feat: Allow users to control Mechanism.Handled for captured exceptions - #5449
feat: Allow users to control Mechanism.Handled for captured exceptions#5449vladbrincoveanu wants to merge 7 commits into
Conversation
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #5449 +/- ##
==========================================
+ Coverage 74.62% 74.64% +0.02%
==========================================
Files 512 512
Lines 18672 18680 +8
Branches 3657 3658 +1
==========================================
+ Hits 13934 13944 +10
+ Misses 3865 3863 -2
Partials 873 873 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
57b519e to
ce1f8c9
Compare
|
We have a warden failing check. |
The Handled-inheritance change in this branch made inner exceptions inherit Handled=false from a non-terminal unhandled outer (e.g. an unobserved-task AggregateException marked terminal:false). Terminal was not inherited, so after the processor strips the outer Data keys, HasUnhandledTerminalException matched the inner (Handled:false, Terminal:null) and IsFromTerminalException() wrongly returned true. Propagate Terminal down the chain symmetric to Handled so the chain stays internally consistent. Terminal is an SDK-internal, non-serialized flag consumed only by the Has*TerminalException scans, so this has no wire/protocol impact. Explicit Terminal flags on any exception still win. Regression flagged by Cursor Bugbot on PR getsentry#5449. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
||
| ex.Data[Mechanism.HandledKey] ??= handled; | ||
| return client.CaptureEvent(new SentryEvent(ex)); | ||
| } |
There was a problem hiding this comment.
Unhandled capture crashes sessions
High Severity
Public CaptureException with handled: false stamps only Mechanism.HandledKey and leaves Terminal unset. Session logic treats unhandled-with-unset-terminal as UnhandledTerminal, so a manual capture ends the session as Crashed and aborts the active transaction even though the process kept running.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 5e8b0b1. Configure here.
There was a problem hiding this comment.
Confirmed, this is the current behavior: handled: false stamps only Mechanism.HandledKey and leaves Terminal unset. A genuinely unhandled exception (e.g. AppDomain.UnhandledException) sets handled: false with no terminal flag, which counts as a Crash (see SentryEvent.HasUnhandledTerminalException).
Adding terminal control to the CaptureException signature (defaulting Terminal = false for user-parameter captures) would expand the API surface beyond what #3383 asks for, so I kept it out of this PR , but happy to open a follow-up issue for it, if you think the session-crash default is too surprising for manual captures. What do you think? @jamescrosswell
(The related WinUI issue CaptureExceptionInternal overwriting the platform-forwarded Handled flag, has since been fixed in 3663ee5; only the session-terminal question above remains open.)
There was a problem hiding this comment.
OK this is going down a bit of a 🐇 🕳️ - apologies in advance for the long reply...
The TLDR; is that I think we need to expose a way to set both Handled and Terminal (not just Handled).
Detailed response
Terminal exists to determine two things:
- A session's end status (Release Health)
a. Only relevant for mobile/desktop/gaming clients (makes no sense for ASP.NET Core) - Whether to abort the active
SentryEvent.GetExceptionType() collapses Handled+Terminal into three buckets:
| Handled | Terminal | ExceptionType | Consequence |
|---|---|---|---|
not false |
(ignored) | Handled |
ReportError() — session stays alive, error count++ |
false |
unset / not false |
UnhandledTerminal |
EndSession(Crashed) and aborts the active transaction |
false |
explicitly false |
UnhandledNonTerminal |
MarkSessionAsUnhandled() — session flagged, but not ended, transaction not aborted |
Session Health
So Terminal only matters when Handled: false and it's basically used to distinguish between unhandled and crashed from the perspective of Release Health. In the .net SDK, that's captured via the the SessionEndStatus in the SentryClient:
sentry-dotnet/src/Sentry/SentryClient.cs
Lines 392 to 409 in f02bc53
And importantly, unhandled exceptions are terminal by default in the .NET SDK:
sentry-dotnet/src/Sentry/SentryEvent.cs
Lines 223 to 265 in 29ef560
Transactions
The impact on transactions is not dependent on the app model and so impacts all apps (including ASP.NET Core).
An UnhandledTerminal exception finishes the current transaction as Aborted (except OTel transactions, which the SpanProcessor handles).
- AppDomainUnhandledExceptionIntegration —
handled: false, no terminal ⇒ Terminal/Crashed. (Console, generic .NET, and the process is genuinely dying.) - WinUIUnhandledExceptionIntegration — passes through the platform's
Handledbool; if!handled, terminal/Crashed. - UnobservedTaskExceptionIntegration — the one built-in case that sets
terminal: false(async-void / unawaitedTask.Run— unhandled but doesn't crash). SentryMiddlewareandAspNetCoreExceptionProcessor— markhandled: false, no terminal flag ⇒ No session tracking in ASP.NET Core and so all this does is terminate the Transaction with an error state, which is the desired behaviour for a 5xx.
What this means for this PR
CaptureException(ex, handled: false) stamps only HandledKey, leaves Terminal unset, so it always classifies Unhandled as UnhandledTerminal = Crash.
- ASP.NET Core / web: Manual
handled: falsecapture will abort the active transaction, which might be surprising to SDK users!!! - MAUI / desktop clients: Manual
handled: falsecapture will end the session as Crashed and open a new one, denting the crash-free-sessions rate — even though nothing actually crashed. Again, possibly surprising to SDK users manually setting handled to false.
So I don't think we can expose handled without also exposing a way to set terminal - since without that you allow SDK users to specify UnhandledTerminal (crash) but no way to specify UnhandledNonTerminal (unhandled but not a crash).
There was a problem hiding this comment.
Hi James,
Thanks for taking the time to explain this.
I will take care of the comments next week Monday. Should I leave the PR in an opened state until then?
There was a problem hiding this comment.
Yeah it's marked as Changes Requested so all good to leave it open.
Thanks @vladbrincoveanu !
6345138 to
801db87
Compare
getsentry#3383) Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
061f5d3 to
0540a78
Compare
Adds SentrySdk-level coverage for the handled parameter, including that an explicit argument overrides a flag previously set via SetSentryMechanism. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The non-handled CaptureException overloads no longer default or preserve a handled flag - they behave exactly as before this feature. Only the new overloads taking an explicit `handled` argument write Mechanism.HandledKey, and CaptureExceptionInternal keeps forcing handled: false for integrations. Tests and the DiagnosticSource/EntityFramework snapshots are updated to match. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
MainExceptionProcessor now marks any exception no integration claimed as handled, instead of inferring it from the presence of a stack trace and leaving never-thrown exceptions as null. This fixes the symptom getsentry#3383 opens with - SentrySdk.CaptureException(new Exception()) reported "Handled --" - and matches the relay protocol docs, which are now cited at the decision site. Doing it in the processor keeps the capture overloads free of any implicit precedence: the plain overloads still write nothing to Exception.Data, and an explicit handled argument (or SetSentryMechanism) always wins. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The preceding two sentences restated what the quoted Relay docs already say, so keep only the authoritative reference. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
There are 2 total unresolved issues (including 1 from previous review).
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 1d662b6. Configure here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…Internal WinUIUnhandledExceptionIntegration forwards the platform's Handled value via SetSentryMechanism before capturing; stamping false unconditionally clobbered it. Only default to unhandled when no flag was declared. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
jamescrosswell
left a comment
There was a problem hiding this comment.
@vladbrincoveanu thank you very much for the PR... it does raise some tricky problems though - see my inline comments below.
| // https://getsentry.github.io/relay/relay_event_schema/protocol/struct.Mechanism.html#structfield.handled | ||
| // "Exceptions captured using capture_exception (called from user code) are handled=true as the user | ||
| // explicitly captured the exception (and therefore kind of handled it)." | ||
| mechanism.Handled = true; |
There was a problem hiding this comment.
This would be a change in behaviour. I think it probably makes sense... we can probably do it without making a major release - arguably it's a bug fix rather than an arbitrary change.
@dingsdax do you have any steer?
| // Integrations stamp the flag via SetSentryMechanism before calling this (e.g. WinUI forwards the | ||
| // platform's Handled value); only default to unhandled when nothing was declared. |
There was a problem hiding this comment.
Suggest we make this a bit clearer. This method is internal and the mechanism is always set by internal integrations before it's called.
We can add it - I think we should note in the comment that it's defensive only, if we do.
| // Integrations stamp the flag via SetSentryMechanism before calling this (e.g. WinUI forwards the | |
| // platform's Handled value); only default to unhandled when nothing was declared. | |
| // Integrations always call `SetSentryMechanism` before calling this method, so | |
| // this fallback is defensive only. In practice, this code never executes. |
| /// <param name="handled">Whether the exception was handled. Recorded on the exception, overriding any flag | ||
| /// previously set on it, including one set via <see cref="SentryExceptionExtensions.SetSentryMechanism"/>.</param> | ||
| /// <returns>The Id of the event</returns> | ||
| public static SentryId CaptureException(this IHub hub, Exception ex, Action<Scope> configureScope, bool handled) |
There was a problem hiding this comment.
I'd suggest reordering the parameters. The action will likely be a lambda function which could be arbitrarily long/complex. Having the bool as the last parameter would make things harder to read.
| public static SentryId CaptureException(this IHub hub, Exception ex, Action<Scope> configureScope, bool handled) | |
| public static SentryId CaptureException(this IHub hub, Exception ex, bool handled, Action<Scope> configureScope) |
| /// previously set on it, including one set via <see cref="SentryExceptionExtensions.SetSentryMechanism"/>.</param> | ||
| /// <returns>The Id of the event.</returns> | ||
| [DebuggerStepThrough] | ||
| public static SentryId CaptureException(Exception exception, Action<Scope> configureScope, bool handled) |
There was a problem hiding this comment.
See related comment above.
| public static SentryId CaptureException(Exception exception, Action<Scope> configureScope, bool handled) | |
| public static SentryId CaptureException(Exception exception, bool handled, Action<Scope> configureScope) |
|
|
||
| ex.Data[Mechanism.HandledKey] ??= handled; | ||
| return client.CaptureEvent(new SentryEvent(ex)); | ||
| } |
There was a problem hiding this comment.
OK this is going down a bit of a 🐇 🕳️ - apologies in advance for the long reply...
The TLDR; is that I think we need to expose a way to set both Handled and Terminal (not just Handled).
Detailed response
Terminal exists to determine two things:
- A session's end status (Release Health)
a. Only relevant for mobile/desktop/gaming clients (makes no sense for ASP.NET Core) - Whether to abort the active
SentryEvent.GetExceptionType() collapses Handled+Terminal into three buckets:
| Handled | Terminal | ExceptionType | Consequence |
|---|---|---|---|
not false |
(ignored) | Handled |
ReportError() — session stays alive, error count++ |
false |
unset / not false |
UnhandledTerminal |
EndSession(Crashed) and aborts the active transaction |
false |
explicitly false |
UnhandledNonTerminal |
MarkSessionAsUnhandled() — session flagged, but not ended, transaction not aborted |
Session Health
So Terminal only matters when Handled: false and it's basically used to distinguish between unhandled and crashed from the perspective of Release Health. In the .net SDK, that's captured via the the SessionEndStatus in the SentryClient:
sentry-dotnet/src/Sentry/SentryClient.cs
Lines 392 to 409 in f02bc53
And importantly, unhandled exceptions are terminal by default in the .NET SDK:
sentry-dotnet/src/Sentry/SentryEvent.cs
Lines 223 to 265 in 29ef560
Transactions
The impact on transactions is not dependent on the app model and so impacts all apps (including ASP.NET Core).
An UnhandledTerminal exception finishes the current transaction as Aborted (except OTel transactions, which the SpanProcessor handles).
- AppDomainUnhandledExceptionIntegration —
handled: false, no terminal ⇒ Terminal/Crashed. (Console, generic .NET, and the process is genuinely dying.) - WinUIUnhandledExceptionIntegration — passes through the platform's
Handledbool; if!handled, terminal/Crashed. - UnobservedTaskExceptionIntegration — the one built-in case that sets
terminal: false(async-void / unawaitedTask.Run— unhandled but doesn't crash). SentryMiddlewareandAspNetCoreExceptionProcessor— markhandled: false, no terminal flag ⇒ No session tracking in ASP.NET Core and so all this does is terminate the Transaction with an error state, which is the desired behaviour for a 5xx.
What this means for this PR
CaptureException(ex, handled: false) stamps only HandledKey, leaves Terminal unset, so it always classifies Unhandled as UnhandledTerminal = Crash.
- ASP.NET Core / web: Manual
handled: falsecapture will abort the active transaction, which might be surprising to SDK users!!! - MAUI / desktop clients: Manual
handled: falsecapture will end the session as Crashed and open a new one, denting the crash-free-sessions rate — even though nothing actually crashed. Again, possibly surprising to SDK users manually setting handled to false.
So I don't think we can expose handled without also exposing a way to set terminal - since without that you allow SDK users to specify UnhandledTerminal (crash) but no way to specify UnhandledNonTerminal (unhandled but not a crash).


Closes #3383
Implements the approach agreed in this comment: SDK users get explicit control over
Mechanism.Handled, and manually captured exceptions default tohandled: true.API changes
New overloads for explicit control:
SentrySdk.CaptureException(Exception exception, bool handled)SentrySdk.CaptureException(Exception exception, Action<Scope> configureScope, bool handled)SentryClientExtensions.CaptureException(this ISentryClient client, Exception ex, bool handled)HubExtensions.CaptureException(this IHub hub, Exception ex, Action<Scope> configureScope, bool handled)The existing signatures are kept, so this is source- and binary-compatible.
handledis a required parameter rather than optional: an optional one would be permanently shadowed by the existing exact-match overloads, and replacing those signatures would be binary-breaking.Behavior
handledrecords that value, overriding any flag previously set on the exception (including viaSetSentryMechanism).Exception.Data, so there is no implicit precedence between a preset flag and a plainCaptureException(ex)call.MainExceptionProcessor. Any exception no integration claimed is now markedhandled: true, replacing the previous "thrown ⇒ true, never thrown ⇒ null" heuristic. This is what fixes the symptom Allow SDK users to controlMechanism.Handledfor captured exceptions #3383 opens with:SentrySdk.CaptureException(new Exception())reportedHandled --and now reportsHandled: true, per the protocol docs ("exceptions captured using capture_exception are handled=true"), which are cited at the decision site. Thrown-and-caught exceptions reportedHandled: truebefore and still do.CaptureExceptionInternal(the integration capture path) stampshandled: false, since integrations capture exceptions that user code did not handle.Terminalsemantics are unchanged. See the discussion below on terminal control for user captures and on the WinUI integration, both are proposed as follow-ups rather than expanding this PR's scope.Testing
Unit tests cover explicit
true/falseon every capture path, explicit values overriding preset flags, the untouched plain overloads, and the integration-path default. Verify snapshots updated: API approvals for all four target frameworks, the aggregate-exception mechanism snapshot, and theSentry.DiagnosticSource/Sentry.EntityFrameworkintegration snapshots, where constructed-and-captured exceptions now reportHandled: true.