@ddwang/magnitude-core 0.3.1-ddwang.2 → 0.3.1-ddwang.21

package/dist/ai/baml_client/watchers.d.ts

@@ -1,32 +0,0 @@
-/*************************************************************************************************
-
-Welcome to Baml! To use this generated code, please run one of the following:
-
-$ npm install @boundaryml/baml
-$ yarn add @boundaryml/baml
-$ pnpm add @boundaryml/baml
-
-*************************************************************************************************/
-export interface VarNotification<T> {
-    variable_name: string;
-    value: T;
-    timestamp: string;
-    function_name: string;
-}
-interface InternalStreamEvent {
-    streamId: string;
-    notificationType: "start" | "update" | "end";
-    value?: any;
-}
-type VarHandler<T> = (event: VarNotification<T>) => void;
-type InternalStreamHandler = (event: InternalStreamEvent) => void;
-export interface InternalEventBindings {
-    functionName: string;
-    vars: Record<string, VarHandler<any>[]>;
-    streams: Record<string, InternalStreamHandler[]>;
-    functions: Record<string, InternalEventBindings>;
-}
-export interface EventCollectorInternal {
-    __handlers(): InternalEventBindings;
-}
-export {};

package/dist/ai/baml_client/watchers.js

@@ -1,50 +0,0 @@
-/*************************************************************************************************
-
-Welcome to Baml! To use this generated code, please run one of the following:
-
-$ npm install @boundaryml/baml
-$ yarn add @boundaryml/baml
-$ pnpm add @boundaryml/baml
-
-*************************************************************************************************/
-// Simple async iterable stream for watch notifications
-class NotificationStream {
-    eventQueue = [];
-    isComplete = false;
-    pushValue(value) {
-        this.eventQueue.push(value);
-    }
-    complete() {
-        this.isComplete = true;
-        this.eventQueue.push(null);
-    }
-    async *[Symbol.asyncIterator]() {
-        while (true) {
-            const event = this.eventQueue.shift();
-            if (event === undefined) {
-                if (this.isComplete) {
-                    break;
-                }
-                // Wait a bit for more events
-                await new Promise(resolve => setTimeout(resolve, 10));
-                continue;
-            }
-            if (event === null) {
-                break;
-            }
-            yield event;
-        }
-    }
-    async getFinalResponse() {
-        // For emit streams, just get the last value
-        let lastValue = null;
-        for await (const value of this) {
-            lastValue = value;
-        }
-        return lastValue;
-    }
-    toStreamable() {
-        throw new Error("toStreamable not implemented for EmitStream");
-    }
-}
-export {};

package/dist/ai/baseten.d.ts

@@ -1,4 +0,0 @@
-import type { BasetenClient } from './types';
-export declare const DEFAULT_BASETEN_MODEL = "deepseek-ai/DeepSeek-V4.1-Flash";
-export declare const DEFAULT_BASETEN_BASE_URL = "https://inference.baseten.co/v1";
-export declare function validateBasetenOptions({ model, reasoningEffort }: BasetenClient['options']): void;

package/dist/ai/baseten.js

@@ -1,10 +0,0 @@
-export const DEFAULT_BASETEN_MODEL = 'deepseek-ai/DeepSeek-V4.1-Flash';
-export const DEFAULT_BASETEN_BASE_URL = 'https://inference.baseten.co/v1';
-export function validateBasetenOptions({ model, reasoningEffort }) {
-    // https://docs.baseten.co/inference/model-apis/reasoning
-    // Other models have different supported levels; leave those to the provider.
-    if (model === DEFAULT_BASETEN_MODEL && reasoningEffort !== undefined
-        && !['none', 'low', 'high', 'max'].includes(reasoningEffort)) {
-        throw new Error(`${model} reasoning effort must be none, low, high, or max.`);
-    }
-}

package/dist/ai/claudeCode.d.ts

@@ -1 +0,0 @@
-export declare function completeClaudeCodeAuthFlow(): Promise<string>;

package/dist/ai/claudeCode.js

@@ -1,186 +0,0 @@
-import { promises as fs } from 'fs';
-import { homedir } from 'os';
-import { join, dirname } from 'path';
-import crypto from 'crypto';
-import { bold, cyanBright } from 'ansis';
-import { exec } from 'node:child_process';
-// Simple cross-platform function to open URLs
-function openUrl(url) {
-    return new Promise((resolve, reject) => {
-        let command;
-        switch (process.platform) {
-            case 'darwin':
-                command = `open "${url}"`;
-                break;
-            case 'win32':
-                command = `start "${url}"`;
-                break;
-            default:
-                command = `xdg-open "${url}"`;
-                break;
-        }
-        exec(command, (error) => {
-            if (error) {
-                reject(error);
-            }
-            else {
-                resolve();
-            }
-        });
-    });
-}
-// Constants
-const CLIENT_ID = '9d1c250a-e61b-44d9-88ed-5944d1962f5e';
-const CREDS_PATH = join(homedir(), '.magnitude', 'credentials', 'claudeCode.json');
-// 1. Generate PKCE pair
-function generatePKCE() {
-    const verifier = crypto.randomBytes(32).toString('base64url');
-    const challenge = crypto
-        .createHash('sha256')
-        .update(verifier)
-        .digest('base64url');
-    return { verifier, challenge };
-}
-// 2. Get OAuth authorization URL
-function getAuthorizationURL(pkce) {
-    const url = new URL('https://claude.ai/oauth/authorize');
-    url.searchParams.set('code', 'true');
-    url.searchParams.set('client_id', CLIENT_ID);
-    url.searchParams.set('response_type', 'code');
-    url.searchParams.set('redirect_uri', 'https://console.anthropic.com/oauth/code/callback');
-    url.searchParams.set('scope', 'org:create_api_key user:profile user:inference');
-    url.searchParams.set('code_challenge', pkce.challenge);
-    url.searchParams.set('code_challenge_method', 'S256');
-    url.searchParams.set('state', pkce.verifier);
-    return url.toString();
-}
-// 3. Exchange authorization code for tokens
-async function exchangeCodeForTokens(code, verifier) {
-    const [authCode, state] = code.split('#');
-    const response = await fetch('https://console.anthropic.com/v1/oauth/token', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-            code: authCode,
-            state: state,
-            grant_type: 'authorization_code',
-            client_id: CLIENT_ID,
-            redirect_uri: 'https://console.anthropic.com/oauth/code/callback',
-            code_verifier: verifier,
-        }),
-    });
-    if (!response.ok) {
-        throw new Error(`Token exchange failed: ${response.statusText}`);
-    }
-    const data = await response.json();
-    return {
-        access_token: data.access_token,
-        refresh_token: data.refresh_token,
-        expires_at: Date.now() + (data.expires_in * 1000),
-    };
-}
-// 4. Refresh access token
-async function refreshAccessToken(refreshToken) {
-    const response = await fetch('https://console.anthropic.com/v1/oauth/token', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({
-            grant_type: 'refresh_token',
-            refresh_token: refreshToken,
-            client_id: CLIENT_ID,
-        }),
-    });
-    if (!response.ok) {
-        throw new Error(`Token refresh failed: ${response.statusText}`);
-    }
-    const data = await response.json();
-    return {
-        access_token: data.access_token,
-        refresh_token: data.refresh_token,
-        expires_at: Date.now() + (data.expires_in * 1000),
-    };
-}
-// 5. Save credentials
-async function saveCredentials(creds) {
-    await fs.mkdir(dirname(CREDS_PATH), { recursive: true });
-    await fs.writeFile(CREDS_PATH, JSON.stringify(creds, null, 2));
-    await fs.chmod(CREDS_PATH, 0o600); // Read/write for owner only
-}
-// 6. Load credentials
-async function loadCredentials() {
-    try {
-        const data = await fs.readFile(CREDS_PATH, 'utf-8');
-        return JSON.parse(data);
-    }
-    catch {
-        return null;
-    }
-}
-// 7. Get valid access token (refresh if needed)
-async function getValidAccessToken() {
-    const creds = await loadCredentials();
-    if (!creds)
-        return null;
-    // If token is still valid, return it
-    if (creds.expires_at > Date.now() + 60000) { // 1 minute buffer
-        return creds.access_token;
-    }
-    // Otherwise, refresh it
-    try {
-        const newCreds = await refreshAccessToken(creds.refresh_token);
-        await saveCredentials(newCreds);
-        return newCreds.access_token;
-    }
-    catch {
-        return null;
-    }
-}
-export async function completeClaudeCodeAuthFlow() {
-    // Try to get existing valid token
-    const existingToken = await getValidAccessToken();
-    if (existingToken)
-        return existingToken;
-    // Otherwise, go through auth flow
-    const pkce = generatePKCE();
-    const authUrl = getAuthorizationURL(pkce);
-    console.log(bold `Claude Code Pro/Max access token missing or expired.`);
-    //console.log(cyanBright`Accounts with Max plan can be used for API access.`)
-    console.log(cyanBright `Opening browser for authentication...`);
-    try {
-        await openUrl(authUrl);
-    }
-    catch (err) {
-        console.log('Could not open browser automatically');
-    }
-    //console.log('Open this URL in your browser:');
-    console.log(bold `\nIf browser did not open, visit:`);
-    console.log(authUrl);
-    console.log(bold `\nPaste the authorization code here:`);
-    const code = await new Promise((resolve) => {
-        process.stdin.once('data', (data) => {
-            resolve(data.toString().trim());
-        });
-    });
-    const creds = await exchangeCodeForTokens(code, pkce.verifier);
-    await saveCredentials(creds);
-    console.log(bold `\nCredentials saved!`);
-    return creds.access_token;
-}
-/*
-// Usage
-const response = await fetch('https://api.anthropic.com/v1/messages', {
-  method: 'POST',
-  headers: {
-    'Content-Type': 'application/json',
-    'Authorization': `Bearer ${accessToken}`,
-    'anthropic-beta': 'oauth-2025-04-20',
-    // make sure to exclude x-api-key
-  },
-  body: JSON.stringify({
-    model: 'claude-3-5-sonnet-20241022',
-    messages: [{ role: 'user', content: 'Hello!' }],
-    max_tokens: 100,
-  }),
-});
-
-*/

package/dist/ai/modelHarness.d.ts

@@ -1,39 +0,0 @@
-import { AgentContext } from "@/ai/baml_client";
-import { ClientRegistry } from "@boundaryml/baml";
-import { LLMClient, ModelUsage } from "@/ai/types";
-import { ActionDefinition } from "@/actions";
-import { Schema, z } from 'zod';
-import { Image } from '@/memory/image';
-import EventEmitter from "eventemitter3";
-import { MultiMediaContentPart } from "@/memory/rendering";
-import { type PlannerResponse } from './plannerResponse';
-interface ModelHarnessOptions {
-    llm: LLMClient;
-}
-export interface ModelHarnessEvents {
-    'tokensUsed': (usage: ModelUsage) => {};
-}
-export declare class ModelHarness {
-    /**
-     * Strong reasoning agent for high level strategy and planning.
-     */
-    readonly events: EventEmitter<ModelHarnessEvents>;
-    private options;
-    private cr;
-    private clientOptions;
-    private baml;
-    private logger;
-    private planner?;
-    constructor(options: ModelHarnessOptions);
-    setup(): Promise<void>;
-    protected createClientRegistry(options: Record<string, any>): ClientRegistry;
-    private clientForSchema;
-    describeModel(): string;
-    private _withUsage;
-    private _reportCallUsage;
-    partialAct<T>(context: AgentContext, // Changed to ModularMemoryContext
-    task: string, data: MultiMediaContentPart[], actionVocabulary: ActionDefinition<T>[]): Promise<PlannerResponse>;
-    extract<T extends Schema>(instructions: string, schema: T, screenshot: Image, domContent: string): Promise<z.infer<T>>;
-    query<T extends Schema>(context: AgentContext, query: string, schema: T): Promise<z.infer<T>>;
-}
-export {};

package/dist/ai/modelHarness.js

@@ -1,324 +0,0 @@
-import { convertToBamlClientOptions } from "./util";
-// Import ModularMemoryContext instead of old MemoryContext
-import { b } from "@/ai/baml_client";
-import { Collector, ClientRegistry, BamlValidationError } from "@boundaryml/baml";
-import logger from "@/logger";
-import TypeBuilder from "./baml_client/type_builder";
-import { z } from 'zod';
-import { convertActionDefinitionsToBaml, convertZodToBaml } from "@/actions/util";
-import EventEmitter from "eventemitter3";
-import { parsePlannerResponse, PlannerResponseError, memoryUpdatesSchema } from './plannerResponse';
-import { anthropicOutputFormat, plannerSchema, usesStructuredOutput } from './structuredOutput';
-import { ModelResponseError } from './modelResponseError';
-import { DEFAULT_BASETEN_MODEL } from './baseten';
-import { checkOperation, operationOptions } from '@/common/operation';
-export class ModelHarness {
-    /**
-     * Strong reasoning agent for high level strategy and planning.
-     */
-    events = new EventEmitter();
-    options;
-    cr;
-    clientOptions;
-    baml;
-    logger;
-    planner;
-    constructor(options) {
-        this.options = {
-            llm: options.llm,
-            //promptCaching: options.promptCaching ?? false
-        };
-        this.logger = logger.child({ name: 'llm' });
-    }
-    async setup() {
-        // Must be called after constructor
-        this.planner = undefined;
-        this.clientOptions = await convertToBamlClientOptions(this.options.llm);
-        this.cr = this.createClientRegistry(this.clientOptions);
-        this.baml = b.withOptions({ clientRegistry: this.cr });
-    }
-    createClientRegistry(options) {
-        const registry = new ClientRegistry();
-        registry.addLlmClient('Magnus', this.options.llm.provider === 'claude-code' ? 'anthropic'
-            : this.options.llm.provider === 'baseten' ? 'openai-generic' : this.options.llm.provider, options, 'DefaultRetryPolicy');
-        registry.setPrimary('Magnus');
-        return registry;
-    }
-    clientForSchema(schema) {
-        if (!usesStructuredOutput(this.options.llm))
-            return this.cr;
-        // Reuse the conservative schema subset across providers; all original
-        // value constraints and notebook semantics remain locally validated.
-        const format = anthropicOutputFormat(schema);
-        if (!format) {
-            this.logger.debug('Schema requires prompt-only output; native structured output cannot represent it');
-            return this.cr;
-        }
-        const outputOptions = this.options.llm.provider === 'baseten'
-            ? { response_format: { type: 'json_schema', json_schema: { name: 'magnitude_response', strict: true, schema: format.schema } } }
-            : { output_config: { format } };
-        return this.createClientRegistry({ ...this.clientOptions, ...outputOptions });
-    }
-    describeModel() {
-        return `${this.options.llm.provider}:${'model' in this.options.llm.options ? this.options.llm.options.model : 'unknown'}`;
-    }
-    async _withUsage(invoke) {
-        checkOperation();
-        // Scope usage to this invocation, including failed parses and provider
-        // retries. A shared cumulative collector can double-count concurrent calls.
-        const collector = new Collector('model-call');
-        try {
-            try {
-                return await invoke(collector);
-            }
-            finally {
-                checkOperation();
-                // Even syntactically complete JSON must not hide a provider
-                // refusal or a truncated batch. Neither gets a format retry.
-                if (this.options.llm.provider === 'anthropic' || this.options.llm.provider === 'claude-code') {
-                    let reason;
-                    try {
-                        reason = collector.last?.calls.at(-1)?.httpResponse?.body.json()?.stop_reason;
-                    }
-                    catch { /* A transport error may not contain JSON. */ }
-                    if (reason === 'refusal' || reason === 'max_tokens')
-                        throw new ModelResponseError(reason);
-                }
-                else if (this.options.llm.provider === 'openai' || this.options.llm.provider === 'baseten') {
-                    let choice;
-                    try {
-                        choice = collector.last?.calls.at(-1)?.httpResponse?.body.json()?.choices?.[0];
-                    }
-                    catch { /* A transport error may not contain JSON. */ }
-                    if (choice?.message?.refusal || choice?.finish_reason === 'content_filter')
-                        throw new ModelResponseError('refusal');
-                    if (choice?.finish_reason === 'length')
-                        throw new ModelResponseError('max_tokens');
-                }
-            }
-        }
-        finally {
-            for (const log of collector.logs) {
-                for (const call of log.calls) {
-                    try {
-                        this._reportCallUsage(call);
-                    }
-                    catch {
-                        this.logger.warn('Unable to report model response usage');
-                    }
-                }
-            }
-        }
-    }
-    _reportCallUsage(call) {
-        let inputTokens = call.usage?.inputTokens;
-        let outputTokens = call.usage?.outputTokens;
-        let cacheWriteInputTokens = 0;
-        let cacheReadInputTokens = 0;
-        if (this.options.llm.provider === 'anthropic' || this.options.llm.provider === 'claude-code') {
-            // Anthropic's input_tokens excludes cache reads/writes. BAML's
-            // normalized usage does not expose that breakdown.
-            try {
-                const usage = call.httpResponse?.body.json()?.usage;
-                if (usage) {
-                    inputTokens = usage.input_tokens ?? inputTokens;
-                    outputTokens = usage.output_tokens ?? outputTokens;
-                    cacheWriteInputTokens = usage.cache_creation_input_tokens ?? 0;
-                    cacheReadInputTokens = usage.cache_read_input_tokens ?? 0;
-                }
-            }
-            catch { /* Non-JSON error response; use per-call usage if available. */ }
-        }
-        else if (this.options.llm.provider === 'openai' || this.options.llm.provider === 'baseten') {
-            try {
-                const usage = call.httpResponse?.body.json()?.usage;
-                if (usage) {
-                    // Both providers include cached input in prompt_tokens and reasoning
-                    // output in completion_tokens. Neither should be counted twice.
-                    cacheReadInputTokens = usage.prompt_tokens_details?.cached_tokens ?? 0;
-                    cacheWriteInputTokens = usage.prompt_tokens_details?.cache_write_tokens ?? 0;
-                    const totalInputTokens = usage.prompt_tokens ?? inputTokens;
-                    if (totalInputTokens != null)
-                        inputTokens = totalInputTokens - cacheReadInputTokens - cacheWriteInputTokens;
-                    outputTokens = usage.completion_tokens ?? outputTokens;
-                }
-            }
-            catch { /* Non-JSON error response; use per-call usage if available. */ }
-        }
-        // A transport failure with no usage is not a paid completion. In
-        // particular, never reuse the preceding successful response's usage.
-        if (inputTokens == null && outputTokens == null)
-            return;
-        inputTokens ??= 0;
-        outputTokens ??= 0;
-        const model = this.options.llm.options.model ?? 'unknown';
-        const isLuna = /^gpt-5\.6-luna(?:-\d{4}-\d{2}-\d{2})?$/.test(model);
-        // Get cost if known
-        const knownCostMap = this.options.llm.provider === 'baseten' ? {
-            // Baseten Model API rates, verified 2026-09-15. Do not inherit other providers' prices.
-            // https://www.baseten.co/library/deepseek-v41-flash/
-            [DEFAULT_BASETEN_MODEL]: { inputTokens: 0.30, outputTokens: 1.20, cacheReadInputTokens: 0.03 },
-        } : {
-            // TODO: track cached savings on Gemini
-            'gemini-2.5-pro': { inputTokens: 1.25, outputTokens: 10.0 },
-            'gemini-2.5-flash': { inputTokens: 0.30, outputTokens: 2.50 },
-            'gemini-2.5-flash-lite': { inputTokens: 0.10, outputTokens: 0.40 },
-            'claude-3.5-sonnet': { inputTokens: 3.00, outputTokens: 15.00, cacheWriteInputTokens: 3.75, cacheReadInputTokens: 0.30 },
-            'claude-3.7-sonnet': { inputTokens: 3.00, outputTokens: 15.00, cacheWriteInputTokens: 3.75, cacheReadInputTokens: 0.30 },
-            'claude-sonnet-4': { inputTokens: 3.00, outputTokens: 15.00, cacheWriteInputTokens: 3.75, cacheReadInputTokens: 0.30 },
-            // Standard API pricing, using the default 5-minute cache TTL:
-            // https://platform.claude.com/docs/en/models/sonnet-5/overview#pricing
-            'claude-sonnet-5': { inputTokens: 2.00, outputTokens: 10.00, cacheWriteInputTokens: 2.50, cacheReadInputTokens: 0.20 },
-            'claude-haiku-4-5': { inputTokens: 1.00, outputTokens: 5.00, cacheWriteInputTokens: 1.25, cacheReadInputTokens: 0.10 },
-            'claude-opus-4': { inputTokens: 15.00, outputTokens: 75.00, cacheWriteInputTokens: 18.75, cacheReadInputTokens: 1.50 },
-            'gpt-4.1': { inputTokens: 2.00, outputTokens: 8.00 },
-            'gpt-4.1-mini': { inputTokens: 0.40, outputTokens: 1.60 },
-            'gpt-4.1-nano': { inputTokens: 0.10, outputTokens: 0.40 },
-            'gpt-4o': { inputTokens: 3.75, outputTokens: 15.00 },
-            // https://developers.openai.com/api/docs/models/gpt-5.6-luna
-            'gpt-5.6-luna': { inputTokens: 0.20, outputTokens: 1.20, cacheWriteInputTokens: 0.25, cacheReadInputTokens: 0.02 },
-            // Assuming Nebius prices, may be higher
-            'qwen2.5-vl-72b': { inputTokens: 0.25, outputTokens: 0.75 }
-        };
-        let inputTokenCost;
-        let outputTokenCost;
-        let cacheWriteInputTokenCost;
-        let cacheReadInputTokenCost;
-        for (const [name, costs] of Object.entries(knownCostMap)) {
-            if (this.options.llm.provider === 'baseten' ? model === name : name === 'gpt-5.6-luna' ? isLuna : model.includes(name)) {
-                inputTokenCost = costs.inputTokens / 1_000_000;
-                outputTokenCost = costs.outputTokens / 1_000_000;
-                cacheReadInputTokenCost = costs.cacheReadInputTokens ? costs.cacheReadInputTokens / 1_000_000 : undefined;
-                cacheWriteInputTokenCost = costs.cacheWriteInputTokens ? costs.cacheWriteInputTokens / 1_000_000 : undefined;
-            }
-        }
-        // Luna long-context pricing applies to the entire request, including
-        // cached input. Output usage already includes billed reasoning tokens.
-        if (isLuna && inputTokens + cacheReadInputTokens + cacheWriteInputTokens > 272_000) {
-            if (inputTokenCost !== undefined)
-                inputTokenCost *= 2;
-            if (cacheWriteInputTokenCost !== undefined)
-                cacheWriteInputTokenCost *= 2;
-            if (cacheReadInputTokenCost !== undefined)
-                cacheReadInputTokenCost *= 2;
-            if (outputTokenCost !== undefined)
-                outputTokenCost *= 1.5;
-        }
-        // console.log("cacheWriteInputTokenCost:", cacheWriteInputTokenCost);
-        // console.log("cacheWriteInputTokens:", cacheWriteInputTokens);
-        const usage = {
-            llm: {
-                provider: this.options.llm.provider,
-                model: model
-            }, //this.options.llm,
-            inputTokens: inputTokens,
-            outputTokens: outputTokens,
-            ...(cacheWriteInputTokens ? { cacheWriteInputTokens } : {}),
-            ...(cacheReadInputTokens ? { cacheReadInputTokens } : {}),
-            ...(inputTokenCost !== undefined && (!cacheWriteInputTokens || cacheWriteInputTokenCost !== undefined) && (!cacheReadInputTokens || cacheReadInputTokenCost !== undefined) ? {
-                inputCost: inputTokens * inputTokenCost +
-                    (cacheWriteInputTokenCost ? cacheWriteInputTokenCost * cacheWriteInputTokens : 0.0) +
-                    (cacheReadInputTokenCost ? cacheReadInputTokenCost * cacheReadInputTokens : 0.0)
-            } : {}),
-            ...(outputTokenCost ? { outputCost: outputTokens * outputTokenCost } : {}),
-            // ...(cacheWriteInputTokenCost ? { : inputTokens * inputTokenCost } : {}),
-            // ...(cacheReadInputTokenCost ? { outputCost: outputTokens * outputTokenCost } : {})
-        };
-        this.events.emit('tokensUsed', usage);
-        //console.log("Usage:", usage);
-    }
-    async partialAct(context, // Changed to ModularMemoryContext
-    task, data, actionVocabulary) {
-        // Notes have one planner path: a required review before browser actions.
-        // Keep memory:note registered on Agent for execution and explicit callers.
-        actionVocabulary = actionVocabulary.filter(action => action.name !== 'memory:note');
-        if (!this.planner || this.planner.vocabulary.length !== actionVocabulary.length
-            || actionVocabulary.some((action, index) => {
-                const saved = this.planner.vocabulary[index];
-                return action.name !== saved.name || action.description !== saved.description || action.schema !== saved.schema;
-            })) {
-            const tb = new TypeBuilder();
-            tb.PartialRecipe.addProperty('memory_updates', convertZodToBaml(tb, memoryUpdatesSchema)).description(memoryUpdatesSchema.description);
-            tb.PartialRecipe.addProperty('actions', tb.list(convertActionDefinitionsToBaml(tb, actionVocabulary))).description('Always provide at least one action');
-            this.planner = {
-                vocabulary: actionVocabulary.map(({ name, description, schema }) => ({ name, description, schema })),
-                tb, clientRegistry: this.clientForSchema(plannerSchema(actionVocabulary)),
-            };
-        }
-        const { tb, clientRegistry } = this.planner;
-        for (let attempt = 0;; attempt++) {
-            try {
-                return await this._withUsage(async (collector) => {
-                    let bamlRejected = false;
-                    try {
-                        await this.baml.CreatePartialRecipe(context, task, data, this.options.llm.provider === 'claude-code', { tb, collector, clientRegistry, signal: operationOptions().signal });
-                    }
-                    catch (error) {
-                        checkOperation();
-                        if (!(error instanceof BamlValidationError))
-                            throw error;
-                        bamlRejected = true;
-                    }
-                    // BAML can fail first or coerce invalid fields. Diagnose its
-                    // raw response locally, but never bypass either validator.
-                    checkOperation();
-                    const plan = parsePlannerResponse(collector.last?.rawLlmResponse ?? null, actionVocabulary);
-                    if (bamlRejected)
-                        throw new PlannerResponseError('$: BAML parser rejected the response despite local validation; return a plan matching the supplied schema');
-                    return plan;
-                });
-            }
-            catch (error) {
-                checkOperation();
-                if (!(error instanceof PlannerResponseError))
-                    throw error;
-                this.logger.warn({ attempt: attempt + 1, diagnostic: error.diagnostic }, attempt === 0
-                    ? 'Invalid planner response; retrying once with the same observations'
-                    : 'Invalid planner response; no format retries remain');
-                if (attempt === 1)
-                    throw new PlannerResponseError(error.diagnostic, 'Planner returned an invalid plan on both attempts');
-                // No invalid output is appended to memory or executed. Keep the
-                // correction short even when the rejected response is enormous.
-                context = { ...context, observationContent: [...context.observationContent, {
-                            role: 'user', cacheControl: false, content: [`Your previous response was rejected as an invalid plan. No actions were executed. Validation: ${error.diagnostic}. Correct the reported fields. Return only one complete JSON object with concise reasoning, a memory_updates array (empty when nothing new needs retaining), and a non-empty actions array matching the schema. No XML, prose, simulated tool calls, or imagined observations. Plan only the next batch from the observations above.`],
-                        }] };
-            }
-        }
-    }
-    async extract(instructions, schema, screenshot, domContent) {
-        const tb = new TypeBuilder();
-        if (schema instanceof z.ZodObject) {
-            // populate ExtractedData with schema KVs instead of wrapping in data key unnecessarily
-            for (const [key, fieldSchema] of Object.entries(schema.shape)) {
-                tb.ExtractedData.addProperty(key, convertZodToBaml(tb, fieldSchema));
-            }
-        }
-        else {
-            // for array or primitive have to wrap data key
-            tb.ExtractedData.addProperty('data', convertZodToBaml(tb, schema));
-        }
-        // } else if (schema instanceof z.ZodArray) {
-        // }
-        const clientRegistry = this.clientForSchema(schema instanceof z.ZodObject ? schema : z.object({ data: schema }));
-        const bamlScreenshot = await screenshot.toBaml();
-        const resp = await this._withUsage(collector => this.baml.ExtractData(instructions, bamlScreenshot, domContent, this.options.llm.provider === 'claude-code', { tb, collector, clientRegistry, signal: operationOptions().signal }));
-        return schema.parse(schema instanceof z.ZodObject ? resp : resp.data);
-    }
-    // ^ extract could prob be a subset of query w trimmed mem
-    async query(context, query, schema) {
-        const tb = new TypeBuilder();
-        if (schema instanceof z.ZodObject) {
-            // populate ExtractedData with schema KVs instead of wrapping in data key unnecessarily
-            for (const [key, fieldSchema] of Object.entries(schema.shape)) {
-                tb.QueryResponse.addProperty(key, convertZodToBaml(tb, fieldSchema));
-            }
-        }
-        else {
-            // for array or primitive have to wrap data key
-            tb.QueryResponse.addProperty('data', convertZodToBaml(tb, schema));
-        }
-        const clientRegistry = this.clientForSchema(schema instanceof z.ZodObject ? schema : z.object({ data: schema }));
-        const resp = await this._withUsage(collector => this.baml.QueryMemory(context, query, this.options.llm.provider === 'claude-code', { tb, collector, clientRegistry, signal: operationOptions().signal }));
-        return schema.parse(schema instanceof z.ZodObject ? resp : resp.data);
-    }
-}

package/dist/ai/modelResponseError.d.ts

@@ -1,4 +0,0 @@
-export declare class ModelResponseError extends Error {
-    readonly reason: 'refusal' | 'max_tokens';
-    constructor(reason: 'refusal' | 'max_tokens');
-}

package/dist/ai/modelResponseError.js

@@ -1,8 +0,0 @@
-export class ModelResponseError extends Error {
-    reason;
-    constructor(reason) {
-        super(reason === 'refusal' ? 'Model refused the request' : 'Model response exceeded its output-token limit');
-        this.reason = reason;
-        this.name = 'ModelResponseError';
-    }
-}