fix(autocontext): add error handling to AutoContextHook.handlePreReasoning()

[xiaoqingwanga]

AgentScope-Java Version

1.0.12-SNAPSHOT

Description

Background

AutoContextHook.handlePreReasoning() does not have error handling. When the blocking model call inside AutoContextMemory.compressIfNeeded() (model.stream().block()) fails due to timeout, service errors, or rate limiting, the exception propagates without being handled, causing the Agent’s reactive pipeline to silently crash. The only trace is an AGENT_CALL_END event in tracing with "exception":"value", preceded by an approximately 10-second blank period, which corresponds to the model call timeout window.

By contrast, handlePreCall() already has a try-catch(Exception) block and continues execution after errors. handlePreReasoning() is missing the same protection.

Changes

AutoContextHook.java

• Wrapped the entire body of handlePreReasoning() in a try-catch block, aligning it with the existing pattern in handlePreCall().
• For non-interruption exceptions, such as model timeouts or rate limiting: log a WARN message and return the original event. The Agent continues running with uncompressed messages, providing graceful degradation.
• For exceptions containing an InterruptedException, where Reactor .block() wraps checked exceptions in a RuntimeException: detect it by walking the cause chain, then return Mono.error(e) to correctly propagate the interruption signal.
• Added a private helper method, isInterruptedException(Throwable), which walks through getCause() level by level to detect InterruptedException.
• Updated the Javadoc for handlePreReasoning() to document the new error-handling behavior.

AutoContextHookTest.java

• Added testPreReasoningEventHandlesCompressionError: uses a FailingModel that throws RuntimeException, triggers strategy-4 compression, and verifies that the hook returns the original event instead of throwing an exception.
• Added testPreReasoningEventPreservesWrappedInterruptedException: uses a FailingModel that throws RuntimeException(new InterruptedException(...)), and verifies that the wrapped interruption exception is correctly propagated through Mono.error().
• Added the FailingModel test helper class.

How to test

# Run the newly added tests
mvn test -pl agentscope-extensions/agentscope-extensions-autocontext-memory \
  -Dtest="AutoContextHookTest#testPreReasoningEventHandlesCompressionError+testPreReasoningEventPreservesWrappedInterruptedException"

# Run the full module test suite: 183 tests
mvn test -pl agentscope-extensions/agentscope-extensions-autocontext-memory

Checklist

• Code has been formatted with mvn spotless:apply
• All tests are passing (mvn test) — 183/183
• Javadoc comments are complete and follow project conventions
• Related documentation has been updated, such as links, examples, etc.
• Code is ready for review

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[LearningGp]

[xiaoqingwanga]

hi, the conflicts are resolved. pls review the changes. @LearningGp

[LearningGp]

Placing onErrorResume after .map(...) inadvertently catches exceptions thrown by buildInputMessages (even if rare) and swallows them with a misleading "Failed to compress context" log. I'd suggest chaining onErrorResume directly after compressIfNeededAsync() so exceptions inside .map can bubble up normally.

[xiaoqingwanga]

Placing onErrorResume after .map(...) inadvertently catches exceptions thrown by buildInputMessages (even if rare) and swallows them with a misleading "Failed to compress context" log. I'd suggest chaining onErrorResume directly after compressIfNeededAsync() so exceptions inside .map can bubble up normally.

Yeah, this solution does help. BUT excptions from buildInputMessages are really rare. And if we apply the optimization, we'll re-introduce the original problem - if an exception happens inside the map, the compression disrupts the agent execution again.

[LearningGp]

LGTM

[LearningGp]

[xiaoqingwanga]

Close this PR because the compaction has been refactored.