Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

[codex] fix(proxy): parse CRLF SSE event terminators (#649)

Yui(ゆい)4b1c449c73af
diff --git a/headroom/proxy/helpers.py b/headroom/proxy/helpers.py
index 1930ec188..baa0e868e 100644
--- a/headroom/proxy/helpers.py
+++ b/headroom/proxy/helpers.py
@@ -669,10 +669,24 @@ def get_body_too_large_status() -> int:
return value
-# Sentinel used by the SSE byte-buffer helper to mark events that have no
-# `event:` line. Per the SSE spec the default event name is "message"; we
-# return ``None`` so callers can decide whether to apply that default.
-_SSE_EVENT_TERMINATOR = b"\n\n"
+# SSE byte-buffer helper supports LF and CRLF event separators. Per the SSE
+# spec the default event name is "message"; we return ``None`` so callers can
+# decide whether to apply that default.
+_SSE_EVENT_TERMINATORS = (b"\n\n", b"\r\n\r\n")
+
+
+def _find_sse_event_terminator(buf: bytearray) -> tuple[int, int] | None:
+ """Return the earliest complete SSE event terminator in ``buf``."""
+ matches = [
+ (idx, len(terminator))
+ for terminator in _SSE_EVENT_TERMINATORS
+ if (idx := buf.find(terminator)) != -1
+ ]
+ if not matches:
+ return None
+ return min(matches, key=lambda match: match[0])
+
+
_SSE_EVENT_LINE_PREFIX = b"event:"
_SSE_DATA_LINE_PREFIX = b"data:"
@@ -718,20 +732,20 @@ def parse_sse_events_from_byte_buffer(
multi-byte characters split across TCP reads will corrupt content.
"""
events: list[tuple[str | None, str]] = []
- terminator = _SSE_EVENT_TERMINATOR
while True:
- idx = buf.find(terminator)
- if idx == -1:
+ terminator_match = _find_sse_event_terminator(buf)
+ if terminator_match is None:
break
+ idx, terminator_len = terminator_match
event_bytes = bytes(buf[:idx])
# Drain the event + the trailing terminator from the buffer.
- del buf[: idx + len(terminator)]
+ del buf[: idx + terminator_len]
# Decoding the COMPLETE event must succeed. If it doesn't, the
# upstream emitted invalid UTF-8 mid-stream — surface loudly.
event_text = event_bytes.decode("utf-8")
event_name: str | None = None
data_lines: list[str] = []
- for line in event_text.split("\n"):
+ for line in event_text.splitlines():
if not line:
continue
# SSE comment line — ignored per spec.
diff --git a/tests/test_sse_utf8_split.py b/tests/test_sse_utf8_split.py
index 989506738..256a44f6e 100644
--- a/tests/test_sse_utf8_split.py
+++ b/tests/test_sse_utf8_split.py
@@ -85,6 +85,16 @@ def test_cjk_split_across_chunks_preserved() -> None:
assert parsed["text"] == "日本語テスト"
+def test_crlf_terminated_event_is_parsed() -> None:
+ """SSE permits CRLF event terminators as well as LF terminators."""
+ buf = bytearray(b'event: message\r\ndata: {"ok": true}\r\n\r\n')
+
+ events = parse_sse_events_from_byte_buffer(buf)
+
+ assert events == [("message", '{"ok": true}')]
+ assert buf == bytearray()
+
+
def test_complete_event_with_invalid_utf8_raises_loud() -> None:
"""Invalid UTF-8 in a *complete* event surfaces loudly (not silent corruption)."""
# Build a complete event whose data field has invalid UTF-8 bytes