fix(core): offload agent execution setup from caller thread - #2464

Review: Approve ✅

Clean fix for #2462 — correctly offloads potentially blocking beforeAgentExecution setup from the caller/event-loop thread onto Schedulers.boundedElastic().

Analysis

Problem: runLifecycleBody called beforeAgentExecution synchronously, which triggers AgentStateStore.get(). Backed by Redis/MySQL/filesystem, this blocks the caller thread — problematic in event-loop style callers.

Solution: Wrap beforeAgentExecution + bindRequestState in Mono.fromCallable(...).subscribeOn(Schedulers.boundedElastic()). The PreparedExecution record cleanly carries the scope out of the callable.

Observations

1. Correctness ✅ — The scope == null path is preserved: PreparedExecution wraps null, and the flatMap branch skips contextWrite when scope is null. Semantically identical to the original.
2. Scheduler choice ✅ — boundedElastic is the right pool for blocking I/O (Redis/MySQL/fs). Does not steal from the computation pool.
3. Per-key serialization gate ✅ — Unchanged; the existing gate still serializes correctly since subscribeOn only affects the fromCallable segment.
4. Tests ✅ — 105 lines of new tests covering the threading behavior. Good coverage.
5. Scope — Minimal, focused change. 2 files, +140/-18.

Note

CLA is not yet signed — author needs to complete the CLA check before merge.

Review: Offload agent execution setup from caller thread

Verdict: Approved ✅ (pending CLA signature)

Good fix for the event-loop blocking issue reported in #2462. The approach is sound and well-targeted.

Analysis

• Root cause correctly identified — ReActAgent.beforeAgentExecution may synchronously load AgentState through AgentStateStore#get, which can be backed by Redis/MySQL/COS/filesystem. Running this on the caller/event-loop thread is a blocking risk.
• Solution — wrapping beforeAgentExecution + bindRequestState in Mono.fromCallable(...).subscribeOn(Schedulers.boundedElastic()) correctly offloads the blocking work. The PreparedExecution record cleanly bundles the result.
• Serialization gate preserved — the per-key lock is maintained outside the deferred block, so concurrent calls for the same session still serialize correctly.
• Null scope case handled — when beforeAgentExecution returns null, the contextWrite is skipped as before.
• Test coverage — 105-line test validating the offload behavior.

Blocker

• ⚠️ CLA not signed — the contributor license agreement check is pending. This needs to be resolved before merge.

LGTM once CLA is signed.

🤖 AI Review

Refactors AgentBase.runLifecycleBody() to offload blocking beforeAgentExecution() hook onto Schedulers.boundedElastic(), wrapping in PreparedExecution record to survive Reactor's null-to-empty semantics. Note: PR title says "checkpoint-based fault recovery" but diff is about lifecycle offloading.

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

🤖 AI Review

Refactors AgentBase.runLifecycleBody() to offload blocking beforeAgentExecution() hook onto Schedulers.boundedElastic(), wrapping in PreparedExecution record to survive Reactor's null-to-empty semantics. Note: PR title says "checkpoint-based fault recovery" but diff is about lifecycle offloading.

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

🤖 AI Review

Refactors AgentBase.runLifecycleBody() to offload blocking beforeAgentExecution() hook onto Schedulers.boundedElastic(), wrapping in PreparedExecution record to survive Reactor's null-to-empty semantics. Note: PR title says "checkpoint-based fault recovery" but diff is about lifecycle offloading.

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