@opengeni/runtime 0.6.1 → 0.7.1

package/dist/sandbox/index.d.ts

@@ -1,6 +1,6 @@
 export { collectSandboxEnvironment, parseExposedPorts } from '@opengeni/config';
 export { CAPABILITY_DESCRIPTORS, CapabilityDescriptor, DESKTOP_STREAM_PORT, StreamTokenPayload, StreamTokenPayload as StreamTokenPayloadType, TERMINAL_STREAM_PORT } from '@opengeni/contracts';
 import '@openai/agents/sandbox';
-export { A as ActiveBackendResolverDeps, a as ActiveBackendUnresolvableError, b as ActivePointer, C as ChannelAConflictError, c as ChannelAEmitter, d as ChannelAExecArgs, e as ChannelAExecResult, f as ChannelANotFoundError, g as ChannelASession, h as ChannelAUnsupportedError, i as ChannelAValidationError, j as ControlRpc, D as DEFAULT_DESKTOP_GEOMETRY, k as DISPLAY_STACK_TIMEOUT_MS, l as DesktopGeometry, m as DisplayStackError, n as DisplayStackUnsupportedError, E as EnsureDisplayStackOptions, o as EnsureDisplayStackResult, p as EnsureTerminalServerOptions, q as EnsureTerminalServerResult, r as EstablishedSandboxSession, s as ExposeStreamPortInput, t as ExposeStreamPortResult, u as ExposedPortEndpoint, F as FinalizeRecordingResult, L as LiveModalSandboxLeaseAttribution, M as MintStreamTokenInput, v as MockAgentResponder, w as MockAgentResponderOptions, x as MockExecHandler, y as ModalModuleLoader, z as ModalOrphanSweepResult, B as ModalSandboxAttribution, N as NatsControlRpc, G as NatsRequestConnection, H as NegotiationContext, I as NumstatEntry, P as PROVIDER_REGISTRY, J as ProviderConstructionContext, K as ProviderRegistration, O as RecordingCodec, Q as RecordingContentType, S as RecordingError, T as RecordingProcess, U as RecordingUnavailableError, V as ResolvedActiveBackend, W as RoutableBackendSession, X as RoutableSandbox, Y as RoutingSandboxSession, Z as RoutingSandboxSessionDeps, _ as RoutingTransitionEvent, $ as RoutingUnsupportedError, a0 as SELFHOSTED_DEFAULT_TIMEOUT_MS, a1 as SELFHOSTED_RECONNECT_WINDOW_MS, a2 as SELFHOSTED_RELAY_STREAM_PATH, a3 as STREAM_PORT, a4 as STREAM_TOKEN_DEFAULT_TTL_SECONDS, a5 as SandboxChannelAService, a6 as SandboxChannelAServiceOptions, a7 as SandboxConfigError, a8 as SandboxCreatedCallback, a9 as SandboxProviderUnavailableError, aa as SelfhostedApplyDiff, ab as SelfhostedControlError, ac as SelfhostedEditor, ad as SelfhostedEnrollment, ae as SelfhostedExecArgs, af as SelfhostedExecResult, ag as SelfhostedImageOutput, ah as SelfhostedLivenessState, ai as SelfhostedNegotiationInput, aj as SelfhostedRelayConfig, ak as SelfhostedSandboxClient, al as SelfhostedSession, am as SelfhostedSessionBuild, an as SelfhostedSessionDeps, ao as SelfhostedSessionState, ap as SelfhostedUnavailableReason, aq as StartRecordingInput, ar as StreamPortUnavailableError, as as TERMINAL_SERVER_TIMEOUT_MS, at as TerminalServerError, au as TerminalServerUnsupportedError, av as agentErrorToControlError, aw as assertDescriptorRegistryInvariants, ax as assertProviderRegistryInvariants, ay as assertSafeRelPath, az as backendSupportsOs, aA as buildDisplayStackScript, aB as buildSelfhostedBackendSession, aC as buildStreamUrl, aD as buildTerminalServerScript, aE as contentTypeForCodec, aF as createSandboxClient, aG as createSandboxClientForBackend, aH as decodeModalSnapshotId, aI as deletePriorPersistedSnapshot, aJ as deleteRecordingArtifacts, aK as deserializeSandboxSessionStateEnvelope, aL as desktopCapableBackend, aM as ensureDisplayStack, aN as ensureModalRegistryImage, aO as ensureTerminalServer, aP as establishSandboxSessionFromEnvelope, aQ as exposeStreamPort, aR as extForCodec, aS as isExecSessionLostBanner, aT as isProviderSandboxNotFoundError, aU as isSelfhostedProviderNotFoundError, aV as isWorkspaceEscapeError, aW as makeActiveBackendResolver, aX as mintStreamToken, aY as modalSandboxAttributionEnvironment, aZ as modalSandboxAttributionTags, a_ as negotiateCapabilities, a$ as negotiateSelfhostedCapabilities, b0 as offlineAgentError, b1 as offlineControlResponse, b2 as parseExecBannerSessionId, b3 as parseNumstatZ, b4 as parsePorcelainV2, b5 as parseUnifiedPatch, b6 as readRecordingBytes, b7 as readWorkspaceArchiveFromEnvelopeSessionState, b8 as recordingStorageKey, b9 as restoredSandboxSessionStateFromEntry, ba as sandboxStateEntryFromRunState, bb as selectBackend, bc as selfhostedLiveness, bd as serializeEstablishedSandboxEnvelope, be as setSelfhostedApplyDiff, bf as startRecording, bg as stopRecording, bh as stripExecBanner, bi as subjectFor, bj as sweepModalOrphanSandboxes, bk as tagModalSandbox, bl as tearDownDisplayStack, bm as tearDownTerminalServer, bn as terminateModalSandboxById, bo as timeoutAgentError, bp as timeoutControlResponse, bq as verifyStreamToken } from '../index-BTo_bZvN.js';
+export { A as ActiveBackendResolverDeps, a as ActiveBackendUnresolvableError, b as ActivePointer, B as BackendUnresolvableCode, C as ChannelAConflictError, c as ChannelAEmitter, d as ChannelAExecArgs, e as ChannelAExecResult, f as ChannelANotFoundError, g as ChannelASession, h as ChannelAUnsupportedError, i as ChannelAValidationError, j as ControlRpc, D as DEFAULT_DESKTOP_GEOMETRY, k as DISPLAY_STACK_TIMEOUT_MS, l as DesktopGeometry, m as DisplayStackError, n as DisplayStackUnsupportedError, E as EnsureDisplayStackOptions, o as EnsureDisplayStackResult, p as EnsureTerminalServerOptions, q as EnsureTerminalServerResult, r as EstablishedSandboxSession, s as ExposeStreamPortInput, t as ExposeStreamPortResult, u as ExposedPortEndpoint, F as FAULT_FIELD_WHAT_HAPPENED, v as FAULT_FIELD_WHAT_PRESERVED, w as FAULT_FIELD_WHAT_TO_TRY, x as FAULT_FIELD_WHICH_LAYER, L as LiveModalSandboxLeaseAttribution, M as MintStreamTokenInput, y as MockAgentResponder, z as MockAgentResponderOptions, G as MockExecHandler, H as ModalModuleLoader, I as ModalOrphanSweepResult, J as ModalSandboxAttribution, N as NEVER_SENT_DETAIL_KEY, K as NatsControlRpc, O as NatsOpStreamConnection, P as NatsOpStreamTransport, Q as NatsRequestConnection, S as NegotiationContext, T as NumstatEntry, U as OP_STREAM_ACK_INTERVAL_MS, V as OP_STREAM_DEFAULT_WINDOW_BYTES, W as OP_STREAM_RECONNECT_HOLD_MS, X as OP_STREAM_SILENCE_TIMEOUT_MS, Y as OpStreamExecOutcome, Z as OpStreamJournal, _ as OpStreamSubscription, $ as OpStreamTransport, a0 as OpStreamUnavailableError, a1 as PROVIDER_REGISTRY, a2 as ProviderConstructionContext, a3 as ProviderRegistration, a4 as REPOSITORY_DISCOVERY_LIMIT, a5 as RecordingArtifactMetadata, a6 as RecordingCodec, a7 as RecordingContentType, a8 as RecordingError, a9 as RecordingProcess, aa as RecordingUnavailableError, ab as RepositoryDiscoveryDegradedReason, ac as RepositoryDiscoveryResult, ad as ResolvedActiveBackend, ae as RoutableBackendSession, af as RoutableSandbox, ag as RoutingSandboxSession, ah as RoutingSandboxSessionDeps, ai as RoutingTransitionEvent, aj as RoutingUnsupportedError, ak as SELFHOSTED_DEFAULT_TIMEOUT_MS, al as SELFHOSTED_DRAINING_MAX_RETRIES, am as SELFHOSTED_EXEC_DRAINING_MAX_RETRIES, an as SELFHOSTED_IDEMPOTENT_READONLY_OPS, ao as SELFHOSTED_INFRASTRUCTURE_FAULT_CLASSES, ap as SELFHOSTED_NEVER_SENT_MAX_RETRIES, aq as SELFHOSTED_RECONNECT_WINDOW_MS, ar as SELFHOSTED_RELAY_STREAM_PATH, as as SELFHOSTED_RETRY_BACKOFF_BASE_MS, at as SELFHOSTED_RETRY_BACKOFF_CAP_MS, au as SELFHOSTED_RETRY_BACKOFF_FACTOR, av as SELFHOSTED_TIMEOUT_MAX_RETRIES, aw as STREAM_PORT, ax as STREAM_TOKEN_DEFAULT_TTL_SECONDS, ay as SandboxChannelAService, az as SandboxChannelAServiceOptions, aA as SandboxConfigError, aB as SandboxCreatedCallback, aC as SandboxProviderUnavailableError, aD as SandboxResumeStateUnavailableError, aE as SelfhostedApplyDiff, aF as SelfhostedControlError, aG as SelfhostedEditor, aH as SelfhostedEnrollment, aI as SelfhostedExecArgs, aJ as SelfhostedExecResult, aK as SelfhostedImageOutput, aL as SelfhostedLivenessState, aM as SelfhostedNegotiationInput, aN as SelfhostedOpObservation, aO as SelfhostedOpObserver, aP as SelfhostedOpStreamDeps, aQ as SelfhostedRelayConfig, aR as SelfhostedRetryClock, aS as SelfhostedRetryDecision, aT as SelfhostedSandboxClient, aU as SelfhostedSession, aV as SelfhostedSessionBuild, aW as SelfhostedSessionDeps, aX as SelfhostedSessionState, aY as SelfhostedUnavailableReason, aZ as StartRecordingInput, a_ as StreamPortUnavailableError, a$ as SwapTargetEstablishability, b0 as TERMINAL_SERVER_TIMEOUT_MS, b1 as TerminalServerError, b2 as TerminalServerUnsupportedError, b3 as agentErrorToControlError, b4 as assertDescriptorRegistryInvariants, b5 as assertProviderRegistryInvariants, b6 as assertSafeRelPath, b7 as backendSupportsOs, b8 as buildDisplayStackScript, b9 as buildSelfhostedBackendSession, ba as buildStreamUrl, bb as buildTerminalServerScript, bc as contentTypeForCodec, bd as createSandboxClient, be as createSandboxClientForBackend, bf as decideSelfhostedRetry, bg as decodeModalSnapshotId, bh as defaultSelfhostedRetryClock, bi as deletePriorPersistedSnapshot, bj as deleteRecordingArtifacts, bk as deserializeSandboxSessionStateEnvelope, bl as desktopCapableBackend, bm as drainingExhaustedError, bn as drainingMessage, bo as ensureDisplayStack, bp as ensureModalRegistryImage, bq as ensureTerminalServer, br as establishSandboxSessionFromEnvelope, bs as execDeadlineHint, bt as exposeStreamPort, bu as extForCodec, bv as inspectRecordingArtifact, bw as isExecSessionLostBanner, bx as isProviderSandboxNotFoundError, by as isSelfhostedProviderNotFoundError, bz as isWorkspaceEscapeError, bA as makeActiveBackendResolver, bB as mintStreamToken, bC as modalSandboxAttributionEnvironment, bD as modalSandboxAttributionTags, bE as negotiateCapabilities, bF as negotiateSelfhostedCapabilities, bG as nextDurableOpId, bH as offlineAgentError, bI as offlineControlResponse, bJ as opAckSubject, bK as opFrameSubject, bL as parseExecBannerSessionId, bM as parseNumstatZ, bN as parsePorcelainV2, bO as parseUnifiedPatch, bP as payloadTooLargeMessage, bQ as readWorkspaceArchiveFromEnvelopeSessionState, bR as recordingStorageKey, bS as renderSelfhostedFault, bT as restoredSandboxSessionStateFromEntry, bU as runWithToolCallCorrelation, bV as sandboxStateEntryFromRunState, bW as sanitizeOpIdToken, bX as selectBackend, bY as selfhostedFaultClass, bZ as selfhostedLiveness, b_ as selfhostedRetryBackoffMs, b$ as serializeEstablishedSandboxEnvelope, c0 as setSelfhostedApplyDiff, c1 as startRecording, c2 as stopRecording, c3 as stripExecBanner, c4 as subjectFor, c5 as swapTargetEstablishability, c6 as sweepModalOrphanSandboxes, c7 as tagModalSandbox, c8 as tearDownDisplayStack, c9 as tearDownTerminalServer, ca as terminateModalSandboxById, cb as timeoutAgentError, cc as timeoutControlResponse, cd as uploadRecordingArtifact, ce as verifyStreamToken } from '../index-DIiiwQY2.js';
 import 'modal';
 import '@opengeni/agent-proto';

package/dist/sandbox/index.js

@@ -10,21 +10,43 @@ import {
   DISPLAY_STACK_TIMEOUT_MS,
   DisplayStackError,
   DisplayStackUnsupportedError,
+  FAULT_FIELD_WHAT_HAPPENED,
+  FAULT_FIELD_WHAT_PRESERVED,
+  FAULT_FIELD_WHAT_TO_TRY,
+  FAULT_FIELD_WHICH_LAYER,
   MockAgentResponder,
+  NEVER_SENT_DETAIL_KEY,
   NatsControlRpc,
+  NatsOpStreamTransport,
+  OP_STREAM_ACK_INTERVAL_MS,
+  OP_STREAM_DEFAULT_WINDOW_BYTES,
+  OP_STREAM_RECONNECT_HOLD_MS,
+  OP_STREAM_SILENCE_TIMEOUT_MS,
+  OpStreamUnavailableError,
   PROVIDER_REGISTRY,
+  REPOSITORY_DISCOVERY_LIMIT,
   RecordingError,
   RecordingUnavailableError,
   RoutingSandboxSession,
   RoutingUnsupportedError,
   SELFHOSTED_DEFAULT_TIMEOUT_MS,
+  SELFHOSTED_DRAINING_MAX_RETRIES,
+  SELFHOSTED_EXEC_DRAINING_MAX_RETRIES,
+  SELFHOSTED_IDEMPOTENT_READONLY_OPS,
+  SELFHOSTED_INFRASTRUCTURE_FAULT_CLASSES,
+  SELFHOSTED_NEVER_SENT_MAX_RETRIES,
   SELFHOSTED_RECONNECT_WINDOW_MS,
   SELFHOSTED_RELAY_STREAM_PATH,
+  SELFHOSTED_RETRY_BACKOFF_BASE_MS,
+  SELFHOSTED_RETRY_BACKOFF_CAP_MS,
+  SELFHOSTED_RETRY_BACKOFF_FACTOR,
+  SELFHOSTED_TIMEOUT_MAX_RETRIES,
   STREAM_PORT,
   STREAM_TOKEN_DEFAULT_TTL_SECONDS,
   SandboxChannelAService,
   SandboxConfigError,
   SandboxProviderUnavailableError,
+  SandboxResumeStateUnavailableError,
   SelfhostedControlError,
   SelfhostedSandboxClient,
   SelfhostedSession,
@@ -47,17 +69,23 @@ import {
   contentTypeForCodec,
   createSandboxClient,
   createSandboxClientForBackend,
+  decideSelfhostedRetry,
   decodeModalSnapshotId,
+  defaultSelfhostedRetryClock,
   deletePriorPersistedSnapshot,
   deleteRecordingArtifacts,
   deserializeSandboxSessionStateEnvelope,
   desktopCapableBackend,
+  drainingExhaustedError,
+  drainingMessage,
   ensureDisplayStack,
   ensureModalRegistryImage,
   ensureTerminalServer,
   establishSandboxSessionFromEnvelope,
+  execDeadlineHint,
   exposeStreamPort,
   extForCodec,
+  inspectRecordingArtifact,
   isExecSessionLostBanner,
   isProviderSandboxNotFoundError,
   isSelfhostedProviderNotFoundError,
@@ -68,26 +96,35 @@ import {
   modalSandboxAttributionTags,
   negotiateCapabilities,
   negotiateSelfhostedCapabilities,
+  nextDurableOpId,
   offlineAgentError,
   offlineControlResponse,
+  opAckSubject,
+  opFrameSubject,
   parseExecBannerSessionId,
   parseExposedPorts,
   parseNumstatZ,
   parsePorcelainV2,
   parseUnifiedPatch,
-  readRecordingBytes,
+  payloadTooLargeMessage,
   readWorkspaceArchiveFromEnvelopeSessionState,
   recordingStorageKey,
+  renderSelfhostedFault,
   restoredSandboxSessionStateFromEntry,
+  runWithToolCallCorrelation,
   sandboxStateEntryFromRunState,
+  sanitizeOpIdToken,
   selectBackend,
+  selfhostedFaultClass,
   selfhostedLiveness,
+  selfhostedRetryBackoffMs,
   serializeEstablishedSandboxEnvelope,
   setSelfhostedApplyDiff,
   startRecording,
   stopRecording,
   stripExecBanner,
   subjectFor,
+  swapTargetEstablishability,
   sweepModalOrphanSandboxes,
   tagModalSandbox,
   tearDownDisplayStack,
@@ -95,8 +132,9 @@ import {
   terminateModalSandboxById,
   timeoutAgentError,
   timeoutControlResponse,
+  uploadRecordingArtifact,
   verifyStreamToken
-} from "../chunk-DYBF7ZEB.js";
+} from "../chunk-RYDCVTMW.js";
 export {
   ActiveBackendUnresolvableError,
   CAPABILITY_DESCRIPTORS,
@@ -109,21 +147,43 @@ export {
   DISPLAY_STACK_TIMEOUT_MS,
   DisplayStackError,
   DisplayStackUnsupportedError,
+  FAULT_FIELD_WHAT_HAPPENED,
+  FAULT_FIELD_WHAT_PRESERVED,
+  FAULT_FIELD_WHAT_TO_TRY,
+  FAULT_FIELD_WHICH_LAYER,
   MockAgentResponder,
+  NEVER_SENT_DETAIL_KEY,
   NatsControlRpc,
+  NatsOpStreamTransport,
+  OP_STREAM_ACK_INTERVAL_MS,
+  OP_STREAM_DEFAULT_WINDOW_BYTES,
+  OP_STREAM_RECONNECT_HOLD_MS,
+  OP_STREAM_SILENCE_TIMEOUT_MS,
+  OpStreamUnavailableError,
   PROVIDER_REGISTRY,
+  REPOSITORY_DISCOVERY_LIMIT,
   RecordingError,
   RecordingUnavailableError,
   RoutingSandboxSession,
   RoutingUnsupportedError,
   SELFHOSTED_DEFAULT_TIMEOUT_MS,
+  SELFHOSTED_DRAINING_MAX_RETRIES,
+  SELFHOSTED_EXEC_DRAINING_MAX_RETRIES,
+  SELFHOSTED_IDEMPOTENT_READONLY_OPS,
+  SELFHOSTED_INFRASTRUCTURE_FAULT_CLASSES,
+  SELFHOSTED_NEVER_SENT_MAX_RETRIES,
   SELFHOSTED_RECONNECT_WINDOW_MS,
   SELFHOSTED_RELAY_STREAM_PATH,
+  SELFHOSTED_RETRY_BACKOFF_BASE_MS,
+  SELFHOSTED_RETRY_BACKOFF_CAP_MS,
+  SELFHOSTED_RETRY_BACKOFF_FACTOR,
+  SELFHOSTED_TIMEOUT_MAX_RETRIES,
   STREAM_PORT,
   STREAM_TOKEN_DEFAULT_TTL_SECONDS,
   SandboxChannelAService,
   SandboxConfigError,
   SandboxProviderUnavailableError,
+  SandboxResumeStateUnavailableError,
   SelfhostedControlError,
   SelfhostedSandboxClient,
   SelfhostedSession,
@@ -146,17 +206,23 @@ export {
   contentTypeForCodec,
   createSandboxClient,
   createSandboxClientForBackend,
+  decideSelfhostedRetry,
   decodeModalSnapshotId,
+  defaultSelfhostedRetryClock,
   deletePriorPersistedSnapshot,
   deleteRecordingArtifacts,
   deserializeSandboxSessionStateEnvelope,
   desktopCapableBackend,
+  drainingExhaustedError,
+  drainingMessage,
   ensureDisplayStack,
   ensureModalRegistryImage,
   ensureTerminalServer,
   establishSandboxSessionFromEnvelope,
+  execDeadlineHint,
   exposeStreamPort,
   extForCodec,
+  inspectRecordingArtifact,
   isExecSessionLostBanner,
   isProviderSandboxNotFoundError,
   isSelfhostedProviderNotFoundError,
@@ -167,26 +233,35 @@ export {
   modalSandboxAttributionTags,
   negotiateCapabilities,
   negotiateSelfhostedCapabilities,
+  nextDurableOpId,
   offlineAgentError,
   offlineControlResponse,
+  opAckSubject,
+  opFrameSubject,
   parseExecBannerSessionId,
   parseExposedPorts,
   parseNumstatZ,
   parsePorcelainV2,
   parseUnifiedPatch,
-  readRecordingBytes,
+  payloadTooLargeMessage,
   readWorkspaceArchiveFromEnvelopeSessionState,
   recordingStorageKey,
+  renderSelfhostedFault,
   restoredSandboxSessionStateFromEntry,
+  runWithToolCallCorrelation,
   sandboxStateEntryFromRunState,
+  sanitizeOpIdToken,
   selectBackend,
+  selfhostedFaultClass,
   selfhostedLiveness,
+  selfhostedRetryBackoffMs,
   serializeEstablishedSandboxEnvelope,
   setSelfhostedApplyDiff,
   startRecording,
   stopRecording,
   stripExecBanner,
   subjectFor,
+  swapTargetEstablishability,
   sweepModalOrphanSandboxes,
   tagModalSandbox,
   tearDownDisplayStack,
@@ -194,6 +269,7 @@ export {
   terminateModalSandboxById,
   timeoutAgentError,
   timeoutControlResponse,
+  uploadRecordingArtifact,
   verifyStreamToken
 };
 //# sourceMappingURL=index.js.map

package/package.json

@@ -1,8 +1,17 @@
 {
   "name": "@opengeni/runtime",
-  "version": "0.6.1",
-  "type": "module",
+  "version": "0.7.1",
   "license": "Apache-2.0",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/Cloudgeni-ai/opengeni.git",
+    "directory": "packages/runtime"
+  },
+  "files": [
+    "dist",
+    "src"
+  ],
+  "type": "module",
   "main": "./dist/index.js",
   "module": "./dist/index.js",
   "types": "./dist/index.d.ts",
@@ -16,32 +25,24 @@
       "import": "./dist/sandbox/index.js"
     }
   },
-  "files": [
-    "dist",
-    "src"
-  ],
   "publishConfig": {
     "access": "public",
     "provenance": true
   },
   "scripts": {
     "build": "tsup",
-    "typecheck": "tsc --noEmit",
+    "typecheck": "tsgo --noEmit",
     "prepublishOnly": "bash ../../scripts/prepublish-guard"
   },
   "dependencies": {
-    "@opengeni/agent-proto": "^0.2.1",
-    "@opengeni/config": "^0.4.0",
-    "@opengeni/contracts": "^0.9.0",
-    "@openai/agents": "^0.11.6",
-    "@openai/agents-extensions": "^0.11.6",
-    "modal": "^0.7.4",
-    "openai": "6.36.0",
-    "@opengeni/codex": "^0.2.1"
-  },
-  "repository": {
-    "type": "git",
-    "url": "git+https://github.com/Cloudgeni-ai/opengeni.git",
-    "directory": "packages/runtime"
+    "@noble/hashes": "^1.8.0",
+    "@openai/agents": "0.13.3",
+    "@openai/agents-extensions": "0.13.3",
+    "@opengeni/agent-proto": "^0.3.0",
+    "@opengeni/codex": "^0.2.3",
+    "@opengeni/config": "^0.5.1",
+    "@opengeni/contracts": "^0.10.0",
+    "modal": "0.7.6",
+    "openai": "6.47.0"
   }
 }

package/src/codex-tool-search.ts

@@ -10,7 +10,7 @@
 // from model context), searches by capability, and only the matched tools are
 // disclosed back and become callable.
 //
-// PROVEN LIVE (Phase 0 probe against /codex/responses on gpt-5.5): the backend
+// PROVEN LIVE (Phase 0 probe against /codex/responses on gpt-5.6-sol): the backend
 // HONORS `defer_loading:true` on function tools — 50 fat tools dropped input_tokens
 // 6766 → 108 — AND makes a tool callable purely because a prior `tool_search_output`
 // disclosed it (V3). The model emits `tool_search_call` on our slug.
@@ -33,13 +33,15 @@ const DEFAULT_SEARCH_LIMIT = 8;
 const MAX_SEARCH_LIMIT = 20;
 
 /** True for a materialized codex_apps connector function tool. */
-export function isCodexAppsFunctionTool(tool: unknown): tool is Tool & { name: string; deferLoading?: boolean } {
+export function isCodexAppsFunctionTool(
+  tool: unknown,
+): tool is Tool & { name: string; deferLoading?: boolean } {
   return (
-    !!tool
-    && typeof tool === "object"
-    && (tool as { type?: unknown }).type === "function"
-    && typeof (tool as { name?: unknown }).name === "string"
-    && (tool as { name: string }).name.startsWith(CODEX_APPS_TOOL_PREFIX)
+    !!tool &&
+    typeof tool === "object" &&
+    (tool as { type?: unknown }).type === "function" &&
+    typeof (tool as { name?: unknown }).name === "string" &&
+    (tool as { name: string }).name.startsWith(CODEX_APPS_TOOL_PREFIX)
   );
 }
 
@@ -47,10 +49,48 @@ export function isCodexAppsFunctionTool(tool: unknown): tool is Tool & { name: s
 // should match on capability words, not drown in glue words (parity with
 // codex-rs, whose search normalizes tokens server-side).
 const STOPWORDS = new Set([
-  "a", "an", "the", "and", "or", "of", "to", "in", "on", "for", "with", "by", "at",
-  "is", "are", "be", "do", "does", "my", "me", "your", "you", "it", "its", "this",
-  "that", "from", "as", "up", "out", "all", "some", "any", "can", "will", "would",
-  "should", "want", "need", "please", "user", "users",
+  "a",
+  "an",
+  "the",
+  "and",
+  "or",
+  "of",
+  "to",
+  "in",
+  "on",
+  "for",
+  "with",
+  "by",
+  "at",
+  "is",
+  "are",
+  "be",
+  "do",
+  "does",
+  "my",
+  "me",
+  "your",
+  "you",
+  "it",
+  "its",
+  "this",
+  "that",
+  "from",
+  "as",
+  "up",
+  "out",
+  "all",
+  "some",
+  "any",
+  "can",
+  "will",
+  "would",
+  "should",
+  "want",
+  "need",
+  "please",
+  "user",
+  "users",
 ]);
 
 /** Light suffix stemmer so "emails"/"email", "creating"/"create" co-match (min-stem guards, no over-stripping). */
@@ -75,9 +115,15 @@ function tokenize(text: string): string[] {
 /** The searchable text of a connector tool: name (weighted ×2) + description + param names. */
 function toolSearchText(tool: Tool): string {
   const raw = (tool as { name?: string }).name ?? "";
-  const name = raw.startsWith(CODEX_APPS_TOOL_PREFIX) ? raw.slice(CODEX_APPS_TOOL_PREFIX.length) : raw;
-  const description = typeof (tool as { description?: unknown }).description === "string" ? (tool as { description: string }).description : "";
-  const params = (tool as { parameters?: { properties?: Record<string, unknown> } }).parameters?.properties;
+  const name = raw.startsWith(CODEX_APPS_TOOL_PREFIX)
+    ? raw.slice(CODEX_APPS_TOOL_PREFIX.length)
+    : raw;
+  const description =
+    typeof (tool as { description?: unknown }).description === "string"
+      ? (tool as { description: string }).description
+      : "";
+  const params = (tool as { parameters?: { properties?: Record<string, unknown> } }).parameters
+    ?.properties;
   const paramNames = params && typeof params === "object" ? Object.keys(params).join(" ") : "";
   return `${name} ${name} ${description} ${paramNames}`;
 }
@@ -127,12 +173,20 @@ export function bm25RankTools(tools: Tool[], query: string, limit: number): Tool
 function parseSearchArgs(raw: unknown): { query: string; limit: number } {
   let obj: Record<string, unknown> = {};
   try {
-    obj = typeof raw === "string" ? (raw.length ? JSON.parse(raw) : {}) : (raw && typeof raw === "object" ? (raw as Record<string, unknown>) : {});
+    obj =
+      typeof raw === "string"
+        ? raw.length
+          ? JSON.parse(raw)
+          : {}
+        : raw && typeof raw === "object"
+          ? (raw as Record<string, unknown>)
+          : {};
   } catch {
     obj = {};
   }
   const query = typeof obj.query === "string" ? obj.query : "";
-  const limitRaw = typeof obj.limit === "number" && Number.isFinite(obj.limit) ? obj.limit : DEFAULT_SEARCH_LIMIT;
+  const limitRaw =
+    typeof obj.limit === "number" && Number.isFinite(obj.limit) ? obj.limit : DEFAULT_SEARCH_LIMIT;
   return { query, limit: Math.max(1, Math.min(MAX_SEARCH_LIMIT, Math.round(limitRaw))) };
 }
 
@@ -147,10 +201,12 @@ function parseSearchArgs(raw: unknown): { query: string; limit: number } {
  */
 export function renderSearchToolDescription(connectorNamespaces: ReadonlySet<string>): string {
   const base =
-    "Search the user's connected app tools by capability. "
-    + "Describe in plain language WHAT you need to do (for example: \"send an email\", \"create a calendar event\") "
-    + "rather than guessing exact tool names. Returns the matching connector tools, which then become callable.";
-  const sources = Array.from(connectorNamespaces).filter((n) => typeof n === "string" && n.length > 0).sort();
+    "Search the user's connected app tools by capability. " +
+    'Describe in plain language WHAT you need to do (for example: "send an email", "create a calendar event") ' +
+    "rather than guessing exact tool names. Returns the matching connector tools, which then become callable.";
+  const sources = Array.from(connectorNamespaces)
+    .filter((n) => typeof n === "string" && n.length > 0)
+    .sort();
   if (sources.length === 0) {
     return `${base}\nConnected sources: none currently available.`;
   }
@@ -160,7 +216,10 @@ export function renderSearchToolDescription(connectorNamespaces: ReadonlySet<str
 const SEARCH_TOOL_PARAMETERS = {
   type: "object",
   properties: {
-    query: { type: "string", description: "Plain-language description of the capability you need." },
+    query: {
+      type: "string",
+      description: "Plain-language description of the capability you need.",
+    },
     limit: { type: "number", description: "Maximum number of tools to return (default 8)." },
   },
   required: ["query"],
@@ -173,7 +232,10 @@ const SEARCH_TOOL_PARAMETERS = {
  * return — the SDK emits the disclosing `tool_search_output` and flips the loaded
  * gate; returning copies would throw). Stateless — reads the pool per call.
  */
-function codexToolSearchExecutor(args: { availableTools?: Tool[]; toolCall?: { arguments?: unknown } }): Tool[] {
+function codexToolSearchExecutor(args: {
+  availableTools?: Tool[];
+  toolCall?: { arguments?: unknown };
+}): Tool[] {
   const deferred = (args.availableTools ?? []).filter(isCodexAppsFunctionTool);
   if (deferred.length === 0) return [];
   const { query, limit } = parseSearchArgs(args.toolCall?.arguments);
@@ -182,16 +244,79 @@ function codexToolSearchExecutor(args: { availableTools?: Tool[]; toolCall?: { a
 
 const NO_NAMESPACES: ReadonlySet<string> = new Set();
 
-/** Build the client-executed tool_search tool that discloses codex_apps connectors on demand. */
-export function buildCodexToolSearchTool(connectorNamespaces: ReadonlySet<string> = NO_NAMESPACES): Tool {
+/** True when the namespace set contains at least one usable (non-empty string) source — mirrors {@link renderSearchToolDescription}'s filter, so it is TRUE exactly when the render produces the "following sources" list and FALSE when it produces "none currently available". */
+function hasConnectorSources(connectorNamespaces: ReadonlySet<string>): boolean {
+  for (const namespace of connectorNamespaces) {
+    if (typeof namespace === "string" && namespace.length > 0) {
+      return true;
+    }
+  }
+  return false;
+}
+
+/**
+ * A per-TURN freeze cell for the tool_search description (AM-8). The connector Set
+ * `installCodexToolSearch` reads is LIVE and by-reference — it fills as the turn's
+ * codex_apps tools/list resolves and can change between a turn's model calls.
+ * Re-rendering the description per call from a changed Set flips the tools block,
+ * and because tools precede the history in the request prefix, ANY change misses the
+ * ENTIRE conversation prefix from that point on — a proven prompt-cache breaker.
+ *
+ * We therefore freeze the description ONCE, at the first model call whose Set has
+ * been populated (connectors discovered), and reuse that frozen string for the rest
+ * of the turn — trading a slightly staler connector list for a byte-stable prefix.
+ * The freeze is created per {@link installCodexToolSearch} (once per turn) and shared
+ * by reference into every clone re-install, so all of a turn's clones agree.
+ */
+export type CodexToolSearchDescriptionFreeze = { value: string | null };
+
+/**
+ * Resolve the tool_search description for this model call, honoring the per-turn
+ * freeze (AM-8):
+ *  - Already frozen ⇒ return the frozen string (byte-stable for the rest of the turn).
+ *  - Not yet frozen + connectors DISCOVERED (non-empty Set) ⇒ render and FREEZE it,
+ *    locking the discovered list against a later same-turn Set mutation.
+ *  - Not yet frozen + Set still EMPTY ⇒ render live ("none currently available") but
+ *    do NOT freeze. This is the load-bearing safety of AM-8: the Set fills lazily and
+ *    best-effort, so an empty Set means "not discovered yet", not "no connectors".
+ *    Freezing "none" here would silently disable the turn's connectors (capability
+ *    loss). Leaving it unfrozen is still byte-stable while the Set stays empty (the
+ *    empty render is a constant), and lets a later call freeze the real list once it
+ *    resolves — at most ONE prefix transition, versus today's per-call churn.
+ * With no freeze cell (the default / direct callers + tests) this is a live render,
+ * byte-for-byte the pre-AM-8 behavior.
+ */
+function resolveSearchToolDescription(
+  connectorNamespaces: ReadonlySet<string>,
+  freeze?: CodexToolSearchDescriptionFreeze,
+): string {
+  if (freeze?.value != null) {
+    return freeze.value;
+  }
+  const rendered = renderSearchToolDescription(connectorNamespaces);
+  if (freeze && hasConnectorSources(connectorNamespaces)) {
+    freeze.value = rendered;
+  }
+  return rendered;
+}
+
+/** Build the client-executed tool_search tool from an already-rendered description. */
+function buildCodexToolSearchToolFromDescription(description: string): Tool {
   return toolSearchTool({
     execution: "client",
-    description: renderSearchToolDescription(connectorNamespaces),
+    description,
     parameters: SEARCH_TOOL_PARAMETERS as unknown as Record<string, unknown>,
     execute: codexToolSearchExecutor as never,
   }) as unknown as Tool;
 }
 
+/** Build the client-executed tool_search tool that discloses codex_apps connectors on demand. */
+export function buildCodexToolSearchTool(
+  connectorNamespaces: ReadonlySet<string> = NO_NAMESPACES,
+): Tool {
+  return buildCodexToolSearchToolFromDescription(renderSearchToolDescription(connectorNamespaces));
+}
+
 /** True for the built-in tool_search tool (so we never add a second one). */
 function isToolSearchTool(tool: unknown): boolean {
   const t = tool as { name?: unknown; providerData?: { type?: unknown } } | null;
@@ -214,16 +339,26 @@ function isToolSearchTool(tool: unknown): boolean {
  * reflects the LIVE connector namespaces, so an empty turn reads
  * "none currently available".
  */
-export function applyCodexToolSearch(tools: Tool[], connectorNamespaces: ReadonlySet<string> = NO_NAMESPACES): Tool[] {
+export function applyCodexToolSearch(
+  tools: Tool[],
+  connectorNamespaces: ReadonlySet<string> = NO_NAMESPACES,
+  descriptionFreeze?: CodexToolSearchDescriptionFreeze,
+): Tool[] {
   for (const tool of tools) {
-    if (isCodexAppsFunctionTool(tool) && (tool as { deferLoading?: boolean }).deferLoading !== true) {
+    if (
+      isCodexAppsFunctionTool(tool) &&
+      (tool as { deferLoading?: boolean }).deferLoading !== true
+    ) {
       (tool as { deferLoading?: boolean }).deferLoading = true;
     }
   }
   if (tools.some(isToolSearchTool)) {
     return tools;
   }
-  return [...tools, buildCodexToolSearchTool(connectorNamespaces)];
+  // AM-8: render through the per-turn freeze so the tools-block prefix stays
+  // byte-stable across a turn's model calls once connectors are discovered.
+  const description = resolveSearchToolDescription(connectorNamespaces, descriptionFreeze);
+  return [...tools, buildCodexToolSearchToolFromDescription(description)];
 }
 
 type CloneCapableAgent = {
@@ -251,15 +386,34 @@ type CloneCapableAgent = {
  * (prepareAgentTools), so by the time the wrapper post-processes getAllTools the
  * current turn's connector sources are known.
  */
-export function installCodexToolSearch(agent: CloneCapableAgent, connectorNamespaces: ReadonlySet<string> = NO_NAMESPACES): void {
+export function installCodexToolSearch(
+  agent: CloneCapableAgent,
+  connectorNamespaces: ReadonlySet<string> = NO_NAMESPACES,
+): void {
+  // ONE freeze cell per install (i.e. per turn), shared BY REFERENCE into every
+  // clone re-install below so the whole turn — the built agent and every clone the
+  // sandbox runtime resolves tools on — agrees on the same frozen description
+  // (AM-8). Created here, at the top-level install, never per clone.
+  installCodexToolSearchWithFreeze(agent, connectorNamespaces, { value: null });
+}
+
+function installCodexToolSearchWithFreeze(
+  agent: CloneCapableAgent,
+  connectorNamespaces: ReadonlySet<string>,
+  descriptionFreeze: CodexToolSearchDescriptionFreeze,
+): void {
   const originalGetAllTools = agent.getAllTools.bind(agent);
   agent.getAllTools = (async (runContext: unknown) =>
-    applyCodexToolSearch(await originalGetAllTools(runContext), connectorNamespaces)) as typeof agent.getAllTools;
+    applyCodexToolSearch(
+      await originalGetAllTools(runContext),
+      connectorNamespaces,
+      descriptionFreeze,
+    )) as typeof agent.getAllTools;
   const originalClone = agent.clone?.bind(agent);
   if (originalClone) {
     const cloneWithToolSearch: NonNullable<CloneCapableAgent["clone"]> = (config: unknown) => {
       const cloned = originalClone(config);
-      installCodexToolSearch(cloned, connectorNamespaces);
+      installCodexToolSearchWithFreeze(cloned, connectorNamespaces, descriptionFreeze);
       return cloned;
     };
     agent.clone = cloneWithToolSearch;