@ponythewhite/base-context 1.0.6 → 1.0.8

package/dist/core/inference-coordinator.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"inference-coordinator.d.ts","sourceRoot":"","sources":["../../src/core/inference-coordinator.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,YAAY,EAAE,kBAAkB,EAAE,QAAQ,EAAE,MAAM,kCAAkC,CAAC;AACnG,OAAO,EACN,KAAK,GAAG,EACR,KAAK,gBAAgB,EAErB,KAAK,OAAO,EAKZ,KAAK,KAAK,EAGV,KAAK,yBAAyB,EAC9B,KAAK,mBAAmB,EAExB,MAAM,+BAA+B,CAAC;AACvC,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,mBAAmB,CAAC;AACzD,OAAO,KAAK,EACX,gBAAgB,EAEhB,+BAA+B,EAC/B,wCAAwC,EACxC,gCAAgC,EAGhC,eAAe,EACf,cAAc,EAGd,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EAKN,KAAK,iBAAiB,EACtB,KAAK,uBAAuB,EAC5B,KAAK,mBAAmB,EAExB,MAAM,6BAA6B,CAAC;AAErC,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,4BAA4B,CAAC;AAQzE,MAAM,WAAW,uBAAuB;IACvC,QAAQ,CAAC,OAAO,EAAE,cAAc,CAAC;IACjC,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,iBAAiB,CAAC,EAAE,MAAM,CAAC;IACpC,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,CAAC;CACjC;AAED,MAAM,WAAW,mBAAmB;IACnC,QAAQ,CAAC,OAAO,EAAE,gBAAgB,CAAC;IACnC,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAC7B,QAAQ,CAAC,UAAU,EAAE,SAAS,MAAM,EAAE,CAAC;IACvC,6EAA6E;IAC7E,QAAQ,CAAC,gBAAgB,EAAE,UAAU,GAAG,eAAe,GAAG,kBAAkB,CAAC;CAC7E;AAED,6FAA6F;AAC7F,MAAM,WAAW,6BAA6B;IAC7C,QAAQ,CAAC,IAAI,EAAE,gBAAgB,CAAC;IAChC,QAAQ,CAAC,MAAM,EAAE,wCAAwC,CAAC;CAC1D;AAYD,MAAM,WAAW,YAAY;IAC5B,QAAQ,CAAC,MAAM,EAAE,OAAO,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC,CAAC;IAC/C,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAC,mBAAmB,CAAC,CAAC;CAC/C;AAED,MAAM,WAAW,sBAAsB;IACtC,QAAQ,CAAC,QAAQ,EAAE,oBAAoB,CAAC;CACxC;AAED,UAAU,sBAAsB;IAC/B,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,EAAE,MAAM,CAAC;IACpB,eAAe,EAAE,MAAM,CAAC;CACxB;AAgBD,sFAAsF;AACtF,eAAO,MAAM,+BAA+B,eAAoD,CAAC;AACjG,oGAAoG;AACpG,eAAO,MAAM,2BAA2B,eAAgD,CAAC;AACzF,wFAAwF;AACxF,eAAO,MAAM,4BAA4B,eAAiD,CAAC;AAC3F,wDAAwD;AACxD,eAAO,MAAM,6BAA6B,eAAkD,CAAC;AA8B7F,kHAAkH;AAClH,wBAAgB,6BAA6B,CAAC,OAAO,EAAE,gBAAgB,GAAG,eAAe,GAAG,SAAS,CAIpG;AAED,gGAAgG;AAChG,wBAAgB,2BAA2B,CAC1C,cAAc,EAAE,CACf,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,KACpC,OAAO,CAAC,mBAAmB,GAAG;IAAE,eAAe,CAAC,EAAE,eAAe,CAAA;CAAE,CAAC,GACvE,QAAQ,CAqBV;AAED,mFAAmF;AACnF,qBAAa,oBAAoB;;IAqI/B,OAAO,CAAC,QAAQ,CAAC,QAAQ;IACzB,OAAO,CAAC,QAAQ,CAAC,KAAK;IAjIvB,OAAO,CAAC,SAAS,CAAS;IAC1B,OAAO,CAAC,mBAAmB,CAAC,CAA8B;IAC1D,OAAO,CAAC,MAAM,CAA4B;IAC1C,OAAO,CAAC,YAAY,CAAC,CAAU;IAC/B,OAAO,CAAC,eAAe,CAAS;IAChC,OAAO,CAAC,oBAAoB,CAAC,CAAa;IAC1C,OAAO,CAAC,QAAQ,CAAS;IACzB,OAAO,CAAC,QAAQ,CAAC,CAAgB;IACjC,OAAO,CAAC,IAAI,CAgBV;IAEF,yBAAyB,CAAC,MAAM,EAAE,MAAM,sBAAsB,GAAG,IAAI,CAEpE;IAED,IAAI,UAAU,IAAI,OAAO,CAExB;IACD,IAAI,YAAY,IAAI,MAAM,CAEzB;IAED,gBAAgB,CAAC,QAAQ,EAAE,MAAM,IAAI,GAAG,MAAM,IAAI,CAGjD;IAED,OAAO,CAAC,cAAc;IAWtB,2GAA2G;IAC3G,gBAAgB,CAAC,OAAO,EAAE,gBAAgB,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC,GAAG,SAAS,CAIlG;IAED,2GAA2G;IAC3G,CAAC,+BAA+B,CAAC,IAAI,oBAAoB,CAKxD;IAED,CAAC,2BAA2B,CAAC,CAAC,IAAI,EAAE,gBAAgB,GAAG,oBAAoB,CAK1E;IAED,gBAAgB,CAAC,UAAU,EAAE,OAAO,CAAC,gBAAgB,CAAC,GAAG,+BAA+B,GAAG,SAAS,CAKnG;IAED,CAAC,4BAA4B,CAAC,CAAC,IAAI,EAAE,gBAAgB,GAAG,oBAAoB,CAK3E;IAED,CAAC,6BAA6B,CAAC,IAAI,oBAAoB,CAItD;IAED,iBAAiB,CAAC,UAAU,EAAE,OAAO,CAAC,gBAAgB,CAAC,GAAG,gCAAgC,GAAG,SAAS,CAKrG;IAED,mGAAmG;IACnG,oBAAoB,CAAC,UAAU,EAAE,OAAO,CAAC,gBAAgB,CAAC,GAAG,6BAA6B,GAAG,SAAS,CAIrG;IAED,6FAA6F;IAC7F,aAAa,IAAI,IAAI,CAIpB;IAEK,WAAW,IAAI,OAAO,CAAC,IAAI,CAAC,CAEjC;IAED,gBAAgB,IAAI,IAAI,CAEvB;IACD,cAAc,IAAI,IAAI,CAErB;IAED,YACkB,QAAQ,EAAE,MAAM,gBAAgB,EAChC,KAAK,GAAE,MAAM,IAAI,CAAC,eAAe,EAAE,WAAW,CAAc,EAC7E,kBAAkB,CAAC,EAAE,yBAAyB,EAM9C;IAED,iFAAiF;IACjF,4BAA4B,IAAI,yBAAyB,GAAG,SAAS,CAEpE;IAED,OAAO,CAAC,eAAe;IAIvB,OAAO,CAAC,OAAO;IASf,OAAO,CAAC,cAAc;IAOtB,oFAAoF;IACpF,OAAO,CAAC,IAAI,CAAC,EAAE,gBAAgB,GAAG,oBAAoB,CA6BrD;IAED,kFAAkF;IAClF,uBAAuB,CACtB,QAAQ,EAAE,SAAS,YAAY,EAAE,EACjC,MAAM,EAAE,iBAAiB,EACzB,QAAQ,CAAC,EAAE,mBAAmB,EAC9B,YAAY,CAAC,EAAE,uBAAuB,GACpC,IAAI,CAKN;IAED,0FAA0F;IACpF,WAAW,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,IAAI,EAAE,sBAAsB,KAAK,OAAO,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,CAAC,CAAC,CAmBnF;IAED,2EAA2E;IAC3E,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAYvB;IAED,OAAO,CAAC,aAAa;YAcP,UAAU;IAwBxB,4EAA4E;IAC5E,OAAO,CAAC,UAAU;IA6BlB,OAAO,CAAC,oBAAoB;YAmBd,OAAO;IA2Tf,KAAK,CACV,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,EACxC,OAAO,EAAE,uBAAuB,GAC9B,OAAO,CAAC,YAAY,CAAC,CAUvB;YAEa,4BAA4B;IA0C1C,QAAQ,CACP,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,EACxC,OAAO,EAAE,uBAAuB,GAC9B,OAAO,CAAC,gBAAgB,CAAC,CAkB3B;IAED,mBAAmB,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,uBAAuB,GAAG,yBAAyB,CAuBnG;IAED,UAAU,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,uBAAuB,GAAG,kBAAkB,CA4CnF;CACD;AAED,iGAAiG;AACjG,wBAAgB,iBAAiB,CAChC,QAAQ,EAAE,oBAAoB,GAAG,SAAS,EAC1C,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,EACxC,OAAO,EAAE,uBAAuB,GAC9B,OAAO,CAAC,gBAAgB,CAAC,CAE3B;AAED,MAAM,MAAM,yBAAyB,GAAG,QAAQ,GAAG;IAClD,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC;IACzB,sBAAsB,CAAC,CAAC,OAAO,EAAE,gBAAgB,EAAE,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;CAC3F,CAAC;AAEF,yFAAyF;AACzF,wBAAgB,4BAA4B,CAC3C,QAAQ,EAAE,QAAQ,EAClB,OAAO,EAAE,uBAAuB,GAC9B,yBAAyB,CAM3B","sourcesContent":["import { randomUUID } from \"node:crypto\";\nimport { setTimeout as delay } from \"node:timers/promises\";\nimport type { AgentMessage, AgentOwnedStreamFn, StreamFn } from \"@ponythewhite/base-context-agent\";\nimport {\n\ttype Api,\n\ttype AssistantMessage,\n\tassertBuiltInAttemptSupport,\n\ttype Context,\n\tcompleteSimple,\n\tinvokeRequestMeasurement,\n\tisLocalFauxStream,\n\tisTransientProviderFailure,\n\ttype Model,\n\ttype ProviderAttemptObserver,\n\tRequestTokenBudget,\n\ttype RequestTokenBudgetOptions,\n\ttype SimpleStreamOptions,\n\tstreamSimple,\n} from \"@ponythewhite/base-context-ai\";\nimport type { AuthSourceToken } from \"./auth-storage.js\";\nimport type {\n\tBoundRequestSink,\n\tContextEpochEntryRef,\n\tNativeBranchRequestOutputWriter,\n\tNativeCompactionRequestOutputAssociation,\n\tNativePlannerRequestOutputWriter,\n\tNativeRequestMetadata,\n\tNativeRequestOutputWriter,\n\tRequestOwnerRef,\n\tRequestPurpose,\n\tResolvedModelContract,\n\tSourceSnapshotRef,\n} from \"./request-events.js\";\nimport {\n\ttype CapturedRequestViewBoundary,\n\tcaptureRequestViewBoundary,\n\tmatchesRequestView,\n\tprepareFixedRequestView,\n\ttype RequestViewCommit,\n\ttype RequestViewFixedPrepare,\n\ttype RequestViewValidate,\n\tselectRequestView,\n} from \"./request-view-selection.js\";\nimport { hashTurnBody, MODEL_REQUEST_ID_HEADER, unwrapSemanticEdgeStreamFn } from \"./semantic-edges.js\";\nimport type { SessionHistoryReadView } from \"./session-history-index.js\";\nimport {\n\tcaptureNativeBranchOutputSource,\n\tcaptureNativeCompactionOutputSource,\n\tcaptureNativePlannerOutputSource,\n\tcaptureNativeRequestOutputSource,\n} from \"./session-manager.js\";\n\nexport interface InferenceRequestOptions {\n\treadonly purpose: RequestPurpose;\n\treadonly purposeDetail?: string;\n\treadonly operationId?: string;\n\treadonly parentOperationId?: string;\n\treadonly semanticEdgeId?: string;\n}\n\nexport interface InferenceSettlement {\n\treadonly message: AssistantMessage;\n\treadonly operationId: string;\n\treadonly attemptIds: readonly string[];\n\t/** Reporting is not a guarantee about an opaque adapter's hidden retries. */\n\treadonly physicalCoverage: \"reported\" | \"unestablished\" | \"local-simulation\";\n}\n\n/** Private completion identity carries this actual source, never descriptive sink labels. */\nexport interface NativeCompactionOutputBinding {\n\treadonly sink: BoundRequestSink;\n\treadonly output: NativeCompactionRequestOutputAssociation;\n}\n\ninterface NativeBranchOutputBinding {\n\treadonly owner: InferenceCoordinator;\n\treadonly write: NativeBranchRequestOutputWriter;\n}\n\ninterface NativePlannerOutputBinding {\n\treadonly owner: InferenceCoordinator;\n\treadonly write: NativePlannerRequestOutputWriter;\n}\n\nexport interface InferenceRun {\n\treadonly events: Awaited<ReturnType<StreamFn>>;\n\treadonly settled: Promise<InferenceSettlement>;\n}\n\nexport interface SessionRuntimeServices {\n\treadonly requests: InferenceCoordinator;\n}\n\ninterface ProviderRecoveryPolicy {\n\tenabled: boolean;\n\tbaseDelayMs: number;\n\tmaxRetryDelayMs: number;\n}\n\ninterface SinkUse {\n\treadonly sink: BoundRequestSink;\n\tactiveUsers: number;\n\tcaptures: number;\n\trelease?: Promise<void>;\n}\n\ninterface BoundOperation {\n\treadonly binding: SinkUse;\n\treadonly owner: Omit<RequestOwnerRef, \"sessionId\">;\n\treadonly metadata: Omit<NativeRequestMetadata, \"source\" | \"owner\" | \"contextEpoch\">;\n}\n\nconst REQUEST_STREAM_BINDING = Symbol(\"base-context.request-stream-binding\");\n/** Internal AgentSession built-in compaction entry, not a public inference option. */\nexport const captureNativeCompactionRequests = Symbol(\"base-context.native-compaction-requests\");\n/** Internal AgentSession built-in branch-summary invocation, never inherited by generic capture. */\nexport const captureNativeBranchRequests = Symbol(\"base-context.native-branch-requests\");\n/** Internal native AgentSession planner capture, never inherited by generic capture. */\nexport const captureNativePlannerRequests = Symbol(\"base-context.native-planner-requests\");\n/** Internal native AgentSession reviewer invocation. */\nexport const captureNativeReviewerRequests = Symbol(\"base-context.native-reviewer-requests\");\ninterface StreamBinding {\n\treadonly coordinator: InferenceCoordinator;\n\treadonly inner: StreamFn;\n}\ntype BoundStreamFn = AgentOwnedStreamFn & { [REQUEST_STREAM_BINDING]?: StreamBinding };\n\nfunction modelContract(model: Model<Api>): ResolvedModelContract {\n\treturn {\n\t\tapi: model.api,\n\t\tprovider: model.provider,\n\t\tmodel: model.id,\n\t\tprofile: { id: \"native-default\", status: \"unvalidated\" },\n\t\tpricing: {\n\t\t\tstatus: \"unvalidated\",\n\t\t\tcurrency: \"USD\",\n\t\t\tunit: \"million-tokens\",\n\t\t\tcatalogRates: {\n\t\t\t\tinput: model.cost.input,\n\t\t\t\toutput: model.cost.output,\n\t\t\t\tcacheRead: model.cost.cacheRead,\n\t\t\t\tcacheWrite: model.cost.cacheWrite,\n\t\t\t},\n\t\t},\n\t};\n}\n\nconst nativeStreamFunctions = new WeakSet<StreamFn>();\nconst nativeAuthSources = new WeakMap<AssistantMessage, AuthSourceToken>();\n\n/** Consume the credential source resolved for this exact native request. Never persisted or sent to providers. */\nexport function takeNativeInferenceAuthSource(message: AssistantMessage): AuthSourceToken | undefined {\n\tconst source = nativeAuthSources.get(message);\n\tnativeAuthSources.delete(message);\n\treturn source;\n}\n\n/** Internal SDK auth/options resolver; transport stays the known instrumented AI dispatcher. */\nexport function createNativeInferenceStream(\n\tresolveOptions: (\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t) => Promise<SimpleStreamOptions & { authSourceToken?: AuthSourceToken }>,\n): StreamFn {\n\tconst dispatch: StreamFn = async (model, context, options) => {\n\t\tconst { authSourceToken, ...resolved } = await resolveOptions(model, context, options);\n\t\tconst events = streamSimple(model, context, {\n\t\t\t...resolved,\n\t\t\tsignal: options?.signal,\n\t\t\tattempts: options?.attempts,\n\t\t\trequireProviderAttempts: true,\n\t\t});\n\t\tif (authSourceToken) {\n\t\t\tconst source = { ...authSourceToken };\n\t\t\tconst result = events.result().then((message) => {\n\t\t\t\tnativeAuthSources.set(message, source);\n\t\t\t\treturn message;\n\t\t\t});\n\t\t\tevents.result = () => result;\n\t\t}\n\t\treturn events;\n\t};\n\tnativeStreamFunctions.add(dispatch);\n\treturn dispatch;\n}\n\n/** Routes physical observations to the source writer. It owns no receipt store. */\nexport class InferenceCoordinator {\n\t#nativeCompactionOutput = false;\n\t#nativeBranchOutput = false;\n\t#nativePlannerOutput = false;\n\t#auxiliaryAttempts?: { remaining: number };\n\tprivate retryTurn = false;\n\tprivate requestViewBoundary?: CapturedRequestViewBoundary;\n\tprivate active = new Set<Promise<void>>();\n\tprivate capturedSink?: SinkUse;\n\tprivate captureReserved = false;\n\tprivate finishCapturePending?: () => void;\n\tprivate disposed = false;\n\tprivate disposal?: Promise<void>;\n\tprivate work = {\n\t\tpending: new Set<Promise<void>>(),\n\t\tsinks: new WeakMap<BoundRequestSink, SinkUse>(),\n\t\tmainOutputs: new WeakMap<AssistantMessage, NativeRequestOutputWriter>(),\n\t\tcompactionSettlements: new WeakMap<InferenceSettlement, NativeCompactionOutputBinding>(),\n\t\tcompactionCompletions: new WeakMap<Promise<AssistantMessage>, NativeCompactionOutputBinding>(),\n\t\tbranchSettlements: new WeakMap<InferenceSettlement, NativeBranchOutputBinding>(),\n\t\tbranchCompletions: new WeakMap<Promise<AssistantMessage>, NativeBranchOutputBinding>(),\n\t\tplannerSettlements: new WeakMap<InferenceSettlement, NativePlannerOutputBinding>(),\n\t\tplannerCompletions: new WeakMap<Promise<AssistantMessage>, NativePlannerOutputBinding>(),\n\t\tlisteners: new Set<() => void>(),\n\t\tadmissionOpen: true,\n\t\tcancellation: new AbortController(),\n\t\tbudget: undefined as RequestTokenBudget | undefined,\n\t\tbudgetMode: undefined as RequestTokenBudgetOptions[\"mode\"] | undefined,\n\t\tproviderRecoveryPolicy: undefined as (() => ProviderRecoveryPolicy) | undefined,\n\t};\n\n\tsetProviderRecoveryPolicy(policy: () => ProviderRecoveryPolicy): void {\n\t\tthis.work.providerRecoveryPolicy = policy;\n\t}\n\n\tget hasPending(): boolean {\n\t\treturn this.work.pending.size > 0;\n\t}\n\tget pendingCount(): number {\n\t\treturn this.work.pending.size;\n\t}\n\n\tonActivityChange(listener: () => void): () => void {\n\t\tthis.work.listeners.add(listener);\n\t\treturn () => this.work.listeners.delete(listener);\n\t}\n\n\tprivate notifyActivity(): void {\n\t\t// Observers can wake an existing owner, but cannot own request settlement.\n\t\tfor (const listener of this.work.listeners) {\n\t\t\ttry {\n\t\t\t\tlistener();\n\t\t\t} catch {\n\t\t\t\t/* Notification is not execution. */\n\t\t\t}\n\t\t}\n\t}\n\n\t/** Consume only this owner's real returned object. Labels and replacement messages remain unassociated. */\n\tappendMainOutput(message: AssistantMessage, owner: object): Promise<string | undefined> | undefined {\n\t\tconst write = this.work.mainOutputs.get(message);\n\t\tthis.work.mainOutputs.delete(message);\n\t\treturn write?.(owner, message);\n\t}\n\n\t/** Opt this native invocation into output association; ordinary captures do not inherit that authority. */\n\t[captureNativeCompactionRequests](): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this);\n\t\tcaptured.#nativeCompactionOutput = true;\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 4 };\n\t\treturn captured;\n\t}\n\n\t[captureNativeBranchRequests](sink: BoundRequestSink): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this, sink);\n\t\tcaptured.#nativeBranchOutput = true;\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 2 };\n\t\treturn captured;\n\t}\n\n\ttakeBranchOutput(completion: Promise<AssistantMessage>): NativeBranchRequestOutputWriter | undefined {\n\t\tconst binding = this.work.branchCompletions.get(completion);\n\t\tthis.work.branchCompletions.delete(completion);\n\t\tif (!binding || !this.#nativeBranchOutput || binding.owner !== this) return undefined;\n\t\treturn binding.write;\n\t}\n\n\t[captureNativePlannerRequests](sink: BoundRequestSink): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this, sink);\n\t\tcaptured.#nativePlannerOutput = true;\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 2 };\n\t\treturn captured;\n\t}\n\n\t[captureNativeReviewerRequests](): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this);\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 2 };\n\t\treturn captured;\n\t}\n\n\ttakePlannerOutput(completion: Promise<AssistantMessage>): NativePlannerRequestOutputWriter | undefined {\n\t\tconst binding = this.work.plannerCompletions.get(completion);\n\t\tthis.work.plannerCompletions.delete(completion);\n\t\tif (!binding || !this.#nativePlannerOutput || binding.owner !== this) return undefined;\n\t\treturn binding.write;\n\t}\n\n\t/** Only the original completion promise and captured source can supply a projected-output link. */\n\ttakeCompactionOutput(completion: Promise<AssistantMessage>): NativeCompactionOutputBinding | undefined {\n\t\tconst binding = this.work.compactionCompletions.get(completion);\n\t\tthis.work.compactionCompletions.delete(completion);\n\t\treturn binding?.sink === this.capturedSink?.sink ? binding : undefined;\n\t}\n\n\t/** Owner teardown stops new sends, but never suppresses an admitted request's settlement. */\n\tstopAdmission(): void {\n\t\tthis.work.admissionOpen = false;\n\t\tthis.work.cancellation.abort();\n\t\tthis.notifyActivity();\n\t}\n\n\tasync waitForIdle(): Promise<void> {\n\t\twhile (this.work.pending.size > 0) await Promise.all([...this.work.pending]);\n\t}\n\n\tprepareTurnRetry(): void {\n\t\tthis.retryTurn = true;\n\t}\n\tclearTurnRetry(): void {\n\t\tthis.retryTurn = false;\n\t}\n\n\tconstructor(\n\t\tprivate readonly bindSink: () => BoundRequestSink,\n\t\tprivate readonly owner: () => Omit<RequestOwnerRef, \"sessionId\"> = () => ({}),\n\t\trequestTokenBudget?: RequestTokenBudgetOptions,\n\t) {\n\t\tif (requestTokenBudget) {\n\t\t\tthis.work.budget = new RequestTokenBudget(requestTokenBudget);\n\t\t\tthis.work.budgetMode = requestTokenBudget.mode;\n\t\t}\n\t}\n\n\t/** Copy the configured policy, not this owner's learned request observations. */\n\tgetRequestTokenBudgetOptions(): RequestTokenBudgetOptions | undefined {\n\t\treturn this.work.budget?.getOptions();\n\t}\n\n\tprivate assertAdmission(): void {\n\t\tif (this.disposed || !this.work.admissionOpen) throw new Error(\"Inference owner is closing\");\n\t}\n\n\tprivate sinkUse(sink: BoundRequestSink): SinkUse {\n\t\tlet binding = this.work.sinks.get(sink);\n\t\tif (!binding) {\n\t\t\tbinding = { sink, activeUsers: 0, captures: 0 };\n\t\t\tthis.work.sinks.set(sink, binding);\n\t\t}\n\t\treturn binding;\n\t}\n\n\tprivate releaseCapture(): void {\n\t\tif (this.captureReserved && this.capturedSink) {\n\t\t\tthis.captureReserved = false;\n\t\t\tthis.capturedSink.captures--;\n\t\t}\n\t}\n\n\t/** Capture an auxiliary operation's subject before asynchronous auth or UI work. */\n\tcapture(sink?: BoundRequestSink): InferenceCoordinator {\n\t\tthis.assertAdmission();\n\t\tconst owner = Object.freeze({ ...this.owner() });\n\t\tconst binding = this.sinkUse(sink ?? this.bindSink());\n\t\tbinding.sink.retain();\n\t\tbinding.release = undefined;\n\t\tbinding.captures++;\n\t\tconst captured = new InferenceCoordinator(\n\t\t\t() => binding.sink,\n\t\t\t() => owner,\n\t\t);\n\t\tcaptured.work = this.work;\n\t\t// Nested captures share the invocation's allowance, not its native output authority.\n\t\tcaptured.#auxiliaryAttempts = this.#auxiliaryAttempts;\n\t\tcaptured.capturedSink = binding;\n\t\tcaptured.captureReserved = true;\n\t\tlet release!: () => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\trelease = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tcaptured.finishCapturePending = () => {\n\t\t\tthis.work.pending.delete(pending);\n\t\t\tcaptured.finishCapturePending = undefined;\n\t\t\trelease();\n\t\t\tthis.notifyActivity();\n\t\t};\n\t\tthis.notifyActivity();\n\t\treturn captured;\n\t}\n\n\t/** Explicit Root-owned epoch boundary, on this same captured MAIN source only. */\n\tbindRequestViewBoundary(\n\t\tmessages: readonly AgentMessage[],\n\t\tcommit: RequestViewCommit,\n\t\tvalidate?: RequestViewValidate,\n\t\tfixedPrepare?: RequestViewFixedPrepare,\n\t): void {\n\t\tthis.assertAdmission();\n\t\tif (!this.capturedSink) throw new Error(\"Request view boundary requires a captured inference owner\");\n\t\tif (this.requestViewBoundary) throw new Error(\"Request view boundary is already bound\");\n\t\tthis.requestViewBoundary = captureRequestViewBoundary(messages, commit, validate, fixedPrepare);\n\t}\n\n\t/** Read through an explicit capture, never by recapturing the mutable current session. */\n\tasync readHistory<T>(read: (view: SessionHistoryReadView) => Promise<T>): Promise<T> {\n\t\tthis.assertAdmission();\n\t\tconst sink = this.capturedSink?.sink;\n\t\tif (!sink?.readHistory) throw new Error(\"Captured canonical history is unavailable\");\n\t\tlet finish!: () => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\tfinish = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tthis.active.add(pending);\n\t\tthis.notifyActivity();\n\t\ttry {\n\t\t\treturn await sink.readHistory(read);\n\t\t} finally {\n\t\t\tthis.active.delete(pending);\n\t\t\tthis.work.pending.delete(pending);\n\t\t\tfinish();\n\t\t\tthis.notifyActivity();\n\t\t}\n\t}\n\n\t/** Release an unused capture, or join this capture's admitted requests. */\n\tdispose(): Promise<void> {\n\t\tthis.disposed = true;\n\t\tthis.disposal ??= (async () => {\n\t\t\tthis.releaseCapture();\n\t\t\ttry {\n\t\t\t\tawait Promise.all([...this.active]);\n\t\t\t\tif (this.capturedSink) await this.finishSink(this.capturedSink, [], false, []);\n\t\t\t} finally {\n\t\t\t\tthis.finishCapturePending?.();\n\t\t\t}\n\t\t})();\n\t\treturn this.disposal;\n\t}\n\n\tprivate bindOperation(model: Model<Api>, request: InferenceRequestOptions, operationId: string): BoundOperation {\n\t\tthis.assertAdmission();\n\t\tconst owner = Object.freeze({ ...this.owner() });\n\t\tconst metadata = Object.freeze({\n\t\t\toperationId,\n\t\t\tparentOperationId: request.parentOperationId,\n\t\t\tsemanticEdgeId: request.semanticEdgeId,\n\t\t\tpurpose: request.purpose,\n\t\t\tpurposeDetail: request.purposeDetail,\n\t\t\tmodelContract: modelContract(model),\n\t\t});\n\t\treturn { binding: this.sinkUse(this.bindSink()), owner, metadata };\n\t}\n\n\tprivate async finishSink(\n\t\tbinding: SinkUse,\n\t\twrites: readonly Promise<void>[],\n\t\tactiveUser: boolean,\n\t\tfailures: unknown[],\n\t): Promise<void> {\n\t\tconst results = await Promise.allSettled([binding.sink.source, ...writes]);\n\t\tfor (const result of results) {\n\t\t\tif (result.status === \"rejected\") failures.push(result.reason);\n\t\t}\n\t\tif (activeUser) binding.activeUsers--;\n\t\ttry {\n\t\t\tif (binding.activeUsers === 0 && binding.captures === 0) {\n\t\t\t\tbinding.release ??= binding.sink.release();\n\t\t\t\tawait binding.release;\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tfailures.push(error);\n\t\t}\n\t\tconst distinct = [...new Set(failures)];\n\t\tif (distinct.length === 1) throw distinct[0];\n\t\tif (distinct.length > 1) throw new AggregateError(distinct, \"Inference request or source cleanup failed\");\n\t}\n\n\t/** Hold one logical operation through settled sends and recovery delays. */\n\tprivate retainWork(binding: SinkUse): (failures?: unknown[]) => Promise<void> {\n\t\tthis.assertAdmission();\n\t\tbinding.sink.retain();\n\t\tbinding.release = undefined;\n\t\tbinding.captures++;\n\t\tlet finish!: () => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\tfinish = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tthis.active.add(pending);\n\t\tthis.notifyActivity();\n\t\tlet release: Promise<void> | undefined;\n\t\treturn (failures = []) => {\n\t\t\trelease ??= (async () => {\n\t\t\t\tbinding.captures--;\n\t\t\t\ttry {\n\t\t\t\t\tawait this.finishSink(binding, [], false, failures);\n\t\t\t\t} finally {\n\t\t\t\t\tthis.active.delete(pending);\n\t\t\t\t\tthis.work.pending.delete(pending);\n\t\t\t\t\tfinish();\n\t\t\t\t\tthis.notifyActivity();\n\t\t\t\t}\n\t\t\t})();\n\t\t\treturn release;\n\t\t};\n\t}\n\n\tprivate providerRecoveryWait(): (message: AssistantMessage, signal?: AbortSignal) => Promise<boolean> {\n\t\tlet nextDelayMs: number | undefined;\n\t\treturn async (message, callerSignal) => {\n\t\t\tconst policy = this.work.providerRecoveryPolicy?.();\n\t\t\tif (!policy?.enabled || !isTransientProviderFailure(message)) return false;\n\t\t\tthis.assertAdmission();\n\t\t\tconst signal = callerSignal\n\t\t\t\t? AbortSignal.any([callerSignal, this.work.cancellation.signal])\n\t\t\t\t: this.work.cancellation.signal;\n\t\t\tsignal.throwIfAborted();\n\t\t\tconst delayMs = Math.min(nextDelayMs ?? policy.baseDelayMs, policy.maxRetryDelayMs);\n\t\t\tawait delay(delayMs, undefined, { signal });\n\t\t\tsignal.throwIfAborted();\n\t\t\tthis.assertAdmission();\n\t\t\tnextDelayMs = Math.min(delayMs * 2, policy.maxRetryDelayMs);\n\t\t\treturn this.work.providerRecoveryPolicy?.().enabled === true;\n\t\t};\n\t}\n\n\tprivate async execute(\n\t\toperation: BoundOperation,\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\tsend: StreamFn,\n\t\tbindMainOutput = false,\n\t): Promise<InferenceRun> {\n\t\tconst outputSource = bindMainOutput ? captureNativeRequestOutputSource(operation.binding.sink) : undefined;\n\t\tconst part =\n\t\t\toperation.metadata.purpose === \"summary\"\n\t\t\t\t? operation.metadata.purposeDetail === \"compaction\"\n\t\t\t\t\t? \"history\"\n\t\t\t\t\t: operation.metadata.purposeDetail === \"compaction-turn-prefix\"\n\t\t\t\t\t\t? \"turn-prefix\"\n\t\t\t\t\t\t: undefined\n\t\t\t\t: undefined;\n\t\tconst compactionSource =\n\t\t\tthis.#nativeCompactionOutput && part && this.capturedSink?.sink === operation.binding.sink\n\t\t\t\t? captureNativeCompactionOutputSource(operation.binding.sink)\n\t\t\t\t: undefined;\n\t\tconst branchSource =\n\t\t\tthis.#nativeBranchOutput &&\n\t\t\toperation.metadata.purpose === \"summary\" &&\n\t\t\toperation.metadata.purposeDetail === \"branch\" &&\n\t\t\tthis.capturedSink?.sink === operation.binding.sink\n\t\t\t\t? captureNativeBranchOutputSource(operation.binding.sink)\n\t\t\t\t: undefined;\n\t\tconst plannerSource =\n\t\t\tthis.#nativePlannerOutput &&\n\t\t\toperation.metadata.purpose === \"refine\" &&\n\t\t\toperation.metadata.purposeDetail === \"plan\" &&\n\t\t\tthis.capturedSink?.sink === operation.binding.sink\n\t\t\t\t? captureNativePlannerOutputSource(operation.binding.sink)\n\t\t\t\t: undefined;\n\t\tlet outputSourceRef: SourceSnapshotRef | undefined;\n\t\tlet release!: (completion?: PromiseLike<void>) => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\trelease = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tthis.active.add(pending);\n\t\tthis.notifyActivity();\n\t\tconst finish = () => {\n\t\t\tthis.work.pending.delete(pending);\n\t\t\tthis.active.delete(pending);\n\t\t\tthis.notifyActivity();\n\t\t};\n\t\tconst admittedAttempts: Array<{ attemptId: string; contextEpoch?: ContextEpochEntryRef }> = [];\n\t\tconst receiptWrites: Promise<void>[] = [];\n\t\tlet activeUser = false;\n\t\tlet localSimulation = false;\n\t\tlet budgetFailure: unknown;\n\t\tconst settle = async (result: Promise<AssistantMessage>): Promise<InferenceSettlement> => {\n\t\t\tconst failures: unknown[] = [];\n\t\t\tlet message: AssistantMessage | undefined;\n\t\t\ttry {\n\t\t\t\tmessage = await result;\n\t\t\t} catch (error) {\n\t\t\t\tfailures.push(error);\n\t\t\t}\n\t\t\tif (budgetFailure !== undefined && !failures.includes(budgetFailure)) failures.push(budgetFailure);\n\t\t\tawait this.finishSink(operation.binding, receiptWrites, activeUser, failures);\n\t\t\tif (message && outputSource && outputSourceRef) {\n\t\t\t\tconst write = outputSource({\n\t\t\t\t\toperationId: operation.metadata.operationId,\n\t\t\t\t\tattemptIds: admittedAttempts.map(({ attemptId }) => attemptId),\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (write) this.work.mainOutputs.set(message, write);\n\t\t\t}\n\t\t\tconst settlement: InferenceSettlement = {\n\t\t\t\tmessage: message!,\n\t\t\t\toperationId: operation.metadata.operationId,\n\t\t\t\tattemptIds: admittedAttempts.map(({ attemptId }) => attemptId),\n\t\t\t\tphysicalCoverage: admittedAttempts.length\n\t\t\t\t\t? \"reported\"\n\t\t\t\t\t: localSimulation\n\t\t\t\t\t\t? \"local-simulation\"\n\t\t\t\t\t\t: \"unestablished\",\n\t\t\t};\n\t\t\tif (part && compactionSource && outputSourceRef) {\n\t\t\t\tconst output = compactionSource({\n\t\t\t\t\toperationId: settlement.operationId,\n\t\t\t\t\tattemptIds: settlement.attemptIds,\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (output)\n\t\t\t\t\tthis.work.compactionSettlements.set(settlement, {\n\t\t\t\t\t\tsink: operation.binding.sink,\n\t\t\t\t\t\toutput: { ...output, part },\n\t\t\t\t\t});\n\t\t\t}\n\t\t\tif (branchSource && outputSourceRef) {\n\t\t\t\tconst write = branchSource({\n\t\t\t\t\toperationId: settlement.operationId,\n\t\t\t\t\tattemptIds: settlement.attemptIds,\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (write) this.work.branchSettlements.set(settlement, { owner: this, write });\n\t\t\t}\n\t\t\tif (plannerSource && outputSourceRef) {\n\t\t\t\tconst write = plannerSource({\n\t\t\t\t\toperationId: settlement.operationId,\n\t\t\t\t\tattemptIds: settlement.attemptIds,\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (write) this.work.plannerSettlements.set(settlement, { owner: this, write });\n\t\t\t}\n\t\t\treturn settlement;\n\t\t};\n\t\ttry {\n\t\t\tthis.assertAdmission();\n\t\t\toperation.binding.sink.retain();\n\t\t\toperation.binding.release = undefined;\n\t\t\toperation.binding.activeUsers++;\n\t\t\tactiveUser = true;\n\t\t\tthis.releaseCapture();\n\t\t\tthis.finishCapturePending?.();\n\t\t\t// Capture the opted-in budgeted invocation before source/auth waits. Callbacks and signals stay live handles.\n\t\t\tif (this.work.budget || (this.requestViewBoundary && operation.metadata.purpose === \"main\")) {\n\t\t\t\tmodel = structuredClone(model);\n\t\t\t\t// Agent tool objects also carry live executors. Snapshot only the native Tool definition.\n\t\t\t\tcontext = structuredClone({\n\t\t\t\t\t...context,\n\t\t\t\t\t...(context.tools\n\t\t\t\t\t\t? {\n\t\t\t\t\t\t\t\ttools: context.tools.map(({ name, description, parameters }) => ({\n\t\t\t\t\t\t\t\t\tname,\n\t\t\t\t\t\t\t\t\tdescription,\n\t\t\t\t\t\t\t\t\tparameters,\n\t\t\t\t\t\t\t\t})),\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t: {}),\n\t\t\t\t});\n\t\t\t\toptions = options\n\t\t\t\t\t? {\n\t\t\t\t\t\t\t...options,\n\t\t\t\t\t\t\theaders: options.headers ? { ...options.headers } : undefined,\n\t\t\t\t\t\t\tthinkingBudgets: options.thinkingBudgets ? { ...options.thinkingBudgets } : undefined,\n\t\t\t\t\t\t\tmetadata: options.metadata ? structuredClone(options.metadata) : undefined,\n\t\t\t\t\t\t}\n\t\t\t\t\t: undefined;\n\t\t\t}\n\t\t\tconst source = Object.freeze({ ...(await operation.binding.sink.source) });\n\t\t\toutputSourceRef = source;\n\t\t\tconst metadata: NativeRequestMetadata = {\n\t\t\t\t...operation.metadata,\n\t\t\t\tsource,\n\t\t\t\towner: Object.freeze({ ...operation.owner, sessionId: source.sessionId }),\n\t\t\t};\n\t\t\tif (send !== streamSimple && !nativeStreamFunctions.has(send)) {\n\t\t\t\tthrow new Error(\n\t\t\t\t\t\"Native inference cannot admit a custom or proxy StreamFn with unestablished physical-attempt coverage. Use the SDK's instrumented built-in API route instead.\",\n\t\t\t\t);\n\t\t\t}\n\t\t\tassertBuiltInAttemptSupport(model.api);\n\t\t\tif (!this.work.admissionOpen) throw new Error(\"Inference owner is closing\");\n\t\t\tconst parentBudget = this.work.budget;\n\t\t\tconst boundary = operation.metadata.purpose === \"main\" ? this.requestViewBoundary : undefined;\n\t\t\tconst budget = boundary ? parentBudget?.capture() : parentBudget;\n\t\t\tlet measuredForAdmission = false;\n\t\t\tconst measureRequest: ProviderAttemptObserver[\"measureRequest\"] =\n\t\t\t\tbudget || boundary?.validate\n\t\t\t\t\t? (representation) => {\n\t\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\t\tconst assessment = budget?.measure(representation);\n\t\t\t\t\t\t\t\tboundary?.validate?.(representation, assessment);\n\t\t\t\t\t\t\t\tif (budget && assessment) budget.assert(assessment);\n\t\t\t\t\t\t\t\tmeasuredForAdmission = true;\n\t\t\t\t\t\t\t\treturn assessment;\n\t\t\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\t\t\tbudgetFailure = error;\n\t\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t: undefined;\n\t\t\tconst canSelect =\n\t\t\t\t(budget || boundary?.fixedPrepare) && boundary && matchesRequestView(boundary, source, context);\n\t\t\tconst requiredPublic = Boolean(boundary?.pendingPublicMessageGroups?.length);\n\t\t\tif (requiredPublic && (!canSelect || boundary?.fixedPrepare))\n\t\t\t\tthrow new Error(\"Tool continuation requires a selectable native public request boundary\");\n\t\t\tlet publicAccepted = false;\n\t\t\tlet selectedContextEpoch: ContextEpochEntryRef | undefined;\n\t\t\tconst attempts: ProviderAttemptObserver = {\n\t\t\t\t...(requiredPublic\n\t\t\t\t\t? {\n\t\t\t\t\t\t\tpendingPublicMessageGroups: boundary!.providerPendingPublicMessageGroups!.map((group) => [\n\t\t\t\t\t\t\t\t...group,\n\t\t\t\t\t\t\t]),\n\t\t\t\t\t\t}\n\t\t\t\t\t: {}),\n\t\t\t\t...(measureRequest ? { measureRequest } : {}),\n\t\t\t\t...(canSelect\n\t\t\t\t\t? ({\n\t\t\t\t\t\t\tprepareRequest: async (representation, projection) => {\n\t\t\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\t\t\tselectedContextEpoch = undefined;\n\t\t\t\t\t\t\t\t\tpublicAccepted = false;\n\t\t\t\t\t\t\t\t\tif (JSON.parse(representation.body!).model !== model.id) return;\n\t\t\t\t\t\t\t\t\tif (boundary!.fixedPrepare) {\n\t\t\t\t\t\t\t\t\t\tconst assessment = budget?.measure(representation);\n\t\t\t\t\t\t\t\t\t\tif (budget && assessment) budget.assert(assessment);\n\t\t\t\t\t\t\t\t\t\tconst accepted = await prepareFixedRequestView(\n\t\t\t\t\t\t\t\t\t\t\tboundary!,\n\t\t\t\t\t\t\t\t\t\t\trepresentation,\n\t\t\t\t\t\t\t\t\t\t\tprojection,\n\t\t\t\t\t\t\t\t\t\t\tassessment,\n\t\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t\t\tif (accepted) selectedContextEpoch = Object.freeze({ ...accepted });\n\t\t\t\t\t\t\t\t\t\treturn;\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\treturn await selectRequestView(\n\t\t\t\t\t\t\t\t\t\t{\n\t\t\t\t\t\t\t\t\t\t\t...boundary!,\n\t\t\t\t\t\t\t\t\t\t\tcommit: async (candidate) => {\n\t\t\t\t\t\t\t\t\t\t\t\tconst accepted = await boundary!.commit(candidate);\n\t\t\t\t\t\t\t\t\t\t\t\tif (accepted) selectedContextEpoch = Object.freeze({ ...accepted });\n\t\t\t\t\t\t\t\t\t\t\t\tpublicAccepted = Boolean(\n\t\t\t\t\t\t\t\t\t\t\t\t\taccepted &&\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcandidate.publicMessages &&\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcandidate.projection.publicWindow &&\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t!candidate.projection.pendingPublicMessageGroups?.length,\n\t\t\t\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t\t\t\t\treturn accepted || undefined;\n\t\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t\trepresentation,\n\t\t\t\t\t\t\t\t\t\tprojection,\n\t\t\t\t\t\t\t\t\t\tbudget!,\n\t\t\t\t\t\t\t\t\t\tthis.work.budgetMode === \"enforce\",\n\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\t\t\t\tbudgetFailure = error;\n\t\t\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies Pick<ProviderAttemptObserver, \"prepareRequest\">)\n\t\t\t\t\t: {}),\n\t\t\t\tadmit: async (descriptor) => {\n\t\t\t\t\tif (requiredPublic && !publicAccepted)\n\t\t\t\t\t\tthrow new Error(\"Unaccepted tool continuation cannot enter native transport\");\n\t\t\t\t\tif (!this.work.admissionOpen) throw new Error(\"Inference owner is closing\");\n\t\t\t\t\t// An adapter without a serializer meter must not bypass an enforced budget.\n\t\t\t\t\tif (measureRequest && !descriptor.requestBudget && (budget || !measuredForAdmission)) {\n\t\t\t\t\t\tconst assessment = invokeRequestMeasurement(attempts, {\n\t\t\t\t\t\t\tapi: descriptor.api,\n\t\t\t\t\t\t\tprovider: descriptor.provider,\n\t\t\t\t\t\t\turl: \"\",\n\t\t\t\t\t\t\tbody: undefined,\n\t\t\t\t\t\t});\n\t\t\t\t\t\tif (assessment) descriptor = { ...descriptor, requestBudget: assessment };\n\t\t\t\t\t}\n\t\t\t\t\tmeasuredForAdmission = false;\n\t\t\t\t\tif (this.#auxiliaryAttempts) {\n\t\t\t\t\t\tif (this.#auxiliaryAttempts.remaining === 0) {\n\t\t\t\t\t\t\tbudgetFailure ??= new Error(\"Native auxiliary operation attempt limit exhausted\");\n\t\t\t\t\t\t\tthrow budgetFailure;\n\t\t\t\t\t\t}\n\t\t\t\t\t\t// Reserve synchronously before persistence; failures and cancellation do not refund it.\n\t\t\t\t\t\tthis.#auxiliaryAttempts.remaining--;\n\t\t\t\t\t}\n\t\t\t\t\tconst attemptId = randomUUID();\n\t\t\t\t\tconst contextEpoch = selectedContextEpoch ? Object.freeze({ ...selectedContextEpoch }) : undefined;\n\t\t\t\t\tconst write = operation.binding.sink.persist({\n\t\t\t\t\t\t...metadata,\n\t\t\t\t\t\t...(contextEpoch ? { contextEpoch } : {}),\n\t\t\t\t\t\ttype: \"attempt_admitted\",\n\t\t\t\t\t\tattemptId,\n\t\t\t\t\t\ttimestamp: Date.now(),\n\t\t\t\t\t\tdescriptor,\n\t\t\t\t\t});\n\t\t\t\t\treceiptWrites.push(write);\n\t\t\t\t\tawait write;\n\t\t\t\t\tadmittedAttempts.push({ attemptId, contextEpoch });\n\t\t\t\t\treturn attemptId;\n\t\t\t\t},\n\t\t\t\tsettle: async (receipt) => {\n\t\t\t\t\tconst contextEpoch = admittedAttempts.find(\n\t\t\t\t\t\t(attempt) => attempt.attemptId === receipt.attemptId,\n\t\t\t\t\t)?.contextEpoch;\n\t\t\t\t\tconst write = operation.binding.sink.persist({\n\t\t\t\t\t\t...metadata,\n\t\t\t\t\t\t...(contextEpoch ? { contextEpoch } : {}),\n\t\t\t\t\t\ttype: \"attempt_settled\",\n\t\t\t\t\t\tattemptId: receipt.attemptId,\n\t\t\t\t\t\ttimestamp: Date.now(),\n\t\t\t\t\t\treceipt,\n\t\t\t\t\t});\n\t\t\t\t\treceiptWrites.push(write);\n\t\t\t\t\tawait write;\n\t\t\t\t\tparentBudget?.observe(receipt);\n\t\t\t\t},\n\t\t\t};\n\t\t\tconst signal = options?.signal\n\t\t\t\t? AbortSignal.any([options.signal, this.work.cancellation.signal])\n\t\t\t\t: this.work.cancellation.signal;\n\t\t\tconst events = await send(model, context, { ...options, signal, attempts, requireProviderAttempts: true });\n\t\t\tlocalSimulation = isLocalFauxStream(events);\n\t\t\tconst settled = settle(events.result());\n\t\t\tevents.result = async () => (await settled).message;\n\t\t\t// Idle tracking observes rejection; callers still receive it through settled/result().\n\t\t\trelease(settled.then(finish, finish));\n\t\t\treturn { events, settled };\n\t\t} catch (error) {\n\t\t\ttry {\n\t\t\t\tawait settle(Promise.reject(error));\n\t\t\t} finally {\n\t\t\t\tfinish();\n\t\t\t\trelease();\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t}\n\n\tasync start(\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\trequest: InferenceRequestOptions,\n\t): Promise<InferenceRun> {\n\t\tconst semanticEdgeId = request.semanticEdgeId ?? options?.headers?.[MODEL_REQUEST_ID_HEADER];\n\t\tconst operationId = request.operationId ?? semanticEdgeId ?? randomUUID();\n\t\treturn this.execute(\n\t\t\tthis.bindOperation(model, { ...request, semanticEdgeId }, operationId),\n\t\t\tmodel,\n\t\t\tcontext,\n\t\t\toptions,\n\t\t\tstreamSimple,\n\t\t);\n\t}\n\n\tprivate async completeWithProviderRecovery(\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\trequest: InferenceRequestOptions,\n\t): Promise<InferenceSettlement> {\n\t\t// Retry the same request, not a caller's later mutable inputs. Keep callbacks/signals live.\n\t\tmodel = structuredClone(model);\n\t\tcontext = structuredClone({\n\t\t\t...context,\n\t\t\t...(context.tools\n\t\t\t\t? { tools: context.tools.map(({ name, description, parameters }) => ({ name, description, parameters })) }\n\t\t\t\t: {}),\n\t\t});\n\t\toptions = options\n\t\t\t? {\n\t\t\t\t\t...options,\n\t\t\t\t\theaders: options.headers ? { ...options.headers } : undefined,\n\t\t\t\t\tthinkingBudgets: options.thinkingBudgets ? { ...options.thinkingBudgets } : undefined,\n\t\t\t\t\tmetadata: options.metadata ? structuredClone(options.metadata) : undefined,\n\t\t\t\t}\n\t\t\t: undefined;\n\t\tconst semanticEdgeId = request.semanticEdgeId ?? options?.headers?.[MODEL_REQUEST_ID_HEADER];\n\t\tconst operationId = request.operationId ?? semanticEdgeId ?? randomUUID();\n\t\tconst operation = this.bindOperation(model, { ...request, semanticEdgeId }, operationId);\n\t\tconst release = this.retainWork(operation.binding);\n\t\tconst recover = this.providerRecoveryWait();\n\t\tlet failure: unknown;\n\t\ttry {\n\t\t\twhile (true) {\n\t\t\t\tconst settlement = await (await this.execute(operation, model, context, options, streamSimple)).settled;\n\t\t\t\t// Rejected execution/settlement, including local allowance exhaustion, never reaches recovery.\n\t\t\t\tif (!(await recover(settlement.message, options?.signal))) return settlement;\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tfailure = error;\n\t\t\tthrow error;\n\t\t} finally {\n\t\t\tawait release(failure === undefined ? [] : [failure]);\n\t\t}\n\t}\n\n\tcomplete(\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\trequest: InferenceRequestOptions,\n\t): Promise<AssistantMessage> {\n\t\tlet completion!: Promise<AssistantMessage>;\n\t\tcompletion = (async () => {\n\t\t\tconst settlement = this.work.providerRecoveryPolicy?.().enabled\n\t\t\t\t? await this.completeWithProviderRecovery(model, context, options, request)\n\t\t\t\t: await (await this.start(model, context, options, request)).settled;\n\t\t\tconst binding = this.work.compactionSettlements.get(settlement);\n\t\t\tthis.work.compactionSettlements.delete(settlement);\n\t\t\tif (binding) this.work.compactionCompletions.set(completion, binding);\n\t\t\tconst branchBinding = this.work.branchSettlements.get(settlement);\n\t\t\tthis.work.branchSettlements.delete(settlement);\n\t\t\tif (branchBinding) this.work.branchCompletions.set(completion, branchBinding);\n\t\t\tconst plannerBinding = this.work.plannerSettlements.get(settlement);\n\t\t\tthis.work.plannerSettlements.delete(settlement);\n\t\t\tif (plannerBinding) this.work.plannerCompletions.set(completion, plannerBinding);\n\t\t\treturn settlement.message;\n\t\t})();\n\t\treturn completion;\n\t}\n\n\tbindAuxiliaryStream(streamFn: StreamFn, request: InferenceRequestOptions): DisposableInferenceStream {\n\t\tconst captured = this.capture();\n\t\tconst release = captured.retainWork(captured.capturedSink!);\n\t\tconst recover = captured.providerRecoveryWait();\n\t\tconst stream = captured.bindStream(streamFn, request);\n\t\tlet disposal: Promise<void> | undefined;\n\t\treturn Object.assign(stream, {\n\t\t\trecoverProviderFailure: async (message: AssistantMessage, signal?: AbortSignal) => {\n\t\t\t\tif (!(await recover(message, signal))) return false;\n\t\t\t\tcaptured.prepareTurnRetry();\n\t\t\t\treturn true;\n\t\t\t},\n\t\t\tdispose: () => {\n\t\t\t\tdisposal ??= (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tawait release();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tawait captured.dispose();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t\treturn disposal;\n\t\t\t},\n\t\t});\n\t}\n\n\tbindStream(streamFn: StreamFn, request: InferenceRequestOptions): AgentOwnedStreamFn {\n\t\tconst unwrapped = unwrapSemanticEdgeStreamFn(streamFn) as BoundStreamFn;\n\t\tconst inner = unwrapped[REQUEST_STREAM_BINDING]?.inner ?? unwrapped;\n\t\t// One Agent stream is serial. Keep only its latest binding for semantic retries.\n\t\t// A side question or child gets its own closure, not the parent's source binding.\n\t\tlet previous: BoundOperation | undefined;\n\t\tlet previousBody: string | undefined;\n\t\tconst wrapped: BoundStreamFn = async (model, context, options, streamContext) => {\n\t\t\tlet captured: InferenceCoordinator | undefined;\n\t\t\tif (streamContext !== undefined) {\n\t\t\t\tif (\n\t\t\t\t\t!(streamContext instanceof InferenceCoordinator) ||\n\t\t\t\t\tstreamContext.work !== this.work ||\n\t\t\t\t\t!streamContext.capturedSink\n\t\t\t\t)\n\t\t\t\t\tthrow new Error(\"Native stream context is not a capture of this inference owner\");\n\t\t\t\tcaptured = streamContext;\n\t\t\t}\n\t\t\tconst executor = captured ?? this;\n\t\t\tconst semanticEdgeId = request.semanticEdgeId ?? options?.headers?.[MODEL_REQUEST_ID_HEADER];\n\t\t\t// Reuse the existing semantic body identity if its best-effort recorder is unavailable.\n\t\t\tconst body = semanticEdgeId ? undefined : hashTurnBody(model, context, options);\n\t\t\tconst retryId =\n\t\t\t\tthis.retryTurn && body !== undefined && body === previousBody ? previous?.metadata.operationId : undefined;\n\t\t\tthis.retryTurn = false;\n\t\t\tconst operationId = request.operationId ?? semanticEdgeId ?? retryId ?? randomUUID();\n\t\t\tpreviousBody = body;\n\t\t\tif (captured || previous?.metadata.operationId !== operationId) {\n\t\t\t\t// A semantic retry can reuse its operation ID, but compilation owns this exact source frontier.\n\t\t\t\tprevious = executor.bindOperation(model, { ...request, semanticEdgeId }, operationId);\n\t\t\t}\n\t\t\treturn (\n\t\t\t\tawait executor.execute(\n\t\t\t\t\tprevious,\n\t\t\t\t\tmodel,\n\t\t\t\t\tcontext,\n\t\t\t\t\toptions,\n\t\t\t\t\tinner,\n\t\t\t\t\tcaptured !== undefined && request.purpose === \"main\",\n\t\t\t\t)\n\t\t\t).events;\n\t\t};\n\t\twrapped[REQUEST_STREAM_BINDING] = { coordinator: this, inner };\n\t\treturn wrapped;\n\t}\n}\n\n/** Standalone helper embedding stays optional; native session callers always supply requests. */\nexport function completeInference(\n\trequests: InferenceCoordinator | undefined,\n\tmodel: Model<Api>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions | undefined,\n\trequest: InferenceRequestOptions,\n): Promise<AssistantMessage> {\n\treturn requests ? requests.complete(model, context, options, request) : completeSimple(model, context, options);\n}\n\nexport type DisposableInferenceStream = StreamFn & {\n\tdispose(): Promise<void>;\n\trecoverProviderFailure?(message: AssistantMessage, signal?: AbortSignal): Promise<boolean>;\n};\n\n/** Side questions keep the subject's writer, but get their own operation and purpose. */\nexport function bindAuxiliaryInferenceStream(\n\tstreamFn: StreamFn,\n\trequest: InferenceRequestOptions,\n): DisposableInferenceStream {\n\tconst inner = unwrapSemanticEdgeStreamFn(streamFn) as BoundStreamFn;\n\tconst binding = inner[REQUEST_STREAM_BINDING];\n\tif (binding) return binding.coordinator.bindAuxiliaryStream(binding.inner, request);\n\tconst stream: StreamFn = (model, context, options) => inner(model, context, options);\n\treturn Object.assign(stream, { dispose: async () => {} });\n}\n"]}
+{"version":3,"file":"inference-coordinator.d.ts","sourceRoot":"","sources":["../../src/core/inference-coordinator.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,YAAY,EAAE,kBAAkB,EAAE,QAAQ,EAAE,MAAM,kCAAkC,CAAC;AACnG,OAAO,EACN,KAAK,GAAG,EACR,KAAK,gBAAgB,EAErB,KAAK,OAAO,EAKZ,KAAK,KAAK,EAMV,KAAK,yBAAyB,EAC9B,KAAK,mBAAmB,EAExB,MAAM,+BAA+B,CAAC;AACvC,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,mBAAmB,CAAC;AACzD,OAAO,KAAK,EACX,gBAAgB,EAEhB,+BAA+B,EAC/B,wCAAwC,EACxC,gCAAgC,EAGhC,eAAe,EACf,cAAc,EAGd,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EAKN,KAAK,iBAAiB,EACtB,KAAK,uBAAuB,EAC5B,KAAK,mBAAmB,EAExB,MAAM,6BAA6B,CAAC;AAErC,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,4BAA4B,CAAC;AAQzE,MAAM,WAAW,uBAAuB;IACvC,QAAQ,CAAC,OAAO,EAAE,cAAc,CAAC;IACjC,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,iBAAiB,CAAC,EAAE,MAAM,CAAC;IACpC,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,CAAC;CACjC;AAED,MAAM,WAAW,mBAAmB;IACnC,QAAQ,CAAC,OAAO,EAAE,gBAAgB,CAAC;IACnC,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAC7B,QAAQ,CAAC,UAAU,EAAE,SAAS,MAAM,EAAE,CAAC;IACvC,6EAA6E;IAC7E,QAAQ,CAAC,gBAAgB,EAAE,UAAU,GAAG,eAAe,GAAG,kBAAkB,CAAC;CAC7E;AAED,6FAA6F;AAC7F,MAAM,WAAW,6BAA6B;IAC7C,QAAQ,CAAC,IAAI,EAAE,gBAAgB,CAAC;IAChC,QAAQ,CAAC,MAAM,EAAE,wCAAwC,CAAC;CAC1D;AAYD,MAAM,WAAW,YAAY;IAC5B,QAAQ,CAAC,MAAM,EAAE,OAAO,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC,CAAC;IAC/C,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAC,mBAAmB,CAAC,CAAC;CAC/C;AAED,MAAM,WAAW,sBAAsB;IACtC,QAAQ,CAAC,QAAQ,EAAE,oBAAoB,CAAC;CACxC;AAED,UAAU,sBAAsB;IAC/B,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,EAAE,MAAM,CAAC;IACpB,eAAe,EAAE,MAAM,CAAC;CACxB;AAgBD,sFAAsF;AACtF,eAAO,MAAM,+BAA+B,eAAoD,CAAC;AACjG,oGAAoG;AACpG,eAAO,MAAM,2BAA2B,eAAgD,CAAC;AACzF,wFAAwF;AACxF,eAAO,MAAM,4BAA4B,eAAiD,CAAC;AAC3F,wDAAwD;AACxD,eAAO,MAAM,6BAA6B,eAAkD,CAAC;AA8B7F,kHAAkH;AAClH,wBAAgB,6BAA6B,CAAC,OAAO,EAAE,gBAAgB,GAAG,eAAe,GAAG,SAAS,CAIpG;AAED,gGAAgG;AAChG,wBAAgB,2BAA2B,CAC1C,cAAc,EAAE,CACf,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,KACpC,OAAO,CAAC,mBAAmB,GAAG;IAAE,eAAe,CAAC,EAAE,eAAe,CAAA;CAAE,CAAC,GACvE,QAAQ,CAqBV;AAED,mFAAmF;AACnF,qBAAa,oBAAoB;;IAuI/B,OAAO,CAAC,QAAQ,CAAC,QAAQ;IACzB,OAAO,CAAC,QAAQ,CAAC,KAAK;IAnIvB,OAAO,CAAC,SAAS,CAAS;IAC1B,OAAO,CAAC,mBAAmB,CAAC,CAA8B;IAC1D,OAAO,CAAC,MAAM,CAA4B;IAC1C,OAAO,CAAC,YAAY,CAAC,CAAU;IAC/B,OAAO,CAAC,eAAe,CAAS;IAChC,OAAO,CAAC,oBAAoB,CAAC,CAAa;IAC1C,OAAO,CAAC,QAAQ,CAAS;IACzB,OAAO,CAAC,QAAQ,CAAC,CAAgB;IACjC,OAAO,CAAC,IAAI,CAkBV;IAEF,yBAAyB,CAAC,MAAM,EAAE,MAAM,sBAAsB,GAAG,IAAI,CAEpE;IAED,IAAI,UAAU,IAAI,OAAO,CAExB;IACD,IAAI,YAAY,IAAI,MAAM,CAEzB;IAED,gBAAgB,CAAC,QAAQ,EAAE,MAAM,IAAI,GAAG,MAAM,IAAI,CAGjD;IAED,OAAO,CAAC,cAAc;IAWtB,2GAA2G;IAC3G,gBAAgB,CAAC,OAAO,EAAE,gBAAgB,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC,GAAG,SAAS,CAIlG;IAED,2GAA2G;IAC3G,CAAC,+BAA+B,CAAC,IAAI,oBAAoB,CAKxD;IAED,CAAC,2BAA2B,CAAC,CAAC,IAAI,EAAE,gBAAgB,GAAG,oBAAoB,CAK1E;IAED,gBAAgB,CAAC,UAAU,EAAE,OAAO,CAAC,gBAAgB,CAAC,GAAG,+BAA+B,GAAG,SAAS,CAKnG;IAED,CAAC,4BAA4B,CAAC,CAAC,IAAI,EAAE,gBAAgB,GAAG,oBAAoB,CAK3E;IAED,CAAC,6BAA6B,CAAC,IAAI,oBAAoB,CAItD;IAED,iBAAiB,CAAC,UAAU,EAAE,OAAO,CAAC,gBAAgB,CAAC,GAAG,gCAAgC,GAAG,SAAS,CAKrG;IAED,mGAAmG;IACnG,oBAAoB,CAAC,UAAU,EAAE,OAAO,CAAC,gBAAgB,CAAC,GAAG,6BAA6B,GAAG,SAAS,CAIrG;IAED,6FAA6F;IAC7F,aAAa,IAAI,IAAI,CAIpB;IAEK,WAAW,IAAI,OAAO,CAAC,IAAI,CAAC,CAEjC;IAED,gBAAgB,IAAI,IAAI,CAEvB;IACD,cAAc,IAAI,IAAI,CAErB;IAED,YACkB,QAAQ,EAAE,MAAM,gBAAgB,EAChC,KAAK,GAAE,MAAM,IAAI,CAAC,eAAe,EAAE,WAAW,CAAc,EAC7E,kBAAkB,CAAC,EAAE,yBAAyB,EAM9C;IAED,iFAAiF;IACjF,4BAA4B,IAAI,yBAAyB,GAAG,SAAS,CAEpE;IAED,OAAO,CAAC,eAAe;IAIvB,OAAO,CAAC,OAAO;IASf,OAAO,CAAC,cAAc;IAOtB,oFAAoF;IACpF,OAAO,CAAC,IAAI,CAAC,EAAE,gBAAgB,GAAG,oBAAoB,CA6BrD;IAED,kFAAkF;IAClF,uBAAuB,CACtB,QAAQ,EAAE,SAAS,YAAY,EAAE,EACjC,MAAM,EAAE,iBAAiB,EACzB,QAAQ,CAAC,EAAE,mBAAmB,EAC9B,YAAY,CAAC,EAAE,uBAAuB,GACpC,IAAI,CAKN;IAED,0FAA0F;IACpF,WAAW,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,IAAI,EAAE,sBAAsB,KAAK,OAAO,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,CAAC,CAAC,CAmBnF;IAED,2EAA2E;IAC3E,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAYvB;IAED,OAAO,CAAC,aAAa;YAcP,UAAU;IAwBxB,4EAA4E;IAC5E,OAAO,CAAC,UAAU;IA6BlB,OAAO,CAAC,oBAAoB;YAmBd,OAAO;IAgaf,KAAK,CACV,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,EACxC,OAAO,EAAE,uBAAuB,GAC9B,OAAO,CAAC,YAAY,CAAC,CAUvB;YAEa,4BAA4B;IA0C1C,QAAQ,CACP,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,EACxC,OAAO,EAAE,uBAAuB,GAC9B,OAAO,CAAC,gBAAgB,CAAC,CAkB3B;IAED,mBAAmB,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,uBAAuB,GAAG,yBAAyB,CAuBnG;IAED,UAAU,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,uBAAuB,GAAG,kBAAkB,CA4CnF;CACD;AAED,iGAAiG;AACjG,wBAAgB,iBAAiB,CAChC,QAAQ,EAAE,oBAAoB,GAAG,SAAS,EAC1C,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,GAAG,SAAS,EACxC,OAAO,EAAE,uBAAuB,GAC9B,OAAO,CAAC,gBAAgB,CAAC,CAE3B;AAED,MAAM,MAAM,yBAAyB,GAAG,QAAQ,GAAG;IAClD,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC;IACzB,sBAAsB,CAAC,CAAC,OAAO,EAAE,gBAAgB,EAAE,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;CAC3F,CAAC;AAEF,yFAAyF;AACzF,wBAAgB,4BAA4B,CAC3C,QAAQ,EAAE,QAAQ,EAClB,OAAO,EAAE,uBAAuB,GAC9B,yBAAyB,CAM3B","sourcesContent":["import { randomUUID } from \"node:crypto\";\nimport { setTimeout as delay } from \"node:timers/promises\";\nimport type { AgentMessage, AgentOwnedStreamFn, StreamFn } from \"@ponythewhite/base-context-agent\";\nimport {\n\ttype Api,\n\ttype AssistantMessage,\n\tassertBuiltInAttemptSupport,\n\ttype Context,\n\tcompleteSimple,\n\tinvokeRequestMeasurement,\n\tisLocalFauxStream,\n\tisTransientProviderFailure,\n\ttype Model,\n\ttype ProviderAttemptObserver,\n\ttype ProviderInputTokenCounter,\n\ttype ProviderRequestRepresentation,\n\ttype RequestTokenAssessment,\n\tRequestTokenBudget,\n\ttype RequestTokenBudgetOptions,\n\ttype SimpleStreamOptions,\n\tstreamSimple,\n} from \"@ponythewhite/base-context-ai\";\nimport type { AuthSourceToken } from \"./auth-storage.js\";\nimport type {\n\tBoundRequestSink,\n\tContextEpochEntryRef,\n\tNativeBranchRequestOutputWriter,\n\tNativeCompactionRequestOutputAssociation,\n\tNativePlannerRequestOutputWriter,\n\tNativeRequestMetadata,\n\tNativeRequestOutputWriter,\n\tRequestOwnerRef,\n\tRequestPurpose,\n\tResolvedModelContract,\n\tSourceSnapshotRef,\n} from \"./request-events.js\";\nimport {\n\ttype CapturedRequestViewBoundary,\n\tcaptureRequestViewBoundary,\n\tmatchesRequestView,\n\tprepareFixedRequestView,\n\ttype RequestViewCommit,\n\ttype RequestViewFixedPrepare,\n\ttype RequestViewValidate,\n\tselectRequestView,\n} from \"./request-view-selection.js\";\nimport { hashTurnBody, MODEL_REQUEST_ID_HEADER, unwrapSemanticEdgeStreamFn } from \"./semantic-edges.js\";\nimport type { SessionHistoryReadView } from \"./session-history-index.js\";\nimport {\n\tcaptureNativeBranchOutputSource,\n\tcaptureNativeCompactionOutputSource,\n\tcaptureNativePlannerOutputSource,\n\tcaptureNativeRequestOutputSource,\n} from \"./session-manager.js\";\n\nexport interface InferenceRequestOptions {\n\treadonly purpose: RequestPurpose;\n\treadonly purposeDetail?: string;\n\treadonly operationId?: string;\n\treadonly parentOperationId?: string;\n\treadonly semanticEdgeId?: string;\n}\n\nexport interface InferenceSettlement {\n\treadonly message: AssistantMessage;\n\treadonly operationId: string;\n\treadonly attemptIds: readonly string[];\n\t/** Reporting is not a guarantee about an opaque adapter's hidden retries. */\n\treadonly physicalCoverage: \"reported\" | \"unestablished\" | \"local-simulation\";\n}\n\n/** Private completion identity carries this actual source, never descriptive sink labels. */\nexport interface NativeCompactionOutputBinding {\n\treadonly sink: BoundRequestSink;\n\treadonly output: NativeCompactionRequestOutputAssociation;\n}\n\ninterface NativeBranchOutputBinding {\n\treadonly owner: InferenceCoordinator;\n\treadonly write: NativeBranchRequestOutputWriter;\n}\n\ninterface NativePlannerOutputBinding {\n\treadonly owner: InferenceCoordinator;\n\treadonly write: NativePlannerRequestOutputWriter;\n}\n\nexport interface InferenceRun {\n\treadonly events: Awaited<ReturnType<StreamFn>>;\n\treadonly settled: Promise<InferenceSettlement>;\n}\n\nexport interface SessionRuntimeServices {\n\treadonly requests: InferenceCoordinator;\n}\n\ninterface ProviderRecoveryPolicy {\n\tenabled: boolean;\n\tbaseDelayMs: number;\n\tmaxRetryDelayMs: number;\n}\n\ninterface SinkUse {\n\treadonly sink: BoundRequestSink;\n\tactiveUsers: number;\n\tcaptures: number;\n\trelease?: Promise<void>;\n}\n\ninterface BoundOperation {\n\treadonly binding: SinkUse;\n\treadonly owner: Omit<RequestOwnerRef, \"sessionId\">;\n\treadonly metadata: Omit<NativeRequestMetadata, \"source\" | \"owner\" | \"contextEpoch\">;\n}\n\nconst REQUEST_STREAM_BINDING = Symbol(\"base-context.request-stream-binding\");\n/** Internal AgentSession built-in compaction entry, not a public inference option. */\nexport const captureNativeCompactionRequests = Symbol(\"base-context.native-compaction-requests\");\n/** Internal AgentSession built-in branch-summary invocation, never inherited by generic capture. */\nexport const captureNativeBranchRequests = Symbol(\"base-context.native-branch-requests\");\n/** Internal native AgentSession planner capture, never inherited by generic capture. */\nexport const captureNativePlannerRequests = Symbol(\"base-context.native-planner-requests\");\n/** Internal native AgentSession reviewer invocation. */\nexport const captureNativeReviewerRequests = Symbol(\"base-context.native-reviewer-requests\");\ninterface StreamBinding {\n\treadonly coordinator: InferenceCoordinator;\n\treadonly inner: StreamFn;\n}\ntype BoundStreamFn = AgentOwnedStreamFn & { [REQUEST_STREAM_BINDING]?: StreamBinding };\n\nfunction modelContract(model: Model<Api>): ResolvedModelContract {\n\treturn {\n\t\tapi: model.api,\n\t\tprovider: model.provider,\n\t\tmodel: model.id,\n\t\tprofile: { id: \"native-default\", status: \"unvalidated\" },\n\t\tpricing: {\n\t\t\tstatus: \"unvalidated\",\n\t\t\tcurrency: \"USD\",\n\t\t\tunit: \"million-tokens\",\n\t\t\tcatalogRates: {\n\t\t\t\tinput: model.cost.input,\n\t\t\t\toutput: model.cost.output,\n\t\t\t\tcacheRead: model.cost.cacheRead,\n\t\t\t\tcacheWrite: model.cost.cacheWrite,\n\t\t\t},\n\t\t},\n\t};\n}\n\nconst nativeStreamFunctions = new WeakSet<StreamFn>();\nconst nativeAuthSources = new WeakMap<AssistantMessage, AuthSourceToken>();\n\n/** Consume the credential source resolved for this exact native request. Never persisted or sent to providers. */\nexport function takeNativeInferenceAuthSource(message: AssistantMessage): AuthSourceToken | undefined {\n\tconst source = nativeAuthSources.get(message);\n\tnativeAuthSources.delete(message);\n\treturn source;\n}\n\n/** Internal SDK auth/options resolver; transport stays the known instrumented AI dispatcher. */\nexport function createNativeInferenceStream(\n\tresolveOptions: (\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t) => Promise<SimpleStreamOptions & { authSourceToken?: AuthSourceToken }>,\n): StreamFn {\n\tconst dispatch: StreamFn = async (model, context, options) => {\n\t\tconst { authSourceToken, ...resolved } = await resolveOptions(model, context, options);\n\t\tconst events = streamSimple(model, context, {\n\t\t\t...resolved,\n\t\t\tsignal: options?.signal,\n\t\t\tattempts: options?.attempts,\n\t\t\trequireProviderAttempts: true,\n\t\t});\n\t\tif (authSourceToken) {\n\t\t\tconst source = { ...authSourceToken };\n\t\t\tconst result = events.result().then((message) => {\n\t\t\t\tnativeAuthSources.set(message, source);\n\t\t\t\treturn message;\n\t\t\t});\n\t\t\tevents.result = () => result;\n\t\t}\n\t\treturn events;\n\t};\n\tnativeStreamFunctions.add(dispatch);\n\treturn dispatch;\n}\n\n/** Routes physical observations to the source writer. It owns no receipt store. */\nexport class InferenceCoordinator {\n\t#nativeCompactionOutput = false;\n\t#nativeBranchOutput = false;\n\t#nativePlannerOutput = false;\n\t#auxiliaryAttempts?: { remaining: number };\n\tprivate retryTurn = false;\n\tprivate requestViewBoundary?: CapturedRequestViewBoundary;\n\tprivate active = new Set<Promise<void>>();\n\tprivate capturedSink?: SinkUse;\n\tprivate captureReserved = false;\n\tprivate finishCapturePending?: () => void;\n\tprivate disposed = false;\n\tprivate disposal?: Promise<void>;\n\tprivate work = {\n\t\tpending: new Set<Promise<void>>(),\n\t\tresponsesMessageIds: new Map<string, number>(),\n\t\tresponsesIdentity: randomUUID(),\n\t\tsinks: new WeakMap<BoundRequestSink, SinkUse>(),\n\t\tmainOutputs: new WeakMap<AssistantMessage, NativeRequestOutputWriter>(),\n\t\tcompactionSettlements: new WeakMap<InferenceSettlement, NativeCompactionOutputBinding>(),\n\t\tcompactionCompletions: new WeakMap<Promise<AssistantMessage>, NativeCompactionOutputBinding>(),\n\t\tbranchSettlements: new WeakMap<InferenceSettlement, NativeBranchOutputBinding>(),\n\t\tbranchCompletions: new WeakMap<Promise<AssistantMessage>, NativeBranchOutputBinding>(),\n\t\tplannerSettlements: new WeakMap<InferenceSettlement, NativePlannerOutputBinding>(),\n\t\tplannerCompletions: new WeakMap<Promise<AssistantMessage>, NativePlannerOutputBinding>(),\n\t\tlisteners: new Set<() => void>(),\n\t\tadmissionOpen: true,\n\t\tcancellation: new AbortController(),\n\t\tbudget: undefined as RequestTokenBudget | undefined,\n\t\tbudgetMode: undefined as RequestTokenBudgetOptions[\"mode\"] | undefined,\n\t\tproviderRecoveryPolicy: undefined as (() => ProviderRecoveryPolicy) | undefined,\n\t};\n\n\tsetProviderRecoveryPolicy(policy: () => ProviderRecoveryPolicy): void {\n\t\tthis.work.providerRecoveryPolicy = policy;\n\t}\n\n\tget hasPending(): boolean {\n\t\treturn this.work.pending.size > 0;\n\t}\n\tget pendingCount(): number {\n\t\treturn this.work.pending.size;\n\t}\n\n\tonActivityChange(listener: () => void): () => void {\n\t\tthis.work.listeners.add(listener);\n\t\treturn () => this.work.listeners.delete(listener);\n\t}\n\n\tprivate notifyActivity(): void {\n\t\t// Observers can wake an existing owner, but cannot own request settlement.\n\t\tfor (const listener of this.work.listeners) {\n\t\t\ttry {\n\t\t\t\tlistener();\n\t\t\t} catch {\n\t\t\t\t/* Notification is not execution. */\n\t\t\t}\n\t\t}\n\t}\n\n\t/** Consume only this owner's real returned object. Labels and replacement messages remain unassociated. */\n\tappendMainOutput(message: AssistantMessage, owner: object): Promise<string | undefined> | undefined {\n\t\tconst write = this.work.mainOutputs.get(message);\n\t\tthis.work.mainOutputs.delete(message);\n\t\treturn write?.(owner, message);\n\t}\n\n\t/** Opt this native invocation into output association; ordinary captures do not inherit that authority. */\n\t[captureNativeCompactionRequests](): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this);\n\t\tcaptured.#nativeCompactionOutput = true;\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 4 };\n\t\treturn captured;\n\t}\n\n\t[captureNativeBranchRequests](sink: BoundRequestSink): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this, sink);\n\t\tcaptured.#nativeBranchOutput = true;\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 2 };\n\t\treturn captured;\n\t}\n\n\ttakeBranchOutput(completion: Promise<AssistantMessage>): NativeBranchRequestOutputWriter | undefined {\n\t\tconst binding = this.work.branchCompletions.get(completion);\n\t\tthis.work.branchCompletions.delete(completion);\n\t\tif (!binding || !this.#nativeBranchOutput || binding.owner !== this) return undefined;\n\t\treturn binding.write;\n\t}\n\n\t[captureNativePlannerRequests](sink: BoundRequestSink): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this, sink);\n\t\tcaptured.#nativePlannerOutput = true;\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 2 };\n\t\treturn captured;\n\t}\n\n\t[captureNativeReviewerRequests](): InferenceCoordinator {\n\t\tconst captured = InferenceCoordinator.prototype.capture.call(this);\n\t\tcaptured.#auxiliaryAttempts ??= { remaining: 2 };\n\t\treturn captured;\n\t}\n\n\ttakePlannerOutput(completion: Promise<AssistantMessage>): NativePlannerRequestOutputWriter | undefined {\n\t\tconst binding = this.work.plannerCompletions.get(completion);\n\t\tthis.work.plannerCompletions.delete(completion);\n\t\tif (!binding || !this.#nativePlannerOutput || binding.owner !== this) return undefined;\n\t\treturn binding.write;\n\t}\n\n\t/** Only the original completion promise and captured source can supply a projected-output link. */\n\ttakeCompactionOutput(completion: Promise<AssistantMessage>): NativeCompactionOutputBinding | undefined {\n\t\tconst binding = this.work.compactionCompletions.get(completion);\n\t\tthis.work.compactionCompletions.delete(completion);\n\t\treturn binding?.sink === this.capturedSink?.sink ? binding : undefined;\n\t}\n\n\t/** Owner teardown stops new sends, but never suppresses an admitted request's settlement. */\n\tstopAdmission(): void {\n\t\tthis.work.admissionOpen = false;\n\t\tthis.work.cancellation.abort();\n\t\tthis.notifyActivity();\n\t}\n\n\tasync waitForIdle(): Promise<void> {\n\t\twhile (this.work.pending.size > 0) await Promise.all([...this.work.pending]);\n\t}\n\n\tprepareTurnRetry(): void {\n\t\tthis.retryTurn = true;\n\t}\n\tclearTurnRetry(): void {\n\t\tthis.retryTurn = false;\n\t}\n\n\tconstructor(\n\t\tprivate readonly bindSink: () => BoundRequestSink,\n\t\tprivate readonly owner: () => Omit<RequestOwnerRef, \"sessionId\"> = () => ({}),\n\t\trequestTokenBudget?: RequestTokenBudgetOptions,\n\t) {\n\t\tif (requestTokenBudget) {\n\t\t\tthis.work.budget = new RequestTokenBudget(requestTokenBudget);\n\t\t\tthis.work.budgetMode = requestTokenBudget.mode;\n\t\t}\n\t}\n\n\t/** Copy the configured policy, not this owner's learned request observations. */\n\tgetRequestTokenBudgetOptions(): RequestTokenBudgetOptions | undefined {\n\t\treturn this.work.budget?.getOptions();\n\t}\n\n\tprivate assertAdmission(): void {\n\t\tif (this.disposed || !this.work.admissionOpen) throw new Error(\"Inference owner is closing\");\n\t}\n\n\tprivate sinkUse(sink: BoundRequestSink): SinkUse {\n\t\tlet binding = this.work.sinks.get(sink);\n\t\tif (!binding) {\n\t\t\tbinding = { sink, activeUsers: 0, captures: 0 };\n\t\t\tthis.work.sinks.set(sink, binding);\n\t\t}\n\t\treturn binding;\n\t}\n\n\tprivate releaseCapture(): void {\n\t\tif (this.captureReserved && this.capturedSink) {\n\t\t\tthis.captureReserved = false;\n\t\t\tthis.capturedSink.captures--;\n\t\t}\n\t}\n\n\t/** Capture an auxiliary operation's subject before asynchronous auth or UI work. */\n\tcapture(sink?: BoundRequestSink): InferenceCoordinator {\n\t\tthis.assertAdmission();\n\t\tconst owner = Object.freeze({ ...this.owner() });\n\t\tconst binding = this.sinkUse(sink ?? this.bindSink());\n\t\tbinding.sink.retain();\n\t\tbinding.release = undefined;\n\t\tbinding.captures++;\n\t\tconst captured = new InferenceCoordinator(\n\t\t\t() => binding.sink,\n\t\t\t() => owner,\n\t\t);\n\t\tcaptured.work = this.work;\n\t\t// Nested captures share the invocation's allowance, not its native output authority.\n\t\tcaptured.#auxiliaryAttempts = this.#auxiliaryAttempts;\n\t\tcaptured.capturedSink = binding;\n\t\tcaptured.captureReserved = true;\n\t\tlet release!: () => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\trelease = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tcaptured.finishCapturePending = () => {\n\t\t\tthis.work.pending.delete(pending);\n\t\t\tcaptured.finishCapturePending = undefined;\n\t\t\trelease();\n\t\t\tthis.notifyActivity();\n\t\t};\n\t\tthis.notifyActivity();\n\t\treturn captured;\n\t}\n\n\t/** Explicit Root-owned epoch boundary, on this same captured MAIN source only. */\n\tbindRequestViewBoundary(\n\t\tmessages: readonly AgentMessage[],\n\t\tcommit: RequestViewCommit,\n\t\tvalidate?: RequestViewValidate,\n\t\tfixedPrepare?: RequestViewFixedPrepare,\n\t): void {\n\t\tthis.assertAdmission();\n\t\tif (!this.capturedSink) throw new Error(\"Request view boundary requires a captured inference owner\");\n\t\tif (this.requestViewBoundary) throw new Error(\"Request view boundary is already bound\");\n\t\tthis.requestViewBoundary = captureRequestViewBoundary(messages, commit, validate, fixedPrepare);\n\t}\n\n\t/** Read through an explicit capture, never by recapturing the mutable current session. */\n\tasync readHistory<T>(read: (view: SessionHistoryReadView) => Promise<T>): Promise<T> {\n\t\tthis.assertAdmission();\n\t\tconst sink = this.capturedSink?.sink;\n\t\tif (!sink?.readHistory) throw new Error(\"Captured canonical history is unavailable\");\n\t\tlet finish!: () => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\tfinish = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tthis.active.add(pending);\n\t\tthis.notifyActivity();\n\t\ttry {\n\t\t\treturn await sink.readHistory(read);\n\t\t} finally {\n\t\t\tthis.active.delete(pending);\n\t\t\tthis.work.pending.delete(pending);\n\t\t\tfinish();\n\t\t\tthis.notifyActivity();\n\t\t}\n\t}\n\n\t/** Release an unused capture, or join this capture's admitted requests. */\n\tdispose(): Promise<void> {\n\t\tthis.disposed = true;\n\t\tthis.disposal ??= (async () => {\n\t\t\tthis.releaseCapture();\n\t\t\ttry {\n\t\t\t\tawait Promise.all([...this.active]);\n\t\t\t\tif (this.capturedSink) await this.finishSink(this.capturedSink, [], false, []);\n\t\t\t} finally {\n\t\t\t\tthis.finishCapturePending?.();\n\t\t\t}\n\t\t})();\n\t\treturn this.disposal;\n\t}\n\n\tprivate bindOperation(model: Model<Api>, request: InferenceRequestOptions, operationId: string): BoundOperation {\n\t\tthis.assertAdmission();\n\t\tconst owner = Object.freeze({ ...this.owner() });\n\t\tconst metadata = Object.freeze({\n\t\t\toperationId,\n\t\t\tparentOperationId: request.parentOperationId,\n\t\t\tsemanticEdgeId: request.semanticEdgeId,\n\t\t\tpurpose: request.purpose,\n\t\t\tpurposeDetail: request.purposeDetail,\n\t\t\tmodelContract: modelContract(model),\n\t\t});\n\t\treturn { binding: this.sinkUse(this.bindSink()), owner, metadata };\n\t}\n\n\tprivate async finishSink(\n\t\tbinding: SinkUse,\n\t\twrites: readonly Promise<void>[],\n\t\tactiveUser: boolean,\n\t\tfailures: unknown[],\n\t): Promise<void> {\n\t\tconst results = await Promise.allSettled([binding.sink.source, ...writes]);\n\t\tfor (const result of results) {\n\t\t\tif (result.status === \"rejected\") failures.push(result.reason);\n\t\t}\n\t\tif (activeUser) binding.activeUsers--;\n\t\ttry {\n\t\t\tif (binding.activeUsers === 0 && binding.captures === 0) {\n\t\t\t\tbinding.release ??= binding.sink.release();\n\t\t\t\tawait binding.release;\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tfailures.push(error);\n\t\t}\n\t\tconst distinct = [...new Set(failures)];\n\t\tif (distinct.length === 1) throw distinct[0];\n\t\tif (distinct.length > 1) throw new AggregateError(distinct, \"Inference request or source cleanup failed\");\n\t}\n\n\t/** Hold one logical operation through settled sends and recovery delays. */\n\tprivate retainWork(binding: SinkUse): (failures?: unknown[]) => Promise<void> {\n\t\tthis.assertAdmission();\n\t\tbinding.sink.retain();\n\t\tbinding.release = undefined;\n\t\tbinding.captures++;\n\t\tlet finish!: () => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\tfinish = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tthis.active.add(pending);\n\t\tthis.notifyActivity();\n\t\tlet release: Promise<void> | undefined;\n\t\treturn (failures = []) => {\n\t\t\trelease ??= (async () => {\n\t\t\t\tbinding.captures--;\n\t\t\t\ttry {\n\t\t\t\t\tawait this.finishSink(binding, [], false, failures);\n\t\t\t\t} finally {\n\t\t\t\t\tthis.active.delete(pending);\n\t\t\t\t\tthis.work.pending.delete(pending);\n\t\t\t\t\tfinish();\n\t\t\t\t\tthis.notifyActivity();\n\t\t\t\t}\n\t\t\t})();\n\t\t\treturn release;\n\t\t};\n\t}\n\n\tprivate providerRecoveryWait(): (message: AssistantMessage, signal?: AbortSignal) => Promise<boolean> {\n\t\tlet nextDelayMs: number | undefined;\n\t\treturn async (message, callerSignal) => {\n\t\t\tconst policy = this.work.providerRecoveryPolicy?.();\n\t\t\tif (!policy?.enabled || !isTransientProviderFailure(message)) return false;\n\t\t\tthis.assertAdmission();\n\t\t\tconst signal = callerSignal\n\t\t\t\t? AbortSignal.any([callerSignal, this.work.cancellation.signal])\n\t\t\t\t: this.work.cancellation.signal;\n\t\t\tsignal.throwIfAborted();\n\t\t\tconst delayMs = Math.min(nextDelayMs ?? policy.baseDelayMs, policy.maxRetryDelayMs);\n\t\t\tawait delay(delayMs, undefined, { signal });\n\t\t\tsignal.throwIfAborted();\n\t\t\tthis.assertAdmission();\n\t\t\tnextDelayMs = Math.min(delayMs * 2, policy.maxRetryDelayMs);\n\t\t\treturn this.work.providerRecoveryPolicy?.().enabled === true;\n\t\t};\n\t}\n\n\tprivate async execute(\n\t\toperation: BoundOperation,\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\tsend: StreamFn,\n\t\tbindMainOutput = false,\n\t): Promise<InferenceRun> {\n\t\tconst outputSource = bindMainOutput ? captureNativeRequestOutputSource(operation.binding.sink) : undefined;\n\t\tconst part =\n\t\t\toperation.metadata.purpose === \"summary\"\n\t\t\t\t? operation.metadata.purposeDetail === \"compaction\"\n\t\t\t\t\t? \"history\"\n\t\t\t\t\t: operation.metadata.purposeDetail === \"compaction-turn-prefix\"\n\t\t\t\t\t\t? \"turn-prefix\"\n\t\t\t\t\t\t: undefined\n\t\t\t\t: undefined;\n\t\tconst compactionSource =\n\t\t\tthis.#nativeCompactionOutput && part && this.capturedSink?.sink === operation.binding.sink\n\t\t\t\t? captureNativeCompactionOutputSource(operation.binding.sink)\n\t\t\t\t: undefined;\n\t\tconst branchSource =\n\t\t\tthis.#nativeBranchOutput &&\n\t\t\toperation.metadata.purpose === \"summary\" &&\n\t\t\toperation.metadata.purposeDetail === \"branch\" &&\n\t\t\tthis.capturedSink?.sink === operation.binding.sink\n\t\t\t\t? captureNativeBranchOutputSource(operation.binding.sink)\n\t\t\t\t: undefined;\n\t\tconst plannerSource =\n\t\t\tthis.#nativePlannerOutput &&\n\t\t\toperation.metadata.purpose === \"refine\" &&\n\t\t\toperation.metadata.purposeDetail === \"plan\" &&\n\t\t\tthis.capturedSink?.sink === operation.binding.sink\n\t\t\t\t? captureNativePlannerOutputSource(operation.binding.sink)\n\t\t\t\t: undefined;\n\t\tlet outputSourceRef: SourceSnapshotRef | undefined;\n\t\tlet release!: (completion?: PromiseLike<void>) => void;\n\t\tconst pending = new Promise<void>((resolve) => {\n\t\t\trelease = resolve;\n\t\t});\n\t\tthis.work.pending.add(pending);\n\t\tthis.active.add(pending);\n\t\tthis.notifyActivity();\n\t\tconst finish = () => {\n\t\t\tthis.work.pending.delete(pending);\n\t\t\tthis.active.delete(pending);\n\t\t\tthis.notifyActivity();\n\t\t};\n\t\tconst admittedAttempts: Array<{ attemptId: string; contextEpoch?: ContextEpochEntryRef }> = [];\n\t\tconst inputCountAttempts = new Map<string, NativeRequestMetadata>();\n\t\tconst receiptWrites: Promise<void>[] = [];\n\t\tlet activeUser = false;\n\t\tlet localSimulation = false;\n\t\tlet budgetFailure: unknown;\n\t\tconst settle = async (result: Promise<AssistantMessage>): Promise<InferenceSettlement> => {\n\t\t\tconst failures: unknown[] = [];\n\t\t\tlet message: AssistantMessage | undefined;\n\t\t\ttry {\n\t\t\t\tmessage = await result;\n\t\t\t} catch (error) {\n\t\t\t\tfailures.push(error);\n\t\t\t}\n\t\t\tif (budgetFailure !== undefined && !failures.includes(budgetFailure)) failures.push(budgetFailure);\n\t\t\tawait this.finishSink(operation.binding, receiptWrites, activeUser, failures);\n\t\t\tif (message && outputSource && outputSourceRef) {\n\t\t\t\tconst write = outputSource({\n\t\t\t\t\toperationId: operation.metadata.operationId,\n\t\t\t\t\tattemptIds: admittedAttempts.map(({ attemptId }) => attemptId),\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (write) this.work.mainOutputs.set(message, write);\n\t\t\t}\n\t\t\tconst settlement: InferenceSettlement = {\n\t\t\t\tmessage: message!,\n\t\t\t\toperationId: operation.metadata.operationId,\n\t\t\t\tattemptIds: admittedAttempts.map(({ attemptId }) => attemptId),\n\t\t\t\tphysicalCoverage: admittedAttempts.length\n\t\t\t\t\t? \"reported\"\n\t\t\t\t\t: localSimulation\n\t\t\t\t\t\t? \"local-simulation\"\n\t\t\t\t\t\t: \"unestablished\",\n\t\t\t};\n\t\t\tif (part && compactionSource && outputSourceRef) {\n\t\t\t\tconst output = compactionSource({\n\t\t\t\t\toperationId: settlement.operationId,\n\t\t\t\t\tattemptIds: settlement.attemptIds,\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (output)\n\t\t\t\t\tthis.work.compactionSettlements.set(settlement, {\n\t\t\t\t\t\tsink: operation.binding.sink,\n\t\t\t\t\t\toutput: { ...output, part },\n\t\t\t\t\t});\n\t\t\t}\n\t\t\tif (branchSource && outputSourceRef) {\n\t\t\t\tconst write = branchSource({\n\t\t\t\t\toperationId: settlement.operationId,\n\t\t\t\t\tattemptIds: settlement.attemptIds,\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (write) this.work.branchSettlements.set(settlement, { owner: this, write });\n\t\t\t}\n\t\t\tif (plannerSource && outputSourceRef) {\n\t\t\t\tconst write = plannerSource({\n\t\t\t\t\toperationId: settlement.operationId,\n\t\t\t\t\tattemptIds: settlement.attemptIds,\n\t\t\t\t\tsource: outputSourceRef,\n\t\t\t\t});\n\t\t\t\tif (write) this.work.plannerSettlements.set(settlement, { owner: this, write });\n\t\t\t}\n\t\t\treturn settlement;\n\t\t};\n\t\ttry {\n\t\t\tthis.assertAdmission();\n\t\t\toperation.binding.sink.retain();\n\t\t\toperation.binding.release = undefined;\n\t\t\toperation.binding.activeUsers++;\n\t\t\tactiveUser = true;\n\t\t\tthis.releaseCapture();\n\t\t\tthis.finishCapturePending?.();\n\t\t\t// Capture the opted-in budgeted invocation before source/auth waits. Callbacks and signals stay live handles.\n\t\t\tif (this.work.budget || (this.requestViewBoundary && operation.metadata.purpose === \"main\")) {\n\t\t\t\tmodel = structuredClone(model);\n\t\t\t\t// Agent tool objects also carry live executors. Snapshot only the native Tool definition.\n\t\t\t\tcontext = structuredClone({\n\t\t\t\t\t...context,\n\t\t\t\t\t...(context.tools\n\t\t\t\t\t\t? {\n\t\t\t\t\t\t\t\ttools: context.tools.map(({ name, description, parameters }) => ({\n\t\t\t\t\t\t\t\t\tname,\n\t\t\t\t\t\t\t\t\tdescription,\n\t\t\t\t\t\t\t\t\tparameters,\n\t\t\t\t\t\t\t\t})),\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t: {}),\n\t\t\t\t});\n\t\t\t\toptions = options\n\t\t\t\t\t? {\n\t\t\t\t\t\t\t...options,\n\t\t\t\t\t\t\theaders: options.headers ? { ...options.headers } : undefined,\n\t\t\t\t\t\t\tthinkingBudgets: options.thinkingBudgets ? { ...options.thinkingBudgets } : undefined,\n\t\t\t\t\t\t\tmetadata: options.metadata ? structuredClone(options.metadata) : undefined,\n\t\t\t\t\t\t}\n\t\t\t\t\t: undefined;\n\t\t\t}\n\t\t\tconst source = Object.freeze({ ...(await operation.binding.sink.source) });\n\t\t\toutputSourceRef = source;\n\t\t\tconst metadata: NativeRequestMetadata = {\n\t\t\t\t...operation.metadata,\n\t\t\t\tsource,\n\t\t\t\towner: Object.freeze({ ...operation.owner, sessionId: source.sessionId }),\n\t\t\t};\n\t\t\tif (send !== streamSimple && !nativeStreamFunctions.has(send)) {\n\t\t\t\tthrow new Error(\n\t\t\t\t\t\"Native inference cannot admit a custom or proxy StreamFn with unestablished physical-attempt coverage. Use the SDK's instrumented built-in API route instead.\",\n\t\t\t\t);\n\t\t\t}\n\t\t\tassertBuiltInAttemptSupport(model.api);\n\t\t\tif (!this.work.admissionOpen) throw new Error(\"Inference owner is closing\");\n\t\t\tconst parentBudget = this.work.budget;\n\t\t\tconst boundary = operation.metadata.purpose === \"main\" ? this.requestViewBoundary : undefined;\n\t\t\tconst budget = boundary ? parentBudget?.capture() : parentBudget;\n\t\t\tlet measuredForAdmission = false;\n\t\t\tlet countingInput = false;\n\t\t\tlet countedInput: { request: ProviderRequestRepresentation; tokens: number } | undefined;\n\t\t\tconst measure = (request: ProviderRequestRepresentation) => {\n\t\t\t\tconst same =\n\t\t\t\t\tcountedInput &&\n\t\t\t\t\t!request.retainedPrefix &&\n\t\t\t\t\tcountedInput.request.api === request.api &&\n\t\t\t\t\tcountedInput.request.provider === request.provider &&\n\t\t\t\t\tcountedInput.request.url === request.url &&\n\t\t\t\t\tcountedInput.request.body === request.body;\n\t\t\t\treturn budget?.measure(\n\t\t\t\t\trequest,\n\t\t\t\t\tsame\n\t\t\t\t\t\t? {\n\t\t\t\t\t\t\t\tinputTokens: countedInput!.tokens,\n\t\t\t\t\t\t\t\tmethod: \"openai-responses-input-count\",\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t: undefined,\n\t\t\t\t);\n\t\t\t};\n\t\t\tconst finalizeMeasurement = async (\n\t\t\t\trequest: ProviderRequestRepresentation,\n\t\t\t\tcountInput?: ProviderInputTokenCounter,\n\t\t\t\tpreparedAssessment?: RequestTokenAssessment,\n\t\t\t) => {\n\t\t\t\tlet assessment = countedInput ? measure(request) : (preparedAssessment ?? measure(request));\n\t\t\t\tconst estimated = assessment?.estimatedInputTokens;\n\t\t\t\tconst available = assessment?.availableInputTokens;\n\t\t\t\tif (\n\t\t\t\t\tcountInput &&\n\t\t\t\t\tassessment?.counter === \"openai-o200k-base-estimate\" &&\n\t\t\t\t\tassessment.status !== \"unknown\" &&\n\t\t\t\t\testimated !== null &&\n\t\t\t\t\testimated !== undefined &&\n\t\t\t\t\tavailable !== null &&\n\t\t\t\t\tavailable !== undefined &&\n\t\t\t\t\tavailable > 0 &&\n\t\t\t\t\testimated >= available * 0.9 &&\n\t\t\t\t\testimated <= available * 1.1\n\t\t\t\t) {\n\t\t\t\t\tthis.assertAdmission();\n\t\t\t\t\toptions?.signal?.throwIfAborted();\n\t\t\t\t\tthis.work.cancellation.signal.throwIfAborted();\n\t\t\t\t\tcountingInput = true;\n\t\t\t\t\ttry {\n\t\t\t\t\t\tconst tokens = await countInput(request);\n\t\t\t\t\t\tcountedInput = { request: Object.freeze({ ...request }), tokens };\n\t\t\t\t\t\tassessment = measure(request);\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tcountingInput = false;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\treturn assessment;\n\t\t\t};\n\t\t\tconst measureRequest: ProviderAttemptObserver[\"measureRequest\"] =\n\t\t\t\tbudget || boundary?.validate\n\t\t\t\t\t? (representation) => {\n\t\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\t\tconst assessment = measure(representation);\n\t\t\t\t\t\t\t\tboundary?.validate?.(representation, assessment);\n\t\t\t\t\t\t\t\tif (budget && assessment) budget.assert(assessment);\n\t\t\t\t\t\t\t\tmeasuredForAdmission = true;\n\t\t\t\t\t\t\t\treturn assessment;\n\t\t\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\t\t\tbudgetFailure = error;\n\t\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t: undefined;\n\t\t\tconst requiredPublic = Boolean(boundary?.pendingPublicMessageGroups?.length);\n\t\t\tconst requiredEpoch = boundary?.requiresEpoch === true;\n\t\t\tconst canSelect =\n\t\t\t\t(budget || boundary?.fixedPrepare || requiredPublic || requiredEpoch) &&\n\t\t\t\tboundary &&\n\t\t\t\tmatchesRequestView(boundary, source, context);\n\t\t\tif (requiredPublic && (!canSelect || boundary?.fixedPrepare))\n\t\t\t\tthrow new Error(\"Tool continuation requires a selectable native public request boundary\");\n\t\t\tconst responsesMessageIds =\n\t\t\t\tcanSelect &&\n\t\t\t\t!boundary.fixedPrepare &&\n\t\t\t\t(model.api === \"openai-responses\" || model.api === \"openai-codex-responses\")\n\t\t\t\t\t? boundary.providerMessageIndices.map((index) => {\n\t\t\t\t\t\t\tconst unit = boundary.units[index];\n\t\t\t\t\t\t\tconst key = JSON.stringify([\n\t\t\t\t\t\t\t\tboundary.source.sessionId,\n\t\t\t\t\t\t\t\tboundary.source.sessionFile,\n\t\t\t\t\t\t\t\tunit.id,\n\t\t\t\t\t\t\t\tunit.sourceRevision,\n\t\t\t\t\t\t\t]);\n\t\t\t\t\t\t\tlet id = this.work.responsesMessageIds.get(key);\n\t\t\t\t\t\t\tif (id === undefined) {\n\t\t\t\t\t\t\t\tid = this.work.responsesMessageIds.size;\n\t\t\t\t\t\t\t\tthis.work.responsesMessageIds.set(key, id);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\treturn id;\n\t\t\t\t\t\t})\n\t\t\t\t\t: undefined;\n\t\t\tlet publicAccepted = false;\n\t\t\tlet selectedContextEpoch: ContextEpochEntryRef | undefined;\n\t\t\tconst attempts: ProviderAttemptObserver = {\n\t\t\t\t...(responsesMessageIds ? { responsesMessageIds } : {}),\n\t\t\t\t...(requiredPublic\n\t\t\t\t\t? {\n\t\t\t\t\t\t\tpendingPublicMessageGroups: boundary!.providerPendingPublicMessageGroups!.map((group) => [\n\t\t\t\t\t\t\t\t...group,\n\t\t\t\t\t\t\t]),\n\t\t\t\t\t\t}\n\t\t\t\t\t: {}),\n\t\t\t\t...(measureRequest ? { measureRequest } : {}),\n\t\t\t\t...(budget || canSelect\n\t\t\t\t\t? ({\n\t\t\t\t\t\t\tprepareRequest: async (representation, projection, countInput) => {\n\t\t\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\t\t\tselectedContextEpoch = undefined;\n\t\t\t\t\t\t\t\t\tpublicAccepted = false;\n\t\t\t\t\t\t\t\t\tif (JSON.parse(representation.body!).model !== model.id) return;\n\t\t\t\t\t\t\t\t\tif (!canSelect) {\n\t\t\t\t\t\t\t\t\t\tconst assessment = await finalizeMeasurement(representation, countInput);\n\t\t\t\t\t\t\t\t\t\tif (budget && assessment) budget.assert(assessment);\n\t\t\t\t\t\t\t\t\t\treturn;\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\t// An unmapped native body must pass the existing final boundary\n\t\t\t\t\t\t\t\t\t// validator before any provider call, including an input count.\n\t\t\t\t\t\t\t\t\tif (!projection) return;\n\t\t\t\t\t\t\t\t\tif (boundary!.fixedPrepare) {\n\t\t\t\t\t\t\t\t\t\tconst assessment = await finalizeMeasurement(representation, countInput);\n\t\t\t\t\t\t\t\t\t\tif (budget && assessment) budget.assert(assessment);\n\t\t\t\t\t\t\t\t\t\tconst accepted = await prepareFixedRequestView(\n\t\t\t\t\t\t\t\t\t\t\tboundary!,\n\t\t\t\t\t\t\t\t\t\t\trepresentation,\n\t\t\t\t\t\t\t\t\t\t\tprojection,\n\t\t\t\t\t\t\t\t\t\t\tassessment,\n\t\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t\t\tif (accepted) selectedContextEpoch = Object.freeze({ ...accepted });\n\t\t\t\t\t\t\t\t\t\treturn;\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\treturn await selectRequestView(\n\t\t\t\t\t\t\t\t\t\t{\n\t\t\t\t\t\t\t\t\t\t\t...boundary!,\n\t\t\t\t\t\t\t\t\t\t\t...(responsesMessageIds ? { responseItemIdentity: this.work.responsesIdentity } : {}),\n\t\t\t\t\t\t\t\t\t\t\tcommit: async (candidate) => {\n\t\t\t\t\t\t\t\t\t\t\t\tconst accepted = await boundary!.commit(candidate);\n\t\t\t\t\t\t\t\t\t\t\t\tif (accepted) selectedContextEpoch = Object.freeze({ ...accepted });\n\t\t\t\t\t\t\t\t\t\t\t\tpublicAccepted = Boolean(\n\t\t\t\t\t\t\t\t\t\t\t\t\taccepted &&\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcandidate.publicMessages &&\n\t\t\t\t\t\t\t\t\t\t\t\t\t\tcandidate.projection.publicWindow &&\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t!candidate.projection.pendingPublicMessageGroups?.length,\n\t\t\t\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t\t\t\t\treturn accepted || undefined;\n\t\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t\t\trepresentation,\n\t\t\t\t\t\t\t\t\t\tprojection,\n\t\t\t\t\t\t\t\t\t\tbudget,\n\t\t\t\t\t\t\t\t\t\tthis.work.budgetMode === \"enforce\",\n\t\t\t\t\t\t\t\t\t\t(request, assessment) => finalizeMeasurement(request, countInput, assessment),\n\t\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\t\t\t\tbudgetFailure = error;\n\t\t\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies Pick<ProviderAttemptObserver, \"prepareRequest\">)\n\t\t\t\t\t: {}),\n\t\t\t\tadmit: async (descriptor) => {\n\t\t\t\t\tconst inputCount = descriptor.kind === \"input-count\";\n\t\t\t\t\tif (inputCount && !countingInput) throw new Error(\"Input count requires its owned prepared request\");\n\t\t\t\t\toptions?.signal?.throwIfAborted();\n\t\t\t\t\tthis.work.cancellation.signal.throwIfAborted();\n\t\t\t\t\tif (!inputCount && requiredEpoch && !selectedContextEpoch)\n\t\t\t\t\t\tthrow new Error(\"Context transition requires an acknowledged native epoch\");\n\t\t\t\t\tif (!inputCount && requiredPublic && !publicAccepted)\n\t\t\t\t\t\tthrow new Error(\"Unaccepted tool continuation cannot enter native transport\");\n\t\t\t\t\tif (!this.work.admissionOpen) throw new Error(\"Inference owner is closing\");\n\t\t\t\t\t// An adapter without a serializer meter must not bypass an enforced budget.\n\t\t\t\t\tif (!inputCount && measureRequest && !descriptor.requestBudget && (budget || !measuredForAdmission)) {\n\t\t\t\t\t\tconst assessment = invokeRequestMeasurement(attempts, {\n\t\t\t\t\t\t\tapi: descriptor.api,\n\t\t\t\t\t\t\tprovider: descriptor.provider,\n\t\t\t\t\t\t\turl: \"\",\n\t\t\t\t\t\t\tbody: undefined,\n\t\t\t\t\t\t});\n\t\t\t\t\t\tif (assessment) descriptor = { ...descriptor, requestBudget: assessment };\n\t\t\t\t\t}\n\t\t\t\t\tmeasuredForAdmission = false;\n\t\t\t\t\tif (this.#auxiliaryAttempts) {\n\t\t\t\t\t\tif (this.#auxiliaryAttempts.remaining === 0) {\n\t\t\t\t\t\t\tbudgetFailure ??= new Error(\"Native auxiliary operation attempt limit exhausted\");\n\t\t\t\t\t\t\tthrow budgetFailure;\n\t\t\t\t\t\t}\n\t\t\t\t\t\t// Reserve synchronously before persistence; failures and cancellation do not refund it.\n\t\t\t\t\t\tthis.#auxiliaryAttempts.remaining--;\n\t\t\t\t\t}\n\t\t\t\t\tconst attemptId = randomUUID();\n\t\t\t\t\tconst contextEpoch =\n\t\t\t\t\t\t!inputCount && selectedContextEpoch ? Object.freeze({ ...selectedContextEpoch }) : undefined;\n\t\t\t\t\tconst attemptMetadata: NativeRequestMetadata = inputCount\n\t\t\t\t\t\t? { ...metadata, purpose: \"native-control\", purposeDetail: \"input-token-count\" }\n\t\t\t\t\t\t: metadata;\n\t\t\t\t\tconst write = operation.binding.sink.persist({\n\t\t\t\t\t\t...attemptMetadata,\n\t\t\t\t\t\t...(contextEpoch ? { contextEpoch } : {}),\n\t\t\t\t\t\ttype: \"attempt_admitted\",\n\t\t\t\t\t\tattemptId,\n\t\t\t\t\t\ttimestamp: Date.now(),\n\t\t\t\t\t\tdescriptor,\n\t\t\t\t\t});\n\t\t\t\t\treceiptWrites.push(write);\n\t\t\t\t\tawait write;\n\t\t\t\t\tif (inputCount) inputCountAttempts.set(attemptId, attemptMetadata);\n\t\t\t\t\telse admittedAttempts.push({ attemptId, contextEpoch });\n\t\t\t\t\treturn attemptId;\n\t\t\t\t},\n\t\t\t\tsettle: async (receipt) => {\n\t\t\t\t\tconst contextEpoch = admittedAttempts.find(\n\t\t\t\t\t\t(attempt) => attempt.attemptId === receipt.attemptId,\n\t\t\t\t\t)?.contextEpoch;\n\t\t\t\t\tconst inputCountMetadata = inputCountAttempts.get(receipt.attemptId);\n\t\t\t\t\tconst write = operation.binding.sink.persist({\n\t\t\t\t\t\t...(inputCountMetadata ?? metadata),\n\t\t\t\t\t\t...(contextEpoch ? { contextEpoch } : {}),\n\t\t\t\t\t\ttype: \"attempt_settled\",\n\t\t\t\t\t\tattemptId: receipt.attemptId,\n\t\t\t\t\t\ttimestamp: Date.now(),\n\t\t\t\t\t\treceipt,\n\t\t\t\t\t});\n\t\t\t\t\treceiptWrites.push(write);\n\t\t\t\t\tawait write;\n\t\t\t\t\tif (!inputCountMetadata) parentBudget?.observe(receipt);\n\t\t\t\t},\n\t\t\t};\n\t\t\tconst signal = options?.signal\n\t\t\t\t? AbortSignal.any([options.signal, this.work.cancellation.signal])\n\t\t\t\t: this.work.cancellation.signal;\n\t\t\tconst events = await send(model, context, { ...options, signal, attempts, requireProviderAttempts: true });\n\t\t\tlocalSimulation = isLocalFauxStream(events);\n\t\t\tconst settled = settle(events.result());\n\t\t\tevents.result = async () => (await settled).message;\n\t\t\t// Idle tracking observes rejection; callers still receive it through settled/result().\n\t\t\trelease(settled.then(finish, finish));\n\t\t\treturn { events, settled };\n\t\t} catch (error) {\n\t\t\ttry {\n\t\t\t\tawait settle(Promise.reject(error));\n\t\t\t} finally {\n\t\t\t\tfinish();\n\t\t\t\trelease();\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t}\n\n\tasync start(\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\trequest: InferenceRequestOptions,\n\t): Promise<InferenceRun> {\n\t\tconst semanticEdgeId = request.semanticEdgeId ?? options?.headers?.[MODEL_REQUEST_ID_HEADER];\n\t\tconst operationId = request.operationId ?? semanticEdgeId ?? randomUUID();\n\t\treturn this.execute(\n\t\t\tthis.bindOperation(model, { ...request, semanticEdgeId }, operationId),\n\t\t\tmodel,\n\t\t\tcontext,\n\t\t\toptions,\n\t\t\tstreamSimple,\n\t\t);\n\t}\n\n\tprivate async completeWithProviderRecovery(\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\trequest: InferenceRequestOptions,\n\t): Promise<InferenceSettlement> {\n\t\t// Retry the same request, not a caller's later mutable inputs. Keep callbacks/signals live.\n\t\tmodel = structuredClone(model);\n\t\tcontext = structuredClone({\n\t\t\t...context,\n\t\t\t...(context.tools\n\t\t\t\t? { tools: context.tools.map(({ name, description, parameters }) => ({ name, description, parameters })) }\n\t\t\t\t: {}),\n\t\t});\n\t\toptions = options\n\t\t\t? {\n\t\t\t\t\t...options,\n\t\t\t\t\theaders: options.headers ? { ...options.headers } : undefined,\n\t\t\t\t\tthinkingBudgets: options.thinkingBudgets ? { ...options.thinkingBudgets } : undefined,\n\t\t\t\t\tmetadata: options.metadata ? structuredClone(options.metadata) : undefined,\n\t\t\t\t}\n\t\t\t: undefined;\n\t\tconst semanticEdgeId = request.semanticEdgeId ?? options?.headers?.[MODEL_REQUEST_ID_HEADER];\n\t\tconst operationId = request.operationId ?? semanticEdgeId ?? randomUUID();\n\t\tconst operation = this.bindOperation(model, { ...request, semanticEdgeId }, operationId);\n\t\tconst release = this.retainWork(operation.binding);\n\t\tconst recover = this.providerRecoveryWait();\n\t\tlet failure: unknown;\n\t\ttry {\n\t\t\twhile (true) {\n\t\t\t\tconst settlement = await (await this.execute(operation, model, context, options, streamSimple)).settled;\n\t\t\t\t// Rejected execution/settlement, including local allowance exhaustion, never reaches recovery.\n\t\t\t\tif (!(await recover(settlement.message, options?.signal))) return settlement;\n\t\t\t}\n\t\t} catch (error) {\n\t\t\tfailure = error;\n\t\t\tthrow error;\n\t\t} finally {\n\t\t\tawait release(failure === undefined ? [] : [failure]);\n\t\t}\n\t}\n\n\tcomplete(\n\t\tmodel: Model<Api>,\n\t\tcontext: Context,\n\t\toptions: SimpleStreamOptions | undefined,\n\t\trequest: InferenceRequestOptions,\n\t): Promise<AssistantMessage> {\n\t\tlet completion!: Promise<AssistantMessage>;\n\t\tcompletion = (async () => {\n\t\t\tconst settlement = this.work.providerRecoveryPolicy?.().enabled\n\t\t\t\t? await this.completeWithProviderRecovery(model, context, options, request)\n\t\t\t\t: await (await this.start(model, context, options, request)).settled;\n\t\t\tconst binding = this.work.compactionSettlements.get(settlement);\n\t\t\tthis.work.compactionSettlements.delete(settlement);\n\t\t\tif (binding) this.work.compactionCompletions.set(completion, binding);\n\t\t\tconst branchBinding = this.work.branchSettlements.get(settlement);\n\t\t\tthis.work.branchSettlements.delete(settlement);\n\t\t\tif (branchBinding) this.work.branchCompletions.set(completion, branchBinding);\n\t\t\tconst plannerBinding = this.work.plannerSettlements.get(settlement);\n\t\t\tthis.work.plannerSettlements.delete(settlement);\n\t\t\tif (plannerBinding) this.work.plannerCompletions.set(completion, plannerBinding);\n\t\t\treturn settlement.message;\n\t\t})();\n\t\treturn completion;\n\t}\n\n\tbindAuxiliaryStream(streamFn: StreamFn, request: InferenceRequestOptions): DisposableInferenceStream {\n\t\tconst captured = this.capture();\n\t\tconst release = captured.retainWork(captured.capturedSink!);\n\t\tconst recover = captured.providerRecoveryWait();\n\t\tconst stream = captured.bindStream(streamFn, request);\n\t\tlet disposal: Promise<void> | undefined;\n\t\treturn Object.assign(stream, {\n\t\t\trecoverProviderFailure: async (message: AssistantMessage, signal?: AbortSignal) => {\n\t\t\t\tif (!(await recover(message, signal))) return false;\n\t\t\t\tcaptured.prepareTurnRetry();\n\t\t\t\treturn true;\n\t\t\t},\n\t\t\tdispose: () => {\n\t\t\t\tdisposal ??= (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tawait release();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tawait captured.dispose();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t\treturn disposal;\n\t\t\t},\n\t\t});\n\t}\n\n\tbindStream(streamFn: StreamFn, request: InferenceRequestOptions): AgentOwnedStreamFn {\n\t\tconst unwrapped = unwrapSemanticEdgeStreamFn(streamFn) as BoundStreamFn;\n\t\tconst inner = unwrapped[REQUEST_STREAM_BINDING]?.inner ?? unwrapped;\n\t\t// One Agent stream is serial. Keep only its latest binding for semantic retries.\n\t\t// A side question or child gets its own closure, not the parent's source binding.\n\t\tlet previous: BoundOperation | undefined;\n\t\tlet previousBody: string | undefined;\n\t\tconst wrapped: BoundStreamFn = async (model, context, options, streamContext) => {\n\t\t\tlet captured: InferenceCoordinator | undefined;\n\t\t\tif (streamContext !== undefined) {\n\t\t\t\tif (\n\t\t\t\t\t!(streamContext instanceof InferenceCoordinator) ||\n\t\t\t\t\tstreamContext.work !== this.work ||\n\t\t\t\t\t!streamContext.capturedSink\n\t\t\t\t)\n\t\t\t\t\tthrow new Error(\"Native stream context is not a capture of this inference owner\");\n\t\t\t\tcaptured = streamContext;\n\t\t\t}\n\t\t\tconst executor = captured ?? this;\n\t\t\tconst semanticEdgeId = request.semanticEdgeId ?? options?.headers?.[MODEL_REQUEST_ID_HEADER];\n\t\t\t// Reuse the existing semantic body identity if its best-effort recorder is unavailable.\n\t\t\tconst body = semanticEdgeId ? undefined : hashTurnBody(model, context, options);\n\t\t\tconst retryId =\n\t\t\t\tthis.retryTurn && body !== undefined && body === previousBody ? previous?.metadata.operationId : undefined;\n\t\t\tthis.retryTurn = false;\n\t\t\tconst operationId = request.operationId ?? semanticEdgeId ?? retryId ?? randomUUID();\n\t\t\tpreviousBody = body;\n\t\t\tif (captured || previous?.metadata.operationId !== operationId) {\n\t\t\t\t// A semantic retry can reuse its operation ID, but compilation owns this exact source frontier.\n\t\t\t\tprevious = executor.bindOperation(model, { ...request, semanticEdgeId }, operationId);\n\t\t\t}\n\t\t\treturn (\n\t\t\t\tawait executor.execute(\n\t\t\t\t\tprevious,\n\t\t\t\t\tmodel,\n\t\t\t\t\tcontext,\n\t\t\t\t\toptions,\n\t\t\t\t\tinner,\n\t\t\t\t\tcaptured !== undefined && request.purpose === \"main\",\n\t\t\t\t)\n\t\t\t).events;\n\t\t};\n\t\twrapped[REQUEST_STREAM_BINDING] = { coordinator: this, inner };\n\t\treturn wrapped;\n\t}\n}\n\n/** Standalone helper embedding stays optional; native session callers always supply requests. */\nexport function completeInference(\n\trequests: InferenceCoordinator | undefined,\n\tmodel: Model<Api>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions | undefined,\n\trequest: InferenceRequestOptions,\n): Promise<AssistantMessage> {\n\treturn requests ? requests.complete(model, context, options, request) : completeSimple(model, context, options);\n}\n\nexport type DisposableInferenceStream = StreamFn & {\n\tdispose(): Promise<void>;\n\trecoverProviderFailure?(message: AssistantMessage, signal?: AbortSignal): Promise<boolean>;\n};\n\n/** Side questions keep the subject's writer, but get their own operation and purpose. */\nexport function bindAuxiliaryInferenceStream(\n\tstreamFn: StreamFn,\n\trequest: InferenceRequestOptions,\n): DisposableInferenceStream {\n\tconst inner = unwrapSemanticEdgeStreamFn(streamFn) as BoundStreamFn;\n\tconst binding = inner[REQUEST_STREAM_BINDING];\n\tif (binding) return binding.coordinator.bindAuxiliaryStream(binding.inner, request);\n\tconst stream: StreamFn = (model, context, options) => inner(model, context, options);\n\treturn Object.assign(stream, { dispose: async () => {} });\n}\n"]}

package/dist/core/inference-coordinator.js

@@ -81,6 +81,8 @@ export class InferenceCoordinator {
     disposal;
     work = {
         pending: new Set(),
+        responsesMessageIds: new Map(),
+        responsesIdentity: randomUUID(),
         sinks: new WeakMap(),
         mainOutputs: new WeakMap(),
         compactionSettlements: new WeakMap(),
@@ -415,6 +417,7 @@ export class InferenceCoordinator {
             this.notifyActivity();
         };
         const admittedAttempts = [];
+        const inputCountAttempts = new Map();
         const receiptWrites = [];
         let activeUser = false;
         let localSimulation = false;
@@ -532,10 +535,55 @@ export class InferenceCoordinator {
             const boundary = operation.metadata.purpose === "main" ? this.requestViewBoundary : undefined;
             const budget = boundary ? parentBudget?.capture() : parentBudget;
             let measuredForAdmission = false;
+            let countingInput = false;
+            let countedInput;
+            const measure = (request) => {
+                const same = countedInput &&
+                    !request.retainedPrefix &&
+                    countedInput.request.api === request.api &&
+                    countedInput.request.provider === request.provider &&
+                    countedInput.request.url === request.url &&
+                    countedInput.request.body === request.body;
+                return budget?.measure(request, same
+                    ? {
+                        inputTokens: countedInput.tokens,
+                        method: "openai-responses-input-count",
+                    }
+                    : undefined);
+            };
+            const finalizeMeasurement = async (request, countInput, preparedAssessment) => {
+                let assessment = countedInput ? measure(request) : (preparedAssessment ?? measure(request));
+                const estimated = assessment?.estimatedInputTokens;
+                const available = assessment?.availableInputTokens;
+                if (countInput &&
+                    assessment?.counter === "openai-o200k-base-estimate" &&
+                    assessment.status !== "unknown" &&
+                    estimated !== null &&
+                    estimated !== undefined &&
+                    available !== null &&
+                    available !== undefined &&
+                    available > 0 &&
+                    estimated >= available * 0.9 &&
+                    estimated <= available * 1.1) {
+                    this.assertAdmission();
+                    options?.signal?.throwIfAborted();
+                    this.work.cancellation.signal.throwIfAborted();
+                    countingInput = true;
+                    try {
+                        const tokens = await countInput(request);
+                        countedInput = { request: Object.freeze({ ...request }), tokens };
+                        assessment = measure(request);
+                    }
+                    finally {
+                        countingInput = false;
+                    }
+                }
+                return assessment;
+            };
             const measureRequest = budget || boundary?.validate
                 ? (representation) => {
                     try {
-                        const assessment = budget?.measure(representation);
+                        const assessment = measure(representation);
                         boundary?.validate?.(representation, assessment);
                         if (budget && assessment)
                             budget.assert(assessment);
@@ -548,13 +596,36 @@ export class InferenceCoordinator {
                     }
                 }
                 : undefined;
-            const canSelect = (budget || boundary?.fixedPrepare) && boundary && matchesRequestView(boundary, source, context);
             const requiredPublic = Boolean(boundary?.pendingPublicMessageGroups?.length);
+            const requiredEpoch = boundary?.requiresEpoch === true;
+            const canSelect = (budget || boundary?.fixedPrepare || requiredPublic || requiredEpoch) &&
+                boundary &&
+                matchesRequestView(boundary, source, context);
             if (requiredPublic && (!canSelect || boundary?.fixedPrepare))
                 throw new Error("Tool continuation requires a selectable native public request boundary");
+            const responsesMessageIds = canSelect &&
+                !boundary.fixedPrepare &&
+                (model.api === "openai-responses" || model.api === "openai-codex-responses")
+                ? boundary.providerMessageIndices.map((index) => {
+                    const unit = boundary.units[index];
+                    const key = JSON.stringify([
+                        boundary.source.sessionId,
+                        boundary.source.sessionFile,
+                        unit.id,
+                        unit.sourceRevision,
+                    ]);
+                    let id = this.work.responsesMessageIds.get(key);
+                    if (id === undefined) {
+                        id = this.work.responsesMessageIds.size;
+                        this.work.responsesMessageIds.set(key, id);
+                    }
+                    return id;
+                })
+                : undefined;
             let publicAccepted = false;
             let selectedContextEpoch;
             const attempts = {
+                ...(responsesMessageIds ? { responsesMessageIds } : {}),
                 ...(requiredPublic
                     ? {
                         pendingPublicMessageGroups: boundary.providerPendingPublicMessageGroups.map((group) => [
@@ -563,16 +634,26 @@ export class InferenceCoordinator {
                     }
                     : {}),
                 ...(measureRequest ? { measureRequest } : {}),
-                ...(canSelect
+                ...(budget || canSelect
                     ? {
-                        prepareRequest: async (representation, projection) => {
+                        prepareRequest: async (representation, projection, countInput) => {
                             try {
                                 selectedContextEpoch = undefined;
                                 publicAccepted = false;
                                 if (JSON.parse(representation.body).model !== model.id)
                                     return;
+                                if (!canSelect) {
+                                    const assessment = await finalizeMeasurement(representation, countInput);
+                                    if (budget && assessment)
+                                        budget.assert(assessment);
+                                    return;
+                                }
+                                // An unmapped native body must pass the existing final boundary
+                                // validator before any provider call, including an input count.
+                                if (!projection)
+                                    return;
                                 if (boundary.fixedPrepare) {
-                                    const assessment = budget?.measure(representation);
+                                    const assessment = await finalizeMeasurement(representation, countInput);
                                     if (budget && assessment)
                                         budget.assert(assessment);
                                     const accepted = await prepareFixedRequestView(boundary, representation, projection, assessment);
@@ -582,6 +663,7 @@ export class InferenceCoordinator {
                                 }
                                 return await selectRequestView({
                                     ...boundary,
+                                    ...(responsesMessageIds ? { responseItemIdentity: this.work.responsesIdentity } : {}),
                                     commit: async (candidate) => {
                                         const accepted = await boundary.commit(candidate);
                                         if (accepted)
@@ -592,7 +674,7 @@ export class InferenceCoordinator {
                                             !candidate.projection.pendingPublicMessageGroups?.length);
                                         return accepted || undefined;
                                     },
-                                }, representation, projection, budget, this.work.budgetMode === "enforce");
+                                }, representation, projection, budget, this.work.budgetMode === "enforce", (request, assessment) => finalizeMeasurement(request, countInput, assessment));
                             }
                             catch (error) {
                                 budgetFailure = error;
@@ -602,12 +684,19 @@ export class InferenceCoordinator {
                     }
                     : {}),
                 admit: async (descriptor) => {
-                    if (requiredPublic && !publicAccepted)
+                    const inputCount = descriptor.kind === "input-count";
+                    if (inputCount && !countingInput)
+                        throw new Error("Input count requires its owned prepared request");
+                    options?.signal?.throwIfAborted();
+                    this.work.cancellation.signal.throwIfAborted();
+                    if (!inputCount && requiredEpoch && !selectedContextEpoch)
+                        throw new Error("Context transition requires an acknowledged native epoch");
+                    if (!inputCount && requiredPublic && !publicAccepted)
                         throw new Error("Unaccepted tool continuation cannot enter native transport");
                     if (!this.work.admissionOpen)
                         throw new Error("Inference owner is closing");
                     // An adapter without a serializer meter must not bypass an enforced budget.
-                    if (measureRequest && !descriptor.requestBudget && (budget || !measuredForAdmission)) {
+                    if (!inputCount && measureRequest && !descriptor.requestBudget && (budget || !measuredForAdmission)) {
                         const assessment = invokeRequestMeasurement(attempts, {
                             api: descriptor.api,
                             provider: descriptor.provider,
@@ -627,9 +716,12 @@ export class InferenceCoordinator {
                         this.#auxiliaryAttempts.remaining--;
                     }
                     const attemptId = randomUUID();
-                    const contextEpoch = selectedContextEpoch ? Object.freeze({ ...selectedContextEpoch }) : undefined;
+                    const contextEpoch = !inputCount && selectedContextEpoch ? Object.freeze({ ...selectedContextEpoch }) : undefined;
+                    const attemptMetadata = inputCount
+                        ? { ...metadata, purpose: "native-control", purposeDetail: "input-token-count" }
+                        : metadata;
                     const write = operation.binding.sink.persist({
-                        ...metadata,
+                        ...attemptMetadata,
                         ...(contextEpoch ? { contextEpoch } : {}),
                         type: "attempt_admitted",
                         attemptId,
@@ -638,13 +730,17 @@ export class InferenceCoordinator {
                     });
                     receiptWrites.push(write);
                     await write;
-                    admittedAttempts.push({ attemptId, contextEpoch });
+                    if (inputCount)
+                        inputCountAttempts.set(attemptId, attemptMetadata);
+                    else
+                        admittedAttempts.push({ attemptId, contextEpoch });
                     return attemptId;
                 },
                 settle: async (receipt) => {
                     const contextEpoch = admittedAttempts.find((attempt) => attempt.attemptId === receipt.attemptId)?.contextEpoch;
+                    const inputCountMetadata = inputCountAttempts.get(receipt.attemptId);
                     const write = operation.binding.sink.persist({
-                        ...metadata,
+                        ...(inputCountMetadata ?? metadata),
                         ...(contextEpoch ? { contextEpoch } : {}),
                         type: "attempt_settled",
                         attemptId: receipt.attemptId,
@@ -653,7 +749,8 @@ export class InferenceCoordinator {
                     });
                     receiptWrites.push(write);
                     await write;
-                    parentBudget?.observe(receipt);
+                    if (!inputCountMetadata)
+                        parentBudget?.observe(receipt);
                 },
             };
             const signal = options?.signal