fix(anthropic): broaden Kimi thinking-suppression to custom endpoints (#17455)

The guard that drops Anthropic's `thinking` kwarg for Kimi endpoints was
matched on `https://api.kimi.com/coding` only.  Users configuring a
custom Kimi-compatible gateway (or an official Moonshot host) with
`api_mode: anthropic_messages` fall through to the generic third-party
path, which strips thinking blocks AND still sends
`thinking={enabled,...}` → upstream rejects with HTTP 400
"reasoning_content is missing in assistant tool call message at index N"
on the next request after a tool call.

Replace `_is_kimi_coding_endpoint` callers (history replay + thinking
kwarg gate) with `_is_kimi_family_endpoint(base_url, model)` that also
matches the `api.kimi.com` / `moonshot.ai` / `moonshot.cn` hosts and
Kimi/Moonshot family model names (`kimi-`, `moonshot-`, `k1.`, `k2.`,
…) for custom / proxied endpoints.  Keeps the UA-header check in
`build_anthropic_client` URL-only — the `claude-code/0.1.0` header is
an official-Kimi contract.

Plumbs optional `model` through `convert_messages_to_anthropic` so
the unsigned reasoning_content→thinking block synthesised for Kimi's
history validation survives the third-party signature-stripping pass
on custom hosts too.

Closes #17057.

tests/agent/test_kimi_coding_anthropic_thinking.py

@@ -94,13 +94,16 @@ class TestKimiCodingSkipsAnthropicThinking:
         )
         assert "thinking" in kwargs
 
-    def test_kimi_root_endpoint_unaffected(self) -> None:
-        """Only the /coding route is special-cased — plain api.kimi.com is not.
+    def test_kimi_root_endpoint_via_anthropic_transport_omits_thinking(self) -> None:
+        """Plain ``api.kimi.com`` hit via the Anthropic transport also omits thinking.
 
-        ``api.kimi.com`` without ``/coding`` uses the chat_completions transport
-        (see runtime_provider._detect_api_mode_for_url); build_anthropic_kwargs
-        should never see it, but if it somehow does we should not suppress
-        thinking there — that path has different semantics.
+        Auto-detection routes ``api.kimi.com/v1`` to ``chat_completions`` by
+        default, but users can explicitly configure
+        ``api_mode: anthropic_messages`` against any Kimi host.  The upstream
+        validation (reasoning_content required on replayed tool-call
+        messages) is the same regardless of URL path, so the thinking
+        suppression must apply to every Kimi host, not just ``/coding``.
+        See #17057.
         """
         from agent.anthropic_adapter import build_anthropic_kwargs
 
@@ -112,4 +115,98 @@ class TestKimiCodingSkipsAnthropicThinking:
             reasoning_config={"enabled": True, "effort": "medium"},
             base_url="https://api.kimi.com/v1",
         )
+        assert "thinking" not in kwargs
+
+    # ── #17057: custom / proxied Kimi-compatible endpoints ──────────
+    @pytest.mark.parametrize(
+        "base_url,model",
+        [
+            # Custom host with Kimi-family model — the reporter's case
+            ("http://my-kimi-proxy.internal", "kimi-2.6"),
+            ("https://llm.example.com/anthropic", "kimi-k2.5"),
+            ("https://llm.example.com/anthropic", "moonshot-v1-8k"),
+            ("https://llm.example.com/anthropic", "kimi_thinking"),
+            ("https://llm.example.com/anthropic", "moonshotai/kimi-k2.5"),
+            # Official Moonshot host (previously uncovered)
+            ("https://api.moonshot.ai/anthropic", "moonshot-v1-32k"),
+            ("https://api.moonshot.cn/anthropic", "moonshot-v1-32k"),
+        ],
+    )
+    def test_kimi_family_custom_endpoint_omits_thinking(
+        self, base_url: str, model: str
+    ) -> None:
+        """Custom / proxied Kimi endpoints must also strip Anthropic thinking."""
+        from agent.anthropic_adapter import build_anthropic_kwargs
+
+        kwargs = build_anthropic_kwargs(
+            model=model,
+            messages=[{"role": "user", "content": "hello"}],
+            tools=None,
+            max_tokens=4096,
+            reasoning_config={"enabled": True, "effort": "medium"},
+            base_url=base_url,
+        )
+        assert "thinking" not in kwargs, (
+            f"Kimi-family endpoint ({base_url}, {model}) must not receive "
+            f"Anthropic thinking — upstream validates reasoning_content on "
+            f"replayed tool-call history we don't preserve."
+        )
+        assert "output_config" not in kwargs
+
+    def test_custom_endpoint_non_kimi_model_keeps_thinking(self) -> None:
+        """Custom endpoint with a non-Kimi model must keep thinking intact.
+
+        Guards against over-broad model-family matching — only model names
+        starting with a Kimi/Moonshot prefix should trigger suppression.
+        """
+        from agent.anthropic_adapter import build_anthropic_kwargs
+
+        kwargs = build_anthropic_kwargs(
+            model="MiniMax-M2.7",
+            messages=[{"role": "user", "content": "hello"}],
+            tools=None,
+            max_tokens=4096,
+            reasoning_config={"enabled": True, "effort": "medium"},
+            base_url="https://my-llm-proxy.example.com/anthropic",
+        )
         assert "thinking" in kwargs
+        assert kwargs["thinking"]["type"] == "enabled"
+
+    def test_kimi_family_replay_preserves_unsigned_thinking(self) -> None:
+        """On a custom Kimi endpoint, unsigned reasoning_content thinking
+        blocks must survive the third-party signature-stripping pass so
+        the upstream's message-history validation passes.
+        """
+        from agent.anthropic_adapter import convert_messages_to_anthropic
+
+        messages = [
+            {"role": "user", "content": "hi"},
+            {
+                "role": "assistant",
+                "reasoning_content": "planning the tool call",
+                "tool_calls": [
+                    {
+                        "id": "call_1",
+                        "type": "function",
+                        "function": {"name": "skill_view", "arguments": "{}"},
+                    }
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "ok"},
+        ]
+        _, converted = convert_messages_to_anthropic(
+            messages,
+            base_url="http://my-kimi-proxy.internal",
+            model="kimi-2.6",
+        )
+        # The assistant message still carries the unsigned thinking block
+        # synthesised from reasoning_content (required by Kimi's history
+        # validation).  A plain third-party endpoint would have stripped it.
+        assistant_msg = next(m for m in converted if m["role"] == "assistant")
+        assistant_blocks = assistant_msg["content"]
+        thinking_blocks = [
+            b for b in assistant_blocks
+            if isinstance(b, dict) and b.get("type") == "thinking"
+        ]
+        assert len(thinking_blocks) == 1
+        assert thinking_blocks[0]["thinking"] == "planning the tool call"