feat(reasoning): add inline think tag extraction and Anthropic thinking_blocks support

Add extract_think() and emit_incremental_think() helpers to extract thinking content from inline <think> and <thought> tags in the content field. This handles models served via Ollama, self-hosted vLLM, or other compatible endpoints that embed reasoning as inline tags instead of using the dedicated reasoning_content API field. Also adds Anthropic thinking_blocks support for extended thinking via the thinking content blocks array. Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent) Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
2026-05-12 23:02:59 +08:00
parent 3a27af0018
commit 3a851f8f8d
5 changed files with 283 additions and 4 deletions
@@ -101,6 +101,132 @@ async def test_runner_preserves_reasoning_fields_and_tool_results():
    )


+@pytest.mark.asyncio
+async def test_runner_emits_anthropic_thinking_blocks():
+    from nanobot.agent.hook import AgentHook, AgentHookContext
+    from nanobot.agent.runner import AgentRunSpec, AgentRunner
+
+    provider = MagicMock()
+    emitted_reasoning: list[str] = []
+
+    async def chat_with_retry(**kwargs):
+        return LLMResponse(
+            content="The answer is 42.",
+            thinking_blocks=[
+                {"type": "thinking", "thinking": "Let me analyze this step by step.", "signature": "sig1"},
+                {"type": "thinking", "thinking": "After careful consideration.", "signature": "sig2"},
+            ],
+            tool_calls=[],
+            usage={"prompt_tokens": 5, "completion_tokens": 3},
+        )
+
+    provider.chat_with_retry = chat_with_retry
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+
+    class ReasoningHook(AgentHook):
+        async def emit_reasoning(self, reasoning_content: str | None) -> None:
+            if reasoning_content:
+                emitted_reasoning.append(reasoning_content)
+
+    runner = AgentRunner(provider)
+    result = await runner.run(AgentRunSpec(
+        initial_messages=[{"role": "user", "content": "question"}],
+        tools=tools,
+        model="test-model",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        hook=ReasoningHook(),
+    ))
+
+    assert result.final_content == "The answer is 42."
+    assert len(emitted_reasoning) == 1
+    assert "Let me analyze this" in emitted_reasoning[0]
+    assert "After careful consideration" in emitted_reasoning[0]
+
+
+@pytest.mark.asyncio
+async def test_runner_emits_inline_think_content_as_reasoning():
+    """Models returning <think>...</think> in content should have thinking extracted and emitted."""
+    from nanobot.agent.hook import AgentHook, AgentHookContext
+    from nanobot.agent.runner import AgentRunSpec, AgentRunner
+
+    provider = MagicMock()
+    emitted_reasoning: list[str] = []
+
+    async def chat_with_retry(**kwargs):
+        return LLMResponse(
+            content="<think>Let me think about this...\nThe answer is 42.</think>The answer is 42.",
+            tool_calls=[],
+            usage={"prompt_tokens": 5, "completion_tokens": 3},
+        )
+
+    provider.chat_with_retry = chat_with_retry
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+
+    class ReasoningHook(AgentHook):
+        async def emit_reasoning(self, reasoning_content: str | None) -> None:
+            if reasoning_content:
+                emitted_reasoning.append(reasoning_content)
+
+    runner = AgentRunner(provider)
+    result = await runner.run(AgentRunSpec(
+        initial_messages=[{"role": "user", "content": "what is the answer?"}],
+        tools=tools,
+        model="test-model",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        hook=ReasoningHook(),
+    ))
+
+    assert result.final_content == "The answer is 42."
+    assert len(emitted_reasoning) == 1
+    assert "Let me think about this" in emitted_reasoning[0]
+    assert "The answer is 42" in emitted_reasoning[0]
+
+
+@pytest.mark.asyncio
+async def test_runner_prefers_reasoning_content_over_inline_think():
+    from nanobot.agent.hook import AgentHook, AgentHookContext
+    from nanobot.agent.runner import AgentRunSpec, AgentRunner
+
+    provider = MagicMock()
+    emitted_reasoning: list[str] = []
+
+    async def chat_with_retry(**kwargs):
+        return LLMResponse(
+            content="<think>inline thinking</think>The answer.",
+            reasoning_content="dedicated reasoning field",
+            tool_calls=[],
+            usage={"prompt_tokens": 5, "completion_tokens": 3},
+        )
+
+    provider.chat_with_retry = chat_with_retry
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+
+    class ReasoningHook(AgentHook):
+        async def emit_reasoning(self, reasoning_content: str | None) -> None:
+            if reasoning_content:
+                emitted_reasoning.append(reasoning_content)
+
+    runner = AgentRunner(provider)
+    result = await runner.run(AgentRunSpec(
+        initial_messages=[{"role": "user", "content": "question"}],
+        tools=tools,
+        model="test-model",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        hook=ReasoningHook(),
+    ))
+
+    assert result.final_content == "The answer."
+    # Only the dedicated field should be emitted, not the inline <think> content
+    assert len(emitted_reasoning) == 1
+    assert emitted_reasoning[0] == "dedicated reasoning field"
+
+
@pytest.mark.asyncio
 async def test_runner_calls_hooks_in_order():
    from nanobot.agent.hook import AgentHook, AgentHookContext
@@ -1,4 +1,4 @@
-from nanobot.utils.helpers import strip_think
+from nanobot.utils.helpers import extract_think, strip_think


 class TestStripThinkTag:
@@ -144,3 +144,84 @@ class TestStripThinkConservativePreserve:
    def test_literal_channel_marker_in_code_block_preserved(self):
        text = "Example:\n```\nif line.startswith('<channel|>'):\n    skip()\n```"
        assert strip_think(text) == text
+
+
+class TestExtractThink:
+
+    def test_no_think_tags(self):
+        thinking, clean = extract_think("Hello World")
+        assert thinking is None
+        assert clean == "Hello World"
+
+    def test_single_think_block(self):
+        text = "Hello <think>reasoning content\nhere</think> World"
+        thinking, clean = extract_think(text)
+        assert thinking == "reasoning content\nhere"
+        assert clean == "Hello  World"
+
+    def test_single_thought_block(self):
+        text = "Hello <thought>reasoning content</thought> World"
+        thinking, clean = extract_think(text)
+        assert thinking == "reasoning content"
+        assert clean == "Hello  World"
+
+    def test_multiple_think_blocks(self):
+        text = "A<think>first</think>B<thought>second</thought>C"
+        thinking, clean = extract_think(text)
+        assert thinking == "first\n\nsecond"
+        assert clean == "ABC"
+
+    def test_think_only_no_content(self):
+        text = "<think>just thinking</think>"
+        thinking, clean = extract_think(text)
+        assert thinking == "just thinking"
+        assert clean == ""
+
+    def test_unclosed_think_not_extracted(self):
+        # Unclosed blocks at start are stripped but NOT extracted
+        text = "<think>unclosed thinking..."
+        thinking, clean = extract_think(text)
+        assert thinking is None
+        assert clean == ""
+
+    def test_empty_think_block(self):
+        text = "Hello <think></think> World"
+        thinking, clean = extract_think(text)
+        # Empty blocks result in empty string after strip
+        assert thinking == ""
+        assert clean == "Hello  World"
+
+    def test_think_with_whitespace_only(self):
+        text = "Hello <think>   \n World"
+        thinking, clean = extract_think(text)
+        assert thinking is None
+        assert clean == "Hello <think>   \n World"
+
+    def test_mixed_think_and_thought(self):
+        text = "Start<think>first reasoning</think>middle<thought>second reasoning</thought>End"
+        thinking, clean = extract_think(text)
+        assert thinking == "first reasoning\n\nsecond reasoning"
+        assert clean == "StartmiddleEnd"
+
+    def test_real_world_ollama_response(self):
+        text = """<think>
+The user is asking about Python list comprehensions.
+Let me explain the syntax and give examples.
+</think>
+
+List comprehensions in Python provide a concise way to create lists. Here's the syntax:
+
+```python
+[expression for item in iterable if condition]
+```
+
+For example:
+```python
+squares = [x**2 for x in range(10)]
+```"""
+        thinking, clean = extract_think(text)
+        assert "list comprehensions" in thinking.lower()
+        assert "Let me explain" in thinking
+        assert "List comprehensions in Python" in clean
+        assert "<think>" not in clean
+        assert "</think>" not in clean