@bastani/pi-ai 0.9.19 → 0.9.20-alpha.10

package/dist/api/openai-codex-responses.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-codex-responses.d.ts","sourceRoot":"","sources":["../../src/api/openai-codex-responses.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAEX,6BAA6B,EAG7B,MAAM,yCAAyC,CAAC;AAIjD,OAAO,KAAK,EAOX,mBAAmB,EACnB,cAAc,EACd,aAAa,EAEb,MAAM,aAAa,CAAC;AAyDrB,MAAM,WAAW,2BAA4B,SAAQ,aAAa;IACjE,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;IACnF,gBAAgB,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,UAAU,GAAG,KAAK,GAAG,IAAI,GAAG,IAAI,CAAC;IACzE,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,aAAa,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IAC1C,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,CAAC;CAC1C;AAwJD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,wBAAwB,EAAE,2BAA2B,CAyQxF,CAAC;AAEF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,wBAAwB,EAAE,mBAAmB,CAqBtF,CAAC;AA6WF,MAAM,WAAW,8BAA8B;IAC9C,QAAQ,EAAE,MAAM,CAAC;IACjB,kBAAkB,EAAE,MAAM,CAAC;IAC3B,iBAAiB,EAAE,MAAM,CAAC;IAC1B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,iBAAiB,EAAE,MAAM,CAAC;IAC1B,mBAAmB,EAAE,MAAM,CAAC;IAC5B,aAAa,EAAE,MAAM,CAAC;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,sBAAsB,CAAC,EAAE,MAAM,CAAC;IAChC,iBAAiB,EAAE,MAAM,CAAC;IAC1B,YAAY,EAAE,MAAM,CAAC;IACrB,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,kBAAkB,CAAC,EAAE,MAAM,CAAC;CAC5B;AA0BD,wBAAgB,iCAAiC,CAAC,SAAS,EAAE,MAAM,GAAG,8BAA8B,GAAG,SAAS,CAG/G;AAED,wBAAgB,mCAAmC,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAQ5E;AAED,wBAAgB,iCAAiC,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAc1E","sourcesContent":["import type * as NodeZlib from \"node:zlib\";\nimport type {\n\tTool as OpenAITool,\n\tResponseCreateParamsStreaming,\n\tResponseInput,\n\tResponseStreamEvent,\n} from \"openai/resources/responses/responses.js\";\n\nimport { clampThinkingLevel } from \"../models.ts\";\nimport { registerSessionResourceCleanup } from \"../session-resources.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tContext,\n\tModel,\n\tProviderEnv,\n\tProviderHeaders,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tUsage,\n} from \"../types.ts\";\nimport { combineAbortSignals } from \"../utils/abort-signals.ts\";\nimport { splitDeferredTools } from \"../utils/deferred-tools.ts\";\nimport {\n\tappendAssistantMessageDiagnostic,\n\tcreateAssistantMessageDiagnostic,\n\tformatThrownValue,\n} from \"../utils/diagnostics.ts\";\nimport { formatProviderError, normalizeProviderError } from \"../utils/error-body.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { headersToRecord } from \"../utils/headers.ts\";\nimport { resolveHttpProxyUrlForTarget } from \"../utils/node-http-proxy.ts\";\nimport { getPiUserAgent } from \"../utils/pi-user-agent.ts\";\nimport { createStreamDeadline, type StreamDeadlineHandle, withStreamDeadline } from \"../utils/stream-deadline.ts\";\nimport { uuidv7 } from \"../utils/uuid.ts\";\nimport { createGrammarToolInputProperties } from \"./constrained-sampling.ts\";\nimport { clampOpenAIPromptCacheKey } from \"./openai-prompt-cache.ts\";\nimport {\n\tassertPayloadPreservesFastRoute,\n\tconvertResponsesMessages,\n\tconvertResponsesTools,\n\tprocessResponsesStream,\n\tresolveRequestedServiceTier,\n} from \"./openai-responses-shared.ts\";\nimport { buildBaseOptions } from \"./simple-options.ts\";\n\n// ============================================================================\n// Configuration\n// ============================================================================\n\nconst DEFAULT_CODEX_BASE_URL = \"https://chatgpt.com/backend-api\";\nconst JWT_CLAIM_PATH = \"https://api.openai.com/auth\" as const;\nconst DEFAULT_MAX_RETRIES = 0;\nconst BASE_DELAY_MS = 1000;\nconst DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\nconst DEFAULT_WEBSOCKET_CONNECT_TIMEOUT_MS = 15_000;\n// The Codex backend accepts zstd-compressed request bodies on the SSE responses\n// endpoint (the same endpoint the official Codex client compresses against).\nconst REQUEST_COMPRESSION_ZSTD_LEVEL = 3;\nconst CODEX_TOOL_CALL_PROVIDERS = new Set([\"openai\", \"openai-codex\", \"opencode\"]);\nconst WEBSOCKET_MESSAGE_TOO_BIG_CLOSE_CODE = 1009;\nconst WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE = \"websocket_connection_limit_reached\";\nconst PREVIOUS_RESPONSE_NOT_FOUND_CODE = \"previous_response_not_found\";\n\nconst CODEX_RESPONSE_STATUSES = new Set<CodexResponseStatus>([\n\t\"completed\",\n\t\"incomplete\",\n\t\"failed\",\n\t\"cancelled\",\n\t\"queued\",\n\t\"in_progress\",\n]);\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface OpenAICodexResponsesOptions extends StreamOptions {\n\treasoningEffort?: \"none\" | \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\n\treasoningSummary?: \"auto\" | \"concise\" | \"detailed\" | \"off\" | \"on\" | null;\n\tserviceTier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\ttextVerbosity?: \"low\" | \"medium\" | \"high\";\n\ttoolChoice?: \"auto\" | \"none\" | \"required\";\n}\n\ntype CodexResponseStatus = \"completed\" | \"incomplete\" | \"failed\" | \"cancelled\" | \"queued\" | \"in_progress\";\n\ninterface RequestBody {\n\tmodel: string;\n\tstore?: boolean;\n\tstream?: boolean;\n\tinstructions?: string;\n\tprevious_response_id?: string;\n\tinput?: ResponseInput;\n\ttools?: OpenAITool[];\n\ttool_choice?: OpenAICodexResponsesOptions[\"toolChoice\"];\n\tparallel_tool_calls?: boolean;\n\ttemperature?: number;\n\treasoning?: { effort?: string; summary?: string };\n\tservice_tier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\ttext?: { verbosity?: string };\n\tinclude?: string[];\n\tprompt_cache_key?: string;\n\t[key: string]: unknown;\n}\n\ntype SuccessfulAssistantMessage = AssistantMessage & { stopReason: \"stop\" | \"length\" | \"toolUse\" };\n\nfunction assertSuccessfulOutput(output: AssistantMessage): asserts output is SuccessfulAssistantMessage {\n\tif (output.stopReason === \"pending\") {\n\t\tthrow new Error(\"Codex stream ended without a stop reason\");\n\t}\n\tif (output.stopReason === \"error\" || output.stopReason === \"aborted\") {\n\t\tthrow new Error(output.errorMessage || \"An unknown error occurred\");\n\t}\n}\n\n// ============================================================================\n// Retry Helpers\n// ============================================================================\n\nfunction isTerminalRateLimitError(errorText: string): boolean {\n\treturn /GoUsageLimitError|FreeUsageLimitError|Monthly usage limit reached|available balance|insufficient_quota|out of budget|quota exceeded|billing/i.test(\n\t\terrorText,\n\t);\n}\n\nfunction isRetryableError(status: number, errorText: string): boolean {\n\tif (status === 429 && isTerminalRateLimitError(errorText)) {\n\t\treturn false;\n\t}\n\tif (status === 429 || status === 500 || status === 502 || status === 503 || status === 504) {\n\t\treturn true;\n\t}\n\treturn /rate.?limit|overloaded|service.?unavailable|upstream.?connect|connection.?refused/i.test(errorText);\n}\n\nfunction getRetryAfterDelayMs(headers: Headers): number | undefined {\n\tconst retryAfterMs = headers.get(\"retry-after-ms\");\n\tif (retryAfterMs !== null) {\n\t\tconst millis = Number(retryAfterMs);\n\t\tif (Number.isFinite(millis)) {\n\t\t\treturn Math.max(0, millis);\n\t\t}\n\t}\n\n\tconst retryAfter = headers.get(\"retry-after\");\n\tif (!retryAfter) {\n\t\treturn undefined;\n\t}\n\n\tconst seconds = Number(retryAfter);\n\tif (Number.isFinite(seconds)) {\n\t\treturn Math.max(0, seconds * 1000);\n\t}\n\n\tconst date = Date.parse(retryAfter);\n\tif (!Number.isNaN(date)) {\n\t\treturn Math.max(0, date - Date.now());\n\t}\n\n\treturn undefined;\n}\n\nclass RetryDelayExceededError extends Error {}\n\nfunction validateRetryDelayMs(delayMs: number, options?: StreamOptions): number {\n\tconst maxRetryDelayMs = options?.maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\tif (maxRetryDelayMs > 0 && delayMs > maxRetryDelayMs) {\n\t\tthrow new RetryDelayExceededError(\n\t\t\t`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxRetryDelayMs / 1000)}s)`,\n\t\t);\n\t}\n\treturn delayMs;\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new Error(\"Request was aborted\"));\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\"abort\", () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(new Error(\"Request was aborted\"));\n\t\t});\n\t});\n}\n\nfunction normalizeTimeoutMs(value: number | undefined): number | undefined {\n\tif (value === undefined) return undefined;\n\tif (!Number.isFinite(value) || value < 0) {\n\t\tthrow new Error(`Invalid timeoutMs: ${String(value)}`);\n\t}\n\treturn Math.floor(value);\n}\n\n// ============================================================================\n// Request Compression\n// ============================================================================\n\ntype ProcessWithBuiltinModule = typeof process & {\n\tgetBuiltinModule?: (id: \"node:zlib\") => typeof NodeZlib;\n};\n\nfunction loadNodeZlib(): typeof NodeZlib | null {\n\tif (typeof process === \"undefined\" || !(process.versions?.node || process.versions?.bun)) {\n\t\treturn null;\n\t}\n\treturn (process as ProcessWithBuiltinModule).getBuiltinModule?.(\"node:zlib\") ?? null;\n}\n\n// Returns the zstd-compressed body bytes, or null when compression is\n// unavailable (browser/Vite builds). Callers fall back to sending the\n// uncompressed JSON when this returns null.\nfunction compressRequestBodyZstd(bodyJson: string): Uint8Array | null {\n\tconst zlib = loadNodeZlib();\n\tif (!zlib || typeof zlib.zstdCompressSync !== \"function\") {\n\t\treturn null;\n\t}\n\ttry {\n\t\tconst compressed = zlib.zstdCompressSync(bodyJson, {\n\t\t\tparams: { [zlib.constants.ZSTD_c_compressionLevel]: REQUEST_COMPRESSION_ZSTD_LEVEL },\n\t\t});\n\t\treturn new Uint8Array(compressed.buffer, compressed.byteOffset, compressed.byteLength);\n\t} catch {\n\t\treturn null;\n\t}\n}\n\n// ============================================================================\n// Main Stream Function\n// ============================================================================\n\nexport const stream: StreamFunction<\"openai-codex-responses\", OpenAICodexResponsesOptions> = (\n\tmodel: Model<\"openai-codex-responses\">,\n\tcontext: Context,\n\toptions?: OpenAICodexResponsesOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: \"openai-codex-responses\" as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"pending\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\tconst streamDeadline = createStreamDeadline(options?.streamDeadlineMs, options?.signal);\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey;\n\t\t\tif (!apiKey) {\n\t\t\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t\t\t}\n\n\t\t\tconst accountId = extractAccountId(apiKey);\n\t\t\tconst grammarToolInputProperties = createGrammarToolInputProperties(\n\t\t\t\tcontext.tools,\n\t\t\t\tmodel.compat?.supportsOpenAIGrammarTools ?? false,\n\t\t\t);\n\t\t\tconst cacheSessionId = options?.cacheRetention === \"none\" ? undefined : options?.sessionId;\n\t\t\tconst codexSessionId = clampOpenAIPromptCacheKey(cacheSessionId);\n\t\t\tlet body = buildRequestBody(model, context, options, codexSessionId, grammarToolInputProperties);\n\t\t\tconst nextBody = await options?.onPayload?.(body, model);\n\t\t\tif (nextBody !== undefined) {\n\t\t\t\tbody = nextBody as RequestBody;\n\t\t\t}\n\t\t\tassertPayloadPreservesFastRoute(model, body);\n\t\t\tconst websocketRequestId = codexSessionId || uuidv7();\n\t\t\tconst sseHeaders = buildSSEHeaders(model.headers, options?.headers, accountId, apiKey, codexSessionId);\n\t\t\tconst websocketHeaders = buildWebSocketHeaders(\n\t\t\t\tmodel.headers,\n\t\t\t\toptions?.headers,\n\t\t\t\taccountId,\n\t\t\t\tapiKey,\n\t\t\t\twebsocketRequestId,\n\t\t\t);\n\t\t\tconst bodyJson = JSON.stringify(body);\n\t\t\tconst httpTimeoutMs = normalizeTimeoutMs(options?.timeoutMs);\n\t\t\tconst websocketConnectTimeoutMs = normalizeTimeoutMs(options?.websocketConnectTimeoutMs);\n\t\t\tconst transport = options?.transport || \"auto\";\n\t\t\tlet startEmitted = false;\n\t\t\tconst websocketDisabledForSession = transport !== \"sse\" && isWebSocketSseFallbackActive(cacheSessionId);\n\t\t\tif (websocketDisabledForSession) {\n\t\t\t\trecordWebSocketSseFallback(cacheSessionId);\n\t\t\t}\n\n\t\t\tif (transport !== \"sse\" && !websocketDisabledForSession) {\n\t\t\t\tlet websocketStarted = false;\n\t\t\t\tlet retriedWebSocketConnectionLimit = false;\n\t\t\t\tlet retriedMissingWebSocketContinuation = false;\n\t\t\t\twhile (true) {\n\t\t\t\t\twebsocketStarted = false;\n\t\t\t\t\ttry {\n\t\t\t\t\t\tawait processWebSocketStream(\n\t\t\t\t\t\t\tresolveCodexWebSocketUrl(model.baseUrl),\n\t\t\t\t\t\t\tbody,\n\t\t\t\t\t\t\twebsocketHeaders,\n\t\t\t\t\t\t\toutput,\n\t\t\t\t\t\t\tstream,\n\t\t\t\t\t\t\tmodel,\n\t\t\t\t\t\t\t() => {\n\t\t\t\t\t\t\t\twebsocketStarted = true;\n\t\t\t\t\t\t\t\tif (!startEmitted) {\n\t\t\t\t\t\t\t\t\tstartEmitted = true;\n\t\t\t\t\t\t\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\tstreamDeadline.deadlineMs,\n\t\t\t\t\t\t\twebsocketConnectTimeoutMs,\n\t\t\t\t\t\t\tcacheSessionId,\n\t\t\t\t\t\t\taccountId,\n\t\t\t\t\t\t\tgrammarToolInputProperties,\n\t\t\t\t\t\t\toptions,\n\t\t\t\t\t\t);\n\n\t\t\t\t\t\tif (options?.signal?.aborted) {\n\t\t\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t\t\t}\n\t\t\t\t\t\tassertSuccessfulOutput(output);\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"done\",\n\t\t\t\t\t\t\treason: output.stopReason,\n\t\t\t\t\t\t\tmessage: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t\tstream.end();\n\t\t\t\t\t\treturn;\n\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\tconst aborted = options?.signal?.aborted;\n\t\t\t\t\t\tconst connectionLimitBeforeStart = !websocketStarted && isWebSocketConnectionLimitReachedError(error);\n\t\t\t\t\t\tconst previousResponseNotFound = isPreviousResponseNotFoundError(error);\n\t\t\t\t\t\tif (!aborted && previousResponseNotFound && !retriedMissingWebSocketContinuation) {\n\t\t\t\t\t\t\tretriedMissingWebSocketContinuation = true;\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tif (!aborted && connectionLimitBeforeStart && !retriedWebSocketConnectionLimit) {\n\t\t\t\t\t\t\tretriedWebSocketConnectionLimit = true;\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tif (aborted || (isCodexNonTransportError(error) && !connectionLimitBeforeStart)) {\n\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tappendAssistantMessageDiagnostic(\n\t\t\t\t\t\t\toutput,\n\t\t\t\t\t\t\tcreateAssistantMessageDiagnostic(\"provider_transport_failure\", error, {\n\t\t\t\t\t\t\t\tconfiguredTransport: transport,\n\t\t\t\t\t\t\t\tfallbackTransport: websocketStarted ? undefined : \"sse\",\n\t\t\t\t\t\t\t\teventsEmitted: websocketStarted,\n\t\t\t\t\t\t\t\tphase: websocketStarted ? \"after_message_stream_start\" : \"before_message_stream_start\",\n\t\t\t\t\t\t\t\trequestBytes: new TextEncoder().encode(bodyJson).byteLength,\n\t\t\t\t\t\t\t}),\n\t\t\t\t\t\t);\n\t\t\t\t\t\trecordWebSocketFailure(cacheSessionId, error);\n\t\t\t\t\t\tif (websocketStarted) {\n\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t}\n\t\t\t\t\t\trecordWebSocketSseFallback(cacheSessionId);\n\t\t\t\t\t\tbreak;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\t// Compress the request body once for the SSE path. The Codex backend\n\t\t\t// decodes Content-Encoding: zstd; the WebSocket transport above sends the\n\t\t\t// uncompressed JSON frame, matching the official Codex client.\n\t\t\tconst compressedBody = compressRequestBodyZstd(bodyJson);\n\t\t\tif (compressedBody) {\n\t\t\t\tsseHeaders.set(\"content-encoding\", \"zstd\");\n\t\t\t}\n\t\t\tconst sseBody: Uint8Array | string = compressedBody ?? bodyJson;\n\n\t\t\t// Fetch with retry logic for rate limits and transient errors\n\t\t\tlet response: Response | undefined;\n\t\t\tlet lastError: Error | undefined;\n\t\t\tconst maxRetries = options?.maxRetries ?? DEFAULT_MAX_RETRIES;\n\n\t\t\tfor (let attempt = 0; attempt <= maxRetries; attempt++) {\n\t\t\t\tif (options?.signal?.aborted) {\n\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t}\n\n\t\t\t\ttry {\n\t\t\t\t\tconst headerTimeoutSignal =\n\t\t\t\t\t\thttpTimeoutMs !== undefined && httpTimeoutMs > 0 ? AbortSignal.timeout(httpTimeoutMs) : undefined;\n\t\t\t\t\tconst combinedSignal = combineAbortSignals([streamDeadline.signal, headerTimeoutSignal]);\n\t\t\t\t\ttry {\n\t\t\t\t\t\tresponse = await (options?.fetch ?? globalThis.fetch)(resolveCodexUrl(model.baseUrl), {\n\t\t\t\t\t\t\tmethod: \"POST\",\n\t\t\t\t\t\t\theaders: sseHeaders,\n\t\t\t\t\t\t\tbody: sseBody,\n\t\t\t\t\t\t\tsignal: combinedSignal.signal,\n\t\t\t\t\t\t});\n\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\tif (headerTimeoutSignal?.aborted && !options?.signal?.aborted) {\n\t\t\t\t\t\t\tthrow new Error(`Codex SSE response headers timed out after ${httpTimeoutMs}ms`);\n\t\t\t\t\t\t}\n\t\t\t\t\t\tthrow error;\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tcombinedSignal.cleanup();\n\t\t\t\t\t}\n\t\t\t\t\tawait options?.onResponse?.(\n\t\t\t\t\t\t{ status: response.status, headers: headersToRecord(response.headers) },\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t);\n\n\t\t\t\t\tif (response.ok) {\n\t\t\t\t\t\tbreak;\n\t\t\t\t\t}\n\n\t\t\t\t\tconst errorText = await response.text();\n\t\t\t\t\tif (attempt < maxRetries && isRetryableError(response.status, errorText)) {\n\t\t\t\t\t\tconst retryAfterDelayMs = getRetryAfterDelayMs(response.headers);\n\t\t\t\t\t\tconst delayMs =\n\t\t\t\t\t\t\tretryAfterDelayMs === undefined\n\t\t\t\t\t\t\t\t? BASE_DELAY_MS * 2 ** attempt\n\t\t\t\t\t\t\t\t: validateRetryDelayMs(retryAfterDelayMs, options);\n\n\t\t\t\t\t\tawait sleep(delayMs, options?.signal);\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Parse error for friendly message on final attempt or non-retryable error\n\t\t\t\t\tconst fakeResponse = new Response(errorText, {\n\t\t\t\t\t\tstatus: response.status,\n\t\t\t\t\t\tstatusText: response.statusText,\n\t\t\t\t\t});\n\t\t\t\t\tconst info = await parseErrorResponse(fakeResponse);\n\t\t\t\t\tthrow new Error(info.friendlyMessage || info.message);\n\t\t\t\t} catch (error) {\n\t\t\t\t\tif (error instanceof Error) {\n\t\t\t\t\t\tif (error.name === \"AbortError\" || error.message === \"Request was aborted\") {\n\t\t\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\tlastError = error instanceof Error ? error : new Error(String(error));\n\t\t\t\t\t// Network errors are retryable\n\t\t\t\t\tif (\n\t\t\t\t\t\tattempt < maxRetries &&\n\t\t\t\t\t\t!(lastError instanceof RetryDelayExceededError) &&\n\t\t\t\t\t\t!lastError.message.includes(\"usage limit\")\n\t\t\t\t\t) {\n\t\t\t\t\t\tconst delayMs = BASE_DELAY_MS * 2 ** attempt;\n\t\t\t\t\t\tawait sleep(delayMs, options?.signal);\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\t\t\t\t\tthrow lastError;\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (!response?.ok) {\n\t\t\t\tthrow lastError ?? new Error(\"Failed after retries\");\n\t\t\t}\n\n\t\t\tif (!response.body) {\n\t\t\t\tthrow new Error(\"No response body\");\n\t\t\t}\n\n\t\t\tif (!startEmitted) {\n\t\t\t\tstartEmitted = true;\n\t\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\t}\n\t\t\tawait processStream(response, output, stream, model, grammarToolInputProperties, options, streamDeadline);\n\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tassertSuccessfulOutput(output);\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\t// Streaming scratch buffers are only used during parsing; never persist them.\n\t\t\t\tdelete (block as { partialJson?: string }).partialJson;\n\t\t\t\tdelete (block as { customInput?: unknown }).customInput;\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = formatProviderError(normalizeProviderError(error));\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t} finally {\n\t\t\tstreamDeadline.cleanup();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimple: StreamFunction<\"openai-codex-responses\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-codex-responses\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey;\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = {\n\t\t...buildBaseOptions(model, context, options, apiKey),\n\t\ttoolChoice: options?.toolChoice,\n\t} satisfies OpenAICodexResponsesOptions;\n\tconst clampedReasoning = options?.reasoning ? clampThinkingLevel(model, options.reasoning) : undefined;\n\tconst reasoningEffort = clampedReasoning === \"off\" ? undefined : clampedReasoning;\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t} satisfies OpenAICodexResponsesOptions);\n};\n\n// ============================================================================\n// Request Building\n// ============================================================================\n\nfunction buildRequestBody(\n\tmodel: Model<\"openai-codex-responses\">,\n\tcontext: Context,\n\toptions: OpenAICodexResponsesOptions | undefined,\n\tcacheSessionId: string | undefined,\n\tgrammarToolInputProperties: ReadonlyMap<string, string> = createGrammarToolInputProperties(\n\t\tcontext.tools,\n\t\tmodel.compat?.supportsOpenAIGrammarTools ?? false,\n\t),\n): RequestBody {\n\tconst supportsStrictMode = model.compat?.supportsStrictMode ?? true;\n\tconst supportsOpenAIGrammarTools = model.compat?.supportsOpenAIGrammarTools ?? false;\n\tconst deferredToolsMode = model.compat?.supportsAdditionalTools\n\t\t? \"additional-tools\"\n\t\t: model.compat?.supportsToolSearch\n\t\t\t? \"tool-search\"\n\t\t\t: undefined;\n\tconst toolPlacement = splitDeferredTools(context, deferredToolsMode !== undefined);\n\tconst messages = convertResponsesMessages(model, context, CODEX_TOOL_CALL_PROVIDERS, {\n\t\tincludeSystemPrompt: false,\n\t\tgrammarToolInputProperties,\n\t\tdeferredTools: toolPlacement.deferred,\n\t\tdeferredToolsMode,\n\t\ttoolOptions: {\n\t\t\tstrict: null,\n\t\t\tsupportsStrictMode,\n\t\t\tsupportsOpenAIGrammarTools,\n\t\t},\n\t});\n\n\tconst body: RequestBody = {\n\t\t// A fast variant keeps its canonical `-fast` id on the model object — that is the identity the\n\t\t// caller selected and records — while routing to the base upstream model plus a service tier.\n\t\tmodel: model.fastRoute?.upstreamModelId ?? model.id,\n\t\tstore: false,\n\t\tstream: true,\n\t\tinstructions: context.systemPrompt || \"You are a helpful assistant.\",\n\t\tinput: messages,\n\t\ttext: { verbosity: options?.textVerbosity || \"low\" },\n\t\tinclude: [\"reasoning.encrypted_content\"],\n\t\tprompt_cache_key: cacheSessionId,\n\t\ttool_choice: options?.toolChoice ?? \"auto\",\n\t\tparallel_tool_calls: true,\n\t};\n\n\tif (options?.temperature !== undefined) {\n\t\tbody.temperature = options.temperature;\n\t}\n\n\t// A fast variant carries its own tier, so a caller that only hands over the model still routes fast.\n\tconst requestedServiceTier = resolveRequestedServiceTier(model, options?.serviceTier);\n\tif (requestedServiceTier !== undefined) {\n\t\tbody.service_tier = requestedServiceTier;\n\t}\n\n\tif (toolPlacement.immediate.length > 0) {\n\t\tbody.tools = convertResponsesTools(toolPlacement.immediate, {\n\t\t\tstrict: null,\n\t\t\tsupportsStrictMode,\n\t\t\tsupportsOpenAIGrammarTools,\n\t\t});\n\t}\n\n\tif (options?.reasoningEffort !== undefined) {\n\t\tconst effort =\n\t\t\toptions.reasoningEffort === \"none\"\n\t\t\t\t? model.thinkingLevelMap?.off === undefined\n\t\t\t\t\t? \"none\"\n\t\t\t\t\t: model.thinkingLevelMap.off\n\t\t\t\t: (model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort);\n\t\tif (effort !== null) {\n\t\t\tbody.reasoning = {\n\t\t\t\teffort,\n\t\t\t\tsummary: options.reasoningSummary ?? \"auto\",\n\t\t\t};\n\t\t}\n\t} else if (model.reasoning && model.thinkingLevelMap?.off !== null) {\n\t\tbody.reasoning = { effort: model.thinkingLevelMap?.off ?? \"none\" };\n\t}\n\n\treturn body;\n}\n\nfunction getServiceTierCostMultiplier(\n\tmodel: Pick<Model<\"openai-codex-responses\">, \"fastRoute\" | \"id\">,\n\tserviceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n): number {\n\t// Price against the model that was actually billed upstream, so a `-fast` variant of a\n\t// per-model rate (gpt-5.5) is not silently charged the generic multiplier.\n\tconst pricedModelId = model.fastRoute?.baseModelId ?? model.id;\n\tswitch (serviceTier) {\n\t\tcase \"flex\":\n\t\t\treturn 0.5;\n\t\tcase \"priority\":\n\t\t\treturn pricedModelId === \"gpt-5.5\" ? 2.5 : 2;\n\t\tdefault:\n\t\t\treturn 1;\n\t}\n}\n\nfunction applyServiceTierPricing(\n\tusage: Usage,\n\tserviceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\tmodel: Pick<Model<\"openai-codex-responses\">, \"fastRoute\" | \"id\">,\n) {\n\tconst multiplier = getServiceTierCostMultiplier(model, serviceTier);\n\tif (multiplier === 1) return;\n\n\tusage.cost.input *= multiplier;\n\tusage.cost.output *= multiplier;\n\tusage.cost.cacheRead *= multiplier;\n\tusage.cost.cacheWrite *= multiplier;\n\tusage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;\n}\n\nfunction resolveCodexServiceTier(\n\tresponseServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\trequestServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n): ResponseCreateParamsStreaming[\"service_tier\"] | undefined {\n\tif (responseServiceTier === \"default\" && (requestServiceTier === \"flex\" || requestServiceTier === \"priority\")) {\n\t\treturn requestServiceTier;\n\t}\n\treturn responseServiceTier ?? requestServiceTier;\n}\n\nfunction resolveCodexUrl(baseUrl?: string): string {\n\tconst raw = baseUrl && baseUrl.trim().length > 0 ? baseUrl : DEFAULT_CODEX_BASE_URL;\n\tconst normalized = raw.replace(/\\/+$/, \"\");\n\tif (normalized.endsWith(\"/codex/responses\")) return normalized;\n\tif (normalized.endsWith(\"/codex\")) return `${normalized}/responses`;\n\treturn `${normalized}/codex/responses`;\n}\n\nfunction resolveCodexWebSocketUrl(baseUrl?: string): string {\n\tconst url = new URL(resolveCodexUrl(baseUrl));\n\tif (url.protocol === \"https:\") url.protocol = \"wss:\";\n\tif (url.protocol === \"http:\") url.protocol = \"ws:\";\n\treturn url.toString();\n}\n\n// ============================================================================\n// Response Processing\n// ============================================================================\n\nasync function processStream(\n\tresponse: Response,\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n\tmodel: Model<\"openai-codex-responses\">,\n\tgrammarToolInputProperties: ReadonlyMap<string, string>,\n\toptions: OpenAICodexResponsesOptions | undefined,\n\tstreamDeadline: StreamDeadlineHandle,\n): Promise<void> {\n\tconst events = mapCodexEvents(parseSSE(response, streamDeadline.signal), output);\n\tawait processResponsesStream(\n\t\twithStreamDeadline(events, streamDeadline.deadlineMs, streamDeadline.abort),\n\t\toutput,\n\t\tstream,\n\t\tmodel,\n\t\t{\n\t\t\tserviceTier: resolveRequestedServiceTier(model, options?.serviceTier),\n\t\t\tgrammarToolInputProperties,\n\t\t\tresolveServiceTier: resolveCodexServiceTier,\n\t\t\tapplyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),\n\t\t},\n\t);\n}\n\nclass CodexApiError extends Error {\n\treadonly code?: string;\n\treadonly payload?: Record<string, unknown>;\n\n\tconstructor(message: string, options?: { code?: string; payload?: Record<string, unknown>; cause?: unknown }) {\n\t\tsuper(message);\n\t\tthis.name = \"CodexApiError\";\n\t\tthis.code = options?.code;\n\t\tthis.payload = options?.payload;\n\t\tthis.cause = options?.cause;\n\t}\n}\n\nclass CodexProtocolError extends Error {\n\treadonly payload?: unknown;\n\n\tconstructor(message: string, options?: { payload?: unknown; cause?: unknown }) {\n\t\tsuper(message);\n\t\tthis.name = \"CodexProtocolError\";\n\t\tthis.payload = options?.payload;\n\t\tthis.cause = options?.cause;\n\t}\n}\n\nfunction isCodexNonTransportError(error: unknown): boolean {\n\treturn error instanceof CodexApiError || error instanceof CodexProtocolError;\n}\n\nfunction isWebSocketConnectionLimitReachedError(error: unknown): boolean {\n\treturn error instanceof CodexApiError && error.code === WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE;\n}\n\nfunction isPreviousResponseNotFoundError(error: unknown): boolean {\n\treturn error instanceof CodexApiError && error.code === PREVIOUS_RESPONSE_NOT_FOUND_CODE;\n}\n\nfunction extractCodexEventError(event: Record<string, unknown>): { code?: string; message?: string } {\n\tconst nested = event.error && typeof event.error === \"object\" ? (event.error as Record<string, unknown>) : undefined;\n\treturn {\n\t\tcode: typeof event.code === \"string\" ? event.code : typeof nested?.code === \"string\" ? nested.code : undefined,\n\t\tmessage:\n\t\t\ttypeof event.message === \"string\"\n\t\t\t\t? event.message\n\t\t\t\t: typeof nested?.message === \"string\"\n\t\t\t\t\t? nested.message\n\t\t\t\t\t: undefined,\n\t};\n}\n\nasync function* mapCodexEvents(\n\tevents: AsyncIterable<Record<string, unknown>>,\n\toutput: AssistantMessage,\n): AsyncGenerator<ResponseStreamEvent> {\n\tfor await (const event of events) {\n\t\tconst type = typeof event.type === \"string\" ? event.type : undefined;\n\t\tif (!type) continue;\n\n\t\tif (type === \"error\") {\n\t\t\tconst { code, message } = extractCodexEventError(event);\n\t\t\tthrow new CodexApiError(`Codex error: ${message || code || JSON.stringify(event)}`, {\n\t\t\t\tcode,\n\t\t\t\tpayload: event,\n\t\t\t});\n\t\t}\n\n\t\tif (type === \"response.failed\") {\n\t\t\tconst response = (event as { response?: { error?: { code?: string; message?: string } } }).response;\n\t\t\tconst code = response?.error?.code;\n\t\t\tconst message = response?.error?.message;\n\t\t\tthrow new CodexApiError(message || \"Codex response failed\", { code, payload: event });\n\t\t}\n\n\t\tif (type === \"response.done\" || type === \"response.completed\" || type === \"response.incomplete\") {\n\t\t\tconst response = (event as { response?: { status?: unknown; end_turn?: unknown } }).response;\n\t\t\tif (typeof response?.end_turn === \"boolean\") {\n\t\t\t\toutput.endTurn = response.end_turn;\n\t\t\t}\n\t\t\tconst normalizedResponse = response\n\t\t\t\t? { ...response, status: normalizeCodexStatus(response.status) }\n\t\t\t\t: response;\n\t\t\tyield { ...event, type: \"response.completed\", response: normalizedResponse } as ResponseStreamEvent;\n\t\t\treturn;\n\t\t}\n\n\t\tyield event as unknown as ResponseStreamEvent;\n\t}\n}\n\nfunction normalizeCodexStatus(status: unknown): CodexResponseStatus | undefined {\n\tif (typeof status !== \"string\") return undefined;\n\treturn CODEX_RESPONSE_STATUSES.has(status as CodexResponseStatus) ? (status as CodexResponseStatus) : undefined;\n}\n\n// ============================================================================\n// SSE Parsing\n// ============================================================================\n\nasync function* parseSSE(response: Response, signal?: AbortSignal): AsyncGenerator<Record<string, unknown>> {\n\tif (!response.body) return;\n\n\tconst reader = response.body.getReader();\n\tconst decoder = new TextDecoder();\n\tlet buffer = \"\";\n\tconst onAbort = () => {\n\t\tvoid reader.cancel().catch(() => {});\n\t};\n\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\n\ttry {\n\t\twhile (true) {\n\t\t\tif (signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tconst { done, value } = await reader.read();\n\t\t\tif (signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tbuffer += done ? decoder.decode() : decoder.decode(value, { stream: true });\n\t\t\t// Treat EOF as terminating the residual SSE frame.\n\t\t\tif (done && buffer.trim()) buffer += \"\\n\\n\";\n\n\t\t\tlet idx = buffer.indexOf(\"\\n\\n\");\n\t\t\twhile (idx !== -1) {\n\t\t\t\tconst chunk = buffer.slice(0, idx);\n\t\t\t\tbuffer = buffer.slice(idx + 2);\n\n\t\t\t\tconst dataLines = chunk\n\t\t\t\t\t.split(\"\\n\")\n\t\t\t\t\t.filter((l) => l.startsWith(\"data:\"))\n\t\t\t\t\t.map((l) => l.slice(5).trim());\n\t\t\t\tif (dataLines.length > 0) {\n\t\t\t\t\tconst data = dataLines.join(\"\\n\").trim();\n\t\t\t\t\tif (data && data !== \"[DONE]\") {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\tyield JSON.parse(data) as Record<string, unknown>;\n\t\t\t\t\t\t} catch (cause) {\n\t\t\t\t\t\t\tthrow new CodexProtocolError(`Invalid Codex SSE JSON: ${formatThrownValue(cause)}`, {\n\t\t\t\t\t\t\t\tcause,\n\t\t\t\t\t\t\t\tpayload: data,\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}\n\t\t\t\tidx = buffer.indexOf(\"\\n\\n\");\n\t\t\t}\n\n\t\t\tif (done) break;\n\t\t}\n\t} finally {\n\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\ttry {\n\t\t\tawait reader.cancel();\n\t\t} catch {}\n\t\ttry {\n\t\t\treader.releaseLock();\n\t\t} catch {}\n\t}\n}\n\n// ============================================================================\n// WebSocket Parsing\n// ============================================================================\n\nconst OPENAI_BETA_RESPONSES_WEBSOCKETS = \"responses_websockets=2026-02-06\";\nconst SESSION_WEBSOCKET_CACHE_TTL_MS = 5 * 60 * 1000;\nconst SESSION_WEBSOCKET_MAX_AGE_MS = 55 * 60 * 1000;\n\ntype WebSocketEventType = \"open\" | \"message\" | \"error\" | \"close\";\ntype WebSocketListener = (event: unknown) => void;\n\ninterface WebSocketLike {\n\tclose(code?: number, reason?: string): void;\n\tsend(data: string): void;\n\taddEventListener(type: WebSocketEventType, listener: WebSocketListener): void;\n\tremoveEventListener(type: WebSocketEventType, listener: WebSocketListener): void;\n}\n\ninterface CachedWebSocketContinuationState {\n\tlastRequestBody: RequestBody;\n\tlastResponseId: string;\n\tlastResponseItems: ResponseInput;\n}\n\ninterface CachedWebSocketConnection {\n\tsocket: WebSocketLike;\n\tbusy: boolean;\n\tcreatedAt: number;\n\tidleTimer?: ReturnType<typeof setTimeout>;\n\tcontinuation?: CachedWebSocketContinuationState;\n}\n\nexport interface OpenAICodexWebSocketDebugStats {\n\trequests: number;\n\tconnectionsCreated: number;\n\tconnectionsReused: number;\n\tcachedContextRequests: number;\n\tstoreTrueRequests: number;\n\tfullContextRequests: number;\n\tdeltaRequests: number;\n\tlastInputItems: number;\n\tlastDeltaInputItems?: number;\n\tlastPreviousResponseId?: string;\n\twebsocketFailures: number;\n\tsseFallbacks: number;\n\twebsocketFallbackActive?: boolean;\n\tlastWebSocketError?: string;\n}\n\nconst websocketSessionCache = new Map<string, Map<string, CachedWebSocketConnection>>();\nconst websocketDebugStats = new Map<string, OpenAICodexWebSocketDebugStats>();\nconst websocketSseFallbackSessions = new Set<string>();\n\nfunction getOrCreateWebSocketDebugStats(sessionId: string): OpenAICodexWebSocketDebugStats {\n\tlet stats = websocketDebugStats.get(sessionId);\n\tif (!stats) {\n\t\tstats = {\n\t\t\trequests: 0,\n\t\t\tconnectionsCreated: 0,\n\t\t\tconnectionsReused: 0,\n\t\t\tcachedContextRequests: 0,\n\t\t\tstoreTrueRequests: 0,\n\t\t\tfullContextRequests: 0,\n\t\t\tdeltaRequests: 0,\n\t\t\tlastInputItems: 0,\n\t\t\twebsocketFailures: 0,\n\t\t\tsseFallbacks: 0,\n\t\t};\n\t\twebsocketDebugStats.set(sessionId, stats);\n\t}\n\treturn stats;\n}\n\nexport function getOpenAICodexWebSocketDebugStats(sessionId: string): OpenAICodexWebSocketDebugStats | undefined {\n\tconst stats = websocketDebugStats.get(sessionId);\n\treturn stats ? { ...stats } : undefined;\n}\n\nexport function resetOpenAICodexWebSocketDebugStats(sessionId?: string): void {\n\tif (sessionId) {\n\t\twebsocketDebugStats.delete(sessionId);\n\t\twebsocketSseFallbackSessions.delete(sessionId);\n\t\treturn;\n\t}\n\twebsocketDebugStats.clear();\n\twebsocketSseFallbackSessions.clear();\n}\n\nexport function closeOpenAICodexWebSocketSessions(sessionId?: string): void {\n\tconst closeEntry = (entry: CachedWebSocketConnection) => {\n\t\tif (entry.idleTimer) clearTimeout(entry.idleTimer);\n\t\tcloseWebSocketSilently(entry.socket, 1000, \"debug_close\");\n\t};\n\tif (sessionId) {\n\t\tfor (const entry of websocketSessionCache.get(sessionId)?.values() ?? []) closeEntry(entry);\n\t\twebsocketSessionCache.delete(sessionId);\n\t\treturn;\n\t}\n\tfor (const accountEntries of websocketSessionCache.values()) {\n\t\tfor (const entry of accountEntries.values()) closeEntry(entry);\n\t}\n\twebsocketSessionCache.clear();\n}\n\nregisterSessionResourceCleanup(closeOpenAICodexWebSocketSessions);\n\nfunction isWebSocketSseFallbackActive(sessionId: string | undefined): boolean {\n\treturn sessionId ? websocketSseFallbackSessions.has(sessionId) : false;\n}\n\nfunction recordWebSocketSseFallback(sessionId: string | undefined): void {\n\tif (!sessionId) return;\n\tconst stats = getOrCreateWebSocketDebugStats(sessionId);\n\tstats.sseFallbacks++;\n\tstats.websocketFallbackActive = isWebSocketSseFallbackActive(sessionId);\n}\n\nfunction recordWebSocketFailure(sessionId: string | undefined, error: unknown): void {\n\tif (!sessionId) return;\n\twebsocketSseFallbackSessions.add(sessionId);\n\n\tconst stats = getOrCreateWebSocketDebugStats(sessionId);\n\tstats.websocketFailures++;\n\tstats.lastWebSocketError = formatThrownValue(error);\n\tstats.websocketFallbackActive = true;\n}\n\ntype WebSocketConstructor = new (\n\turl: string,\n\tprotocols?: string | string[] | { headers?: Record<string, string> },\n) => WebSocketLike;\n\nlet _cachedWebsocket: WebSocketConstructor | null = null;\nasync function getWebSocketConstructor(env?: ProviderEnv): Promise<WebSocketConstructor | null> {\n\tif (!env && _cachedWebsocket) return _cachedWebsocket;\n\n\t// bun doesn't respect http proxy envs, ref: https://github.com/oven-sh/bun/issues/15489\n\t// TODO: remove this when bun supports proxy envs in websocket.\n\tif (typeof process !== \"undefined\" && process.versions?.bun) {\n\t\tconst WebSocketWithProxy = class extends WebSocket {\n\t\t\tconstructor(url: string | URL, options?: string | string[] | Record<string, unknown>) {\n\t\t\t\tlet _opts: Record<string, unknown> = {};\n\t\t\t\tif (Array.isArray(options) || typeof options === \"string\") {\n\t\t\t\t\t_opts = { protocols: options };\n\t\t\t\t} else {\n\t\t\t\t\t_opts = { ...options };\n\t\t\t\t}\n\n\t\t\t\tconst proxyUrl = resolveHttpProxyUrlForTarget(\n\t\t\t\t\turl.toString().replace(/^wss:/, \"https:\").replace(/^ws:/, \"http:\"),\n\t\t\t\t\tenv,\n\t\t\t\t);\n\t\t\t\tsuper(url, { ..._opts, ...(proxyUrl ? { proxy: proxyUrl.toString() } : {}) } as any);\n\t\t\t}\n\t\t};\n\t\tif (!env) {\n\t\t\t_cachedWebsocket = WebSocketWithProxy;\n\t\t}\n\t\treturn WebSocketWithProxy;\n\t}\n\n\tconst ctor = (globalThis as { WebSocket?: unknown }).WebSocket;\n\tif (typeof ctor !== \"function\") return null;\n\treturn ctor as unknown as WebSocketConstructor;\n}\n\nclass WebSocketCloseError extends Error {\n\treadonly code?: number;\n\treadonly reason?: string;\n\treadonly wasClean?: boolean;\n\n\tconstructor(message: string, options?: { code?: number; reason?: string; wasClean?: boolean }) {\n\t\tsuper(message);\n\t\tthis.name = \"WebSocketCloseError\";\n\t\tthis.code = options?.code;\n\t\tthis.reason = options?.reason;\n\t\tthis.wasClean = options?.wasClean;\n\t}\n}\n\nfunction getWebSocketReadyState(socket: WebSocketLike): number | undefined {\n\tconst readyState = (socket as { readyState?: unknown }).readyState;\n\treturn typeof readyState === \"number\" ? readyState : undefined;\n}\n\nfunction isWebSocketReusable(socket: WebSocketLike): boolean {\n\tconst readyState = getWebSocketReadyState(socket);\n\t// If readyState is unavailable, assume the runtime keeps it open/reusable.\n\treturn readyState === undefined || readyState === 1;\n}\n\nfunction isWebSocketSessionExpired(entry: CachedWebSocketConnection): boolean {\n\treturn Date.now() - entry.createdAt >= SESSION_WEBSOCKET_MAX_AGE_MS;\n}\n\nfunction closeWebSocketSilently(socket: WebSocketLike, code = 1000, reason = \"done\"): void {\n\ttry {\n\t\tsocket.close(code, reason);\n\t} catch {}\n}\n\nfunction scheduleSessionWebSocketExpiry(sessionId: string, accountId: string, entry: CachedWebSocketConnection): void {\n\tif (entry.idleTimer) {\n\t\tclearTimeout(entry.idleTimer);\n\t}\n\tentry.idleTimer = setTimeout(() => {\n\t\tif (entry.busy) return;\n\t\tcloseWebSocketSilently(entry.socket, 1000, \"idle_timeout\");\n\t\tconst accountEntries = websocketSessionCache.get(sessionId);\n\t\tif (accountEntries?.get(accountId) === entry) accountEntries.delete(accountId);\n\t\tif (accountEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t}, SESSION_WEBSOCKET_CACHE_TTL_MS);\n}\n\nasync function connectWebSocket(\n\turl: string,\n\theaders: Headers,\n\tsignal?: AbortSignal,\n\tconnectTimeoutMs = DEFAULT_WEBSOCKET_CONNECT_TIMEOUT_MS,\n\tenv?: ProviderEnv,\n): Promise<WebSocketLike> {\n\tconst WebSocketCtor = await getWebSocketConstructor(env);\n\tif (!WebSocketCtor) {\n\t\tthrow new Error(\"WebSocket transport is not available in this runtime\");\n\t}\n\n\tconst wsHeaders = headersToRecord(headers);\n\tdelete wsHeaders[\"OpenAI-Beta\"];\n\n\treturn new Promise<WebSocketLike>((resolve, reject) => {\n\t\tlet settled = false;\n\t\tlet timeout: ReturnType<typeof setTimeout> | undefined;\n\t\tlet socket: WebSocketLike;\n\n\t\ttry {\n\t\t\tsocket = new WebSocketCtor(url, { headers: wsHeaders });\n\t\t} catch (error) {\n\t\t\treject(error instanceof Error ? error : new Error(String(error)));\n\t\t\treturn;\n\t\t}\n\n\t\tconst cleanup = () => {\n\t\t\tif (timeout) {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\ttimeout = undefined;\n\t\t\t}\n\t\t\tsocket.removeEventListener(\"open\", onOpen);\n\t\t\tsocket.removeEventListener(\"error\", onError);\n\t\t\tsocket.removeEventListener(\"close\", onClose);\n\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t};\n\t\tconst fail = (error: Error, closeReason?: string) => {\n\t\t\tif (settled) return;\n\t\t\tsettled = true;\n\t\t\tcleanup();\n\t\t\tif (closeReason) {\n\t\t\t\tcloseWebSocketSilently(socket, 1000, closeReason);\n\t\t\t}\n\t\t\treject(error);\n\t\t};\n\t\tconst onOpen: WebSocketListener = () => {\n\t\t\tif (settled) return;\n\t\t\tsettled = true;\n\t\t\tcleanup();\n\t\t\tresolve(socket);\n\t\t};\n\t\tconst onError: WebSocketListener = (event) => {\n\t\t\tfail(extractWebSocketError(event));\n\t\t};\n\t\tconst onClose: WebSocketListener = (event) => {\n\t\t\tfail(extractWebSocketCloseError(event));\n\t\t};\n\t\tconst onAbort = () => {\n\t\t\tfail(new Error(\"Request was aborted\"), \"aborted\");\n\t\t};\n\n\t\tsocket.addEventListener(\"open\", onOpen);\n\t\tsocket.addEventListener(\"error\", onError);\n\t\tsocket.addEventListener(\"close\", onClose);\n\t\tsignal?.addEventListener(\"abort\", onAbort);\n\n\t\tif (connectTimeoutMs > 0) {\n\t\t\ttimeout = setTimeout(() => {\n\t\t\t\tfail(new Error(`WebSocket connect timeout after ${connectTimeoutMs}ms`), \"connect_timeout\");\n\t\t\t}, connectTimeoutMs);\n\t\t}\n\t\tif (signal?.aborted) {\n\t\t\tonAbort();\n\t\t}\n\t});\n}\n\nasync function acquireWebSocket(\n\turl: string,\n\theaders: Headers,\n\tsessionId: string | undefined,\n\taccountId: string,\n\tsignal?: AbortSignal,\n\tconnectTimeoutMs?: number,\n\tenv?: ProviderEnv,\n): Promise<{\n\tsocket: WebSocketLike;\n\tentry?: CachedWebSocketConnection;\n\treused: boolean;\n\trelease: (options?: { keep?: boolean }) => void;\n}> {\n\tif (!sessionId) {\n\t\tconst socket = await connectWebSocket(url, headers, signal, connectTimeoutMs, env);\n\t\treturn {\n\t\t\tsocket,\n\t\t\treused: false,\n\t\t\trelease: () => closeWebSocketSilently(socket),\n\t\t};\n\t}\n\n\tlet accountEntries = websocketSessionCache.get(sessionId);\n\tconst cached = accountEntries?.get(accountId);\n\tif (cached) {\n\t\tif (cached.idleTimer) {\n\t\t\tclearTimeout(cached.idleTimer);\n\t\t\tcached.idleTimer = undefined;\n\t\t}\n\t\tif (!cached.busy && isWebSocketSessionExpired(cached)) {\n\t\t\tcloseWebSocketSilently(cached.socket, 1000, \"connection_age_limit\");\n\t\t\taccountEntries?.delete(accountId);\n\t\t\tif (accountEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t} else if (!cached.busy && isWebSocketReusable(cached.socket)) {\n\t\t\tcached.busy = true;\n\t\t\treturn {\n\t\t\t\tsocket: cached.socket,\n\t\t\t\tentry: cached,\n\t\t\t\treused: true,\n\t\t\t\trelease: ({ keep } = {}) => {\n\t\t\t\t\tif (!keep || !isWebSocketReusable(cached.socket)) {\n\t\t\t\t\t\tcloseWebSocketSilently(cached.socket);\n\t\t\t\t\t\tconst currentEntries = websocketSessionCache.get(sessionId);\n\t\t\t\t\t\tif (currentEntries?.get(accountId) === cached) currentEntries.delete(accountId);\n\t\t\t\t\t\tif (currentEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\tcached.busy = false;\n\t\t\t\t\tscheduleSessionWebSocketExpiry(sessionId, accountId, cached);\n\t\t\t\t},\n\t\t\t};\n\t\t}\n\t\tif (cached.busy) {\n\t\t\tconst socket = await connectWebSocket(url, headers, signal, connectTimeoutMs, env);\n\t\t\treturn {\n\t\t\t\tsocket,\n\t\t\t\treused: false,\n\t\t\t\trelease: () => {\n\t\t\t\t\tcloseWebSocketSilently(socket);\n\t\t\t\t},\n\t\t\t};\n\t\t}\n\t\tif (!isWebSocketReusable(cached.socket)) {\n\t\t\tcloseWebSocketSilently(cached.socket);\n\t\t\taccountEntries?.delete(accountId);\n\t\t\tif (accountEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t}\n\t}\n\n\tconst socket = await connectWebSocket(url, headers, signal, connectTimeoutMs, env);\n\tconst entry: CachedWebSocketConnection = { socket, busy: true, createdAt: Date.now() };\n\taccountEntries = websocketSessionCache.get(sessionId);\n\tif (!accountEntries) {\n\t\taccountEntries = new Map();\n\t\twebsocketSessionCache.set(sessionId, accountEntries);\n\t}\n\taccountEntries.set(accountId, entry);\n\treturn {\n\t\tsocket,\n\t\tentry,\n\t\treused: false,\n\t\trelease: ({ keep } = {}) => {\n\t\t\tif (!keep || !isWebSocketReusable(entry.socket)) {\n\t\t\t\tcloseWebSocketSilently(entry.socket);\n\t\t\t\tif (entry.idleTimer) clearTimeout(entry.idleTimer);\n\t\t\t\tconst currentEntries = websocketSessionCache.get(sessionId);\n\t\t\t\tif (currentEntries?.get(accountId) === entry) currentEntries.delete(accountId);\n\t\t\t\tif (currentEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tentry.busy = false;\n\t\t\tscheduleSessionWebSocketExpiry(sessionId, accountId, entry);\n\t\t},\n\t};\n}\n\nfunction extractWebSocketError(event: unknown): Error {\n\tif (event && typeof event === \"object\") {\n\t\tconst message = \"message\" in event ? (event as { message?: unknown }).message : undefined;\n\t\tif (typeof message === \"string\" && message.length > 0) {\n\t\t\treturn new Error(message);\n\t\t}\n\n\t\tconst nestedError = \"error\" in event ? (event as { error?: unknown }).error : undefined;\n\t\tif (nestedError instanceof Error && nestedError.message.length > 0) {\n\t\t\treturn nestedError;\n\t\t}\n\t\tif (nestedError && typeof nestedError === \"object\" && \"message\" in nestedError) {\n\t\t\tconst nestedMessage = (nestedError as { message?: unknown }).message;\n\t\t\tif (typeof nestedMessage === \"string\" && nestedMessage.length > 0) {\n\t\t\t\treturn new Error(nestedMessage);\n\t\t\t}\n\t\t}\n\t}\n\treturn new Error(\"WebSocket error\");\n}\n\nfunction extractWebSocketCloseError(event: unknown): Error {\n\tif (event && typeof event === \"object\") {\n\t\tconst code = \"code\" in event ? (event as { code?: unknown }).code : undefined;\n\t\tconst reason = \"reason\" in event ? (event as { reason?: unknown }).reason : undefined;\n\t\tconst wasClean = \"wasClean\" in event ? (event as { wasClean?: unknown }).wasClean : undefined;\n\t\tconst codeText = typeof code === \"number\" ? ` ${code}` : \"\";\n\t\tlet reasonText = typeof reason === \"string\" && reason.length > 0 ? ` ${reason}` : \"\";\n\t\tif (!reasonText && code === WEBSOCKET_MESSAGE_TOO_BIG_CLOSE_CODE) {\n\t\t\treasonText = \" message too big\";\n\t\t}\n\t\treturn new WebSocketCloseError(`WebSocket closed${codeText}${reasonText}`.trim(), {\n\t\t\tcode: typeof code === \"number\" ? code : undefined,\n\t\t\treason: typeof reason === \"string\" && reason.length > 0 ? reason : undefined,\n\t\t\twasClean: typeof wasClean === \"boolean\" ? wasClean : undefined,\n\t\t});\n\t}\n\treturn new Error(\"WebSocket closed\");\n}\n\nasync function decodeWebSocketData(data: unknown): Promise<string | null> {\n\tif (typeof data === \"string\") return data;\n\tif (data instanceof ArrayBuffer) {\n\t\treturn new TextDecoder().decode(new Uint8Array(data));\n\t}\n\tif (ArrayBuffer.isView(data)) {\n\t\tconst view = data as ArrayBufferView;\n\t\treturn new TextDecoder().decode(new Uint8Array(view.buffer, view.byteOffset, view.byteLength));\n\t}\n\tif (data && typeof data === \"object\" && \"arrayBuffer\" in data) {\n\t\tconst blobLike = data as { arrayBuffer: () => Promise<ArrayBuffer> };\n\t\tconst arrayBuffer = await blobLike.arrayBuffer();\n\t\treturn new TextDecoder().decode(new Uint8Array(arrayBuffer));\n\t}\n\treturn null;\n}\n\nasync function* parseWebSocket(\n\tsocket: WebSocketLike,\n\tsignal?: AbortSignal,\n\tidleTimeoutMs?: number,\n): AsyncGenerator<Record<string, unknown>> {\n\tconst queue: Record<string, unknown>[] = [];\n\tlet pending: (() => void) | null = null;\n\tlet done = false;\n\tlet failed: Error | null = null;\n\tlet sawCompletion = false;\n\n\tconst wake = () => {\n\t\tif (!pending) return;\n\t\tconst resolve = pending;\n\t\tpending = null;\n\t\tresolve();\n\t};\n\n\tconst onMessage: WebSocketListener = (event) => {\n\t\tvoid (async () => {\n\t\t\tlet text: string | null = null;\n\t\t\ttry {\n\t\t\t\tif (!event || typeof event !== \"object\" || !(\"data\" in event)) return;\n\t\t\t\ttext = await decodeWebSocketData((event as { data?: unknown }).data);\n\t\t\t\tif (!text) return;\n\t\t\t\tconst parsed = JSON.parse(text) as Record<string, unknown>;\n\t\t\t\tconst type = typeof parsed.type === \"string\" ? parsed.type : \"\";\n\t\t\t\tif (type === \"response.completed\" || type === \"response.done\" || type === \"response.incomplete\") {\n\t\t\t\t\tsawCompletion = true;\n\t\t\t\t\tdone = true;\n\t\t\t\t}\n\t\t\t\tqueue.push(parsed);\n\t\t\t\twake();\n\t\t\t} catch (cause) {\n\t\t\t\tfailed = new CodexProtocolError(`Invalid Codex WebSocket JSON: ${formatThrownValue(cause)}`, {\n\t\t\t\t\tcause,\n\t\t\t\t\tpayload: text,\n\t\t\t\t});\n\t\t\t\tdone = true;\n\t\t\t\twake();\n\t\t\t}\n\t\t})();\n\t};\n\n\tconst onError: WebSocketListener = (event) => {\n\t\tfailed = extractWebSocketError(event);\n\t\tdone = true;\n\t\twake();\n\t};\n\n\tconst onClose: WebSocketListener = (event) => {\n\t\tif (sawCompletion) {\n\t\t\tdone = true;\n\t\t\twake();\n\t\t\treturn;\n\t\t}\n\t\tif (!failed) {\n\t\t\tfailed = extractWebSocketCloseError(event);\n\t\t}\n\t\tdone = true;\n\t\twake();\n\t};\n\n\tconst onAbort = () => {\n\t\tfailed = new Error(\"Request was aborted\");\n\t\tdone = true;\n\t\twake();\n\t};\n\n\tsocket.addEventListener(\"message\", onMessage);\n\tsocket.addEventListener(\"error\", onError);\n\tsocket.addEventListener(\"close\", onClose);\n\tsignal?.addEventListener(\"abort\", onAbort);\n\n\ttry {\n\t\twhile (true) {\n\t\t\tif (signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (queue.length > 0) {\n\t\t\t\tyield queue.shift()!;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (done) break;\n\t\t\tlet timeout: ReturnType<typeof setTimeout> | undefined;\n\t\t\tawait new Promise<void>((resolve, reject) => {\n\t\t\t\tpending = resolve;\n\t\t\t\tif (idleTimeoutMs !== undefined && idleTimeoutMs > 0) {\n\t\t\t\t\ttimeout = setTimeout(() => {\n\t\t\t\t\t\tconst error = new Error(`WebSocket idle timeout after ${idleTimeoutMs}ms`);\n\t\t\t\t\t\tfailed = error;\n\t\t\t\t\t\tdone = true;\n\t\t\t\t\t\tpending = null;\n\t\t\t\t\t\tcloseWebSocketSilently(socket, 1000, \"idle_timeout\");\n\t\t\t\t\t\treject(error);\n\t\t\t\t\t}, idleTimeoutMs);\n\t\t\t\t}\n\t\t\t}).finally(() => {\n\t\t\t\tif (timeout) {\n\t\t\t\t\tclearTimeout(timeout);\n\t\t\t\t}\n\t\t\t});\n\t\t}\n\n\t\tif (failed) {\n\t\t\tthrow failed;\n\t\t}\n\t\tif (!sawCompletion) {\n\t\t\tthrow new Error(\"WebSocket stream closed before response.completed\");\n\t\t}\n\t} finally {\n\t\tsocket.removeEventListener(\"message\", onMessage);\n\t\tsocket.removeEventListener(\"error\", onError);\n\t\tsocket.removeEventListener(\"close\", onClose);\n\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t}\n}\n\nfunction requestBodyWithoutInput(body: RequestBody): RequestBody {\n\tconst { input: _input, previous_response_id: _previousResponseId, ...rest } = body;\n\treturn rest;\n}\n\nfunction responseInputsEqual(a: ResponseInput | undefined, b: ResponseInput | undefined): boolean {\n\treturn JSON.stringify(a ?? []) === JSON.stringify(b ?? []);\n}\n\nfunction requestBodiesMatchExceptInput(a: RequestBody, b: RequestBody): boolean {\n\treturn JSON.stringify(requestBodyWithoutInput(a)) === JSON.stringify(requestBodyWithoutInput(b));\n}\n\nfunction getCachedWebSocketInputDelta(\n\tbody: RequestBody,\n\tcontinuation: CachedWebSocketContinuationState,\n): ResponseInput | undefined {\n\tif (!requestBodiesMatchExceptInput(body, continuation.lastRequestBody)) {\n\t\treturn undefined;\n\t}\n\n\tconst currentInput = body.input ?? [];\n\tconst baseline = [...(continuation.lastRequestBody.input ?? []), ...continuation.lastResponseItems];\n\tif (currentInput.length < baseline.length) {\n\t\treturn undefined;\n\t}\n\n\tconst prefix = currentInput.slice(0, baseline.length);\n\tif (!responseInputsEqual(prefix, baseline)) {\n\t\treturn undefined;\n\t}\n\n\treturn currentInput.slice(baseline.length);\n}\n\nfunction buildCachedWebSocketRequestBody(entry: CachedWebSocketConnection, body: RequestBody): RequestBody {\n\tconst continuation = entry.continuation;\n\tif (!continuation) {\n\t\treturn body;\n\t}\n\n\tconst delta = getCachedWebSocketInputDelta(body, continuation);\n\tif (!delta || !continuation.lastResponseId) {\n\t\tentry.continuation = undefined;\n\t\treturn body;\n\t}\n\n\treturn {\n\t\t...body,\n\t\tprevious_response_id: continuation.lastResponseId,\n\t\tinput: delta,\n\t};\n}\n\nasync function* startWebSocketOutputOnFirstEvent(\n\tevents: AsyncIterable<ResponseStreamEvent>,\n\tonStart: () => void,\n): AsyncGenerator<ResponseStreamEvent> {\n\tlet started = false;\n\tfor await (const event of events) {\n\t\tif (!started) {\n\t\t\tstarted = true;\n\t\t\tonStart();\n\t\t}\n\t\tyield event;\n\t}\n}\n\nasync function processWebSocketStream(\n\turl: string,\n\tbody: RequestBody,\n\theaders: Headers,\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n\tmodel: Model<\"openai-codex-responses\">,\n\tonStart: () => void,\n\tidleTimeoutMs: number | undefined,\n\twebsocketConnectTimeoutMs: number | undefined,\n\tcacheSessionId: string | undefined,\n\taccountId: string,\n\tgrammarToolInputProperties: ReadonlyMap<string, string>,\n\toptions?: OpenAICodexResponsesOptions,\n): Promise<void> {\n\tconst { socket, entry, reused, release } = await acquireWebSocket(\n\t\turl,\n\t\theaders,\n\t\tcacheSessionId,\n\t\taccountId,\n\t\toptions?.signal,\n\t\twebsocketConnectTimeoutMs,\n\t\toptions?.env,\n\t);\n\tlet keepConnection = true;\n\tconst useCachedContext = options?.transport === \"websocket-cached\" || options?.transport === \"auto\";\n\t// ChatGPT Codex Responses rejects `store: true` (\"Store must be set to false\").\n\t// WebSocket continuation still works via connection-scoped previous_response_id state.\n\tconst fullBody = body;\n\tconst requestBody = useCachedContext && entry ? buildCachedWebSocketRequestBody(entry, fullBody) : fullBody;\n\tconst stats = cacheSessionId ? getOrCreateWebSocketDebugStats(cacheSessionId) : undefined;\n\tif (stats) {\n\t\tstats.requests++;\n\t\tif (reused) stats.connectionsReused++;\n\t\telse stats.connectionsCreated++;\n\t\tif (useCachedContext) stats.cachedContextRequests++;\n\t\tif (requestBody.store === true) stats.storeTrueRequests++;\n\t\tstats.lastInputItems = requestBody.input?.length ?? 0;\n\t\tif (requestBody.previous_response_id) {\n\t\t\tstats.deltaRequests++;\n\t\t\tstats.lastDeltaInputItems = requestBody.input?.length ?? 0;\n\t\t\tstats.lastPreviousResponseId = requestBody.previous_response_id;\n\t\t} else {\n\t\t\tstats.fullContextRequests++;\n\t\t\tstats.lastDeltaInputItems = undefined;\n\t\t\tstats.lastPreviousResponseId = undefined;\n\t\t}\n\t}\n\ttry {\n\t\tsocket.send(JSON.stringify({ type: \"response.create\", ...requestBody }));\n\t\tawait processResponsesStream(\n\t\t\tstartWebSocketOutputOnFirstEvent(\n\t\t\t\tmapCodexEvents(parseWebSocket(socket, options?.signal, idleTimeoutMs), output),\n\t\t\t\tonStart,\n\t\t\t),\n\t\t\toutput,\n\t\t\tstream,\n\t\t\tmodel,\n\t\t\t{\n\t\t\t\tserviceTier: resolveRequestedServiceTier(model, options?.serviceTier),\n\t\t\t\tgrammarToolInputProperties,\n\t\t\t\tresolveServiceTier: resolveCodexServiceTier,\n\t\t\t\tapplyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),\n\t\t\t},\n\t\t);\n\t\tif (options?.signal?.aborted) {\n\t\t\tkeepConnection = false;\n\t\t} else if (useCachedContext && entry && output.responseId) {\n\t\t\tconst responseItems = convertResponsesMessages(model, { messages: [output] }, CODEX_TOOL_CALL_PROVIDERS, {\n\t\t\t\tincludeSystemPrompt: false,\n\t\t\t\tgrammarToolInputProperties,\n\t\t\t}).filter((item) => item.type !== \"function_call_output\" && item.type !== \"custom_tool_call_output\");\n\t\t\tentry.continuation = {\n\t\t\t\tlastRequestBody: fullBody,\n\t\t\t\tlastResponseId: output.responseId,\n\t\t\t\tlastResponseItems: responseItems,\n\t\t\t};\n\t\t}\n\t} catch (error) {\n\t\tif (entry) {\n\t\t\tentry.continuation = undefined;\n\t\t}\n\t\tkeepConnection = false;\n\t\tthrow error;\n\t} finally {\n\t\trelease({ keep: keepConnection });\n\t}\n}\n\n// ============================================================================\n// Error Handling\n// ============================================================================\n\nasync function parseErrorResponse(response: Response): Promise<{ message: string; friendlyMessage?: string }> {\n\tconst raw = await response.text();\n\tlet message = raw || response.statusText || \"Request failed\";\n\tlet friendlyMessage: string | undefined;\n\n\ttry {\n\t\tconst parsed = JSON.parse(raw) as {\n\t\t\terror?: { code?: string; type?: string; message?: string; plan_type?: string; resets_at?: number };\n\t\t};\n\t\tconst err = parsed?.error;\n\t\tif (err) {\n\t\t\tconst code = err.code || err.type || \"\";\n\t\t\tif (/usage_limit_reached|usage_not_included|rate_limit_exceeded/i.test(code) || response.status === 429) {\n\t\t\t\tconst plan = err.plan_type ? ` (${err.plan_type.toLowerCase()} plan)` : \"\";\n\t\t\t\tconst mins = err.resets_at\n\t\t\t\t\t? Math.max(0, Math.round((err.resets_at * 1000 - Date.now()) / 60000))\n\t\t\t\t\t: undefined;\n\t\t\t\tconst when = mins !== undefined ? ` Try again in ~${mins} min.` : \"\";\n\t\t\t\tfriendlyMessage = `You have hit your ChatGPT usage limit${plan}.${when}`.trim();\n\t\t\t}\n\t\t\tmessage = err.message || friendlyMessage || message;\n\t\t}\n\t} catch {}\n\n\treturn { message, friendlyMessage };\n}\n\n// ============================================================================\n// Auth & Headers\n// ============================================================================\n\nfunction extractAccountId(token: string): string {\n\ttry {\n\t\tconst parts = token.split(\".\");\n\t\tif (parts.length !== 3) throw new Error(\"Invalid token\");\n\t\tconst payload = JSON.parse(atob(parts[1]));\n\t\tconst accountId = payload?.[JWT_CLAIM_PATH]?.chatgpt_account_id;\n\t\tif (!accountId) throw new Error(\"No account ID in token\");\n\t\treturn accountId;\n\t} catch {\n\t\tthrow new Error(\"Failed to extract accountId from token\");\n\t}\n}\n\nfunction buildBaseCodexHeaders(\n\tinitHeaders: Record<string, string> | undefined,\n\tadditionalHeaders: ProviderHeaders | undefined,\n\taccountId: string,\n\ttoken: string,\n): Headers {\n\tconst headers = new Headers(initHeaders);\n\tfor (const [key, value] of Object.entries(additionalHeaders || {})) {\n\t\tif (value === null) {\n\t\t\theaders.delete(key);\n\t\t} else {\n\t\t\theaders.set(key, value);\n\t\t}\n\t}\n\theaders.set(\"Authorization\", `Bearer ${token}`);\n\theaders.set(\"chatgpt-account-id\", accountId);\n\theaders.set(\"originator\", \"pi\");\n\theaders.set(\"User-Agent\", getPiUserAgent());\n\treturn headers;\n}\n\nfunction buildSSEHeaders(\n\tinitHeaders: Record<string, string> | undefined,\n\tadditionalHeaders: ProviderHeaders | undefined,\n\taccountId: string,\n\ttoken: string,\n\tsessionId?: string,\n): Headers {\n\tconst headers = buildBaseCodexHeaders(initHeaders, additionalHeaders, accountId, token);\n\theaders.set(\"OpenAI-Beta\", \"responses=experimental\");\n\theaders.set(\"accept\", \"text/event-stream\");\n\theaders.set(\"content-type\", \"application/json\");\n\n\tif (sessionId) {\n\t\theaders.set(\"session-id\", sessionId);\n\t\theaders.set(\"x-client-request-id\", sessionId);\n\t}\n\n\treturn headers;\n}\n\nfunction buildWebSocketHeaders(\n\tinitHeaders: Record<string, string> | undefined,\n\tadditionalHeaders: ProviderHeaders | undefined,\n\taccountId: string,\n\ttoken: string,\n\trequestId: string,\n): Headers {\n\tconst headers = buildBaseCodexHeaders(initHeaders, additionalHeaders, accountId, token);\n\theaders.delete(\"accept\");\n\theaders.delete(\"content-type\");\n\theaders.delete(\"OpenAI-Beta\");\n\theaders.delete(\"openai-beta\");\n\theaders.set(\"OpenAI-Beta\", OPENAI_BETA_RESPONSES_WEBSOCKETS);\n\theaders.set(\"x-client-request-id\", requestId);\n\theaders.set(\"session-id\", requestId);\n\treturn headers;\n}\n"]}
+{"version":3,"file":"openai-codex-responses.d.ts","sourceRoot":"","sources":["../../src/api/openai-codex-responses.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAEX,6BAA6B,EAG7B,MAAM,yCAAyC,CAAC;AAIjD,OAAO,KAAK,EAMX,mBAAmB,EACnB,cAAc,EACd,aAAa,EAGb,MAAM,aAAa,CAAC;AAgErB,MAAM,WAAW,2BAA4B,SAAQ,aAAa;IACjE,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;IACnF,gBAAgB,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,UAAU,GAAG,KAAK,GAAG,IAAI,GAAG,IAAI,CAAC;IACzE,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,aAAa,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IAC1C,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,CAAC;CAC1C;AAwJD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,wBAAwB,EAAE,2BAA2B,CA0QxF,CAAC;AAEF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,wBAAwB,EAAE,mBAAmB,CAqBtF,CAAC;AAkYF,MAAM,WAAW,8BAA8B;IAC9C,QAAQ,EAAE,MAAM,CAAC;IACjB,kBAAkB,EAAE,MAAM,CAAC;IAC3B,iBAAiB,EAAE,MAAM,CAAC;IAC1B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,iBAAiB,EAAE,MAAM,CAAC;IAC1B,mBAAmB,EAAE,MAAM,CAAC;IAC5B,aAAa,EAAE,MAAM,CAAC;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,sBAAsB,CAAC,EAAE,MAAM,CAAC;IAChC,iBAAiB,EAAE,MAAM,CAAC;IAC1B,YAAY,EAAE,MAAM,CAAC;IACrB,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,kBAAkB,CAAC,EAAE,MAAM,CAAC;CAC5B;AA0BD,wBAAgB,iCAAiC,CAAC,SAAS,EAAE,MAAM,GAAG,8BAA8B,GAAG,SAAS,CAG/G;AAED,wBAAgB,mCAAmC,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAQ5E;AAED,wBAAgB,iCAAiC,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAc1E","sourcesContent":["import type * as NodeZlib from \"node:zlib\";\nimport type {\n\tTool as OpenAITool,\n\tResponseCreateParamsStreaming,\n\tResponseInput,\n\tResponseStreamEvent,\n} from \"openai/resources/responses/responses.js\";\n\nimport { clampThinkingLevel } from \"../models.ts\";\nimport { registerSessionResourceCleanup } from \"../session-resources.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tModel,\n\tProviderEnv,\n\tProviderHeaders,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tTranscriptContext,\n\tUsage,\n} from \"../types.ts\";\nimport { combineAbortSignals } from \"../utils/abort-signals.ts\";\nimport {\n\tappendAssistantMessageDiagnostic,\n\tcreateAssistantMessageDiagnostic,\n\tformatThrownValue,\n} from \"../utils/diagnostics.ts\";\nimport { formatProviderError, normalizeProviderError } from \"../utils/error-body.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { headersToRecord } from \"../utils/headers.ts\";\nimport { resolveHttpProxyUrlForTarget } from \"../utils/node-http-proxy.ts\";\nimport { getPiUserAgent } from \"../utils/pi-user-agent.ts\";\nimport { createStreamDeadline, type StreamDeadlineHandle, withStreamDeadline } from \"../utils/stream-deadline.ts\";\nimport { getSystemMessageText } from \"../utils/text.ts\";\nimport {\n\tgetDeclaredTools,\n\tgetInitialSystemMessage,\n\tnormalizeContext,\n\tresolveTranscript,\n\tresolveTranscriptTools,\n} from \"../utils/transcript.ts\";\nimport { uuidv7 } from \"../utils/uuid.ts\";\nimport { createGrammarToolInputProperties } from \"./constrained-sampling.ts\";\nimport { clampOpenAIPromptCacheKey } from \"./openai-prompt-cache.ts\";\nimport {\n\tassertPayloadPreservesFastRoute,\n\tconvertResponsesMessages,\n\tconvertResponsesTools,\n\tprocessResponsesStream,\n\tresolveRequestedServiceTier,\n} from \"./openai-responses-shared.ts\";\nimport { buildBaseOptions } from \"./simple-options.ts\";\n\n// ============================================================================\n// Configuration\n// ============================================================================\n\nconst DEFAULT_CODEX_BASE_URL = \"https://chatgpt.com/backend-api\";\nconst JWT_CLAIM_PATH = \"https://api.openai.com/auth\" as const;\nconst DEFAULT_MAX_RETRIES = 0;\nconst BASE_DELAY_MS = 1000;\nconst DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\nconst DEFAULT_WEBSOCKET_CONNECT_TIMEOUT_MS = 15_000;\n// The Codex backend accepts zstd-compressed request bodies on the SSE responses\n// endpoint (the same endpoint the official Codex client compresses against).\nconst REQUEST_COMPRESSION_ZSTD_LEVEL = 3;\nconst CODEX_TOOL_CALL_PROVIDERS = new Set([\"openai\", \"openai-codex\", \"opencode\"]);\nconst WEBSOCKET_MESSAGE_TOO_BIG_CLOSE_CODE = 1009;\nconst WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE = \"websocket_connection_limit_reached\";\nconst PREVIOUS_RESPONSE_NOT_FOUND_CODE = \"previous_response_not_found\";\n\nconst CODEX_RESPONSE_STATUSES = new Set<CodexResponseStatus>([\n\t\"completed\",\n\t\"incomplete\",\n\t\"failed\",\n\t\"cancelled\",\n\t\"queued\",\n\t\"in_progress\",\n]);\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface OpenAICodexResponsesOptions extends StreamOptions {\n\treasoningEffort?: \"none\" | \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\";\n\treasoningSummary?: \"auto\" | \"concise\" | \"detailed\" | \"off\" | \"on\" | null;\n\tserviceTier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\ttextVerbosity?: \"low\" | \"medium\" | \"high\";\n\ttoolChoice?: \"auto\" | \"none\" | \"required\";\n}\n\ntype CodexResponseStatus = \"completed\" | \"incomplete\" | \"failed\" | \"cancelled\" | \"queued\" | \"in_progress\";\n\ninterface RequestBody {\n\tmodel: string;\n\tstore?: boolean;\n\tstream?: boolean;\n\tinstructions?: string;\n\tprevious_response_id?: string;\n\tinput?: ResponseInput;\n\ttools?: OpenAITool[];\n\ttool_choice?: OpenAICodexResponsesOptions[\"toolChoice\"];\n\tparallel_tool_calls?: boolean;\n\ttemperature?: number;\n\treasoning?: { effort?: string; summary?: string };\n\tservice_tier?: ResponseCreateParamsStreaming[\"service_tier\"];\n\ttext?: { verbosity?: string };\n\tinclude?: string[];\n\tprompt_cache_key?: string;\n\t[key: string]: unknown;\n}\n\ntype SuccessfulAssistantMessage = AssistantMessage & { stopReason: \"stop\" | \"length\" | \"toolUse\" };\n\nfunction assertSuccessfulOutput(output: AssistantMessage): asserts output is SuccessfulAssistantMessage {\n\tif (output.stopReason === \"pending\") {\n\t\tthrow new Error(\"Codex stream ended without a stop reason\");\n\t}\n\tif (output.stopReason === \"error\" || output.stopReason === \"aborted\") {\n\t\tthrow new Error(output.errorMessage || \"An unknown error occurred\");\n\t}\n}\n\n// ============================================================================\n// Retry Helpers\n// ============================================================================\n\nfunction isTerminalRateLimitError(errorText: string): boolean {\n\treturn /GoUsageLimitError|FreeUsageLimitError|Monthly usage limit reached|available balance|insufficient_quota|out of budget|quota exceeded|billing/i.test(\n\t\terrorText,\n\t);\n}\n\nfunction isRetryableError(status: number, errorText: string): boolean {\n\tif (status === 429 && isTerminalRateLimitError(errorText)) {\n\t\treturn false;\n\t}\n\tif (status === 429 || status === 500 || status === 502 || status === 503 || status === 504) {\n\t\treturn true;\n\t}\n\treturn /rate.?limit|overloaded|service.?unavailable|upstream.?connect|connection.?refused/i.test(errorText);\n}\n\nfunction getRetryAfterDelayMs(headers: Headers): number | undefined {\n\tconst retryAfterMs = headers.get(\"retry-after-ms\");\n\tif (retryAfterMs !== null) {\n\t\tconst millis = Number(retryAfterMs);\n\t\tif (Number.isFinite(millis)) {\n\t\t\treturn Math.max(0, millis);\n\t\t}\n\t}\n\n\tconst retryAfter = headers.get(\"retry-after\");\n\tif (!retryAfter) {\n\t\treturn undefined;\n\t}\n\n\tconst seconds = Number(retryAfter);\n\tif (Number.isFinite(seconds)) {\n\t\treturn Math.max(0, seconds * 1000);\n\t}\n\n\tconst date = Date.parse(retryAfter);\n\tif (!Number.isNaN(date)) {\n\t\treturn Math.max(0, date - Date.now());\n\t}\n\n\treturn undefined;\n}\n\nclass RetryDelayExceededError extends Error {}\n\nfunction validateRetryDelayMs(delayMs: number, options?: StreamOptions): number {\n\tconst maxRetryDelayMs = options?.maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\tif (maxRetryDelayMs > 0 && delayMs > maxRetryDelayMs) {\n\t\tthrow new RetryDelayExceededError(\n\t\t\t`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxRetryDelayMs / 1000)}s)`,\n\t\t);\n\t}\n\treturn delayMs;\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new Error(\"Request was aborted\"));\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\"abort\", () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(new Error(\"Request was aborted\"));\n\t\t});\n\t});\n}\n\nfunction normalizeTimeoutMs(value: number | undefined): number | undefined {\n\tif (value === undefined) return undefined;\n\tif (!Number.isFinite(value) || value < 0) {\n\t\tthrow new Error(`Invalid timeoutMs: ${String(value)}`);\n\t}\n\treturn Math.floor(value);\n}\n\n// ============================================================================\n// Request Compression\n// ============================================================================\n\ntype ProcessWithBuiltinModule = typeof process & {\n\tgetBuiltinModule?: (id: \"node:zlib\") => typeof NodeZlib;\n};\n\nfunction loadNodeZlib(): typeof NodeZlib | null {\n\tif (typeof process === \"undefined\" || !(process.versions?.node || process.versions?.bun)) {\n\t\treturn null;\n\t}\n\treturn (process as ProcessWithBuiltinModule).getBuiltinModule?.(\"node:zlib\") ?? null;\n}\n\n// Returns the zstd-compressed body bytes, or null when compression is\n// unavailable (browser/Vite builds). Callers fall back to sending the\n// uncompressed JSON when this returns null.\nfunction compressRequestBodyZstd(bodyJson: string): Uint8Array | null {\n\tconst zlib = loadNodeZlib();\n\tif (!zlib || typeof zlib.zstdCompressSync !== \"function\") {\n\t\treturn null;\n\t}\n\ttry {\n\t\tconst compressed = zlib.zstdCompressSync(bodyJson, {\n\t\t\tparams: { [zlib.constants.ZSTD_c_compressionLevel]: REQUEST_COMPRESSION_ZSTD_LEVEL },\n\t\t});\n\t\treturn new Uint8Array(compressed.buffer, compressed.byteOffset, compressed.byteLength);\n\t} catch {\n\t\treturn null;\n\t}\n}\n\n// ============================================================================\n// Main Stream Function\n// ============================================================================\n\nexport const stream: StreamFunction<\"openai-codex-responses\", OpenAICodexResponsesOptions> = (\n\tmodel: Model<\"openai-codex-responses\">,\n\tcontext: TranscriptContext,\n\toptions?: OpenAICodexResponsesOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\tconst normalizedContext = resolveTranscript(context, model.compat?.supportsMidConvoSystemMessages);\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: \"openai-codex-responses\" as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"pending\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\tconst streamDeadline = createStreamDeadline(options?.streamDeadlineMs, options?.signal);\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey;\n\t\t\tif (!apiKey) {\n\t\t\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t\t\t}\n\n\t\t\tconst accountId = extractAccountId(apiKey);\n\t\t\tconst grammarToolInputProperties = createGrammarToolInputProperties(\n\t\t\t\tgetDeclaredTools(normalizedContext.messages),\n\t\t\t\tmodel.compat?.supportsOpenAIGrammarTools ?? false,\n\t\t\t);\n\t\t\tconst cacheSessionId = options?.cacheRetention === \"none\" ? undefined : options?.sessionId;\n\t\t\tconst codexSessionId = clampOpenAIPromptCacheKey(cacheSessionId);\n\t\t\tlet body = buildRequestBody(model, normalizedContext, options, codexSessionId, grammarToolInputProperties);\n\t\t\tconst nextBody = await options?.onPayload?.(body, model);\n\t\t\tif (nextBody !== undefined) {\n\t\t\t\tbody = nextBody as RequestBody;\n\t\t\t}\n\t\t\tassertPayloadPreservesFastRoute(model, body);\n\t\t\tconst websocketRequestId = codexSessionId || uuidv7();\n\t\t\tconst sseHeaders = buildSSEHeaders(model.headers, options?.headers, accountId, apiKey, codexSessionId);\n\t\t\tconst websocketHeaders = buildWebSocketHeaders(\n\t\t\t\tmodel.headers,\n\t\t\t\toptions?.headers,\n\t\t\t\taccountId,\n\t\t\t\tapiKey,\n\t\t\t\twebsocketRequestId,\n\t\t\t);\n\t\t\tconst bodyJson = JSON.stringify(body);\n\t\t\tconst httpTimeoutMs = normalizeTimeoutMs(options?.timeoutMs);\n\t\t\tconst websocketConnectTimeoutMs = normalizeTimeoutMs(options?.websocketConnectTimeoutMs);\n\t\t\tconst transport = options?.transport || \"auto\";\n\t\t\tlet startEmitted = false;\n\t\t\tconst websocketDisabledForSession = transport !== \"sse\" && isWebSocketSseFallbackActive(cacheSessionId);\n\t\t\tif (websocketDisabledForSession) {\n\t\t\t\trecordWebSocketSseFallback(cacheSessionId);\n\t\t\t}\n\n\t\t\tif (transport !== \"sse\" && !websocketDisabledForSession) {\n\t\t\t\tlet websocketStarted = false;\n\t\t\t\tlet retriedWebSocketConnectionLimit = false;\n\t\t\t\tlet retriedMissingWebSocketContinuation = false;\n\t\t\t\twhile (true) {\n\t\t\t\t\twebsocketStarted = false;\n\t\t\t\t\ttry {\n\t\t\t\t\t\tawait processWebSocketStream(\n\t\t\t\t\t\t\tresolveCodexWebSocketUrl(model.baseUrl),\n\t\t\t\t\t\t\tbody,\n\t\t\t\t\t\t\twebsocketHeaders,\n\t\t\t\t\t\t\toutput,\n\t\t\t\t\t\t\tstream,\n\t\t\t\t\t\t\tmodel,\n\t\t\t\t\t\t\t() => {\n\t\t\t\t\t\t\t\twebsocketStarted = true;\n\t\t\t\t\t\t\t\tif (!startEmitted) {\n\t\t\t\t\t\t\t\t\tstartEmitted = true;\n\t\t\t\t\t\t\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\tstreamDeadline.deadlineMs,\n\t\t\t\t\t\t\twebsocketConnectTimeoutMs,\n\t\t\t\t\t\t\tcacheSessionId,\n\t\t\t\t\t\t\taccountId,\n\t\t\t\t\t\t\tgrammarToolInputProperties,\n\t\t\t\t\t\t\toptions,\n\t\t\t\t\t\t);\n\n\t\t\t\t\t\tif (options?.signal?.aborted) {\n\t\t\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t\t\t}\n\t\t\t\t\t\tassertSuccessfulOutput(output);\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"done\",\n\t\t\t\t\t\t\treason: output.stopReason,\n\t\t\t\t\t\t\tmessage: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t\tstream.end();\n\t\t\t\t\t\treturn;\n\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\tconst aborted = options?.signal?.aborted;\n\t\t\t\t\t\tconst connectionLimitBeforeStart = !websocketStarted && isWebSocketConnectionLimitReachedError(error);\n\t\t\t\t\t\tconst previousResponseNotFound = isPreviousResponseNotFoundError(error);\n\t\t\t\t\t\tif (!aborted && previousResponseNotFound && !retriedMissingWebSocketContinuation) {\n\t\t\t\t\t\t\tretriedMissingWebSocketContinuation = true;\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tif (!aborted && connectionLimitBeforeStart && !retriedWebSocketConnectionLimit) {\n\t\t\t\t\t\t\tretriedWebSocketConnectionLimit = true;\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tif (aborted || (isCodexNonTransportError(error) && !connectionLimitBeforeStart)) {\n\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tappendAssistantMessageDiagnostic(\n\t\t\t\t\t\t\toutput,\n\t\t\t\t\t\t\tcreateAssistantMessageDiagnostic(\"provider_transport_failure\", error, {\n\t\t\t\t\t\t\t\tconfiguredTransport: transport,\n\t\t\t\t\t\t\t\t...(websocketStarted ? {} : { fallbackTransport: \"sse\" }),\n\t\t\t\t\t\t\t\teventsEmitted: websocketStarted,\n\t\t\t\t\t\t\t\tphase: websocketStarted ? \"after_message_stream_start\" : \"before_message_stream_start\",\n\t\t\t\t\t\t\t\trequestBytes: new TextEncoder().encode(bodyJson).byteLength,\n\t\t\t\t\t\t\t}),\n\t\t\t\t\t\t);\n\t\t\t\t\t\trecordWebSocketFailure(cacheSessionId, error);\n\t\t\t\t\t\tif (websocketStarted) {\n\t\t\t\t\t\t\tthrow error;\n\t\t\t\t\t\t}\n\t\t\t\t\t\trecordWebSocketSseFallback(cacheSessionId);\n\t\t\t\t\t\tbreak;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\t// Compress the request body once for the SSE path. The Codex backend\n\t\t\t// decodes Content-Encoding: zstd; the WebSocket transport above sends the\n\t\t\t// uncompressed JSON frame, matching the official Codex client.\n\t\t\tconst compressedBody = compressRequestBodyZstd(bodyJson);\n\t\t\tif (compressedBody) {\n\t\t\t\tsseHeaders.set(\"content-encoding\", \"zstd\");\n\t\t\t}\n\t\t\tconst sseBody: Uint8Array | string = compressedBody ?? bodyJson;\n\n\t\t\t// Fetch with retry logic for rate limits and transient errors\n\t\t\tlet response: Response | undefined;\n\t\t\tlet lastError: Error | undefined;\n\t\t\tconst maxRetries = options?.maxRetries ?? DEFAULT_MAX_RETRIES;\n\n\t\t\tfor (let attempt = 0; attempt <= maxRetries; attempt++) {\n\t\t\t\tif (options?.signal?.aborted) {\n\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t}\n\n\t\t\t\ttry {\n\t\t\t\t\tconst headerTimeoutSignal =\n\t\t\t\t\t\thttpTimeoutMs !== undefined && httpTimeoutMs > 0 ? AbortSignal.timeout(httpTimeoutMs) : undefined;\n\t\t\t\t\tconst combinedSignal = combineAbortSignals([streamDeadline.signal, headerTimeoutSignal]);\n\t\t\t\t\ttry {\n\t\t\t\t\t\tresponse = await (options?.fetch ?? globalThis.fetch)(resolveCodexUrl(model.baseUrl), {\n\t\t\t\t\t\t\tmethod: \"POST\",\n\t\t\t\t\t\t\theaders: sseHeaders,\n\t\t\t\t\t\t\tbody: sseBody,\n\t\t\t\t\t\t\tsignal: combinedSignal.signal,\n\t\t\t\t\t\t});\n\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\tif (headerTimeoutSignal?.aborted && !options?.signal?.aborted) {\n\t\t\t\t\t\t\tthrow new Error(`Codex SSE response headers timed out after ${httpTimeoutMs}ms`);\n\t\t\t\t\t\t}\n\t\t\t\t\t\tthrow error;\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tcombinedSignal.cleanup();\n\t\t\t\t\t}\n\t\t\t\t\tawait options?.onResponse?.(\n\t\t\t\t\t\t{ status: response.status, headers: headersToRecord(response.headers) },\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t);\n\n\t\t\t\t\tif (response.ok) {\n\t\t\t\t\t\tbreak;\n\t\t\t\t\t}\n\n\t\t\t\t\tconst errorText = await response.text();\n\t\t\t\t\tif (attempt < maxRetries && isRetryableError(response.status, errorText)) {\n\t\t\t\t\t\tconst retryAfterDelayMs = getRetryAfterDelayMs(response.headers);\n\t\t\t\t\t\tconst delayMs =\n\t\t\t\t\t\t\tretryAfterDelayMs === undefined\n\t\t\t\t\t\t\t\t? BASE_DELAY_MS * 2 ** attempt\n\t\t\t\t\t\t\t\t: validateRetryDelayMs(retryAfterDelayMs, options);\n\n\t\t\t\t\t\tawait sleep(delayMs, options?.signal);\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Parse error for friendly message on final attempt or non-retryable error\n\t\t\t\t\tconst fakeResponse = new Response(errorText, {\n\t\t\t\t\t\tstatus: response.status,\n\t\t\t\t\t\tstatusText: response.statusText,\n\t\t\t\t\t});\n\t\t\t\t\tconst info = await parseErrorResponse(fakeResponse);\n\t\t\t\t\tthrow new Error(info.friendlyMessage || info.message);\n\t\t\t\t} catch (error) {\n\t\t\t\t\tif (error instanceof Error) {\n\t\t\t\t\t\tif (error.name === \"AbortError\" || error.message === \"Request was aborted\") {\n\t\t\t\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\tlastError = error instanceof Error ? error : new Error(String(error));\n\t\t\t\t\t// Network errors are retryable\n\t\t\t\t\tif (\n\t\t\t\t\t\tattempt < maxRetries &&\n\t\t\t\t\t\t!(lastError instanceof RetryDelayExceededError) &&\n\t\t\t\t\t\t!lastError.message.includes(\"usage limit\")\n\t\t\t\t\t) {\n\t\t\t\t\t\tconst delayMs = BASE_DELAY_MS * 2 ** attempt;\n\t\t\t\t\t\tawait sleep(delayMs, options?.signal);\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\t\t\t\t\tthrow lastError;\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (!response?.ok) {\n\t\t\t\tthrow lastError ?? new Error(\"Failed after retries\");\n\t\t\t}\n\n\t\t\tif (!response.body) {\n\t\t\t\tthrow new Error(\"No response body\");\n\t\t\t}\n\n\t\t\tif (!startEmitted) {\n\t\t\t\tstartEmitted = true;\n\t\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\t}\n\t\t\tawait processStream(response, output, stream, model, grammarToolInputProperties, options, streamDeadline);\n\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tassertSuccessfulOutput(output);\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\t// Streaming scratch buffers are only used during parsing; never persist them.\n\t\t\t\tdelete (block as { partialJson?: string }).partialJson;\n\t\t\t\tdelete (block as { customInput?: unknown }).customInput;\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = formatProviderError(normalizeProviderError(error));\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t} finally {\n\t\t\tstreamDeadline.cleanup();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimple: StreamFunction<\"openai-codex-responses\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-codex-responses\">,\n\tcontext: TranscriptContext,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey;\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = {\n\t\t...buildBaseOptions(model, context, options, apiKey),\n\t\ttoolChoice: options?.toolChoice,\n\t} satisfies OpenAICodexResponsesOptions;\n\tconst clampedReasoning = options?.reasoning ? clampThinkingLevel(model, options.reasoning) : undefined;\n\tconst reasoningEffort = clampedReasoning === \"off\" ? undefined : clampedReasoning;\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t} satisfies OpenAICodexResponsesOptions);\n};\n\n// ============================================================================\n// Request Building\n// ============================================================================\n\nfunction buildRequestBody(\n\tmodel: Model<\"openai-codex-responses\">,\n\tcontext: TranscriptContext,\n\toptions: OpenAICodexResponsesOptions | undefined,\n\tcacheSessionId: string | undefined,\n\tgrammarToolInputProperties: ReadonlyMap<string, string> = createGrammarToolInputProperties(\n\t\tgetDeclaredTools(context.messages),\n\t\tmodel.compat?.supportsOpenAIGrammarTools ?? false,\n\t),\n): RequestBody {\n\tconst supportsStrictMode = model.compat?.supportsStrictMode ?? true;\n\tconst supportsOpenAIGrammarTools = model.compat?.supportsOpenAIGrammarTools ?? false;\n\tconst supportsAdditionalTools = model.compat?.supportsAdditionalTools ?? false;\n\tconst supportsToolSearch = model.compat?.supportsToolSearch ?? false;\n\tconst transcriptTools = resolveTranscriptTools(context.messages, supportsAdditionalTools || supportsToolSearch);\n\tconst messages = convertResponsesMessages(model, context, CODEX_TOOL_CALL_PROVIDERS, {\n\t\tincludeSystemPrompt: false,\n\t\tgrammarToolInputProperties,\n\t\tsupportsMidConvoSystemMessages: model.compat?.supportsMidConvoSystemMessages ?? false,\n\t\tsupportsAdditionalTools,\n\t\tsupportsToolSearch,\n\t\ttoolOptions: { strict: null, supportsStrictMode, supportsOpenAIGrammarTools },\n\t});\n\n\tconst initialSystemMessage = getInitialSystemMessage(context.messages);\n\tconst instructions = initialSystemMessage ? getSystemMessageText(initialSystemMessage) : \"\";\n\tconst body: RequestBody = {\n\t\t// A fast variant keeps its canonical `-fast` id on the model object — that is the identity the\n\t\t// caller selected and records — while routing to the base upstream model plus a service tier.\n\t\tmodel: model.fastRoute?.upstreamModelId ?? model.id,\n\t\tstore: false,\n\t\tstream: true,\n\t\tinstructions: instructions || \"You are a helpful assistant.\",\n\t\tinput: messages,\n\t\ttext: { verbosity: options?.textVerbosity || \"low\" },\n\t\tinclude: [\"reasoning.encrypted_content\"],\n\t\tprompt_cache_key: cacheSessionId,\n\t\ttool_choice: options?.toolChoice ?? \"auto\",\n\t\tparallel_tool_calls: true,\n\t};\n\n\tif (options?.temperature !== undefined) {\n\t\tbody.temperature = options.temperature;\n\t}\n\n\t// A fast variant carries its own tier, so a caller that only hands over the model still routes fast.\n\tconst requestedServiceTier = resolveRequestedServiceTier(model, options?.serviceTier);\n\tif (requestedServiceTier !== undefined) {\n\t\tbody.service_tier = requestedServiceTier;\n\t}\n\n\tif (transcriptTools.requestTools.length > 0) {\n\t\tbody.tools = convertResponsesTools(transcriptTools.requestTools, {\n\t\t\tstrict: null,\n\t\t\tsupportsStrictMode,\n\t\t\tsupportsOpenAIGrammarTools,\n\t\t});\n\t}\n\n\tif (options?.reasoningEffort !== undefined) {\n\t\tconst effort =\n\t\t\toptions.reasoningEffort === \"none\"\n\t\t\t\t? model.thinkingLevelMap?.off === undefined\n\t\t\t\t\t? \"none\"\n\t\t\t\t\t: model.thinkingLevelMap.off\n\t\t\t\t: (model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort);\n\t\tif (effort !== null) {\n\t\t\tbody.reasoning = {\n\t\t\t\teffort,\n\t\t\t\tsummary: options.reasoningSummary ?? \"auto\",\n\t\t\t};\n\t\t}\n\t} else if (model.reasoning && model.thinkingLevelMap?.off !== null) {\n\t\tbody.reasoning = { effort: model.thinkingLevelMap?.off ?? \"none\" };\n\t}\n\n\treturn body;\n}\n\nfunction getServiceTierCostMultiplier(\n\tmodel: Pick<Model<\"openai-codex-responses\">, \"fastRoute\" | \"id\">,\n\tserviceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n): number {\n\t// Price against the model that was actually billed upstream, so a `-fast` variant of a\n\t// per-model rate (gpt-5.5) is not silently charged the generic multiplier.\n\tconst pricedModelId = model.fastRoute?.baseModelId ?? model.id;\n\tswitch (serviceTier) {\n\t\tcase \"flex\":\n\t\t\treturn 0.5;\n\t\tcase \"priority\":\n\t\t\treturn pricedModelId === \"gpt-5.5\" ? 2.5 : 2;\n\t\tdefault:\n\t\t\treturn 1;\n\t}\n}\n\nfunction applyServiceTierPricing(\n\tusage: Usage,\n\tserviceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\tmodel: Pick<Model<\"openai-codex-responses\">, \"fastRoute\" | \"id\">,\n) {\n\tconst multiplier = getServiceTierCostMultiplier(model, serviceTier);\n\tif (multiplier === 1) return;\n\n\tusage.cost.input *= multiplier;\n\tusage.cost.output *= multiplier;\n\tusage.cost.cacheRead *= multiplier;\n\tusage.cost.cacheWrite *= multiplier;\n\tusage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;\n}\n\nfunction resolveCodexServiceTier(\n\tresponseServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n\trequestServiceTier: ResponseCreateParamsStreaming[\"service_tier\"] | undefined,\n): ResponseCreateParamsStreaming[\"service_tier\"] | undefined {\n\tif (responseServiceTier === \"default\" && (requestServiceTier === \"flex\" || requestServiceTier === \"priority\")) {\n\t\treturn requestServiceTier;\n\t}\n\treturn responseServiceTier ?? requestServiceTier;\n}\n\nfunction resolveCodexUrl(baseUrl?: string): string {\n\tconst raw = baseUrl && baseUrl.trim().length > 0 ? baseUrl : DEFAULT_CODEX_BASE_URL;\n\tconst normalized = raw.replace(/\\/+$/, \"\");\n\tif (normalized.endsWith(\"/codex/responses\")) return normalized;\n\tif (normalized.endsWith(\"/codex\")) return `${normalized}/responses`;\n\treturn `${normalized}/codex/responses`;\n}\n\nfunction resolveCodexWebSocketUrl(baseUrl?: string): string {\n\tconst url = new URL(resolveCodexUrl(baseUrl));\n\tif (url.protocol === \"https:\") url.protocol = \"wss:\";\n\tif (url.protocol === \"http:\") url.protocol = \"ws:\";\n\treturn url.toString();\n}\n\n// ============================================================================\n// Response Processing\n// ============================================================================\n\nasync function processStream(\n\tresponse: Response,\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n\tmodel: Model<\"openai-codex-responses\">,\n\tgrammarToolInputProperties: ReadonlyMap<string, string>,\n\toptions: OpenAICodexResponsesOptions | undefined,\n\tstreamDeadline: StreamDeadlineHandle,\n): Promise<void> {\n\tconst events = mapCodexEvents(\n\t\tparseSSE(response, streamDeadline.signal),\n\t\toutput,\n\t\tmodel,\n\t\toptions?.onProviderStreamEvent,\n\t);\n\tawait processResponsesStream(\n\t\twithStreamDeadline(events, streamDeadline.deadlineMs, streamDeadline.abort),\n\t\toutput,\n\t\tstream,\n\t\tmodel,\n\t\t{\n\t\t\tserviceTier: resolveRequestedServiceTier(model, options?.serviceTier),\n\t\t\tgrammarToolInputProperties,\n\t\t\tresolveServiceTier: resolveCodexServiceTier,\n\t\t\tapplyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),\n\t\t},\n\t);\n}\n\nclass CodexApiError extends Error {\n\treadonly code?: string;\n\treadonly payload?: Record<string, unknown>;\n\n\tconstructor(message: string, options?: { code?: string; payload?: Record<string, unknown>; cause?: unknown }) {\n\t\tsuper(message);\n\t\tthis.name = \"CodexApiError\";\n\t\tthis.code = options?.code;\n\t\tthis.payload = options?.payload;\n\t\tthis.cause = options?.cause;\n\t}\n}\n\nclass CodexProtocolError extends Error {\n\treadonly payload?: unknown;\n\n\tconstructor(message: string, options?: { payload?: unknown; cause?: unknown }) {\n\t\tsuper(message);\n\t\tthis.name = \"CodexProtocolError\";\n\t\tthis.payload = options?.payload;\n\t\tthis.cause = options?.cause;\n\t}\n}\n\nclass ProviderStreamEventCallbackError extends Error {\n\tconstructor(cause: unknown) {\n\t\tsuper(formatThrownValue(cause));\n\t\tthis.name = \"ProviderStreamEventCallbackError\";\n\t\tthis.cause = cause;\n\t}\n}\n\nfunction isCodexNonTransportError(error: unknown): boolean {\n\treturn (\n\t\terror instanceof CodexApiError ||\n\t\terror instanceof CodexProtocolError ||\n\t\terror instanceof ProviderStreamEventCallbackError\n\t);\n}\n\nfunction isWebSocketConnectionLimitReachedError(error: unknown): boolean {\n\treturn error instanceof CodexApiError && error.code === WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE;\n}\n\nfunction isPreviousResponseNotFoundError(error: unknown): boolean {\n\treturn error instanceof CodexApiError && error.code === PREVIOUS_RESPONSE_NOT_FOUND_CODE;\n}\n\nfunction extractCodexEventError(event: Record<string, unknown>): { code?: string; message?: string } {\n\tconst nested = event.error && typeof event.error === \"object\" ? (event.error as Record<string, unknown>) : undefined;\n\treturn {\n\t\tcode: typeof event.code === \"string\" ? event.code : typeof nested?.code === \"string\" ? nested.code : undefined,\n\t\tmessage:\n\t\t\ttypeof event.message === \"string\"\n\t\t\t\t? event.message\n\t\t\t\t: typeof nested?.message === \"string\"\n\t\t\t\t\t? nested.message\n\t\t\t\t\t: undefined,\n\t};\n}\n\nasync function* mapCodexEvents(\n\tevents: AsyncIterable<Record<string, unknown>>,\n\toutput: AssistantMessage,\n\tmodel: Model<\"openai-codex-responses\">,\n\tonProviderStreamEvent: StreamOptions[\"onProviderStreamEvent\"] | undefined,\n): AsyncGenerator<ResponseStreamEvent> {\n\tfor await (const event of events) {\n\t\ttry {\n\t\t\tawait onProviderStreamEvent?.(event, model);\n\t\t} catch (error) {\n\t\t\t// Keep callback failures out of Codex's WebSocket retry and SSE fallback path.\n\t\t\tthrow new ProviderStreamEventCallbackError(error);\n\t\t}\n\t\tconst type = typeof event.type === \"string\" ? event.type : undefined;\n\t\tif (!type) continue;\n\n\t\tif (type === \"error\") {\n\t\t\tconst { code, message } = extractCodexEventError(event);\n\t\t\tthrow new CodexApiError(`Codex error: ${message || code || JSON.stringify(event)}`, {\n\t\t\t\tcode,\n\t\t\t\tpayload: event,\n\t\t\t});\n\t\t}\n\n\t\tif (type === \"response.failed\") {\n\t\t\tconst response = (event as { response?: { error?: { code?: string; message?: string } } }).response;\n\t\t\tconst code = response?.error?.code;\n\t\t\tconst message = response?.error?.message;\n\t\t\tthrow new CodexApiError(message || \"Codex response failed\", { code, payload: event });\n\t\t}\n\n\t\tif (type === \"response.done\" || type === \"response.completed\" || type === \"response.incomplete\") {\n\t\t\tconst response = (event as { response?: { status?: unknown; end_turn?: unknown } }).response;\n\t\t\tif (typeof response?.end_turn === \"boolean\") {\n\t\t\t\toutput.endTurn = response.end_turn;\n\t\t\t}\n\t\t\tconst normalizedResponse = response\n\t\t\t\t? { ...response, status: normalizeCodexStatus(response.status) }\n\t\t\t\t: response;\n\t\t\tyield { ...event, type: \"response.completed\", response: normalizedResponse } as ResponseStreamEvent;\n\t\t\treturn;\n\t\t}\n\n\t\tyield event as unknown as ResponseStreamEvent;\n\t}\n}\n\nfunction normalizeCodexStatus(status: unknown): CodexResponseStatus | undefined {\n\tif (typeof status !== \"string\") return undefined;\n\treturn CODEX_RESPONSE_STATUSES.has(status as CodexResponseStatus) ? (status as CodexResponseStatus) : undefined;\n}\n\n// ============================================================================\n// SSE Parsing\n// ============================================================================\n\nasync function* parseSSE(response: Response, signal?: AbortSignal): AsyncGenerator<Record<string, unknown>> {\n\tif (!response.body) return;\n\n\tconst reader = response.body.getReader();\n\tconst decoder = new TextDecoder();\n\tlet buffer = \"\";\n\tconst onAbort = () => {\n\t\tvoid reader.cancel().catch(() => {});\n\t};\n\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\n\ttry {\n\t\twhile (true) {\n\t\t\tif (signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tconst { done, value } = await reader.read();\n\t\t\tif (signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tbuffer += done ? decoder.decode() : decoder.decode(value, { stream: true });\n\t\t\t// Treat EOF as terminating the residual SSE frame.\n\t\t\tif (done && buffer.trim()) buffer += \"\\n\\n\";\n\n\t\t\tlet idx = buffer.indexOf(\"\\n\\n\");\n\t\t\twhile (idx !== -1) {\n\t\t\t\tconst chunk = buffer.slice(0, idx);\n\t\t\t\tbuffer = buffer.slice(idx + 2);\n\n\t\t\t\tconst dataLines = chunk\n\t\t\t\t\t.split(\"\\n\")\n\t\t\t\t\t.filter((l) => l.startsWith(\"data:\"))\n\t\t\t\t\t.map((l) => l.slice(5).trim());\n\t\t\t\tif (dataLines.length > 0) {\n\t\t\t\t\tconst data = dataLines.join(\"\\n\").trim();\n\t\t\t\t\tif (data && data !== \"[DONE]\") {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\tyield JSON.parse(data) as Record<string, unknown>;\n\t\t\t\t\t\t} catch (cause) {\n\t\t\t\t\t\t\tthrow new CodexProtocolError(`Invalid Codex SSE JSON: ${formatThrownValue(cause)}`, {\n\t\t\t\t\t\t\t\tcause,\n\t\t\t\t\t\t\t\tpayload: data,\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}\n\t\t\t\tidx = buffer.indexOf(\"\\n\\n\");\n\t\t\t}\n\n\t\t\tif (done) break;\n\t\t}\n\t} finally {\n\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\ttry {\n\t\t\tawait reader.cancel();\n\t\t} catch {}\n\t\ttry {\n\t\t\treader.releaseLock();\n\t\t} catch {}\n\t}\n}\n\n// ============================================================================\n// WebSocket Parsing\n// ============================================================================\n\nconst OPENAI_BETA_RESPONSES_WEBSOCKETS = \"responses_websockets=2026-02-06\";\nconst SESSION_WEBSOCKET_CACHE_TTL_MS = 5 * 60 * 1000;\nconst SESSION_WEBSOCKET_MAX_AGE_MS = 55 * 60 * 1000;\n\ntype WebSocketEventType = \"open\" | \"message\" | \"error\" | \"close\";\ntype WebSocketListener = (event: unknown) => void;\n\ninterface WebSocketLike {\n\tclose(code?: number, reason?: string): void;\n\tsend(data: string): void;\n\taddEventListener(type: WebSocketEventType, listener: WebSocketListener): void;\n\tremoveEventListener(type: WebSocketEventType, listener: WebSocketListener): void;\n}\n\ninterface CachedWebSocketContinuationState {\n\tlastRequestBody: RequestBody;\n\tlastResponseId: string;\n\tlastResponseItems: ResponseInput;\n}\n\ninterface CachedWebSocketConnection {\n\tsocket: WebSocketLike;\n\tbusy: boolean;\n\tcreatedAt: number;\n\tidleTimer?: ReturnType<typeof setTimeout>;\n\tcontinuation?: CachedWebSocketContinuationState;\n}\n\nexport interface OpenAICodexWebSocketDebugStats {\n\trequests: number;\n\tconnectionsCreated: number;\n\tconnectionsReused: number;\n\tcachedContextRequests: number;\n\tstoreTrueRequests: number;\n\tfullContextRequests: number;\n\tdeltaRequests: number;\n\tlastInputItems: number;\n\tlastDeltaInputItems?: number;\n\tlastPreviousResponseId?: string;\n\twebsocketFailures: number;\n\tsseFallbacks: number;\n\twebsocketFallbackActive?: boolean;\n\tlastWebSocketError?: string;\n}\n\nconst websocketSessionCache = new Map<string, Map<string, CachedWebSocketConnection>>();\nconst websocketDebugStats = new Map<string, OpenAICodexWebSocketDebugStats>();\nconst websocketSseFallbackSessions = new Set<string>();\n\nfunction getOrCreateWebSocketDebugStats(sessionId: string): OpenAICodexWebSocketDebugStats {\n\tlet stats = websocketDebugStats.get(sessionId);\n\tif (!stats) {\n\t\tstats = {\n\t\t\trequests: 0,\n\t\t\tconnectionsCreated: 0,\n\t\t\tconnectionsReused: 0,\n\t\t\tcachedContextRequests: 0,\n\t\t\tstoreTrueRequests: 0,\n\t\t\tfullContextRequests: 0,\n\t\t\tdeltaRequests: 0,\n\t\t\tlastInputItems: 0,\n\t\t\twebsocketFailures: 0,\n\t\t\tsseFallbacks: 0,\n\t\t};\n\t\twebsocketDebugStats.set(sessionId, stats);\n\t}\n\treturn stats;\n}\n\nexport function getOpenAICodexWebSocketDebugStats(sessionId: string): OpenAICodexWebSocketDebugStats | undefined {\n\tconst stats = websocketDebugStats.get(sessionId);\n\treturn stats ? { ...stats } : undefined;\n}\n\nexport function resetOpenAICodexWebSocketDebugStats(sessionId?: string): void {\n\tif (sessionId) {\n\t\twebsocketDebugStats.delete(sessionId);\n\t\twebsocketSseFallbackSessions.delete(sessionId);\n\t\treturn;\n\t}\n\twebsocketDebugStats.clear();\n\twebsocketSseFallbackSessions.clear();\n}\n\nexport function closeOpenAICodexWebSocketSessions(sessionId?: string): void {\n\tconst closeEntry = (entry: CachedWebSocketConnection) => {\n\t\tif (entry.idleTimer) clearTimeout(entry.idleTimer);\n\t\tcloseWebSocketSilently(entry.socket, 1000, \"debug_close\");\n\t};\n\tif (sessionId) {\n\t\tfor (const entry of websocketSessionCache.get(sessionId)?.values() ?? []) closeEntry(entry);\n\t\twebsocketSessionCache.delete(sessionId);\n\t\treturn;\n\t}\n\tfor (const accountEntries of websocketSessionCache.values()) {\n\t\tfor (const entry of accountEntries.values()) closeEntry(entry);\n\t}\n\twebsocketSessionCache.clear();\n}\n\nregisterSessionResourceCleanup(closeOpenAICodexWebSocketSessions);\n\nfunction isWebSocketSseFallbackActive(sessionId: string | undefined): boolean {\n\treturn sessionId ? websocketSseFallbackSessions.has(sessionId) : false;\n}\n\nfunction recordWebSocketSseFallback(sessionId: string | undefined): void {\n\tif (!sessionId) return;\n\tconst stats = getOrCreateWebSocketDebugStats(sessionId);\n\tstats.sseFallbacks++;\n\tstats.websocketFallbackActive = isWebSocketSseFallbackActive(sessionId);\n}\n\nfunction recordWebSocketFailure(sessionId: string | undefined, error: unknown): void {\n\tif (!sessionId) return;\n\twebsocketSseFallbackSessions.add(sessionId);\n\n\tconst stats = getOrCreateWebSocketDebugStats(sessionId);\n\tstats.websocketFailures++;\n\tstats.lastWebSocketError = formatThrownValue(error);\n\tstats.websocketFallbackActive = true;\n}\n\ntype WebSocketConstructor = new (\n\turl: string,\n\tprotocols?: string | string[] | { headers?: Record<string, string> },\n) => WebSocketLike;\n\nlet _cachedWebsocket: WebSocketConstructor | null = null;\nasync function getWebSocketConstructor(env?: ProviderEnv): Promise<WebSocketConstructor | null> {\n\tif (!env && _cachedWebsocket) return _cachedWebsocket;\n\n\t// bun doesn't respect http proxy envs, ref: https://github.com/oven-sh/bun/issues/15489\n\t// TODO: remove this when bun supports proxy envs in websocket.\n\tif (typeof process !== \"undefined\" && process.versions?.bun) {\n\t\tconst WebSocketWithProxy = class extends WebSocket {\n\t\t\tconstructor(url: string | URL, options?: string | string[] | Record<string, unknown>) {\n\t\t\t\tlet _opts: Record<string, unknown> = {};\n\t\t\t\tif (Array.isArray(options) || typeof options === \"string\") {\n\t\t\t\t\t_opts = { protocols: options };\n\t\t\t\t} else {\n\t\t\t\t\t_opts = { ...options };\n\t\t\t\t}\n\n\t\t\t\tconst proxyUrl = resolveHttpProxyUrlForTarget(\n\t\t\t\t\turl.toString().replace(/^wss:/, \"https:\").replace(/^ws:/, \"http:\"),\n\t\t\t\t\tenv,\n\t\t\t\t);\n\t\t\t\tsuper(url, { ..._opts, ...(proxyUrl ? { proxy: proxyUrl.toString() } : {}) } as any);\n\t\t\t}\n\t\t};\n\t\tif (!env) {\n\t\t\t_cachedWebsocket = WebSocketWithProxy;\n\t\t}\n\t\treturn WebSocketWithProxy;\n\t}\n\n\tconst ctor = (globalThis as { WebSocket?: unknown }).WebSocket;\n\tif (typeof ctor !== \"function\") return null;\n\treturn ctor as unknown as WebSocketConstructor;\n}\n\nclass WebSocketCloseError extends Error {\n\treadonly code?: number;\n\treadonly reason?: string;\n\treadonly wasClean?: boolean;\n\n\tconstructor(message: string, options?: { code?: number; reason?: string; wasClean?: boolean }) {\n\t\tsuper(message);\n\t\tthis.name = \"WebSocketCloseError\";\n\t\tthis.code = options?.code;\n\t\tthis.reason = options?.reason;\n\t\tthis.wasClean = options?.wasClean;\n\t}\n}\n\nfunction getWebSocketReadyState(socket: WebSocketLike): number | undefined {\n\tconst readyState = (socket as { readyState?: unknown }).readyState;\n\treturn typeof readyState === \"number\" ? readyState : undefined;\n}\n\nfunction isWebSocketReusable(socket: WebSocketLike): boolean {\n\tconst readyState = getWebSocketReadyState(socket);\n\t// If readyState is unavailable, assume the runtime keeps it open/reusable.\n\treturn readyState === undefined || readyState === 1;\n}\n\nfunction isWebSocketSessionExpired(entry: CachedWebSocketConnection): boolean {\n\treturn Date.now() - entry.createdAt >= SESSION_WEBSOCKET_MAX_AGE_MS;\n}\n\nfunction closeWebSocketSilently(socket: WebSocketLike, code = 1000, reason = \"done\"): void {\n\ttry {\n\t\tsocket.close(code, reason);\n\t} catch {}\n}\n\nfunction scheduleSessionWebSocketExpiry(sessionId: string, accountId: string, entry: CachedWebSocketConnection): void {\n\tif (entry.idleTimer) {\n\t\tclearTimeout(entry.idleTimer);\n\t}\n\tentry.idleTimer = setTimeout(() => {\n\t\tif (entry.busy) return;\n\t\tcloseWebSocketSilently(entry.socket, 1000, \"idle_timeout\");\n\t\tconst accountEntries = websocketSessionCache.get(sessionId);\n\t\tif (accountEntries?.get(accountId) === entry) accountEntries.delete(accountId);\n\t\tif (accountEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t}, SESSION_WEBSOCKET_CACHE_TTL_MS);\n}\n\nasync function connectWebSocket(\n\turl: string,\n\theaders: Headers,\n\tsignal?: AbortSignal,\n\tconnectTimeoutMs = DEFAULT_WEBSOCKET_CONNECT_TIMEOUT_MS,\n\tenv?: ProviderEnv,\n): Promise<WebSocketLike> {\n\tconst WebSocketCtor = await getWebSocketConstructor(env);\n\tif (!WebSocketCtor) {\n\t\tthrow new Error(\"WebSocket transport is not available in this runtime\");\n\t}\n\n\tconst wsHeaders = headersToRecord(headers);\n\tdelete wsHeaders[\"OpenAI-Beta\"];\n\n\treturn new Promise<WebSocketLike>((resolve, reject) => {\n\t\tlet settled = false;\n\t\tlet timeout: ReturnType<typeof setTimeout> | undefined;\n\t\tlet socket: WebSocketLike;\n\n\t\ttry {\n\t\t\tsocket = new WebSocketCtor(url, { headers: wsHeaders });\n\t\t} catch (error) {\n\t\t\treject(error instanceof Error ? error : new Error(String(error)));\n\t\t\treturn;\n\t\t}\n\n\t\tconst cleanup = () => {\n\t\t\tif (timeout) {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\ttimeout = undefined;\n\t\t\t}\n\t\t\tsocket.removeEventListener(\"open\", onOpen);\n\t\t\tsocket.removeEventListener(\"error\", onError);\n\t\t\tsocket.removeEventListener(\"close\", onClose);\n\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t};\n\t\tconst fail = (error: Error, closeReason?: string) => {\n\t\t\tif (settled) return;\n\t\t\tsettled = true;\n\t\t\tcleanup();\n\t\t\tif (closeReason) {\n\t\t\t\tcloseWebSocketSilently(socket, 1000, closeReason);\n\t\t\t}\n\t\t\treject(error);\n\t\t};\n\t\tconst onOpen: WebSocketListener = () => {\n\t\t\tif (settled) return;\n\t\t\tsettled = true;\n\t\t\tcleanup();\n\t\t\tresolve(socket);\n\t\t};\n\t\tconst onError: WebSocketListener = (event) => {\n\t\t\tfail(extractWebSocketError(event));\n\t\t};\n\t\tconst onClose: WebSocketListener = (event) => {\n\t\t\tfail(extractWebSocketCloseError(event));\n\t\t};\n\t\tconst onAbort = () => {\n\t\t\tfail(new Error(\"Request was aborted\"), \"aborted\");\n\t\t};\n\n\t\tsocket.addEventListener(\"open\", onOpen);\n\t\tsocket.addEventListener(\"error\", onError);\n\t\tsocket.addEventListener(\"close\", onClose);\n\t\tsignal?.addEventListener(\"abort\", onAbort);\n\n\t\tif (connectTimeoutMs > 0) {\n\t\t\ttimeout = setTimeout(() => {\n\t\t\t\tfail(new Error(`WebSocket connect timeout after ${connectTimeoutMs}ms`), \"connect_timeout\");\n\t\t\t}, connectTimeoutMs);\n\t\t}\n\t\tif (signal?.aborted) {\n\t\t\tonAbort();\n\t\t}\n\t});\n}\n\nasync function acquireWebSocket(\n\turl: string,\n\theaders: Headers,\n\tsessionId: string | undefined,\n\taccountId: string,\n\tsignal?: AbortSignal,\n\tconnectTimeoutMs?: number,\n\tenv?: ProviderEnv,\n): Promise<{\n\tsocket: WebSocketLike;\n\tentry?: CachedWebSocketConnection;\n\treused: boolean;\n\trelease: (options?: { keep?: boolean }) => void;\n}> {\n\tif (!sessionId) {\n\t\tconst socket = await connectWebSocket(url, headers, signal, connectTimeoutMs, env);\n\t\treturn {\n\t\t\tsocket,\n\t\t\treused: false,\n\t\t\trelease: () => closeWebSocketSilently(socket),\n\t\t};\n\t}\n\n\tlet accountEntries = websocketSessionCache.get(sessionId);\n\tconst cached = accountEntries?.get(accountId);\n\tif (cached) {\n\t\tif (cached.idleTimer) {\n\t\t\tclearTimeout(cached.idleTimer);\n\t\t\tcached.idleTimer = undefined;\n\t\t}\n\t\tif (!cached.busy && isWebSocketSessionExpired(cached)) {\n\t\t\tcloseWebSocketSilently(cached.socket, 1000, \"connection_age_limit\");\n\t\t\taccountEntries?.delete(accountId);\n\t\t\tif (accountEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t} else if (!cached.busy && isWebSocketReusable(cached.socket)) {\n\t\t\tcached.busy = true;\n\t\t\treturn {\n\t\t\t\tsocket: cached.socket,\n\t\t\t\tentry: cached,\n\t\t\t\treused: true,\n\t\t\t\trelease: ({ keep } = {}) => {\n\t\t\t\t\tif (!keep || !isWebSocketReusable(cached.socket)) {\n\t\t\t\t\t\tcloseWebSocketSilently(cached.socket);\n\t\t\t\t\t\tconst currentEntries = websocketSessionCache.get(sessionId);\n\t\t\t\t\t\tif (currentEntries?.get(accountId) === cached) currentEntries.delete(accountId);\n\t\t\t\t\t\tif (currentEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\tcached.busy = false;\n\t\t\t\t\tscheduleSessionWebSocketExpiry(sessionId, accountId, cached);\n\t\t\t\t},\n\t\t\t};\n\t\t}\n\t\tif (cached.busy) {\n\t\t\tconst socket = await connectWebSocket(url, headers, signal, connectTimeoutMs, env);\n\t\t\treturn {\n\t\t\t\tsocket,\n\t\t\t\treused: false,\n\t\t\t\trelease: () => {\n\t\t\t\t\tcloseWebSocketSilently(socket);\n\t\t\t\t},\n\t\t\t};\n\t\t}\n\t\tif (!isWebSocketReusable(cached.socket)) {\n\t\t\tcloseWebSocketSilently(cached.socket);\n\t\t\taccountEntries?.delete(accountId);\n\t\t\tif (accountEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t}\n\t}\n\n\tconst socket = await connectWebSocket(url, headers, signal, connectTimeoutMs, env);\n\tconst entry: CachedWebSocketConnection = { socket, busy: true, createdAt: Date.now() };\n\taccountEntries = websocketSessionCache.get(sessionId);\n\tif (!accountEntries) {\n\t\taccountEntries = new Map();\n\t\twebsocketSessionCache.set(sessionId, accountEntries);\n\t}\n\taccountEntries.set(accountId, entry);\n\treturn {\n\t\tsocket,\n\t\tentry,\n\t\treused: false,\n\t\trelease: ({ keep } = {}) => {\n\t\t\tif (!keep || !isWebSocketReusable(entry.socket)) {\n\t\t\t\tcloseWebSocketSilently(entry.socket);\n\t\t\t\tif (entry.idleTimer) clearTimeout(entry.idleTimer);\n\t\t\t\tconst currentEntries = websocketSessionCache.get(sessionId);\n\t\t\t\tif (currentEntries?.get(accountId) === entry) currentEntries.delete(accountId);\n\t\t\t\tif (currentEntries?.size === 0) websocketSessionCache.delete(sessionId);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tentry.busy = false;\n\t\t\tscheduleSessionWebSocketExpiry(sessionId, accountId, entry);\n\t\t},\n\t};\n}\n\nfunction extractWebSocketError(event: unknown): Error {\n\tif (event && typeof event === \"object\") {\n\t\tconst message = \"message\" in event ? (event as { message?: unknown }).message : undefined;\n\t\tif (typeof message === \"string\" && message.length > 0) {\n\t\t\treturn new Error(message);\n\t\t}\n\n\t\tconst nestedError = \"error\" in event ? (event as { error?: unknown }).error : undefined;\n\t\tif (nestedError instanceof Error && nestedError.message.length > 0) {\n\t\t\treturn nestedError;\n\t\t}\n\t\tif (nestedError && typeof nestedError === \"object\" && \"message\" in nestedError) {\n\t\t\tconst nestedMessage = (nestedError as { message?: unknown }).message;\n\t\t\tif (typeof nestedMessage === \"string\" && nestedMessage.length > 0) {\n\t\t\t\treturn new Error(nestedMessage);\n\t\t\t}\n\t\t}\n\t}\n\treturn new Error(\"WebSocket error\");\n}\n\nfunction extractWebSocketCloseError(event: unknown): Error {\n\tif (event && typeof event === \"object\") {\n\t\tconst code = \"code\" in event ? (event as { code?: unknown }).code : undefined;\n\t\tconst reason = \"reason\" in event ? (event as { reason?: unknown }).reason : undefined;\n\t\tconst wasClean = \"wasClean\" in event ? (event as { wasClean?: unknown }).wasClean : undefined;\n\t\tconst codeText = typeof code === \"number\" ? ` ${code}` : \"\";\n\t\tlet reasonText = typeof reason === \"string\" && reason.length > 0 ? ` ${reason}` : \"\";\n\t\tif (!reasonText && code === WEBSOCKET_MESSAGE_TOO_BIG_CLOSE_CODE) {\n\t\t\treasonText = \" message too big\";\n\t\t}\n\t\treturn new WebSocketCloseError(`WebSocket closed${codeText}${reasonText}`.trim(), {\n\t\t\tcode: typeof code === \"number\" ? code : undefined,\n\t\t\treason: typeof reason === \"string\" && reason.length > 0 ? reason : undefined,\n\t\t\twasClean: typeof wasClean === \"boolean\" ? wasClean : undefined,\n\t\t});\n\t}\n\treturn new Error(\"WebSocket closed\");\n}\n\nasync function decodeWebSocketData(data: unknown): Promise<string | null> {\n\tif (typeof data === \"string\") return data;\n\tif (data instanceof ArrayBuffer) {\n\t\treturn new TextDecoder().decode(new Uint8Array(data));\n\t}\n\tif (ArrayBuffer.isView(data)) {\n\t\tconst view = data as ArrayBufferView;\n\t\treturn new TextDecoder().decode(new Uint8Array(view.buffer, view.byteOffset, view.byteLength));\n\t}\n\tif (data && typeof data === \"object\" && \"arrayBuffer\" in data) {\n\t\tconst blobLike = data as { arrayBuffer: () => Promise<ArrayBuffer> };\n\t\tconst arrayBuffer = await blobLike.arrayBuffer();\n\t\treturn new TextDecoder().decode(new Uint8Array(arrayBuffer));\n\t}\n\treturn null;\n}\n\nasync function* parseWebSocket(\n\tsocket: WebSocketLike,\n\tsignal?: AbortSignal,\n\tidleTimeoutMs?: number,\n): AsyncGenerator<Record<string, unknown>> {\n\tconst queue: Record<string, unknown>[] = [];\n\tlet pending: (() => void) | null = null;\n\tlet done = false;\n\tlet failed: Error | null = null;\n\tlet sawCompletion = false;\n\n\tconst wake = () => {\n\t\tif (!pending) return;\n\t\tconst resolve = pending;\n\t\tpending = null;\n\t\tresolve();\n\t};\n\n\tconst onMessage: WebSocketListener = (event) => {\n\t\tvoid (async () => {\n\t\t\tlet text: string | null = null;\n\t\t\ttry {\n\t\t\t\tif (!event || typeof event !== \"object\" || !(\"data\" in event)) return;\n\t\t\t\ttext = await decodeWebSocketData((event as { data?: unknown }).data);\n\t\t\t\tif (!text) return;\n\t\t\t\tconst parsed = JSON.parse(text) as Record<string, unknown>;\n\t\t\t\tconst type = typeof parsed.type === \"string\" ? parsed.type : \"\";\n\t\t\t\tif (type === \"response.completed\" || type === \"response.done\" || type === \"response.incomplete\") {\n\t\t\t\t\tsawCompletion = true;\n\t\t\t\t\tdone = true;\n\t\t\t\t}\n\t\t\t\tqueue.push(parsed);\n\t\t\t\twake();\n\t\t\t} catch (cause) {\n\t\t\t\tfailed = new CodexProtocolError(`Invalid Codex WebSocket JSON: ${formatThrownValue(cause)}`, {\n\t\t\t\t\tcause,\n\t\t\t\t\tpayload: text,\n\t\t\t\t});\n\t\t\t\tdone = true;\n\t\t\t\twake();\n\t\t\t}\n\t\t})();\n\t};\n\n\tconst onError: WebSocketListener = (event) => {\n\t\tfailed = extractWebSocketError(event);\n\t\tdone = true;\n\t\twake();\n\t};\n\n\tconst onClose: WebSocketListener = (event) => {\n\t\tif (sawCompletion) {\n\t\t\tdone = true;\n\t\t\twake();\n\t\t\treturn;\n\t\t}\n\t\tif (!failed) {\n\t\t\tfailed = extractWebSocketCloseError(event);\n\t\t}\n\t\tdone = true;\n\t\twake();\n\t};\n\n\tconst onAbort = () => {\n\t\tfailed = new Error(\"Request was aborted\");\n\t\tdone = true;\n\t\twake();\n\t};\n\n\tsocket.addEventListener(\"message\", onMessage);\n\tsocket.addEventListener(\"error\", onError);\n\tsocket.addEventListener(\"close\", onClose);\n\tsignal?.addEventListener(\"abort\", onAbort);\n\n\ttry {\n\t\twhile (true) {\n\t\t\tif (signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (queue.length > 0) {\n\t\t\t\tyield queue.shift()!;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (done) break;\n\t\t\tlet timeout: ReturnType<typeof setTimeout> | undefined;\n\t\t\tawait new Promise<void>((resolve, reject) => {\n\t\t\t\tpending = resolve;\n\t\t\t\tif (idleTimeoutMs !== undefined && idleTimeoutMs > 0) {\n\t\t\t\t\ttimeout = setTimeout(() => {\n\t\t\t\t\t\tconst error = new Error(`WebSocket idle timeout after ${idleTimeoutMs}ms`);\n\t\t\t\t\t\tfailed = error;\n\t\t\t\t\t\tdone = true;\n\t\t\t\t\t\tpending = null;\n\t\t\t\t\t\tcloseWebSocketSilently(socket, 1000, \"idle_timeout\");\n\t\t\t\t\t\treject(error);\n\t\t\t\t\t}, idleTimeoutMs);\n\t\t\t\t}\n\t\t\t}).finally(() => {\n\t\t\t\tif (timeout) {\n\t\t\t\t\tclearTimeout(timeout);\n\t\t\t\t}\n\t\t\t});\n\t\t}\n\n\t\tif (failed) {\n\t\t\tthrow failed;\n\t\t}\n\t\tif (!sawCompletion) {\n\t\t\tthrow new Error(\"WebSocket stream closed before response.completed\");\n\t\t}\n\t} finally {\n\t\tsocket.removeEventListener(\"message\", onMessage);\n\t\tsocket.removeEventListener(\"error\", onError);\n\t\tsocket.removeEventListener(\"close\", onClose);\n\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t}\n}\n\nfunction requestBodyWithoutInput(body: RequestBody): RequestBody {\n\tconst { input: _input, previous_response_id: _previousResponseId, ...rest } = body;\n\treturn rest;\n}\n\nfunction responseInputsEqual(a: ResponseInput | undefined, b: ResponseInput | undefined): boolean {\n\treturn JSON.stringify(a ?? []) === JSON.stringify(b ?? []);\n}\n\nfunction requestBodiesMatchExceptInput(a: RequestBody, b: RequestBody): boolean {\n\treturn JSON.stringify(requestBodyWithoutInput(a)) === JSON.stringify(requestBodyWithoutInput(b));\n}\n\nfunction getCachedWebSocketInputDelta(\n\tbody: RequestBody,\n\tcontinuation: CachedWebSocketContinuationState,\n): ResponseInput | undefined {\n\tif (!requestBodiesMatchExceptInput(body, continuation.lastRequestBody)) {\n\t\treturn undefined;\n\t}\n\n\tconst currentInput = body.input ?? [];\n\tconst baseline = [...(continuation.lastRequestBody.input ?? []), ...continuation.lastResponseItems];\n\tif (currentInput.length < baseline.length) {\n\t\treturn undefined;\n\t}\n\n\tconst prefix = currentInput.slice(0, baseline.length);\n\tif (!responseInputsEqual(prefix, baseline)) {\n\t\treturn undefined;\n\t}\n\n\treturn currentInput.slice(baseline.length);\n}\n\nfunction buildCachedWebSocketRequestBody(entry: CachedWebSocketConnection, body: RequestBody): RequestBody {\n\tconst continuation = entry.continuation;\n\tif (!continuation) {\n\t\treturn body;\n\t}\n\n\tconst delta = getCachedWebSocketInputDelta(body, continuation);\n\tif (!delta || !continuation.lastResponseId) {\n\t\tentry.continuation = undefined;\n\t\treturn body;\n\t}\n\n\treturn {\n\t\t...body,\n\t\tprevious_response_id: continuation.lastResponseId,\n\t\tinput: delta,\n\t};\n}\n\nasync function* startWebSocketOutputOnFirstEvent(\n\tevents: AsyncIterable<ResponseStreamEvent>,\n\tonStart: () => void,\n): AsyncGenerator<ResponseStreamEvent> {\n\tlet started = false;\n\tfor await (const event of events) {\n\t\tif (!started) {\n\t\t\tstarted = true;\n\t\t\tonStart();\n\t\t}\n\t\tyield event;\n\t}\n}\n\nasync function processWebSocketStream(\n\turl: string,\n\tbody: RequestBody,\n\theaders: Headers,\n\toutput: AssistantMessage,\n\tstream: AssistantMessageEventStream,\n\tmodel: Model<\"openai-codex-responses\">,\n\tonStart: () => void,\n\tidleTimeoutMs: number | undefined,\n\twebsocketConnectTimeoutMs: number | undefined,\n\tcacheSessionId: string | undefined,\n\taccountId: string,\n\tgrammarToolInputProperties: ReadonlyMap<string, string>,\n\toptions?: OpenAICodexResponsesOptions,\n): Promise<void> {\n\tconst { socket, entry, reused, release } = await acquireWebSocket(\n\t\turl,\n\t\theaders,\n\t\tcacheSessionId,\n\t\taccountId,\n\t\toptions?.signal,\n\t\twebsocketConnectTimeoutMs,\n\t\toptions?.env,\n\t);\n\tlet keepConnection = true;\n\tconst useCachedContext = options?.transport === \"websocket-cached\" || options?.transport === \"auto\";\n\t// ChatGPT Codex Responses rejects `store: true` (\"Store must be set to false\").\n\t// WebSocket continuation still works via connection-scoped previous_response_id state.\n\tconst fullBody = body;\n\tconst requestBody = useCachedContext && entry ? buildCachedWebSocketRequestBody(entry, fullBody) : fullBody;\n\tconst stats = cacheSessionId ? getOrCreateWebSocketDebugStats(cacheSessionId) : undefined;\n\tif (stats) {\n\t\tstats.requests++;\n\t\tif (reused) stats.connectionsReused++;\n\t\telse stats.connectionsCreated++;\n\t\tif (useCachedContext) stats.cachedContextRequests++;\n\t\tif (requestBody.store === true) stats.storeTrueRequests++;\n\t\tstats.lastInputItems = requestBody.input?.length ?? 0;\n\t\tif (requestBody.previous_response_id) {\n\t\t\tstats.deltaRequests++;\n\t\t\tstats.lastDeltaInputItems = requestBody.input?.length ?? 0;\n\t\t\tstats.lastPreviousResponseId = requestBody.previous_response_id;\n\t\t} else {\n\t\t\tstats.fullContextRequests++;\n\t\t\tstats.lastDeltaInputItems = undefined;\n\t\t\tstats.lastPreviousResponseId = undefined;\n\t\t}\n\t}\n\ttry {\n\t\tsocket.send(JSON.stringify({ type: \"response.create\", ...requestBody }));\n\t\tawait processResponsesStream(\n\t\t\tstartWebSocketOutputOnFirstEvent(\n\t\t\t\tmapCodexEvents(\n\t\t\t\t\tparseWebSocket(socket, options?.signal, idleTimeoutMs),\n\t\t\t\t\toutput,\n\t\t\t\t\tmodel,\n\t\t\t\t\toptions?.onProviderStreamEvent,\n\t\t\t\t),\n\t\t\t\tonStart,\n\t\t\t),\n\t\t\toutput,\n\t\t\tstream,\n\t\t\tmodel,\n\t\t\t{\n\t\t\t\tserviceTier: resolveRequestedServiceTier(model, options?.serviceTier),\n\t\t\t\tgrammarToolInputProperties,\n\t\t\t\tresolveServiceTier: resolveCodexServiceTier,\n\t\t\t\tapplyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),\n\t\t\t},\n\t\t);\n\t\tif (options?.signal?.aborted) {\n\t\t\tkeepConnection = false;\n\t\t} else if (useCachedContext && entry && output.responseId) {\n\t\t\tconst responseItems = convertResponsesMessages(\n\t\t\t\tmodel,\n\t\t\t\tnormalizeContext({ messages: [output] }),\n\t\t\t\tCODEX_TOOL_CALL_PROVIDERS,\n\t\t\t\t{\n\t\t\t\t\tincludeSystemPrompt: false,\n\t\t\t\t\tgrammarToolInputProperties,\n\t\t\t\t},\n\t\t\t).filter((item) => item.type !== \"function_call_output\" && item.type !== \"custom_tool_call_output\");\n\t\t\tentry.continuation = {\n\t\t\t\tlastRequestBody: fullBody,\n\t\t\t\tlastResponseId: output.responseId,\n\t\t\t\tlastResponseItems: responseItems,\n\t\t\t};\n\t\t}\n\t} catch (error) {\n\t\tif (entry) {\n\t\t\tentry.continuation = undefined;\n\t\t}\n\t\tkeepConnection = false;\n\t\tthrow error;\n\t} finally {\n\t\trelease({ keep: keepConnection });\n\t}\n}\n\n// ============================================================================\n// Error Handling\n// ============================================================================\n\nasync function parseErrorResponse(response: Response): Promise<{ message: string; friendlyMessage?: string }> {\n\tconst raw = await response.text();\n\tlet message = raw || response.statusText || \"Request failed\";\n\tlet friendlyMessage: string | undefined;\n\n\ttry {\n\t\tconst parsed = JSON.parse(raw) as {\n\t\t\terror?: { code?: string; type?: string; message?: string; plan_type?: string; resets_at?: number };\n\t\t};\n\t\tconst err = parsed?.error;\n\t\tif (err) {\n\t\t\tconst code = err.code || err.type || \"\";\n\t\t\tif (/usage_limit_reached|usage_not_included|rate_limit_exceeded/i.test(code) || response.status === 429) {\n\t\t\t\tconst plan = err.plan_type ? ` (${err.plan_type.toLowerCase()} plan)` : \"\";\n\t\t\t\tconst mins = err.resets_at\n\t\t\t\t\t? Math.max(0, Math.round((err.resets_at * 1000 - Date.now()) / 60000))\n\t\t\t\t\t: undefined;\n\t\t\t\tconst when = mins !== undefined ? ` Try again in ~${mins} min.` : \"\";\n\t\t\t\tfriendlyMessage = `You have hit your ChatGPT usage limit${plan}.${when}`.trim();\n\t\t\t}\n\t\t\tmessage = err.message || friendlyMessage || message;\n\t\t}\n\t} catch {}\n\n\treturn { message, friendlyMessage };\n}\n\n// ============================================================================\n// Auth & Headers\n// ============================================================================\n\nfunction extractAccountId(token: string): string {\n\ttry {\n\t\tconst parts = token.split(\".\");\n\t\tif (parts.length !== 3) throw new Error(\"Invalid token\");\n\t\tconst payload = JSON.parse(atob(parts[1]));\n\t\tconst accountId = payload?.[JWT_CLAIM_PATH]?.chatgpt_account_id;\n\t\tif (!accountId) throw new Error(\"No account ID in token\");\n\t\treturn accountId;\n\t} catch {\n\t\tthrow new Error(\"Failed to extract accountId from token\");\n\t}\n}\n\nfunction buildBaseCodexHeaders(\n\tinitHeaders: Record<string, string> | undefined,\n\tadditionalHeaders: ProviderHeaders | undefined,\n\taccountId: string,\n\ttoken: string,\n): Headers {\n\tconst headers = new Headers(initHeaders);\n\tfor (const [key, value] of Object.entries(additionalHeaders || {})) {\n\t\tif (value === null) {\n\t\t\theaders.delete(key);\n\t\t} else {\n\t\t\theaders.set(key, value);\n\t\t}\n\t}\n\theaders.set(\"Authorization\", `Bearer ${token}`);\n\theaders.set(\"chatgpt-account-id\", accountId);\n\theaders.set(\"originator\", \"pi\");\n\theaders.set(\"User-Agent\", getPiUserAgent());\n\treturn headers;\n}\n\nfunction buildSSEHeaders(\n\tinitHeaders: Record<string, string> | undefined,\n\tadditionalHeaders: ProviderHeaders | undefined,\n\taccountId: string,\n\ttoken: string,\n\tsessionId?: string,\n): Headers {\n\tconst headers = buildBaseCodexHeaders(initHeaders, additionalHeaders, accountId, token);\n\theaders.set(\"OpenAI-Beta\", \"responses=experimental\");\n\theaders.set(\"accept\", \"text/event-stream\");\n\theaders.set(\"content-type\", \"application/json\");\n\n\tif (sessionId) {\n\t\theaders.set(\"session-id\", sessionId);\n\t\theaders.set(\"x-client-request-id\", sessionId);\n\t}\n\n\treturn headers;\n}\n\nfunction buildWebSocketHeaders(\n\tinitHeaders: Record<string, string> | undefined,\n\tadditionalHeaders: ProviderHeaders | undefined,\n\taccountId: string,\n\ttoken: string,\n\trequestId: string,\n): Headers {\n\tconst headers = buildBaseCodexHeaders(initHeaders, additionalHeaders, accountId, token);\n\theaders.delete(\"accept\");\n\theaders.delete(\"content-type\");\n\theaders.delete(\"OpenAI-Beta\");\n\theaders.delete(\"openai-beta\");\n\theaders.set(\"OpenAI-Beta\", OPENAI_BETA_RESPONSES_WEBSOCKETS);\n\theaders.set(\"x-client-request-id\", requestId);\n\theaders.set(\"session-id\", requestId);\n\treturn headers;\n}\n"]}

package/dist/api/openai-codex-responses.js

@@ -1,7 +1,6 @@
 import { clampThinkingLevel } from "../models.js";
 import { registerSessionResourceCleanup } from "../session-resources.js";
 import { combineAbortSignals } from "../utils/abort-signals.js";
-import { splitDeferredTools } from "../utils/deferred-tools.js";
 import { appendAssistantMessageDiagnostic, createAssistantMessageDiagnostic, formatThrownValue, } from "../utils/diagnostics.js";
 import { formatProviderError, normalizeProviderError } from "../utils/error-body.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
@@ -9,6 +8,8 @@ import { headersToRecord } from "../utils/headers.js";
 import { resolveHttpProxyUrlForTarget } from "../utils/node-http-proxy.js";
 import { getPiUserAgent } from "../utils/pi-user-agent.js";
 import { createStreamDeadline, withStreamDeadline } from "../utils/stream-deadline.js";
+import { getSystemMessageText } from "../utils/text.js";
+import { getDeclaredTools, getInitialSystemMessage, normalizeContext, resolveTranscript, resolveTranscriptTools, } from "../utils/transcript.js";
 import { uuidv7 } from "../utils/uuid.js";
 import { createGrammarToolInputProperties } from "./constrained-sampling.js";
 import { clampOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
@@ -142,6 +143,7 @@ function compressRequestBodyZstd(bodyJson) {
 // ============================================================================
 export const stream = (model, context, options) => {
     const stream = new AssistantMessageEventStream();
+    const normalizedContext = resolveTranscript(context, model.compat?.supportsMidConvoSystemMessages);
     (async () => {
         const output = {
             role: "assistant",
@@ -167,10 +169,10 @@ export const stream = (model, context, options) => {
                 throw new Error(`No API key for provider: ${model.provider}`);
             }
             const accountId = extractAccountId(apiKey);
-            const grammarToolInputProperties = createGrammarToolInputProperties(context.tools, model.compat?.supportsOpenAIGrammarTools ?? false);
+            const grammarToolInputProperties = createGrammarToolInputProperties(getDeclaredTools(normalizedContext.messages), model.compat?.supportsOpenAIGrammarTools ?? false);
             const cacheSessionId = options?.cacheRetention === "none" ? undefined : options?.sessionId;
             const codexSessionId = clampOpenAIPromptCacheKey(cacheSessionId);
-            let body = buildRequestBody(model, context, options, codexSessionId, grammarToolInputProperties);
+            let body = buildRequestBody(model, normalizedContext, options, codexSessionId, grammarToolInputProperties);
             const nextBody = await options?.onPayload?.(body, model);
             if (nextBody !== undefined) {
                 body = nextBody;
@@ -231,7 +233,7 @@ export const stream = (model, context, options) => {
                         }
                         appendAssistantMessageDiagnostic(output, createAssistantMessageDiagnostic("provider_transport_failure", error, {
                             configuredTransport: transport,
-                            fallbackTransport: websocketStarted ? undefined : "sse",
+                            ...(websocketStarted ? {} : { fallbackTransport: "sse" }),
                             eventsEmitted: websocketStarted,
                             phase: websocketStarted ? "after_message_stream_start" : "before_message_stream_start",
                             requestBytes: new TextEncoder().encode(bodyJson).byteLength,
@@ -374,33 +376,29 @@ export const streamSimple = (model, context, options) => {
 // ============================================================================
 // Request Building
 // ============================================================================
-function buildRequestBody(model, context, options, cacheSessionId, grammarToolInputProperties = createGrammarToolInputProperties(context.tools, model.compat?.supportsOpenAIGrammarTools ?? false)) {
+function buildRequestBody(model, context, options, cacheSessionId, grammarToolInputProperties = createGrammarToolInputProperties(getDeclaredTools(context.messages), model.compat?.supportsOpenAIGrammarTools ?? false)) {
     const supportsStrictMode = model.compat?.supportsStrictMode ?? true;
     const supportsOpenAIGrammarTools = model.compat?.supportsOpenAIGrammarTools ?? false;
-    const deferredToolsMode = model.compat?.supportsAdditionalTools
-        ? "additional-tools"
-        : model.compat?.supportsToolSearch
-            ? "tool-search"
-            : undefined;
-    const toolPlacement = splitDeferredTools(context, deferredToolsMode !== undefined);
+    const supportsAdditionalTools = model.compat?.supportsAdditionalTools ?? false;
+    const supportsToolSearch = model.compat?.supportsToolSearch ?? false;
+    const transcriptTools = resolveTranscriptTools(context.messages, supportsAdditionalTools || supportsToolSearch);
     const messages = convertResponsesMessages(model, context, CODEX_TOOL_CALL_PROVIDERS, {
         includeSystemPrompt: false,
         grammarToolInputProperties,
-        deferredTools: toolPlacement.deferred,
-        deferredToolsMode,
-        toolOptions: {
-            strict: null,
-            supportsStrictMode,
-            supportsOpenAIGrammarTools,
-        },
+        supportsMidConvoSystemMessages: model.compat?.supportsMidConvoSystemMessages ?? false,
+        supportsAdditionalTools,
+        supportsToolSearch,
+        toolOptions: { strict: null, supportsStrictMode, supportsOpenAIGrammarTools },
     });
+    const initialSystemMessage = getInitialSystemMessage(context.messages);
+    const instructions = initialSystemMessage ? getSystemMessageText(initialSystemMessage) : "";
     const body = {
         // A fast variant keeps its canonical `-fast` id on the model object — that is the identity the
         // caller selected and records — while routing to the base upstream model plus a service tier.
         model: model.fastRoute?.upstreamModelId ?? model.id,
         store: false,
         stream: true,
-        instructions: context.systemPrompt || "You are a helpful assistant.",
+        instructions: instructions || "You are a helpful assistant.",
         input: messages,
         text: { verbosity: options?.textVerbosity || "low" },
         include: ["reasoning.encrypted_content"],
@@ -416,8 +414,8 @@ function buildRequestBody(model, context, options, cacheSessionId, grammarToolIn
     if (requestedServiceTier !== undefined) {
         body.service_tier = requestedServiceTier;
     }
-    if (toolPlacement.immediate.length > 0) {
-        body.tools = convertResponsesTools(toolPlacement.immediate, {
+    if (transcriptTools.requestTools.length > 0) {
+        body.tools = convertResponsesTools(transcriptTools.requestTools, {
             strict: null,
             supportsStrictMode,
             supportsOpenAIGrammarTools,
@@ -491,7 +489,7 @@ function resolveCodexWebSocketUrl(baseUrl) {
 // Response Processing
 // ============================================================================
 async function processStream(response, output, stream, model, grammarToolInputProperties, options, streamDeadline) {
-    const events = mapCodexEvents(parseSSE(response, streamDeadline.signal), output);
+    const events = mapCodexEvents(parseSSE(response, streamDeadline.signal), output, model, options?.onProviderStreamEvent);
     await processResponsesStream(withStreamDeadline(events, streamDeadline.deadlineMs, streamDeadline.abort), output, stream, model, {
         serviceTier: resolveRequestedServiceTier(model, options?.serviceTier),
         grammarToolInputProperties,
@@ -519,8 +517,17 @@ class CodexProtocolError extends Error {
         this.cause = options?.cause;
     }
 }
+class ProviderStreamEventCallbackError extends Error {
+    constructor(cause) {
+        super(formatThrownValue(cause));
+        this.name = "ProviderStreamEventCallbackError";
+        this.cause = cause;
+    }
+}
 function isCodexNonTransportError(error) {
-    return error instanceof CodexApiError || error instanceof CodexProtocolError;
+    return (error instanceof CodexApiError ||
+        error instanceof CodexProtocolError ||
+        error instanceof ProviderStreamEventCallbackError);
 }
 function isWebSocketConnectionLimitReachedError(error) {
     return error instanceof CodexApiError && error.code === WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE;
@@ -539,8 +546,15 @@ function extractCodexEventError(event) {
                 : undefined,
     };
 }
-async function* mapCodexEvents(events, output) {
+async function* mapCodexEvents(events, output, model, onProviderStreamEvent) {
     for await (const event of events) {
+        try {
+            await onProviderStreamEvent?.(event, model);
+        }
+        catch (error) {
+            // Keep callback failures out of Codex's WebSocket retry and SSE fallback path.
+            throw new ProviderStreamEventCallbackError(error);
+        }
         const type = typeof event.type === "string" ? event.type : undefined;
         if (!type)
             continue;
@@ -1202,7 +1216,7 @@ async function processWebSocketStream(url, body, headers, output, stream, model,
     }
     try {
         socket.send(JSON.stringify({ type: "response.create", ...requestBody }));
-        await processResponsesStream(startWebSocketOutputOnFirstEvent(mapCodexEvents(parseWebSocket(socket, options?.signal, idleTimeoutMs), output), onStart), output, stream, model, {
+        await processResponsesStream(startWebSocketOutputOnFirstEvent(mapCodexEvents(parseWebSocket(socket, options?.signal, idleTimeoutMs), output, model, options?.onProviderStreamEvent), onStart), output, stream, model, {
             serviceTier: resolveRequestedServiceTier(model, options?.serviceTier),
             grammarToolInputProperties,
             resolveServiceTier: resolveCodexServiceTier,
@@ -1212,7 +1226,7 @@ async function processWebSocketStream(url, body, headers, output, stream, model,
             keepConnection = false;
         }
         else if (useCachedContext && entry && output.responseId) {
-            const responseItems = convertResponsesMessages(model, { messages: [output] }, CODEX_TOOL_CALL_PROVIDERS, {
+            const responseItems = convertResponsesMessages(model, normalizeContext({ messages: [output] }), CODEX_TOOL_CALL_PROVIDERS, {
                 includeSystemPrompt: false,
                 grammarToolInputProperties,
             }).filter((item) => item.type !== "function_call_output" && item.type !== "custom_tool_call_output");