@opengeni/runtime 4.6.0 → 4.7.0-canary.37090756480001

package/src/context-compaction.ts

@@ -3,9 +3,9 @@
  *
  * The checkpoint model sees the current active history plus one fixed
  * checkpoint prompt, then the active history is rebuilt from the newest real
- * user messages within one cumulative 20k-token budget plus one summary.
- * Assistant messages, tool calls/results, reasoning, and images are removed
- * from the active model-facing history; the database audit rows remain.
+ * user/system input messages within one cumulative 20k-token budget plus one summary.
+ * Assistant messages, tool calls/results, and reasoning are removed from the
+ * active model-facing history; retained input images and database audit rows remain.
  */
 
 import {
@@ -118,11 +118,16 @@ function itemRole(item: unknown): string | undefined {
   return typeof role === "string" ? role : undefined;
 }
 
-/** A user-authored `message` item is the only legal turn boundary. */
+/** A real user-role `message` item. */
 export function isUserMessage(item: unknown): boolean {
   return itemType(item) === "message" && itemRole(item) === "user";
 }
 
+/** Accepted machine-input batches are canonical system messages, not user intent. */
+function isPortableInputMessage(item: unknown): boolean {
+  return isUserMessage(item) || (itemType(item) === "message" && itemRole(item) === "system");
+}
+
 /** True for our synthetic compaction summary item. */
 export function isCompactionSummary(item: unknown): boolean {
   return (
@@ -1617,7 +1622,7 @@ function oldestLogicalUnitCuts(items: readonly CompactionItem[]): number[] {
 
 /**
  * Build the active history after compaction:
- * the newest real user messages that fit one cumulative, model-bounded budget
+ * the newest user/system input messages that fit one cumulative, model-bounded budget
  * (prior summaries excluded, retained images preserved) plus one marked summary item.
  */
 export function buildCompactionReplacementHistory(
@@ -1633,7 +1638,7 @@ export function buildCompactionReplacementHistory(
   );
   for (let index = items.length - 1; index >= 0 && remaining > 0; index -= 1) {
     const item = items[index]!;
-    if (!isUserMessage(item) || isCompactionSummary(item) || isAttachmentCatalog(item)) {
+    if (!isPortableInputMessage(item) || isCompactionSummary(item) || isAttachmentCatalog(item)) {
       continue;
     }
     const textTokens = estimateTextTokens(messageText(item));
@@ -1667,18 +1672,20 @@ export function isRemoteCompactionItem(item: unknown): item is CompactionItem {
   );
 }
 
-/** Messages retained beside a remote v2 compaction blob (user + developer). */
+/** Messages retained beside a remote v2 compaction blob (user + system + developer). */
 export function isRetainedRemoteV2Message(item: unknown): boolean {
   if (isCompactionSummary(item) || isAttachmentCatalog(item) || itemType(item) === "compaction") {
     return false;
   }
   const role = itemRole(item);
-  return itemType(item) === "message" && (role === "user" || role === "developer");
+  return (
+    itemType(item) === "message" && (role === "user" || role === "system" || role === "developer")
+  );
 }
 
 /**
  * Build the active history after Codex remote compaction v2:
- * newest retained user/developer messages within the CLI 64k budget plus the
+ * newest retained user/system/developer messages within the CLI 64k budget plus the
  * opaque `{ type: "compaction", encrypted_content }` item.
  *
  * Both modes preserve retained image parts. Charge their projected image

package/src/index.ts

@@ -6,11 +6,13 @@ import {
 } from "./prepared-compaction-request";
 export { preparedCompactionRequest, queuePreparedCompaction } from "./prepared-compaction-request";
 import { AnthropicMessagesModel } from "./anthropic-messages";
+export { AnthropicProviderRejection } from "./anthropic-messages";
 import { instrumentedModelFetch } from "./model-provider-client";
 import type { ModelProviderApi, ResolvedModelProvider, Settings } from "@opengeni/config";
 import { isRunMcpCredentialError, RunMcpCredentials } from "./mcp-run-credentials";
 import { normalizeCredentialProviderMcpUrl } from "@opengeni/contracts";
 export { RunMcpCredentials, RunMcpCredentialError } from "./mcp-run-credentials";
+export { AnthropicRequestError } from "./anthropic-request-error";
 import { executeCommandReadWithRefresh } from "./command-read-refresh";
 import {
   captureMcpOperationDispatch,
@@ -266,6 +268,7 @@ import {
 import { AsyncLocalStorage } from "node:async_hooks";
 import { createHash, randomUUID } from "node:crypto";
 import { dirname, isAbsolute, join, posix as posixPath } from "node:path";
+import { gzipSync } from "node:zlib";
 
 import { z } from "zod";
 
@@ -312,6 +315,7 @@ import {
   createSandboxClient,
   isModalTaskExecStartPreDispatchUnavailableError,
   isModalCommandStartOutcomeUnknownError,
+  isProviderCommandObservationUnavailableError,
   isRoutingMutationOutcomeUnknownError,
   renderRoutingMutationOutcomeUnknownToolResult,
   repairSerializedRunStateExposedPorts,
@@ -3974,6 +3978,9 @@ function buildAgentCapabilitiesFromComposition(
       // Preserve that behavior except for client-side, pre-dispatch Modal
       // readiness proof, which reaches bounded same-turn recovery.
       execCommandErrorFunction: (_context, error) => {
+        if (isProviderCommandObservationUnavailableError(error)) {
+          return "Managed sandbox command observation unavailable. Outcome unknown. Do not replay the command or resend stdin; observe the existing invocation.";
+        }
         if (isModalTaskExecStartPreDispatchUnavailableError(error)) throw error;
         if (isRoutingMutationOutcomeUnknownError(error)) {
           // The outer physical fence must retain the exact process before
@@ -11766,11 +11773,11 @@ const RIG_SETUP_PROVIDER_IMAGE_MARKER_ROOT = "/var/opengeni";
 // Modal's command transport caps aggregate argv at 64 KiB. Cancellation and
 // run-as wrappers duplicate/expand this command, so stage moderate scripts too.
 const RIG_SETUP_INLINE_COMMAND_MAX_BYTES = 4 * 1024;
-// The cancellation fence embeds a lifecycle command twice, then the current
-// runAs wrapper repeats it across several execution branches. Keep each base64
-// chunk below Modal's 64-KiB aggregate argument ceiling after both wrappers.
-const RIG_SETUP_PAYLOAD_CHUNK_CHARS = 7 * 1024;
+// Both cancellation and the SDK run-as wrapper repeat the payload three times.
+// Leave room for their fixed shell programs under Modal's 64-KiB argv ceiling.
+const RIG_SETUP_PAYLOAD_CHUNK_CHARS = 2 * 1024;
 const RIG_SETUP_PAYLOAD_ROOT = "/tmp/opengeni/rig-setup-payloads";
+const RIG_SETUP_GZIP_SENTINEL = "__OPENGENI_SETUP_GZIP__";
 
 export type RigSetupScriptCommandOptions = {
   timeoutMs?: number;
@@ -11907,22 +11914,55 @@ async function stageRigSetupScript(
   session: SandboxSessionLike,
   script: string,
   context: SandboxLifecycleHookContext,
+  options: { payloadRoot?: string; label?: string } = {},
 ): Promise<string> {
-  const payloadPath = `${RIG_SETUP_PAYLOAD_ROOT}/${randomUUID()}.sh`;
+  const payloadRoot = options.payloadRoot ?? RIG_SETUP_PAYLOAD_ROOT;
+  const payloadPath = `${payloadRoot}/${randomUUID()}.sh`;
   const encodedPath = `${payloadPath}.b64`;
-  const encoded = Buffer.from(script, "utf8").toString("base64");
-  const commands = [
-    `set -eu\numask 077\nmkdir -p ${shellQuote(RIG_SETUP_PAYLOAD_ROOT)}\n: > ${shellQuote(encodedPath)}`,
-  ];
-  for (let offset = 0; offset < encoded.length; offset += RIG_SETUP_PAYLOAD_CHUNK_CHARS) {
+  const bytes = Buffer.from(script, "utf8");
+  const compressed = gzipSync(bytes);
+  const compressionUseful = compressed.length < bytes.length;
+  try {
+    // Repeated shell programs compress well. Probe in the existing bootstrap
+    // call, retaining the same bounded transfer on machines without gzip.
+    const bootstrap = await runSandboxLifecycleCommand(
+      session,
+      {
+        cmd: [
+          "set -eu",
+          "umask 077",
+          `mkdir -p ${shellQuote(payloadRoot)}`,
+          `: > ${shellQuote(encodedPath)}`,
+          ...(compressionUseful
+            ? [
+                `if command -v gzip >/dev/null 2>&1; then printf '%s\\n' ${shellQuote(RIG_SETUP_GZIP_SENTINEL)}; fi`,
+              ]
+            : []),
+        ].join("\n"),
+        workdir: "/workspace",
+        ...(context.runAs ? { runAs: context.runAs } : {}),
+        yieldTimeMs: SANDBOX_LIFECYCLE_COMMAND_TIMEOUT_MS,
+        maxOutputTokens: 4_000,
+      },
+      context.commandRunner,
+    );
+    assertSandboxCommandSucceeded(
+      bootstrap,
+      options.label ?? "Sandbox Environment setup payload staging",
+    );
+    const useCompression =
+      compressionUseful &&
+      sandboxCommandOutput(bootstrap).split(/\r?\n/u).includes(RIG_SETUP_GZIP_SENTINEL);
+    const encoded = (useCompression ? compressed : bytes).toString("base64");
+    const commands: string[] = [];
+    for (let offset = 0; offset < encoded.length; offset += RIG_SETUP_PAYLOAD_CHUNK_CHARS) {
+      commands.push(
+        `printf '%s' ${shellQuote(encoded.slice(offset, offset + RIG_SETUP_PAYLOAD_CHUNK_CHARS))} >> ${shellQuote(encodedPath)}`,
+      );
+    }
     commands.push(
-      `printf '%s' ${shellQuote(encoded.slice(offset, offset + RIG_SETUP_PAYLOAD_CHUNK_CHARS))} >> ${shellQuote(encodedPath)}`,
+      `set -eu\nbase64 -d < ${shellQuote(encodedPath)}${useCompression ? " | gzip -dc" : ""} > ${shellQuote(payloadPath)}\nchmod 0700 ${shellQuote(payloadPath)}\nrm -f ${shellQuote(encodedPath)}`,
     );
-  }
-  commands.push(
-    `set -eu\nbase64 -d ${shellQuote(encodedPath)} > ${shellQuote(payloadPath)}\nchmod 0700 ${shellQuote(payloadPath)}\nrm -f ${shellQuote(encodedPath)}`,
-  );
-  try {
     for (const command of commands) {
       const result = await runSandboxLifecycleCommand(
         session,
@@ -11935,7 +11975,10 @@ async function stageRigSetupScript(
         },
         context.commandRunner,
       );
-      assertSandboxCommandSucceeded(result, "Sandbox Environment setup payload staging");
+      assertSandboxCommandSucceeded(
+        result,
+        options.label ?? "Sandbox Environment setup payload staging",
+      );
     }
     return payloadPath;
   } catch (error) {
@@ -12168,6 +12211,7 @@ export async function runRepositoryCloneHook(
     editor: null,
     staged: [],
   };
+  let stagedCloneScript: string | null = null;
   try {
     // Direct provider tokens retain the established off-manifest per-exec seed.
     // Smart-Git broker bearers take a stricter path: stage opaque bytes through
@@ -12195,9 +12239,19 @@ export async function runRepositoryCloneHook(
       stagedBrokerSeeds.staged,
       options,
     );
-    const command = sandboxGitProvisioningCommand(
+    let command = sandboxGitProvisioningCommand(
       seedPrefix ? `${seedPrefix}\n${cloneCommand}` : cloneCommand,
     );
+    // SDK setup also wraps run-as commands and cancellation can expand them.
+    // Reuse the bounded script transport instead of sending the whole clone
+    // program through those nested shell arguments.
+    if (Buffer.byteLength(command, "utf8") > RIG_SETUP_INLINE_COMMAND_MAX_BYTES) {
+      stagedCloneScript = await stageRigSetupScript(session, command, context, {
+        payloadRoot: "/tmp/opengeni/repository-setup-payloads",
+        label: "Repository setup payload staging",
+      });
+      command = `exec /bin/sh ${shellQuote(stagedCloneScript)}`;
+    }
     const result = await runSandboxLifecycleCommand(
       session,
       {
@@ -12237,6 +12291,19 @@ export async function runRepositoryCloneHook(
     });
     throw error;
   } finally {
+    if (stagedCloneScript) {
+      await runSandboxLifecycleCommand(
+        session,
+        {
+          cmd: `rm -f ${shellQuote(stagedCloneScript)} ${shellQuote(`${stagedCloneScript}.b64`)}`,
+          workdir: "/workspace",
+          ...(context.runAs ? { runAs: context.runAs } : {}),
+          yieldTimeMs: SANDBOX_LIFECYCLE_COMMAND_TIMEOUT_MS,
+          maxOutputTokens: 1_000,
+        },
+        context.commandRunner,
+      ).catch(() => undefined);
+    }
     if (stagedBrokerSeeds.editor) {
       await cleanupStagedGitCredentialSeeds(stagedBrokerSeeds.editor, stagedBrokerSeeds.staged);
     }

package/src/interaction-tools.ts

@@ -9,6 +9,10 @@ import {
   BrowserActionReceipt,
   BrowserClipboard,
   BrowserDiagnosticBatch,
+  BrowserDownload,
+  BrowserDownloadListResponse,
+  BrowserDownloadSaveRequest,
+  BrowserDownloadSaveResponse,
   BrowserDomReadResponse,
   BrowserDomReadLocator,
   BrowserDomReadSelector,
@@ -279,6 +283,8 @@ const TOOL_PERMISSION = {
   browser_act: "sessions:control",
   browser_clipboard: "sessions:read",
   browser_debug: "sessions:read",
+  browser_downloads: "sessions:read",
+  browser_download_save: ["sessions:control", "files:upload"],
   browser_auth: "sessions:control",
   interaction_request_human: "sessions:control",
   browser_identity: "sessions:control",
@@ -290,7 +296,7 @@ const TOOL_PERMISSION = {
   computer_clipboard: "sessions:read",
   computer_act: "sessions:control",
   computer_lifecycle: "sessions:control",
-} as const satisfies Record<InteractionAttemptToolName, Permission>;
+} as const satisfies Record<InteractionAttemptToolName, Permission | readonly Permission[]>;
 
 const DiscoveryInput = z
   .object({
@@ -618,6 +624,26 @@ const ComputerLifecycleInput = z
 
 const TERMINAL_LIFECYCLES = new Set(["ended", "failed"]);
 
+const BrowserDownloadsInput = z
+  .object({
+    browserSessionId: z.string().uuid(),
+    operation: z.enum(["list", "get"]).default("list"),
+    downloadId: z.string().uuid().optional(),
+  })
+  .strict()
+  .superRefine((value, context) => {
+    if ((value.operation === "get") !== (value.downloadId !== undefined)) {
+      context.addIssue({
+        code: "custom",
+        path: ["downloadId"],
+        message: "downloadId is required only for operation=get",
+      });
+    }
+  });
+const BrowserDownloadSaveInput = BrowserDownloadSaveRequest.omit({ operationId: true })
+  .extend({ browserSessionId: z.string().uuid(), downloadId: z.string().uuid() })
+  .strict();
+
 export type CreateInteractionAttemptToolsInput = {
   transport: InteractionTransport;
   workspaceId: string;
@@ -1039,6 +1065,82 @@ export function createInteractionAttemptToolDefinitions(
       ),
   });
 
+  add({
+    name: "browser_downloads",
+    codemodePath: ["interaction", "browser", "downloads"],
+    title: "Inspect browser downloads",
+    description:
+      "List browser-produced files or get one exact download by id. Metadata identifies completed bytes and their SHA-256; diagnostics alone do not expose a download id or file bytes. Use browser_download_save to materialize a completed download in its source session's workspace. Attached browsers and Lightpanda do not expose managed downloads.",
+    input: BrowserDownloadsInput,
+    output: z.union([BrowserDownloadListResponse, BrowserDownload]),
+    readOnly: true,
+    idempotent: true,
+    execute: async (value) => {
+      if (value.operation === "get") {
+        const download = await input.transport.getBrowserDownload(
+          input.workspaceId,
+          value.browserSessionId,
+          value.downloadId!,
+        );
+        if (
+          download.browserSessionId !== value.browserSessionId ||
+          download.id !== value.downloadId
+        ) {
+          throw new Error("Browser download belongs to another resource");
+        }
+        return download;
+      }
+      const response = await input.transport.listBrowserDownloads(
+        input.workspaceId,
+        value.browserSessionId,
+      );
+      if (
+        response.browserSessionId !== value.browserSessionId ||
+        response.downloads.some(
+          (download) =>
+            download.browserSessionId !== value.browserSessionId ||
+            download.controllerGeneration !== response.controllerGeneration,
+        )
+      ) {
+        throw new Error("Browser downloads belong to another session binding");
+      }
+      return response;
+    },
+  });
+
+  add({
+    name: "browser_download_save",
+    codemodePath: ["interaction", "browser", "downloadSave"],
+    title: "Save browser download to workspace",
+    description:
+      "Save one exact completed managed browser download to a portable relative path in the browser's source session workspace. Requires sessions:control and files:upload. Returns the materialized destinationPath, fileId and integrity metadata; read the saved bytes with ordinary workspace file tools. Existing files are protected unless overwrite=true. The attempt operation id fences retries; uncertain outcomes must reconcile that same operation. Attached browsers and Lightpanda cannot publish managed downloads.",
+    input: BrowserDownloadSaveInput,
+    output: BrowserDownloadSaveResponse,
+    readOnly: false,
+    idempotent: true,
+    execute: async (value, context) => {
+      const response = await input.transport.saveBrowserDownload(
+        input.workspaceId,
+        value.browserSessionId,
+        value.downloadId,
+        {
+          operationId: context.operationId,
+          destinationPath: value.destinationPath,
+          overwrite: value.overwrite,
+        },
+      );
+      if (
+        response.download.browserSessionId !== value.browserSessionId ||
+        response.download.id !== value.downloadId ||
+        response.operationId !== context.operationId ||
+        response.destinationPath !== value.destinationPath
+      ) {
+        throw new Error("Browser download save returned another operation binding");
+      }
+      return response;
+    },
+  });
+
   add({
     name: "browser_auth",
     codemodePath: ["interaction", "browser", "auth"],
@@ -1992,8 +2094,16 @@ function jsonSchema(schema: z.ZodType): AttemptToolJsonSchema {
   return z.toJSONSchema(schema, { target: "draft-2020-12" }) as AttemptToolJsonSchema;
 }
 
-function hasToolPermission(permissions: readonly Permission[], required: Permission): boolean {
-  return permissions.includes(required) || permissions.includes("workspace:admin");
+function hasToolPermission(
+  permissions: readonly Permission[],
+  required: Permission | readonly Permission[],
+): boolean {
+  return (
+    permissions.includes("workspace:admin") ||
+    (typeof required === "string" ? [required] : required).every((permission) =>
+      permissions.includes(permission),
+    )
+  );
 }
 
 function firstPartyApiBaseUrl(settings: Settings, workspaceId: string): string {

package/src/lazy-tool-transport.ts

@@ -477,10 +477,10 @@ export class LazyToolRuntime {
           typeof args.limit === "number" && Number.isFinite(args.limit)
             ? Math.max(1, Math.min(40, Math.floor(args.limit)))
             : 20;
-        const tools = [...new Set(this.searchableTools(this.currentTools))]
+        const authorizedTools = [...new Set(this.searchableTools(this.currentTools))]
           .filter(isFunctionTool)
-          .filter((tool) => tool.name.startsWith(prefix))
           .sort((a, b) => (a.name < b.name ? -1 : a.name > b.name ? 1 : 0));
+        const tools = authorizedTools.filter((tool) => tool.name.startsWith(prefix));
         const cursorIndex =
           cursor === undefined ? -1 : tools.findIndex((tool) => tool.name === cursor);
         if (cursor !== undefined && cursorIndex === -1) {
@@ -513,17 +513,33 @@ export class LazyToolRuntime {
           }
           descriptors.push(descriptor);
         }
-        return JSON.stringify({
+        const result = {
           tools: descriptors,
           total: tools.length,
           nextCursor: index < tools.length ? descriptors.at(-1)!.name : null,
-          ...(prefix && tools.length === 0
-            ? {
-                message:
-                  "No tool names match this literal prefix. Retry tool_list without namePrefix to browse the authorized catalog; tool names can include a server namespace.",
-              }
-            : {}),
-        });
+        };
+        if (!prefix || tools.length > 0) return JSON.stringify(result);
+
+        // Preserve literal-prefix pagination. Recovery hints are descriptors
+        // from the same authorized deferred pool, never schemas or new grants.
+        const recovery = {
+          ...result,
+          message:
+            "No tool names match this literal prefix. This does not establish capability absence. Load any suggested exact names with tool_search, or retry tool_list without namePrefix to browse the authorized catalog.",
+          suggestions: [] as { name: string; description: string }[],
+        };
+        for (const candidate of authorizedTools) {
+          if (!candidate.name.includes(prefix)) continue;
+          const suggestion = {
+            name: candidate.name,
+            description: Array.from(candidate.description).slice(0, 160).join(""),
+          };
+          const next = { ...recovery, suggestions: [...recovery.suggestions, suggestion] };
+          if (Buffer.byteLength(JSON.stringify(next)) > 16 * 1024) continue;
+          recovery.suggestions.push(suggestion);
+          if (recovery.suggestions.length >= Math.min(limit, 8)) break;
+        }
+        return JSON.stringify(recovery);
       },
     }) as unknown as Tool;
   }

package/src/mcp-network.ts

@@ -17,7 +17,9 @@ export const MCP_MAX_RESPONSE_BYTES = 8 * 1024 * 1024;
 // a lower HTTP limit rejects valid calls before the tool can execute and turns
 // a deterministic transport refusal into apparent outcome uncertainty.
 export const MCP_MAX_INBOUND_REQUEST_BYTES = CODEMODE_ARGUMENTS_MAX_BYTES + 64 * 1024;
-export const MCP_MAX_TOOL_DEFINITION_BYTES = 128 * 1024;
+// Rich nested schemas fit within one bounded discovery/disclosure envelope.
+// Whole-server and aggregate limits still bound the complete catalog.
+export const MCP_MAX_TOOL_DEFINITION_BYTES = 512 * 1024;
 export const MCP_MAX_TOOL_LIST_BYTES = 4 * 1024 * 1024;
 export const MCP_MAX_AGGREGATE_TOOL_LIST_ENTRIES = MCP_MAX_CATALOG_TOOL_ENTRIES;
 // One provider may use the available catalog allowance. The shared budget
@@ -25,7 +27,7 @@ export const MCP_MAX_AGGREGATE_TOOL_LIST_ENTRIES = MCP_MAX_CATALOG_TOOL_ENTRIES;
 // silently excluded otherwise bounded catalogs from best-effort discovery.
 export const MCP_MAX_TOOL_LIST_ENTRIES = MCP_MAX_AGGREGATE_TOOL_LIST_ENTRIES;
 export const MCP_MAX_TOOL_RESULT_BYTES = 1024 * 1024;
-export const MCP_MAX_TOOL_SEARCH_DISCLOSURE_BYTES = 256 * 1024;
+export const MCP_MAX_TOOL_SEARCH_DISCLOSURE_BYTES = MCP_MAX_TOOL_DEFINITION_BYTES + 16 * 1024;
 export const MCP_MAX_SELECTED_SERVERS = 64;
 export const MCP_MAX_CONCURRENT_SERVER_OPERATIONS = 8;
 // @openai/agents has a separate lifecycle fence around MCPServer.connect().

package/src/model-provider-client.ts

@@ -30,7 +30,7 @@ import { recordModelTransportStarted } from "./model-preparation-diagnostics";
 import { captureProviderRequestBody } from "./model-request-capture";
 import { withoutQuotaExhaustedRetries } from "./provider-quota";
 import {
-  observeClaudeUsageResponse,
+  captureClaudeRequestToken,
   prepareClaudeSubscriptionRequest,
 } from "./claude-subscription-usage";
 
@@ -266,6 +266,12 @@ function withoutAuthenticationHeaders(inner: typeof fetch): typeof fetch {
 }
 
 export function buildProviderClient(provider: ResolvedModelProvider, settings: Settings): OpenAI {
+  if (
+    (provider.kind === "direct-openai-workspace" || provider.kind === "direct-azure-workspace") &&
+    !provider.apiKey?.trim()
+  ) {
+    throw new Error("OpenAI or Azure OpenAI workspace key is unavailable");
+  }
   const workspaceGateway = provider.kind === "vercel-gateway-workspace";
   const scopedCredentialProvider =
     provider.credentialSource?.kind === "workspace_connection" ||
@@ -403,7 +409,9 @@ export function instrumentedModelFetch(provider: string, inner: typeof fetch): t
     if (!isModelCallFetch(input)) {
       return await inner(input, init);
     }
-    init = await prepareClaudeSubscriptionRequest(provider, input, init);
+    const prepared = await prepareClaudeSubscriptionRequest(provider, input, init);
+    init = prepared.init;
+    const claudeRequestToken = captureClaudeRequestToken(input, init);
     // The attempt-local observer durably checkpoints provider dispatch before
     // this process can place request bytes on the network.
     await recordModelTransportStarted();
@@ -411,7 +419,7 @@ export function instrumentedModelFetch(provider: string, inner: typeof fetch): t
     const started = performance.now();
     try {
       const response = await inner(input, capture.init);
-      observeClaudeUsageResponse(provider, response);
+      prepared.observe(response, claudeRequestToken);
       recordModelCallMetric(provider, response.ok ? "completed" : "failed", started);
       return response;
     } catch (error) {

package/src/model-provider-routing.ts

@@ -1,3 +1,4 @@
+import { isDirectModelId } from "@opengeni/contracts";
 import type { ConfiguredModel, ResolvedModelProvider, Settings } from "@opengeni/config";
 import { configuredProviders, resolveModelProvider } from "@opengeni/config";
 import {
@@ -334,6 +335,9 @@ export class MultiProviderModelProvider implements ModelProvider {
         throw new XaiSubscriptionUnavailableError(modelName);
       }
     }
+    if (modelName && isDirectModelId(modelName)) {
+      throw new Error("The selected OpenAI or Azure OpenAI connection is unavailable");
+    }
     // Preserve the legacy unlisted-model fallback, but bind it through the same
     // typed request-policy model as every configured Responses call. This keeps
     // Azure wire normalization at the object stage instead of reintroducing a

package/src/runtime-skills.ts

@@ -50,6 +50,7 @@ export type NativeToolSkillSet = Readonly<{
   defaults?: boolean;
   editableArtifacts: boolean;
   sites?: boolean;
+  schedules?: boolean;
   videoGeneration: boolean;
 }>;
 
@@ -121,6 +122,7 @@ export function loadNativeToolSkillArtifacts(
 ): readonly RuntimeSkillArtifact[] {
   const directories: string[] = nativeTools.defaults === false ? [] : ["bundled_default_skills"];
   if (nativeTools.projects) directories.push("bundled_project_skills");
+  if (nativeTools.schedules) directories.push("bundled_schedule_skills");
   if (nativeTools.editableArtifacts) directories.push("bundled_artifact_skills");
   if (nativeTools.sites) directories.push("bundled_site_skills");
   if (nativeTools.videoGeneration) directories.push("bundled_video_skills");
@@ -391,6 +393,12 @@ function nativeToolSkillSources(nativeTools: NativeToolSkillSet): Array<{
       reason: "bundled Site authoring skill",
     });
   }
+  if (nativeTools.schedules) {
+    sources.push({
+      names: skillDirNames(packagedSkillDirectory("bundled_schedule_skills")),
+      reason: "native scheduled-task tool surface",
+    });
+  }
   if (nativeTools.videoGeneration) {
     sources.push({
       names: skillDirNames(packagedSkillDirectory("bundled_video_skills")),

package/src/sandbox/channel-a.ts

@@ -178,6 +178,7 @@ export type ChannelASession = ProviderCommandSession & {
     chars?: string;
     yieldTimeMs?: number;
     maxOutputTokens?: number;
+    signal?: AbortSignal;
   }): Promise<string>;
   writeStdinForProcessMutation?(args: {
     sessionId: number;
@@ -193,6 +194,8 @@ export type ChannelASession = ProviderCommandSession & {
   }): Promise<string>;
   cancelExecCommand?(opId: string): Promise<boolean>;
   hasRetainedProcess?(providerSessionId: number): boolean;
+  /** Cleanup only: forget the local route after exact durable terminal proof. */
+  reconcileRetainedProcess?(providerSessionId: number): Promise<boolean>;
   retainedProcessHasTypedHandleLoss?(providerSessionId: number): boolean;
   execCommandForProcessControl?(providerSessionId: number, args: ChannelAExecArgs): Promise<string>;
   createEditor?(runAs?: string): ChannelAEditor;

package/src/sandbox/index.ts

@@ -2,7 +2,11 @@ export type {
   ProviderCommandPersistence,
   ProviderCommandSession,
 } from "./provider-command-session";
-export { ProviderCommandStartOutcomeUnknownError } from "./provider-command-session";
+export {
+  ProviderCommandStartOutcomeUnknownError,
+  ProviderCommandObservationUnavailableError,
+  isProviderCommandObservationUnavailableError,
+} from "./provider-command-session";
 // @opengeni/runtime/sandbox — the agent-loop-free sandbox leaf.
 //
 // This module is the load-bearing pre-req for the API-direct control plane
@@ -26,6 +30,12 @@ export { ProviderCommandStartOutcomeUnknownError } from "./provider-command-sess
 import type { Settings } from "@opengeni/config";
 import { collectSandboxEnvironment, parseExposedPorts } from "@opengeni/config";
 export type { WorkspaceArchiveSpool, VerifiedHostWorkspaceArchive } from "./archive-spool";
+export { reduceModalRawOutputPage } from "./providers/modal-command-raw-page";
+export type {
+  ModalRawOutputPage,
+  ModalRawOutputStreams,
+  ModalRawOutputExit,
+} from "./providers/modal-command-raw-page";
 import type { WorkspaceArchiveSpool } from "./archive-spool";
 import { restoreHostWorkspaceArchive } from "./host-archive-spool";
 export {