fix(sandbox): stop splitting Cloudflare SSE events at chunk boundaries

[abhay-codes07]

Summary

The Cloudflare Worker's server-sent event stream is split into lines by a hand-rolled decoder that classifies a chunk ending on CR immediately: it emits the line and keeps a bare CR as the buffer. A CR at the end of the available bytes cannot be classified yet, because the next chunk may begin with LF and CRLF is a single terminator. The stored CR is then read as the start of a new line, so the following LF produces a blank line, and a blank line dispatches an SSE event. One multi-line event arrives as several.

Transport chunk boundaries are arbitrary, so which events an application receives depended on how the response happened to be framed rather than on its content:

stream = "data: a\r\ndata: b\r\n\r\n"

one chunk        ['data: a', 'data: b', '']          -> [('message', 'a\nb')]
split on the CR  ['data: a', '', 'data: b', '', '']  -> [('message', 'a'), ('message', 'b')]

flush() had the matching problem and appended a blank line for a stream ending on CR. The CR-only and trailing-CR shapes were also wrong when decoded in a single call, independently of chunking.

The decoder is adapted from the httpx line decoder, which defers a trailing CR into the next iteration and documents that it matches str.splitlines(). This restores that behavior: keep the unterminated line, CR included, until the next chunk decides, and let flush close the pending line instead of adding a blank one.

Measured against str.splitlines() as the oracle:

	before	after
stream shapes disagreeing at some chunk size	6 of 7	0 of 7
randomized streams whose events changed with chunking	191 of 400	0 of 400

Test plan

The decoders had no tests, which is how this survived. Added to tests/extensions/sandbox/test_cloudflare.py:

• test_sse_line_decoder_matches_splitlines_for_every_chunk_boundary, parametrized over LF, CRLF, CR, mixed, empty-data, trailing-CR and bare-CRLF streams, asserting agreement with str.splitlines() at every chunk size from 1 to the full length.
• test_sse_event_is_not_split_when_crlf_straddles_a_chunk_boundary, asserting the decoded events are identical for every chunk size.
• test_sse_line_decoder_flush_does_not_emit_a_phantom_line_for_a_trailing_cr.

8 fail on main and pass with the fix. The LF-only case passes in both runs, which is the control confirming the tests are not simply rejecting the decoder wholesale:

# main
FAILED ...test_sse_line_decoder_matches_splitlines_for_every_chunk_boundary[crlf]
FAILED ...test_sse_line_decoder_matches_splitlines_for_every_chunk_boundary[cr]
FAILED ...test_sse_line_decoder_matches_splitlines_for_every_chunk_boundary[mixed]
FAILED ...test_sse_line_decoder_matches_splitlines_for_every_chunk_boundary[empty_data]
FAILED ...test_sse_line_decoder_matches_splitlines_for_every_chunk_boundary[trailing_cr]
FAILED ...test_sse_line_decoder_matches_splitlines_for_every_chunk_boundary[bare_crlf]
FAILED ...test_sse_event_is_not_split_when_crlf_straddles_a_chunk_boundary
FAILED ...test_sse_line_decoder_flush_does_not_emit_a_phantom_line_for_a_trailing_cr
8 failed, 3 passed, 68 deselected

Verification from the repository root:

Command	Result
make format	clean
make lint	all checks passed
make mypy	5 errors, all pre-existing on main, none in the touched files
make pyright	1 error, pre-existing on main (src/agents/sandbox/util/tar_utils.py:161)
uv run pytest tests/extensions/sandbox/test_cloudflare.py	79 passed
make tests	5827 passed

The full suite run was done on Windows, where some sandbox symlink and tracing timing tests fail independently of this change. I diffed the failing set against a clean main checkout in the same environment and the two sets are identical, 54 versus 54, with no differences in either direction.

Issue number

Closes #4214

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

The verification script is a bash script that shells out to make. I ran the underlying steps individually instead, with the results above.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.