Fix DeepSeek cache usage accounting (#937)

ARCHITECTURE.md

@@ -352,7 +352,10 @@ Native Anthropic providers call the native request policy for raw request
 dumping, default tokens, stream flags, thinking payloads, and `extra_body`
 handling. Concrete provider packages keep only true upstream quirks such as
 Gemini thought signatures, NIM tool-schema aliases and retry downgrades, or
-DeepSeek attachment/tool/thinking compatibility.
+DeepSeek attachment/tool/thinking compatibility. DeepSeek intentionally uses its
+OpenAI-compatible Chat Completions endpoint because that is the endpoint that
+reports prompt-cache hit/miss counters; the provider maps those counters back
+into Anthropic usage fields for Claude-compatible clients.
 
 Shared provider responsibilities include upstream rate limiting, model listing,
 safe error mapping, transport cleanup, thinking/tool handling, retry or recovery

README.md

@@ -197,7 +197,7 @@ Get a key at [platform.deepseek.com/api_keys](https://platform.deepseek.com/api_
 
 In the Admin UI, paste it into `DEEPSEEK_API_KEY`, then set `MODEL` to a DeepSeek slug such as `deepseek/deepseek-chat`.
 
-This provider uses DeepSeek's Anthropic-compatible endpoint, not the OpenAI chat-completions endpoint.
+FCC uses DeepSeek's OpenAI-compatible Chat Completions endpoint so DeepSeek's prompt-cache hit/miss counters can be mapped into Claude-compatible usage metadata.
 
 ### 5. [Mistral La Plateforme](https://console.mistral.ai/)
 
@@ -570,8 +570,8 @@ Important pieces:
 - Responses requests convert to Anthropic Messages internally, then share the same model router, normalizer, and provider adapters.
 - `fcc-codex` registers a custom `fcc` provider that points Codex at the local proxy's `/v1/responses` endpoint.
 - Model routing resolves Claude model names to `MODEL_OPUS`, `MODEL_SONNET`, `MODEL_HAIKU`, or `MODEL`.
-- NIM, OpenCode Zen, and OpenCode Go use OpenAI chat streaming translated into Anthropic SSE.
-- Wafer, OpenRouter, DeepSeek, Kimi, Fireworks AI, Cloudflare, Z.ai, LM Studio, llama.cpp, and Ollama use Anthropic Messages style transports where applicable (with provider-specific quirks and model-list URLs).
+- NIM, DeepSeek, OpenCode Zen, and OpenCode Go use OpenAI chat streaming translated into Anthropic SSE.
+- Wafer, OpenRouter, Kimi, Fireworks AI, Cloudflare, Z.ai, LM Studio, llama.cpp, and Ollama use Anthropic Messages style transports where applicable (with provider-specific quirks and model-list URLs).
 - The proxy normalizes thinking blocks, tool calls, token usage metadata, and provider errors into the shape each client expects.
 - Request optimizations answer trivial Claude Code probes locally to save latency and quota.
 

config/provider_catalog.py

@@ -16,10 +16,8 @@ NVIDIA_NIM_DEFAULT_BASE = "https://integrate.api.nvidia.com/v1"
 # Moonshot Kimi Anthropic-compatible Messages API (POST …/messages).
 KIMI_DEFAULT_BASE = "https://api.moonshot.ai/anthropic/v1"
 WAFER_DEFAULT_BASE = "https://pass.wafer.ai/v1"
-# DeepSeek Anthropic-compatible Messages API (not OpenAI ``/v1`` chat completions).
-DEEPSEEK_ANTHROPIC_DEFAULT_BASE = "https://api.deepseek.com/anthropic"
-# Historical export name: DeepSeek upstream is the native Anthropic path above.
-DEEPSEEK_DEFAULT_BASE = DEEPSEEK_ANTHROPIC_DEFAULT_BASE
+# DeepSeek Chat Completions API; cache usage is reported on this endpoint.
+DEEPSEEK_DEFAULT_BASE = "https://api.deepseek.com"
 FIREWORKS_DEFAULT_BASE = "https://api.fireworks.ai/inference/v1"
 # Cloudflare account-scoped AI REST root; provider appends /accounts/{id}/ai/v1.
 CLOUDFLARE_AI_REST_ROOT = "https://api.cloudflare.com/client/v4"
@@ -94,12 +92,12 @@ PROVIDER_CATALOG: dict[str, ProviderDescriptor] = {
     "deepseek": ProviderDescriptor(
         provider_id="deepseek",
         display_name="DeepSeek",
-        transport_type="anthropic_messages",
+        transport_type="openai_chat",
         credential_env="DEEPSEEK_API_KEY",
         credential_url="https://platform.deepseek.com/api_keys",
         credential_attr="deepseek_api_key",
-        default_base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
-        capabilities=("chat", "streaming", "tools", "thinking", "native_anthropic"),
+        default_base_url=DEEPSEEK_DEFAULT_BASE,
+        capabilities=("chat", "streaming", "tools", "thinking", "rate_limit"),
     ),
     "mistral": ProviderDescriptor(
         provider_id="mistral",

core/anthropic/streaming/ledger.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 import hashlib
 import json
 import uuid
-from collections.abc import Iterator
+from collections.abc import Iterator, Mapping
 from contextlib import suppress
 from dataclasses import dataclass, field
 from typing import Any
@@ -207,16 +207,31 @@ class AnthropicStreamLedger:
             },
         )
 
-    def message_delta(self, stop_reason: str, output_tokens: int | None) -> str:
+    def message_delta(
+        self,
+        stop_reason: str,
+        output_tokens: int | None,
+        *,
+        usage_fields: Mapping[str, int] | None = None,
+    ) -> str:
         self.stop_reason = stop_reason
         safe_in = _safe_usage_int(self.input_tokens)
         safe_out = output_tokens if isinstance(output_tokens, int) else 0
+        usage = {"input_tokens": safe_in, "output_tokens": safe_out}
+        if usage_fields:
+            usage.update(
+                {
+                    key: value
+                    for key, value in usage_fields.items()
+                    if isinstance(value, int)
+                }
+            )
         return self._emitter.event(
             "message_delta",
             {
                 "type": "message_delta",
                 "delta": {"stop_reason": stop_reason, "stop_sequence": None},
-                "usage": {"input_tokens": safe_in, "output_tokens": safe_out},
+                "usage": usage,
             },
         )
 

providers/deepseek/__init__.py

@@ -1,11 +1,10 @@
 """DeepSeek provider exports."""
 
-from providers.defaults import DEEPSEEK_ANTHROPIC_DEFAULT_BASE, DEEPSEEK_DEFAULT_BASE
+from providers.defaults import DEEPSEEK_DEFAULT_BASE
 
 from .client import DeepSeekProvider
 
 __all__ = [
-    "DEEPSEEK_ANTHROPIC_DEFAULT_BASE",
     "DEEPSEEK_DEFAULT_BASE",
     "DeepSeekProvider",
 ]

providers/deepseek/client.py

@@ -1,26 +1,26 @@
-"""DeepSeek provider implementation (native Anthropic-compatible Messages)."""
+"""DeepSeek provider implementation (OpenAI-compatible Chat Completions)."""
 
 from __future__ import annotations
 
+from collections.abc import Mapping
 from typing import Any
 
-import httpx
-
 from providers.base import ProviderConfig
-from providers.defaults import DEEPSEEK_ANTHROPIC_DEFAULT_BASE
-from providers.transports.anthropic_messages import AnthropicMessagesTransport
+from providers.defaults import DEEPSEEK_DEFAULT_BASE
+from providers.transports.openai_chat import OpenAIChatTransport
 
 from .compat import build_deepseek_request_body
 
 
-class DeepSeekProvider(AnthropicMessagesTransport):
-    """DeepSeek using ``https://api.deepseek.com/anthropic`` (Anthropic Messages API)."""
+class DeepSeekProvider(OpenAIChatTransport):
+    """DeepSeek using ``https://api.deepseek.com`` Chat Completions."""
 
     def __init__(self, config: ProviderConfig):
         super().__init__(
             config,
             provider_name="DEEPSEEK",
-            default_base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+            base_url=config.base_url or DEEPSEEK_DEFAULT_BASE,
+            api_key=config.api_key,
         )
 
     def _build_request_body(
@@ -31,21 +31,26 @@ class DeepSeekProvider(AnthropicMessagesTransport):
             thinking_enabled=self._is_thinking_enabled(request, thinking_enabled),
         )
 
-    def _request_headers(self) -> dict[str, str]:
-        return {
-            "Accept": "text/event-stream",
-            "Content-Type": "application/json",
-            "x-api-key": self._api_key,
-        }
+    def _anthropic_usage_fields(self, usage_info: Any) -> dict[str, int]:
+        usage_fields: dict[str, int] = {}
+        cache_hit_tokens = _usage_int(usage_info, "prompt_cache_hit_tokens")
+        if cache_hit_tokens is not None:
+            usage_fields["cache_read_input_tokens"] = cache_hit_tokens
+        cache_miss_tokens = _usage_int(usage_info, "prompt_cache_miss_tokens")
+        if cache_miss_tokens is not None:
+            usage_fields["cache_creation_input_tokens"] = cache_miss_tokens
+        return usage_fields
 
-    async def _send_model_list_request(self) -> httpx.Response:
-        """DeepSeek lists models from the OpenAI-format root, not /anthropic."""
-        url = str(
-            httpx.URL(self._base_url).copy_with(
-                path="/models", query=None, fragment=None
-            )
-        )
-        return await self._client.get(url, headers=self._model_list_headers())
 
-    def _model_list_headers(self) -> dict[str, str]:
-        return {"Authorization": f"Bearer {self._api_key}"}
+def _usage_int(usage_info: Any, key: str) -> int | None:
+    if usage_info is None:
+        return None
+    if isinstance(usage_info, Mapping):
+        value = usage_info.get(key)
+    else:
+        value = getattr(usage_info, key, None)
+        if value is None:
+            extra = getattr(usage_info, "model_extra", None)
+            if isinstance(extra, Mapping):
+                value = extra.get(key)
+    return value if isinstance(value, int) else None

providers/deepseek/compat.py

@@ -1,8 +1,9 @@
-"""DeepSeek native Anthropic compatibility request policy."""
+"""DeepSeek Anthropic-to-OpenAI chat request policy."""
 
 from __future__ import annotations
 
 from collections.abc import Mapping
+from types import SimpleNamespace
 from typing import Any
 
 from loguru import logger
@@ -11,6 +12,15 @@ from config.constants import ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS
 from core.anthropic import serialize_tool_result_content
 from core.anthropic.native_messages_request import dump_raw_messages_request
 from providers.exceptions import InvalidRequestError
+from providers.transports.openai_chat import (
+    OpenAIChatRequestPolicy,
+    build_openai_chat_request_body,
+)
+
+_REQUEST_POLICY = OpenAIChatRequestPolicy(
+    provider_name="DEEPSEEK",
+    include_extra_body=True,
+)
 
 _UNSUPPORTED_MESSAGE_BLOCK_TYPES = frozenset(
     {
@@ -29,9 +39,9 @@ _OMITTED_ATTACHMENT_BLOCK = {"type": "text", "text": _OMITTED_ATTACHMENT_TEXT}
 
 
 def build_deepseek_request_body(request_data: Any, *, thinking_enabled: bool) -> dict:
-    """Build a DeepSeek ``/v1/messages`` JSON body (Anthropic format)."""
+    """Build a DeepSeek Chat Completions body from an Anthropic request."""
     logger.debug(
-        "DEEPSEEK_REQUEST: native build model={} msgs={}",
+        "DEEPSEEK_REQUEST: chat build model={} msgs={}",
         getattr(request_data, "model", "?"),
         len(getattr(request_data, "messages", [])),
     )
@@ -39,8 +49,7 @@ def build_deepseek_request_body(request_data: Any, *, thinking_enabled: bool) ->
     data = dump_raw_messages_request(request_data)
     if "messages" in data:
         data["messages"] = _strip_unsupported_attachment_blocks(data["messages"])
-    _validate_deepseek_native_request_dict(data)
-    data.pop("extra_body", None)
+    _validate_deepseek_request_dict(data)
     _downgrade_forced_tool_choice(data)
 
     has_tool_history = _has_tool_history(data)
@@ -74,41 +83,37 @@ def build_deepseek_request_body(request_data: Any, *, thinking_enabled: bool) ->
                 len(data.get("tools", [])),
             )
 
-    thinking_cfg = data.pop("thinking", None)
-    if effective_thinking_enabled and isinstance(thinking_cfg, dict):
-        thinking_payload: dict[str, Any] = {"type": "enabled"}
-        budget_tokens = thinking_cfg.get("budget_tokens")
-        if isinstance(budget_tokens, int):
-            thinking_payload["budget_tokens"] = budget_tokens
-        data["thinking"] = thinking_payload
-
     if "messages" in data:
-        data["messages"] = _strip_reasoning_content_when_native(
-            _normalize_tool_result_content(
-                sanitize_deepseek_messages_for_native(
-                    data["messages"],
-                    thinking_enabled=effective_thinking_enabled,
-                )
+        data["messages"] = _normalize_tool_result_content(
+            sanitize_deepseek_messages_for_openai(
+                data["messages"],
+                thinking_enabled=effective_thinking_enabled,
             )
         )
-    if "max_tokens" not in data or data.get("max_tokens") is None:
-        data["max_tokens"] = ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS
 
-    data["stream"] = True
+    sanitized_request = _request_from_dict(request_data, data)
+    body = build_openai_chat_request_body(
+        sanitized_request,
+        thinking_enabled=effective_thinking_enabled,
+        policy=_REQUEST_POLICY,
+        postprocessors=(_apply_deepseek_chat_extras,),
+    )
+    if "max_tokens" not in body or body.get("max_tokens") is None:
+        body["max_tokens"] = ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS
 
     logger.debug(
         "DEEPSEEK_REQUEST: build done model={} msgs={} tools={}",
-        data.get("model"),
-        len(data.get("messages", [])),
-        len(data.get("tools", [])),
+        body.get("model"),
+        len(body.get("messages", [])),
+        len(body.get("tools", [])),
     )
-    return data
+    return body
 
 
-def sanitize_deepseek_messages_for_native(
+def sanitize_deepseek_messages_for_openai(
     messages: Any, *, thinking_enabled: bool
 ) -> Any:
-    """Filter assistant content for DeepSeek's partial native Anthropic support."""
+    """Filter assistant content before converting to DeepSeek Chat Completions."""
     if not isinstance(messages, list):
         return messages
 
@@ -244,19 +249,17 @@ def _walk_block_list_for_unsupported(blocks: Any, *, where: str) -> None:
             )
 
 
-def _validate_deepseek_native_request_dict(data: dict[str, Any]) -> None:
+def _validate_deepseek_request_dict(data: dict[str, Any]) -> None:
     mcp = data.get("mcp_servers")
     if mcp:
-        raise InvalidRequestError(
-            "DeepSeek native does not support mcp_servers on requests."
-        )
+        raise InvalidRequestError("DeepSeek does not support mcp_servers on requests.")
 
     for tool in data.get("tools") or ():
         if not isinstance(tool, dict):
             continue
         if _is_server_listed_tool(tool):
             raise InvalidRequestError(
-                "DeepSeek native does not support listed Anthropic server tools "
+                "DeepSeek does not support listed Anthropic server tools "
                 "(web_search / web_fetch). Remove them or use a different provider."
             )
 
@@ -401,20 +404,6 @@ def _normalize_tool_result_content(messages: Any) -> Any:
     return normalized
 
 
-def _strip_reasoning_content_when_native(messages: Any) -> Any:
-    if not isinstance(messages, list):
-        return messages
-    out: list[Any] = []
-    for message in messages:
-        if not isinstance(message, dict):
-            out.append(message)
-            continue
-        out.append(
-            {key: value for key, value in message.items() if key != "reasoning_content"}
-        )
-    return out
-
-
 def _downgrade_forced_tool_choice(data: dict[str, Any]) -> None:
     tool_choice = data.get("tool_choice")
     if not isinstance(tool_choice, dict):
@@ -429,3 +418,24 @@ def _downgrade_forced_tool_choice(data: dict[str, Any]) -> None:
         tool_choice["name"],
     )
     data["tool_choice"] = {"type": "auto"}
+
+
+def _request_from_dict(request_data: Any, data: dict[str, Any]) -> Any:
+    validator = getattr(type(request_data), "model_validate", None)
+    if callable(validator):
+        return validator(data)
+    return SimpleNamespace(**data)
+
+
+def _apply_deepseek_chat_extras(
+    body: dict[str, Any], _request_data: Any, thinking_enabled: bool
+) -> None:
+    stream_options = body.setdefault("stream_options", {})
+    if isinstance(stream_options, dict):
+        stream_options["include_usage"] = True
+
+    if not thinking_enabled or body.get("model") == "deepseek-reasoner":
+        return
+    extra_body = body.setdefault("extra_body", {})
+    if isinstance(extra_body, dict):
+        extra_body.setdefault("thinking", {"type": "enabled"})

providers/defaults.py

@@ -4,7 +4,6 @@ from config.provider_catalog import (
     CEREBRAS_DEFAULT_BASE,
     CLOUDFLARE_AI_REST_ROOT,
     CODESTRAL_DEFAULT_BASE,
-    DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
     DEEPSEEK_DEFAULT_BASE,
     GEMINI_DEFAULT_BASE,
     GROQ_DEFAULT_BASE,
@@ -25,7 +24,6 @@ __all__ = (
     "CEREBRAS_DEFAULT_BASE",
     "CLOUDFLARE_AI_REST_ROOT",
     "CODESTRAL_DEFAULT_BASE",
-    "DEEPSEEK_ANTHROPIC_DEFAULT_BASE",
     "DEEPSEEK_DEFAULT_BASE",
     "GEMINI_DEFAULT_BASE",
     "GROQ_DEFAULT_BASE",

providers/transports/openai_chat/stream.py

@@ -380,6 +380,7 @@ class OpenAIChatStreamAdapter:
             ledger.message_delta(
                 ledger.final_stop_reason(map_stop_reason(finish_reason)),
                 output_tokens,
+                usage_fields=self._transport._anthropic_usage_fields(usage_info),
             )
         ):
             yield event

providers/transports/openai_chat/transport.py

@@ -107,6 +107,10 @@ class OpenAIChatTransport(BaseProvider):
         """Return provider-specific per-tool argument aliases for this request."""
         return {}
 
+    def _anthropic_usage_fields(self, usage_info: Any) -> dict[str, int]:
+        """Return provider-specific Anthropic usage fields for final SSE usage."""
+        return {}
+
     async def _create_stream(self, body: dict) -> tuple[Any, dict]:
         """Create a streaming chat completion, optionally retrying once."""
         try:

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "free-claude-code"
-version = "2.4.0"
+version = "2.4.1"
 description = "Middleware between Claude Code CLI (Anthropic API) and NVIDIA NIM"
 readme = "README.md"
 requires-python = ">=3.14.0"

tests/providers/test_deepseek.py

@@ -1,8 +1,9 @@
-"""Tests for DeepSeek native Anthropic Messages provider."""
+"""Tests for DeepSeek OpenAI-compatible Chat Completions provider."""
 
 import logging
 from contextlib import asynccontextmanager
-from unittest.mock import AsyncMock, MagicMock, patch
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
 
 import pytest
 
@@ -13,9 +14,9 @@ from api.models.anthropic import (
     Tool,
 )
 from config.constants import ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS
+from core.anthropic.stream_contracts import parse_sse_text
 from providers.base import ProviderConfig
 from providers.deepseek import (
-    DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
     DEEPSEEK_DEFAULT_BASE,
     DeepSeekProvider,
 )
@@ -26,7 +27,7 @@ from providers.exceptions import InvalidRequestError
 def deepseek_config():
     return ProviderConfig(
         api_key="test_deepseek_key",
-        base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+        base_url=DEEPSEEK_DEFAULT_BASE,
         rate_limit=10,
         rate_window=60,
         enable_thinking=True,
@@ -39,9 +40,7 @@ def mock_rate_limiter():
     async def _slot():
         yield
 
-    with patch(
-        "providers.transports.anthropic_messages.transport.GlobalRateLimiter"
-    ) as mock:
+    with patch("providers.transports.openai_chat.transport.GlobalRateLimiter") as mock:
         instance = mock.get_scoped_instance.return_value
 
         async def _passthrough(fn, *args, **kwargs):
@@ -58,26 +57,18 @@ def deepseek_provider(deepseek_config):
 
 
 def test_default_base_url_alias():
-    assert DEEPSEEK_DEFAULT_BASE == DEEPSEEK_ANTHROPIC_DEFAULT_BASE
-    assert DEEPSEEK_ANTHROPIC_DEFAULT_BASE == "https://api.deepseek.com/anthropic"
+    assert DEEPSEEK_DEFAULT_BASE == "https://api.deepseek.com"
 
 
 def test_init(deepseek_config):
-    with patch("httpx.AsyncClient") as mock_client:
+    with patch("providers.transports.openai_chat.transport.AsyncOpenAI") as mock_client:
         provider = DeepSeekProvider(deepseek_config)
     assert provider._api_key == "test_deepseek_key"
-    assert provider._base_url == "https://api.deepseek.com/anthropic"
+    assert provider._base_url == "https://api.deepseek.com"
     assert mock_client.called
 
 
-def test_request_headers_includes_x_api_key(deepseek_provider):
-    h = deepseek_provider._request_headers()
-    assert h["x-api-key"] == "test_deepseek_key"
-    assert h["Content-Type"] == "application/json"
-    assert h["Accept"] == "text/event-stream"
-
-
-def test_build_request_body_native_shape(deepseek_provider):
+def test_build_request_body_openai_chat_shape(deepseek_provider):
     request = MessagesRequest(
         model="deepseek-v4-pro",
         max_tokens=100,
@@ -86,14 +77,12 @@ def test_build_request_body_native_shape(deepseek_provider):
     )
     body = deepseek_provider._build_request_body(request)
     assert body["model"] == "deepseek-v4-pro"
-    assert body["stream"] is True
-    assert body["messages"][0]["role"] == "user"
-    assert body["messages"][0]["content"] in (
-        "Hello",
-        [{"type": "text", "text": "Hello"}],
-    )
-    assert body["system"] == "S"
+    assert "stream" not in body
+    assert body["messages"][0] == {"role": "system", "content": "S"}
+    assert body["messages"][1]["role"] == "user"
+    assert body["messages"][1] == {"role": "user", "content": "Hello"}
     assert body["max_tokens"] == 100
+    assert body["stream_options"] == {"include_usage": True}
 
 
 def test_build_request_body_default_max_tokens(deepseek_provider):
@@ -114,8 +103,7 @@ def test_build_request_body_thinking_enabled(deepseek_provider):
         }
     )
     body = deepseek_provider._build_request_body(request)
-    assert body["thinking"] == {"type": "enabled", "budget_tokens": 2000}
-    assert "extra_body" not in body
+    assert body["extra_body"]["thinking"] == {"type": "enabled"}
 
 
 def test_build_request_body_tool_list_keeps_thinking(deepseek_provider):
@@ -136,8 +124,8 @@ def test_build_request_body_tool_list_keeps_thinking(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    assert body["thinking"] == {"type": "enabled", "budget_tokens": 2000}
-    assert body["tools"][0]["name"] == "Read"
+    assert body["extra_body"]["thinking"] == {"type": "enabled"}
+    assert body["tools"][0]["function"]["name"] == "Read"
 
 
 def test_build_request_body_tool_choice_keeps_thinking(deepseek_provider):
@@ -152,8 +140,8 @@ def test_build_request_body_tool_choice_keeps_thinking(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    assert body["thinking"] == {"type": "enabled", "budget_tokens": 2000}
-    assert body["tool_choice"] == {"type": "auto"}
+    assert body["extra_body"]["thinking"] == {"type": "enabled"}
+    assert body["tool_choice"] == "auto"
 
 
 def test_build_request_body_forced_tool_choice_downgrades_to_auto(
@@ -177,15 +165,15 @@ def test_build_request_body_forced_tool_choice_downgrades_to_auto(
 
     body = deepseek_provider._build_request_body(request)
 
-    assert body["thinking"] == {"type": "enabled", "budget_tokens": 2000}
-    assert body["tool_choice"] == {"type": "auto"}
+    assert body["extra_body"]["thinking"] == {"type": "enabled"}
+    assert body["tool_choice"] == "auto"
 
 
 def test_build_request_body_respects_global_thinking_disable():
     provider = DeepSeekProvider(
         ProviderConfig(
             api_key="k",
-            base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+            base_url=DEEPSEEK_DEFAULT_BASE,
             rate_limit=1,
             rate_window=1,
             enable_thinking=False,
@@ -199,7 +187,8 @@ def test_build_request_body_respects_global_thinking_disable():
         }
     )
     body = provider._build_request_body(request)
-    assert "thinking" not in body
+    assert "extra_body" not in body
+    assert body["stream_options"] == {"include_usage": True}
 
 
 def test_preserve_unsigned_thinking_when_thinking_on(deepseek_provider):
@@ -222,10 +211,8 @@ def test_preserve_unsigned_thinking_when_thinking_on(deepseek_provider):
         }
     )
     body = deepseek_provider._build_request_body(request)
-    blocks = body["messages"][0]["content"]
-    assert len(blocks) == 2
-    assert blocks[0]["type"] == "thinking"
-    assert blocks[0]["thinking"] == "plain"
+    assert body["messages"][0]["content"] == "out"
+    assert body["messages"][0]["reasoning_content"] == "plain"
 
 
 def test_strip_redacted_thinking_when_thinking_on(deepseek_provider):
@@ -244,9 +231,7 @@ def test_strip_redacted_thinking_when_thinking_on(deepseek_provider):
         }
     )
     body = deepseek_provider._build_request_body(request)
-    types = {b["type"] for b in body["messages"][0]["content"]}
-    assert "redacted_thinking" not in types
-    assert "text" in types
+    assert body["messages"][0] == {"role": "assistant", "content": "out"}
 
 
 def test_tool_history_with_replayable_thinking_preserves_thinking(deepseek_provider):
@@ -292,17 +277,19 @@ def test_tool_history_with_replayable_thinking_preserves_thinking(deepseek_provi
 
     body = deepseek_provider._build_request_body(request)
 
-    assert body["thinking"] == {"type": "enabled", "budget_tokens": 2000}
-    assert body["context_management"] == {
-        "edits": [{"type": "clear_thinking_20251015", "keep": "all"}]
+    assert body["extra_body"]["thinking"] == {"type": "enabled"}
+    assert "context_management" not in body
+    assert "output_config" not in body
+    assistant = body["messages"][0]
+    assert assistant["content"] == ""
+    assert assistant["reasoning_content"] == "hidden"
+    assert assistant["tool_calls"][0]["function"]["name"] == "Read"
+    assert assistant["tool_calls"][0]["function"]["arguments"] == '{"file_path": "x"}'
+    assert body["messages"][1] == {
+        "role": "tool",
+        "tool_call_id": "t1",
+        "content": "ok",
     }
-    assert body["output_config"] == {"effort": "high"}
-    assistant_blocks = body["messages"][0]["content"]
-    assert [block["type"] for block in assistant_blocks] == ["thinking", "tool_use"]
-    assert assistant_blocks[0]["thinking"] == "hidden"
-    assert assistant_blocks[0]["signature"] == "sig_123"
-    assert assistant_blocks[1]["name"] == "Read"
-    assert body["messages"][1]["content"][0]["type"] == "tool_result"
 
 
 def test_tool_history_with_unsigned_thinking_preserves_thinking(deepseek_provider):
@@ -339,11 +326,8 @@ def test_tool_history_with_unsigned_thinking_preserves_thinking(deepseek_provide
 
     body = deepseek_provider._build_request_body(request)
 
-    assert body["thinking"] == {"type": "enabled"}
-    assert body["messages"][0]["content"][0] == {
-        "type": "thinking",
-        "thinking": "plain",
-    }
+    assert body["extra_body"]["thinking"] == {"type": "enabled"}
+    assert body["messages"][0]["reasoning_content"] == "plain"
 
 
 def test_tool_history_without_thinking_disables_thinking_and_hints(deepseek_provider):
@@ -395,16 +379,13 @@ def test_tool_history_without_thinking_disables_thinking_and_hints(deepseek_prov
 
     body = deepseek_provider._build_request_body(request)
 
-    assert "thinking" not in body
-    assert body["context_management"] == {
-        "edits": [{"type": "other_edit", "keep": "all"}],
-        "other": True,
-    }
-    assert body["output_config"] == {"format": "text"}
-    assert body["tools"][0]["name"] == "Read"
-    assert body["tool_choice"] == {"type": "auto"}
-    assert body["messages"][0]["content"][0]["type"] == "tool_use"
-    assert body["messages"][1]["content"][0]["type"] == "tool_result"
+    assert "extra_body" not in body
+    assert "context_management" not in body
+    assert "output_config" not in body
+    assert body["tools"][0]["function"]["name"] == "Read"
+    assert body["tool_choice"] == "auto"
+    assert body["messages"][0]["tool_calls"][0]["function"]["name"] == "Read"
+    assert body["messages"][1]["role"] == "tool"
 
 
 def test_tool_history_with_empty_thinking_disables_thinking(deepseek_provider):
@@ -441,15 +422,16 @@ def test_tool_history_with_empty_thinking_disables_thinking(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    assert "thinking" not in body
-    assert [block["type"] for block in body["messages"][0]["content"]] == ["tool_use"]
+    assert "extra_body" not in body
+    assert "reasoning_content" not in body["messages"][0]
+    assert body["messages"][0]["tool_calls"][0]["function"]["name"] == "Read"
 
 
 def test_thinking_off_strips_thinking_history():
     provider = DeepSeekProvider(
         ProviderConfig(
             api_key="k",
-            base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+            base_url=DEEPSEEK_DEFAULT_BASE,
             rate_limit=1,
             rate_window=1,
             enable_thinking=False,
@@ -470,8 +452,7 @@ def test_thinking_off_strips_thinking_history():
         }
     )
     body = provider._build_request_body(request)
-    for b in body["messages"][0]["content"]:
-        assert b["type"] != "thinking"
+    assert "reasoning_content" not in body["messages"][0]
     assert "sec" not in str(body["messages"])
 
 
@@ -505,8 +486,8 @@ def test_passthrough_tool_use_and_result(deepseek_provider):
         }
     )
     body = deepseek_provider._build_request_body(request)
-    assert body["messages"][0]["content"][0]["type"] == "tool_use"
-    assert body["messages"][1]["content"][0]["type"] == "tool_result"
+    assert body["messages"][0]["tool_calls"][0]["function"]["name"] == "n"
+    assert body["messages"][1]["role"] == "tool"
 
 
 def test_preflight_strips_user_image():
@@ -532,7 +513,7 @@ def test_preflight_strips_user_image():
     provider = DeepSeekProvider(
         ProviderConfig(
             api_key="k",
-            base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+            base_url=DEEPSEEK_DEFAULT_BASE,
             rate_limit=1,
             rate_window=1,
         )
@@ -541,8 +522,8 @@ def test_preflight_strips_user_image():
     provider.preflight_stream(request, thinking_enabled=True)
     body = provider._build_request_body(request)
     content = body["messages"][0]["content"]
-    block_types = [b["type"] for b in content] if isinstance(content, list) else []
-    assert "image" not in block_types
+    assert "attachment omitted" in content.lower()
+    assert "image or document inputs" in content.lower()
 
 
 def test_preflight_rejects_mcp_servers():
@@ -554,7 +535,7 @@ def test_preflight_rejects_mcp_servers():
     provider = DeepSeekProvider(
         ProviderConfig(
             api_key="k",
-            base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+            base_url=DEEPSEEK_DEFAULT_BASE,
             rate_limit=1,
             rate_window=1,
         )
@@ -572,7 +553,7 @@ def test_preflight_rejects_listed_server_tools_in_tools_list():
     provider = DeepSeekProvider(
         ProviderConfig(
             api_key="k",
-            base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+            base_url=DEEPSEEK_DEFAULT_BASE,
             rate_limit=1,
             rate_window=1,
         )
@@ -608,7 +589,7 @@ def test_preflight_rejects_server_tool_result_blocks():
     provider = DeepSeekProvider(
         ProviderConfig(
             api_key="k",
-            base_url=DEEPSEEK_ANTHROPIC_DEFAULT_BASE,
+            base_url=DEEPSEEK_DEFAULT_BASE,
             rate_limit=1,
             rate_window=1,
         )
@@ -617,8 +598,7 @@ def test_preflight_rejects_server_tool_result_blocks():
         provider.preflight_stream(request)
 
 
-def test_strip_reasoning_content_not_forwarded(deepseek_provider):
-    # ``reasoning_content`` is for OpenAI helpers only, not in native body.
+def test_reasoning_content_replayed_to_openai_chat(deepseek_provider):
     request = MessagesRequest(
         model="m",
         messages=[
@@ -630,39 +610,77 @@ def test_strip_reasoning_content_not_forwarded(deepseek_provider):
         ],
     )
     body = deepseek_provider._build_request_body(request)
-    assert "reasoning_content" not in body["messages"][0]
+    assert body["messages"][0]["reasoning_content"] == "r"
 
 
 @pytest.mark.asyncio
-async def test_stream_uses_post_messages_path(deepseek_provider):
+async def test_stream_uses_chat_completions_and_maps_cache_usage(deepseek_provider):
     request = MessagesRequest(
         model="m",
         messages=[Message(role="user", content="hi")],
     )
-    called: dict[str, str] = {}
 
-    async def fake_send(request, *args, **kwargs):
-        called["path"] = request.url.path
-        mock_resp = MagicMock()
-        mock_resp.status_code = 200
-        mock_resp.is_closed = False
-        mock_resp.raise_for_status = lambda: None
+    async def fake_stream():
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=SimpleNamespace(
+                        content="hello", reasoning_content=None, tool_calls=None
+                    ),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=SimpleNamespace(
+                        content=None, reasoning_content=None, tool_calls=None
+                    ),
+                    finish_reason="stop",
+                )
+            ],
+            usage=None,
+        )
+        yield SimpleNamespace(
+            choices=[],
+            usage=SimpleNamespace(
+                completion_tokens=3,
+                prompt_tokens=30,
+                prompt_cache_hit_tokens=10,
+                prompt_cache_miss_tokens=20,
+            ),
+        )
 
-        async def aiter():
-            if False:  # pragma: no cover
-                yield ""
+    create = AsyncMock(return_value=fake_stream())
+    with patch.object(deepseek_provider._client.chat.completions, "create", create):
+        chunks = [
+            chunk
+            async for chunk in deepseek_provider.stream_response(
+                request, input_tokens=7, request_id="r1"
+            )
+        ]
 
-        mock_resp.aiter_lines = aiter
-        mock_resp.aclose = AsyncMock()
-        return mock_resp
-
-    deepseek_provider._client.send = fake_send
-    _ = [x async for x in deepseek_provider.stream_response(request, request_id="r1")]
-
-    assert called["path"] == "/anthropic/messages"
+    create.assert_awaited_once()
+    await_args = create.await_args
+    assert await_args is not None
+    assert await_args.kwargs["model"] == "m"
+    assert await_args.kwargs["stream"] is True
+    assert await_args.kwargs["stream_options"] == {"include_usage": True}
+    parsed = parse_sse_text("".join(chunks))
+    usage = next(
+        event.data["usage"] for event in parsed if event.event == "message_delta"
+    )
+    assert usage == {
+        "input_tokens": 7,
+        "output_tokens": 3,
+        "cache_read_input_tokens": 10,
+        "cache_creation_input_tokens": 20,
+    }
 
 
-def test_drops_extra_body_from_canonical_request(deepseek_provider):
+def test_preserves_extra_body_for_openai_chat_request(deepseek_provider):
     raw = {
         "model": "m",
         "max_tokens": 3,
@@ -671,7 +689,7 @@ def test_drops_extra_body_from_canonical_request(deepseek_provider):
     }
     r = MessagesRequest.model_validate(raw)
     body = deepseek_provider._build_request_body(r)
-    assert "extra_body" not in body
+    assert body["extra_body"] == {"note": 1, "thinking": {"type": "enabled"}}
 
 
 def test_normalizes_tool_result_content_array_to_string(deepseek_provider):
@@ -710,10 +728,8 @@ def test_normalizes_tool_result_content_array_to_string(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    # Verify tool_result content is now a string
-    user_msg = body["messages"][1]
-    tool_result = user_msg["content"][0]
-    assert tool_result["type"] == "tool_result"
+    tool_result = body["messages"][1]
+    assert tool_result["role"] == "tool"
     assert isinstance(tool_result["content"], str)
     assert "file1.txt" in tool_result["content"]
     assert "file2.txt" in tool_result["content"]
@@ -746,11 +762,11 @@ def test_strips_document_blocks_for_deepseek(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    # Document block should be stripped; tool_result preserved
-    content = body["messages"][0]["content"]
-    block_types = [block["type"] for block in content]
-    assert "document" not in block_types
-    assert "tool_result" in block_types
+    assert body["messages"][0] == {
+        "role": "tool",
+        "tool_call_id": "t1",
+        "content": "PDF text extracted",
+    }
 
 
 def test_strips_image_blocks_for_deepseek(deepseek_provider):
@@ -779,10 +795,7 @@ def test_strips_image_blocks_for_deepseek(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    content = body["messages"][0]["content"]
-    block_types = [block["type"] for block in content]
-    assert "image" not in block_types
-    assert "text" in block_types
+    assert body["messages"][0] == {"role": "user", "content": "describe this"}
 
 
 def test_normalizes_tool_result_content_dict_to_string(deepseek_provider):
@@ -818,10 +831,8 @@ def test_normalizes_tool_result_content_dict_to_string(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    # Verify tool_result content is now a JSON string
-    user_msg = body["messages"][1]
-    tool_result = user_msg["content"][0]
-    assert tool_result["type"] == "tool_result"
+    tool_result = body["messages"][1]
+    assert tool_result["role"] == "tool"
     assert isinstance(tool_result["content"], str)
     assert "status" in tool_result["content"]
     assert "success" in tool_result["content"]
@@ -870,8 +881,8 @@ def test_strips_image_block_inside_tool_result(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    tool_result = body["messages"][1]["content"][0]
-    assert tool_result["type"] == "tool_result"
+    tool_result = body["messages"][1]
+    assert tool_result["role"] == "tool"
     # After stripping + string-normalization, no base64/image marker survives.
     assert isinstance(tool_result["content"], str)
     assert "screenshot saved" in tool_result["content"]
@@ -921,8 +932,8 @@ def test_image_only_tool_result_replaced_with_placeholder(deepseek_provider):
 
     body = deepseek_provider._build_request_body(request)
 
-    tool_result = body["messages"][1]["content"][0]
-    assert tool_result["type"] == "tool_result"
+    tool_result = body["messages"][1]
+    assert tool_result["role"] == "tool"
     assert isinstance(tool_result["content"], str)
     assert tool_result["content"] != ""
     assert "attachment omitted" in tool_result["content"].lower()
@@ -972,8 +983,8 @@ def test_document_only_tool_result_replaced_with_generic_placeholder(
 
     body = deepseek_provider._build_request_body(request)
 
-    tool_result = body["messages"][1]["content"][0]
-    assert tool_result["type"] == "tool_result"
+    tool_result = body["messages"][1]
+    assert tool_result["role"] == "tool"
     assert isinstance(tool_result["content"], str)
     assert "attachment omitted" in tool_result["content"].lower()
     assert "document inputs" in tool_result["content"].lower()
@@ -1006,10 +1017,8 @@ def test_image_only_message_replaced_with_placeholder(deepseek_provider):
     body = deepseek_provider._build_request_body(request)
 
     content = body["messages"][0]["content"]
-    assert len(content) == 1
-    assert content[0]["type"] == "text"
-    assert "attachment omitted" in content[0]["text"].lower()
-    assert "image or document inputs" in content[0]["text"].lower()
+    assert "attachment omitted" in content.lower()
+    assert "image or document inputs" in content.lower()
 
 
 def test_document_only_message_replaced_with_placeholder(deepseek_provider):
@@ -1034,10 +1043,8 @@ def test_document_only_message_replaced_with_placeholder(deepseek_provider):
     body = deepseek_provider._build_request_body(request)
 
     content = body["messages"][0]["content"]
-    assert len(content) == 1
-    assert content[0]["type"] == "text"
-    assert "attachment omitted" in content[0]["text"].lower()
-    assert "document inputs" in content[0]["text"].lower()
+    assert "attachment omitted" in content.lower()
+    assert "document inputs" in content.lower()
 
 
 def test_warns_when_stripping_attachment_blocks(deepseek_provider, caplog):

tests/providers/test_model_validation.py

@@ -94,17 +94,14 @@ async def test_native_openai_compatible_provider_lists_model_ids() -> None:
 async def test_deepseek_lists_models_from_root_endpoint() -> None:
     provider = DeepSeekProvider(ProviderConfig(api_key="deepseek-key"))
     with patch.object(
-        provider._client,
-        "get",
+        provider._client.models,
+        "list",
         new_callable=AsyncMock,
-        return_value=_response(200, {"data": [{"id": "deepseek-chat"}]}),
-    ) as mock_get:
+        return_value=SimpleNamespace(data=[SimpleNamespace(id="deepseek-chat")]),
+    ) as mock_list:
         assert await provider.list_model_ids() == frozenset({"deepseek-chat"})
 
-    mock_get.assert_awaited_once_with(
-        "https://api.deepseek.com/models",
-        headers={"Authorization": "Bearer deepseek-key"},
-    )
+    mock_list.assert_awaited_once_with()
 
 
 @pytest.mark.asyncio

uv.lock

@@ -561,7 +561,7 @@ wheels = [
 
 [[package]]
 name = "free-claude-code"
-version = "2.4.0"
+version = "2.4.1"
 source = { editable = "." }
 dependencies = [
     { name = "aiohttp" },