feat: preserve governed OpenAI tool messages in 1.4.0

Add faithful single-choice chat results, tool stream activity, error-frame retries, and isolated protocol caching. Reject empty cached completions with the same value predicate as transport, add regression coverage, and adopt Apache-2.0.
2026-10-01 14:32:19 -04:00
parent 0e29deac1a
commit 9185eab8c9
42 changed files with 2227 additions and 78 deletions
@@ -15,6 +15,7 @@ from polygateway.types import (
    ChatRequest,
    Effort,
    LLMResponse,
+    OpenAIChatCompletion,
    SourceConfig,
    ThinkingObservation,
 )
@@ -891,3 +892,113 @@ class TestCallStatsNotPoisoned:
        without = ChatRequest(messages=_MSGS)
        assert with_ctx.cache_namespace == without.cache_namespace
        assert digest_messages(with_ctx.messages) == digest_messages(without.messages)
+
+
+class TestOpenAICompletionCache:
+    """协议缓存不得混模式或把字典冒充公共值对象。"""
+
+    def test_mode_isolation_preserves_text_key(self):
+        assert build_cache_key("m", _MSGS, "proj", None) == build_cache_key(
+            "m", _MSGS, "proj", None, response_mode="text"
+        )
+        assert build_cache_key("m", _MSGS, "proj", None) != build_cache_key(
+            "m", _MSGS, "proj", None, response_mode="openai"
+        )
+
+    @pytest.mark.parametrize(
+        "completion",
+        [
+            None,
+            {},
+            {
+                "message": {"content": "ok"},
+                "finish_reason": "stop",
+                "response_id": None,
+                "usage": None,
+            },
+        ],
+    )
+    async def test_missing_or_bad_completion_misses(self, completion):
+        backend = InMemoryCache()
+        mw = CacheMW(backend=backend, model_fingerprint="m", default_namespace="proj", ttl_s=60)
+        raw = json.loads(mw._serialize(_resp()))
+        raw["completion"] = completion
+        assert (
+            mw._rehydrate(json.dumps(raw), ChatRequest(messages=_MSGS, response_mode="openai"))
+            is None
+        )
+
+    async def test_serialization_failure_does_not_fail_call(self):
+        mw = CacheMW(
+            backend=InMemoryCache(), model_fingerprint="m", default_namespace="proj", ttl_s=60
+        )
+        response = _resp(cost=object())
+
+        async def terminal(_):
+            return response
+
+        assert await mw(ChatRequest(messages=_MSGS), terminal) is response
+
+
+@pytest.mark.parametrize(
+    "message",
+    [
+        {"role": "assistant"},
+        {"role": "assistant", "content": None, "tool_calls": []},
+        {"role": "assistant", "content": "", "tool_calls": []},
+        {"role": "assistant", "content": " \n", "tool_calls": []},
+        {"role": "assistant", "reasoning_content": "reason only"},
+    ],
+)
+@pytest.mark.parametrize("cached_content", ["", "stale projection"])
+async def test_openai_cache_without_output_misses_and_refills(message, cached_content):
+    """空协议产物不能靠旧投影正文伪装成功；miss 后回源并覆写。"""
+    backend = InMemoryCache()
+    mw = _mw(backend)
+    request = ChatRequest(messages=_MSGS, response_mode="openai")
+    key = build_cache_key("m", _MSGS, "proj", None, response_mode="openai")
+    bad = _resp(
+        content=cached_content,
+        completion=OpenAIChatCompletion(message, "stop", "bad", None),
+    )
+    await backend.set(key, mw._serialize(bad), ttl_s=60)
+    fresh = _resp(
+        completion=OpenAIChatCompletion(
+            {"role": "assistant", "content": "cached"}, "stop", "fresh", None
+        )
+    )
+    terminal = _Terminal(fresh)
+    result = await mw(request, terminal)
+    assert terminal.calls == 1 and result is fresh and not result.cache_hit
+    hit = await mw(request, terminal)
+    assert terminal.calls == 1 and hit.cache_hit
+    assert hit.completion.response_id == "fresh"
+
+
+async def test_openai_tool_only_cache_hits_without_origin():
+    """有效工具本身是产物，空文本投影不妨碍命中。"""
+    backend = InMemoryCache()
+    mw = _mw(backend)
+    completion = OpenAIChatCompletion(
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": "call-1",
+                    "type": "function",
+                    "function": {"name": "echo", "arguments": "{}"},
+                }
+            ],
+        },
+        "tool_calls",
+        "tool-response",
+        None,
+    )
+    request = ChatRequest(messages=_MSGS, response_mode="openai")
+    key = build_cache_key("m", _MSGS, "proj", None, response_mode="openai")
+    await backend.set(key, mw._serialize(_resp(content="", completion=completion)), ttl_s=60)
+    terminal = _Terminal(_resp())
+    hit = await mw(request, terminal)
+    assert hit.cache_hit and terminal.calls == 0
+    assert hit.completion == completion