@tanstack-ai-browser/wllama 0.2.1 → 0.2.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tanstack-ai-browser/wllama",
-  "version": "0.2.1",
+  "version": "0.2.2",
   "description": "TanStack AI adapter for GGUF models running in the browser with wllama (llama.cpp in WebAssembly)",
   "license": "MIT",
   "author": "daveycodez",
@@ -55,11 +55,11 @@
     "@tanstack/ai": "1.4.0",
     "esbuild": "0.28.2"
   },
-  "scripts": {
-    "test": "node --import @swc-node/register/esm-register --test \"src/**/*.test.ts\""
-  },
   "dependencies": {
     "@tanstack/openai-base": "^1.0.4",
     "@wllama/wllama": "3.13.0"
+  },
+  "scripts": {
+    "test": "node --import @swc-node/register/esm-register --test \"src/**/*.test.ts\""
   }
-}
+}

package/dist/model-memory.d.ts

@@ -1,46 +0,0 @@
-/**
- * The memory the page's models take, shared by this project's runtimes
- * (Transformers.js, WebGL, wllama), each of which loads models in its own
- * worker. Before a runtime loads a model it claims the memory: other
- * runtimes' least recently used models are unloaded until everything fits
- * the budget. A phone can't hold two 1 GB models (an iPhone Safari tab has
- * 2 GB), and small models still stay loaded side by side.
- *
- * The packages each carry this file and share one registry through
- * globalThis, so they find each other even when bundled apart. Keep the
- * copies identical (packages/{transformers-js,webgl,wllama}/src).
- */
-/**
- * The default budget: 1.5 GB on iOS (a Safari tab gets 2 GB, the page and
- * Safari take the rest), half the device's memory where the browser says
- * (Chrome: up to 8 GB), else 4 GB.
- */
-export declare function defaultMemoryBudget(): number;
-/**
- * Sets how many bytes the page's models may take together, across this
- * project's runtimes, from the next load on; undefined for the default.
- */
-export declare function setMemoryBudget(bytes: number | undefined): void;
-/**
- * Makes room for `owner`'s model of about `bytes`, unloading other runtimes'
- * models, least recently used first, until it fits the budget (models kept
- * loaded stay), then records it in place of `owner`'s previous one.
- * Resolves to false, recording nothing, if `signal` aborted meanwhile.
- */
-export declare function claimMemory(model: {
-    owner: string;
-    model: string;
-    bytes: number;
-    keepLoaded: boolean;
-    unload: () => unknown;
-    /**
-     * The load the claim is for: once it's aborted (canceled or replaced),
-     * nothing more is unloaded for it and nothing is recorded.
-     */
-    signal?: AbortSignal;
-}): Promise<boolean>;
-/** Marks `owner`'s model as just used (for least recently used order). */
-export declare function touchMemory(owner: string): void;
-/** Forgets `owner`'s model, once it's unloaded or failed to load. */
-export declare function releaseMemory(owner: string): void;
-//# sourceMappingURL=model-memory.d.ts.map

package/dist/model-memory.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"model-memory.d.ts","sourceRoot":"","sources":["../src/model-memory.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AA6BH;;;;GAIG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,CAS5C;AAED;;;GAGG;AACH,wBAAgB,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,SAAS,QAExD;AAED;;;;;GAKG;AACH,wBAAsB,WAAW,CAAC,KAAK,EAAE;IACvC,KAAK,EAAE,MAAM,CAAA;IACb,KAAK,EAAE,MAAM,CAAA;IACb,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,EAAE,OAAO,CAAA;IACnB,MAAM,EAAE,MAAM,OAAO,CAAA;IACrB;;;OAGG;IACH,MAAM,CAAC,EAAE,WAAW,CAAA;CACrB,GAAG,OAAO,CAAC,OAAO,CAAC,CAkBnB;AAED,0EAA0E;AAC1E,wBAAgB,WAAW,CAAC,KAAK,EAAE,MAAM,QAGxC;AAED,qEAAqE;AACrE,wBAAgB,aAAa,CAAC,KAAK,EAAE,MAAM,QAE1C"}

package/dist/model-memory.js

@@ -1,79 +0,0 @@
-/**
- * The memory the page's models take, shared by this project's runtimes
- * (Transformers.js, WebGL, wllama), each of which loads models in its own
- * worker. Before a runtime loads a model it claims the memory: other
- * runtimes' least recently used models are unloaded until everything fits
- * the budget. A phone can't hold two 1 GB models (an iPhone Safari tab has
- * 2 GB), and small models still stay loaded side by side.
- *
- * The packages each carry this file and share one registry through
- * globalThis, so they find each other even when bundled apart. Keep the
- * copies identical (packages/{transformers-js,webgl,wllama}/src).
- */
-const KEY = Symbol.for("@tanstack-ai-browser/model-memory");
-function registry() {
-    const scope = globalThis;
-    return (scope[KEY] ??= { version: 1, held: new Map() });
-}
-/**
- * The default budget: 1.5 GB on iOS (a Safari tab gets 2 GB, the page and
- * Safari take the rest), half the device's memory where the browser says
- * (Chrome: up to 8 GB), else 4 GB.
- */
-export function defaultMemoryBudget() {
-    if (typeof navigator === "undefined")
-        return Number.POSITIVE_INFINITY;
-    const { userAgent, maxTouchPoints } = navigator;
-    const ios = /iPhone|iPad|iPod/.test(userAgent) ||
-        (/Macintosh/.test(userAgent) && maxTouchPoints > 1);
-    if (ios)
-        return 1.5e9;
-    const gigabytes = navigator.deviceMemory;
-    return gigabytes ? gigabytes * 0.5e9 : 4e9;
-}
-/**
- * Sets how many bytes the page's models may take together, across this
- * project's runtimes, from the next load on; undefined for the default.
- */
-export function setMemoryBudget(bytes) {
-    registry().budget = bytes;
-}
-/**
- * Makes room for `owner`'s model of about `bytes`, unloading other runtimes'
- * models, least recently used first, until it fits the budget (models kept
- * loaded stay), then records it in place of `owner`'s previous one.
- * Resolves to false, recording nothing, if `signal` aborted meanwhile.
- */
-export async function claimMemory(model) {
-    const { signal, ...entry } = model;
-    if (signal?.aborted)
-        return false;
-    const { held, budget } = registry();
-    const limit = budget ?? defaultMemoryBudget();
-    held.delete(entry.owner);
-    const others = [...held.values()].sort((a, b) => a.lastUsed - b.lastUsed);
-    let total = others.reduce((sum, h) => sum + h.bytes, entry.bytes);
-    for (const other of others) {
-        if (total <= limit)
-            break;
-        if (other.keepLoaded)
-            continue;
-        held.delete(other.owner);
-        total -= other.bytes;
-        await other.unload();
-        if (signal?.aborted)
-            return false;
-    }
-    held.set(entry.owner, { ...entry, lastUsed: Date.now() });
-    return true;
-}
-/** Marks `owner`'s model as just used (for least recently used order). */
-export function touchMemory(owner) {
-    const held = registry().held.get(owner);
-    if (held)
-        held.lastUsed = Date.now();
-}
-/** Forgets `owner`'s model, once it's unloaded or failed to load. */
-export function releaseMemory(owner) {
-    registry().held.delete(owner);
-}

package/dist/model-meta.d.ts

@@ -1,7 +0,0 @@
-/**
- * The GGUF models this package lists, with what each supports and its
- * reasoning levels.
- */
-export { WLLAMA_CHAT_MODELS, WLLAMA_MODEL_META, WLLAMA_MODEL_REASONING } from "./catalog.js";
-export type { WllamaChatModel, WllamaModelMeta, WllamaModelReasoningByName } from "./catalog.js";
-//# sourceMappingURL=model-meta.d.ts.map

package/dist/model-meta.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"model-meta.d.ts","sourceRoot":"","sources":["../src/model-meta.ts"],"names":[],"mappings":"AAAA;;;GAGG;AACH,OAAO,EACL,kBAAkB,EAClB,iBAAiB,EACjB,sBAAsB,EACvB,MAAM,cAAc,CAAA;AACrB,YAAY,EACV,eAAe,EACf,eAAe,EACf,0BAA0B,EAC3B,MAAM,cAAc,CAAA"}

package/dist/model-meta.js

@@ -1,5 +0,0 @@
-/**
- * The GGUF models this package lists, with what each supports and its
- * reasoning levels.
- */
-export { WLLAMA_CHAT_MODELS, WLLAMA_MODEL_META, WLLAMA_MODEL_REASONING } from "./catalog.js";

package/dist/shared-load.d.ts

@@ -1,48 +0,0 @@
-/**
- * The callers waiting on one shared model load: their progress callbacks
- * and how many still wait. A caller whose signal aborts stops waiting at
- * once and hears nothing more; when the last caller still waiting aborts,
- * the load is cancelled. (Kept in step with the same file in the other
- * runtime packages.)
- */
-/** A download's progress, from 0 to 1. */
-type DownloadProgressCallback = (progress: number) => void;
-/** What an aborted `create()` rejects with. */
-export declare function abortError(): DOMException;
-/** Calls a progress callback; one that throws doesn't stop the load. */
-export declare function callProgress(listener: DownloadProgressCallback, progress: number): void;
-export type WaitOptions = {
-    onDownloadProgress?: DownloadProgressCallback;
-    signal?: AbortSignal;
-    /**
-     * Whether this caller aborting last cancels the load (true by default).
-     * A chat's own wait passes false: it stops waiting, and the download goes
-     * on for the next chat.
-     */
-    cancels?: boolean;
-    /** What this caller rejects with when it aborts: an AbortError by default. */
-    reason?: (signal: AbortSignal) => unknown;
-};
-export declare class LoadWaiters {
-    private readonly cancel;
-    /** Each callback, with how many waiting callers use it: called once per report. */
-    private readonly callbacks;
-    private waiting;
-    private settled;
-    /** `cancel` stops the load: called when the last caller still waiting aborts. */
-    constructor(cancel: () => void);
-    /** Reports progress to the callers still waiting. */
-    report(progress: number): void;
-    /** The load settled (ready, failed or replaced): nothing more to report or cancel. */
-    settle(): void;
-    /**
-     * `load`'s result for one more caller. Aborting `signal` rejects this
-     * caller at once and stops its reports; if it was the last caller waiting
-     * (and `cancels`), the load is cancelled. The promise is fine to leave
-     * unawaited. Check the signal before starting the load: an aborted one
-     * rejects here without waiting.
-     */
-    wait<T>(load: Promise<T>, options?: WaitOptions): Promise<T>;
-}
-export {};
-//# sourceMappingURL=shared-load.d.ts.map

package/dist/shared-load.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"shared-load.d.ts","sourceRoot":"","sources":["../src/shared-load.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,0CAA0C;AAC1C,KAAK,wBAAwB,GAAG,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,CAAA;AAE1D,+CAA+C;AAC/C,wBAAgB,UAAU,IAAI,YAAY,CAEzC;AAED,wEAAwE;AACxE,wBAAgB,YAAY,CAAC,QAAQ,EAAE,wBAAwB,EAAE,QAAQ,EAAE,MAAM,QAMhF;AAED,MAAM,MAAM,WAAW,GAAG;IACxB,kBAAkB,CAAC,EAAE,wBAAwB,CAAA;IAC7C,MAAM,CAAC,EAAE,WAAW,CAAA;IACpB;;;;OAIG;IACH,OAAO,CAAC,EAAE,OAAO,CAAA;IACjB,8EAA8E;IAC9E,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,WAAW,KAAK,OAAO,CAAA;CAC1C,CAAA;AAED,qBAAa,WAAW;IAOV,OAAO,CAAC,QAAQ,CAAC,MAAM;IANnC,mFAAmF;IACnF,OAAO,CAAC,QAAQ,CAAC,SAAS,CAA8C;IACxE,OAAO,CAAC,OAAO,CAAI;IACnB,OAAO,CAAC,OAAO,CAAQ;IAEvB,iFAAiF;IACjF,YAA6B,MAAM,EAAE,MAAM,IAAI,EAAI;IAEnD,qDAAqD;IACrD,MAAM,CAAC,QAAQ,EAAE,MAAM,QAEtB;IAED,sFAAsF;IACtF,MAAM,SAIL;IAED;;;;;;OAMG;IACH,IAAI,CAAC,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,CAAC,CAAC,EAAE,OAAO,GAAE,WAAgB,GAAG,OAAO,CAAC,CAAC,CAAC,CA6C/D;CACF"}

package/dist/shared-load.js

@@ -1,103 +0,0 @@
-/**
- * The callers waiting on one shared model load: their progress callbacks
- * and how many still wait. A caller whose signal aborts stops waiting at
- * once and hears nothing more; when the last caller still waiting aborts,
- * the load is cancelled. (Kept in step with the same file in the other
- * runtime packages.)
- */
-/** What an aborted `create()` rejects with. */
-export function abortError() {
-    return new DOMException("The model load was aborted.", "AbortError");
-}
-/** Calls a progress callback; one that throws doesn't stop the load. */
-export function callProgress(listener, progress) {
-    try {
-        listener(progress);
-    }
-    catch (error) {
-        console.error(error);
-    }
-}
-export class LoadWaiters {
-    cancel;
-    /** Each callback, with how many waiting callers use it: called once per report. */
-    callbacks = new Map();
-    waiting = 0;
-    settled = false;
-    /** `cancel` stops the load: called when the last caller still waiting aborts. */
-    constructor(cancel) {
-        this.cancel = cancel;
-    }
-    /** Reports progress to the callers still waiting. */
-    report(progress) {
-        for (const listener of [...this.callbacks.keys()])
-            callProgress(listener, progress);
-    }
-    /** The load settled (ready, failed or replaced): nothing more to report or cancel. */
-    settle() {
-        this.settled = true;
-        this.callbacks.clear();
-        this.waiting = 0;
-    }
-    /**
-     * `load`'s result for one more caller. Aborting `signal` rejects this
-     * caller at once and stops its reports; if it was the last caller waiting
-     * (and `cancels`), the load is cancelled. The promise is fine to leave
-     * unawaited. Check the signal before starting the load: an aborted one
-     * rejects here without waiting.
-     */
-    wait(load, options = {}) {
-        const { onDownloadProgress, signal, cancels = true, reason = abortError } = options;
-        if (signal?.aborted)
-            return quiet(Promise.reject(reason(signal)));
-        if (this.settled)
-            return quiet(load);
-        this.waiting++;
-        if (onDownloadProgress) {
-            this.callbacks.set(onDownloadProgress, (this.callbacks.get(onDownloadProgress) ?? 0) + 1);
-        }
-        let left = false;
-        const leave = () => {
-            if (left)
-                return;
-            left = true;
-            if (this.settled)
-                return;
-            this.waiting--;
-            if (!onDownloadProgress)
-                return;
-            const count = (this.callbacks.get(onDownloadProgress) ?? 0) - 1;
-            if (count > 0)
-                this.callbacks.set(onDownloadProgress, count);
-            else
-                this.callbacks.delete(onDownloadProgress);
-        };
-        if (!signal) {
-            load.then(leave, leave);
-            return quiet(load);
-        }
-        return quiet(new Promise((resolve, reject) => {
-            const onAbort = () => {
-                leave();
-                reject(reason(signal));
-                if (cancels && !this.settled && this.waiting === 0)
-                    this.cancel();
-            };
-            signal.addEventListener("abort", onAbort, { once: true });
-            load.then((value) => {
-                signal.removeEventListener("abort", onAbort);
-                leave();
-                resolve(value);
-            }, (error) => {
-                signal.removeEventListener("abort", onAbort);
-                leave();
-                reject(error);
-            });
-        }));
-    }
-}
-/** `promise`, with a rejection that's fine to leave unhandled. */
-function quiet(promise) {
-    promise.catch(() => undefined);
-    return promise;
-}

package/dist/store.d.ts

@@ -1,8 +0,0 @@
-/** A value that can be read, set and watched. Framework-free. */
-export type Store<T> = {
-    get: () => T;
-    set: (value: T) => void;
-    subscribe: (listener: () => void) => () => void;
-};
-export declare function createStore<T>(initial: T): Store<T>;
-//# sourceMappingURL=store.d.ts.map

package/dist/store.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"store.d.ts","sourceRoot":"","sources":["../src/store.ts"],"names":[],"mappings":"AAAA,iEAAiE;AACjE,MAAM,MAAM,KAAK,CAAC,CAAC,IAAI;IACrB,GAAG,EAAE,MAAM,CAAC,CAAA;IACZ,GAAG,EAAE,CAAC,KAAK,EAAE,CAAC,KAAK,IAAI,CAAA;IACvB,SAAS,EAAE,CAAC,QAAQ,EAAE,MAAM,IAAI,KAAK,MAAM,IAAI,CAAA;CAChD,CAAA;AAED,wBAAgB,WAAW,CAAC,CAAC,EAAE,OAAO,EAAE,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,CAgBnD"}

package/dist/store.js

@@ -1,17 +0,0 @@
-export function createStore(initial) {
-    let value = initial;
-    const listeners = new Set();
-    return {
-        get: () => value,
-        set: (next) => {
-            value = next;
-            listeners.forEach((listener) => listener());
-        },
-        subscribe: (listener) => {
-            listeners.add(listener);
-            return () => {
-                listeners.delete(listener);
-            };
-        }
-    };
-}

package/dist/template.d.ts

@@ -1,26 +0,0 @@
-/** A chat turn as LFM2.5's chat template takes it. */
-export type ChatTurn = {
-    role: "system" | "user" | "assistant" | "tool";
-    /**
-     * The text. An assistant turn's is the reply as the model wrote it:
-     * thinking (`<think>…</think>`) and tool-call markup included. The
-     * template drops the thinking of all but the last assistant turn.
-     */
-    content: string;
-};
-/** A tool as chat templates take it (OpenAI's function format). */
-export type ToolSpec = {
-    type: "function";
-    function: {
-        name: string;
-        description: string;
-        parameters: unknown;
-    };
-};
-/**
- * The prompt for `chat`, ending where the model's reply starts (or where
- * the chat ends, without `reply`). The template's BOS token is left out:
- * llama.cpp adds it when it reads the prompt.
- */
-export declare function renderChat(source: string, chat: Array<ChatTurn>, tools: Array<ToolSpec>, reply?: boolean): string;
-//# sourceMappingURL=template.d.ts.map

package/dist/template.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"template.d.ts","sourceRoot":"","sources":["../src/template.ts"],"names":[],"mappings":"AAMA,sDAAsD;AACtD,MAAM,MAAM,QAAQ,GAAG;IACrB,IAAI,EAAE,QAAQ,GAAG,MAAM,GAAG,WAAW,GAAG,MAAM,CAAA;IAC9C;;;;OAIG;IACH,OAAO,EAAE,MAAM,CAAA;CAChB,CAAA;AAED,mEAAmE;AACnE,MAAM,MAAM,QAAQ,GAAG;IACrB,IAAI,EAAE,UAAU,CAAA;IAChB,QAAQ,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,OAAO,CAAA;KAAE,CAAA;CACrE,CAAA;AAID;;;;GAIG;AACH,wBAAgB,UAAU,CACxB,MAAM,EAAE,MAAM,EACd,IAAI,EAAE,KAAK,CAAC,QAAQ,CAAC,EACrB,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,EACtB,KAAK,UAAO,UAcb"}

package/dist/template.js

@@ -1,25 +0,0 @@
-/**
- * The prompt for a chat, written by the model's own chat template (the
- * Jinja template in its GGUF file), the way Transformers renders it.
- */
-import { Template } from "@huggingface/jinja";
-const templates = new Map();
-/**
- * The prompt for `chat`, ending where the model's reply starts (or where
- * the chat ends, without `reply`). The template's BOS token is left out:
- * llama.cpp adds it when it reads the prompt.
- */
-export function renderChat(source, chat, tools, reply = true) {
-    let template = templates.get(source);
-    if (!template) {
-        template = new Template(source);
-        templates.set(source, template);
-    }
-    return template.render({
-        messages: chat,
-        ...(tools.length > 0 && { tools }),
-        add_generation_prompt: reply,
-        bos_token: "",
-        eos_token: ""
-    });
-}

package/dist/thinking.d.ts

@@ -1,39 +0,0 @@
-/**
- * Splitting a reply into thinking and answer. Copied from
- * @tanstack-ai-browser/transformers-js (thinking.ts), so this package
- * doesn't depend on that one and its worker.
- */
-/** Where a model's thinking starts and ends in its reply. */
-export type ThinkingMarkers = {
-    open: string;
-    close: string;
-};
-/** `<think>…</think>`, as LFM models (and most others) write it. */
-export declare const THINK_TAGS: ThinkingMarkers;
-/**
- * Splits a model's reply into its thinking and its answer. Thinking is any
- * `<think>…</think>` block (or the format's own markers), or everything
- * before a lone close (models prompted with an empty `<think></think>`
- * sometimes still think a little and close it). An open with no close yet
- * means it's still thinking.
- */
-export declare function splitThinking(text: string, { open: OPEN, close: CLOSE }?: ThinkingMarkers): {
-    thinking: string;
-    answer: string;
-    isThinking: boolean;
-};
-export type ReplyPiece = {
-    kind: "thinking" | "answer";
-    text: string;
-};
-/**
- * Splits a reply into thinking and answer while it streams. A reply that
- * starts with the open marker (`<think>`) is thinking until the close
- * (`</think>`); the rest is the answer. Markers split across tokens are held
- * back until they are complete.
- */
-export declare function createThinkingSplitter({ open: OPEN, close: CLOSE }?: ThinkingMarkers): {
-    push: (text: string) => Array<ReplyPiece>;
-    flush: () => Array<ReplyPiece>;
-};
-//# sourceMappingURL=thinking.d.ts.map

package/dist/thinking.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"thinking.d.ts","sourceRoot":"","sources":["../src/thinking.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,6DAA6D;AAC7D,MAAM,MAAM,eAAe,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAAA;AAE7D,oEAAoE;AACpE,eAAO,MAAM,UAAU,EAAE,eAGxB,CAAA;AAED;;;;;;GAMG;AACH,wBAAgB,aAAa,CAC3B,IAAI,EAAE,MAAM,EACZ,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,KAAK,EAAE,GAAE,eAA4B;;;;EAmC3D;AAUD,MAAM,MAAM,UAAU,GAAG;IAAE,IAAI,EAAE,UAAU,GAAG,QAAQ,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAA;AAEtE;;;;;GAKG;AACH,wBAAgB,sBAAsB,CAAC,EACrC,IAAI,EAAE,IAAI,EACV,KAAK,EAAE,KAAK,EACb,GAAE,eAA4B;iBAeT,MAAM,KAAG,KAAK,CAAC,UAAU,CAAC;iBA6C5B,KAAK,CAAC,UAAU,CAAC;EAUpC"}

package/dist/thinking.js

@@ -1,140 +0,0 @@
-/**
- * Splitting a reply into thinking and answer. Copied from
- * @tanstack-ai-browser/transformers-js (thinking.ts), so this package
- * doesn't depend on that one and its worker.
- */
-/** `<think>…</think>`, as LFM models (and most others) write it. */
-export const THINK_TAGS = {
-    open: "<think>",
-    close: "</think>"
-};
-/**
- * Splits a model's reply into its thinking and its answer. Thinking is any
- * `<think>…</think>` block (or the format's own markers), or everything
- * before a lone close (models prompted with an empty `<think></think>`
- * sometimes still think a little and close it). An open with no close yet
- * means it's still thinking.
- */
-export function splitThinking(text, { open: OPEN, close: CLOSE } = THINK_TAGS) {
-    const thoughts = [];
-    let rest = text;
-    let isThinking = false;
-    for (;;) {
-        const open = rest.indexOf(OPEN);
-        const close = rest.indexOf(CLOSE);
-        if (close !== -1 && (open === -1 || close < open)) {
-            // A close with no open: everything before it was thinking.
-            thoughts.push(rest.slice(0, close));
-            rest = rest.slice(close + CLOSE.length);
-            continue;
-        }
-        if (open === -1)
-            break;
-        const before = rest.slice(0, open);
-        const after = rest.slice(open + OPEN.length);
-        const end = after.indexOf(CLOSE);
-        if (end === -1) {
-            thoughts.push(after);
-            rest = before;
-            isThinking = true;
-            break;
-        }
-        thoughts.push(after.slice(0, end));
-        rest = before + after.slice(end + CLOSE.length);
-    }
-    return {
-        thinking: thoughts
-            .map((t) => t.trim())
-            .filter(Boolean)
-            .join("\n\n"),
-        answer: rest.trim(),
-        isThinking
-    };
-}
-/** Length of the longest end of `text` that could be the start of `tag`. */
-function partialTag(text, tag) {
-    for (let n = Math.min(tag.length - 1, text.length); n > 0; n--) {
-        if (tag.startsWith(text.slice(-n)))
-            return n;
-    }
-    return 0;
-}
-/**
- * Splits a reply into thinking and answer while it streams. A reply that
- * starts with the open marker (`<think>`) is thinking until the close
- * (`</think>`); the rest is the answer. Markers split across tokens are held
- * back until they are complete.
- */
-export function createThinkingSplitter({ open: OPEN, close: CLOSE } = THINK_TAGS) {
-    let state = "start";
-    let buffer = "";
-    /** Whether the thinking has had any text yet (to drop a leading newline). */
-    let thought = false;
-    /** Whether the answer has had any text yet (to drop leading blank lines). */
-    let answered = false;
-    const answer = (text) => {
-        const out = answered ? text : text.trimStart();
-        if (!out)
-            return [];
-        answered = true;
-        return [{ kind: "answer", text: out }];
-    };
-    function push(text) {
-        buffer += text;
-        const pieces = [];
-        for (;;) {
-            if (state === "start") {
-                const lead = buffer.trimStart();
-                if (lead.startsWith(OPEN)) {
-                    buffer = lead.slice(OPEN.length);
-                    state = "thinking";
-                }
-                else if (OPEN.startsWith(lead)) {
-                    return pieces;
-                }
-                else {
-                    state = "answer";
-                }
-            }
-            else if (state === "thinking") {
-                const end = buffer.indexOf(CLOSE);
-                if (end === -1) {
-                    const keep = partialTag(buffer, CLOSE);
-                    const done = buffer.slice(0, buffer.length - keep);
-                    buffer = buffer.slice(done.length);
-                    const text = thought ? done : done.trimStart();
-                    if (text) {
-                        thought = true;
-                        pieces.push({ kind: "thinking", text });
-                    }
-                    return pieces;
-                }
-                const text = thought
-                    ? buffer.slice(0, end)
-                    : buffer.slice(0, end).trimStart();
-                if (text) {
-                    thought = true;
-                    pieces.push({ kind: "thinking", text });
-                }
-                buffer = buffer.slice(end + CLOSE.length);
-                state = "answer";
-            }
-            else {
-                pieces.push(...answer(buffer));
-                buffer = "";
-                return pieces;
-            }
-        }
-    }
-    /** Whatever is still held back, once the reply has ended. */
-    function flush() {
-        const rest = buffer;
-        buffer = "";
-        if (!rest)
-            return [];
-        return state === "thinking"
-            ? [{ kind: "thinking", text: rest }]
-            : answer(rest);
-    }
-    return { push, flush };
-}

package/dist/tool-calls.d.ts

@@ -1,33 +0,0 @@
-/**
- * Reading and writing tool calls the way LFM2 and LFM2.5 write them:
- * Pythonic calls between markers,
- * `<|tool_call_start|>[get_weather(city="Tokyo")]<|tool_call_end|>`.
- *
- * Copied from @tanstack-ai-browser/transformers-js (tool-calls.ts, its LFM2
- * parts), so this package doesn't depend on that one and its worker.
- */
-export type ParsedToolCall = {
-    name: string;
-    args: Record<string, unknown>;
-};
-/** How a model writes tool calls. */
-export type ToolFormat = "lfm2";
-export declare const TOOL_CALL_START = "<|tool_call_start|>";
-export declare const TOOL_CALL_END = "<|tool_call_end|>";
-/** Every tool call in `text`. */
-export declare function parseToolCalls(text: string): Array<ParsedToolCall>;
-/**
- * Splits a streaming answer into the text before the first tool call (passed
- * through as it arrives) and the tool-call markup after it (held back and
- * parsed once the reply ends).
- */
-export declare function createToolCallSplitter(markers?: Array<string>): {
-    push: (text: string) => string;
-    flush: () => {
-        text: string;
-        markup: string | null;
-    };
-};
-/** A past tool call in LFM2.5's own format, for the assistant turn's text. */
-export declare function formatLfm2ToolCalls(calls: Array<ParsedToolCall>): string;
-//# sourceMappingURL=tool-calls.d.ts.map

package/dist/tool-calls.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"tool-calls.d.ts","sourceRoot":"","sources":["../src/tool-calls.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,MAAM,cAAc,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAAE,CAAA;AAE5E,qCAAqC;AACrC,MAAM,MAAM,UAAU,GAAG,MAAM,CAAA;AAE/B,eAAO,MAAM,eAAe,wBAAwB,CAAA;AACpD,eAAO,MAAM,aAAa,sBAAsB,CAAA;AA+GhD,iCAAiC;AACjC,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,KAAK,CAAC,cAAc,CAAC,CAQlE;AAWD;;;;GAIG;AACH,wBAAgB,sBAAsB,CACpC,OAAO,GAAE,KAAK,CAAC,MAAM,CAAqB;iBAMtB,MAAM;;;;;EA6B3B;AAUD,8EAA8E;AAC9E,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,KAAK,CAAC,cAAc,CAAC,UAU/D"}

package/dist/tool-calls.js

@@ -1,203 +0,0 @@
-/**
- * Reading and writing tool calls the way LFM2 and LFM2.5 write them:
- * Pythonic calls between markers,
- * `<|tool_call_start|>[get_weather(city="Tokyo")]<|tool_call_end|>`.
- *
- * Copied from @tanstack-ai-browser/transformers-js (tool-calls.ts, its LFM2
- * parts), so this package doesn't depend on that one and its worker.
- */
-export const TOOL_CALL_START = "<|tool_call_start|>";
-export const TOOL_CALL_END = "<|tool_call_end|>";
-/** A Python-style or JSON value: "x", 'x', 1.5, True, None, [1, 2]… */
-function parseValue(raw) {
-    const s = raw.trim();
-    if (/^(["']).*\1$/s.test(s)) {
-        try {
-            return JSON.parse(s[0] === "'" ? `"${s.slice(1, -1)}"` : s);
-        }
-        catch {
-            return s.slice(1, -1);
-        }
-    }
-    if (s === "True" || s === "true")
-        return true;
-    if (s === "False" || s === "false")
-        return false;
-    if (s === "None" || s === "null")
-        return null;
-    if (s !== "" && !Number.isNaN(Number(s)))
-        return Number(s);
-    try {
-        return JSON.parse(s);
-    }
-    catch {
-        return s;
-    }
-}
-/** After a value: the next argument (`, name=`), or the end of the call. */
-const VALUE_END = /^\s*(?:,\s*[A-Za-z_]\w*\s*=|,?\s*\)|$)/;
-/**
- * Where the string starting at `body[start]` ends (the index of its closing
- * quote). The model doesn't escape quotes inside its strings ('JWST's
- * findings'), so a quote only closes the string when the argument ends
- * after it; otherwise it's part of the text.
- */
-function stringEnd(body, start) {
-    const quote = body[start];
-    let last = -1;
-    for (let i = start + 1; i < body.length; i++) {
-        if (body[i] === "\\")
-            i++;
-        else if (body[i] === quote) {
-            if (VALUE_END.test(body.slice(i + 1)))
-                return i;
-            last = i;
-        }
-    }
-    return last;
-}
-/** A Python string literal's text, `\'` and `\"` included. */
-function stringText(raw) {
-    const inner = raw.slice(1, -1);
-    try {
-        const escaped = inner.replace(/\\'/g, "'").replace(/(^|[^\\])"/g, '$1\\"');
-        return JSON.parse(`"${escaped}"`);
-    }
-    catch {
-        return inner;
-    }
-}
-/** Where a `[…]` or `{…}` starting at `body[start]` ends, strings skipped. */
-function bracketEnd(body, start) {
-    let depth = 0;
-    for (let i = start; i < body.length; i++) {
-        const c = body[i];
-        if (c === '"' || c === "'") {
-            const end = stringEnd(body, i);
-            if (end < 0)
-                return -1;
-            i = end;
-        }
-        else if (c === "[" || c === "{")
-            depth++;
-        else if ((c === "]" || c === "}") && --depth === 0)
-            return i;
-    }
-    return -1;
-}
-/** `[name(a="x", b=2), other()]` → calls. */
-function parsePythonic(body) {
-    const calls = [];
-    const callStart = /([A-Za-z_][\w.-]*)\(/g;
-    let m;
-    while ((m = callStart.exec(body))) {
-        const args = {};
-        let i = m.index + m[0].length;
-        for (;;) {
-            const key = /^\s*([A-Za-z_]\w*)\s*=\s*/.exec(body.slice(i));
-            if (!key)
-                break;
-            i += key[0].length;
-            const first = body[i];
-            let end;
-            if (first === '"' || first === "'") {
-                end = stringEnd(body, i);
-                if (end < 0)
-                    break;
-                args[key[1]] = stringText(body.slice(i, end + 1));
-            }
-            else if (first === "[" || first === "{") {
-                end = bracketEnd(body, i);
-                if (end < 0)
-                    break;
-                args[key[1]] = parseValue(body.slice(i, end + 1));
-            }
-            else {
-                const rest = /^[^,)]*/.exec(body.slice(i))?.[0] ?? "";
-                end = i + rest.length - 1;
-                args[key[1]] = parseValue(rest);
-            }
-            i = end + 1;
-            const comma = /^\s*,/.exec(body.slice(i));
-            if (!comma)
-                break;
-            i += comma[0].length;
-        }
-        calls.push({ name: m[1], args });
-        // The next call starts after this one's closing parenthesis.
-        const close = body.indexOf(")", i);
-        callStart.lastIndex = close < 0 ? body.length : close + 1;
-    }
-    return calls;
-}
-/** Every tool call in `text`. */
-export function parseToolCalls(text) {
-    const calls = [];
-    for (const m of text.matchAll(/<\|tool_call_start\|>([\s\S]*?)(?:<\|tool_call_end\|>|$)/g)) {
-        calls.push(...parsePythonic(m[1]));
-    }
-    return calls;
-}
-/** Length of the longest end of `text` that could start one of `markers`. */
-function partialMarker(text, markers) {
-    for (let n = Math.min(text.length, 18); n > 0; n--) {
-        const end = text.slice(-n);
-        if (markers.some((m) => m.startsWith(end)))
-            return n;
-    }
-    return 0;
-}
-/**
- * Splits a streaming answer into the text before the first tool call (passed
- * through as it arrives) and the tool-call markup after it (held back and
- * parsed once the reply ends).
- */
-export function createToolCallSplitter(markers = [TOOL_CALL_START]) {
-    let held = "";
-    let markup = null;
-    /** Text that can be shown now. */
-    function push(text) {
-        if (markup !== null) {
-            markup += text;
-            return "";
-        }
-        held += text;
-        const at = Math.min(...markers.map((m) => held.indexOf(m)).filter((i) => i >= 0));
-        if (Number.isFinite(at)) {
-            const before = held.slice(0, at);
-            markup = held.slice(at);
-            held = "";
-            return before;
-        }
-        const keep = partialMarker(held, markers);
-        const out = held.slice(0, held.length - keep);
-        held = held.slice(out.length);
-        return out;
-    }
-    /** The rest of the text, and the tool-call markup if any. */
-    function flush() {
-        const rest = held;
-        held = "";
-        return { text: rest, markup };
-    }
-    return { push, flush };
-}
-/** A value written the way Python (and LFM2.5) writes it. */
-function pythonValue(value) {
-    if (value === true)
-        return "True";
-    if (value === false)
-        return "False";
-    if (value === null || value === undefined)
-        return "None";
-    return JSON.stringify(value);
-}
-/** A past tool call in LFM2.5's own format, for the assistant turn's text. */
-export function formatLfm2ToolCalls(calls) {
-    const list = calls
-        .map((c) => `${c.name}(${Object.entries(c.args)
-        .map(([k, v]) => `${k}=${pythonValue(v)}`)
-        .join(", ")})`)
-        .join(", ");
-    return `${TOOL_CALL_START}[${list}]${TOOL_CALL_END}`;
-}