@opengeni/runtime 4.3.0-canary.36855269964001 → 4.3.0

package/dist/claude-subscription-usage.d.ts

@@ -1,7 +1,5 @@
 type Observer = (providerId: string, response: Response) => void;
-type Prepare = (providerId: string, headers: Headers) => Promise<Headers>;
 /** Like Codex usage headers: free observations within this turn's provider context. */
-export declare function withClaudeUsageObserver<T>(observer: Observer, run: () => Promise<T>, prepare?: Prepare): Promise<T>;
-export declare function prepareClaudeSubscriptionRequest(providerId: string, input: Parameters<typeof fetch>[0], init?: RequestInit): Promise<RequestInit | undefined>;
+export declare function withClaudeUsageObserver<T>(observer: Observer, run: () => Promise<T>): Promise<T>;
 export declare function observeClaudeUsageResponse(providerId: string, response: Response): void;
 export {};

package/dist/index.d.ts

@@ -35,15 +35,13 @@ import { type WorkspaceSkillSearchPath } from "./workspace-skills.js";
 import { type EffectiveSkillSelection, type RuntimeSkillActivation, type RuntimeSkillIndexEntry } from "./runtime-skills.js";
 export { composeRuntimeSkills, loadNativeToolSkillArtifacts, readRuntimeSkill, skillCatalogFromComposition, type EffectiveSkillSelection, type InstalledSkillActivation, type NativeToolSkillSet, type RuntimeSkillActivation, type RuntimeSkillArtifact, type RuntimeSkillArtifactFile, type RuntimeSkillComposition, type RuntimeSkillDescriptor, type RuntimeSkillIndexEntry, type SessionSkillActivation, } from "./runtime-skills.js";
 import { type PersistentAgentInstructionInspection } from "./model-context-inspector.js";
-import { type ModelCallLifecycle } from "./model-request-capture.js";
-export { beforeModelRequest as awaitModelCallAdmission } from "./model-request-capture.js";
 import { type ModelPreparationMeasurement } from "./model-preparation-diagnostics.js";
 import type { ModelResponseUsage, NormalizedRuntimeEvent, SerializedHumanInputInterruption, SerializedInteractionInterventionInterruption } from "./run-events.js";
 export { buildPortableSkillArtifact, getSkillLibraryEntry, isSkillLibraryEntryId, listSkillLibraryEntries, loadSkillLibrarySkill, PORTABLE_SKILL_MAX_FILE_BYTES, PORTABLE_SKILL_MAX_FILES, PORTABLE_SKILL_MAX_TOTAL_BYTES, parsePortableSkillFrontmatter, skillLibraryRepositoryUrl, skillArtifactContentSha256, type PortableSkillArtifact, type SkillLibraryEntry, type SkillLibraryFile, type SkillLibrarySkill, } from "./skill-library.js";
 export { MCP_LIFECYCLE_OUTCOMES, MCP_LIFECYCLE_PHASES, MCP_LIFECYCLE_POLICIES, MCP_TOOL_CALL_OUTCOMES, SANDBOX_READINESS_REPLACEMENT_OUTCOMES, type McpLifecycleOutcome, type McpLifecyclePhase, type McpLifecyclePolicy, type McpToolCallOutcome, type RuntimeMetricsHooks, type SandboxReadinessReplacementOutcome, } from "./metrics.js";
 export type { ModelPreparationMeasurement, ModelPreparationPhase, } from "./model-preparation-diagnostics.js";
 export { markModelPreparationFirstSandboxOperation, recordModelPreparationMeasurement, } from "./model-preparation-diagnostics.js";
-export { CodexSubscriptionUnavailableError, OrganizationGatewayUnavailableError, OrganizationOpenRouterUnavailableError, MultiProviderModelProvider, OpenGeniChatCompletionsModel, OpenGeniResponsesModel, ResponsesStreamingTerminalError, UNKNOWN_MODEL_FINISH_REASON_CODE, UnknownModelFinishReasonError, WorkspaceGatewayUnavailableError, WorkspaceOpenRouterUnavailableError, WorkspaceModelPolicyBlockedError, XaiSubscriptionUnavailableError, azureOpenAIDefaultQuery, buildModelInstance, buildOpenAIClientFromSettings, buildProviderClient, configureOpenAI, configureRuntimeMetricsHooks, modelRequestPolicyForProvider, normalizeVercelGatewayRequestBody, resolveTurnModel, vercelGatewayRoutingFetch, } from "./model-provider.js";
+export { CodexSubscriptionUnavailableError, OrganizationGatewayUnavailableError, OrganizationOpenRouterUnavailableError, MultiProviderModelProvider, OpenGeniChatCompletionsModel, OpenGeniResponsesModel, UNKNOWN_MODEL_FINISH_REASON_CODE, UnknownModelFinishReasonError, WorkspaceGatewayUnavailableError, WorkspaceOpenRouterUnavailableError, WorkspaceModelPolicyBlockedError, XaiSubscriptionUnavailableError, azureOpenAIDefaultQuery, buildModelInstance, buildOpenAIClientFromSettings, buildProviderClient, configureOpenAI, configureRuntimeMetricsHooks, modelRequestPolicyForProvider, normalizeVercelGatewayRequestBody, resolveTurnModel, vercelGatewayRoutingFetch, } from "./model-provider.js";
 export { appendWorkspaceGovernance, CompanyProfilePromptLimitError, hasActiveWorkspaceInstructionPolicy, renderWorkspaceGovernanceContext, WorkspaceGovernancePromptLimitError, type WorkspaceGovernanceContext, } from "./workspace-governance.js";
 export { buildModelContextSnapshot, buildModelContextSnapshotFromRequest, createModelVisibleContextCaptureFilter, joinPersistentAgentInstructionLayers, type PersistentAgentInstructionInspection, type PersistentAgentInstructionLayerDraft, } from "./model-context-inspector.js";
 export { createTurnToolCancellationController, TurnSandboxCommandCancelledError, } from "./sandbox/turn-tool-cancellation.js";
@@ -1041,10 +1039,6 @@ export declare function restoreInterruptedRunState(agent: Agent<any, any>, seria
 export declare const TURN_OPERATIONAL_NOTICE_PREFIX = "[OpenGeni turn-scoped operational notice]\n";
 export declare function prepareRunInput(agent: Agent<any, any>, input: AgentSegmentInput, options?: PrepareInputOptions): Promise<PreparedAgentInput>;
 export type RunAgentStreamOptions = {
-    /** Producer-side gate before calling any resolved model/provider. */
-    beforeModelRequest?: ModelCallLifecycle["beforeModelRequest"];
-    /** Register terminal response settlement before the next producer model request. */
-    onModelResponse?: ModelCallLifecycle["onModelResponse"];
     /** Abort the provider/tool loop when the owning activity is cancelled. */
     signal?: AbortSignal;
     /** Nonblocking phase measurements for request preparation before provider I/O. */

package/dist/index.js

@@ -52,7 +52,6 @@ import {
   REMOTE_COMPACTION_V2_BETA_FEATURE,
   REMOTE_COMPACTION_V2_IMPLEMENTATION,
   REMOTE_V2_RETAINED_MESSAGE_TOKEN_BUDGET,
-  ResponsesStreamingTerminalError,
   RunMcpCredentialError,
   RunMcpCredentials,
   SANDBOX_READINESS_REPLACEMENT_OUTCOMES,
@@ -88,7 +87,6 @@ import {
   attemptToolCallMeta,
   azureCliLoginCommand,
   azureOpenAIDefaultQuery,
-  beforeModelRequest,
   boundModelToolOutputsFilterForSettings,
   buildAgentCapabilities,
   buildCompactionPromptInput,
@@ -285,7 +283,7 @@ import {
   workspaceEnvironmentInstructions,
   wrapAttemptToolDefinitions,
   wrapAttemptToolExecute
-} from "./chunk-PEEYZDLS.js";
+} from "./chunk-22M75RJF.js";
 import {
   ActiveBackendUnresolvableError,
   BROWSER_CONTROL_ADMIN_TOKEN_FILE,
@@ -496,7 +494,6 @@ import {
   recordingStorageKey,
   redactOpenSandboxSignedUriPath,
   relativeConnectedMachinePath,
-  renderRoutingMutationOutcomeUnknownToolResult,
   renderSelfhostedFault,
   renewSandboxProviderExpiration,
   repairSerializedRunStateExposedPorts,
@@ -563,7 +560,7 @@ import {
   withRunCredentialsClient,
   withRunCredentialsSession,
   withoutSandboxProviderIdentity
-} from "./chunk-BK3S65CI.js";
+} from "./chunk-JUFVYQCY.js";
 import "./chunk-7KVMU7O3.js";
 import "./chunk-4A4XTKKX.js";
 import "./chunk-GUJMCB7F.js";
@@ -707,7 +704,6 @@ export {
   REPOSITORY_DISCOVERY_LIMIT,
   RecordingError,
   RecordingUnavailableError,
-  ResponsesStreamingTerminalError,
   RoutingActiveRouteChangedError,
   RoutingBackendRecoveryRequiredError,
   RoutingMutationOutcomeUnknownError,
@@ -794,7 +790,6 @@ export {
   assertProviderRegistryInvariants,
   assertSafeRelPath,
   attemptToolCallMeta,
-  beforeModelRequest as awaitModelCallAdmission,
   azureCliLoginCommand,
   azureOpenAIDefaultQuery,
   backendSupportsOs,
@@ -1045,7 +1040,6 @@ export {
   relativeConnectedMachinePath,
   releaseMcpResultCustomDataFromSdkEvent,
   renderCompactionPromptInputForChat,
-  renderRoutingMutationOutcomeUnknownToolResult,
   renderSelfhostedFault,
   renderSessionGoalContext,
   renderWorkspaceGovernanceContext,

package/dist/input-wait-yield.d.ts

@@ -44,10 +44,6 @@ export declare class InputWaitYield {
      * rolls back the remote mutation nor replaces its durable attempt authority.
      */
     sealForSettlement(signal?: AbortSignal | undefined): Promise<void>;
-    /** A settled stream may need one final-reply handoff before attempt closure.
-     * Drain already-admitted receipts without reopening that stream's admission.
-     */
-    drainForHandoff(signal?: AbortSignal | undefined): Promise<void>;
     private endStream;
     private assertCurrent;
     private openTools;

package/dist/mcp-run-credentials.d.ts

@@ -1,8 +1,7 @@
-import { CredentialProviderMcpMaterial, type McpServerConnectionRef } from "@opengeni/contracts";
+import { CredentialProviderMcpMaterial } from "@opengeni/contracts";
 export type RunMcpCredentialTarget = {
     id: string;
     url: string;
-    connectionRef?: McpServerConnectionRef | undefined;
 };
 export type RunMcpCredentialMaterial = {
     mcp?: CredentialProviderMcpMaterial;
@@ -27,7 +26,7 @@ export declare class RunMcpCredentials {
     /** Validate a complete replacement before changing any live request headers. */
     prepare(material: RunMcpCredentialMaterial | null): () => void;
     replace(material: RunMcpCredentialMaterial | null): void;
-    /** Refuse credentials for native, omitted, local, or rewritten server routes. */
+    /** Refuse credentials for omitted, local, or rewritten server routes. */
     assertRemoteTargets(targets: readonly RunMcpCredentialTarget[]): void;
     excludeLocalTarget(id: string): void;
     has(id: string): boolean;

package/dist/model-preparation-diagnostics.d.ts

@@ -6,21 +6,12 @@ export type ModelPreparationMeasurement = {
     count?: number;
 };
 type ModelPreparationObserver = (measurement: ModelPreparationMeasurement) => void;
-type ModelTransportFetch = (...args: Parameters<typeof fetch>) => ReturnType<typeof fetch>;
 export declare function withModelPreparationObserver<T>(observer: ModelPreparationObserver | undefined, callback: () => T): T;
 /** Bind one attempt-local durable checkpoint immediately before generic model
  * transport enters fetch. Cached provider clients are process-global, so this
  * authority must be async-local rather than stored on the client instance. */
 export declare function withModelTransportStartedObserver<T>(observer: (() => Promise<void> | void) | undefined, callback: () => T): T;
 export declare function recordModelTransportStarted(): Promise<void>;
-/** The OpenAI SDK retries thrown fetch errors and replaces their identity with
- * APIConnectionError. Carry only our pre-wire refusal through its HTTP-error
- * path, which honors the retry veto. Native transports keep the original throw.
- * Each fetch owns its context even when concurrent turns share a cached client.
- */
-export declare function sdkModelTransportAdmissionFetch(inner: ModelTransportFetch): ModelTransportFetch;
-/** Restore the exact host-owned error after the SDK has suppressed retries. */
-export declare function rethrowModelTransportAdmissionRefusal(headers: Headers): void;
 /** Record the first routed sandbox operation boundary without publishing an
  * overlapping parent duration. The observer receives only exclusive leaf
  * buckets; this boundary exists solely to split the surrounding SDK work. */

package/dist/model-provider-routing.d.ts

@@ -16,10 +16,6 @@ export declare class OpenGeniChatCompletionsModel extends OpenAIChatCompletionsM
 export declare class OpenGeniResponsesModel extends AppendOnlyOpenAIResponsesModel {
     protected readonly provider: ResolvedModelProvider;
     constructor(client: OpenAI, model: string, provider: ResolvedModelProvider);
-    protected _fetchResponse(request: ModelRequest, stream: false): Promise<OpenAI.Responses.Response>;
-    protected _fetchResponse(request: ModelRequest, stream: true): Promise<AsyncIterable<OpenAI.Responses.ResponseStreamEvent>>;
-    private classifiedResponseStream;
-    private ownsResponsesTerminalClassification;
     protected _buildResponsesCreateRequest(request: ModelRequest, stream: boolean): {
         requestData: Record<string, any>;
         sdkRequestHeaders: Record<string, string | null>;

package/dist/model-provider.d.ts

@@ -1,5 +1,4 @@
 import type { Settings } from "@opengeni/config";
-export { ResponsesStreamingTerminalError } from "./responses-terminal-error.js";
 export { CodexSubscriptionUnavailableError, OrganizationGatewayUnavailableError, OrganizationOpenRouterUnavailableError, UNKNOWN_MODEL_FINISH_REASON_CODE, UnknownModelFinishReasonError, WorkspaceGatewayUnavailableError, WorkspaceOpenRouterUnavailableError, WorkspaceModelPolicyBlockedError, XaiSubscriptionUnavailableError, } from "./model-provider-errors.js";
 export { azureOpenAIDefaultQuery, buildOpenAIClientFromSettings, buildProviderClient, configureRuntimeMetricsHooks, recordRuntimeMcpLifecycleMetric, recordRuntimeMcpToolCallMetric, } from "./model-provider-client.js";
 export { modelRequestPolicyForProvider, normalizeVercelGatewayRequestBody, } from "./model-provider-request-policy.js";

package/dist/model-request-capture.d.ts

@@ -3,15 +3,6 @@ export type ModelRequestCapture = ((request: ModelRequest) => void | Promise<voi
     nextProviderRequestIndex?: () => number;
     onProviderRequest?: (provider: string, body: string | null, reason?: string, index?: number) => void | Promise<void>;
 };
-/** Awaited producer-side authority, separate from observational request capture. */
-export type ModelCallLifecycle = {
-    beforeModelRequest?: () => Promise<void>;
-    /** Register before yielding so the next model entry waits for consumer settlement. */
-    onModelResponse?: (event: StreamEvent) => Promise<void>;
-};
-export declare function withModelCallLifecycle<T>(lifecycle: ModelCallLifecycle, fn: () => T): T;
-export declare function beforeModelRequest(): Promise<void>;
-export declare function modelResponseSettlement(event: StreamEvent): Promise<void> | undefined;
 /** The same agent can re-enter runAgentStream after in-activity compaction. */
 export declare function nextModelContextCaptureIndex(agent: object): number;
 export declare function withModelRequestCapture<T>(capture: ModelRequestCapture | undefined, fn: () => T): T;

package/dist/replayable-json-body.d.ts

@@ -1,4 +1,4 @@
-import OpenAI, { type APIError } from "openai";
+import OpenAI from "openai";
 import type { APIPromise } from "openai/core/api-promise";
 /** Internal fetch-init hook used only when a transport performs its own retry. */
 export declare const REPLAYABLE_REQUEST_BODY_FACTORY: unique symbol;
@@ -46,7 +46,6 @@ export declare function requestBodyText(body: BodyInit | null | undefined): Prom
 export declare class ReplayableJsonOpenAI extends OpenAI {
     private readonly modelRequestPolicy;
     constructor(options: OpenAIOptions, hooks?: ReplayableJsonOpenAIHooks);
-    protected makeStatusError(status: number, error: NonNullable<Parameters<typeof APIError.generate>[1]>, message: string | undefined, headers: Headers): APIError;
     post<Rsp>(path: string, opts?: OpenAIPostOptions): APIPromise<Rsp>;
     buildRequest(inputOptions: OpenAIBuildOptions, context?: OpenAIBuildContext): Promise<OpenAIBuildResult>;
 }

package/dist/sandbox/index.d.ts

@@ -49,7 +49,7 @@ export { nextDurableOpId, runWithToolCallCorrelation, sanitizeOpIdToken } from "
 export { negotiateSelfhostedCapabilities, selfhostedLiveness, selfhostedHeartbeatLiveness, SELFHOSTED_RECONNECT_WINDOW_MS, type SelfhostedNegotiationInput, type SelfhostedLivenessState, type SelfhostedEnrollment, } from "./selfhosted/capabilities.js";
 export { MockAgentResponder, type MockAgentResponderOptions, type MockExecHandler, } from "./selfhosted/testing.js";
 export { createMockSelfhostedOpStream, FakeOpRunner, type FakeOpScript, InMemoryOpStreamTransport, } from "./selfhosted/op-testing.js";
-export { isRoutingMutationOutcomeUnknownError, RoutingBackendRecoveryRequiredError, RoutingMutationOutcomeUnknownError, renderRoutingMutationOutcomeUnknownToolResult, RoutingRetainedProcessNotFoundError, RoutingSandboxSession, RoutingWorkspaceRootChangedError, RoutingActiveRouteChangedError, RoutingUnsupportedError, type ActivePointer, type DefaultBackendLossResult, type RoutableBackendSession, type ResolvedActiveBackend, type RoutingMutationSettlementResult, type RoutingSandboxFirstOperationObservation, type RoutingSandboxFirstOperationObserver, type RoutingSandboxFirstOperationPhase, type RoutingSandboxFirstOperationPhaseObservation, type RoutingSandboxOperationObservation, type RoutingSandboxOperationObserver, type RoutingSandboxCaptureWaitObserver, type RoutingSandboxPhaseOutcome, type RoutingSandboxWaitObservation, type RoutingRetainedProcess, type RoutingRetainedProcessAdoption, type RoutingRetainedProcessTerminalProof, type RoutingSandboxSessionDeps, type RoutingTransitionEvent, } from "./routing/routing-session.js";
+export { isRoutingMutationOutcomeUnknownError, RoutingBackendRecoveryRequiredError, RoutingMutationOutcomeUnknownError, RoutingRetainedProcessNotFoundError, RoutingSandboxSession, RoutingWorkspaceRootChangedError, RoutingActiveRouteChangedError, RoutingUnsupportedError, type ActivePointer, type DefaultBackendLossResult, type RoutableBackendSession, type ResolvedActiveBackend, type RoutingMutationSettlementResult, type RoutingSandboxFirstOperationObservation, type RoutingSandboxFirstOperationObserver, type RoutingSandboxFirstOperationPhase, type RoutingSandboxFirstOperationPhaseObservation, type RoutingSandboxOperationObservation, type RoutingSandboxOperationObserver, type RoutingSandboxCaptureWaitObserver, type RoutingSandboxPhaseOutcome, type RoutingSandboxWaitObservation, type RoutingRetainedProcess, type RoutingRetainedProcessAdoption, type RoutingRetainedProcessTerminalProof, type RoutingSandboxSessionDeps, type RoutingTransitionEvent, } from "./routing/routing-session.js";
 export { makeActiveBackendResolver, ActiveBackendUnresolvableError, swapTargetEstablishability, type ActiveBackendResolverDeps, type RoutableSandbox, type BackendUnresolvableCode, type SwapTargetEstablishability, } from "./routing/backend-resolver.js";
 /**
  * Construct the raw provider SandboxClient for the configured backend. Registry-

package/dist/sandbox/index.js

@@ -206,7 +206,6 @@ import {
   recordingStorageKey,
   redactOpenSandboxSignedUriPath,
   relativeConnectedMachinePath,
-  renderRoutingMutationOutcomeUnknownToolResult,
   renderSelfhostedFault,
   renewSandboxProviderExpiration,
   repairSerializedRunStateExposedPorts,
@@ -273,7 +272,7 @@ import {
   withRunCredentialsClient,
   withRunCredentialsSession,
   withoutSandboxProviderIdentity
-} from "../chunk-BK3S65CI.js";
+} from "../chunk-JUFVYQCY.js";
 import "../chunk-7KVMU7O3.js";
 import "../chunk-4A4XTKKX.js";
 import "../chunk-GUJMCB7F.js";
@@ -487,7 +486,6 @@ export {
   recordingStorageKey,
   redactOpenSandboxSignedUriPath,
   relativeConnectedMachinePath,
-  renderRoutingMutationOutcomeUnknownToolResult,
   renderSelfhostedFault,
   renewSandboxProviderExpiration,
   repairSerializedRunStateExposedPorts,

package/dist/sandbox/provider-command-session.d.ts

@@ -37,12 +37,6 @@ export type ProviderCommandPersistence = {
 export declare class ProviderCommandStartRejectedError extends Error {
     constructor(cause: unknown);
 }
-/** Start was attempted exactly once, but its acknowledgement is unavailable.
- * The client-chosen locator is reconciliation authority, never replay authority. */
-export declare class ProviderCommandStartOutcomeUnknownError extends Error {
-    readonly command: ModalRouterProviderCommand;
-    constructor(command: ModalRouterProviderCommand, cause: unknown);
-}
 export type ProviderCommandSession = {
     verifyCommandSupervisionCapability?(): Promise<{
         sandboxId: string;

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

@@ -21,11 +21,6 @@ export declare class ModalCommandStartPreDispatchUnavailableError extends Error
     private constructor();
     static ensureReady(client: Client, signal?: AbortSignal): Promise<void>;
 }
-/** Local cancellation/closure before Start dispatch. This permits exact
- * never-started reservation settlement, NOT another launch or turn recovery. */
-export declare class ModalCommandStartNotDispatchedError extends Error {
-    constructor(cause: unknown);
-}
 export type ModalRouterIdentity = {
     taskId: string;
     execId: string;

package/dist/sandbox/routing/routing-session.d.ts

@@ -175,11 +175,11 @@ export interface RoutingSandboxSessionDeps {
         op: string;
         backend: ResolvedActiveBackend;
         admission: unknown;
-        outcome: "resolved" | "rejected" | "outcome_unknown";
-        /** Provider result for a resolved call. Absent when Start is unknown. */
+        outcome: "resolved" | "rejected";
+        /** Provider result for a resolved call. Never present for rejection. */
         result?: unknown;
-        /** Exact candidate for a yielded process or a client-chosen unknown Start.
-         * Unknown Start settlement retains this writer; it does not prove exit. */
+        /** Stable candidate generated before durable promotion is attempted. It is
+         * supplied only when the provider returned a positive yielded-session id. */
         retainedProcess?: RoutingRetainedProcess;
     }) => Promise<void | RoutingMutationSettlementResult>;
     /** Admit one model/user-visible stdin mutation under the already-durable
@@ -343,7 +343,6 @@ export declare class RoutingRetainedProcessNotFoundError extends Error {
     readonly retryable = false;
     constructor(providerSessionId: number);
 }
-export declare function renderRoutingMutationOutcomeUnknownToolResult(error: RoutingMutationOutcomeUnknownError): string;
 /**
  * ONE stable session-shaped object the SDK binds to. Every method re-reads the
  * pointer, resolves the active backend (cached by epoch), and dispatches. A

package/dist/workspace-tool-gateway.js

@@ -1,7 +1,7 @@
 import {
   prepareAgentTools
-} from "./chunk-PEEYZDLS.js";
-import "./chunk-BK3S65CI.js";
+} from "./chunk-22M75RJF.js";
+import "./chunk-JUFVYQCY.js";
 import "./chunk-7KVMU7O3.js";
 import "./chunk-4A4XTKKX.js";
 import "./chunk-GUJMCB7F.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@opengeni/runtime",
-  "version": "4.3.0-canary.36855269964001",
+  "version": "4.3.0",
   "license": "Apache-2.0",
   "repository": {
     "type": "git",
@@ -63,17 +63,17 @@
     "@openai/agents": "0.14.3",
     "@openai/agents-core": "0.14.3",
     "@openai/agents-extensions": "0.14.3",
-    "@opengeni/agent-proto": "^0.6.0-canary.36855269964001",
-    "@opengeni/capabilities": "^0.3.4-canary.36855269964001",
-    "@opengeni/codemode": "^0.6.5-canary.36855269964001",
-    "@opengeni/codex": "^0.2.29-canary.36855269964001",
-    "@opengeni/config": "^3.1.1-canary.36855269964001",
-    "@opengeni/contracts": "^5.4.0-canary.36855269964001",
-    "@opengeni/network": "^0.3.1-canary.36855269964001",
-    "@opengeni/observability": "^0.8.35-canary.36855269964001",
-    "@opengeni/sdk": "^7.4.0-canary.36855269964001",
-    "@opengeni/tool-gateway": "^0.1.16-canary.36855269964001",
-    "@opengeni/xai-subscription": "^0.1.4-canary.36855269964001",
+    "@opengeni/agent-proto": "^0.6.0",
+    "@opengeni/capabilities": "^0.3.4",
+    "@opengeni/codemode": "^0.6.5",
+    "@opengeni/codex": "^0.2.29",
+    "@opengeni/config": "^3.1.1",
+    "@opengeni/contracts": "^5.4.0",
+    "@opengeni/network": "^0.3.1",
+    "@opengeni/observability": "^0.8.35",
+    "@opengeni/sdk": "^7.4.0",
+    "@opengeni/tool-gateway": "^0.1.16",
+    "@opengeni/xai-subscription": "^0.1.4",
     "debug": "4.4.3",
     "modal": "0.9.0",
     "openai": "6.47.0",

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

@@ -184,11 +184,6 @@ Read selectively: [Product integration shapes](references/product-integration-sh
 [Data tools and credentials](references/data-tools-and-credentials.md), and
 [External users and embedded connection setup](references/external-users-and-connect.md).
 
-For per-seat included usage, administrator splits, top-ups, team budgets, or
-browser progress meters, read [Usage allowances](references/usage-allowances.md).
-Keep organization-budget writes on the backend; the conversation proxy exposes
-only own usage. These are post-call ceilings, not prepaid reservations.
-
 ## Build Gotchas
 
 - Always pass `baseUrl` (`process.env.OPENGENI_API_BASE_URL`); the chat facade

package/src/bundled_default_skills/opengeni-client/references/api-workflows.md

@@ -25,8 +25,6 @@ Only add `x-opengeni-access-key` when the operator says the deployment
 shared-key boundary is enabled. It is not a replacement for organization API
 keys in managed SaaS.
 
-A full organization API key can provision workspaces, members and asUser sessions; /v1/access/me reports this as credential.effectiveWorkspacePermissions.
-
 ## Minimal Server-Side Session Client
 
 ```ts
@@ -249,12 +247,6 @@ Never distribute an OpenGeni credential to a Connected Machine or try to inject
 
 ## Billing And Limits
 
-For per-seat plans, team budgets, administrator splits, and top-ups, read
-[Usage allowances](usage-allowances.md). Amounts are integer USD micros,
-configuration/member writes are versioned, grants are operation-keyed, and
-the conversation proxy exposes only `/usage/me`. Verify actual debit/reset
-and webhook behavior before promising enforcement or notifications.
-
 Managed SaaS uses prepaid Stripe credits and local usage/cost accounting. Client behavior should be simple:
 
 - Show billing/credit status from `/v1/billing` when the user has billing permission.

package/src/bundled_default_skills/opengeni-client/references/product-shapes-and-ui.md

@@ -31,22 +31,9 @@ console routes and 404 on the product origin. Never pass them to a raw `<a>`.
   proxy `sandboxFiles: true` (default off) plus `files:read`; they are confined
   to the selected working directory, including on Connected Machines, and
   refuse symlinked paths. Disabled sandbox links render unavailable.
-- Artifacts and Sites: reuse the OpenGeni surfaces, do not rebuild them.
-  `opengeni-site` fences render the inline Site preview automatically. Mount
-  `SessionArtifactViewer` (`@opengeni/react/artifacts`) in a host container
-  (for example the main area beside an assistant panel, a full-screen sheet on
-  phones) and pass `onOpenArtifact` to `SessionConversation` (or
-  `viewerLinkResolver` to `MessageTimeline`). Enable the proxy's
-  `artifacts: true`; editors also need the document, spreadsheet, and
-  presentation runtimes (`@opengeni/artifact-kernel-wasm-document`,
-  `@opengeni/artifact-kernel-wasm-spreadsheet`,
-  `@opengeni/artifact-kernel-wasm-presentation`) plus the SDK Worker URL
-  (`editableRuntimes`). A custom host proxy reports the viewer capability with
-  `artifactViewerCapability` from `@opengeni/sdk`. Close the viewer when the session changes.
-  Translate its copy with `labels` (partial `ArtifactLabels`) or
-  `ArtifactLabelsProvider`.
-- Other routing: `resolveLink={(target) => ... ({ href } | { open } | null)}`
-  on `SessionConversation` or `MessageTimeline`, or `OpenGeniLinkProvider` for a
+- Route editable artifacts and Sites to the product's own UI with
+  `resolveLink={(target) => ... ({ href } | { open } | null)}` on
+  `SessionConversation` or `MessageTimeline`, or `OpenGeniLinkProvider` for a
   subtree; it also covers `Markdown` inside a custom `renderMessageText`.
   Unresolved targets render as unavailable text, not broken links.
 - `MessageTimeline` alone has no defaults; add

package/src/claude-subscription-usage.ts

@@ -1,34 +1,15 @@
 import { AsyncLocalStorage } from "node:async_hooks";
 
 type Observer = (providerId: string, response: Response) => void;
-type Prepare = (providerId: string, headers: Headers) => Promise<Headers>;
-const observers = new AsyncLocalStorage<{
-  observe: Observer;
-  prepare?: Prepare;
-}>();
+const observers = new AsyncLocalStorage<Observer>();
 
 /** Like Codex usage headers: free observations within this turn's provider context. */
-export function withClaudeUsageObserver<T>(
-  observer: Observer,
-  run: () => Promise<T>,
-  prepare?: Prepare,
-): Promise<T> {
-  return observers.run({ observe: observer, ...(prepare ? { prepare } : {}) }, run);
-}
-export async function prepareClaudeSubscriptionRequest(
-  providerId: string,
-  input: Parameters<typeof fetch>[0],
-  init?: RequestInit,
-) {
-  const prepare = observers.getStore()?.prepare;
-  if (!prepare) return init;
-  const headers = new Headers(input instanceof Request ? input.headers : undefined);
-  new Headers(init?.headers).forEach((value, key) => headers.set(key, value));
-  return { ...init, headers: await prepare(providerId, headers) };
+export function withClaudeUsageObserver<T>(observer: Observer, run: () => Promise<T>): Promise<T> {
+  return observers.run(observer, run);
 }
 export function observeClaudeUsageResponse(providerId: string, response: Response): void {
   try {
-    observers.getStore()?.observe(providerId, response);
+    observers.getStore()?.(providerId, response);
   } catch {
     // Usage telemetry must never change or consume a model response.
   }

package/src/index.ts

@@ -281,7 +281,6 @@ import {
   createSandboxClient,
   isModalTaskExecStartPreDispatchUnavailableError,
   isRoutingMutationOutcomeUnknownError,
-  renderRoutingMutationOutcomeUnknownToolResult,
   repairSerializedRunStateExposedPorts,
   restoredSandboxSessionStateFromEntry,
   setOpenSandboxApplyDiff,
@@ -370,13 +369,10 @@ import {
   ModelRequestCaptureProvider,
   notifyModelRequestCapture,
   withModelRequestCapture,
-  withModelCallLifecycle,
-  type ModelCallLifecycle,
   type ModelRequestCapture,
   nextModelContextCaptureIndex,
 } from "./model-request-capture";
 import { decodeValidatedViewImageDataUrl } from "./view-image-validation";
-export { beforeModelRequest as awaitModelCallAdmission } from "./model-request-capture";
 import {
   baseModelInputFilterForSettings,
   boundModelToolOutputsFilterForSettings,
@@ -462,7 +458,6 @@ export {
   MultiProviderModelProvider,
   OpenGeniChatCompletionsModel,
   OpenGeniResponsesModel,
-  ResponsesStreamingTerminalError,
   UNKNOWN_MODEL_FINISH_REASON_CODE,
   UnknownModelFinishReasonError,
   WorkspaceGatewayUnavailableError,
@@ -2327,7 +2322,6 @@ export function coreInstructions(
 ): string[] {
   return [
     "If the session has a goal, you own it: keep working until you call opengeni__goal_complete with concrete evidence or opengeni__goal_pause with a rationale; revise it with opengeni__goal_update; create one with opengeni__goal_set when given a long-running objective. Resume a paused goal with opengeni__goal_resume when the user asks you to continue, regardless of who paused it, or when the blocker you paused for has cleared. A question alone is not such a request: answer it and leave the goal paused.",
-    "Goal completion records short ledger proof, not the user-facing deliverable. After goal_complete succeeds, finish the same turn with the requested answer, or a concise summary and retained artifact link. Never use evidence as the final reply. A later child result after completion is context to integrate, not a reason to stay silent or restart the completed goal.",
     'Choose durable storage by purpose, including requests to "remember this" or keep something "for future sessions". Knowledge is retrieval-only information, not standing behavior. Use instruction_policy_get then instruction_policy_save for short always-on workspace rules, such as "Keep replies concise; expand when asked." Use Skills for reusable procedures and context-specific or personal behavioral preferences; make the Skill description state when it applies so the prompt index can guide skill_read. Do not save behavioral preferences as Knowledge by rephrasing them as facts like "the user prefers concise replies". Preserve the intended scope: do not turn a personal preference into a workspace-wide rule; use an authorized personal Skill when appropriate, or explain the unavailable scope. Split mixed requests into a short rule and detailed Skill steps without duplicating them in Knowledge.',
     "Before changing workspace instructions, read the current instruction and exact baseline, preserve unrelated rules, append new rules, and use only a localized exact anchored edit for updates or removals. Agents cannot replace the complete instruction; whole-policy rewrites belong in the manual workspace editor. Do not bypass a destination's Agent learning setting, review, scope, limits, or unavailable tools by saving to another destination. Report the actual destination and receipt: active, pending review, or not saved. Do not promise future behavior from a Knowledge save.",
     "Use knowledge_search and knowledge_get when retained information is relevant. For questions about what the workspace knows, ground the answer in authorized Knowledge and attached sources. Missing internal facts remain unknown; do not infer a person's role or cite unrelated public search results as evidence. Keep any relevant external research clearly separate from workspace records. Default retrieval returns published information. Before retaining or correcting anything, also search view=needs_review for existing pending entries and collections; read those with knowledge_get view=needs_review. Pending means unapproved: you may inspect and improve it, but do not present it as accepted knowledge or activate behavioral guidance from it. Reuse the existing entryId and current version instead of creating another proposal each run. Save durable facts, decisions, requirements, incidents, fixes and outcomes autonomously with knowledge_save only when they help a plausible future task. Do not retain routine approvals, acknowledgments, temporary task instructions, status chatter or raw screenshots merely because they occurred. Before saving, use knowledge_prepare_save when available to fetch the authorized collection map with descriptions and parent IDs plus related published and pending entries; otherwise search both views and browse collections. Skip unchanged duplicates, improve the existing entry when appropriate, and create a new entry only for distinct useful information. Read the current entry before correcting it and preserve uncertainty, conflicting evidence and existing relationships. When a user supplies or confirms a durable fact, use knowledge_retain_message to retain the actual message and cite its returned entryId/revisionId in the finding evidence. Preserve existing source evidence and relationships when correcting an entry; do not replace them with empty arrays merely because the user confirmed a new value. Explicit confirmation can supersede a conflicting pending proposal, but explain that outcome to the user. Preserve uncertainty and exact supporting evidence; the fact label is not a verification claim. Chat attachments retain their original files for the conversation without automatically publishing OCR or source text into Knowledge. Inspect an image visually in the current task; save only a useful supported observation, requirement or incident when warranted. Use knowledge_retain_file purpose=evidence only when a selected finding needs the original as supporting evidence, or purpose=reference when deliberately retaining a reusable source document. Supporting sources remain accessible through evidence links and explicit includeEvidence searches, but do not fill ordinary Knowledge discovery. An upload alone is not a reason to save Knowledge. A source entry contains retained original text; a finding states a useful conclusion and cites the exact source revision. Do not create both when they would simply repeat the same text. Reuse collections for customers, products, systems or subjects across sources, and link one entry to multiple collections instead of copying it. Technical incidents belong with the affected system and should include cause, fix and outcome when known. Reorganize references when useful; do not erase source evidence or revision history. Private tasks author personal Knowledge; shared tasks author workspace Knowledge. Automatic publishes immediately, Review first keeps a pending proposal without pausing your task, and Off prevents authoring while permitting retrieval. Never bypass review by making a new entry, switching tools, or asking for another approval. Reuse the same operationId and exact request after an uncertain save, including recovery. Use task_note_save for temporary coordination in this task tree. Conversation history is separate. Workspace instructions are concise standing rules; Skills are reusable procedures with their own Agent learning settings. Do not save the same content in multiple authorities.",
@@ -3879,12 +3873,6 @@ function buildAgentCapabilitiesFromComposition(
       // readiness proof, which reaches bounded same-turn recovery.
       execCommandErrorFunction: (_context, error) => {
         if (isModalTaskExecStartPreDispatchUnavailableError(error)) throw error;
-        if (isRoutingMutationOutcomeUnknownError(error)) {
-          // The outer physical fence must retain the exact process before
-          // rendering uncertainty. Platform/setup calls still throw normally.
-          if (toolCancellation) throw error;
-          return renderRoutingMutationOutcomeUnknownToolResult(error);
-        }
         const details = error instanceof Error ? error.toString() : String(error);
         return `An error occurred while running the tool. Please try again. Error: ${details}`;
       },
@@ -4390,7 +4378,6 @@ export async function prepareAgentTools(
     settings.mcpServers.filter(
       (config) =>
         tools.some((tool) => tool.id === config.id) &&
-        !config.connectionRef &&
         !localRegistry.has(config.id) &&
         !isFirstPartyMcpServer(settings, config) &&
         !isCodexAppsMcpServer(config),
@@ -5014,7 +5001,6 @@ export function selectedSessionRemoteMcpTargets(
     if (configs.length !== 1) return [];
     const config = configs[0]!;
     if (
-      config.connectionRef ||
       isFirstPartyMcpServer(settings, config) ||
       isCodexAppsMcpServer(config) ||
       BUILT_IN_MCP_BRIDGE_ADAPTERS.some((adapter) =>
@@ -8064,10 +8050,6 @@ export async function prepareRunInput(
 }
 
 export type RunAgentStreamOptions = {
-  /** Producer-side gate before calling any resolved model/provider. */
-  beforeModelRequest?: ModelCallLifecycle["beforeModelRequest"];
-  /** Register terminal response settlement before the next producer model request. */
-  onModelResponse?: ModelCallLifecycle["onModelResponse"];
   /** Abort the provider/tool loop when the owning activity is cancelled. */
   signal?: AbortSignal;
   /** Nonblocking phase measurements for request preparation before provider I/O. */
@@ -8311,17 +8293,8 @@ export async function runAgentStream(
   const scope = gate?.beginStream(overrides.signal);
   try {
     if (scope) agent.toolUseBehavior = scope.toolUseBehavior;
-    const stream = await withModelCallLifecycle(
-      {
-        ...(overrides.beforeModelRequest
-          ? { beforeModelRequest: overrides.beforeModelRequest }
-          : {}),
-        ...(overrides.onModelResponse ? { onModelResponse: overrides.onModelResponse } : {}),
-      },
-      () =>
-        withPreparedCompactionRequest(agent, () =>
-          runAgentStreamInternal(agent, input, settings, overrides, scope),
-        ),
+    const stream = await withPreparedCompactionRequest(agent, () =>
+      runAgentStreamInternal(agent, input, settings, overrides, scope),
     );
     // Observe the SDK's own settlement promise before exposing the stream. Do
     // not wrap/replace SDK history, errors, cancellation, or stream iteration.

package/src/input-wait-yield.ts

@@ -127,13 +127,6 @@ export class InputWaitYield {
     await this.drain(this.scope, signal);
   }
 
-  /** A settled stream may need one final-reply handoff before attempt closure.
-   * Drain already-admitted receipts without reopening that stream's admission.
-   */
-  async drainForHandoff(signal = this.scope.signal): Promise<void> {
-    await this.drain(this.scope, signal);
-  }
-
   private endStream(scope: StreamScope): void {
     scope.phase = "settled";
     scope.ended = true;

package/src/interaction-tools.ts

@@ -681,7 +681,6 @@ export function createInteractionAttemptToolDefinitions(
           raw,
           context,
           options.execute,
-          options.readOnly,
         ),
     });
   };
@@ -1916,7 +1915,6 @@ async function safeInteractionExecution<TInput extends z.ZodType, TOutput extend
     value: z.output<TInput>,
     context: AttemptToolExecutionContext,
   ) => Promise<z.input<TOutput> | InteractionExecutionResult<z.input<TOutput>>>,
-  readOnly: boolean,
 ): Promise<AttemptToolResultValue> {
   try {
     const value = inputSchema.parse(raw);
@@ -1938,18 +1936,11 @@ async function safeInteractionExecution<TInput extends z.ZodType, TOutput extend
     if (error instanceof z.ZodError) {
       return interactionErrorResult("invalid_arguments", "Interaction tool arguments are invalid.");
     }
-    // A failed read is still a useful tool result. Preserve the API's public
-    // explanation; mutations and uncertain outcomes retain their failure path.
-    if (
-      error instanceof OpenGeniApiError &&
-      !error.outcomeUnknown &&
-      (error.status < 500 || readOnly)
-    ) {
+    if (error instanceof OpenGeniApiError && !error.outcomeUnknown && error.status < 500) {
       return interactionErrorResult(
         error.code ?? `http_${error.status}`,
         boundedErrorMessage(error.message),
         error.retryable,
-        error.correlationId,
       );
     }
     throw error;
@@ -1974,9 +1965,8 @@ function interactionErrorResult(
   code: string,
   message: string,
   retryable = false,
-  requestId?: string,
 ): AttemptToolResultValue {
-  const error = { code, message, retryable, ...(requestId ? { requestId } : {}) };
+  const error = { code, message, retryable };
   return {
     isError: true,
     content: [{ type: "text" as const, text: JSON.stringify({ error }) }],