Add the retry overlay file and register it in workflow.yaml

Completes the prior commit: 408/429 join 5XX via x-speakeasy-retries overlay.
Verified offline with `speakeasy overlay apply`: statusCodes -> [5XX, 408, 429].

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

.speakeasy/overlays/retry-transient-status-codes.overlay.yaml

@@ -0,0 +1,11 @@
+overlay: 1.0.0
+x-speakeasy-jsonpath: rfc9535
+info:
+  title: Retry request-timeout and rate-limit responses
+  version: 0.0.0
+actions:
+  - target: $["x-speakeasy-retries"].statusCodes
+    description: Also retry 408 (request timeout) and 429 (rate limit), not just 5XX
+    update:
+      - "408"
+      - "429"

.speakeasy/workflow.yaml

@@ -11,6 +11,7 @@ sources:
             - location: .speakeasy/overlays/allof-simplify.overlay.yaml
             - location: .speakeasy/overlays/boolean-query-params.overlay.yaml
             - location: .speakeasy/overlays/fix-nullable-pagination.overlay.yaml
+            - location: .speakeasy/overlays/retry-transient-status-codes.overlay.yaml
         output: .speakeasy/out.openapi.yaml
         registry:
             location: registry.speakeasyapi.dev/openrouter/sdk/open-router-chat-completions-api

src/openrouter/utils/retries.py

@@ -1,7 +1,6 @@
 """Code generated by Speakeasy (https://speakeasy.com). DO NOT EDIT."""
 
 import asyncio
-import json
 import random
 import time
 from datetime import datetime
@@ -151,24 +150,6 @@ def _parse_retry_after_ms_header(response: httpx.Response) -> Optional[int]:
     return None
 
 
-TRANSIENT_INNER_ERROR_CODES = {408, 429, 500, 502, 503, 504, 524, 529}
-
-
-def _wraps_transient_inner_code(response: httpx.Response) -> bool:
-    """OpenRouter (a middleman) sometimes wraps a transient upstream failure in an HTTP
-    400 whose JSON error.code holds the real status; retry those. Ordinary 400
-    invalid-request responses (including inner error.code == 400) stay non-retryable."""
-    try:
-        body = response.json()
-    except (json.JSONDecodeError, httpx.ResponseNotRead):
-        return False
-    error = body.get("error") if isinstance(body, dict) else None
-    code = error.get("code") if isinstance(error, dict) else None
-    if isinstance(code, str) and code.isdigit():
-        code = int(code)
-    return code in TRANSIENT_INNER_ERROR_CODES
-
-
 def _get_sleep_interval(
     exception: Exception,
     initial_interval: int,
@@ -226,12 +207,6 @@ def retry(func, retries: Retries):
 
                         if res.status_code == parsed_code:
                             raise TemporaryError(res)
-
-                if res.status_code == 400:
-                    if not res.is_closed:
-                        res.read()
-                    if _wraps_transient_inner_code(res):
-                        raise TemporaryError(res)
             except (httpx.NetworkError, httpx.TimeoutException) as exception:
                 if retries.config.retry_connection_errors:
                     raise
@@ -277,12 +252,6 @@ async def retry_async(func, retries: Retries):
 
                         if res.status_code == parsed_code:
                             raise TemporaryError(res)
-
-                if res.status_code == 400:
-                    if not res.is_closed:
-                        await res.aread()
-                    if _wraps_transient_inner_code(res):
-                        raise TemporaryError(res)
             except (httpx.NetworkError, httpx.TimeoutException) as exception:
                 if retries.config.retry_connection_errors:
                     raise

tests/test_retries.py

@@ -1,93 +0,0 @@
-import asyncio
-import unittest
-
-import httpx
-
-from openrouter.utils.retries import (
-    BackoffStrategy,
-    RetryConfig,
-    Retries,
-    _wraps_transient_inner_code,
-    retry,
-    retry_async,
-)
-
-
-def response(status_code, json_body=None):
-    request = httpx.Request("POST", "https://openrouter.ai/api/v1/chat/completions")
-    if json_body is None:
-        return httpx.Response(status_code, request=request)
-    return httpx.Response(status_code, json=json_body, request=request)
-
-
-def fast_retries():
-    # Jitter-free, 1ms intervals. Budget is generous so these tests terminate on
-    # success / non-retryable status rather than on the elapsed-time timeout, which
-    # keeps `calls` counts deterministic regardless of CI scheduling overhead.
-    backoff = BackoffStrategy(
-        initial_interval=1, max_interval=1, exponent=1.0, max_elapsed_time=10000, jitter_ms=0
-    )
-    return Retries(RetryConfig("backoff", backoff, retry_connection_errors=True), ["5XX"])
-
-
-class WrapsTransientInnerCodeTests(unittest.TestCase):
-    def test_400_with_transient_inner_code(self):
-        self.assertTrue(_wraps_transient_inner_code(response(400, {"error": {"code": 502}})))
-
-    def test_400_with_string_inner_code(self):
-        self.assertTrue(_wraps_transient_inner_code(response(400, {"error": {"code": "529"}})))
-
-    def test_400_invalid_request_inner_400_not_retried(self):
-        self.assertFalse(_wraps_transient_inner_code(response(400, {"error": {"code": 400}})))
-
-    def test_400_without_inner_code(self):
-        self.assertFalse(_wraps_transient_inner_code(response(400, {"error": {"message": "bad"}})))
-
-    def test_400_non_json_body(self):
-        self.assertFalse(_wraps_transient_inner_code(response(400)))
-
-
-class RetryIntegrationTests(unittest.TestCase):
-    def test_retries_400_wrapping_transient_code_then_succeeds(self):
-        calls = 0
-
-        def func():
-            nonlocal calls
-            calls += 1
-            if calls < 3:
-                return response(400, {"error": {"message": "Provider returned error", "code": 502}})
-            return response(200, {"ok": True})
-
-        result = retry(func, fast_retries())
-        self.assertEqual(result.status_code, 200)
-        self.assertEqual(calls, 3)  # retried twice, then succeeded
-
-    def test_does_not_retry_plain_400(self):
-        calls = 0
-
-        def func():
-            nonlocal calls
-            calls += 1
-            return response(400, {"error": {"message": "logprobs must be 0-5", "code": 400}})
-
-        result = retry(func, fast_retries())
-        self.assertEqual(result.status_code, 400)
-        self.assertEqual(calls, 1)  # not retried
-
-    def test_async_retries_400_wrapping_transient_code_then_succeeds(self):
-        calls = 0
-
-        async def func():
-            nonlocal calls
-            calls += 1
-            if calls < 3:
-                return response(400, {"error": {"code": 503}})
-            return response(200, {"ok": True})
-
-        result = asyncio.run(retry_async(func, fast_retries()))
-        self.assertEqual(result.status_code, 200)
-        self.assertEqual(calls, 3)
-
-
-if __name__ == "__main__":
-    unittest.main()