Skip to content

DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.

fix(subagent): retain long subprocess output lines

dajiaohuang41bac2f27427
diff --git a/deeptutor/services/subagent/process.py b/deeptutor/services/subagent/process.py
index 1bbe321d9..2955e5909 100644
--- a/deeptutor/services/subagent/process.py
+++ b/deeptutor/services/subagent/process.py
@@ -89,11 +89,20 @@ async def stream_process_lines(
await queue.put(None)
return
try:
+ pending = bytearray()
while True:
- raw = await stream.readline()
+ # JSONL backends can emit whole answers or tool results in one
+ # frame, exceeding StreamReader.readline's default 64 KiB limit.
+ raw = await stream.read(65536)
if not raw:
break
- await queue.put((channel, raw.decode("utf-8", "replace").rstrip("\r\n")))
+ pending.extend(raw)
+ while (boundary := pending.find(b"\n")) != -1:
+ line = bytes(pending[:boundary])
+ del pending[: boundary + 1]
+ await queue.put((channel, line.decode("utf-8", "replace").rstrip("\r\n")))
+ if pending:
+ await queue.put((channel, pending.decode("utf-8", "replace").rstrip("\r\n")))
except Exception: # pragma: no cover - defensive: a broken pipe must not hang the queue
logger.debug("subagent %s pump failed", channel, exc_info=True)
finally:
diff --git a/tests/services/test_subagent_long_lines.py b/tests/services/test_subagent_long_lines.py
new file mode 100644
index 000000000..e74e8eab3
--- /dev/null
+++ b/tests/services/test_subagent_long_lines.py
@@ -0,0 +1,53 @@
+"""Native subprocess output must survive large, message-granular frames."""
+
+import asyncio
+import sys
+
+import pytest
+
+from deeptutor.services.subagent.openclaw import OpenClawBackend
+from deeptutor.services.subagent.process import stream_process_lines
+
+
+async def test_large_native_lines_preserve_both_channels_and_final_fragment():
+ text = "学习" * 40000
+ script = (
+ "import sys; "
+ "sys.stdout.buffer.write(('学习' * 40000 + '\\n').encode()); "
+ "sys.stdout.flush(); sys.stderr.write('diagnostic\\n'); "
+ "sys.stderr.flush(); sys.stdout.buffer.write('尾部'.encode()); sys.stdout.flush()"
+ )
+
+ async def collect():
+ return [item async for item in stream_process_lines([sys.executable, "-c", script])]
+
+ # The unfixed reader can stop draining and leave the child blocked on its pipe.
+ output = await asyncio.wait_for(collect(), timeout=5)
+ assert [line for channel, line in output if channel == "stdout"] == [text, "尾部"]
+ assert ("stderr", "diagnostic") in output
+ assert output[-1] == ("exit", "0")
+
+
+async def test_openclaw_receives_large_native_json_answer(monkeypatch):
+ answer = "x" * 100000
+ script = (
+ "import json; print(json.dumps({'status': 'ok', "
+ "'result': {'payloads': [{'text': 'x' * 100000}]}}))"
+ )
+ backend = OpenClawBackend()
+ monkeypatch.setattr(
+ backend,
+ "_build_command",
+ lambda *args, **kwargs: [sys.executable, "-c", script],
+ )
+ events = []
+
+ async def on_event(event):
+ events.append(event)
+
+ result = await asyncio.wait_for(backend.consult("question", on_event=on_event), timeout=5)
+ assert result.success, result.error
+ assert result.final_text == answer
+ assert any(event.text == answer for event in events)