[xihaian251]

Fixes #999.

Application.astream_result calls post_end_stream from its error path, but the
callback is only meaningful after pre_start_stream has run. That ordering does not
hold in two cases:

• the halt action is not streaming — the if not next_action.streaming branch
  delegates to _astep and returns early on success, so pre_start_stream never runs;
• input processing fails before the streaming phase.

Hooks that keep per-stream state cannot pair the callback. LocalTrackingClient and
S3TrackingClient index self.stream_state[app_id, action, partition_key], which only
pre_start_stream populates, so post_end_stream raises KeyError from inside the
except block and replaces the exception the application raised:

tracked=False: ValueError: original failure
tracked=True:  KeyError: ('<app_id>', 'fail', None)

Net effect: attaching a tracker changes which exception the caller sees.

Changes

• track whether pre_start_stream completed, and only call post_end_stream from the
  error path when it did;
• add a core test that a failing non-streaming action does not fire post_end_stream
  without pre_start_stream;
• add a tracking test that a failing action keeps its original exception with
  LocalTrackingClient attached.

Semantics of the guard

stream_started is set only after the pre_start_stream hook call returns, so a hook
that raises during initialization is not followed by post_end_stream. The alternative
(set it before the call, so every adapter is paired even on partial initialization) is
also defensible; I picked "post only for a completed pre" because it guarantees the
error path can never mask the original exception. Happy to switch if you prefer the
other contract.

How I tested this

• new tests: 2 passed
• tests/core/test_application.py: 143 passed (142 before)
• tests/tracking/test_local_tracking_client.py: 18 passed (17 before)
• black 23.11.0 --line-length=100, isort 5.12.0 (--profile black), flake8 6.1.0:
  clean on the three changed files
• verified that both new tests fail on 8bfd7268 without the source change

Not run: the full pre-commit suite (frontend lint needs npm), and the whole tests/
tree. Linux / Python 3.9–3.12 not tested (local environment: Windows 11, Python 3.13.1).

Notes

• S3TrackingClient.post_end_stream has the same lookup; it is covered by the same fix.
• stream_result (sync) does not call post_end_stream from its error path, so it never
  had this masking problem — the async twin was the outlier.
• This is independent of astep() passes unprocessed inputs to pre_run_step hooks #979 / fix(core): process inputs before astep lifecycle hooks #980 / fix(core): process and validate inputs before pre_run_step in astep (#979) #995: the code path is untouched by those PRs.

[Copilot]

🟢 Approval recommended

No active review issues were identified for the change guarding stream cleanup when initialization has not completed.

No open issues 🎉

What changed in this PR

Fixes #999 by preventing Application.astream_result() from ending a stream whose initialization never completed, preserving the original exception when tracking is enabled.

Changes:

• Guard error-path post_end_stream calls on successful completion of pre_start_stream.
• Add regression tests for lifecycle-hook pairing and original-exception propagation with local tracking.

File	Description
tests/​tracking/​test_local_tracking_client.py	Verifies local tracking preserves a failing action’s original exception.
tests/​core/​test_application.py	Verifies a failing non-streaming action triggers neither stream lifecycle hook.
burr/​core/​application.py	Calls error-path stream cleanup only after stream initialization completes.

_{Review effort: Balanced}

[xihaian251]

CI failure analysis
I investigated the failures in Build Burr, attempt #2, and confirmed that they are unrelated to the three files changed by this PR.

1. Python 3.9 jobs (4 failures)
   All four jobs fail during actions/setup-python@v7, before dependency installation or test execution:
   The version '3.9' with architecture 'x64' was not found for Ubuntu 26.04.
   The latest upstream main is affected by the same failures. This appears to be a runner compatibility issue rather than a test regression. Pinning the Python 3.9 matrix entries to ubuntu-24.04 may resolve it.
2. Pre-commit failure
   Black reformatted two example files unrelated to this PR:
   examples/image-telephone/application.py examples/integrations/hamilton/image-telephone/application.py
   Both files were unchanged from the original PR base. Upstream has since addressed their formatting in commit 5e934dde, and pre-commit now passes on main.
3. Passing checks
   Python 3.10, 3.11, and 3.12 test matrices passed.
   Release Validation passed, including its installation and packaging checks.
   The two new regression tests passed locally, and both failed on the unpatched baseline.
   I'll keep this PR focused on fixing astream_result() masks action exceptions with a tracker KeyError when the action is non-streaming #999 rather than including unrelated CI configuration changes.
   Thanks for reviewing!

[YongGoose]

I checked this on top of current main (36e3ff3, which now includes #1027) in a Linux container with Python 3.12. The patch applies cleanly.

• With the local tracker, the repro from #999 raises KeyError: ('<app_id>', 'fail', None) on main and the original ValueError with this change.
• test_astream_result_failing_non_streaming_action_does_not_end_stream and test_astream_result_failing_action_keeps_original_exception fail on main and pass here. test_astream_result_failing_streaming_action_ends_the_stream_it_started passes on both, as expected.
• tests/core and tests/tracking: 507 passed, 1 skipped.
• black 23.11.0, isort 5.12.0 and flake8 6.1.0 with the pre-commit settings are clean on the three changed files.

#1016 makes the same source change for the same issue (a stream_started flag set after pre_start_stream and checked before post_end_stream). This one was opened first and also tests that a stream that did start still gets its post_end_stream.