Add trigger_exception_handler() for post-mortem debugging of caught exceptions

[nshepperd]

Adds a convenient trigger_exception_handler() function (previously named post_mortem(); renamed per review discussion) that can be called to immediately trigger the graphical postmortem debugger given any exc_info structure. We need to suspend sys.monitoring callbacks while doing this, because the built in reentrancy protection doesn't apply when pydevd suspends directly from normal code instead of a monitoring callback. Functions were added in _pydevd_sys_monitoring to do that.

[nshepperd]

See #1530.

[rchiodo]

Thanks for the submission but I think this would need some sort of automated test before we'd accept it. And maybe a section in the readme.md explaining how to use it.

[nshepperd]

Actually I was wondering how to setup tests for this. I figured it out today. Hooray!

[rchiodo]

/azp run

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[rchiodo]

I'm confused about something. pdb.post_mortem attaches the debugger to the process and starts debugging it. This won't. You have to be already attached, right?

What would be the point of this then? Is it just so you don't have to break on exceptions in the debugger, but instead control when the break happens?

[rchiodo]

I think ideally this would actually start VS code and open the debugger. Which might be doable, but would probably have to be configurable somehow.

[nshepperd]

Yes, the point is so that you can eg. make a context manager or function decorator to debug exceptions that bubble past it instead of setting bespoke breakpoint, which isn't really directly possible for this. I find something like this indispensable for debugging things like api servers or unit tests (pytest) that have exception handlers that eat exceptions. With this you can just wrap your route handler or set an exception handler with conftest.py and launch the program under the debugger and everything just works.

I think previously debugpy had the 'user uncaught exceptions' breakpoint added for similar use cases, but that's much less specific and in some cases has too many false positives to be useful.

You're right this is a bit different from pdb.post_mortem() in that that actually starts the debugger. I'm not sure about starting vscode though, heh. Maybe it's possible via some sort of vscode:// url? What we could do is at least run debugpy.listen and wait for the debugger to attach? I'm not sure i would actually use that though, usually if i'd want to debug something with vscode I'm already debugging it (eg. I press the 'debug unit test' button) and the issue is just getting the breakpoint.

[nshepperd]

Either way each use cases should work, i think. If you wanted to launch vscode and attach the debugger you can use existing debugpy functions for that and then call post_mortem. Or if we include that in the function you can check if debugpy.is_client_connected() before calling it if you don't want that.

[rchiodo]

With this you can just wrap your route handler or set an exception handler with conftest.py and launch the program under the debugger and everything just works.

If this is the intended use case, maybe the name is off then. It's not post_mortem, it's more like 'force_exception_handler' or something like that. I feel like 'post_mortem' is intended to be like pdb means it. It starts a debugger even when it's not already attached.

[rchiodo]

Your unit tests also seem to be failing. Not sure if you can see the errors or not

[nshepperd]

trigger_exception_handler(), maybe?

I see the test failures. Cython issue. I forgot to regenerate the cython files. Confusing because it worked locally.

[rchiodo]

trigger_exception_handler sounds good to me :)

[rchiodo]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[rchiodo]

LGTM. I'll get somebody else to take a look.

[rchiodo]

🔒 Automated review in progress — @rchiodo is auto-reviewing this PR.

[heejaechang]

🔒 Automated review in progress — @heejaechang is auto-reviewing this PR.

[heejaechang]

Approved via Review Center.

[rchiodo]

📍 src/debugpy/_vendored/pydevd/_pydevd_sys_monitoring/_pydevd_sys_monitoring.py:1875
When the except branch freshly creates a ThreadInfo via _get_thread_info(True, 1) (tracing enabled), it still reports previous_state = True. The caller's finally then sees saved == True and re-enables tracing on a thread that was not previously traced, contradicting "restore previous state" and potentially causing unexpected breakpoint stops on background threads. When the ThreadInfo is created here rather than found, return False so the caller skips the resume. Mirror the same fix in the .pyx (and regenerate the .c).

[nshepperd]

I think how it is currently is actually correct. True is the default state (for everything except pydevd-daemon threads), in the sense that every other monitoring event sets thread_info like that if it doesn't exist. So that's what we should restore.

[rchiodo]

Approved via Review Center.

[heejaechang]

Approved via Review Center.

[rchiodo]

📍 tests/debugpy/test_trigger_exception_handler.py:40

The Rules reviewer flagged that four tests (_basic, _basic_with_exception, _excinfo, _not_as_uncaught) assert with some.dict.containing({...}) and some.str.matching(r"(.+\.)?ValueError") without justification, violating docs/pylancewiki/review/tests-no-partial-asserts.md (partial-dict/substring matches require a comment explaining why the result cannot be made deterministic). This is a confirmed rule violation and is non-downgradeable. These some.* matchers are debugpy's idiomatic harness, and exceptionId legitimately varies across Python versions (ValueError vs builtins.ValueError) — so the minimal compliant fix is to add a short comment stating why exact equality isn't possible (and tighten to == on the parts of the exceptionInfo response that are deterministic).

[verified]

[rchiodo]

📍 src/debugpy/server/api.py:380

Unresolved from the prior round. The identical multi-line "excinfo must be an exception instance or a (type, value, traceback) tuple ..." message is raised by both the tuple-shape guard and the element-type guard. When the second guard fires the tuple shape is already correct, so the message is misleading. Hoist it into a single constant, or give the element-type guard a distinct message (e.g. reporting the actual element types).

[verified]

[nshepperd]

seems to be a pylance thing

[rchiodo]

📍 src/debugpy/_vendored/pydevd/_pydevd_sys_monitoring/_pydevd_sys_monitoring.py:1887
The except branch still creates thread_info via _get_thread_info(True, 1) (default trace=True) and returns previous_state = True, contradicting the docstring's "returns the previous state" for a thread that was never traced (Skeptic, Architect, and Rules all re-confirm this). However, the Advocate's caller-trace shows the literal return False requested last round is not a clean fix: returning False before setting thread_info.trace = False skips suppression during the stop (defeating the PR's PEP-669 purpose on this path), while returning False after suspending would leave the fresh thread permanently at trace=False (a suppression leak). The current behavior — suppress during the stop, then resume to the thread's True default — is actually functionally correct. Resolve the contract honestly: either clarify the docstring that for a freshly-created ThreadInfo "previous state" is the thread's default (enabled), or, if a fresh thread should end untraced, suspend during the stop AND return False so resume is skipped. Mirror whatever you choose in the .pyx and regenerate the .c.

[verified]

[nshepperd]

I think this is good now. Thank Claude for letting me stop procrastinating on this. I'll get CLA permission from my employer.

[rchiodo]

Looks good overall. A couple of low-priority, non-blocking notes on a docstring/contract mismatch and a duplicated error message; safe to merge.

[rchiodo]

Approved via Review Center.

[heejaechang]

Approved via Review Center.

[rchiodo]

On Python 3.14, an existing __exception__ local makes remove_exception_from_frame() raise ValueError, aborting the default stop and potentially masking the caught exception after an unconditional stop. Preserve and restore any previous binding rather than unconditionally removing it, with regression tests for both as_uncaught modes.

[nshepperd]

Oh, you're right. It does fail on 3.14. I've fixed this with a context manager that restores the previous state, using it in both places.

[nshepperd]

Interestingly, Claude says this is also a problem on HEAD. It shows me this repro testcase for the userUnhandled exception handler, which seemingly hangs the unit test unless the context manager/restore method is used there:

import pytest

from tests import debug
from tests.patterns import some
from tests.timeline import Event


def test_uu_repro(pyfile, target, run):
    @pyfile
    def code_to_debug():
        import threading

        import debuggee

        debuggee.setup()

        def thread_target():
            __exception__ = "sentinel"  # real local in the user boundary frame
            raise ValueError("boom")  # @raise

        t = threading.Thread(target=thread_target)
        t.start()
        t.join()
        print("main continued after thread")  # @done

    with debug.Session() as session:
        session.config["justMyCode"] = True
        session.expected_exit_code = some.int
        with run(session, target(code_to_debug)):
            session.request(
                "setExceptionBreakpoints", {"filters": ["userUnhandled"]}
            )
            session.set_breakpoints(code_to_debug, ["done"])

        session.wait_for_stop("exception")
        session.request_continue()
        session.wait_for_next(Event("exited") | Event("terminated"))

[rchiodo]

The tests do run in 3.14 as far as I know, so it should be working in main

[nshepperd]

Right, but you didn't have a unit test that puts an existing local named __exception__, as far as i can see.

[rchiodo]

Using get_file_type(...) is None excludes LIB_FILE frames as well as debugger-internal frames; a library-only traceback therefore never stops despite as_uncaught=False. Exclude only PYDEV_FILE frames so unconditional mode can select ordinary library frames.

[nshepperd]

Sure.

[rchiodo]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[nshepperd]

@microsoft-github-policy-service agree company="Anlatan"

[rchiodo]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[rchiodo]

Hmm, I'm wondering if we can just stick the exception on the frame as an attribute. Modifying the locals is usually not so great because then the user sees it in the locals list (and the potential conflict). Although it did that before. But that might explain the error you were talking about with 3.14.

Something like this instead:

self.frame.__stored_exception = exception_info

Then removing it would just be deleting that attribute.

[nshepperd]

TBH I'm not sure what __exception__ is used for. I assumed the point was for users to see it in the locals list (eg. vscode debugging terminal). Or maybe it's a legacy thing used by Eclipse?

[rchiodo]

Looking at it, it seems it takes special precautions to remove it from the locals just for that reason. So that users don't see it. I think we can probably just change addExceptionToFrame and removeExceptionFromFrame by just adding an attribute on it. Then we wouldn't have to special case locals.