fix(providers): address Codex review on PR #67

Two P2 issues caught during PR #67 review (and surfaced by tests that
landed on main after this branch forked).

1. **OpenAI**: stop auto-injecting ``max_tokens`` / ``temperature`` into
   the default kwargs. Two new tests on main
   (``test_payload_quirk_default_no_rewrite`` /
   ``test_payload_quirk_no_max_completion_tokens_noop``) assert that
   the provider does not stamp ``max_tokens`` on the payload — the
   caller is responsible via ``on_payload`` or the
   ``max_completion_tokens_alias`` quirk. The unconditional default
   also caused new-API providers (those requiring
   ``max_completion_tokens``) to reject requests that contained both
   limit fields.

2. **Anthropic**: skip ``temperature`` when extended thinking is on.
   Anthropic disallows temperature modifications with thinking; the
   previous unconditional default made any reasoning-enabled request
   fail. See:
   https://platform.claude.com/docs/en/build-with-claude/extended-thinking#feature-compatibility

All 499 cubepi tests pass; cubebox cubepi+OpenAI-compat E2E still
green.

Author: xfgong

cubepi/providers/anthropic.py

@@ -107,8 +107,14 @@ async def stream(
             "model": model.id,
             "messages": api_messages,
             "max_tokens": max_tokens,
-            "temperature": model.temperature,
         }
+        # Anthropic disallows temperature modifications when extended
+        # thinking is enabled — see
+        # https://platform.claude.com/docs/en/build-with-claude/extended-thinking#feature-compatibility
+        # The provider request will fail with thinking on if we send a
+        # non-default temperature, so skip the field unless thinking is off.
+        if thinking == "off":
+            kwargs["temperature"] = model.temperature
         if system_prompt:
             if cache_control and self._cache_policy.mark_system():
                 kwargs["system"] = [

cubepi/providers/openai.py

@@ -77,8 +77,6 @@ async def stream(
             "model": model.id,
             "messages": api_messages,
             "stream": True,
-            "max_tokens": model.max_tokens,
-            "temperature": model.temperature,
         }
         if tools:
             kwargs["tools"] = [self._convert_tool(t) for t in tools]