From 975a6806c318976efd1d9adc6ee44aed4d8ff8d4 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 20 Aug 2026 04:11:08 -0700
Subject: [PATCH] test(e2e): request reasoning explicitly on the reasoning-cost
 assertions

The two tests that assert on reasoning cost read reasoning_tokens off the
response and required it to be nonzero, without ever asking the model to
reason. Both now send reasoning_effort, so the assertion rests on a
parameter the test sets rather than on the model's default behavior.

The cache-breakdown test sends it on its prime call too: OpenAI's prefix
cache keys on the reasoning setting as well as the tokens, so priming at
a different effort never produces a read.
---
 .../test_cache_cost_accounting_e2e.py         | 30 +++++++++++++++++--
 .../test_service_tier_pricing_e2e.py          |  7 ++++-
 2 files changed, 33 insertions(+), 4 deletions(-)

diff --git a/tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py b/tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py
index ca5985fcda6..c50ec3d902f 100644
--- a/tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py
+++ b/tests/e2e/quota_management/spend_tracking/test_cache_cost_accounting_e2e.py
@@ -30,6 +30,12 @@ OpenAI caching is best-effort, so each test retries with a fresh prefix (new
 marker = brand-new cache identity) up to three times before failing; the prime and
 measured calls share the prefix but differ in the trailing question, which defeats
 the proxy's own response cache without touching the provider's prefix cache.
+
+The test that asserts on reasoning cost requests reasoning explicitly with
+`reasoning_effort`, so that assertion rests on a parameter the test sets rather
+than on whatever the model happens to do by default. Its prime call carries the
+same value: OpenAI's prefix cache keys on the reasoning setting as well as the
+tokens, so a prime at a different effort never produces a read.
 """
 
 import pytest
@@ -67,6 +73,7 @@ CACHE_WRITE_RATE = 5e-05
 
 PRIME_QUESTION = "Reply with the single word ready."
 REASONING_QUESTION = "Compute 47*83 - 19*7 step by step, then reply with just the final number."
+REASONING_EFFORT = "high"
 
 
 class _StreamChunk(BaseModel):
@@ -84,12 +91,15 @@ def _cache_priced_params(backend: str) -> LiteLLMParamsBody:
     )
 
 
-def _chat_body(model: str, content: str, *, stream: bool = False) -> ChatBody:
+def _chat_body(
+    model: str, content: str, *, stream: bool = False, reasoning_effort: str | None = None
+) -> ChatBody:
     return ChatBody(
         model=model,
         messages=[ChatMessage(role="user", content=content)],
         stream=stream,
         max_completion_tokens=4000,
+        reasoning_effort=reasoning_effort,
     )
 
 
@@ -151,9 +161,23 @@ class TestCacheCostAccounting:
 
         for _ in range(CACHE_ATTEMPTS):
             prefix = cacheable_prefix(unique_marker())
-            unwrap(client.proxy.chat(scoped_key, _chat_body(model, f"{prefix}\n{PRIME_QUESTION}")))
+            unwrap(
+                client.proxy.chat(
+                    scoped_key,
+                    _chat_body(
+                        model, f"{prefix}\n{PRIME_QUESTION}", reasoning_effort=REASONING_EFFORT
+                    ),
+                )
+            )
             chat = unwrap(
-                client.proxy.chat(scoped_key, _chat_body(model, f"{prefix}\n{REASONING_QUESTION}"))
+                client.proxy.chat(
+                    scoped_key,
+                    _chat_body(
+                        model,
+                        f"{prefix}\n{REASONING_QUESTION}",
+                        reasoning_effort=REASONING_EFFORT,
+                    ),
+                )
             )
             assert chat.id, f"chat response carried no id: {chat}"
             row = _require_row(client, chat.id)
diff --git a/tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py b/tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py
index 171c849fb4c..770c5699b4e 100644
--- a/tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py
+++ b/tests/e2e/quota_management/spend_tracking/test_service_tier_pricing_e2e.py
@@ -10,7 +10,9 @@ computed from the wrong tier (or a mix) cannot match the expected numbers. The
 prompt is a fresh unique marker per run, keeping cached tokens out of the math.
 The response's own `service_tier` echo is asserted first: if OpenAI ever declined
 priority processing and served the default tier, the test fails there instead of
-producing a vacuous rate comparison.
+producing a vacuous rate comparison. Reasoning is requested explicitly with
+`reasoning_effort`, so the reasoning-rate assertion rests on a parameter the test
+sets rather than on whatever the model happens to do by default.
 """
 
 import pytest
@@ -38,6 +40,8 @@ OUTPUT_RATE = 8e-05
 PRIORITY_INPUT_RATE = 6e-05
 PRIORITY_OUTPUT_RATE = 1.6e-04
 
+REASONING_EFFORT = "high"
+
 
 class TestServiceTierPricing:
     @pytest.mark.covers("quota_management.spend_tracking.service_tier.bills_tier_rates")
@@ -74,6 +78,7 @@ class TestServiceTierPricing:
                     ],
                     max_completion_tokens=4000,
                     service_tier="priority",
+                    reasoning_effort=REASONING_EFFORT,
                 ),
             )
         )