Skip to content

Use Claude Code, Codex, VSCode, Pi, and OpenCode (and 6 other harnesses) for free (1.3B+ free tokens) from your terminal, app, IDE, or phone, and now from the browser with native browser sessions (multi-harness + multi-model) like OpenClaw (voice supported + ToS friendly)

Python

patch: preserve classifier policy across model fallbacks (#2056)

Ali Khokhar14c9f85c504a
diff --git a/src/free_claude_code/application/reasoning.py b/src/free_claude_code/application/reasoning.py
index 32c76928..f682d9fd 100644
--- a/src/free_claude_code/application/reasoning.py
+++ b/src/free_claude_code/application/reasoning.py
@@ -47,6 +47,9 @@ def resolve_fallback_reasoning(
) -> ReasoningPolicy:
"""Resolve a candidate from original intent, never a previous fallback."""
+ # Classifier requests keep the handler's tolerant policy on every candidate.
+ if original.control is ReasoningControl.PREFER_OFF:
+ return original
if preference is ReasoningPreference.INHERIT:
return original
if preference is ReasoningPreference.CLIENT:
diff --git a/tests/api/test_classifier_provider_compatibility.py b/tests/api/test_classifier_provider_compatibility.py
index e0a065b1..de115972 100644
--- a/tests/api/test_classifier_provider_compatibility.py
+++ b/tests/api/test_classifier_provider_compatibility.py
@@ -12,6 +12,7 @@ from openai import APIError, AsyncOpenAI, BadRequestError
from free_claude_code.api.handlers import MessagesHandler
from free_claude_code.application.model_metadata import ProviderModelInfo
from free_claude_code.config.constants import ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS
+from free_claude_code.config.reasoning import FALLBACK_REASONING_PREFERENCES
from free_claude_code.config.settings import Settings
from free_claude_code.core.anthropic.models import MessagesRequest
from free_claude_code.core.anthropic.stream_contracts import (
@@ -21,6 +22,10 @@ from free_claude_code.core.anthropic.stream_contracts import (
)
from free_claude_code.core.reasoning import ReasoningCapability
from free_claude_code.providers.open_router import OpenRouterProvider
+from tests.api.model_fallback_support import (
+ ControlledFallbackProvider,
+ execution_failure,
+)
from tests.providers.support import (
SDKStreamDouble,
immediate_admission,
@@ -173,18 +178,32 @@ async def test_openrouter_numeric_sse_rejection_uses_classifier_correction(
@pytest.mark.asyncio
@pytest.mark.parametrize("stream", [False, True])
[email protected]("fallback_effort", [None, *FALLBACK_REASONING_PREFERENCES])
@pytest.mark.parametrize(
"reject_off",
[False, True, "stream"],
ids=["off-ignored", "off-rejected", "off-stream-rejected"],
)
-async def test_classifier_mandatory_reasoning_still_returns_verdict(reject_off, stream):
+async def test_classifier_mandatory_reasoning_still_returns_verdict(
+ reject_off, stream, fallback_effort
+):
provider = OpenRouterProvider(
make_provider_config(api_key="test", base_url="https://openrouter.test/v1"),
admission=immediate_admission(),
)
bodies = []
streams = []
+ primary = ControlledFallbackProvider(failure=execution_failure("primary failed"))
+ settings = Settings()
+ request = classifier_request().model_copy(update={"stream": stream})
+ if fallback_effort is not None:
+ settings = Settings(
+ MODEL="nvidia_nim/primary",
+ MODEL_FALLBACKS=[
+ {"model": "open_router/dynamic-route", "reasoning": fallback_effort}
+ ],
+ )
+ request = request.model_copy(update={"model": "nvidia_nim/primary"})
async def upstream(**body):
bodies.append(body)
@@ -227,8 +246,12 @@ async def test_classifier_mandatory_reasoning_still_returns_verdict(reject_off,
try:
handler = MessagesHandler(
- Settings(),
- provider_resolver=AsyncMock(side_effect=lambda _: provider),
+ settings,
+ provider_resolver=AsyncMock(
+ side_effect=lambda provider_id: (
+ primary if provider_id == "nvidia_nim" else provider
+ )
+ ),
web_tools=StubWebToolsClient(),
)
with patch.object(
@@ -236,9 +259,7 @@ async def test_classifier_mandatory_reasoning_still_returns_verdict(reject_off,
"create",
new=AsyncMock(side_effect=upstream),
):
- response = await handler.create(
- classifier_request().model_copy(update={"stream": stream})
- )
+ response = await handler.create(request)
assert isinstance(response, (JSONResponse, StreamingResponse))
assert response.status_code == 200
if stream:
diff --git a/tests/api/test_model_fallback_wire_controls.py b/tests/api/test_model_fallback_wire_controls.py
index 81e3810e..02ac1be7 100644
--- a/tests/api/test_model_fallback_wire_controls.py
+++ b/tests/api/test_model_fallback_wire_controls.py
@@ -3,11 +3,15 @@
import json
from unittest.mock import AsyncMock
+import httpx
import httpx2
import pytest
+from fastapi.responses import JSONResponse
+from free_claude_code.api.handlers import MessagesHandler
from free_claude_code.application.execution import ProviderExecutor
from free_claude_code.application.routing import ModelRouter
+from free_claude_code.config.reasoning import FALLBACK_REASONING_PREFERENCES
from free_claude_code.config.settings import Settings
from free_claude_code.core.anthropic import MessagesRequest
from free_claude_code.core.anthropic.stream_contracts import parse_sse_text
@@ -19,6 +23,7 @@ from tests.api.model_fallback_support import (
ControlledFallbackProvider,
execution_failure,
)
+from tests.api.test_classifier_provider_compatibility import classifier_request
from tests.api.test_delivered_stream_recovery import (
assert_completed,
public_text,
@@ -27,6 +32,58 @@ from tests.api.test_delivered_stream_recovery import (
from tests.api.test_hidden_stream_retries import delivered
from tests.providers.support import immediate_admission
from tests.providers.test_custom_provider import completion, definition, provider
+from tests.web_tools_support import StubWebToolsClient
+
+
[email protected]("effort", FALLBACK_REASONING_PREFERENCES)
+async def test_classifier_fallback_keeps_manual_thinking_compatible(effort):
+ captured = []
+
+ def reply(request):
+ captured.append(json.loads(request.content))
+ return httpx.Response(
+ 200,
+ text=completion("anthropic_messages").replace(
+ "Hello", "<severity>0</severity>"
+ ),
+ headers={"content-type": "text/event-stream"},
+ )
+
+ configuration = definition(
+ api_format="anthropic_messages", reasoning_format="messages_manual"
+ )
+ fallback = provider(configuration, messages_handler=reply)
+ primary = ControlledFallbackProvider(failure=execution_failure("primary failed"))
+ settings = Settings(
+ MODEL="nvidia_nim/primary",
+ custom_providers=(configuration,),
+ MODEL_FALLBACKS=[
+ {"model": f"{configuration.provider_id}/backup", "reasoning": effort}
+ ],
+ )
+ handler = MessagesHandler(
+ settings,
+ provider_resolver=AsyncMock(
+ side_effect=lambda provider_id: (
+ primary if provider_id == "nvidia_nim" else fallback
+ )
+ ),
+ web_tools=StubWebToolsClient(),
+ )
+ try:
+ response = await handler.create(
+ classifier_request().model_copy(update={"model": "nvidia_nim/primary"})
+ )
+ assert isinstance(response, JSONResponse) and response.status_code == 200
+ assert json.loads(bytes(response.body))["content"] == [
+ {"type": "text", "text": "<severity>0</severity>"}
+ ]
+ assert len(captured) == 1
+ assert captured[0]["thinking"] == {"type": "disabled"}
+ assert captured[0]["max_tokens"] > 64
+ finally:
+ await fallback.cleanup()
@pytest.mark.asyncio