[Bug][2.5.0 Regression] Workflow replay mixes events across invocation_ids and breaks consecutive HITL runs

[jordi-ydteam]

ADK 2.5.0 regression: workflow replay mixes events across invocations

I am reporting a reproducible workflow replay regression introduced in Google ADK 2.5.0.

Summary

The first execution of a human-in-the-loop workflow can pause, resume and complete successfully. Starting the same workflow again in the same ADK session can fail with:

RuntimeError: Replay divergence detected: Timed out waiting for sequence key 'fake_planner@1' to be unblocked.

Version comparison

google-adk==2.4.0: PASS
google-adk==2.5.0: FAIL
google-adk==2.5.0 using a new session for the second run: PASS

Suspected regression

The behavior appears to have been introduced by commit:

64a7448fc6e05f37e6dedadd6c0e65bc19264057

This commit optimized workflow rehydration by building event indexes.

In ADK 2.4.0, replay sequence scanning explicitly ignored events whose invocation_id differed from the current invocation.

In ADK 2.5.0:

• node state reconstruction remains scoped to the current invocation_id;
• _build_event_index() receives invocation_id, but does not appear to use it;
• sequence scanning consumes indexed events from the entire session;
• deterministic sequence keys such as fake_planner@1 are reused by subsequent invocations.

As a result, events from an earlier invocation may be included in the replay sequence barrier for the current invocation, causing the reused sequence key to remain blocked until the replay timeout.

Expected behavior

Each workflow invocation should reconstruct and scan its replay sequence using only events belonging to its own invocation_id.

Starting the same HITL workflow twice in one session should work.

Actual behavior

The second HITL execution in the same session can fail with replay divergence.

Suggested fix

Either:

1. filter events by invocation_id when building the event index;
2. filter them inside _scan_sequence() as in ADK 2.4.0; or
3. key the replay index by both invocation_id and workflow path.

A regression test should execute the same interruptible workflow twice using the same session and assert that both invocations complete successfully.

I have temporarily pinned the application to google-adk==2.4.0, where the reproduction passes.

Could a maintainer confirm whether this is a known regression and convert this discussion into an issue?