Renderer and extension host OOM when running multiple concurrent Copilot chat/agent sessions (Windows)

Type: Bug

- VS Code Version: 1.135.0 (x64)
- OS Version: Windows 11, 10.0.26200
- Machine: 16 logical CPUs, 64 GB RAM
- Feature: agent mode + Copilot CLI chat sessions (`copilotcli` agent provider)
- Does this issue occur when all extensions are disabled?: Not verified — the repro requires Copilot Chat.

### Describe the bug

Running **two or more concurrent Copilot chat/agent sessions** (multiple windows, agent mode, long
transcripts with many rendered code blocks) reliably drives the renderer process to an
out-of-memory kill within a couple of hours. The machine has 64 GB RAM with >20 GB free at the time
of the crash, so this is a **per-process V8/renderer heap exhaustion, not system memory pressure**.

Two distinct crash signatures occur, often in the same session:

1. **Extension host** — `crashed with code -36861` (`exited with code: 4294930435`, i.e. `0xFFFF7003`),
reported by Electron as `reason: 'crashed'`.
2. **Renderer** — `renderer process gone (reason: oom, code: -536870904)` (`0xE0000008`, PartitionAlloc
`kOomExceptionCode`).

Once the renderer starts OOM-ing it enters a crash loop — three OOM kills in 70 seconds — and the
window has to be closed.

### Steps to Reproduce

1. Open 2+ VS Code windows, each with an active Copilot agent-mode chat session.
2. Drive long agent conversations that render many markdown code blocks and use editing/tool calls.
3. Leave the sessions open and keep interacting for ~1–2 hours.
4. Observe escalating `potential listener LEAK detected` errors in the renderer log, followed by
`renderer process gone (reason: oom, ...)`.

### Crash timeline

Excerpt from `main.log` across several sessions (PIDs redacted):

```
2026-08-23 12:57:28.636 [info] Extension host with pid exited with code: 4294930435, signal: unknown.
2026-08-23 12:57:28.640 [error] [UtilityProcess id: 4, type: extensionHost, pid: ]: crashed with code -36861 and reason 'crashed'
2026-08-25 12:28:35.539 [error] CodeWindow: renderer process gone (reason: crashed, code: -36861)
2026-08-26 09:42:03.153 [error] CodeWindow: renderer process gone (reason: crashed, code: -36861)
2026-08-26 12:15:01.259 [error] CodeWindow: renderer process gone (reason: oom, code: -536870904)
2026-08-26 13:17:38.679 [error] ptyHost terminated unexpectedly with code 4294930435
2026-08-26 14:14:02.368 [info] Extension host with pid exited with code: 4294930435, signal: unknown.
2026-08-26 14:14:02.369 [error] [UtilityProcess id: 3, type: extensionHost, pid: ]: crashed with code -36861 and reason 'crashed'
2026-08-26 14:34:14.132 [error] CodeWindow: detected unresponsive
2026-08-26 15:25:34.541 [warning] No ptyHost heartbeat after 6 seconds
2026-08-26 15:25:38.795 [error] CodeWindow: renderer process gone (reason: crashed, code: -36861)
2026-08-26 15:26:56.602 [error] CodeWindow: renderer process gone (reason: oom, code: -536870904)
2026-08-26 15:27:30.298 [error] CodeWindow: renderer process gone (reason: oom, code: -536870904)
2026-08-26 15:28:05.884 [error] CodeWindow: renderer process gone (reason: oom, code: -536870904)
2026-08-26 15:28:06.789 [error] [UtilityProcessWorker]: terminated unexpectedly with code 3221226505, signal: unknown
```

Preceding the crash, the extension host repeatedly goes unresponsive:

```
[info] Extension host (LocalProcess pid: ) is unresponsive.
[info] UNRESPONSIVE extension host: starting to profile NOW
[info] Extension host (LocalProcess pid: ) is responsive.
```

### Leak evidence — chat code-block editor pool

`renderer.log` shows `potential listener LEAK detected` growing steadily, and the count scales with
the number/duration of concurrent chat sessions. Counts of leak warnings per window log:

```
3, 6, 8, 13, 13, 14, 16, 16, 16, 25, 37, 39, 39, 47, 67, 149, 218
```

The heaviest windows (149, 218) are the ones that went on to OOM.

**Leak stack 1 — chat code block renderer / editor pool:**

```
[error] potential listener LEAK detected, popular: Error
at s.create (workbench.desktop.main.js:431:18494)
at _Tt._event (workbench.desktop.main.js:433:2398)
at new HE (workbench.desktop.main.js:2760:31045)
at s._createInstance (workbench.desktop.main.js:2098:2000)
at s.createInstance (workbench.desktop.main.js:2098:1440)
at CCe._pool._register.Rti.maxIdleSize [as _itemFactory] (workbench.desktop.main.js:2872:11122)
at Rti.get (workbench.desktop.main.js:2727:52291)
at CCe.get (workbench.desktop.main.js:2872:11257)
at xw.renderCodeBlock (workbench.desktop.main.js:2871:4543)
at Object.codeBlockRendererSync (workbench.desktop.main.js:2871:1632)
at o.codeBlockRendererSync.t.code (workbench.desktop.main.js:822:2688)
at s.parse (workbench.desktop.main.js:817:1812)
at i6.parse (workbench.desktop.main.js:818:5470)
at not.render (workbench.desktop.main.js:937:207936)
at v5.render (workbench.desktop.main.js:2723:44)
at new xw (workbench.desktop.main.js:2871:3169)
```

**Leak stack 2 — chat editing modified-file entries leaking `onDidChangeFilesAssociation`:**

```
[error] potential listener LEAK detected, popular: Error
at s.create (workbench.desktop.main.js:431:18494)
at qL._event [as onDidChangeFilesAssociation] (workbench.desktop.main.js:433:2398)
at lV.registerListeners (workbench.desktop.main.js:1980:12895)
at new lV (workbench.desktop.main.js:1980:12556)
at s._createInstance (workbench.desktop.main.js:2098:2000)
at NYe.doResolve (workbench.desktop.main.js:5392:8693)
at Wht.doCreateReferencedObject (workbench.desktop.main.js:1980:28085)
at Wht.acquire (workbench.desktop.main.js:431:3513)
at PTe.acquire (workbench.desktop.main.js:431:3832)
at Bht.createModelReference (workbench.desktop.main.js:1980:30381)
at mJ._createModifiedFileEntry (workbench.desktop.main.js:3475:38300)
at mJ._getOrCreateModifiedFileEntry (workbench.desktop.main.js:3475:37372)
at mJ._initEntries (workbench.desktop.main.js:3475:35210)
at async mJ._init (workbench.desktop.main.js:3475:24485)
```

Both stacks point at chat surfaces creating text-model references / editor instances per rendered
item and registering listeners that outlive the chat turn. With N concurrent sessions the growth is
roughly N× per unit time, which matches the observed correlation between session count and
time-to-OOM.

### Analysis

- `-536870904` = `0xE0000008` = `partition_alloc::win::kOomExceptionCode` — the renderer really did
exhaust its heap.
- `-36861` = `0xFFFF7003` — V8/Node heap exhaustion on a different termination path, so Electron
labels it `reason: 'crashed'` rather than `'oom'`.
- Free system RAM at crash time was ~22 GB of 64 GB, and commit had ~49 GB free, so this is not
system-wide memory pressure.

### Related issues

- microsoft/vscode#265548 — Copilot + very large file → `-36861` (open)
- microsoft/vscode#303526 — renderer repeatedly crashes with `reason: oom, code: -536870904` during
heavy workspace + chat sessions (open)
- microsoft/vscode#170407 — extension conflict → `-36861` (open)
- microsoft/vscode#332408 — **merged**, fixes extension host OOM when listing Copilot CLI sessions by
throttling concurrent session deserialization (milestone 1.136.0). This likely addresses the
*extension host* half of this report but not the *renderer* OOM.

### Notes on attachments

No Crashpad minidump is attached: `~/.vscode/argv.json` had `"enable-crash-reporter": false` and
`telemetry.telemetryLevel` was `"off"`, so no dump was produced for these crashes. The crash reporter
has now been enabled and a minidump will be attached once the crash reproduces.

Windows Error Reporting did record earlier `Code.exe` faults on this machine:

- `EventType=BEX64`, `Exception Code=c0000409` (STATUS_STACK_BUFFER_OVERRUN / fastfail),
fault module `Code.exe`, app version 1.130.0.0
- `EventType=AppHangB1`, `Hang Signature a67e`, app versions 1.130.0.0 / 1.131.0.0

All logs above have been redacted of usernames, machine names, file paths, workspace names, chat
content and session identifiers.