[Elioooon]

Summary

• detect JSON-RPC messages split over more than two physical stdout lines
• add a v2 fixture that reproduces the transport classification gap

Validation

• uv run pytest tests/v2/test_cli.py -k 'transport or message_split_across_multiple_lines' -q
• uv run --with ruff ruff check src/tck/v2/conformance/test_transport.py tests/fixtures/agents/v2/message_split_across_multiple_lines.py
• git diff --check

[EugeneTheDev]

Problem: the scan stops at the first following line that parses as JSON on its own, and it has no other stop. That gives a false PASS in one case and a very slow run in another.

How it happens:

1. An agent pretty-prints a message that has an array of strings, e.g. "tags": ["a", "b"] with indent=2.
2. The line "b" is valid JSON by itself, so following.parse_error is None hits break before the closing lines are joined.
3. ACP-TRANSPORT-203 PASSes for a message split over many lines. I ran such an agent: 201 FAIL, 203 PASS.

Second case: when the run of unparsable lines is long (an agent that logs to stdout), every start index rescans the rest of the run and decodes a growing buffer, so the cost is about n^3. With a conforming agent that prints log lines after EOF, the whole run took 1.5s for 1000 lines and 21s for 3000 lines. On main it takes 1.3s for 3000 lines.

Impact: a framing violation can go unreported, and noisy agents can hit --test-timeout. Both come from this PR's loop.

Fix: do not stop on a valid line. Stop when the joined text is already malformed before its end, because more lines cannot repair it:

for following in received[index + 1 :]:
    combined.extend(b"\n")
    combined.extend(following.raw)
    try:
        text = combined.decode("utf-8")
        json.loads(text)
    except UnicodeDecodeError:
        break
    except json.JSONDecodeError as exc:
        if exc.pos < len(text.rstrip()):
            break  # malformed before the end; more lines cannot fix it
        continue
    pytest.fail(...)

I applied this temporarily: the string-array agent now FAILs 203, the PR's fixture still FAILs, the 3000-line agent PASSes in 0.13s, garbage_after_response.py and conforming.py PASS, and the related tests/v2/test_cli.py tests pass. The new fixture has no array, so it would not catch the first case; adding a string array to its message (e.g. in _meta) would.

[EugeneTheDev]

The docstring above (lines 84-88) still says the reassembly is "the next entry's raw bytes". It now joins any number of following lines. Please update it.

[Elioooon]

Addressed on the current head (bba543a). The reassembly scan no longer stops merely because an individual following line parses as JSON; it keeps joining lines and stops only once a JSON decode error occurs before the trailing whitespace, which cannot be repaired by appending more input. The split-message fixture now includes _meta.tags: ["a", "b"], covering the string-array early-stop case, and the docstring now describes reassembly across one or more following entries.

Validation completed here: python3 -m compileall -q src/tck/v2/conformance/test_transport.py tests/fixtures/agents/v2/message_split_across_multiple_lines.py and git diff --check origin/main...HEAD. I could not complete the focused pytest run on this host: its installed Python is 3.7.3 and uv was still obtaining a compatible interpreter. This PR currently has no GitHub check runs.