[alexcstark]

The lax6e server's memory grew from 0.6 GB to 6.8 GB over six days of uptime, almost all of it native heap ([heap] 5.4 GB; the V8 heap was 1.35 GB). Process spawns copy the server's page tables, so that growth made every git spawn about 10× slower and drove the event-loop stalls that left uploads and threads stuck.

Cause
The Cursor chat path (CursorAgentSdk.ts) passes no local.store. In that case the SDK opens a default SqliteLocalAgentStore per workspace and caches it for the life of the process (a module-level Map in @cursor/sdk 1.0.35, unchanged in 1.0.38). Every agent created or resumed through it keeps its checkpoint blobs in memory as Uint8Array copies. Resuming an agent loads all blobs it ever wrote. Only deleting the agent evicts them, and agent.close() doesn't. Text generation already avoided this by owning and disposing its store (pingdotgg#19310). The chat path didn't.

Evidence (lax6e, upstream 0.1.0, measured 2026-11-04)

• The old server held 91 Cursor agent store.db handles open. Exactly 91 agent stores had been written since that server started, holding 1.83 GB of blob data.
• Hourly RSS growth fits about 4.4 MB per Cursor run vs about 0.6 MB per Claude or Codex run (correlation 0.44 vs 0.22), over 69 intervals.
• Offline reproduction with the real SDK 1.0.35: seed 3 agents with 100 MB of checkpoints each, then resume and close them.
  • Default store: arrayBuffers 315 MB, still retained after close and GC.
  • Owned store disposed on close: arrayBuffers 1 MB.

Fix
The runner opens a SqliteLocalAgentStore per session when the caller passes none, and uses it for the agent and its run lookups (listRuns/cancelRun, which would otherwise reopen the default store). It disposes the store when the session closes, or when opening the agent fails. SqliteLocalAgentStore.open({ workspaceRef: cwd }) resolves the same default state root the SDK uses, so existing agents still resume; the reproduction resumed seeded agents through it.

Proof

• CursorAgentSdk.test.ts: the close-order cases now also expect the store to be passed to the SDK and disposed after agent.close, including when close fails. The resume-retry cases expect the store in run lookups. With origin/main's runner, 4 of 7 fail; with this change, 7/7 pass.
• provider-cursor tests: 131/131. tsc-rs passes for provider-cursor and apps/server. vp fmt --check and vp lint pass on the changed files.
• Not yet verified on lax6e: its new build (b603e5bdfb) predates this change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01StkPGjC6GMoMkcwbiPzuUv

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.9 KiB	20.9 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.2 KiB	21.2 KiB	0 B (0.0%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: 99c96b0 · PR result: f96f083 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.