Fix SyntaxError crash line to show file, line, and column

[SemTiOne]

SyntaxError crash lines now use the error's own file, line, and column instead of the location of the traceback entry where it was raised.

Before:

    E     File "package/lib.py", line 1
    E       def broken(:
    E                  ^
    E   SyntaxError: invalid syntax
    package/lib.py:1: SyntaxError

After:

    E     File "package/lib.py", line 1
    E       def broken(:
    E                  ^
    E   SyntaxError: invalid syntax
    package/lib.py:1:12: SyntaxError: invalid syntax

Collection errors append the same file:line:col crash line when precise location is available. The ReprFileLocation dataclass gains an optional column field (column: int | None = None); existing output is unchanged when the field is absent.

Note: This PR doesn't solve the removal of the now-redundant File..., line N block. This is a deliberate design decision; it comes from stdlib format_exception_only(); deferred to a follow-up issue/discussion.

• Include documentation when adding new features.
• Include new tests or update existing tests when applicable.
• Allow maintainers to push and squash when merging my commits. Please uncheck this if you prefer to squash the commits yourself.
• Closes Format SyntaxErrors like Tracebacklines #2388
• If AI agents were used, they are credited in Co-authored-by commit trailers.
• Create a new changelog file in the changelog directory, with a name like <ISSUE NUMBER>.<TYPE>.rst. See changelog/README.rst for details.
• Add yourself to AUTHORS in alphabetical order.

[SemTiOne]

Hello @RonnyPfannschmidt, could you please review this PR when you have a moment? Thank you for your time.

[nicoddemus]

Why the reprcrash.column guard? If not None, we call str(reprcrash) anyway.

[SemTiOne]

You are right. str(reprcrash) is safe with column=None, but the guard isn't about safety. Dropping the guard here would append a near-duplicate location line in the no-column case. The column is the signal that we have the SyntaxError's own precise location worth appending, without it we keep the existing traceback line. What do you think?

[nicoddemus]

We should make the intent more explicit then, preferably without just slapping a comment there... 🤔

[SemTiOne]

Of course, here's my code proposal, wdyt?

msg = str(repr_)
syntax_error_has_precise_location = reprcrash is not None and reprcrash.column is not None
if syntax_error_has_precise_location:
    msg += "\n" + str(reprcrash) 

[nicoddemus]

Most of the tests are constructing a SyntexError manually... can we construct them instead via actual code and exec()? Seems more resilient/reliable that way.

[SemTiOne]

Yes I agree. But manual tests here have some purposes. One, they pin the exact file:line:col format. Two, without_filename is unreachable here via compile() which will raise TypeError before the parsing starts, also for without_lineno, it has no known practicable path that produces it either, they test the defensive fallback logic in _syntax_error_location. And yes, without_offset can be produced via exec() (null byte), but I think it's an obscure trigger, because that ties to a CPython implementation detail without adding resilience over the manual construction. So I suggest... that we combine it, exec() is for happy-path but keep the manual tests. How about that?

[nicoddemus]

Why are we raising manual SyntaxErrors here, instead of just writing code with syntax errors?

[SemTiOne]

So that it stays a runtime failure pinned to file.py:1:5. Broken code either collects it differently or can't be relied on to give offset 5.

[RonnyPfannschmidt]

Thanks for picking this up — the _getreprcrash() half of this PR is exactly the right fix for #2388, and it verifies nicely: --tb=line now gives gen.py:1:12: SyntaxError: invalid syntax where main gives the useless test_raise.py:2: File "gen.py", line 1, and the short summary improves the same way. --tb=native is correctly untouched, and report serialization round-trips the new column field.

The repr_traceback_entry change, however, needs another look: it overwrites the last traceback entry's file:line with the SyntaxError's own location while the source snippet and in <funcname> still describe the real frame — so the one line editors parse for jump-to-location now points somewhere that doesn't contain the displayed code.

Reproducer (validated against the current head, 4d8030b, and against main):

# test_pr14899_repro.py
def helper():
    compile("def broken(:", "gen.py", "exec")

def test_issue1_wrong_frame_attribution():
    helper()

def test_issue3_chained_cause_link():
    try:
        compile("def broken(:", "gen.py", "exec")
    except SyntaxError as e:
        raise RuntimeError("wrapper") from e

# test_issue2_broken_syntax.py  (a real collection-time SyntaxError)
def broken(:
    pass

# test_issue7_import_no_offset.py  (import-time SyntaxError without offset)
raise SyntaxError("custom import error", ("weird.py", 3, None, "src", 3, None))

Issue 1 — frame attribution (--tb=short)

Actual (this PR):

gen.py:1:12: in helper
    compile("def broken(:", "gen.py", "exec")
E     File "gen.py", line 1
...

helper is not in gen.py, and gen.py:1 does not contain that source line. Expected — the frame keeps its real location (as on main), since the E-block and the crash line already carry the syntax-error location:

test_pr14899_repro.py:2: in helper
    compile("def broken(:", "gen.py", "exec")
E     File "gen.py", line 1
...

The same mislabelled line appears as the trailing gen.py:1:12: SyntaxError under the last frame in --tb=long and --fulltrace, and inside the "direct cause" section for the chained case (gen.py:1:12: in test_issue3_chained_cause_link).

Issue 2 — collection error mixes three files into one entry

Actual (this PR, pytest test_issue2_broken_syntax.py):

test_issue2_broken_syntax.py:1:12: in parse
    return compile(source, filename, mode, flags,
E     File ".../test_issue2_broken_syntax.py", line 1
E       def broken(:
E                  ^
E   SyntaxError: invalid syntax
test_issue2_broken_syntax.py:1:12: SyntaxError: invalid syntax

The entry line fuses the user file's location, stdlib ast.parse's function name, and stdlib source — and the location then appears three times. Expected: entry lines keep their real frames; the new file:line:col crash line appended once at the end is the actual improvement:

.../ast.py:54: in parse
    return compile(source, filename, mode, flags,
E     File ".../test_issue2_broken_syntax.py", line 1
E       def broken(:
E                  ^
E   SyntaxError: invalid syntax
test_issue2_broken_syntax.py:1:12: SyntaxError: invalid syntax

Issue 3 — the crash-line append in python.py is gated on column

test_issue7_import_no_offset.py (filename and lineno present, offset absent) gets no appended crash line at all, while issue 2's case gets one — the if reprcrash.column is not None gate should be if reprcrash is not None, and arguably the append belongs in the repr machinery rather than in importtestmodule. Expected final line:

weird.py:3: SyntaxError: custom import error

Smaller points

• raise SyntaxError(None, ("file.py", 1, 5, ...)) renders the crash line as SyntaxError: None (the E-block correctly says <no detail available>).
• offset=0 passes the is not None guard and prints a zero column (file.py:1:0: ...); 0 should probably be treated like None.
• filename or path in repr_traceback_entry is dead code — _syntax_error_location already guarantees a truthy filename.

My suggestion: drop (or narrow to the bare-typename long-style case) the repr_traceback_entry rewrite, keep _getreprcrash(), and make the collection append unconditional. That keeps every win shown in the PR description's after-example except the rewritten per-frame line, without the misattribution.

---

This analysis and comment were produced by an AI (Claude, via Claude Code) at my request; I have reviewed the findings. — Ronny

[SemTiOne]

Thanks, these are valid and will be addressed. But for issue 3, I changed the guard to if reprcrash is not None: and ran your test_issue7_import_no_offset.py reproducer. Without offset, _getreprcrash() doesn't fall back to the SyntaxError's own file/lineno because that would need all three fields in _syntax_error_location. It falls back to the traceback entry instead, producing this:

...
src\_pytest\assertion\rewrite.py:187: in exec_module
    exec(co, module.__dict__)
test_issue7_import_no_offset.py:1: in <module>
    raise SyntaxError("custom import error", ("weird.py", 3, None, "src", 3, None))
E     File "weird.py", line 3
E       src
E   SyntaxError: custom import error
D:\pytest\test_issue7_import_no_offset.py:1:   File "weird.py", line 3

Not weird.py:3: SyntaxError: custom import error like you expected. I believe getting that output would need _syntax_error_location to handle missing offset, which is a bigger change because that function is shared with repr_traceback_entry, so relaxing it isn't scoped to python.py alone. Perhaps scope it as a follow-up?

[user01010111]

@SemTiOne Would you be happy for me to take the missing-offset case as a separate follow-up after this PR lands? The proposed scope is to retain a SyntaxError's filename and line when offset is None, append that location during collection, and leave traceback frame locations unchanged. I'll wait for confirmation before preparing a separate patch.

[SemTiOne]

Sure, feel free to open it. I agree with the proposed scope.

[SemTiOne]

@nicoddemus @RonnyPfannschmidt PTAL when you have a moment. 🙏

[RonnyPfannschmidt]

I prompted an AI agent (Claude, via Claude Code) to re-check this PR against my earlier review; the verification runs, the reproducers and the prose below are its work. I have read it and I am approving on the strength of it. — Ronny

---

Every blocking point from my earlier review is fixed at head (7489cf7). Dropping the repr_traceback_entry rewrite was the right call — the diff is now just three things, and each of them earns its place.

Re-verified in a clean worktree, Python 3.14.3:

• Frame attribution — --tb=short gives test_repro.py:2: in helper, the real frame. The chained-cause section is correct too.
• Collection — entries keep their real frames (ast.py:46: in parse), with one appended crash line at the end instead of three copies of a fused location.
• Zero offset — handled by the exc.offset truthiness check, pinned by a test.
• msg=None — renders SyntaxError: <no detail available>.
• filename or path — gone with the rewrite.

The win, measured against main on a real file:

main:  test_relpath.py:6:   File "sub/mod.py", line 1
PR:    sub/mod.py:1:12: SyntaxError: invalid syntax

The short summary line improves identically, which is the part I care most about — FAILED ... - File "sub/mod.py", line 1 was actively misleading. --tb=native is untouched.

testing/code/ test_reports.py test_collection.py test_terminal.py test_junitxml.py → 784 passed, 14 skipped, 1 xfailed. pre-commit run -a clean, mypy included.

The missing-offset case (weird.py:3: with no column) is still absent; that is a non-regression — main prints nothing there either — and @user01010111 has it as an agreed follow-up.

Approving. The inline comments are nits only; take them or leave them, none of them should hold this up.

[SemTiOne]

@RonnyPfannschmidt All nits addressed except nit 2 & 4.