fix(context_engine): preserve reasoning fields in history projection

raven/context_engine/history_trimmer.py

@@ -31,8 +31,18 @@
 
 # Provider-safe message keys. Anything else on a session message
 # (timestamps, internal ids, manifest annotations) is dropped before
-# the dict reaches the LLM.
-_ALLOWED_KEYS = {"role", "content", "tool_calls", "tool_call_id", "name"}
+# the dict reaches the LLM. reasoning_content / thinking_blocks must survive
+# so multi-turn reasoning contracts (e.g. DeepSeek thinking mode) hold; the
+# provider gate strips thinking_blocks for non-Anthropic targets downstream.
+_ALLOWED_KEYS = {
+    "role",
+    "content",
+    "tool_calls",
+    "tool_call_id",
+    "name",
+    "reasoning_content",
+    "thinking_blocks",
+}
 
 
 @dataclass

raven/context_engine/segments/curator.py

@@ -293,7 +293,7 @@ def _system_prompt() -> str:
 
     @staticmethod
     def _history_from_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
-        allowed = {"role", "content", "tool_calls", "tool_call_id", "name"}
+        allowed = {"role", "content", "tool_calls", "tool_call_id", "name", "reasoning_content", "thinking_blocks"}
         out: list[dict[str, Any]] = []
         for message in messages:
             entry = {k: v for k, v in message.items() if k in allowed}

raven/utils/helpers.py

@@ -117,6 +117,11 @@ def estimate_prompt_tokens(
                 parts.append(value)
         if msg.get("tool_calls"):
             parts.append(json.dumps(msg["tool_calls"], ensure_ascii=False))
+        reasoning = msg.get("reasoning_content")
+        if isinstance(reasoning, str) and reasoning:
+            parts.append(reasoning)
+        if msg.get("thinking_blocks"):
+            parts.append(json.dumps(msg["thinking_blocks"], ensure_ascii=False))
 
     if tools:
         parts.append(json.dumps(tools, ensure_ascii=False))
@@ -154,6 +159,11 @@ def estimate_message_tokens(message: dict[str, Any]) -> int:
             parts.append(value)
     if message.get("tool_calls"):
         parts.append(json.dumps(message["tool_calls"], ensure_ascii=False))
+    reasoning = message.get("reasoning_content")
+    if isinstance(reasoning, str) and reasoning:
+        parts.append(reasoning)
+    if message.get("thinking_blocks"):
+        parts.append(json.dumps(message["thinking_blocks"], ensure_ascii=False))
 
     payload = "\n".join(parts)
     if not payload:

tests/test_curator_context_engine.py

@@ -8,6 +8,7 @@
 from raven.agent.loop import AgentLoop
 from raven.config import ContextConfig
 from raven.context_engine import ContextAssembler, TurnContext
+from raven.context_engine.segments.curator import CuratorSegmentBuilder
 from raven.memory_engine.base import TokenBudget
 from raven.providers.base import LLMProvider, LLMResponse, ToolCallRequest
 from raven.spine.message import ChatType, Source
@@ -195,3 +196,20 @@ async def test_process_message_records_main_and_curator_trajectories(tmp_path: P
     trace_text = traces[0].read_text(encoding="utf-8")
     assert "curator_llm_request" in trace_text
     assert "main_agent_result" in trace_text
+
+
+def test_history_from_messages_preserves_reasoning_fields():
+    messages = [
+        {"role": "user", "content": "hi"},
+        {
+            "role": "assistant",
+            "content": "answer",
+            "reasoning_content": "chain of thought",
+            "thinking_blocks": [{"thinking": "block"}],
+        },
+    ]
+
+    history = CuratorSegmentBuilder._history_from_messages(messages)
+
+    assert history[1]["reasoning_content"] == "chain of thought"
+    assert history[1]["thinking_blocks"] == [{"thinking": "block"}]

tests/test_history_trimmer.py

@@ -0,0 +1,30 @@
+from __future__ import annotations
+
+from raven.context_engine.history_trimmer import HistoryTrimmer
+
+
+def test_history_from_ids_preserves_reasoning_fields():
+    messages = [
+        {"role": "user", "content": "hi"},
+        {
+            "role": "assistant",
+            "content": "answer",
+            "reasoning_content": "chain of thought",
+            "thinking_blocks": [{"thinking": "block"}],
+        },
+    ]
+
+    history = HistoryTrimmer.history_from_ids(messages, [0, 1])
+
+    assert history[1]["reasoning_content"] == "chain of thought"
+    assert history[1]["thinking_blocks"] == [{"thinking": "block"}]
+
+
+def test_history_from_ids_drops_non_provider_keys():
+    messages = [
+        {"role": "user", "content": "hi", "timestamp": "2026-07-08T00:00:00"},
+    ]
+
+    history = HistoryTrimmer.history_from_ids(messages, [0])
+
+    assert history == [{"role": "user", "content": "hi"}]

tests/test_token_estimation.py

@@ -45,6 +45,27 @@ def test_estimate_prompt_tokens_counts_assistant_tool_calls(
     assert prompt_tokens > 10_000
 
 
+def test_estimate_counts_reasoning_fields(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(
+        helpers.tiktoken,
+        "get_encoding",
+        lambda _name: _FakeEncoding(),
+    )
+    base = {"role": "assistant", "content": "answer"}
+    with_reasoning = {
+        "role": "assistant",
+        "content": "answer",
+        "reasoning_content": "r" * 5_000,
+        "thinking_blocks": [{"thinking": "t" * 5_000}],
+    }
+
+    assert helpers.estimate_message_tokens(with_reasoning) > helpers.estimate_message_tokens(base)
+    assert helpers.estimate_prompt_tokens([with_reasoning]) > helpers.estimate_prompt_tokens([base])
+    assert helpers.estimate_message_tokens(with_reasoning) > 10_000
+
+
 def test_estimate_prompt_tokens_fallback_counts_tool_calls(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None: