emrg: persist and pass back DeepSeek thinking-mode reasoning

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1008) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1011) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (259: 45 daemon_client + 20 conn-manager + 22 app-commands + 129 renderer smoke + 15 i18n + 8 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -2539,18 +2539,24 @@ async def _run_tool_loop(
  reasoning=full_reasoning,
  )
 
- # Persist assistant message
+ # Persist assistant message (reasoning kept for DeepSeek
+ # thinking-mode pass-back, rant 2026-08-22T17:25:02)
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the
  # LLM context stays coherent when queued messages are
  # injected after this round (mirrors Case 2's assistant
  # tool_calls message).
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round (after the
  # round-top drain) must not end the turn — inject and continue.
@@ -2605,13 +2611,18 @@ async def _run_tool_loop(
  },
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if full_reasoning:
+ # DeepSeek thinking mode: reasoning must be passed back
+ # verbatim on the next round (rant 2026-08-22T17:25:02).
+ assistant_msg["reasoning_content"] = full_reasoning
  messages.append(assistant_msg)
 
  # Persist assistant message WITH embedded tool_calls
  session.append_message({
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  "tool_calls": [
  {
  "id": tc.get("id", ""),
@@ -2734,12 +2745,17 @@ async def _run_tool_loop(
  "type": "message",
  "role": "assistant",
  "content": full_content,
+ **({"reasoning": full_reasoning} if full_reasoning else {}),
  })
 
  # Append the assistant reply to the local messages so the LLM
  # context stays coherent when queued messages are injected
  # after this round.
- messages.append({"role": "assistant", "content": full_content})
+ messages.append({
+ "role": "assistant",
+ "content": full_content,
+ **({"reasoning_content": full_reasoning} if full_reasoning else {}),
+ })
 
  # P1 (rant 21:55:37): messages queued mid-round must not end the
  # turn — inject and continue (injection round does not consume
@@ -3736,6 +3752,10 @@ async def _reflect():
  # IMPORTANT: assistant message with tool_calls must come BEFORE
  # tool result messages (OpenAI/DeepSeek API requirement).
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:
@@ -3848,6 +3868,10 @@ async def _consolidate_session_memories(
  "function": {"name": fn.get("name", ""), "arguments": fn.get("arguments", "")},
  })
  assistant_msg["tool_calls"] = openai_tool_calls
+ if msg.get("reasoning_content") or msg.get("reasoning"):
+ assistant_msg["reasoning_content"] = (
+ msg.get("reasoning_content") or msg.get("reasoning")
+ )
  messages.append(assistant_msg)
 
  for tc in tool_calls:

emrg/session.py

@@ -323,6 +323,12 @@ def get_messages_for_llm(self) -> list[dict]:
  if r.get("type") == "message":
  msg: dict = {"role": r["role"], "content": r.get("content")}
 
+ # DeepSeek thinking mode: assistant reasoning must be passed
+ # back verbatim to the API (rant 2026-08-22T17:25:02). Old
+ # records without the field are skipped naturally.
+ if r.get("role") == "assistant" and r.get("reasoning"):
+ msg["reasoning_content"] = r["reasoning"]
+
  # Check for embedded tool_calls (current format)
  embedded_tc = r.get("tool_calls")
  if embedded_tc and r["role"] == "assistant":

tests/test_session.py

@@ -641,6 +641,55 @@ def test_messages_with_embedded_tool_calls(self, tmp_path):
  assert result[2]["role"] == "tool"
  assert result[2]["tool_call_id"] == "call_1"
 
+ def test_assistant_reasoning_passed_back(self, tmp_path):
+ """get_messages_for_llm() maps persisted reasoning to reasoning_content
+ (DeepSeek thinking-mode pass-back, rant 2026-08-22T17:25:02)."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": "answer",
+ "reasoning": "think step by step",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "think step by step"
+
+ def test_assistant_reasoning_with_tool_calls_passed_back(self, tmp_path):
+ """get_messages_for_llm() keeps reasoning_content on tool-call rounds."""
+ session = Session.create(tmp_path)
+ session.append_message({
+ "type": "message",
+ "role": "assistant",
+ "content": None,
+ "reasoning": "need to read the file",
+ "tool_calls": [
+ {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}},
+ ],
+ })
+ session.append_message({
+ "type": "tool_result",
+ "role": "tool",
+ "tool_call_id": "call_1",
+ "content": "file content",
+ })
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 2
+ assert result[0]["reasoning_content"] == "need to read the file"
+ assert "tool_calls" in result[0]
+
+ def test_old_records_without_reasoning_untouched(self, tmp_path):
+ """get_messages_for_llm() skips reasoning for old records without the field."""
+ session = Session.create(tmp_path)
+ session.append_message({"type": "message", "role": "assistant", "content": "old answer"})
+
+ result = session.get_messages_for_llm()
+ assert len(result) == 1
+ assert "reasoning_content" not in result[0]
+ assert result[0] == {"role": "assistant", "content": "old answer"}
+
 
 class TestSessionListSessions:
  """Tests for Session.list_sessions()."""