From 425aab2ab951ac492efcba6ea9c5aab72eb8f16e Mon Sep 17 00:00:00 2001
From: Pravin Kamble <iampbkamble@gmail.com>
Date: Tue, 10 Nov 2026 09:31:29 +0530
Subject: [PATCH] DRU-613 -- Skip unparseable lines in the pi stream

---
 backend/druks/harnesses/pi.py |  8 ++++++-
 backend/tests/test_pi.py      | 41 +++++++++++++++++++++++++----------
 2 files changed, 36 insertions(+), 13 deletions(-)

diff --git a/backend/druks/harnesses/pi.py b/backend/druks/harnesses/pi.py
index 6b84a411..72b4a245 100644
--- a/backend/druks/harnesses/pi.py
+++ b/backend/druks/harnesses/pi.py
@@ -1,3 +1,4 @@
+import contextlib
 import json
 import re
 import shlex
@@ -127,7 +128,12 @@ async def build_invocation(
 
     def parse(self, result: HarnessRunResult, *, artifact_dir: Path, run_id: str) -> Any:
         try:
-            events = [json.loads(line) for line in result.stdout.splitlines()]
+            events = []
+            # result.stdout keeps only the last STDOUT_BUFFER_LIMIT bytes, so the first
+            # line can be cut, even inside a UTF-8 character.
+            for line in result.stdout.splitlines():
+                with contextlib.suppress(ValueError):
+                    events.append(json.loads(line))
             # pi repeats an assistant message on turn_end and agent_end;
             # message_end is the one place each appears once.
             messages = [
diff --git a/backend/tests/test_pi.py b/backend/tests/test_pi.py
index c63c74ca..652e18ca 100644
--- a/backend/tests/test_pi.py
+++ b/backend/tests/test_pi.py
@@ -302,13 +302,15 @@ def test_parse_treats_a_missing_result_file_as_invalid_output(tmp_path: Path) ->
         _parse(b'{"type":"agent_settled"}', tmp_path)
 
 
-@pytest.mark.parametrize(
-    ("stdout", "expected_cost"),
-    [
-        (b"not json", None),
+def test_parse_skips_the_line_the_buffer_trim_cut(tmp_path: Path) -> None:
+    (tmp_path / "run-1").mkdir()
+    (tmp_path / "run-1" / "output.json").write_text('{"answer": "done", "count": 3}')
+    stdout = b"\n".join(
         (
+            b'\x99 done"}]}}',
             _message_end(
                 "assistant",
+                stopReason="stop",
                 usage={
                     "input": 1,
                     "output": 1,
@@ -317,18 +319,33 @@ def test_parse_treats_a_missing_result_file_as_invalid_output(tmp_path: Path) ->
                     "cost": {"total": 0.25},
                 },
             ),
-            0.25,
-        ),
-    ],
-)
-def test_parse_treats_a_broken_stream_as_invalid_output(
-    tmp_path: Path, stdout: bytes, expected_cost: float | None
-) -> None:
+        )
+    )
+
+    output = _parse(stdout, tmp_path)
+
+    assert _Contract.model_validate(output) == _Contract(answer="done", count=3)
+    cost_usd, _ = read_cost(tmp_path / "run-1")
+    assert cost_usd == 0.25
+
+
+def test_parse_treats_a_broken_stream_as_invalid_output(tmp_path: Path) -> None:
+    stdout = _message_end(
+        "assistant",
+        usage={
+            "input": 1,
+            "output": 1,
+            "cacheRead": 0,
+            "cacheWrite": 0,
+            "cost": {"total": 0.25},
+        },
+    )
+
     with pytest.raises(HarnessInvalidOutputError, match="no usable stream"):
         _parse(stdout, tmp_path)
 
     cost_usd, _ = read_cost(tmp_path / "run-1")
-    assert cost_usd == expected_cost
+    assert cost_usd == 0.25
 
 
 async def test_a_pasted_key_becomes_the_catalog_entry(client, druks_db) -> None: