From fe53ba509625a90fa83a49ba9030a0c123419f5b Mon Sep 17 00:00:00 2001
From: Amir Fathi <amirfathi.me@gmail.com>
Date: Sun, 23 Aug 2026 11:32:06 +0000
Subject: [PATCH] fix: accumulate all text blocks in Anthropic non-streaming
 response

The non-streaming _query loop set llm_response.completion_text on every
text content block instead of appending to it, so only the last text
block in a multi-block Message.content survived; earlier blocks were
silently dropped. The streaming path already accumulates correctly
with final_text +=. A Message can carry more than one text block (the
Citations API and extended-thinking flows both produce this shape).

Adds a regression test that fails on the old overwrite behavior and
passes with the fix.
---
 .../core/provider/sources/anthropic_source.py |  5 +--
 tests/test_anthropic_kimi_code_provider.py    | 34 +++++++++++++++++++
 2 files changed, 37 insertions(+), 2 deletions(-)

diff --git a/astrbot/core/provider/sources/anthropic_source.py b/astrbot/core/provider/sources/anthropic_source.py
index 27cc459622..f4e425e3eb 100644
--- a/astrbot/core/provider/sources/anthropic_source.py
+++ b/astrbot/core/provider/sources/anthropic_source.py
@@ -540,11 +540,11 @@ async def _query(
             )
 
         llm_response = LLMResponse(role="assistant")
+        completion_text_parts: list[str] = []
 
         for content_block in completion.content:
             if content_block.type == "text":
-                completion_text = str(content_block.text).strip()
-                llm_response.completion_text = completion_text
+                completion_text_parts.append(str(content_block.text))
 
             if content_block.type == "thinking":
                 reasoning_content = str(content_block.thinking).strip()
@@ -556,6 +556,7 @@ async def _query(
                 llm_response.tools_call_name.append(content_block.name)
                 llm_response.tools_call_ids.append(content_block.id)
 
+        llm_response.completion_text = "".join(completion_text_parts).strip()
         llm_response.id = completion.id
         llm_response.usage = self._extract_usage(completion.usage)
 
diff --git a/tests/test_anthropic_kimi_code_provider.py b/tests/test_anthropic_kimi_code_provider.py
index 0dc33f58ba..42e8855249 100644
--- a/tests/test_anthropic_kimi_code_provider.py
+++ b/tests/test_anthropic_kimi_code_provider.py
@@ -736,6 +736,40 @@ async def fake_create(**kwargs):
     assert llm_response.usage.output == 0
 
 
+@pytest.mark.asyncio
+async def test_query_accumulates_multiple_text_blocks(monkeypatch):
+    """A Message with more than one text content block (e.g. Citations API
+    output) must concatenate every block, not keep only the last one."""
+    provider = _setup_provider_with_mock_client(monkeypatch)
+
+    class _FakeMessageBlock:
+        def __init__(self, text: str):
+            self.type = "text"
+            self.text = text
+
+    class _FakeMessage:
+        def __init__(self):
+            self.id = "msg_multi_block"
+            self.content = [
+                _FakeMessageBlock("First block: the important part."),
+                _FakeMessageBlock("Second block: a trailing citation note."),
+            ]
+            self.stop_reason = "end_turn"
+            self.usage = None
+
+    async def fake_create(**kwargs):
+        return _FakeMessage()
+
+    monkeypatch.setattr(anthropic_source, "Message", _FakeMessage)
+    provider.client.messages.create = fake_create
+
+    llm_response = await provider.text_chat(prompt="test")
+
+    assert llm_response.completion_text == (
+        "First block: the important part.Second block: a trailing citation note."
+    )
+
+
 @pytest.mark.asyncio
 async def test_tool_choice_auto_converts_to_dict(monkeypatch):
     """tool_choice='auto' 应转换为 {'type': 'auto'}"""