@opengeni/runtime 4.0.1 → 4.0.3-canary.36199476632001

package/dist/sandbox/browser-control-client.d.ts

@@ -1,4 +1,5 @@
-import { BROWSER_PROFILE_ARTIFACT_FORMAT, type BrowserActionCommand as BrowserActionCommandValue, type BrowserActionReceipt as BrowserActionReceiptValue, type BrowserClipboard as BrowserClipboardValue, type BrowserDiagnosticBatch as BrowserDiagnosticBatchValue, type BrowserDownload as BrowserDownloadValue, type BrowserDownloadExportReceipt as BrowserDownloadExportReceiptValue, type BrowserDownloadExportRequest as BrowserDownloadExportRequestValue, type BrowserDownloadListResponse as BrowserDownloadListResponseValue, type BrowserExternalAuthCommand as BrowserExternalAuthCommandValue, type BrowserExternalAuthResult as BrowserExternalAuthResultValue, type BrowserDiagnosticKind, type BrowserObservation as BrowserObservationValue, type BrowserProtectedAuthFillCommand as BrowserProtectedAuthFillCommandValue, type BrowserProtectedAuthFillReceipt as BrowserProtectedAuthFillReceiptValue, type BrowserRevisionMaterialization as BrowserRevisionMaterializationValue, type BrowserTarget as BrowserTargetValue, type BrowserWorkspaceFileStageRequest as BrowserWorkspaceFileStageRequestValue, type BrowserWorkspaceFileStageResponse as BrowserWorkspaceFileStageResponseValue, type ComputerActionCommand as ComputerActionCommandValue, type ComputerActionReceipt as ComputerActionReceiptValue, type ComputerClipboard as ComputerClipboardValue, type ComputerObservation as ComputerObservationValue, type ComputerSessionCapabilities as ComputerSessionCapabilitiesValue, type ComputerTarget as ComputerTargetValue, type InteractionError as InteractionErrorValue, type NetworkRouteConsistency as NetworkRouteConsistencyValue } from "@opengeni/contracts";
+import { type BrowserFrameMetadata } from "@opengeni/sdk";
+import { BROWSER_PROFILE_ARTIFACT_FORMAT, type BrowserActionCommand as BrowserActionCommandValue, type BrowserActionReceipt as BrowserActionReceiptValue, type BrowserClipboard as BrowserClipboardValue, type BrowserDiagnosticBatch as BrowserDiagnosticBatchValue, type BrowserDownload as BrowserDownloadValue, type BrowserDownloadExportReceipt as BrowserDownloadExportReceiptValue, type BrowserDownloadExportRequest as BrowserDownloadExportRequestValue, type BrowserDownloadListResponse as BrowserDownloadListResponseValue, type BrowserDomReadRequest as BrowserDomReadRequestValue, type BrowserDomReadResponse as BrowserDomReadResponseValue, type BrowserExternalAuthCommand as BrowserExternalAuthCommandValue, type BrowserExternalAuthResult as BrowserExternalAuthResultValue, type BrowserDiagnosticKind, type BrowserObservation as BrowserObservationValue, type BrowserProtectedAuthFillCommand as BrowserProtectedAuthFillCommandValue, type BrowserProtectedAuthFillReceipt as BrowserProtectedAuthFillReceiptValue, type BrowserRevisionMaterialization as BrowserRevisionMaterializationValue, type BrowserTarget as BrowserTargetValue, type BrowserTargetState as BrowserTargetStateValue, type BrowserWorkspaceFileStageRequest as BrowserWorkspaceFileStageRequestValue, type BrowserWorkspaceFileStageResponse as BrowserWorkspaceFileStageResponseValue, type ComputerActionCommand as ComputerActionCommandValue, type ComputerActionReceipt as ComputerActionReceiptValue, type ComputerClipboard as ComputerClipboardValue, type ComputerObservation as ComputerObservationValue, type ComputerSessionCapabilities as ComputerSessionCapabilitiesValue, type ComputerTarget as ComputerTargetValue, type InteractionError as InteractionErrorValue, type NetworkRouteConsistency as NetworkRouteConsistencyValue } from "@opengeni/contracts";
 import type { BrowserControlEnsureRequest, BrowserControlEnsureResponse, BrowserFramesOpenRequest, ComputerFramesOpenRequest, StreamChannel } from "@opengeni/agent-proto";
 import { type EnsureBrowserControlServerResult } from "./browser-control-server.js";
 import { type ExposedPortEndpoint } from "./stream-port.js";
@@ -70,6 +71,22 @@ export type ComputerControlFrame = {
     metadataHeader: string;
     metadata: ComputerControlFrameMetadata;
 };
+export type BrowserControlFrame = {
+    data: Uint8Array;
+    mediaType: "image/jpeg" | "image/png";
+    metadataHeader: string;
+    metadata: BrowserFrameMetadata;
+};
+export type BrowserControlScreenshotOptions = {
+    fullPage?: boolean;
+    format?: "jpeg" | "png";
+    quality?: number;
+};
+type ExpectedBrowserFrameEvidence = {
+    browserSessionId: string;
+    controllerGeneration: string;
+    targetId: string;
+};
 export type ComputerControlFrameMetadata = {
     frameId: string;
     computerSessionId: string;
@@ -378,6 +395,13 @@ export declare class BrowserControlClient {
         expectedFrameEvidence: ExpectedComputerFrameEvidence;
         timeoutMs?: number;
     }): Promise<ComputerControlFrame>;
+    requestBrowserBytesForSession(input: {
+        method: "GET";
+        path: string;
+        token: string;
+        expectedFrameEvidence: ExpectedBrowserFrameEvidence;
+        timeoutMs?: number;
+    }): Promise<BrowserControlFrame>;
     private requestJson;
     private requestBytes;
     private controllerPort;
@@ -400,6 +424,9 @@ export declare class BrowserControlSessionClient {
     selectTarget(targetId: string): Promise<BrowserObservationValue>;
     closeTarget(targetId: string): Promise<BrowserTargetValue[]>;
     observe(targetId: string): Promise<BrowserObservationValue>;
+    targetState(targetId: string): Promise<BrowserTargetStateValue>;
+    readDom(targetId: string, requestInput: BrowserDomReadRequestValue): Promise<BrowserDomReadResponseValue>;
+    capture(targetId: string, options?: BrowserControlScreenshotOptions): Promise<BrowserControlFrame>;
     readClipboard(): Promise<BrowserClipboardValue>;
     action(command: BrowserActionCommandValue): Promise<BrowserActionReceiptValue>;
     /** API-broker-only file authority path. Signed read URLs are materialized by

package/dist/sandbox/channel-a.d.ts

@@ -187,6 +187,25 @@ export type RepositoryDiscoveryResult = {
     complete: boolean;
     degradedReason: RepositoryDiscoveryDegradedReason | null;
 };
+/** Raw ripgrep stdout kept on the box before encoding. */
+export declare const CODE_SEARCH_RG_MAX_BYTES: number;
+export type CodeSearchRipgrepOutcome = {
+    /** False when ripgrep is not installed on the box. */
+    available: boolean;
+    stdout: string;
+    /** rg exit status (0 matches, 1 none, 2 error); null after a timeout. */
+    exitCode: number | null;
+    /** Output hit the byte cap and was cut at a line boundary. */
+    truncated: boolean;
+    timedOut: boolean;
+};
+/**
+ * Accept only the ripgrep arguments `code_search` needs. Flags that run a
+ * program (`--pre`, `-z`) or read other files, and paths that leave the
+ * workspace, are rejected before a command is built. `--no-config` stops a
+ * `RIPGREP_CONFIG_PATH` on the box from adding flags.
+ */
+export declare function validateCodeSearchRipgrepArgs(args: readonly string[], workspaceRoot?: string): string[];
 export declare class SandboxChannelAService {
     private readonly session;
     private readonly workspaceRoot;
@@ -287,6 +306,30 @@ export declare class SandboxChannelAService {
     detectReposDetailed(): Promise<RepositoryDiscoveryResult>;
     /** Detect repo roots within the workspace (for the Git.repos capability). */
     detectRepos(): Promise<string[]>;
+    /**
+     * Run ripgrep read-only in the workspace root for the Jev `code_search` tool.
+     * Arguments are checked against a closed allowlist (no `--pre`, `-z` or other
+     * flags that run programs), so the command is provably read-only and skips
+     * durable mutation admission. stdout is capped at `maxBytes`, gzip+base64
+     * framed so providers that drop newlines or retain limited output cannot
+     * corrupt it, and bounded by a wall-clock watchdog (not GNU `timeout`, which
+     * stock macOS lacks). Resolves `available: false` when `rg` is not installed,
+     * and throws when the box could not store the output whole (a full
+     * temporary directory) rather than returning a prefix as complete.
+     */
+    codeSearchRipgrep(args: readonly string[], options: {
+        timeoutMs: number;
+        maxBytes: number;
+        maxTransferBytes?: number;
+    }): Promise<CodeSearchRipgrepOutcome>;
+    /**
+     * Fetch one chunk of compressed ripgrep output kept on the box by a previous
+     * codeSearchRipgrep call. The last fetch deletes the file. Null when the
+     * file is gone.
+     */
+    private codeSearchRipgrepChunk;
+    /** Classify workspace-relative paths for `code_search` path filters. */
+    codeSearchPathKinds(paths: readonly string[]): Promise<Record<string, "file" | "directory" | "missing">>;
     /** Run a bounded command to physical completion and return buffered output.
      *  A deadline cancels the exact provider operation, proves its process group
      *  absent, and settles any retained-process admission before returning 124. */

package/dist/sandbox/index.d.ts

@@ -18,7 +18,7 @@ export { SandboxConfigError, SandboxExactResumeInstanceUnavailableError, Sandbox
 export { PROVIDER_REGISTRY, assertProviderRegistryInvariants, buildImmutableProviderImage, isProviderApiThrottleError, prepareProviderForTeardownAfterCapture, providerSupportsImmutableImageBuild, renewSandboxProviderExpiration, providerWorkspaceCapturePolicy, type ProviderRegistration, type ProviderConstructionContext, type ProviderExactResumeMode, type ProviderImmutableImageBuildInput, type ProviderImmutableImageBuildResult, type ProviderExpirationRenewalInput, type ProviderWorkspaceCapturePolicy, type ProviderWorkspaceCaptureTakeover, } from "./providers/index.js";
 export { classifyProviderSandboxFailure, isProviderSandboxGoneDuringRoutedOperation, isProviderSandboxNotFoundError, isProviderSandboxTransientError, type ProviderSandboxFailure, type ProviderSandboxFailureKind, } from "./provider-errors.js";
 export { WORKSPACE_ARCHIVE_DESCRIPTOR_VERSION, WorkspaceArchiveIntegrityError, captureVerifiedWorkspaceArchive, decodeNativeSnapshotRef, encodeNativeSnapshotRef, describeLegacyNativeSnapshotArchive, describeNativeSnapshotArchive, fingerprintSandboxWorkspace, inlineWorkspaceArchiveForRestore, parseWorkspaceArchiveDescriptor, readVerifiedWorkspaceArchive, verifyRestoredWorkspace, type NativeSnapshotDescriptor, type NativeSnapshotProvider, type NativeSnapshotRef, type TarWorkspaceArchiveDescriptor, type VerifiedWorkspaceArchive, type WorkspaceArchiveDescriptor, type WorkspaceArchiveIntegrityCode, type WorkspaceTreeFingerprint, } from "./workspace-archive.js";
-export { ensureModalRegistryImage, deleteModalCheckpointSnapshot, inspectModalSandboxLifecycle, isModalTaskExecStartDnsResolutionError, modalSessionMatchesCheckpointProviderBinding, modalSandboxAttributionEnvironment, modalSandboxAttributionTags, resolveModalCheckpointProviderBinding, resolveModalCheckpointProviderBindingForLiveSandbox, resolveModalCheckpointProviderBindingForSession, sweepModalOrphanSandboxes, tagModalSandbox, terminateModalSandboxById, type LiveModalSandboxLeaseAttribution, type ModalModuleLoader, type ModalCheckpointProviderBinding, type ModalOrphanSweepResult, type ModalOrphanSweepTermination, type ModalSandboxAttribution, type RevalidateModalOrphanTermination, } from "./providers/modal.js";
+export { ensureModalRegistryImage, deleteModalCheckpointSnapshot, inspectModalSandboxLifecycle, isModalTaskExecStartPreDispatchUnavailableError, modalSessionMatchesCheckpointProviderBinding, modalSandboxAttributionEnvironment, modalSandboxAttributionTags, resolveModalCheckpointProviderBinding, resolveModalCheckpointProviderBindingForLiveSandbox, resolveModalCheckpointProviderBindingForSession, sweepModalOrphanSandboxes, tagModalSandbox, terminateModalSandboxById, type LiveModalSandboxLeaseAttribution, type ModalModuleLoader, type ModalCheckpointProviderBinding, type ModalOrphanSweepResult, type ModalOrphanSweepTermination, type ModalSandboxAttribution, type RevalidateModalOrphanTermination, } from "./providers/modal.js";
 export { OpenSandboxClient, OpenSandboxSession, setOpenSandboxApplyDiff, type OpenSandboxApplyDiff, type OpenSandboxClientOptions, type OpenSandboxEditor, type OpenSandboxSessionState, } from "./providers/opensandbox-adapter.js";
 export { selectBackend, sdkBackendIdForSandboxBackend, sandboxBackendForSdkBackendId, backendSupportsOs, desktopCapableBackend, negotiateCapabilities, type NegotiationContext, } from "./select.js";
 export { STREAM_TOKEN_DEFAULT_TTL_SECONDS, mintStreamToken, verifyStreamToken, StreamTokenPayload, type MintStreamTokenInput, type StreamTokenPayloadType, } from "./stream-token.js";
@@ -28,13 +28,13 @@ export { createTurnToolCancellationController, TurnSandboxCommandCancelledError,
 export type { TurnSandboxCommandArgs, TurnSandboxCommandSession, TurnToolCancellationController, TurnToolCancellationFence, } from "./turn-tool-cancellation.js";
 export { TERMINAL_STREAM_PORT, TERMINAL_SERVER_TIMEOUT_MS, TerminalServerError, TerminalServerUnsupportedError, buildTerminalServerScript, ensureTerminalServer, tearDownTerminalServer, type EnsureTerminalServerOptions, type EnsureTerminalServerResult, } from "./terminal-server.js";
 export { BROWSER_CONTROL_PORT, BROWSER_CONTROL_SERVER_TIMEOUT_MS, BrowserControlServerError, BrowserControlServerUnsupportedError, buildBrowserControlServerScript, ensureBrowserControlServer, tearDownBrowserControlServer, type EnsureBrowserControlServerOptions, type EnsureBrowserControlServerResult, } from "./browser-control-server.js";
-export { BROWSER_CONTROL_ADMIN_TOKEN_FILE, BrowserControlClient, BrowserControlProtocolError, BrowserControlRequestError, BrowserControlSessionClient, BrowserControlTransportError, BrowserControlUnsupportedError, ComputerFrameEvidenceMismatchError, decodeComputerControlFrameMetadataHeader, provisionBrowserControlClient, validateComputerControlFrameEvidence, type BrowserControlPlacementSession, type BrowserStateDownloadGrant, type BrowserStateUploadGrant, type BrowserViewGrant, type CapturePlacementBrowserStateInput, type CreatePlacementBrowserSessionInput, type PlacementBrowserSession, type PlacementBrowserSessionReference, type PlacementBrowserNetworkRoute, type PlacementBrowserTransport, type PlacementBrowserStateCaptureReceipt, type ComputerControlFrame, type ComputerControlFrameMetadata, type ComputerFrameEvidenceMismatchReason, type ExpectedComputerFrameEvidence, type ProvisionBrowserControlClientInput, type ProvisionBrowserControlClientResult, type RestorePlacementBrowserStateInput, } from "./browser-control-client.js";
+export { BROWSER_CONTROL_ADMIN_TOKEN_FILE, BrowserControlClient, BrowserControlProtocolError, BrowserControlRequestError, BrowserControlSessionClient, BrowserControlTransportError, BrowserControlUnsupportedError, ComputerFrameEvidenceMismatchError, decodeComputerControlFrameMetadataHeader, provisionBrowserControlClient, validateComputerControlFrameEvidence, type BrowserControlPlacementSession, type BrowserControlScreenshotOptions, type BrowserControlFrame, type BrowserStateDownloadGrant, type BrowserStateUploadGrant, type BrowserViewGrant, type CapturePlacementBrowserStateInput, type CreatePlacementBrowserSessionInput, type PlacementBrowserSession, type PlacementBrowserSessionReference, type PlacementBrowserNetworkRoute, type PlacementBrowserTransport, type PlacementBrowserStateCaptureReceipt, type ComputerControlFrame, type ComputerControlFrameMetadata, type ComputerFrameEvidenceMismatchReason, type ExpectedComputerFrameEvidence, type ProvisionBrowserControlClientInput, type ProvisionBrowserControlClientResult, type RestorePlacementBrowserStateInput, } from "./browser-control-client.js";
 export { RunCredentialValidationError, normalizeRunCredentialsResolution, runCredentialRoot, runCredentialPointerFile, materializeRunCredentials, clearRunCredentials, clearRunCredentialsForAttempt, withRunCredentialEnvironment, withRunCredentialsSession, withRunCredentialsClient, type RunCredentialExpectedScope, type NormalizedRunCredentialMaterial, type RunCredentialCommandSession, type RunCredentialCommandRunner, type RunCredentialSessionReady, type MaterializeRunCredentialsOptions, } from "./run-credentials.js";
 export { CodemodeTokenPathError, codemodeTokenFileForSession, codemodeTokenFileFromEnvironment, withCodemodeTokenEnvironment, withCodemodeTokenSession, withCodemodeTokenClient, } from "./codemode-token.js";
 export { CODEMODE_TOKEN_TTL_SECONDS, mintSandboxCodemodeToken, type CodemodeAuthorityScope, type SandboxCodemodeAuthority, type MintedSandboxCodemodeToken, } from "./codemode-authority.js";
 export { exposeStreamPort, buildStreamUrl, exposedPortAllowsHostFetch, exposedPortEndpointFromUrl, isOpenSandboxLifecycleProxyPath, joinExposedPortPath, parseOpenSandboxSignedUriPath, redactOpenSandboxSignedUriPath, signedEndpointExpiresAtMs, signedEndpointNeedsRefresh, StreamPortUnavailableError, type ExposedPortEndpoint, type ExposeStreamPortInput, type ExposeStreamPortResult, type OpenSandboxSignedUriPath, } from "./stream-port.js";
 export { startRecording, stopRecording, inspectRecordingArtifact, uploadRecordingArtifact, deleteRecordingArtifacts, recordingStorageKey, contentTypeForCodec, extForCodec, RecordingUnavailableError, RecordingError, type RecordingCodec, type RecordingContentType, type StartRecordingInput, type RecordingProcess, type RecordingArtifactMetadata, } from "./recording.js";
-export { SandboxChannelAService, ChannelAPartialMutationError, ChannelAValidationError, ChannelAUnavailableError, ChannelAConflictError, ChannelAFileSystemRouteChangedError, ChannelANotFoundError, ChannelAUnsupportedError, stripExecBanner, parseExecBannerSessionId, parseExecBannerExitCode, isExecSessionLostBanner, assertSafeRelPath, parsePorcelainV2, parseNumstatZ, parseUnifiedPatch, REPOSITORY_DISCOVERY_LIMIT, type ChannelASession, type ChannelAExecArgs, type ChannelAExecResult, type ChannelAEmitter, type SandboxChannelAServiceOptions, type RepositoryDiscoveryDegradedReason, type RepositoryDiscoveryResult, type NumstatEntry, } from "./channel-a.js";
+export { SandboxChannelAService, ChannelAPartialMutationError, ChannelAValidationError, ChannelAUnavailableError, ChannelAConflictError, ChannelAFileSystemRouteChangedError, ChannelANotFoundError, ChannelAUnsupportedError, stripExecBanner, parseExecBannerSessionId, parseExecBannerExitCode, isExecSessionLostBanner, assertSafeRelPath, parsePorcelainV2, parseNumstatZ, parseUnifiedPatch, REPOSITORY_DISCOVERY_LIMIT, CODE_SEARCH_RG_MAX_BYTES, validateCodeSearchRipgrepArgs, type CodeSearchRipgrepOutcome, type ChannelASession, type ChannelAExecArgs, type ChannelAExecResult, type ChannelAEmitter, type SandboxChannelAServiceOptions, type RepositoryDiscoveryDegradedReason, type RepositoryDiscoveryResult, type NumstatEntry, } from "./channel-a.js";
 export { type ControlRpc, NatsControlRpc, SelfhostedControlError, agentErrorToControlError, subjectFor, offlineControlResponse, timeoutControlResponse, offlineAgentError, timeoutAgentError, drainingExhaustedError, payloadTooLargeMessage, drainingMessage, execDeadlineHint, NEVER_SENT_DETAIL_KEY, type NatsRequestConnection, type SelfhostedUnavailableReason, } from "./selfhosted/control-rpc.js";
 export type { SelfhostedOpObserver, SelfhostedOpObservation } from "./selfhosted/op-observer.js";
 export { selfhostedFaultClass, SELFHOSTED_INFRASTRUCTURE_FAULT_CLASSES, renderSelfhostedFault, FAULT_FIELD_WHAT_HAPPENED, FAULT_FIELD_WHICH_LAYER, FAULT_FIELD_WHAT_PRESERVED, FAULT_FIELD_WHAT_TO_TRY, } from "./selfhosted/fault-rendering.js";

package/dist/sandbox/index.js

@@ -13,6 +13,7 @@ import {
   BrowserControlUnsupportedError,
   CAPABILITY_DESCRIPTORS,
   CODEMODE_TOKEN_TTL_SECONDS,
+  CODE_SEARCH_RG_MAX_BYTES,
   ChannelAConflictError,
   ChannelAFileSystemRouteChangedError,
   ChannelANotFoundError,
@@ -158,7 +159,7 @@ import {
   inspectRecordingArtifact,
   isConnectedMachineAbsolutePath,
   isExecSessionLostBanner,
-  isModalTaskExecStartDnsResolutionError,
+  isModalTaskExecStartPreDispatchUnavailableError,
   isOpenSandboxLifecycleProxyPath,
   isProviderApiThrottleError,
   isProviderSandboxGoneDuringRoutedOperation,
@@ -258,6 +259,7 @@ import {
   timeoutAgentError,
   timeoutControlResponse,
   uploadRecordingArtifact,
+  validateCodeSearchRipgrepArgs,
   validateComputerControlFrameEvidence,
   verifyRestoredWorkspace,
   verifySandboxExecReadiness,
@@ -269,11 +271,11 @@ import {
   withRunCredentialsClient,
   withRunCredentialsSession,
   withoutSandboxProviderIdentity
-} from "../chunk-J5AR5XAJ.js";
-import "../chunk-SPT3PLMX.js";
+} from "../chunk-FAFZCSN7.js";
+import "../chunk-PRHMSM6C.js";
 import "../chunk-4A4XTKKX.js";
-import "../chunk-ZG2YH25H.js";
-import "../chunk-UNICCRW5.js";
+import "../chunk-JAYHA6BF.js";
+import "../chunk-LZIXCZUV.js";
 import "../chunk-PZ5AY32C.js";
 export {
   ActiveBackendUnresolvableError,
@@ -290,6 +292,7 @@ export {
   BrowserControlUnsupportedError,
   CAPABILITY_DESCRIPTORS,
   CODEMODE_TOKEN_TTL_SECONDS,
+  CODE_SEARCH_RG_MAX_BYTES,
   ChannelAConflictError,
   ChannelAFileSystemRouteChangedError,
   ChannelANotFoundError,
@@ -435,7 +438,7 @@ export {
   inspectRecordingArtifact,
   isConnectedMachineAbsolutePath,
   isExecSessionLostBanner,
-  isModalTaskExecStartDnsResolutionError,
+  isModalTaskExecStartPreDispatchUnavailableError,
   isOpenSandboxLifecycleProxyPath,
   isProviderApiThrottleError,
   isProviderSandboxGoneDuringRoutedOperation,
@@ -535,6 +538,7 @@ export {
   timeoutAgentError,
   timeoutControlResponse,
   uploadRecordingArtifact,
+  validateCodeSearchRipgrepArgs,
   validateComputerControlFrameEvidence,
   verifyRestoredWorkspace,
   verifySandboxExecReadiness,

package/dist/sandbox/providers/modal-command-router-wire.d.ts

@@ -1,3 +1,4 @@
+import { Client } from "@grpc/grpc-js";
 import protobuf from "protobufjs";
 import { ProviderCommandStartRejectedError } from "../provider-command-session.js";
 export declare const modalRouterWire: protobuf.Root;
@@ -7,6 +8,13 @@ export declare class ModalCommandStartRejectedError extends ProviderCommandStart
     readonly code: number;
     constructor(code: number, cause: unknown);
 }
+/** Only a local channel-readiness failure BEFORE TaskExecStart is issued can
+ * prove that start was never dispatched. RPC status/details, even if they look
+ * like a DNS resolver error, can originate from an accepting server. */
+export declare class ModalCommandStartPreDispatchUnavailableError extends Error {
+    private constructor();
+    static ensureReady(client: Client, signal?: AbortSignal): Promise<void>;
+}
 export type ModalRouterIdentity = {
     taskId: string;
     execId: string;

package/dist/sandbox/providers/modal.d.ts

@@ -27,13 +27,12 @@ export type RevalidateModalOrphanTermination = (candidate: ModalOrphanSweepTermi
 export declare function modalSandboxAttributionEnvironment(input: ModalSandboxAttribution): Record<string, string>;
 export declare function modalSandboxAttributionTags(input: ModalSandboxAttribution): Record<string, string>;
 /**
- * Modal exhausted its own retries before failing to resolve the exact command
- * router DNS name, so TaskExecStart never connected and replay is safe. Keep
- * this fail-closed across Agents SDK wrappers: every reachable structural leaf
- * must be that exact ClientError with no HTTP status metadata, and any incomplete
- * or mixed graph is rejected.
+ * Only the native client's pre-dispatch readiness gate proves Start was not
+ * sent. Server-controlled gRPC/SDK status and DNS-looking details prove nothing.
+ * Any retained or outcome-unknown routing boundary vetoes recovery even when
+ * it contains a genuine pre-dispatch error among its causes.
  */
-export declare function isModalTaskExecStartDnsResolutionError(error: unknown): boolean;
+export declare function isModalTaskExecStartPreDispatchUnavailableError(error: unknown): boolean;
 /**
  * Modal proves that the exact exec has already terminated with a typed
  * FAILED_PRECONDITION from its stdin-write RPC. Keep this deliberately

package/dist/workspace-tool-gateway.js

@@ -1,14 +1,14 @@
 import {
   prepareAgentTools
-} from "./chunk-J6J3VYFW.js";
-import "./chunk-J5AR5XAJ.js";
-import "./chunk-SPT3PLMX.js";
+} from "./chunk-P4V7JGQ2.js";
+import "./chunk-FAFZCSN7.js";
+import "./chunk-PRHMSM6C.js";
 import "./chunk-4A4XTKKX.js";
-import "./chunk-ZG2YH25H.js";
+import "./chunk-JAYHA6BF.js";
 import "./chunk-TE27J65Q.js";
 import "./chunk-UZ66CGC2.js";
 import "./chunk-YU4VRTR4.js";
-import "./chunk-UNICCRW5.js";
+import "./chunk-LZIXCZUV.js";
 import "./chunk-PZ5AY32C.js";
 export {
   prepareAgentTools as prepareWorkspaceToolGatewayTools

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@opengeni/runtime",
-  "version": "4.0.1",
+  "version": "4.0.3-canary.36199476632001",
   "license": "Apache-2.0",
   "repository": {
     "type": "git",
@@ -63,16 +63,16 @@
     "@openai/agents": "0.14.3",
     "@openai/agents-core": "0.14.3",
     "@openai/agents-extensions": "0.14.3",
-    "@opengeni/agent-proto": "^0.6.0",
-    "@opengeni/capabilities": "^0.3.4",
-    "@opengeni/codemode": "^0.6.1",
-    "@opengeni/codex": "^0.2.25",
-    "@opengeni/config": "^2.1.0",
-    "@opengeni/contracts": "^5.1.0",
-    "@opengeni/network": "^0.3.1",
-    "@opengeni/sdk": "^7.1.0",
-    "@opengeni/tool-gateway": "^0.1.12",
-    "@opengeni/xai-subscription": "^0.1.4",
+    "@opengeni/agent-proto": "^0.6.0-canary.36199476632001",
+    "@opengeni/capabilities": "^0.3.4-canary.36199476632001",
+    "@opengeni/codemode": "^0.6.2-canary.36199476632001",
+    "@opengeni/codex": "^0.2.26-canary.36199476632001",
+    "@opengeni/config": "^2.1.1-canary.36199476632001",
+    "@opengeni/contracts": "^5.1.1-canary.36199476632001",
+    "@opengeni/network": "^0.3.1-canary.36199476632001",
+    "@opengeni/sdk": "^7.1.1-canary.36199476632001",
+    "@opengeni/tool-gateway": "^0.1.13-canary.36199476632001",
+    "@opengeni/xai-subscription": "^0.1.4-canary.36199476632001",
     "debug": "4.4.3",
     "modal": "0.9.0",
     "openai": "6.47.0",

package/src/browser-action-json-schema.ts

@@ -0,0 +1,10 @@
+import type { AttemptToolJsonSchema } from "@opengeni/contracts";
+import { z } from "zod";
+
+/** Keep repeated browser action and locator variants once in the model schema. */
+export function browserActionInputJsonSchema(schema: z.ZodType): AttemptToolJsonSchema {
+  return z.toJSONSchema(schema, {
+    target: "draft-2020-12",
+    reused: "ref",
+  }) as AttemptToolJsonSchema;
+}

package/src/bundled_artifact_skills/opengeni-documents/SKILL.md

@@ -44,7 +44,10 @@ Skill was installed or explicitly checked out to the filesystem.
 2. Make the smallest coherent edit. One `opengeni__editable_artifact_apply`
    call is one atomic command batch. Use stable ids from inspection for existing
    objects and `openGeni.artifacts.ids.document(...)` for new objects in
-   CodeMode. A direct call must pass the inspected `headSequence` and
+   CodeMode. On a Connected Machine without the JS client, use the offline native
+   `codemode document-id <kind> <namespace>` helper described in `references/api.md`;
+   copy the exact namespace from the inspected summary, never invent it.
+   A direct call must pass the inspected `headSequence` and
    `stateHash`; CodeMode carries its last read head automatically.
 3. For one simple edit, call the artifact tools directly. For loops, several
    inspections, generated ids, or a multi-part batch, write auditable Bun code

package/src/bundled_artifact_skills/opengeni-documents/references/api.md

@@ -37,6 +37,26 @@ artifactId: document.id, inspectionReceiptId }]` with the completion evidence to
 If the document changes, inspect it again and use the new receipt. Include the
 returned artifact reference in the user handoff.
 
+## Native Connected Machine ID helper
+
+When the JavaScript client is unavailable, the native client can generate IDs
+offline. Inspect the document summary with the artifact tools first. Copy its
+exact decimal `idNamespace` string into `DOCUMENT_NAMESPACE`, then run:
+
+```sh
+"$OPENGENI_CODEMODE_NATIVE_CLIENT" codemode document-id paragraph "$DOCUMENT_NAMESPACE"
+```
+
+The result is JSON with an `id` field. Kinds are `paragraph`, `table`, `page-break`,
+`section`, `header`, `footer`, `comment`, and `tracked-change`. The helper uses the
+same ID format as `openGeni.artifacts.ids.document`; it needs no credentials or
+network and does not mutate an artifact. Keep uint64 namespaces as decimal strings,
+not JavaScript numbers. Direct apply calls still require the inspected
+`headSequence` and `stateHash`. Reuse inspected IDs for existing objects.
+If the installed native client lacks `document-id`, use the deployment-selected
+JavaScript helper when available; otherwise report the missing client capability
+rather than inventing IDs or claiming an edit succeeded.
+
 ## Authoring example
 
 ```js

package/src/bundled_default_skills/opengeni-client/SKILL.md

@@ -55,7 +55,10 @@ Embedded products may narrow bundled guidance with `bundledSkillIds`.
   missing access is configured. Read
   [Discovery and autonomy](references/discovery-and-autonomy.md) for that workflow.
 - Ask only for consequential product choices or external authority that cannot
-  be inferred. Do not ask the customer to restate facts the system proves.
+  be inferred. Offer a recommended setup and use the existing structured question
+  UI, when available, for the few unresolved choices about chat sharing, learning
+  across chats, or data access. Skip choices already settled; do not impose an
+  onboarding questionnaire. See [Discovery and autonomy](references/discovery-and-autonomy.md).
 - Use a reversible, clearly stated default when an unresolved choice is
   low-risk. Resolve privacy, tenant authority, data writes, cost exposure, and
   ambiguous external mutations before crossing those boundaries.

package/src/bundled_default_skills/opengeni-client/references/discovery-and-autonomy.md

@@ -52,6 +52,37 @@ Group tightly related unresolved decisions when that makes them easier to answer
 
 For a missing privacy answer, default provisionally to the smaller sharing boundary and explain the operational cost. Do not silently weaken isolation to reduce workspace count.
 
+### Keep product choices lightweight
+
+When meaningful choices remain, group them into one short question interaction
+using the host's existing structured human-input UI when available, or a concise
+chat question otherwise. Recommend the setup that fits the product and let the
+user accept it or adjust individual choices. Do not build a new questionnaire or
+ask every integration the same questions. A suggested answer is not consent to
+send data, share private content, or perform an external action.
+
+Use product language for the relevant unresolved choices:
+
+- **Learning across chats:** no new lasting learning, remember for each person,
+  or shared team knowledge. Explain briefly that chat history/retention is separate
+  and disabling learning does not delete history or existing authorized Knowledge.
+- **Who can open chats:** only their owner, the team, or a choice on each chat.
+  Keep human visibility separate from agent access and Knowledge scope. Choose the
+  workspace mapping from the actual sharing boundary; private chats alone do not
+  require a workspace per person.
+- **How the agent gets data:** current-page snapshots, read-only tools to fetch
+  more reports, or controlled queries for deeper analysis. State the meaningful
+  limitation of the recommendation. Confirm broader access or writes separately
+  only when they are part of the requested product.
+
+For example, if all three choices are unresolved for a simple dashboard, propose
+“Private chats, no learning between chats, and current-page data only” with a short
+explanation that the agent cannot fetch another report on its own. Do not reuse
+that default for a team assistant whose requirements already imply shared work.
+Continue independent discovery while awaiting an answer; ask again only when new
+information introduces a material decision. Summarize any provisional choices in
+the handoff so they do not become invisible product decisions.
+
 ## Follow the wanted autonomy
 
 Infer the delivery mode from explicit user language first, then repository guidance and established team workflow:

package/src/bundled_default_skills/opengeni-client/references/runtime-profile-and-verification.md

@@ -154,4 +154,36 @@ silently linking identities or widening membership. Complete discoverable setup
 before handoff; identify operator-only blockers separately and do not depend on
 internal source, cluster access, prior chat memory, or undocumented local helpers.
 
+### Make the next step easy
+
+Keep the user-facing handoff short: what works, what was verified, and the next
+action needed to make it usable. Put detailed commands and configuration in the
+repository setup guide and link it. When setup is blocked, show a compact checklist
+in the conversation using existing text/link UI; no new setup-card renderer is
+required.
+
+For missing OpenGeni credentials, point to the selected deployment's Organization
+settings: the overview shows **Organization ID** with a copy control, and
+**Developer** contains **Create Organization API Key** plus another ID copy
+control. Prefer a verified settings link or returned setup action over an invented
+URL. Verify older/self-hosted UI availability before promising these controls.
+Explain the key type and required access for the implemented calls: workspace
+provisioning and chat writes need a write-capable organization key, not a read-only
+key. State its actual scope without inventing granular permission switches.
+
+Name the exact server configuration variables and secret-store destination. Fill
+verified non-secret IDs yourself when authorized and unambiguous. Give a secure
+interactive command or secret-manager action for the key; never ask the user to
+paste a secret into chat or put it in a browser variable. Follow with the actual
+deployment command/target and a first-question smoke test, including model/billing
+readiness when unverified. Complete authorized steps yourself; clearly identify
+any remaining owner login, secret entry, or deployment approval.
+
+Resolve ordinary missing tools using the project's supported installation or
+package-runner path where possible. Investigate dependency warnings enough to
+state their effect and available fix. Surface only material unresolved risks or
+user actions in plain language; keep routine audit counts and logs in the setup
+or verification notes. Do not claim a warning is harmless without evidence, or
+make unrelated upgrades merely to silence the report.
+
 If a durable customer integration Skill would reduce future rediscovery, generate one beside the integration code containing only stable, non-secret project facts and smoke probes. Do not turn the generic implementation Skill into the customer's analytics prompt, and do not make generated runtime behavior depend on the implementation workspace retaining this implementation Skill.

package/src/codex-tool-search.ts

@@ -185,6 +185,41 @@ export function bm25RankTools(tools: Tool[], query: string, limit: number): Tool
   return hits.slice(0, limit).map((s) => s.tool); // no hits ⇒ [] (codex-rs parity; see doc)
 }
 
+/**
+ * Tools a free-text query names exactly, in query order. A query token that
+ * contains "_" names a tool when it equals the tool's full model name or the
+ * part after its last "__" server separator (`goal_complete` and
+ * `opengeni__goal_complete` both name `opengeni__goal_complete`). Single words
+ * such as "read" are never treated as names. Agents often search by exact name,
+ * and BM25 alone ranked a neighbouring tool first for about a fifth of those
+ * queries (`goal_complete` returned `goal_set` first).
+ */
+export function toolsNamedInQuery<T extends Tool & { name: string }>(
+  tools: T[],
+  query: string,
+): T[] {
+  const byName = new Map<string, T[]>();
+  for (const tool of tools) {
+    const full = tool.name.toLowerCase();
+    const separator = full.lastIndexOf(MCP_TOOL_NAME_SEPARATOR);
+    const short = separator >= 0 ? full.slice(separator + MCP_TOOL_NAME_SEPARATOR.length) : full;
+    for (const key of new Set([full, short])) {
+      const named = byName.get(key) ?? [];
+      named.push(tool);
+      byName.set(key, named);
+    }
+  }
+  const named: T[] = [];
+  for (const raw of query.toLowerCase().split(/[^a-z0-9_]+/)) {
+    const token = raw.replace(/^_+|_+$/g, "");
+    if (!token.includes("_")) continue;
+    for (const tool of byName.get(token) ?? []) {
+      if (!named.includes(tool)) named.push(tool);
+    }
+  }
+  return named;
+}
+
 /** Parse ranked or exact-name discovery arguments (string or object). */
 function parseSearchArgs(raw: unknown): { query: string; limit: number; names?: string[] } {
   let obj: Record<string, unknown> = {};
@@ -231,12 +266,25 @@ export function searchToolPool(availableTools: Tool[], rawArguments: unknown): T
       tool.type === "function" && typeof (tool as { name?: unknown }).name === "string",
   );
   if (searchable.length === 0) return [];
-  const { query, limit, names } = parseSearchArgs(rawArguments);
-  // Exact disclosure is a lookup in this already-authorized pool, not a fuzzy
-  // search or another executor registry. Keep the SDK's original references.
-  const ranked = names
-    ? searchable.filter((tool) => names.includes(tool.name))
-    : bm25RankTools(searchable, query, searchable.length);
+  const { query, limit: requestedLimit, names } = parseSearchArgs(rawArguments);
+  let ranked: Tool[];
+  let limit = requestedLimit;
+  if (names) {
+    // Exact disclosure is a lookup in this already-authorized pool, not a fuzzy
+    // search or another executor registry. Keep the SDK's original references.
+    ranked = searchable.filter((tool) => names.includes(tool.name));
+  } else {
+    // Every tool the query names comes first, even beyond the requested limit;
+    // BM25 fills any remaining places.
+    const named = toolsNamedInQuery(searchable, query);
+    ranked = [
+      ...named,
+      ...bm25RankTools(searchable, query, searchable.length).filter(
+        (tool) => !named.includes(tool as Tool & { name: string }),
+      ),
+    ];
+    limit = Math.min(MAX_SEARCH_LIMIT, Math.max(requestedLimit, named.length));
+  }
   const bounded: Tool[] = [];
   let disclosedBytes = 0;
   for (const tool of ranked) {

package/src/context-compaction.ts

@@ -43,6 +43,10 @@ export type CompactionItem = Record<string, unknown>;
 export const COMPACTION_SUMMARY_MARKER = "opengeni_context_summary";
 
 export const SUMMARY_BUFFER_TOKENS = 20_000;
+/** Leave input room on models whose entire context is smaller than the normal reserve. */
+export function compactionSummaryOutputTokens(contextWindowTokens: number): number {
+  return Math.min(SUMMARY_BUFFER_TOKENS, Math.max(1, Math.floor(contextWindowTokens / 4)));
+}
 // A single cumulative budget for all retained real user messages, matching
 // Codex core's build_compacted_history_with_limit (not a per-message allowance).
 export const COMPACT_USER_MESSAGE_MAX_TOKENS = 20_000;
@@ -1243,7 +1247,7 @@ export class EmptyCompactionSummaryError extends Error {
   constructor(diagnostics: Record<string, unknown> = {}) {
     const compact = JSON.stringify(diagnostics).slice(0, 2_000);
     super(
-      `Compaction summarizer returned no assistant text; active history was preserved${compact ? ` (${compact})` : ""}`,
+      `Compaction could not produce a usable checkpoint; active history was preserved${compact ? ` (${compact})` : ""}`,
     );
     this.name = "EmptyCompactionSummaryError";
     this.diagnostics = diagnostics;
@@ -1520,10 +1524,18 @@ export function omitOpaqueArtifactsFromPortableCompactionHistory(
     }
     if (type === "reasoning" && hasOpaqueProviderArtifact(item)) {
       const projected = projectRejectedReasoningArtifact(item);
+      if (Object.keys(projected).length === 1) {
+        changed = true;
+        continue;
+      }
       changed ||= projected !== item;
       out.push(projected as CompactionItem);
       continue;
     }
+    if (type === "reasoning" && Object.keys(item).length === 1) {
+      changed = true;
+      continue;
+    }
     out.push(item);
   }
   return changed ? out : (items as CompactionItem[]);
@@ -1605,16 +1617,20 @@ function oldestLogicalUnitCuts(items: readonly CompactionItem[]): number[] {
 
 /**
  * Build the active history after compaction:
- * the newest real user messages that fit one cumulative 20k-token budget
+ * the newest real user messages that fit one cumulative, model-bounded budget
  * (prior summaries excluded, retained images preserved) plus one marked summary item.
  */
 export function buildCompactionReplacementHistory(
   items: readonly CompactionItem[],
   summaryBody: string,
   retainedItemTokens: (item: CompactionItem) => number = (item) => estimateTokens([item]),
+  retainedMessageBudgetTokens = COMPACT_USER_MESSAGE_MAX_TOKENS,
 ): CompactionItem[] {
   const retainedReversed: CompactionItem[] = [];
-  let remaining = COMPACT_USER_MESSAGE_MAX_TOKENS;
+  let remaining = Math.max(
+    0,
+    Math.min(COMPACT_USER_MESSAGE_MAX_TOKENS, retainedMessageBudgetTokens),
+  );
   for (let index = items.length - 1; index >= 0 && remaining > 0; index -= 1) {
     const item = items[index]!;
     if (!isUserMessage(item) || isCompactionSummary(item) || isAttachmentCatalog(item)) {