File size: 15,435 Bytes
a9e27c0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
"""Claude Auto mode exercises the real handler/provider request boundary."""

import json
from types import SimpleNamespace
from unittest.mock import AsyncMock, patch

import httpx as native_httpx
import httpx2 as httpx
import pytest
from fastapi.responses import JSONResponse, StreamingResponse
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 (
    parse_sse_text,
    text_content,
    thinking_content,
)
from free_claude_code.core.reasoning import ReasoningCapability
from free_claude_code.providers.gemini import GeminiProvider
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,
    make_provider_config,
    sdk_chat_chunk,
)
from tests.providers.test_gemini import _sse
from tests.web_tools_support import StubWebToolsClient


class ClassifierStream(SDKStreamDouble):
    def __init__(self, *, starved: bool = False, error: Exception | None = None):
        self.starved = starved
        self.error = error
        self.closed = False
        super().__init__(self._iterate(), close=self.aclose)

    async def _iterate(self):
        if self.error is not None:
            raise self.error
        for text, reasoning, finish in (
            (None, "Internal provider analysis", None),
            (None if self.starved else "<severity>0</severity>", None, None),
            (None, None, "length" if self.starved else "stop"),
        ):
            yield sdk_chat_chunk(
                SimpleNamespace(
                    choices=[
                        SimpleNamespace(
                            delta=SimpleNamespace(
                                content=text,
                                reasoning=reasoning,
                                reasoning_content=reasoning,
                                tool_calls=None,
                            ),
                            finish_reason=finish,
                        )
                    ],
                    usage=None,
                )
            )

    async def aclose(self):
        self.closed = True


def classifier_request():
    return MessagesRequest.model_validate(
        {
            "model": "open_router/dynamic-route",
            "system": "Output <severity>N</severity> for this command's safety.",
            "messages": [
                {
                    "role": "user",
                    "content": "<transcript>Read the requested file.</transcript>",
                }
            ],
            "max_tokens": 64,
            "thinking": {"type": "disabled"},
            "stop_sequences": ["</severity>"],
            "stream": False,
        }
    )


@pytest.mark.asyncio
@pytest.mark.parametrize(
    ("capability", "off", "limit"),
    [
        ("none", False, 64),
        ("optional", True, 64),
        ("required", False, None),
        ("unknown", True, None),
    ],
)
@pytest.mark.parametrize("rejection", [None, "mandatory", "field", "unrelated"])
async def test_native_gemini_classifier_preserves_reasoning_preparation(
    capability, off, limit, rejection
):
    bodies = []

    def reply(request):
        body = json.loads(request.content)
        bodies.append(body)
        config = body["generationConfig"]
        if config["thinkingConfig"].get("thinkingBudget") == 0 and rejection:
            error = {
                "mandatory": {"message": "Thinking cannot be disabled"},
                "field": {
                    "message": "Value is not supported",
                    "param": "generationConfig.thinkingConfig.thinkingBudget",
                },
                "unrelated": {"message": "Invalid tool schema", "param": "tools"},
            }[rejection]
            return native_httpx.Response(400, json={"error": error})
        return _sse(
            [
                {"text": "Private classifier reasoning", "thought": True},
                {"text": "<severity>0</severity>"},
            ]
        )

    provider = GeminiProvider(
        make_provider_config("test", "https://gemini.invalid/v1beta"),
        admission=immediate_admission(),
        transport=native_httpx.MockTransport(reply),
    )
    try:
        response = await MessagesHandler(
            Settings(),
            provider_resolver=AsyncMock(return_value=provider),
            model_info_lookup=lambda *_: ProviderModelInfo(
                "models/gemini-2.5-flash",
                reasoning_capability=ReasoningCapability(capability),
            ),
            web_tools=StubWebToolsClient(),
        ).create(
            classifier_request().model_copy(
                update={"model": "gemini/models/gemini-2.5-flash"}
            )
        )
        assert isinstance(response, JSONResponse)
        if off and rejection == "unrelated":
            assert response.status_code >= 400
            assert len(bodies) == 1
        else:
            assert response.status_code == 200
            assert json.loads(bytes(response.body))["content"] == [
                {"type": "text", "text": "<severity>0</severity>"}
            ]
            assert len(bodies) == (2 if off and rejection else 1)
        config = bodies[0]["generationConfig"]
        assert config.get("maxOutputTokens") == limit
        assert (config["thinkingConfig"].get("thinkingBudget") == 0) is off
        if len(bodies) == 2:
            retried = bodies[1]["generationConfig"]
            assert "maxOutputTokens" not in retried
            assert "thinkingBudget" not in retried["thinkingConfig"]
    finally:
        await provider.cleanup()


@pytest.mark.asyncio
@pytest.mark.parametrize(
    ("error", "corrects"),
    [
        ({"code": 400, "message": "Reasoning cannot be disabled."}, True),
        (
            {
                "code": 400,
                "message": "Reasoning cannot be disabled.",
                "metadata": {"error_type": "invalid_request"},
            },
            True,
        ),
        ({"code": 400, "message": "Messages contain an invalid role."}, False),
        (
            {"code": 400, "param": "tools", "message": "Reasoning is required."},
            False,
        ),
        ({"code": 403, "message": "Reasoning cannot be disabled."}, False),
    ],
    ids=["numeric", "metadata", "unrelated", "other-field", "permission"],
)
async def test_openrouter_numeric_sse_rejection_uses_classifier_correction(
    error, corrects
):
    bodies = []

    def upstream(request):
        bodies.append(json.loads(request.content))
        chunk = {
            "id": "gen-test",
            "object": "chat.completion.chunk",
            "created": 1,
            "model": "dynamic-route",
            "provider": "test",
            "choices": [
                {
                    "index": 0,
                    "delta": {"content": "<severity>0</severity>"},
                    "finish_reason": "stop",
                }
            ],
        }
        if len(bodies) == 1:
            chunk["error"] = error
            chunk["choices"] = [
                {"index": 0, "delta": {"content": ""}, "finish_reason": "error"}
            ]
        return httpx.Response(
            200,
            headers={"content-type": "text/event-stream"},
            text=f"data: {json.dumps(chunk)}\n\ndata: [DONE]\n\n",
        )

    with patch(
        "free_claude_code.providers.openai_chat.client.AsyncOpenAI",
        return_value=AsyncOpenAI(
            api_key="test",
            base_url="https://provider.invalid/v1",
            max_retries=0,
            http_client=httpx.AsyncClient(transport=httpx.MockTransport(upstream)),
        ),
    ):
        provider = OpenRouterProvider(
            make_provider_config("test", "https://provider.invalid/v1"),
            admission=immediate_admission(),
        )
    try:
        response = await MessagesHandler(
            Settings(),
            provider_resolver=AsyncMock(
                side_effect=lambda _, _api_key_number: provider
            ),
            web_tools=StubWebToolsClient(),
        ).create(classifier_request())
        assert isinstance(response, JSONResponse)
        assert len(bodies) == (2 if corrects else 1)
        if corrects:
            assert response.status_code == 200
            assert json.loads(bytes(response.body))["content"] == [
                {"type": "text", "text": "<severity>0</severity>"}
            ]
            assert bodies[0]["reasoning"] == {"enabled": False}
            assert "reasoning" not in bodies[1]
            assert bodies[1]["max_tokens"] == 81920
        else:
            assert response.status_code >= 400
    finally:
        await provider.cleanup()


@pytest.mark.asyncio
@pytest.mark.parametrize("stream", [False, True])
@pytest.mark.parametrize("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, 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)
        if (
            reject_off
            and body.get("extra_body", {}).get("reasoning", {}).get("enabled") is False
        ):
            error = {
                "error": {
                    "code": "invalid_request_error",
                    "message": "Reasoning is mandatory for this endpoint and cannot be disabled.",
                }
            }
            if reject_off == "stream":
                rejected = ClassifierStream(
                    error=APIError(
                        error["error"]["message"],
                        request=httpx.Request(
                            "POST", "https://openrouter.test/v1/chat/completions"
                        ),
                        body=error["error"],
                    )
                )
                streams.append(rejected)
                return rejected
            raise BadRequestError(
                error["error"]["message"],
                response=httpx.Response(
                    400,
                    request=httpx.Request(
                        "POST", "https://openrouter.test/v1/chat/completions"
                    ),
                    json=error,
                ),
                body=error,
            )
        result = ClassifierStream(starved=body.get("max_tokens", 0) < 2048)
        streams.append(result)
        return result

    try:
        handler = MessagesHandler(
            settings,
            provider_resolver=AsyncMock(
                side_effect=lambda provider_id, _api_key_number: (
                    primary if provider_id == "nvidia_nim" else provider
                )
            ),
            web_tools=StubWebToolsClient(),
        )
        with patch.object(
            provider._client.chat.completions,
            "create",
            new=AsyncMock(side_effect=upstream),
        ):
            response = await handler.create(request)
            assert isinstance(response, (JSONResponse, StreamingResponse))
            assert response.status_code == 200
            if stream:
                assert isinstance(response, StreamingResponse)
                raw = "".join([str(chunk) async for chunk in response.body_iterator])
                events = parse_sse_text(raw)
                assert text_content(events) == "<severity>0</severity>"
                assert thinking_content(events) == ""
                assert events[-1].event == "message_stop"
            else:
                assert isinstance(response, JSONResponse)
                message = json.loads(bytes(response.body))
                assert message["content"] == [
                    {"type": "text", "text": "<severity>0</severity>"}
                ]
                assert message["stop_reason"] == "end_turn"
        assert bodies[0]["max_tokens"] == ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS
        assert len(bodies) == (2 if reject_off else 1)
        assert all(stream.closed for stream in streams)
    finally:
        await provider.cleanup()


@pytest.mark.asyncio
@pytest.mark.parametrize(
    ("capability", "off", "limit"),
    [
        ("none", False, 64),
        ("optional", True, 64),
        ("required", False, 81920),
        ("unknown", True, 81920),
    ],
)
async def test_handler_uses_cached_capabilities_for_the_selected_provider(
    capability, off, limit
):
    provider = OpenRouterProvider(
        make_provider_config("test", "https://provider.invalid/v1"),
        admission=immediate_admission(),
    )
    bodies = []

    async def upstream(**body):
        bodies.append(body)
        return ClassifierStream()

    try:
        handler = MessagesHandler(
            Settings(),
            provider_resolver=AsyncMock(
                side_effect=lambda _, _api_key_number: provider
            ),
            model_info_lookup=lambda provider_id, model_id: next(
                (
                    info
                    for info in (
                        ProviderModelInfo(
                            "open_router/dynamic-route",
                            reasoning_capability=ReasoningCapability(capability),
                        ),
                    )
                    if info.model_id == f"{provider_id}/{model_id}"
                ),
                None,
            ),
            web_tools=StubWebToolsClient(),
        )
        with patch.object(
            provider._client.chat.completions,
            "create",
            new=AsyncMock(side_effect=upstream),
        ):
            response = await handler.create(classifier_request())
        assert isinstance(response, JSONResponse) and response.status_code == 200
        assert json.loads(bytes(response.body))["content"] == [
            {"type": "text", "text": "<severity>0</severity>"}
        ]
        assert (
            bodies[0].get("extra_body", {}).get("reasoning", {}).get("enabled") is False
        ) is off
        assert bodies[0]["max_tokens"] == limit
    finally:
        await provider.cleanup()