shortcutxl 0.3.76 → 0.3.78

package/dist/embedded-agent/bootstrap.js

@@ -1,10 +1,10 @@
-import { DEFAULT_THINKING_LEVEL } from '../core/defaults.js';
+import { resolveShortcutModelDefaultReasoningLevel } from '../shortcut-model-catalog.js';
 import { ShortcutXLAgentProviderId } from './model-registry.js';
 export function createShortcutXLDefaultSessionContext(modelId, provider = ShortcutXLAgentProviderId.Shortcut) {
     return {
         messages: [],
         model: { modelId, provider },
-        thinkingLevel: DEFAULT_THINKING_LEVEL
+        thinkingLevel: resolveShortcutModelDefaultReasoningLevel(modelId) ?? 'high'
     };
 }
 export function createShortcutXLAgentSettingsStore() {

package/dist/embedded-agent/compose.js

@@ -3,6 +3,7 @@
  * loop, session persistence, compaction, run stats, credit gate, suggestions,
  * and pending host-tool requests into the public runtime facade.
  */
+import { resolveModelReasoningLevel } from '../ai/models.js';
 import { AGENT_TOOL_IN_FLIGHT_DISPLAY_DELAY_MS, AgentEventKind, AgentMainMessageStreamEventType, AgentTaskControlAction, AgentTelemetryEventType, AgentTelemetryTerminalPhase } from '../contracts/agent-api.js';
 import { AgentStopReason } from '../contracts/agent-snapshot.js';
 import { AgentController } from '../core/agent-controller.js';
@@ -34,7 +35,7 @@ import { buildUserMessage } from './prompt/build-user-message.js';
 import { createEmbeddedRecoveryWiring } from './recovery-wiring.js';
 import { createRunStatsAccumulator } from './run-stats.js';
 import { createRunTelemetryFinalizer } from './run-telemetry-finalizer.js';
-import { ShortcutThinkingLevelToEffort } from './stream.js';
+import { toShortcutEffortLevel } from './stream.js';
 import { runSuggestionsViaBridge } from './suggestions.js';
 import { EmbeddedTaskNotificationTurnQueue } from './task-notification-turn-queue.js';
 import { EmbeddedSubmitOverlapPolicy, EmbeddedTurnLifecycle, EmbeddedTurnPhase, getSubmitOverlapPolicy } from './turn-lifecycle.js';
@@ -205,12 +206,12 @@ function applySubmitSelections(agent, modelRegistry, sessionStore, submitOptions
         }
         agent.setModel(overrideModel);
     }
-    if (submitOptions?.thinkingLevel !== undefined) {
-        if (agent.state.thinkingLevel !== submitOptions.thinkingLevel) {
-            sessionStore.appendThinkingLevelChange(submitOptions.thinkingLevel);
-        }
-        agent.setThinkingLevel(submitOptions.thinkingLevel);
+    const requestedThinking = submitOptions?.thinkingLevel ?? agent.state.thinkingLevel;
+    const effectiveThinking = resolveModelReasoningLevel(agent.state.model, requestedThinking);
+    if (agent.state.thinkingLevel !== effectiveThinking) {
+        sessionStore.appendThinkingLevelChange(effectiveThinking);
     }
+    agent.setThinkingLevel(effectiveThinking);
     if (submitOptions?.speed !== undefined) {
         if (agent.speed !== submitOptions.speed) {
             sessionStore.appendSpeedChange(submitOptions.speed);
@@ -658,7 +659,7 @@ export async function composeEmbeddedAgentRuntimeInternal(options) {
         // actually streams against, including host-degraded Pivot runs.
         // `RunMetricsDropdown` keys its display off this value.
         resolveRunConfig: () => {
-            const effortLevel = ShortcutThinkingLevelToEffort[agent.state.thinkingLevel];
+            const effortLevel = toShortcutEffortLevel(agent.state.thinkingLevel);
             return {
                 modelName: agent.state.model?.id,
                 effortLevel,

package/dist/embedded-agent/gateway-base-url.d.ts

@@ -13,6 +13,7 @@
  * transports and the Shortcut api-v2 dispatcher route through it so they cannot
  * disagree (which previously broke GLM 5.2 over the OpenAI-completions path).
  */
+import type { Api, Model } from '../ai/types.js';
 /** Ensure the base URL ends with exactly one "/v1" segment. */
 export declare function withV1Segment(baseUrl: string): string;
 /** Ensure the base URL does NOT end with a "/v1" segment. */
@@ -24,4 +25,6 @@ export declare function withoutV1Segment(baseUrl: string): string;
 export type GatewayTransport = 'anthropic-messages' | 'openai-completions' | 'openai-responses' | 'google-vertex';
 /** Shape the gateway base URL the way the given transport's SDK expects. */
 export declare function gatewayBaseUrlForTransport(baseUrl: string, transport: GatewayTransport): string;
+/** Present a gateway model to the shared provider layer with its native identity. */
+export declare function gatewayModelForTransport<TApi extends GatewayTransport>(model: Model<Api>, baseUrl: string, transport: TApi): Model<TApi>;
 //# sourceMappingURL=gateway-base-url.d.ts.map

package/dist/embedded-agent/gateway-base-url.js

@@ -38,4 +38,39 @@ export function gatewayBaseUrlForTransport(baseUrl, transport) {
             return stripTrailingSlashes(baseUrl);
     }
 }
+function providerForTransport(transport) {
+    switch (transport) {
+        case 'anthropic-messages':
+            return 'anthropic';
+        case 'google-vertex':
+            return 'google';
+        case 'openai-completions':
+        case 'openai-responses':
+            return 'openai';
+    }
+}
+/** Present a gateway model to the shared provider layer with its native identity. */
+export function gatewayModelForTransport(model, baseUrl, transport) {
+    return {
+        ...model,
+        api: transport,
+        baseUrl: gatewayBaseUrlForTransport(baseUrl, transport),
+        provider: providerForTransport(transport),
+        // Dynamic Anthropic aliases use the current adaptive-effort contract.
+        // Curated catalog metadata still wins, including Haiku's budget control.
+        ...(transport === 'anthropic-messages' && !model.capabilities?.reasoning
+            ? {
+                capabilities: {
+                    ...model.capabilities,
+                    reasoning: {
+                        control: 'adaptive-effort',
+                        defaultLevel: 'high',
+                        levels: ['low', 'medium', 'high', 'xhigh', 'max'],
+                        supportsOff: true
+                    }
+                }
+            }
+            : {})
+    };
+}
 //# sourceMappingURL=gateway-base-url.js.map

package/dist/embedded-agent/google-vertex-transport.d.ts

@@ -1,17 +1,8 @@
 import { type ShortcutXLModelInvokeResponse } from '../contracts/model-call.js';
 import type { ShortcutXLAiModelInvokeRequest, ShortcutXLAiModelStreamRequest } from '../contracts/wire.js';
-import type { ShortcutXLModelUsageStatusLogger } from './model-call-orchestrator.js';
+import { type ProviderGatewayOptions, type ShortcutXLModelUsageStatusLogger } from './provider-native-request.js';
 import type { RunShortcutXLModelStreamRequestOptions } from './stream.js';
-export interface GoogleVertexGatewayAuth {
-    readonly headerName: 'Authorization' | 'x-api-key';
-    readonly token: string;
-}
-export interface GoogleVertexTransportOptions {
-    readonly auth: GoogleVertexGatewayAuth;
-    readonly baseUrl: string;
-    readonly fetchImpl?: typeof fetch;
-    readonly fetchOptions?: RequestInit;
-    readonly headers?: Readonly<Record<string, string | null>>;
+export interface GoogleVertexTransportOptions extends ProviderGatewayOptions {
     readonly usageStatusLogger?: ShortcutXLModelUsageStatusLogger;
 }
 export declare function streamAiGoogleVertexModelRequest(request: ShortcutXLAiModelStreamRequest, options: GoogleVertexTransportOptions & {

package/dist/embedded-agent/google-vertex-transport.js

@@ -1,111 +1,9 @@
-import { convertMessages, convertTools, mapStopReasonString, mapToolChoice } from '../ai/providers/google-shared.js';
-import { sanitizeSurrogates } from '../ai/utils/sanitize-unicode.js';
+import { mapStopReasonString } from '../ai/providers/google-shared.js';
+import { buildGoogleVertexRequest, googleVertexGatewayEndpoint } from '../ai/providers/google-vertex.js';
 import { ShortcutXLModelCallTransportKind, ShortcutXLProviderUsageStatus } from '../contracts/model-call.js';
 import { toProviderUsageFromAi, toShortcutXLModelStreamEvent } from './anthropic-messages-transport.js';
+import { applyGatewayProviderPayload, concreteProviderHeaders, createGatewayProviderModel, createGatewayProviderOptions } from './provider-native-request.js';
 import { normalizeTransportFetchError, normalizeTransportStreamError, throwInvalidResponse, throwTransportHttpError } from './transport-errors.js';
-const ApiV2GoogleProject = '-';
-const ApiV2GoogleLocation = 'global';
-function normalizeBaseUrl(baseUrl) {
-    return baseUrl.replace(/\/+$/, '');
-}
-function createAuthHeader(auth) {
-    if (auth.headerName === 'Authorization') {
-        return {
-            Authorization: auth.token.startsWith('Bearer ') ? auth.token : `Bearer ${auth.token}`
-        };
-    }
-    return { 'x-api-key': auth.token };
-}
-function concreteHeaders(headers) {
-    const result = {};
-    for (const [key, value] of Object.entries(headers ?? {})) {
-        if (value !== null)
-            result[key] = value;
-    }
-    return result;
-}
-function buildGatewayAiModel(request, baseUrl) {
-    return {
-        api: 'google-vertex',
-        baseUrl: normalizeBaseUrl(baseUrl),
-        contextWindow: request.model.contextWindow,
-        cost: { cacheRead: 0, cacheWrite: 0, input: 0, output: 0 },
-        id: request.model.id,
-        input: request.model.input ?? ['text', 'image'],
-        maxTokens: request.model.maxTokens,
-        name: request.model.name,
-        provider: 'google',
-        reasoning: request.model.reasoning
-    };
-}
-function mapThinkingLevel(effort) {
-    switch (effort) {
-        case 'minimal':
-            return 'MINIMAL';
-        case 'low':
-            return 'LOW';
-        case 'medium':
-            return 'MEDIUM';
-        case 'high':
-        case 'xhigh':
-            return 'HIGH';
-    }
-}
-function buildToolConfig(request) {
-    const toolChoice = request.options && 'toolChoice' in request.options ? request.options.toolChoice : undefined;
-    if (!request.context.tools || request.context.tools.length === 0 || !toolChoice) {
-        return undefined;
-    }
-    if (typeof toolChoice === 'object') {
-        return {
-            functionCallingConfig: {
-                allowedFunctionNames: [toolChoice.name],
-                mode: 'ANY'
-            }
-        };
-    }
-    return {
-        functionCallingConfig: {
-            mode: mapToolChoice(toolChoice === 'required' ? 'any' : toolChoice)
-        }
-    };
-}
-function buildRequestBody(request, baseUrl) {
-    const model = buildGatewayAiModel(request, baseUrl);
-    const generationConfig = {};
-    if (request.options?.maxTokens !== undefined) {
-        generationConfig.maxOutputTokens = request.options.maxTokens;
-    }
-    if (request.options?.temperature !== undefined) {
-        generationConfig.temperature = request.options.temperature;
-    }
-    if (request.options?.reasoning && model.reasoning) {
-        generationConfig.thinkingConfig = {
-            includeThoughts: true,
-            thinkingLevel: mapThinkingLevel(request.options.reasoning)
-        };
-    }
-    const toolConfig = buildToolConfig(request);
-    return {
-        contents: convertMessages(model, request.context),
-        ...(request.context.systemPrompt
-            ? {
-                systemInstruction: {
-                    parts: [{ text: sanitizeSurrogates(request.context.systemPrompt) }]
-                }
-            }
-            : {}),
-        ...(request.context.tools && request.context.tools.length > 0
-            ? { tools: convertTools(request.context.tools) }
-            : {}),
-        ...(toolConfig ? { toolConfig } : {}),
-        ...(Object.keys(generationConfig).length > 0 ? { generationConfig } : {})
-    };
-}
-function googleEndpoint(baseUrl, modelId, operation) {
-    const model = encodeURIComponent(modelId);
-    return `${normalizeBaseUrl(baseUrl)}/v1/projects/${ApiV2GoogleProject}/locations/${ApiV2GoogleLocation}/publishers/google/models/${model}:${operation}`;
-}
 function toUsage(usage) {
     const input = usage?.promptTokenCount ?? 0;
     const output = (usage?.candidatesTokenCount ?? 0) + (usage?.thoughtsTokenCount ?? 0);
@@ -231,16 +129,18 @@ function createInitialMessage(model) {
 }
 export async function streamAiGoogleVertexModelRequest(request, options) {
     const fetchImpl = options.fetchImpl ?? fetch;
-    const model = buildGatewayAiModel(request, options.baseUrl);
+    const model = createGatewayProviderModel(request, 'google-vertex', options.baseUrl);
+    const requestOptions = createGatewayProviderOptions(request, 'google-vertex', options, {
+        signal: options.signal
+    });
     let response;
     try {
-        response = await fetchImpl(`${googleEndpoint(options.baseUrl, model.id, 'streamGenerateContent')}?alt=sse`, {
-            body: JSON.stringify(buildRequestBody(request, options.baseUrl)),
+        response = await fetchImpl(`${googleVertexGatewayEndpoint(options.baseUrl, model.id, 'streamGenerateContent')}?alt=sse`, {
+            body: JSON.stringify(applyGatewayProviderPayload(request, 'google-vertex', buildGoogleVertexRequest(model, request.context, requestOptions))),
             credentials: options.fetchOptions?.credentials,
-            headers: concreteHeaders({
+            headers: concreteProviderHeaders({
                 'content-type': 'application/json',
-                ...options.headers,
-                ...createAuthHeader(options.auth)
+                ...requestOptions.headers
             }),
             method: 'POST',
             signal: options.signal
@@ -390,16 +290,19 @@ export async function streamAiGoogleVertexModelRequest(request, options) {
 }
 export async function invokeGoogleVertexModel(request, options) {
     const fetchImpl = options.fetchImpl ?? fetch;
-    const model = buildGatewayAiModel(request, options.baseUrl);
+    const model = createGatewayProviderModel(request, 'google-vertex', options.baseUrl);
+    const requestOptions = createGatewayProviderOptions(request, 'google-vertex', options, {
+        maxTokens: 4096,
+        signal: options.signal
+    });
     let response;
     try {
-        response = await fetchImpl(googleEndpoint(options.baseUrl, model.id, 'generateContent'), {
-            body: JSON.stringify(buildRequestBody(request, options.baseUrl)),
+        response = await fetchImpl(googleVertexGatewayEndpoint(options.baseUrl, model.id, 'generateContent'), {
+            body: JSON.stringify(applyGatewayProviderPayload(request, 'google-vertex', buildGoogleVertexRequest(model, request.context, requestOptions))),
             credentials: options.fetchOptions?.credentials,
-            headers: concreteHeaders({
+            headers: concreteProviderHeaders({
                 'content-type': 'application/json',
-                ...options.headers,
-                ...createAuthHeader(options.auth)
+                ...requestOptions.headers
             }),
             method: 'POST',
             signal: options.signal

package/dist/embedded-agent/host-runtime-options.d.ts

@@ -1,11 +1,3 @@
-/**
- * Composition-root helper for both the production worker boot and the
- * in-process test transport. Everything the embedded runtime needs that
- * doesn't depend on host-specific infrastructure (credit-balance fetchers,
- * host prompt platform) is wired here so the two transports cannot
- * drift on lock-step state, the `onAgentStart` reset hook, and tool
- * composition. Host-specific bits stay at the call sites.
- */
 import type { Api, Model } from '../ai/types.js';
 import { type AgentConfiguration, type ChildAgentEventEmitter } from '../contracts/agent-api.js';
 import { AgentExecuteCodeMode } from '../contracts/agent-execute-code.js';

package/dist/embedded-agent/host-runtime-options.js

@@ -1,6 +1,14 @@
+/**
+ * Composition-root helper for both the production worker boot and the
+ * in-process test transport. Everything the embedded runtime needs that
+ * doesn't depend on host-specific infrastructure (credit-balance fetchers,
+ * host prompt platform) is wired here so the two transports cannot
+ * drift on lock-step state, the `onAgentStart` reset hook, and tool
+ * composition. Host-specific bits stay at the call sites.
+ */
+import { resolveModelReasoningLevel } from '../ai/models.js';
 import { AgentTaskControlAction } from '../contracts/agent-api.js';
 import { AgentExecuteCodeMode } from '../contracts/agent-execute-code.js';
-import { DEFAULT_THINKING_LEVEL } from '../core/defaults.js';
 import { createShortcutXLAgentSettingsStore } from './bootstrap.js';
 import { buildSubagentRuntimeOptions } from './host-tools/build-tool-list.js';
 import { createDefaultHostTools, createSessionToolInfoState } from './host-tools/index.js';
@@ -133,7 +141,7 @@ export function buildHostEmbeddedAgentOptions(input) {
         controlTask: (request, context) => controlEmbeddedTask(request, createSubagentRuntimeOptions(input, taskManager, sessionToolInfoState, context)),
         streamFn: input.streamFn,
         compactionCompleteFn: createStreamBackedCompactionComplete(input.streamFn),
-        thinkingLevel: DEFAULT_THINKING_LEVEL,
+        thinkingLevel: resolveModelReasoningLevel(input.model, undefined),
         ...(suggestionsSkillsPrompt?.trim() ? { suggestionsSkillsPrompt } : {}),
         ...(input.creditGate ? { creditGate: input.creditGate } : {}),
         ...(input.extraTools ? { tools: [...input.extraTools] } : {})

package/dist/embedded-agent/host-tools/invoke-llm/contract.d.ts

@@ -1,6 +1,6 @@
 /**
  * Bridge contract for `invoke_llm` — non-streaming LLM call dispatched to the
- * client-side `/api/llm/invoke` handler. Not a user-callable tool; the runtime
+ * client-side provider-native model worker. Not a user-callable tool; the runtime
  * uses this for internal one-shot LLM jobs (e.g. follow-up suggestions).
  */
 import type { AssistantMessage, ToolResultMessage, UserMessage } from '../../../ai/types.js';
@@ -23,7 +23,7 @@ export interface InvokeLlmParams {
     readonly messages: readonly InvokeLlmMessage[];
     readonly tools?: readonly InvokeLlmToolSchema[];
     readonly tool_choice?: InvokeLlmToolChoice;
-    readonly effort_level?: 'low' | 'medium' | 'high';
+    readonly effort_level?: 'low' | 'medium' | 'high' | 'xhigh' | 'max';
     readonly thinking_type?: 'on' | 'off';
     readonly max_tokens?: number;
 }

package/dist/embedded-agent/host-tools/invoke-llm/contract.js

@@ -1,6 +1,6 @@
 /**
  * Bridge contract for `invoke_llm` — non-streaming LLM call dispatched to the
- * client-side `/api/llm/invoke` handler. Not a user-callable tool; the runtime
+ * client-side provider-native model worker. Not a user-callable tool; the runtime
  * uses this for internal one-shot LLM jobs (e.g. follow-up suggestions).
  */
 import { Type } from '@sinclair/typebox';
@@ -107,7 +107,9 @@ function validateParams(value) {
     if (candidate.effort_level !== undefined &&
         candidate.effort_level !== 'low' &&
         candidate.effort_level !== 'medium' &&
-        candidate.effort_level !== 'high') {
+        candidate.effort_level !== 'high' &&
+        candidate.effort_level !== 'xhigh' &&
+        candidate.effort_level !== 'max') {
         return false;
     }
     if (candidate.thinking_type !== undefined &&

package/dist/embedded-agent/host-tools/sandbox-llm/contract.d.ts

@@ -7,15 +7,16 @@ import { type ShortcutXLModelStreamErrorDetails } from '../../../contracts/model
 export declare const SANDBOX_LLM_KIND: "sandboxLlm";
 export type SandboxLlmKind = typeof SANDBOX_LLM_KIND;
 export declare const SANDBOX_LLM_GEMINI_MODEL: "gemini-3-flash-preview";
-export declare const SANDBOX_LLM_GEMINI_LITE_MODEL: "gemini-3.5-flash-lite";
-export declare const SANDBOX_LLM_GEMINI_36_MODEL: "gemini-3.6-flash";
 export declare const SANDBOX_LLM_HAIKU_MODEL: "claude-haiku-4-5-20251001";
-export declare const SANDBOX_LLM_MAX_FILE_PATHS = 20;
+export declare const SANDBOX_LLM_MAX_FILE_PATHS = 10;
+export declare const SANDBOX_LLM_MAX_COMBINED_FILE_SIZE_MIB = 45;
+export declare const SANDBOX_LLM_MAX_COMBINED_FILE_BYTES: number;
 export declare const SANDBOX_LLM_OUTPUT_TOKENS = 32768;
 export declare const SANDBOX_LLM_TIMEOUT_SECONDS = 300;
 export declare const SANDBOX_LLM_PDF_PAGE_LIMIT = 100;
 export declare const SANDBOX_LLM_PDF_SIZE_LIMIT_MB = 10;
 export declare const SANDBOX_UPLOADS_PATH = "/workspace";
+export declare const SANDBOX_LLM_INPUT_TOO_LARGE_MESSAGE = "Sandbox LLM input limits exceeded: maximum 10 files and 45 MiB combined.";
 export declare const SANDBOX_LLM_IMAGE_EXTENSIONS: {
     readonly '.png': "image/png";
     readonly '.jpg': "image/jpeg";
@@ -36,12 +37,16 @@ export interface SandboxLlmDetails {
     readonly body?: string;
     /** Renderer flips to error styling when true. */
     readonly error?: boolean;
+    /** Number of requested files when a local input limit rejects the request. */
+    readonly file_count?: number;
     /** Structured model/helper failure details, when available. */
     readonly llm_error?: ShortcutXLModelStreamErrorDetails;
     /** Sandbox path the analysis was saved to (success path). */
     readonly save_path?: string;
     /** Resolved model id actually invoked. */
     readonly model?: string;
+    /** Raw bytes read when a local input limit rejects the request. */
+    readonly total_bytes?: number;
 }
 export type SandboxLlmRequest = HostToolRequest & {
     readonly kind: SandboxLlmKind;

package/dist/embedded-agent/host-tools/sandbox-llm/contract.js

@@ -10,15 +10,16 @@ import { withAbort } from '../../../utils/abort.js';
 import { SANDBOX_LLM } from '../tool-names.js';
 export const SANDBOX_LLM_KIND = 'sandboxLlm';
 export const SANDBOX_LLM_GEMINI_MODEL = SHORTCUT_MODEL_ID.Gemini3Flash;
-export const SANDBOX_LLM_GEMINI_LITE_MODEL = SHORTCUT_MODEL_ID.Gemini35FlashLite;
-export const SANDBOX_LLM_GEMINI_36_MODEL = SHORTCUT_MODEL_ID.Gemini36Flash;
 export const SANDBOX_LLM_HAIKU_MODEL = 'claude-haiku-4-5-20251001';
-export const SANDBOX_LLM_MAX_FILE_PATHS = 20;
+export const SANDBOX_LLM_MAX_FILE_PATHS = 10;
+export const SANDBOX_LLM_MAX_COMBINED_FILE_SIZE_MIB = 45;
+export const SANDBOX_LLM_MAX_COMBINED_FILE_BYTES = SANDBOX_LLM_MAX_COMBINED_FILE_SIZE_MIB * 1024 * 1024;
 export const SANDBOX_LLM_OUTPUT_TOKENS = 32_768;
 export const SANDBOX_LLM_TIMEOUT_SECONDS = 300;
 export const SANDBOX_LLM_PDF_PAGE_LIMIT = 100;
 export const SANDBOX_LLM_PDF_SIZE_LIMIT_MB = 10;
 export const SANDBOX_UPLOADS_PATH = '/workspace';
+export const SANDBOX_LLM_INPUT_TOO_LARGE_MESSAGE = `Sandbox LLM input limits exceeded: maximum ${SANDBOX_LLM_MAX_FILE_PATHS} files and ${SANDBOX_LLM_MAX_COMBINED_FILE_SIZE_MIB} MiB combined.`;
 export const SANDBOX_LLM_IMAGE_EXTENSIONS = {
     '.png': 'image/png',
     '.jpg': 'image/jpeg',
@@ -55,9 +56,11 @@ export const SANDBOX_LLM_TEXT_EXTENSIONS = new Set([
 const sandboxLlmDetailsSchema = Type.Object({
     body: Type.Optional(Type.String()),
     error: Type.Optional(Type.Boolean()),
+    file_count: Type.Optional(Type.Integer({ minimum: 0 })),
     llm_error: Type.Optional(Type.Unknown()),
     save_path: Type.Optional(Type.String()),
-    model: Type.Optional(Type.String())
+    model: Type.Optional(Type.String()),
+    total_bytes: Type.Optional(Type.Integer({ minimum: 0 }))
 });
 const sandboxLlmDetailsHelpers = createResultDetailsSchemaHelpers(sandboxLlmDetailsSchema);
 const LLM_ERROR_OBSERVABILITY_FIELDS = {
@@ -97,8 +100,6 @@ Also supports: .pdf
 
 MODEL SELECTION:
 - "${SANDBOX_LLM_GEMINI_MODEL}" (default): Best for careful extraction, tables, and especially images
-- "${SANDBOX_LLM_GEMINI_LITE_MODEL}": Fast, cost-efficient document parsing and structured extraction
-- "${SANDBOX_LLM_GEMINI_36_MODEL}": Complex visual and multimodal reasoning
 - "${SANDBOX_LLM_HAIKU_MODEL}": Quick scanning, structure discovery, fallback to ${SANDBOX_LLM_GEMINI_MODEL} if it doesn't work.
 
 FOR EXTRACTION TASKS:
@@ -109,12 +110,13 @@ Output filename is always saved to the file's stem folder:
 - For example, blah/report.pdf -> blah/report/analysis.md
 - uploads/report/analysis.pdf -> uploads/report/analysis/final_analysis.md
 
-⚠️ PDF SIZE LIMIT: Files cannot exceed ${SANDBOX_LLM_PDF_SIZE_LIMIT_MB} MB (exceeding causes LLM failures). Slice the PDF into smaller chunks or downsample to 150 dpi (use ghostscript for PDFs).
+⚠️ INPUT LIMITS: Each request accepts at most ${SANDBOX_LLM_MAX_FILE_PATHS} files and ${SANDBOX_LLM_MAX_COMBINED_FILE_SIZE_MIB} MiB of raw file data combined.
+⚠️ PDF SIZE LIMIT: Each PDF cannot exceed ${SANDBOX_LLM_PDF_SIZE_LIMIT_MB} MiB (exceeding causes LLM failures). Slice the PDF into smaller chunks or downsample to 150 dpi (use ghostscript for PDFs).
 ⚠️ PDF PAGE LIMIT: Claude Haiku has a ${SANDBOX_LLM_PDF_PAGE_LIMIT} page HARD LIMIT (exceeding causes instant failure).`;
 export const SANDBOX_LLM_DESCRIPTION_DESCRIPTION = "Brief human-readable description of what this analysis will do (e.g., 'Extracting sales data from Q4 report')";
 export const SANDBOX_LLM_FILE_PATHS_DESCRIPTION = `List of file paths in sandbox (e.g., ['${SANDBOX_UPLOADS_PATH}/chart.png']). Omit if the query already contains all needed information.`;
 export const SANDBOX_LLM_QUERY_DESCRIPTION = "The question/instruction for the LLM about the content. Can also be a file path ending in .txt (e.g., '/uploads/report/query.txt') — the file contents will be used as the query. Useful for batch/parallel calls to programmatically write .txt query files with variables.";
-export const SANDBOX_LLM_MODEL_DESCRIPTION = `Model to use: '${SANDBOX_LLM_GEMINI_MODEL}' (default), '${SANDBOX_LLM_GEMINI_LITE_MODEL}' (fast, cost-efficient extraction), '${SANDBOX_LLM_GEMINI_36_MODEL}' (complex visual reasoning), or '${SANDBOX_LLM_HAIKU_MODEL}' (fallback/verifier)`;
+export const SANDBOX_LLM_MODEL_DESCRIPTION = `Model to use: '${SANDBOX_LLM_GEMINI_MODEL}' (default) or '${SANDBOX_LLM_HAIKU_MODEL}' (fallback/verifier)`;
 export const SANDBOX_LLM_OUTPUT_FILENAME_DESCRIPTION = "Optional output filename (e.g., 'analysis.md'). If empty, auto-generates a unique filename.";
 const sandboxLlmLlmSchema = Type.Object({
     description: Type.String({
@@ -205,6 +207,8 @@ function getResultObservabilityFields(details) {
     const metadata = {
         ...(details.model !== undefined ? { model: details.model } : {}),
         ...(details.error !== undefined ? { error: details.error } : {}),
+        ...(details.file_count !== undefined ? { file_count: details.file_count } : {}),
+        ...(details.total_bytes !== undefined ? { total_bytes: details.total_bytes } : {}),
         ...getLlmErrorObservabilityFields(details.llm_error)
     };
     return Object.keys(metadata).length > 0 ? metadata : undefined;

package/dist/embedded-agent/host-tools/sandbox-llm/index.d.ts

@@ -1,3 +1,3 @@
 /** Public entrypoint for the `sandbox_llm` host tool. */
-export { SANDBOX_LLM_DESCRIPTION_DESCRIPTION, SANDBOX_LLM_FILE_PATHS_DESCRIPTION, SANDBOX_LLM_GEMINI_36_MODEL, SANDBOX_LLM_GEMINI_LITE_MODEL, SANDBOX_LLM_GEMINI_MODEL, SANDBOX_LLM_HAIKU_MODEL, SANDBOX_LLM_IMAGE_EXTENSIONS, SANDBOX_LLM_KIND, SANDBOX_LLM_MAX_FILE_PATHS, SANDBOX_LLM_MODEL_DESCRIPTION, SANDBOX_LLM_OUTPUT_FILENAME_DESCRIPTION, SANDBOX_LLM_OUTPUT_TOKENS, SANDBOX_LLM_PDF_PAGE_LIMIT, SANDBOX_LLM_PDF_SIZE_LIMIT_MB, SANDBOX_LLM_QUERY_DESCRIPTION, SANDBOX_LLM_SYSTEM_PROMPT, SANDBOX_LLM_TEXT_EXTENSIONS, SANDBOX_LLM_TIMEOUT_SECONDS, SANDBOX_LLM_TOOL_DESCRIPTION, SANDBOX_LLM_TOOL_SUMMARY, SANDBOX_UPLOADS_PATH, sandboxLlmContract, type SandboxLlmDetails, type SandboxLlmKind, type SandboxLlmParams, type SandboxLlmRequest } from './contract.js';
+export { SANDBOX_LLM_DESCRIPTION_DESCRIPTION, SANDBOX_LLM_FILE_PATHS_DESCRIPTION, SANDBOX_LLM_GEMINI_MODEL, SANDBOX_LLM_HAIKU_MODEL, SANDBOX_LLM_IMAGE_EXTENSIONS, SANDBOX_LLM_INPUT_TOO_LARGE_MESSAGE, SANDBOX_LLM_KIND, SANDBOX_LLM_MAX_COMBINED_FILE_BYTES, SANDBOX_LLM_MAX_COMBINED_FILE_SIZE_MIB, SANDBOX_LLM_MAX_FILE_PATHS, SANDBOX_LLM_MODEL_DESCRIPTION, SANDBOX_LLM_OUTPUT_FILENAME_DESCRIPTION, SANDBOX_LLM_OUTPUT_TOKENS, SANDBOX_LLM_PDF_PAGE_LIMIT, SANDBOX_LLM_PDF_SIZE_LIMIT_MB, SANDBOX_LLM_QUERY_DESCRIPTION, SANDBOX_LLM_SYSTEM_PROMPT, SANDBOX_LLM_TEXT_EXTENSIONS, SANDBOX_LLM_TIMEOUT_SECONDS, SANDBOX_LLM_TOOL_DESCRIPTION, SANDBOX_LLM_TOOL_SUMMARY, SANDBOX_UPLOADS_PATH, sandboxLlmContract, type SandboxLlmDetails, type SandboxLlmKind, type SandboxLlmParams, type SandboxLlmRequest } from './contract.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/embedded-agent/host-tools/sandbox-llm/index.js

@@ -1,3 +1,3 @@
 /** Public entrypoint for the `sandbox_llm` host tool. */
-export { SANDBOX_LLM_DESCRIPTION_DESCRIPTION, SANDBOX_LLM_FILE_PATHS_DESCRIPTION, SANDBOX_LLM_GEMINI_36_MODEL, SANDBOX_LLM_GEMINI_LITE_MODEL, SANDBOX_LLM_GEMINI_MODEL, SANDBOX_LLM_HAIKU_MODEL, SANDBOX_LLM_IMAGE_EXTENSIONS, SANDBOX_LLM_KIND, SANDBOX_LLM_MAX_FILE_PATHS, SANDBOX_LLM_MODEL_DESCRIPTION, SANDBOX_LLM_OUTPUT_FILENAME_DESCRIPTION, SANDBOX_LLM_OUTPUT_TOKENS, SANDBOX_LLM_PDF_PAGE_LIMIT, SANDBOX_LLM_PDF_SIZE_LIMIT_MB, SANDBOX_LLM_QUERY_DESCRIPTION, SANDBOX_LLM_SYSTEM_PROMPT, SANDBOX_LLM_TEXT_EXTENSIONS, SANDBOX_LLM_TIMEOUT_SECONDS, SANDBOX_LLM_TOOL_DESCRIPTION, SANDBOX_LLM_TOOL_SUMMARY, SANDBOX_UPLOADS_PATH, sandboxLlmContract } from './contract.js';
+export { SANDBOX_LLM_DESCRIPTION_DESCRIPTION, SANDBOX_LLM_FILE_PATHS_DESCRIPTION, SANDBOX_LLM_GEMINI_MODEL, SANDBOX_LLM_HAIKU_MODEL, SANDBOX_LLM_IMAGE_EXTENSIONS, SANDBOX_LLM_INPUT_TOO_LARGE_MESSAGE, SANDBOX_LLM_KIND, SANDBOX_LLM_MAX_COMBINED_FILE_BYTES, SANDBOX_LLM_MAX_COMBINED_FILE_SIZE_MIB, SANDBOX_LLM_MAX_FILE_PATHS, SANDBOX_LLM_MODEL_DESCRIPTION, SANDBOX_LLM_OUTPUT_FILENAME_DESCRIPTION, SANDBOX_LLM_OUTPUT_TOKENS, SANDBOX_LLM_PDF_PAGE_LIMIT, SANDBOX_LLM_PDF_SIZE_LIMIT_MB, SANDBOX_LLM_QUERY_DESCRIPTION, SANDBOX_LLM_SYSTEM_PROMPT, SANDBOX_LLM_TEXT_EXTENSIONS, SANDBOX_LLM_TIMEOUT_SECONDS, SANDBOX_LLM_TOOL_DESCRIPTION, SANDBOX_LLM_TOOL_SUMMARY, SANDBOX_UPLOADS_PATH, sandboxLlmContract } from './contract.js';
 //# sourceMappingURL=index.js.map

package/dist/embedded-agent/host-tools/shared/index.d.ts

@@ -1,5 +1,5 @@
 export { base64ToBytes, bytesToBase64, bytesToUtf8 } from './base64.js';
-export { LlmStreamError, readLlmStreamText, type LlmStreamRequest } from './llm-stream.js';
+export { LlmStreamError } from './llm-stream-error.js';
 export { MultimodalContentBlockFormat, buildAnthropicBlock, buildContentBlock, buildGeminiBlock } from './multimodal.js';
 export type { MultimodalBlock, MultimodalBlockInput, MultimodalBlockResult } from './multimodal.js';
 export { baseName, dirName, getFileDataDir, lowerExt, stem } from './sandbox-paths.js';

package/dist/embedded-agent/host-tools/shared/index.js

@@ -1,5 +1,5 @@
 export { base64ToBytes, bytesToBase64, bytesToUtf8 } from './base64.js';
-export { LlmStreamError, readLlmStreamText } from './llm-stream.js';
+export { LlmStreamError } from './llm-stream-error.js';
 export { MultimodalContentBlockFormat, buildAnthropicBlock, buildContentBlock, buildGeminiBlock } from './multimodal.js';
 export { baseName, dirName, getFileDataDir, lowerExt, stem } from './sandbox-paths.js';
 //# sourceMappingURL=index.js.map

package/dist/embedded-agent/host-tools/shared/llm-stream-error.d.ts

@@ -0,0 +1,7 @@
+import type { ShortcutXLModelStreamErrorDetails } from '../../../contracts/model-stream.js';
+/** Structured provider-native model-stream failure surfaced to helper tools. */
+export declare class LlmStreamError extends Error {
+    readonly details?: ShortcutXLModelStreamErrorDetails | undefined;
+    constructor(message: string, details?: ShortcutXLModelStreamErrorDetails | undefined);
+}
+//# sourceMappingURL=llm-stream-error.d.ts.map

package/dist/embedded-agent/host-tools/shared/llm-stream-error.js

@@ -0,0 +1,10 @@
+/** Structured provider-native model-stream failure surfaced to helper tools. */
+export class LlmStreamError extends Error {
+    details;
+    constructor(message, details) {
+        super(message);
+        this.details = details;
+        this.name = 'LlmStreamError';
+    }
+}
+//# sourceMappingURL=llm-stream-error.js.map

package/dist/embedded-agent/host-tools/take-screenshot/contract.d.ts

@@ -8,8 +8,6 @@ import { type HostToolRequest, type ToolContract, type ToolLlmOptionsBase } from
 export declare const TAKE_SCREENSHOT_KIND: "takeScreenshot";
 export type TakeScreenshotKind = typeof TAKE_SCREENSHOT_KIND;
 export declare const TAKE_SCREENSHOT_GEMINI_MODEL: "gemini-3-flash-preview";
-export declare const TAKE_SCREENSHOT_GEMINI_LITE_MODEL: "gemini-3.5-flash-lite";
-export declare const TAKE_SCREENSHOT_GEMINI_36_MODEL: "gemini-3.6-flash";
 export declare const TAKE_SCREENSHOT_HAIKU_MODEL: "claude-haiku-4-5-20251001";
 export declare const TAKE_SCREENSHOT_MIN_RANGES = 1;
 export declare const TAKE_SCREENSHOT_MAX_RANGES = 10;

package/dist/embedded-agent/host-tools/take-screenshot/contract.js

@@ -12,8 +12,6 @@ import { withAbort } from '../../../utils/abort.js';
 import { TAKE_SCREENSHOT } from '../tool-names.js';
 export const TAKE_SCREENSHOT_KIND = 'takeScreenshot';
 export const TAKE_SCREENSHOT_GEMINI_MODEL = SHORTCUT_MODEL_ID.Gemini3Flash;
-export const TAKE_SCREENSHOT_GEMINI_LITE_MODEL = SHORTCUT_MODEL_ID.Gemini35FlashLite;
-export const TAKE_SCREENSHOT_GEMINI_36_MODEL = SHORTCUT_MODEL_ID.Gemini36Flash;
 export const TAKE_SCREENSHOT_HAIKU_MODEL = 'claude-haiku-4-5-20251001';
 export const TAKE_SCREENSHOT_MIN_RANGES = 1;
 export const TAKE_SCREENSHOT_MAX_RANGES = 10;
@@ -25,7 +23,7 @@ export const TAKE_SCREENSHOT_RANGES_DESCRIPTION = 'List of sheet:range pairs to
 export const TAKE_SCREENSHOT_SHEET_DESCRIPTION = 'Sheet name to capture.';
 export const TAKE_SCREENSHOT_CELL_RANGE_DESCRIPTION = 'Range to capture, e.g. "A1:G10".';
 export const TAKE_SCREENSHOT_QUERY_DESCRIPTION = 'What to analyze about the screenshots.';
-export const TAKE_SCREENSHOT_MODEL_DESCRIPTION = `Vision model to use for analysis. Default: '${TAKE_SCREENSHOT_GEMINI_MODEL}'. Use '${TAKE_SCREENSHOT_GEMINI_LITE_MODEL}' for fast, cost-efficient extraction or '${TAKE_SCREENSHOT_GEMINI_36_MODEL}' for complex visual reasoning.`;
+export const TAKE_SCREENSHOT_MODEL_DESCRIPTION = `Vision model to use for analysis. Default: '${TAKE_SCREENSHOT_GEMINI_MODEL}'.`;
 const screenshotRangeSchema = Type.Object({
     sheet: Type.String({ description: TAKE_SCREENSHOT_SHEET_DESCRIPTION }),
     cell_range: Type.String({ description: TAKE_SCREENSHOT_CELL_RANGE_DESCRIPTION })

package/dist/embedded-agent/host-tools/take-screenshot/index.d.ts

@@ -1,5 +1,5 @@
 /**
  * Public entrypoint for the `take_screenshot` host tool.
  */
-export { TAKE_SCREENSHOT_CELL_RANGE_DESCRIPTION, TAKE_SCREENSHOT_GEMINI_36_MODEL, TAKE_SCREENSHOT_GEMINI_LITE_MODEL, TAKE_SCREENSHOT_GEMINI_MODEL, TAKE_SCREENSHOT_HAIKU_MODEL, TAKE_SCREENSHOT_KIND, TAKE_SCREENSHOT_MAX_RANGES, TAKE_SCREENSHOT_MIN_RANGES, TAKE_SCREENSHOT_MODEL_DESCRIPTION, TAKE_SCREENSHOT_OUTPUT_TOKENS, TAKE_SCREENSHOT_QUERY_DESCRIPTION, TAKE_SCREENSHOT_RANGES_DESCRIPTION, TAKE_SCREENSHOT_SHEET_DESCRIPTION, TAKE_SCREENSHOT_SYSTEM_PROMPT, TAKE_SCREENSHOT_TIMEOUT_SECONDS, TAKE_SCREENSHOT_TOOL_DESCRIPTION, takeScreenshotContract, type ScreenshotRange, type TakeScreenshotDetails, type TakeScreenshotKind, type TakeScreenshotParams, type TakeScreenshotRequest } from './contract.js';
+export { TAKE_SCREENSHOT_CELL_RANGE_DESCRIPTION, TAKE_SCREENSHOT_GEMINI_MODEL, TAKE_SCREENSHOT_HAIKU_MODEL, TAKE_SCREENSHOT_KIND, TAKE_SCREENSHOT_MAX_RANGES, TAKE_SCREENSHOT_MIN_RANGES, TAKE_SCREENSHOT_MODEL_DESCRIPTION, TAKE_SCREENSHOT_OUTPUT_TOKENS, TAKE_SCREENSHOT_QUERY_DESCRIPTION, TAKE_SCREENSHOT_RANGES_DESCRIPTION, TAKE_SCREENSHOT_SHEET_DESCRIPTION, TAKE_SCREENSHOT_SYSTEM_PROMPT, TAKE_SCREENSHOT_TIMEOUT_SECONDS, TAKE_SCREENSHOT_TOOL_DESCRIPTION, takeScreenshotContract, type ScreenshotRange, type TakeScreenshotDetails, type TakeScreenshotKind, type TakeScreenshotParams, type TakeScreenshotRequest } from './contract.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/embedded-agent/host-tools/take-screenshot/index.js

@@ -1,5 +1,5 @@
 /**
  * Public entrypoint for the `take_screenshot` host tool.
  */
-export { TAKE_SCREENSHOT_CELL_RANGE_DESCRIPTION, TAKE_SCREENSHOT_GEMINI_36_MODEL, TAKE_SCREENSHOT_GEMINI_LITE_MODEL, TAKE_SCREENSHOT_GEMINI_MODEL, TAKE_SCREENSHOT_HAIKU_MODEL, TAKE_SCREENSHOT_KIND, TAKE_SCREENSHOT_MAX_RANGES, TAKE_SCREENSHOT_MIN_RANGES, TAKE_SCREENSHOT_MODEL_DESCRIPTION, TAKE_SCREENSHOT_OUTPUT_TOKENS, TAKE_SCREENSHOT_QUERY_DESCRIPTION, TAKE_SCREENSHOT_RANGES_DESCRIPTION, TAKE_SCREENSHOT_SHEET_DESCRIPTION, TAKE_SCREENSHOT_SYSTEM_PROMPT, TAKE_SCREENSHOT_TIMEOUT_SECONDS, TAKE_SCREENSHOT_TOOL_DESCRIPTION, takeScreenshotContract } from './contract.js';
+export { TAKE_SCREENSHOT_CELL_RANGE_DESCRIPTION, TAKE_SCREENSHOT_GEMINI_MODEL, TAKE_SCREENSHOT_HAIKU_MODEL, TAKE_SCREENSHOT_KIND, TAKE_SCREENSHOT_MAX_RANGES, TAKE_SCREENSHOT_MIN_RANGES, TAKE_SCREENSHOT_MODEL_DESCRIPTION, TAKE_SCREENSHOT_OUTPUT_TOKENS, TAKE_SCREENSHOT_QUERY_DESCRIPTION, TAKE_SCREENSHOT_RANGES_DESCRIPTION, TAKE_SCREENSHOT_SHEET_DESCRIPTION, TAKE_SCREENSHOT_SYSTEM_PROMPT, TAKE_SCREENSHOT_TIMEOUT_SECONDS, TAKE_SCREENSHOT_TOOL_DESCRIPTION, takeScreenshotContract } from './contract.js';
 //# sourceMappingURL=index.js.map

package/dist/embedded-agent/host-tools/task/agents/clone.js

@@ -5,7 +5,7 @@
  * can spawn leaf subagents, but cannot recursively spawn another clone.
  */
 import { RUNTIME_THINKING_LEVEL } from '../../../../contracts/agent-settings.js';
-import { AGENT_PHILOSOPHY, COMMUNICATION, DATA_FORMATTING, TASK_EXECUTION, WRITING_CODE } from '../../../../core/prompts/agent-guidelines.js';
+import { COMMUNICATION, DATA_FORMATTING, TASK_EXECUTION, WRITING_CODE } from '../../../../core/prompts/agent-guidelines.js';
 import { buildSubagentApiReferenceSection } from '../../../../utils/api-reference.js';
 import { buildActionModeMainSection } from '../../../prompt/modes/action.js';
 import { API_HIERARCHY_SECTION, SANDBOX_FILESYSTEM, getPlatformIdentity } from '../../../prompt/sections.js';
@@ -33,7 +33,6 @@ function buildSystemPrompt(ctx) {
     const apiReference = buildSubagentApiReferenceSection(ctx.platform, ctx.spreadsheetReadOnly);
     return [
         buildActionModeMainSection(platformIdentity, currentDate),
-        AGENT_PHILOSOPHY,
         TASK_EXECUTION,
         COMMUNICATION,
         SANDBOX_FILESYSTEM,

package/dist/embedded-agent/host-tools/task/contract.js

@@ -126,7 +126,7 @@ function buildTaskLlmSchema(allowedSubagents) {
         run_in_background: Type.Optional(Type.Boolean({ description: TASK_RUN_IN_BACKGROUND_DESCRIPTION }))
     }, { additionalProperties: false });
 }
-/** Task spawns never require host approval — the parent already cleared the gate. */
+/** Task spawns do not require host approval because children inherit the parent boundary. */
 function evaluateApproval() {
     return { required: false };
 }