (typeof err === 'object' && err !== null) {
+        const message = err.message;
+        if (typeof message === 'string' && message.length > 0) {
+            return message.slice(0, ERROR_TEXT_CAP);
+        }
+    }
+    return '';
+}
+/**
+ * What a mid-stream frame says went wrong — '' when nothing did.
+ *
+ * Two spellings on this dialect: a `data:` frame carrying `error`, and an
+ * `event: error` frame whose payload names the failure at the top level.
+ * Both mean the generation failed; neither carries `choices`, which is
+ * exactly why an unchecked one looks like an ordinary skippable frame.
+ */
+function frameFailureText(frame) {
+    const data = typeof frame.data === 'object' && frame.data !== null
+        ? frame.data
+        : undefined;
+    const carriesError = data?.error !== undefined && data?.error !== null;
+    const isErrorEvent = frame.event === 'error';
+    if (!carriesError && !isErrorEvent)
+        return '';
+    const detail = (carriesError ? extractErrorPayload(data?.error) : '') ||
+        (typeof data?.message === 'string' ? data.message.slice(0, ERROR_TEXT_CAP) : '');
+    return detail || 'the service reported an error but named no reason';
+}
+/**
+ * The provider's own labelled error — one shape for every failure that is
+ * not one of the two typed, actionable ones.
+ */
+function providerError(message, status) {
+    return Object.assign(new Error(message), {
+        name: 'FoundryLocalProviderError',
+        ...(status !== undefined && { status }),
+    });
+}
+/**
+ * An abort is the CALLER's word, so it must reach them AS an abort — never
+ * dressed up as a service failure. The signal's own `reason` is used when it
+ * is an Error (what every runtime supplies: a DOMException named
+ * 'AbortError'); anything else becomes one, keeping the reason as `cause`.
+ */
+function asAbortError(reason) {
+    if (reason instanceof Error)
+        return reason;
+    return Object.assign(new Error('This operation was aborted'), {
+        name: 'AbortError',
+        ...(reason !== undefined && { cause: reason }),
+    });
+}
+/** Stop right here when the caller's signal has already fired. */
+function throwIfAborted(signal) {
+    if (signal?.aborted)
+        throw asAbortError(signal.reason);
+}
+/**
+ * `promise`, except that a caller abort ends the wait immediately.
+ *
+ * `fetch` resolves on HEADERS. Everything after that — a streamed body, a
+ * large non-streaming JSON read — used to be deaf to the caller's signal, so
+ * a cancelled turn kept the local model generating to completion. This is
+ * what makes `LLMRequest.signal` honest for the whole call rather than only
+ * until the headers land.
+ */
+function untilAborted(promise, signal) {
+    if (!signal)
+        return promise;
+    if (signal.aborted) {
+        void promise.catch(() => undefined); // the abandoned work must not warn
+        return Promise.reject(asAbortError(signal.reason));
+    }
+    return new Promise((resolve, reject) => {
+        const onAbort = () => reject(asAbortError(signal.reason));
+        signal.addEventListener('abort', onAbort, { once: true });
+        promise.then(resolve, reject).finally(() => {
+            signal.removeEventListener('abort', onAbort);
+        });
+    });
+}
+//# sourceMappingURL=FoundryLocalProvider.js.map