nexus-ai-pro 2.2.0 → 2.4.0

package/dist/agent/middleware/limits.js

@@ -0,0 +1,52 @@
+import { modelCallsInTurn } from './shared.js';
+export class ModelCallLimitError extends Error {
+    scope;
+    limit;
+    code = 'MODEL_CALL_LIMIT';
+    constructor(scope, limit) {
+        super(`The agent reached its limit of ${limit} model calls per ${scope}`);
+        this.scope = scope;
+        this.limit = limit;
+        this.name = 'ModelCallLimitError';
+    }
+}
+const NAMESPACE = ['nexus', 'agent', 'model-calls'];
+export function modelCallLimit(options) {
+    if (options.run === undefined && options.thread === undefined) {
+        throw new RangeError('modelCallLimit() needs a run limit, a thread limit, or both');
+    }
+    const local = new Map();
+    return {
+        name: 'model-call-limit',
+        async wrapModelCall(request, next, context) {
+            let reached;
+            if (options.run !== undefined && modelCallsInTurn(context.state.messages) >= options.run) {
+                reached = ['run', options.run];
+            }
+            let used = 0;
+            if (!reached && options.thread !== undefined) {
+                used = context.store
+                    ? ((await context.store.get(NAMESPACE, context.threadId))?.value ?? 0)
+                    : (local.get(context.threadId) ?? 0);
+                if (used >= options.thread)
+                    reached = ['thread', options.thread];
+            }
+            if (reached) {
+                const [scope, limit] = reached;
+                if (options.onLimit === 'error')
+                    throw new ModelCallLimitError(scope, limit);
+                return context.stop(options.message ?? `I stopped here: this ${scope} reached its limit of ${limit} model calls.`);
+            }
+            if (options.thread !== undefined) {
+                if (context.store)
+                    await context.store.put(NAMESPACE, context.threadId, used + 1);
+                else {
+                    local.set(context.threadId, used + 1);
+                    if (local.size > 10_000)
+                        local.delete(local.keys().next().value);
+                }
+            }
+            return next(request);
+        },
+    };
+}

package/dist/agent/middleware/models.d.ts

@@ -0,0 +1,48 @@
+import type { CompletionRequest } from '../../types/messages.js';
+import type { AgentMiddleware, AgentMiddlewareContext, AgentModelClient } from '../create-agent.js';
+/** A model to fall back to: a model name for the agent's own client, or another client. */
+export type ModelFallbackTarget = string | {
+    client: AgentModelClient;
+    model?: string;
+};
+/** One fallback about to happen. */
+export interface ModelFallbackEvent {
+    /** The model that failed. */
+    from: string;
+    /** The model tried next. */
+    to: string;
+    /** What the failed call threw. */
+    error: unknown;
+}
+/** Options for `modelFallback()`. */
+export interface ModelFallbackOptions {
+    /** Whether an error moves on to the next model. Defaults to every error but a cancellation. */
+    fallbackOn?: (error: unknown, model: string) => boolean;
+    /** Called before each fallback, for logs and metrics. */
+    onFallback?: (event: ModelFallbackEvent) => void;
+}
+/**
+ * Tries other models, in order, when the model call fails.
+ *
+ * The client's router already falls back between providers for one request; this is the fallback
+ * it cannot make: to a client of a different kind, or once the client's own routes are exhausted. A
+ * target with its own client is called through `complete()`, never streamed, and middleware after
+ * this one does not wrap it.
+ */
+export declare function modelFallback(models: readonly ModelFallbackTarget[], options?: ModelFallbackOptions): AgentMiddleware;
+/** What `dynamicModel()` decides from: the step's context and the request about to be sent. */
+export type DynamicModelContext = AgentMiddlewareContext & {
+    request: CompletionRequest;
+};
+/**
+ * Chooses the model for each call from the run so far.
+ *
+ * A cheap model for the first steps and a stronger one once the task proves hard, a model with a
+ * longer context once the transcript grows, or a tenant's own model. Returning nothing keeps the
+ * model the request already names.
+ *
+ * ```ts
+ * dynamicModel(({ state }) => (state.iterations >= 4 ? 'gpt-5' : 'gpt-5-mini'))
+ * ```
+ */
+export declare function dynamicModel(select: (context: DynamicModelContext) => string | undefined | Promise<string | undefined>): AgentMiddleware;

package/dist/agent/middleware/models.js

@@ -0,0 +1,44 @@
+import { isCancellation } from './shared.js';
+export function modelFallback(models, options = {}) {
+    if (models.length === 0)
+        throw new RangeError('modelFallback() needs at least one model to fall back to');
+    return {
+        name: 'model-fallback',
+        async wrapModelCall(request, next, context) {
+            try {
+                return await next(request);
+            }
+            catch (first) {
+                let error = first;
+                let from = request.model;
+                for (const target of models) {
+                    if (isCancellation(error, context.signal) || !(options.fallbackOn?.(error, from) ?? true))
+                        throw error;
+                    const model = typeof target === 'string' ? target : (target.model ?? request.model);
+                    options.onFallback?.({ from, to: model, error });
+                    try {
+                        return typeof target === 'string'
+                            ? await next({ ...request, model })
+                            : await target.client.complete({ ...request, model });
+                    }
+                    catch (failure) {
+                        error = failure;
+                        from = model;
+                    }
+                }
+                throw error;
+            }
+        },
+    };
+}
+export function dynamicModel(select) {
+    return {
+        name: 'dynamic-model',
+        async beforeModel(context) {
+            const model = await select(context);
+            if (!model || model === context.request.model)
+                return;
+            return { ...context.request, model };
+        },
+    };
+}

package/dist/agent/middleware/pii.d.ts

@@ -0,0 +1,67 @@
+import type { AgentMiddleware } from '../create-agent.js';
+/** The personal data and secrets `piiMiddleware()` finds on its own. */
+export type PiiKind = 'email' | 'credit-card' | 'ssn' | 'phone' | 'iban' | 'ip-address' | 'secret';
+/**
+ * What happens to a match: `redact` replaces it with its kind, `mask` keeps its last four
+ * characters, `hash` replaces it with a stable token so two values stay distinguishable, and `block`
+ * refuses the whole message.
+ */
+export type PiiStrategy = 'redact' | 'mask' | 'hash' | 'block';
+/** Where a match was found. */
+export type PiiWhere = 'input' | 'output' | 'tool';
+/** One match, as `onDetect` sees it: never the value itself. */
+export interface PiiFinding {
+    /** Its kind: a built-in kind or a name from `patterns`. */
+    kind: string;
+    /** Where it was found. */
+    where: PiiWhere;
+    /** What was done with it. */
+    strategy: PiiStrategy;
+}
+/** Options for `piiMiddleware()`. */
+export interface PiiMiddlewareOptions {
+    /** The built-in kinds to find. Defaults to all of them. */
+    kinds?: readonly PiiKind[];
+    /** Kinds of your own, by name: `{ 'employee-id': /\bEMP-\d{6}\b/ }`. */
+    patterns?: Readonly<Record<string, RegExp>>;
+    /**
+     * What happens to a match, for every kind or per kind: `'mask'`, or
+     * `{ email: 'hash', 'credit-card': 'block' }`. Kinds not named are redacted.
+     */
+    strategy?: PiiStrategy | Readonly<Record<string, PiiStrategy>>;
+    /**
+     * Where to look. `input` is every message sent to the model but the system prompt; `output` is
+     * the model's answer; `toolResults` cleans a result before it enters the transcript, so it is never
+     * checkpointed either. Defaults to input and tool results.
+     */
+    apply?: {
+        input?: boolean;
+        output?: boolean;
+        toolResults?: boolean;
+    };
+    /** Called for each match, for an audit trail. */
+    onDetect?: (finding: PiiFinding) => void;
+}
+/** Thrown when a match under the `block` strategy is found in the model's input or output. */
+export declare class PiiBlockedError extends Error {
+    /** The kind that was found. */
+    readonly kind: string;
+    /** Where it was found. */
+    readonly where: PiiWhere;
+    /** Always `PII_BLOCKED`. */
+    readonly code = "PII_BLOCKED";
+    constructor(
+    /** The kind that was found. */
+    kind: string, 
+    /** Where it was found. */
+    where: PiiWhere);
+}
+/**
+ * Keeps personal data and secrets away from the model, the transcript, or both.
+ *
+ * Finds emails, card numbers (checked with Luhn), US social security numbers, phone numbers, IBANs
+ * (checked with their checksum), IP addresses, and common API keys and private keys, plus any kinds
+ * of your own; and redacts, masks, hashes, or blocks each. The validated kinds keep false positives
+ * low: an order number is not a card number unless its checksum says so.
+ */
+export declare function piiMiddleware(options?: PiiMiddlewareOptions): AgentMiddleware;

package/dist/agent/middleware/pii.js

@@ -0,0 +1,142 @@
+import { ibanChecksum, luhn, phoneNumber } from '../../utils/checksums.js';
+import { shortHash } from './shared.js';
+export class PiiBlockedError extends Error {
+    kind;
+    where;
+    code = 'PII_BLOCKED';
+    constructor(kind, where) {
+        super(`The ${where === 'input' ? 'request' : 'response'} contains ${kind}, which is blocked`);
+        this.kind = kind;
+        this.where = where;
+        this.name = 'PiiBlockedError';
+    }
+}
+const DETECTORS = {
+    secret: {
+        pattern: /-----BEGIN [A-Z ]*PRIVATE KEY-----[\s\S]*?-----END [A-Z ]*PRIVATE KEY-----|\b(?:sk-[A-Za-z0-9_-]{20,}|AKIA[0-9A-Z]{16}|gh[pousr]_[A-Za-z0-9]{36,}|xox[abprs]-[A-Za-z0-9-]{10,}|AIza[0-9A-Za-z_-]{35})\b/g,
+    },
+    email: { pattern: /\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}\b/g },
+    'credit-card': { pattern: /\b\d(?:[ -]?\d){12,18}\b/g, valid: luhn },
+    iban: { pattern: /\b[A-Z]{2}\d{2}(?: ?[A-Z0-9]{4}){2,7}(?: ?[A-Z0-9]{1,4})?\b/g, valid: ibanChecksum },
+    ssn: { pattern: /\b(?!000|666|9\d\d)\d{3}-(?!00)\d{2}-(?!0000)\d{4}\b/g },
+    'ip-address': { pattern: /\b(?:(?:25[0-5]|2[0-4]\d|1?\d?\d)\.){3}(?:25[0-5]|2[0-4]\d|1?\d?\d)\b/g },
+    phone: {
+        pattern: /(?<![\w+])\+?(?:\(\d{1,4}\)|\d{1,4})(?:[ .-]?(?:\(\d{1,4}\)|\d{2,4})){2,4}(?!\w)/g,
+        valid: phoneNumber,
+    },
+};
+const ORDER = ['secret', 'email', 'credit-card', 'iban', 'ssn', 'ip-address', 'phone'];
+export function piiMiddleware(options = {}) {
+    const kinds = new Set(options.kinds ?? ORDER);
+    const detectors = [
+        ...ORDER.filter((kind) => kinds.has(kind)).map((kind) => [kind, DETECTORS[kind]]),
+        ...Object.entries(options.patterns ?? {}).map(([kind, pattern]) => [
+            kind,
+            { pattern: pattern.flags.includes('g') ? pattern : new RegExp(pattern.source, `${pattern.flags}g`) },
+        ]),
+    ];
+    const apply = { input: true, output: false, toolResults: true, ...options.apply };
+    const strategyOf = (kind) => typeof options.strategy === 'string' ? options.strategy : (options.strategy?.[kind] ?? 'redact');
+    const clean = (text, where) => {
+        let current = text;
+        for (const [kind, detector] of detectors) {
+            const strategy = strategyOf(kind);
+            let blocked = false;
+            current = current.replace(detector.pattern, (match) => {
+                if (detector.valid && !detector.valid(match))
+                    return match;
+                options.onDetect?.({ kind, where, strategy });
+                if (strategy === 'block')
+                    blocked = true;
+                return replace(match, kind, strategy);
+            });
+            if (blocked)
+                return { blocked: kind };
+        }
+        return { text: current };
+    };
+    const cleanMessage = (message, where) => {
+        if (typeof message.content === 'string') {
+            const result = clean(message.content, where);
+            if ('blocked' in result)
+                throw new PiiBlockedError(result.blocked, where);
+            return result.text === message.content ? message : { ...message, content: result.text };
+        }
+        let changed = false;
+        const content = message.content.map((part) => {
+            if (part.type !== 'text')
+                return part;
+            const result = clean(part.text, where);
+            if ('blocked' in result)
+                throw new PiiBlockedError(result.blocked, where);
+            if (result.text === part.text)
+                return part;
+            changed = true;
+            return { ...part, text: result.text };
+        });
+        return changed ? { ...message, content } : message;
+    };
+    return {
+        name: 'pii',
+        beforeModel({ request }) {
+            if (!apply.input)
+                return;
+            return {
+                ...request,
+                messages: request.messages.map((message) => message.role === 'system' ? message : cleanMessage(message, 'input')),
+            };
+        },
+        afterModel({ response }) {
+            if (!apply.output || !response.content)
+                return;
+            const result = clean(response.content, 'output');
+            if ('blocked' in result)
+                throw new PiiBlockedError(result.blocked, 'output');
+            return result.text === response.content ? undefined : { ...response, content: result.text };
+        },
+        async wrapToolCall(_call, next) {
+            const result = await next();
+            if (!apply.toolResults || !result.ok)
+                return result;
+            let blocked;
+            const walk = (value) => {
+                if (blocked)
+                    return value;
+                if (typeof value === 'string') {
+                    const cleaned = clean(value, 'tool');
+                    if ('blocked' in cleaned) {
+                        blocked = cleaned.blocked;
+                        return value;
+                    }
+                    return cleaned.text;
+                }
+                if (Array.isArray(value))
+                    return value.map(walk);
+                if (isToolOutput(value)) {
+                    const content = value.content.map((part) => part.type === 'text' ? { ...part, text: walk(part.text) } : part);
+                    return { ...value, content };
+                }
+                if (value && typeof value === 'object' && Object.getPrototypeOf(value) === Object.prototype) {
+                    return Object.fromEntries(Object.entries(value).map(([key, entry]) => [key, walk(entry)]));
+                }
+                return value;
+            };
+            const cleaned = walk(result.result);
+            if (blocked)
+                return { ok: false, error: `The tool's result was withheld: it contains ${blocked}` };
+            return { ...result, result: cleaned };
+        },
+    };
+}
+function replace(match, kind, strategy) {
+    if (strategy === 'mask') {
+        const visible = match.replace(/[^A-Za-z0-9]/g, '').slice(-4);
+        return `${'*'.repeat(Math.max(4, match.length - visible.length))}${visible}`;
+    }
+    if (strategy === 'hash')
+        return `[${kind}:${shortHash(match)}]`;
+    return `[${kind}]`;
+}
+function isToolOutput(value) {
+    return value?.type === 'tool_output' && Array.isArray(value.content);
+}

package/dist/agent/middleware/retry.d.ts

@@ -0,0 +1,66 @@
+import type { NexusResponse } from '../../types/response.js';
+import type { AgentMiddleware, AgentToolCall, AgentToolResult } from '../create-agent.js';
+import { type RetryBackoff } from './shared.js';
+/** One retry about to happen. */
+export interface RetryEvent {
+    /** The attempt that failed, from 1. */
+    attempt: number;
+    /** How long until the next attempt, in milliseconds. */
+    delayMs: number;
+    /** The error a model call threw, when it threw. */
+    error?: unknown;
+    /** The failed tool call's result, for a tool retry. */
+    result?: AgentToolResult;
+    /** The tool's name, for a tool retry. */
+    tool?: string;
+}
+/** Options for `modelRetry()`. */
+export interface ModelRetryOptions {
+    /** Attempts in all, the first included. Defaults to 3. */
+    attempts?: number;
+    /** How long to wait between attempts. */
+    backoff?: RetryBackoff;
+    /**
+     * Whether an error is worth another attempt. Defaults to every error not marked
+     * `retryable: false`, such as a provider's authentication error, and never a cancellation.
+     */
+    retryOn?: (error: unknown, attempt: number) => boolean;
+    /** Whether a response that did arrive is tried again, such as an empty one. Defaults to never. */
+    retryResponse?: (response: NexusResponse, attempt: number) => boolean;
+    /** Called before each new attempt, for logs and metrics. */
+    onRetry?: (event: RetryEvent) => void;
+}
+/**
+ * Tries a failed model call again, with exponential backoff.
+ *
+ * The client already retries one provider request; this retries the agent's step as a whole, after
+ * the client gave up, and can retry a response that arrived but is unusable. A streamed call that
+ * fails halfway streams again from the start. Waits stop when the run is cancelled.
+ */
+export declare function modelRetry(options?: ModelRetryOptions): AgentMiddleware;
+/** Options for `toolRetry()`. */
+export interface ToolRetryOptions {
+    /** Attempts in all, the first included. Defaults to 3. */
+    attempts?: number;
+    /** How long to wait between attempts. */
+    backoff?: RetryBackoff;
+    /**
+     * Which calls may run again: tool names, `'all'`, or a function of the call. Defaults to calls
+     * that are safe to repeat, whose tool declares capabilities and none of them writes, runs a
+     * command or code, reaches the network, or is the application's own. A call that may already have
+     * changed something is retried only when you say so.
+     */
+    tools?: readonly string[] | 'all' | ((call: AgentToolCall) => boolean);
+    /** Whether a result is a failure worth another attempt. Defaults to any result with `ok: false`. */
+    retryOn?: (result: AgentToolResult, call: AgentToolCall, attempt: number) => boolean;
+    /** Called before each new attempt, for logs and metrics. */
+    onRetry?: (event: RetryEvent) => void;
+}
+/**
+ * Runs a failed tool call again, with exponential backoff.
+ *
+ * By default only a call that cannot have changed anything is repeated: retrying a payment or a
+ * write that timed out after it landed would do it twice. Name the tools that are idempotent to
+ * retry those too.
+ */
+export declare function toolRetry(options?: ToolRetryOptions): AgentMiddleware;

package/dist/agent/middleware/retry.js

@@ -0,0 +1,59 @@
+import { isSensitiveCapability } from '../capabilities.js';
+import { backoffDelay, isCancellation, sleep } from './shared.js';
+export function modelRetry(options = {}) {
+    const attempts = Math.max(1, options.attempts ?? 3);
+    const retryOn = options.retryOn ?? ((error) => error?.retryable !== false);
+    return {
+        name: 'model-retry',
+        async wrapModelCall(request, next, context) {
+            for (let attempt = 1;; attempt += 1) {
+                try {
+                    const response = await next(request);
+                    if (attempt < attempts && options.retryResponse?.(response, attempt)) {
+                        const delayMs = backoffDelay(attempt, options.backoff);
+                        options.onRetry?.({ attempt, delayMs });
+                        await sleep(delayMs, context.signal);
+                        continue;
+                    }
+                    return response;
+                }
+                catch (error) {
+                    if (attempt >= attempts || isCancellation(error, context.signal) || !retryOn(error, attempt))
+                        throw error;
+                    const delayMs = backoffDelay(attempt, options.backoff);
+                    options.onRetry?.({ attempt, delayMs, error });
+                    await sleep(delayMs, context.signal);
+                }
+            }
+        },
+    };
+}
+export function toolRetry(options = {}) {
+    const attempts = Math.max(1, options.attempts ?? 3);
+    const retryOn = options.retryOn ?? ((result) => !result.ok);
+    const eligible = (call) => {
+        const tools = options.tools;
+        if (tools === 'all')
+            return true;
+        if (typeof tools === 'function')
+            return tools(call);
+        if (tools)
+            return tools.includes(call.name);
+        return call.capabilities !== undefined && !call.capabilities.some(isSensitiveCapability);
+    };
+    return {
+        name: 'tool-retry',
+        async wrapToolCall(call, next, context) {
+            if (!eligible(call))
+                return next();
+            for (let attempt = 1;; attempt += 1) {
+                const result = await next();
+                if (attempt >= attempts || context.signal.aborted || !retryOn(result, call, attempt))
+                    return result;
+                const delayMs = backoffDelay(attempt, options.backoff);
+                options.onRetry?.({ attempt, delayMs, result, tool: call.name });
+                await sleep(delayMs, context.signal);
+            }
+        },
+    };
+}

package/dist/agent/middleware/shared.d.ts

@@ -0,0 +1,41 @@
+import type { Message } from '../../types/messages.js';
+import type { AgentToolCall } from '../create-agent.js';
+/**
+ * What the middleware share. Every count here is read from the transcript rather than kept in
+ * memory, so it survives a restart, a resumed approval, and a second process, and two threads never
+ * share one.
+ */
+/**
+ * The current turn: every message after the latest user message. One `invoke()` is one turn; a run
+ * that resumes after an approval is the same turn.
+ */
+export declare function turnOf(messages: readonly Message[]): readonly Message[];
+/** Model calls made in the current turn: its assistant messages. */
+export declare function modelCallsInTurn(messages: readonly Message[]): number;
+/**
+ * Where a call sits among the current turn's calls to the same tool, from 1. Calls the model made in
+ * parallel are ordered as the model listed them, so the position never depends on which finished
+ * first.
+ */
+export declare function callPosition(messages: readonly Message[], call: Pick<AgentToolCall, 'id' | 'name'>): number;
+/** A message's text: the string, or its text parts joined. */
+export declare function textOf(content: Message['content']): string;
+/** How long to wait between attempts. */
+export interface RetryBackoff {
+    /** The wait before the second attempt, in milliseconds. Defaults to 500. */
+    initialMs?: number;
+    /** The longest wait, in milliseconds. Defaults to 10,000. */
+    maxMs?: number;
+    /** What each wait is multiplied by. Defaults to 2. */
+    factor?: number;
+    /** Spreads each wait between half and all of it, so runs that failed together retry apart. Defaults to true. */
+    jitter?: boolean;
+}
+/** The wait before attempt `attempt + 1`, after `attempt` failed. */
+export declare function backoffDelay(attempt: number, backoff?: RetryBackoff): number;
+/** Waits, or rejects with the signal's reason as soon as it aborts. */
+export declare function sleep(ms: number, signal?: AbortSignal): Promise<void>;
+/** True for a cancellation: never worth retrying or falling back from. */
+export declare function isCancellation(error: unknown, signal?: AbortSignal): boolean;
+/** A short, stable hash of a text: cyrb53 as 8 hex digits. Not a security boundary. */
+export declare function shortHash(text: string): string;

package/dist/agent/middleware/shared.js

@@ -0,0 +1,73 @@
+export function turnOf(messages) {
+    for (let index = messages.length - 1; index >= 0; index -= 1) {
+        if (messages[index]?.role === 'user')
+            return messages.slice(index + 1);
+    }
+    return messages;
+}
+export function modelCallsInTurn(messages) {
+    return turnOf(messages).filter((message) => message.role === 'assistant').length;
+}
+export function callPosition(messages, call) {
+    let position = 0;
+    for (const message of turnOf(messages)) {
+        for (const made of message.toolCalls ?? []) {
+            if (made.function.name !== call.name)
+                continue;
+            position += 1;
+            if (made.id === call.id)
+                return position;
+        }
+    }
+    return position + 1;
+}
+export function textOf(content) {
+    if (typeof content === 'string')
+        return content;
+    return content
+        .filter((part) => part.type === 'text')
+        .map((part) => part.text)
+        .join('\n');
+}
+export function backoffDelay(attempt, backoff = {}) {
+    const initial = backoff.initialMs ?? 500;
+    const delay = Math.min(backoff.maxMs ?? 10_000, initial * (backoff.factor ?? 2) ** Math.max(0, attempt - 1));
+    return backoff.jitter === false ? delay : delay / 2 + Math.random() * (delay / 2);
+}
+export function sleep(ms, signal) {
+    if (ms <= 0)
+        return signal?.aborted ? Promise.reject(signal.reason) : Promise.resolve();
+    return new Promise((resolve, reject) => {
+        if (signal?.aborted) {
+            reject(signal.reason);
+            return;
+        }
+        const onAbort = () => {
+            clearTimeout(timer);
+            reject(signal?.reason);
+        };
+        const timer = setTimeout(() => {
+            signal?.removeEventListener('abort', onAbort);
+            resolve();
+        }, ms);
+        signal?.addEventListener('abort', onAbort, { once: true });
+    });
+}
+export function isCancellation(error, signal) {
+    if (signal?.aborted)
+        return true;
+    const name = error?.name;
+    return name === 'AbortError';
+}
+export function shortHash(text) {
+    let h1 = 0xdeadbeef;
+    let h2 = 0x41c6ce57;
+    for (let index = 0; index < text.length; index += 1) {
+        const code = text.charCodeAt(index);
+        h1 = Math.imul(h1 ^ code, 2654435761);
+        h2 = Math.imul(h2 ^ code, 1597334677);
+    }
+    h1 = Math.imul(h1 ^ (h1 >>> 16), 2246822507) ^ Math.imul(h2 ^ (h2 >>> 13), 3266489909);
+    h2 = Math.imul(h2 ^ (h2 >>> 16), 2246822507) ^ Math.imul(h1 ^ (h1 >>> 13), 3266489909);
+    return (4294967296 * (2097151 & h2) + (h1 >>> 0)).toString(16).padStart(14, '0').slice(-8);
+}

package/dist/agent/middleware/tool-selector.d.ts

@@ -0,0 +1,64 @@
+import type { CompletionRequest, ToolDefinition } from '../../types/messages.js';
+import type { AgentMiddleware, AgentMiddlewareContext } from '../create-agent.js';
+/** What a rule chooses tools from. */
+export interface ToolSelectionContext extends AgentMiddlewareContext {
+    /** The request about to be sent. */
+    request: CompletionRequest;
+    /** Every tool the request offers. */
+    tools: readonly ToolDefinition[];
+    /** What the turn is about: the text `query` returned. */
+    query: string;
+}
+/** Options for `toolSelector()`. Give it `select`, `embed`, or both. */
+export interface ToolSelectorOptions {
+    /**
+     * A rule: the names of the tools to send. With `embed` too, the rule narrows the candidates and
+     * similarity ranks what is left.
+     */
+    select?: (context: ToolSelectionContext) => readonly string[] | Promise<readonly string[]>;
+    /**
+     * Embeds texts, for choosing by similarity: any embedding function, such as one from
+     * `toEmbeddingFunction()`. Each tool is embedded once, from `describe`; the query once per turn.
+     */
+    embed?: (texts: string[]) => Promise<readonly (readonly number[])[]> | readonly (readonly number[])[];
+    /** The most tools sent, `always` included. Defaults to 12. */
+    maxTools?: number;
+    /** The lowest similarity a tool needs to be sent, from -1 to 1. Defaults to none. */
+    minScore?: number;
+    /** Tools sent on every call, whatever the turn is about. */
+    always?: readonly string[];
+    /**
+     * Keeps every tool the model already called this turn, so a tool it is halfway through using
+     * never disappears from under it. Defaults to true.
+     */
+    keepUsed?: boolean;
+    /** What the turn is about. Defaults to the latest user message. */
+    query?: (context: AgentMiddlewareContext & {
+        request: CompletionRequest;
+    }) => string;
+    /** The text a tool is embedded from. Defaults to its name and description. */
+    describe?: (tool: ToolDefinition) => string;
+    /** Called with each selection, for logs and for measuring what the selector saves. */
+    onSelect?: (selection: ToolSelection) => void;
+}
+/** What `toolSelector()` sent for one call. */
+export interface ToolSelection {
+    /** The query the tools were chosen for. */
+    query: string;
+    /** The names of the tools sent. */
+    selected: string[];
+    /** How many tools the request offered. */
+    offered: number;
+    /** Each ranked tool's similarity to the query, when tools were chosen by similarity. */
+    scores?: Record<string, number>;
+}
+/**
+ * Sends a model only the tools relevant to the turn.
+ *
+ * An agent with 150 tools otherwise sends 150 schemas on every call: the tokens are paid each time,
+ * the call is slower, and a model choosing among that many picks worse. A rule, embedding
+ * similarity, or both choose a dozen; tools named in `always`, the tool a `toolChoice` forces, and
+ * the tools the model already used this turn are always kept. The agent can still run every tool —
+ * only what the model is offered changes.
+ */
+export declare function toolSelector(options: ToolSelectorOptions): AgentMiddleware;