Truncate junitxml failure and error message attributes

[Cherith1222]

Summary

--junitxml wrote the full assertion message into the message attribute of <failure> and <error>. With -vv, that attribute can be tens of thousands of characters. The attribute is now shortened from the original text, before XML escaping: whole lines are kept up to a 500-character budget. A single line longer than the budget is cut at 500 characters and then escaped, so an escape such as #x1B is not split. When anything is dropped, the attribute ends with a note that the full report is in the element text. The element body still contains the full longrepr.

Fixes #12223.

The commit already includes changelog/12223.bugfix.rst. AUTHORS was not changed.

Tests

On Windows, Python 3.13.3, in this checkout, via the project virtualenv:

• Before this revision, python -m pytest testing/test_junitxml.py -k message_is_truncated -p no:cacheprovider was 3 passed, 140 deselected, and -k failure_verbose_message was 2 passed, 141 deselected.
• After this revision: -k TestShortenFailureMessage was 4 passed; -k message_is_truncated was 3 passed; -k failure_message_cuts_on_line_boundary was 1 passed; -k raises_match_message_is_not_shortened was 1 passed; -k failure_verbose_message was 2 passed.
• ruff 0.16.9 format and check passed on src/_pytest/junitxml.py and testing/test_junitxml.py.

Not run: the rest of the pytest suite, mypy, and Python 3.11.

AI assistance

Drafted with Cursor. The account owner authorized opening this pull request. A human has not reviewed the diff. There is no Co-authored-by, Reviewed-by, or Signed-off-by trailer.

Checklist

• Include new tests or update existing tests when applicable.
• Allow maintainers to push and squash when merging my commits.
• Add text like Fixes #12223 to the PR description. The commit message already says the same.
• Changelog fragment: changelog/12223.bugfix.rst.
• AUTHORS was not updated.
• AI assistance is disclosed in this description and in the commit message. A Co-authored-by trailer was not added.

[RonnyPfannschmidt]

🤖 Written by Claude Opus 5.5 via Claude Code for the pytest maintainers; I prompted it, it did the work, I read it.

Thanks for picking this up. The goal is right: a short message attribute, with the full report in the element text, which already carries it. I'd suggest changing how the attribute is shortened.

Why the current approach falls short

• It cuts at 1000 chars with a bare .... A reader of the attribute has no hint that the full report is in the element text.
• It cuts after bin_xml_escape, so it can split an escape like #x1B. Shorten the raw message first, then escape it.
• This isn't only a -vv problem. A plain raise ValueError(5000 * "z") produces a 5,018-char attribute at default verbosity.

What common failures look like on main. I measured message for 16 typical failures: bare, == and in asserts on int/str/dict/list/approx, pytest.fail, pytest.raises (no raise and bad match), KeyError, AttributeError, a multiline exception, ExceptionGroup, and a fixture setup error.

• At -q every one is ≤ 283 chars and ≤ 7 lines.
• At -vv only container diffs grow: dict 326, list(range(200)) 4,426, and the #12223 example about 25k.
• Keeping only the first line is too aggressive. For pytest.raises(match=...), line 1 is just Regex pattern did not match., and the expected regex and actual message are on lines 2–3. String diffs and multiline exceptions lose their useful part the same way.

Suggestion: keep whole lines up to a budget. If anything was dropped, end with a pointer line:

_MESSAGE_BUDGET = 300  # or ~500

def _shorten_message(msg: str) -> str:
    lines = msg.splitlines() or [""]
    kept: list[str] = []
    size = 0
    for line in lines:
        if kept and size + len(line) + 1 > _MESSAGE_BUDGET:
            break
        kept.append(line)
        size += len(line) + 1
    out = "\n".join(kept)
    cut = len(out) > _MESSAGE_BUDGET
    if cut:
        out = out[:_MESSAGE_BUDGET] + "..."
    rest = len(lines) - len(kept)
    if not cut and not rest:
        return out
    note = f"+{rest} more lines" if rest else "truncated"
    return f"{out}\n[{note}; full report in element text]"

It would be applied as bin_xml_escape(_shorten_message(message)) in append_failure and append_error.

With a budget of 300, all 16 cases stay byte-identical at -q. At -vv only the dict and list diffs get cut, for example 4,426 → 350 chars ending in [+207 more lines; full report in element text]. A budget around 500 would keep the small -vv dict intact too. The common case doesn't change, and the pointer only shows up when something was actually dropped.

On tests: it would be good to cover the line-boundary cut, the hard cut of a single overlong first line, and an unchanged short multiline message (e.g. the pytest.raises(match=...) case).

---

Generated by Claude Code

[Cherith1222]

The raw message is shortened before bin_xml_escape. Whole lines are kept up to a 500-character budget. A single line longer than that budget is cut at 500 characters, and only then is the text escaped. When anything is dropped, the attribute ends with a note that the full report is in the element text: [+N more lines; full report in element text], or [truncated; full report in element text] when the first line itself was cut.

I used 500 rather than 300 so a small -vv container diff can still fit. A message that already fits is left unchanged, including a short multiline pytest.raises(match=...) failure.

This revision was drafted with Cursor. A human has not line-reviewed the new diff.

[RonnyPfannschmidt]

🤖 Written by Claude Opus 5.5 via Claude Code for the pytest maintainers; I prompted it, it did the work, I read it.

Closing in favor of #15150. That PR is a clean reimplementation of the line-budget approach for #12223, and its CI is green.

---

Generated by Claude Code