Agent Host session load causes child-chat subscription storm and memory spike

[sandy081]

Summary

Loading and reopening a large Agent Host session can generate thousands of duplicate child-chat subscriptions, causing a sharp renderer/transport memory spike and hundreds of MB of AHP traffic.

Reproduction

Using the session Agent Host session migration (39 subagent chats):

1. Open the session and wait for history/Changes to settle.
2. Open another session.
3. Reopen the original session.

Controlled runs with isolated copies of the same profile:

Build	AHP bytes	Child-chat snapshots
43c9cf468f20 (good Insiders)	42.3 MB	78
fc62850c022b (bad Insiders)	343.2+ MB	1,005 retained after rotation

Bisect and root cause

A source-level bisect identified commit 38220eab177286ceaf47098a843396d351464247 (#331079):

Source revision	AHP bytes	Child-chat snapshots
Parent 30ad9a507bd	42.9 MB	85
38220eab177	367.7 MB	864

The triggering change is the new synchronous external-changes scope check in AgentFeedbackService._isFileInSessionScope:

if (session.externalChanges?.get().some(file => isEqual(file.uri, resourceUri))) {
    return true;
}

externalChanges is subscription-backed. Calling .get() while it is unobserved recomputes its dependencies for each visible changed-file editor, transiently acquiring and releasing all child-chat subscriptions. The target session has 39 child chats, so repeated file scope checks create a large overlapping subscribe/unsubscribe storm.

Line-level verification on 38220eab177:

• Removing only the new externalChanges.get() check: 43.0 MB / 87 child snapshots (baseline restored).
• Removing the separate new subagent annotation-restore await instead: 375.1 MB / 866 child snapshots (still reproduces).

Confidence in this causal attribution is very high; the immediate-parent, child, and single-expression A/B tests were all run against the same profile and automated scenario.

[benibenj]

#332521