fix(harness): detach memory flush and maintenance from agent call response

[birdie7761]

AgentScope-Java Version

 2.0.1

Description

Fixes #2276, #2225

MemoryFlushMiddleware and MemoryMaintenanceMiddleware appended their LLM-backed work via concatWith onto the returned Flux. Because ReActAgent.callInternal ends with takeLast(1), which cannot emit until the upstream signals
onComplete, callers that consume the agent response to completion (blockLast(), takeLast(1), WebFlux controllers awaiting Mono<Msg>) ended up waiting for the full memory flush LLM call (19–27s per call) and the consolidation LLM call
(~44s on first run).

The original implementation used doOnComplete (fire-and-forget). PR #1802 (RC4) inadvertently swapped it for concatWith, introducing this regression.

This PR restores the fire-and-forget behavior: both middlewares now subscribe their work independently of the returned Flux via doOnComplete().subscribe(), so the Flux completes as soon as the underlying agent call does.

Additional safeguards

• Both middlewares implement AutoCloseable and track pending Disposables in a Set guarded by synchronized(pending). A pre-subscribe closed check prevents new work, and a post-subscribe check disposes any subscription that raced with
  close(). HarnessAgent.close() drains them (bounded by 5s) before workspace teardown to prevent races with temp directory deletion in tests and CLI shutdown.
• captureFlushRequest is wrapped in try/catch so exceptions (e.g. from List.copyOf or resolveAgentState) cannot escape into the doOnComplete callback and turn an already-completed Flux into an error.
• MemoryFlushMiddleware: flush LLM call is capped by a 5-minute timeout (after subscribeOn, measuring execution not queue wait) so a hung model provider cannot tie up a boundedElastic worker.
• MemoryMaintenanceMiddleware: the entire maintenance run (file I/O + consolidation LLM) is capped by a 5-minute MAINTENANCE_TIMEOUT; the consolidation LLM itself is capped by a slightly shorter 4m30s CONSOLIDATION_TIMEOUT so logs can
  distinguish "consolidation timed out" from "entire maintenance timed out".
• Both middlewares snapshot the full RuntimeContext via RuntimeContext.builder().from(rc).build() before background execution, preserving all fields that filesystem operations or NamespaceFactory implementations may depend on while
  isolating attribute maps from concurrent mutations.

Regression tests

• onAgent_completesBeforeSlowFlushFinishes / completesBeforeSlowConsolidationFinishes: mock a slow LLM, assert the returned Flux completes in <1s while the LLM is still running in the background.
• onAgent_flushError_doesNotPropagateToFlux / maintenanceError_doesNotPropagateToFlux: verify errors in detached work never reach the caller.
• onAgent_afterClose_doesNotScheduleNewFlush / doesNotScheduleNewMaintenance: verify close() prevents new work; uses polling instead of fixed sleep for deterministic verification on slow CI.
• close_waitsForPendingFlush_thenReturns / waitsForPendingMaintenance_thenReturns: verify close() drains in-flight work before returning.
• close_disposesHungFlush_andReturnsPromptly / close_disposesHungConsolidation_andReturnsPromptly: verify close() disposes a hanging model and returns within CLOSE_AWAIT_TIMEOUT.

Checklist

• Code has been formatted with mvn spotless:apply

• All tests are passing (mvn test)

• Javadoc comments are complete and follow project conventions

• Related documentation has been updated

• Code is ready for review

  Fixes [Bug]:harnessAgent编排模式,Memory 的 concatWith 后置钩子阻塞了 call() Flux complete处理，导致编排长时间等待？是否有其他方案 #2276
  Fixes [Bug]:MemoryConsolidator blocks agent call via concatWith + .block() #2225

[codecov]

Codecov Report

❌ Patch coverage is 81.30841% with 20 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...arness/agent/middleware/MemoryFlushMiddleware.java	79.31%	11 Missing and 1 partial ⚠️
.../agent/middleware/MemoryMaintenanceMiddleware.java	83.67%	7 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[AgentScopeJavaBot]

🤖 AI Review

This PR fixes a real and impactful latency bug: MemoryFlushMiddleware and MemoryMaintenanceMiddleware previously used concatWith to append their LLM-backed work onto the returned Flux, forcing callers (e.g. ReActAgent.callInternal with takeLast(1)) to block until the memory flush/maintenance LLM call completed. The fix correctly switches to doOnComplete(() -> subscribe()) for genuine fire-and-forget behavior, adds AutoCloseable lifecycle management with bounded drain on shutdown, and includes thorough test coverage for the async semantics. The concurrency design (synchronized pending-set pattern, double-check on closed flag, timeout caps) is well thought-out. Two areas warrant attention: a subtle race in the subscribe-then-add-to-pending pattern, and missing stack traces in close-failure logs.

(inline comments could not be attached — line numbers fell outside PR hunks. See archived report.)

[AgentScopeJavaBot]

🤖 AI Review

This PR fixes a real and impactful latency bug: MemoryFlushMiddleware and MemoryMaintenanceMiddleware previously used concatWith to append their LLM-backed work onto the returned Flux, forcing callers (e.g. ReActAgent.callInternal with takeLast(1)) to block until the memory flush/maintenance LLM call completed. The fix correctly switches to doOnComplete(() -> subscribe()) for genuine fire-and-forget behavior, adds AutoCloseable lifecycle management with bounded drain on shutdown, and includes thorough test coverage for the async semantics. The concurrency design (synchronized pending-set pattern, double-check on closed flag, timeout caps) is well thought-out. Two areas warrant attention: a subtle race in the subscribe-then-add-to-pending pattern, and missing stack traces in close-failure logs.

(inline comments could not be attached — line numbers fell outside PR hunks. See archived report.)