@bastani/pi-ai 0.9.28 → 0.9.29-alpha.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (64) hide show
  1. package/CHANGELOG.md +11 -0
  2. package/README.md +29 -2
  3. package/dist/api/openai-codex-responses.d.ts.map +1 -1
  4. package/dist/api/openai-codex-responses.js +182 -146
  5. package/dist/api/openai-codex-responses.js.map +1 -1
  6. package/dist/api/openai-images.d.ts +17 -0
  7. package/dist/api/openai-images.d.ts.map +1 -0
  8. package/dist/api/openai-images.js +114 -0
  9. package/dist/api/openai-images.js.map +1 -0
  10. package/dist/api/openai-images.lazy.d.ts +3 -0
  11. package/dist/api/openai-images.lazy.d.ts.map +1 -0
  12. package/dist/api/openai-images.lazy.js +4 -0
  13. package/dist/api/openai-images.lazy.js.map +1 -0
  14. package/dist/api/openai-responses-shared.d.ts +26 -0
  15. package/dist/api/openai-responses-shared.d.ts.map +1 -1
  16. package/dist/api/openai-responses-shared.js +81 -0
  17. package/dist/api/openai-responses-shared.js.map +1 -1
  18. package/dist/api/openai-responses.d.ts.map +1 -1
  19. package/dist/api/openai-responses.js +102 -10
  20. package/dist/api/openai-responses.js.map +1 -1
  21. package/dist/env-api-keys.d.ts.map +1 -1
  22. package/dist/env-api-keys.js +1 -0
  23. package/dist/env-api-keys.js.map +1 -1
  24. package/dist/index.d.ts +1 -0
  25. package/dist/index.d.ts.map +1 -1
  26. package/dist/index.js.map +1 -1
  27. package/dist/models.d.ts +4 -2
  28. package/dist/models.d.ts.map +1 -1
  29. package/dist/models.generated.d.ts +4 -0
  30. package/dist/models.generated.d.ts.map +1 -1
  31. package/dist/models.generated.js +4 -0
  32. package/dist/models.generated.js.map +1 -1
  33. package/dist/models.js +2 -1
  34. package/dist/models.js.map +1 -1
  35. package/dist/providers/all.d.ts.map +1 -1
  36. package/dist/providers/all.js +2 -0
  37. package/dist/providers/all.js.map +1 -1
  38. package/dist/providers/data/.manifest.json +1 -1
  39. package/dist/providers/data/baseten.json +1 -1
  40. package/dist/providers/data/cloudflare-ai-gateway.json +1 -1
  41. package/dist/providers/data/cloudflare-workers-ai.json +1 -1
  42. package/dist/providers/data/google-vertex.json +1 -1
  43. package/dist/providers/data/openai-codex.json +1 -1
  44. package/dist/providers/data/openai-images.json +1 -0
  45. package/dist/providers/data/openai.json +1 -1
  46. package/dist/providers/data/opencode.json +1 -1
  47. package/dist/providers/data/openrouter.json +1 -1
  48. package/dist/providers/data/vercel-ai-gateway.json +1 -1
  49. package/dist/providers/images/register-builtins.d.ts +1 -0
  50. package/dist/providers/images/register-builtins.d.ts.map +1 -1
  51. package/dist/providers/images/register-builtins.js +18 -0
  52. package/dist/providers/images/register-builtins.js.map +1 -1
  53. package/dist/providers/openai-images.d.ts +3 -0
  54. package/dist/providers/openai-images.d.ts.map +1 -0
  55. package/dist/providers/openai-images.js +19 -0
  56. package/dist/providers/openai-images.js.map +1 -0
  57. package/dist/providers/openai-images.models.d.ts +5 -0
  58. package/dist/providers/openai-images.models.d.ts.map +1 -0
  59. package/dist/providers/openai-images.models.js +8 -0
  60. package/dist/providers/openai-images.models.js.map +1 -0
  61. package/dist/types.d.ts +16 -2
  62. package/dist/types.d.ts.map +1 -1
  63. package/dist/types.js.map +1 -1
  64. package/package.json +1 -1
package/CHANGELOG.md CHANGED
@@ -4,6 +4,17 @@ This package is a Bastani fork of `@earendil-works/pi-ai`. Upstream history at t
4
4
 
5
5
  ## [Unreleased]
6
6
 
7
+ ## [0.9.29-alpha.1] - 2026-10-08
8
+
9
+ ### Added
10
+
11
+ - Added the `ultrafast` service tier to `gpt-6.1-sol` and `gpt-5.6-sol` on `openai` and `openai-api`, and to `gpt-6.1-sol` on `openai-codex`. GPT-6.1 Sol uses OpenAI's published Ultrafast rates ($12 input, $60 output per million tokens), with cache rates derived at the same six-times-Standard ratio; GPT-5.6 Sol Ultrafast is a preview without published rates and is priced at six times its Standard rates ([#3529](https://github.com/bastani-inc/atomic/issues/3529)).
12
+ - Added the `openai-images` provider (`openaiImagesProvider()`) and the `openai-images` image API for OpenAI's Images API, with `gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, `gpt-image-2`, `gpt-image-1.5`, `gpt-image-1`, and `gpt-image-1-mini`. Text input becomes the prompt; image inputs send the request to `/images/edits`. `OpenAIImagesOptions` adds `size`, `quality`, `n`, and `background`, and `Models.generateImages()` options are now typed per image API through `ImageApiOptionsMap`. Like `openai-decisions`, the provider has no login of its own: it reuses the API key stored for `openai-api`, then the one stored for `openai`, then `OPENAI_API_KEY`, and skips a ChatGPT sign-in on `openai` ([#3527](https://github.com/bastani-inc/atomic/issues/3527)).
13
+
14
+ ### Changed
15
+
16
+ - OpenAI Responses and OpenAI Codex Responses requests now retry once at the default tier when the server rejects the requested `service_tier` before producing output, and record a `service_tier_unavailable` warning diagnostic on the assistant message once the default-tier request is accepted. On OpenAI Responses with an API key, a response that reports a slower tier than requested records the same warning. Requests made with a ChatGPT sign-in token, on `openai` or `openai-codex`, do not warn on a reported `default` and are priced at the requested tier, because the ChatGPT backend reports `default` for every tier it applies. Otherwise pricing follows the tier the server reports ([#3529](https://github.com/bastani-inc/atomic/issues/3529)).
17
+
7
18
  ## [0.9.28] - 2026-10-08
8
19
 
9
20
  ### Breaking Changes
package/README.md CHANGED
@@ -404,7 +404,7 @@ if (modelAuth) {
404
404
 
405
405
  Both overloads resolve credentials, refresh expired OAuth when necessary, and may return an auth-derived `apiKey`, `headers`, or `baseUrl`. `getAuth()` resolves `undefined` for unconfigured providers and rejects with `ModelsError` when something is actually broken (`"oauth"`: token refresh failed, credential preserved for re-login; `"auth"`: key resolution or credential store failure). Request paths surface the same failures as stream errors.
406
406
 
407
- A provider can reuse other providers' stored API keys instead of offering its own login. List their IDs in `ApiKeyAuth.borrowCredentialsFrom`, or build the auth with `sharedApiKeyAuth(name, envVars, providerIds)`. The provider then has no stored credential of its own: the first listed provider with a stored API key supplies it, in the order given. Stored OAuth credentials are skipped, and the environment variables apply when none qualifies. `openai-decisions` uses this with `['openai-api', 'openai']`.
407
+ A provider can reuse other providers' stored API keys instead of offering its own login. List their IDs in `ApiKeyAuth.borrowCredentialsFrom`, or build the auth with `sharedApiKeyAuth(name, envVars, providerIds)`. The provider then has no stored credential of its own: the first listed provider with a stored API key supplies it, in the order given. Stored OAuth credentials are skipped, and the environment variables apply when none qualifies. `openai-decisions` and `openai-images` use this with `['openai-api', 'openai']`.
408
408
 
409
409
  `getAuth()`, `checkAuth()`, `getAvailable()`, login, and logout accept optional caller cancellation through their existing options or interaction objects and remain unbounded when no signal is supplied. Provider `login`, `ApiKeyAuth.check`, `ApiKeyAuth.resolve`, and `OAuthAuth.refresh` implementations always receive a concrete signal and must honor it for blocking work.
410
410
 
@@ -828,7 +828,7 @@ const pixels = createProvider({
828
828
  models.setProvider(pixels);
829
829
  ```
830
830
 
831
- The old global API (`getImageModel()` / `getImageModels()` / `getImageProviders()` / `generateImages()`) remains available on the [compat entrypoint](#migrating-from-the-old-global-api):
831
+ The old global API (`getImageModel()` / `getImageModels()` / `getImageProviders()` / `generateImages()`) remains available on the [compat entrypoint](#migrating-from-the-old-global-api). Its `generateImages()` dispatches on `model.api` and takes `apiKey` explicitly:
832
832
 
833
833
  ```typescript
834
834
  import { getImageModel, generateImages } from '@bastani/pi-ai/compat';
@@ -841,6 +841,18 @@ const result = await generateImages(model, {
841
841
  });
842
842
  ```
843
843
 
844
+ `registerImagesApiProvider({ api, generateImages })` from the compat entrypoint is the extension point for image backends that are not built in. It registers an implementation for an image API ID, and the global `generateImages()` then dispatches models with that `api` to it:
845
+
846
+ ```typescript
847
+ import { generateImages, registerImagesApiProvider } from '@bastani/pi-ai/compat';
848
+
849
+ registerImagesApiProvider({
850
+ api: 'pixels-images',
851
+ generateImages: async (model, context, options) => { /* call the backend with options?.apiKey */ },
852
+ });
853
+ const result = await generateImages(pixelsModel, context, { apiKey: process.env.PIXELS_API_KEY });
854
+ ```
855
+
844
856
  Some models also support image input:
845
857
 
846
858
  ```typescript
@@ -862,6 +874,21 @@ console.log(model.input); // ['text'] or ['text', 'image']
862
874
  console.log(model.output); // ['image'] or ['image', 'text']
863
875
  ```
864
876
 
877
+ ### OpenAI Images API
878
+
879
+ The `openai-images` provider serves OpenAI's GPT Image models (`gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, `gpt-image-2`, `gpt-image-1.5`, `gpt-image-1`, `gpt-image-1-mini`) through the Images API. It borrows the API key stored for `openai-api`, then the one stored for `openai`, then `OPENAI_API_KEY`; a ChatGPT OAuth credential on `openai` is skipped, so with only that credential the models are unavailable and `generateImages()` returns `Provider is not configured: openai-images`. Text blocks become the `prompt`. Without image blocks the request goes to `/images/generations`; with them it goes to `/images/edits`, uploading each image block. `OpenAIImagesOptions` adds `size`, `quality`, `n`, and `background`, each sent only when set:
880
+
881
+ ```typescript
882
+ import { getBuiltinImageModel } from '@bastani/pi-ai/providers/all';
883
+
884
+ const flare = getBuiltinImageModel('openai-images', 'gpt-image-2.5-flare');
885
+ const result = await models.generateImages(flare, {
886
+ input: [{ type: 'text', text: 'A red fox in the snow, watercolor' }]
887
+ }, { size: '1024x1024', quality: 'low' });
888
+ ```
889
+
890
+ Each returned image is an `ImageContent` block whose `mimeType` follows the response's `output_format`.
891
+
865
892
  ### Notes and Limitations
866
893
 
867
894
  - Image models and chat models share `Models` and `Provider`; list them with `getModelsOfType('image')` and run them with `generateImages()`, never the chat/stream APIs.
@@ -1 +1 @@
1
- {"version":3,"file":"openai-codex-responses.d.ts","sourceRoot":"","sources":["../../src/api/openai-codex-responses.ts"],"names":[],"mappings":"AAKA,OAAO,KAAK,EAMX,mBAAmB,EACnB,cAAc,EACd,aAAa,EAGb,MAAM,aAAa,CAAC;AAwBrB,OAAO,EAON,KAAK,oBAAoB,EAEzB,MAAM,8BAA8B,CAAC;AAkCtC,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,oBAAoB,CAAC;IACnC,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;AA2WF,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"}
1
+ {"version":3,"file":"openai-codex-responses.d.ts","sourceRoot":"","sources":["../../src/api/openai-codex-responses.ts"],"names":[],"mappings":"AAKA,OAAO,KAAK,EAMX,mBAAmB,EACnB,cAAc,EACd,aAAa,EAGb,MAAM,aAAa,CAAC;AAwBrB,OAAO,EASN,KAAK,oBAAoB,EAGzB,MAAM,8BAA8B,CAAC;AAkCtC,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,oBAAoB,CAAC;IACnC,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,CA8TxF,CAAC;AAEF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,wBAAwB,EAAE,mBAAmB,CAqBtF,CAAC;AAqXF,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"}
@@ -13,7 +13,7 @@ import { getDeclaredTools, getInitialSystemMessage, normalizeContext, resolveTra
13
13
  import { uuidv7 } from "../utils/uuid.js";
14
14
  import { createGrammarToolInputProperties } from "./constrained-sampling.js";
15
15
  import { clampOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
16
- import { applyServiceTierPricing, assertPayloadPreservesFastRoute, codexServiceTierForRequest, convertResponsesMessages, convertResponsesTools, processResponsesStream, resolveRequestedServiceTier, } from "./openai-responses-shared.js";
16
+ import { appendServiceTierRejectedWarning, applyServiceTierPricing, assertPayloadPreservesFastRoute, codexServiceTierForRequest, convertResponsesMessages, convertResponsesTools, isServiceTierRejection, processResponsesStream, resolveRequestedServiceTier, resolveChatGPTBackendServiceTier, } from "./openai-responses-shared.js";
17
17
  import { buildBaseOptions } from "./simple-options.js";
18
18
  // ============================================================================
19
19
  // Configuration
@@ -181,161 +181,184 @@ export const stream = (model, context, options) => {
181
181
  const websocketRequestId = codexSessionId || uuidv7();
182
182
  const sseHeaders = buildSSEHeaders(model.headers, options?.headers, accountId, apiKey, codexSessionId);
183
183
  const websocketHeaders = buildWebSocketHeaders(model.headers, options?.headers, accountId, apiKey, websocketRequestId);
184
- const bodyJson = JSON.stringify(body);
185
184
  const httpTimeoutMs = normalizeTimeoutMs(options?.timeoutMs);
186
185
  const websocketConnectTimeoutMs = normalizeTimeoutMs(options?.websocketConnectTimeoutMs);
187
186
  const transport = options?.transport || "auto";
188
187
  let startEmitted = false;
189
- const websocketDisabledForSession = transport !== "sse" && isWebSocketSseFallbackActive(cacheSessionId);
190
- if (websocketDisabledForSession) {
191
- recordWebSocketSseFallback(cacheSessionId);
192
- }
193
- if (transport !== "sse" && !websocketDisabledForSession) {
194
- let websocketStarted = false;
195
- let retriedWebSocketConnectionLimit = false;
196
- let retriedMissingWebSocketContinuation = false;
197
- while (true) {
198
- websocketStarted = false;
188
+ const emitStart = () => {
189
+ if (!startEmitted) {
190
+ startEmitted = true;
191
+ stream.push({ type: "start", partial: output });
192
+ }
193
+ };
194
+ // One request over WebSocket (falling back to SSE), processed through to a successful stop reason.
195
+ const runRequest = async (requestBody, requestSseHeaders, requestWebsocketHeaders, requestServiceTier) => {
196
+ const bodyJson = JSON.stringify(requestBody);
197
+ const websocketDisabledForSession = transport !== "sse" && isWebSocketSseFallbackActive(cacheSessionId);
198
+ if (websocketDisabledForSession) {
199
+ recordWebSocketSseFallback(cacheSessionId);
200
+ }
201
+ if (transport !== "sse" && !websocketDisabledForSession) {
202
+ let websocketStarted = false;
203
+ let retriedWebSocketConnectionLimit = false;
204
+ let retriedMissingWebSocketContinuation = false;
205
+ while (true) {
206
+ websocketStarted = false;
207
+ try {
208
+ await processWebSocketStream(resolveCodexWebSocketUrl(model.baseUrl), requestBody, requestWebsocketHeaders, output, stream, model, () => {
209
+ websocketStarted = true;
210
+ emitStart();
211
+ }, streamDeadline.deadlineMs, websocketConnectTimeoutMs, cacheSessionId, accountId, grammarToolInputProperties, requestServiceTier, options);
212
+ if (options?.signal?.aborted) {
213
+ throw new Error("Request was aborted");
214
+ }
215
+ assertSuccessfulOutput(output);
216
+ return;
217
+ }
218
+ catch (error) {
219
+ const aborted = options?.signal?.aborted;
220
+ const connectionLimitBeforeStart = !websocketStarted && isWebSocketConnectionLimitReachedError(error);
221
+ const previousResponseNotFound = isPreviousResponseNotFoundError(error);
222
+ if (!aborted && previousResponseNotFound && !retriedMissingWebSocketContinuation) {
223
+ retriedMissingWebSocketContinuation = true;
224
+ continue;
225
+ }
226
+ if (!aborted && connectionLimitBeforeStart && !retriedWebSocketConnectionLimit) {
227
+ retriedWebSocketConnectionLimit = true;
228
+ continue;
229
+ }
230
+ if (aborted || (isCodexNonTransportError(error) && !connectionLimitBeforeStart)) {
231
+ throw error;
232
+ }
233
+ appendAssistantMessageDiagnostic(output, createAssistantMessageDiagnostic("provider_transport_failure", error, {
234
+ configuredTransport: transport,
235
+ ...(websocketStarted ? {} : { fallbackTransport: "sse" }),
236
+ eventsEmitted: websocketStarted,
237
+ phase: websocketStarted ? "after_message_stream_start" : "before_message_stream_start",
238
+ requestBytes: new TextEncoder().encode(bodyJson).byteLength,
239
+ }));
240
+ recordWebSocketFailure(cacheSessionId, error);
241
+ if (websocketStarted) {
242
+ throw error;
243
+ }
244
+ recordWebSocketSseFallback(cacheSessionId);
245
+ break;
246
+ }
247
+ }
248
+ }
249
+ // Compress the request body once for the SSE path. The Codex backend
250
+ // decodes Content-Encoding: zstd; the WebSocket transport above sends the
251
+ // uncompressed JSON frame, matching the official Codex client.
252
+ const compressedBody = compressRequestBodyZstd(bodyJson);
253
+ if (compressedBody) {
254
+ requestSseHeaders.set("content-encoding", "zstd");
255
+ }
256
+ const sseBody = compressedBody ?? bodyJson;
257
+ // Fetch with retry logic for rate limits and transient errors
258
+ let response;
259
+ let lastError;
260
+ const maxRetries = options?.maxRetries ?? DEFAULT_MAX_RETRIES;
261
+ for (let attempt = 0; attempt <= maxRetries; attempt++) {
262
+ if (options?.signal?.aborted) {
263
+ throw new Error("Request was aborted");
264
+ }
199
265
  try {
200
- await processWebSocketStream(resolveCodexWebSocketUrl(model.baseUrl), body, websocketHeaders, output, stream, model, () => {
201
- websocketStarted = true;
202
- if (!startEmitted) {
203
- startEmitted = true;
204
- stream.push({ type: "start", partial: output });
266
+ const headerTimeoutSignal = httpTimeoutMs !== undefined && httpTimeoutMs > 0 ? AbortSignal.timeout(httpTimeoutMs) : undefined;
267
+ const combinedSignal = combineAbortSignals([streamDeadline.signal, headerTimeoutSignal]);
268
+ try {
269
+ response = await (options?.fetch ?? globalThis.fetch)(resolveCodexUrl(model.baseUrl), {
270
+ method: "POST",
271
+ headers: requestSseHeaders,
272
+ body: sseBody,
273
+ signal: combinedSignal.signal,
274
+ });
275
+ }
276
+ catch (error) {
277
+ if (headerTimeoutSignal?.aborted && !options?.signal?.aborted) {
278
+ throw new Error(`Codex SSE response headers timed out after ${httpTimeoutMs}ms`);
205
279
  }
206
- }, streamDeadline.deadlineMs, websocketConnectTimeoutMs, cacheSessionId, accountId, grammarToolInputProperties, options);
207
- if (options?.signal?.aborted) {
208
- throw new Error("Request was aborted");
280
+ throw error;
281
+ }
282
+ finally {
283
+ combinedSignal.cleanup();
284
+ }
285
+ await options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);
286
+ if (response.ok) {
287
+ break;
288
+ }
289
+ const errorText = await response.text();
290
+ if (attempt < maxRetries && isRetryableError(response.status, errorText)) {
291
+ const retryAfterDelayMs = getRetryAfterDelayMs(response.headers);
292
+ const delayMs = retryAfterDelayMs === undefined
293
+ ? BASE_DELAY_MS * 2 ** attempt
294
+ : validateRetryDelayMs(retryAfterDelayMs, options);
295
+ await sleep(delayMs, options?.signal);
296
+ continue;
209
297
  }
210
- assertSuccessfulOutput(output);
211
- stream.push({
212
- type: "done",
213
- reason: output.stopReason,
214
- message: output,
298
+ // Parse error for friendly message on final attempt or non-retryable error
299
+ const fakeResponse = new Response(errorText, {
300
+ status: response.status,
301
+ statusText: response.statusText,
215
302
  });
216
- stream.end();
217
- return;
303
+ const info = await parseErrorResponse(fakeResponse);
304
+ throw new Error(info.friendlyMessage || info.message);
218
305
  }
219
306
  catch (error) {
220
- const aborted = options?.signal?.aborted;
221
- const connectionLimitBeforeStart = !websocketStarted && isWebSocketConnectionLimitReachedError(error);
222
- const previousResponseNotFound = isPreviousResponseNotFoundError(error);
223
- if (!aborted && previousResponseNotFound && !retriedMissingWebSocketContinuation) {
224
- retriedMissingWebSocketContinuation = true;
225
- continue;
307
+ if (error instanceof Error) {
308
+ if (error.name === "AbortError" || error.message === "Request was aborted") {
309
+ throw new Error("Request was aborted");
310
+ }
226
311
  }
227
- if (!aborted && connectionLimitBeforeStart && !retriedWebSocketConnectionLimit) {
228
- retriedWebSocketConnectionLimit = true;
312
+ lastError = error instanceof Error ? error : new Error(String(error));
313
+ // Network errors are retryable
314
+ if (attempt < maxRetries &&
315
+ !(lastError instanceof RetryDelayExceededError) &&
316
+ !lastError.message.includes("usage limit")) {
317
+ const delayMs = BASE_DELAY_MS * 2 ** attempt;
318
+ await sleep(delayMs, options?.signal);
229
319
  continue;
230
320
  }
231
- if (aborted || (isCodexNonTransportError(error) && !connectionLimitBeforeStart)) {
232
- throw error;
233
- }
234
- appendAssistantMessageDiagnostic(output, createAssistantMessageDiagnostic("provider_transport_failure", error, {
235
- configuredTransport: transport,
236
- ...(websocketStarted ? {} : { fallbackTransport: "sse" }),
237
- eventsEmitted: websocketStarted,
238
- phase: websocketStarted ? "after_message_stream_start" : "before_message_stream_start",
239
- requestBytes: new TextEncoder().encode(bodyJson).byteLength,
240
- }));
241
- recordWebSocketFailure(cacheSessionId, error);
242
- if (websocketStarted) {
243
- throw error;
244
- }
245
- recordWebSocketSseFallback(cacheSessionId);
246
- break;
321
+ throw lastError;
247
322
  }
248
323
  }
249
- }
250
- // Compress the request body once for the SSE path. The Codex backend
251
- // decodes Content-Encoding: zstd; the WebSocket transport above sends the
252
- // uncompressed JSON frame, matching the official Codex client.
253
- const compressedBody = compressRequestBodyZstd(bodyJson);
254
- if (compressedBody) {
255
- sseHeaders.set("content-encoding", "zstd");
256
- }
257
- const sseBody = compressedBody ?? bodyJson;
258
- // Fetch with retry logic for rate limits and transient errors
259
- let response;
260
- let lastError;
261
- const maxRetries = options?.maxRetries ?? DEFAULT_MAX_RETRIES;
262
- for (let attempt = 0; attempt <= maxRetries; attempt++) {
324
+ if (!response?.ok) {
325
+ throw lastError ?? new Error("Failed after retries");
326
+ }
327
+ if (!response.body) {
328
+ throw new Error("No response body");
329
+ }
330
+ emitStart();
331
+ await processStream(response, output, stream, model, grammarToolInputProperties, requestServiceTier, options, streamDeadline);
263
332
  if (options?.signal?.aborted) {
264
333
  throw new Error("Request was aborted");
265
334
  }
335
+ assertSuccessfulOutput(output);
336
+ };
337
+ try {
338
+ await runRequest(body, sseHeaders, websocketHeaders, resolveCodexRequestServiceTier(model, options?.serviceTier));
339
+ }
340
+ catch (error) {
341
+ const rejectedServiceTier = body.service_tier;
342
+ if (rejectedServiceTier == null ||
343
+ options?.signal?.aborted ||
344
+ output.content.length > 0 ||
345
+ !isCodexServiceTierRejection(error)) {
346
+ throw error;
347
+ }
348
+ // The backend refused the tier before producing output: retry once at the default tier, with no
349
+ // service_tier and no tier routing hint.
350
+ const { service_tier: _rejectedServiceTier, ...defaultTierBody } = body;
351
+ let ranAtDefault = false;
266
352
  try {
267
- const headerTimeoutSignal = httpTimeoutMs !== undefined && httpTimeoutMs > 0 ? AbortSignal.timeout(httpTimeoutMs) : undefined;
268
- const combinedSignal = combineAbortSignals([streamDeadline.signal, headerTimeoutSignal]);
269
- try {
270
- response = await (options?.fetch ?? globalThis.fetch)(resolveCodexUrl(model.baseUrl), {
271
- method: "POST",
272
- headers: sseHeaders,
273
- body: sseBody,
274
- signal: combinedSignal.signal,
275
- });
276
- }
277
- catch (error) {
278
- if (headerTimeoutSignal?.aborted && !options?.signal?.aborted) {
279
- throw new Error(`Codex SSE response headers timed out after ${httpTimeoutMs}ms`);
280
- }
281
- throw error;
282
- }
283
- finally {
284
- combinedSignal.cleanup();
285
- }
286
- await options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);
287
- if (response.ok) {
288
- break;
289
- }
290
- const errorText = await response.text();
291
- if (attempt < maxRetries && isRetryableError(response.status, errorText)) {
292
- const retryAfterDelayMs = getRetryAfterDelayMs(response.headers);
293
- const delayMs = retryAfterDelayMs === undefined
294
- ? BASE_DELAY_MS * 2 ** attempt
295
- : validateRetryDelayMs(retryAfterDelayMs, options);
296
- await sleep(delayMs, options?.signal);
297
- continue;
298
- }
299
- // Parse error for friendly message on final attempt or non-retryable error
300
- const fakeResponse = new Response(errorText, {
301
- status: response.status,
302
- statusText: response.statusText,
303
- });
304
- const info = await parseErrorResponse(fakeResponse);
305
- throw new Error(info.friendlyMessage || info.message);
353
+ await runRequest(defaultTierBody, withoutCodexRoutingHint(sseHeaders), withoutCodexRoutingHint(websocketHeaders), undefined);
354
+ ranAtDefault = true;
306
355
  }
307
- catch (error) {
308
- if (error instanceof Error) {
309
- if (error.name === "AbortError" || error.message === "Request was aborted") {
310
- throw new Error("Request was aborted");
311
- }
312
- }
313
- lastError = error instanceof Error ? error : new Error(String(error));
314
- // Network errors are retryable
315
- if (attempt < maxRetries &&
316
- !(lastError instanceof RetryDelayExceededError) &&
317
- !lastError.message.includes("usage limit")) {
318
- const delayMs = BASE_DELAY_MS * 2 ** attempt;
319
- await sleep(delayMs, options?.signal);
320
- continue;
356
+ finally {
357
+ if (ranAtDefault || output.content.length > 0) {
358
+ appendServiceTierRejectedWarning(output, model, rejectedServiceTier);
321
359
  }
322
- throw lastError;
323
360
  }
324
361
  }
325
- if (!response?.ok) {
326
- throw lastError ?? new Error("Failed after retries");
327
- }
328
- if (!response.body) {
329
- throw new Error("No response body");
330
- }
331
- if (!startEmitted) {
332
- startEmitted = true;
333
- stream.push({ type: "start", partial: output });
334
- }
335
- await processStream(response, output, stream, model, grammarToolInputProperties, options, streamDeadline);
336
- if (options?.signal?.aborted) {
337
- throw new Error("Request was aborted");
338
- }
339
362
  assertSuccessfulOutput(output);
340
363
  stream.push({ type: "done", reason: output.stopReason, message: output });
341
364
  stream.end();
@@ -441,13 +464,6 @@ function buildRequestBody(model, context, options, cacheSessionId, grammarToolIn
441
464
  function resolveCodexRequestServiceTier(model, optionsServiceTier) {
442
465
  return codexServiceTierForRequest(model, resolveRequestedServiceTier(model, optionsServiceTier));
443
466
  }
444
- function resolveCodexServiceTier(responseServiceTier, requestServiceTier) {
445
- if (responseServiceTier === "default" &&
446
- (requestServiceTier === "flex" || requestServiceTier === "priority" || requestServiceTier === "ultrafast")) {
447
- return requestServiceTier;
448
- }
449
- return responseServiceTier ?? requestServiceTier;
450
- }
451
467
  function resolveCodexUrl(baseUrl) {
452
468
  const raw = baseUrl && baseUrl.trim().length > 0 ? baseUrl : DEFAULT_CODEX_BASE_URL;
453
469
  const normalized = raw.replace(/\/+$/, "");
@@ -468,12 +484,12 @@ function resolveCodexWebSocketUrl(baseUrl) {
468
484
  // ============================================================================
469
485
  // Response Processing
470
486
  // ============================================================================
471
- async function processStream(response, output, stream, model, grammarToolInputProperties, options, streamDeadline) {
487
+ async function processStream(response, output, stream, model, grammarToolInputProperties, requestServiceTier, options, streamDeadline) {
472
488
  const events = mapCodexEvents(parseSSE(response, streamDeadline.signal), output, model, options?.onProviderStreamEvent);
473
489
  await processResponsesStream(withStreamDeadline(events, streamDeadline.deadlineMs, streamDeadline.abort), output, stream, model, {
474
- serviceTier: resolveCodexRequestServiceTier(model, options?.serviceTier),
490
+ serviceTier: requestServiceTier,
475
491
  grammarToolInputProperties,
476
- resolveServiceTier: resolveCodexServiceTier,
492
+ resolveServiceTier: resolveChatGPTBackendServiceTier,
477
493
  applyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),
478
494
  });
479
495
  }
@@ -515,6 +531,26 @@ function isWebSocketConnectionLimitReachedError(error) {
515
531
  function isPreviousResponseNotFoundError(error) {
516
532
  return error instanceof CodexApiError && error.code === PREVIOUS_RESPONSE_NOT_FOUND_CODE;
517
533
  }
534
+ /**
535
+ * Whether the Codex backend refused the request's `service_tier`: an HTTP 400 or a stream `error` event whose
536
+ * message names the tier, or an error event with `param: "service_tier"`.
537
+ */
538
+ function isCodexServiceTierRejection(error) {
539
+ if (isServiceTierRejection(error))
540
+ return true;
541
+ if (!(error instanceof CodexApiError) || error.payload === undefined)
542
+ return false;
543
+ const nested = error.payload.error;
544
+ const nestedParam = typeof nested === "object" && nested !== null ? nested.param : undefined;
545
+ return error.payload.param === "service_tier" || nestedParam === "service_tier";
546
+ }
547
+ /** Atomic's Codex fast routing names the requested tier in this header; a default-tier retry must not carry it. */
548
+ const CODEX_ROUTING_HINT_HEADER = "x-codex-routing-hint";
549
+ function withoutCodexRoutingHint(headers) {
550
+ const defaultTierHeaders = new Headers(headers);
551
+ defaultTierHeaders.delete(CODEX_ROUTING_HINT_HEADER);
552
+ return defaultTierHeaders;
553
+ }
518
554
  function extractCodexEventError(event) {
519
555
  const nested = event.error && typeof event.error === "object" ? event.error : undefined;
520
556
  return {
@@ -1163,7 +1199,7 @@ async function* startWebSocketOutputOnFirstEvent(events, onStart) {
1163
1199
  yield event;
1164
1200
  }
1165
1201
  }
1166
- async function processWebSocketStream(url, body, headers, output, stream, model, onStart, idleTimeoutMs, websocketConnectTimeoutMs, cacheSessionId, accountId, grammarToolInputProperties, options) {
1202
+ async function processWebSocketStream(url, body, headers, output, stream, model, onStart, idleTimeoutMs, websocketConnectTimeoutMs, cacheSessionId, accountId, grammarToolInputProperties, requestServiceTier, options) {
1167
1203
  const { socket, entry, reused, release } = await acquireWebSocket(url, headers, cacheSessionId, accountId, options?.signal, websocketConnectTimeoutMs, options?.env);
1168
1204
  let keepConnection = true;
1169
1205
  const useCachedContext = options?.transport === "websocket-cached" || options?.transport === "auto";
@@ -1197,9 +1233,9 @@ async function processWebSocketStream(url, body, headers, output, stream, model,
1197
1233
  try {
1198
1234
  socket.send(JSON.stringify({ type: "response.create", ...requestBody }));
1199
1235
  await processResponsesStream(startWebSocketOutputOnFirstEvent(mapCodexEvents(parseWebSocket(socket, options?.signal, idleTimeoutMs), output, model, options?.onProviderStreamEvent), onStart), output, stream, model, {
1200
- serviceTier: resolveCodexRequestServiceTier(model, options?.serviceTier),
1236
+ serviceTier: requestServiceTier,
1201
1237
  grammarToolInputProperties,
1202
- resolveServiceTier: resolveCodexServiceTier,
1238
+ resolveServiceTier: resolveChatGPTBackendServiceTier,
1203
1239
  applyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),
1204
1240
  });
1205
1241
  if (options?.signal?.aborted) {