[alexcstark]

On Linux, Node starts children with fork() (libuv 1.52 still does; posix_spawn is Apple-only), which copies the parent's page tables on the main thread. The cost grows with the server's memory. A server that had run 6 days and grown to 6.8 GB blocked the main thread 64–223 ms per spawn. With about 2 git spawns a second, that was 22% of wall-clock time, and HTTP/WebSocket handling froze for 2–11 s: image uploads sat at 0%, and requests took up to 11.5 s.

Evidence (lax6e, 2026-11-04)

• perf: 73% of samples were the main thread in the kernel, mainly copy_present_pte (30%) and zap_pte_range (28%), reached through uv_spawn → _Fork → clone → dup_mmap.
• bpftrace: 222 forks in 120 s (218 git), each blocking 64–223 ms, 26.7 s in total.
• Spawn cost vs parent memory (Node on lax6e): 46 MB → 0.7 ms, 1.1 GB → 25 ms, 3.3 GB → 68 ms.

Fix

• Children start in a small helper process (t3 __spawn-helper, a fast path in bin.ts that loads only Node built-ins), launched once at boot while the server is small.
• The server sends spawn requests over IPC (advanced serialization). The helper relays each child's output back, and stdin writes the other way.
  • Bytes are relayed rather than pipe handles passed: a passed pipe loses whatever the helper already read, and its hand-over races a fast exit. The new concurrency test caught exactly that.
• @effect/platform-node-shared gets a small patch that exposes the spawn call (layerWithSpawn), so Effect's ChildProcessSpawner and everything on it (ProcessRunner, git, provider CLIs) is reused unchanged. server.ts swaps it in, on Linux only.
• Children get the server's current env/cwd, not the helper's startup snapshot.
• The helper's heap is capped (NODE_OPTIONS), because V8 collects lazily and an uncapped helper drifts upward.
• If the helper dies, its in-flight children are ended and reported, not left hanging. If it can't launch, spawns fall back to running directly.

Proof

• SpawnHelper.test.ts (6, it.live, through the real bin.ts fast path):
  • children run under the helper (their parent PID isn't the server's);
  • output, stderr and exit codes come back, stdin writes through, and timeouts kill the child;
  • 40 concurrent children plus a 1 MB output all arrive intact;
  • a dead helper ends its children and a new helper starts;
  • a helper that can't launch falls back to direct spawns.
• Each of these was made to fail against an injected defect: spawning locally, dropping the child on exit, skipping the in-flight cleanup, and removing the launch fallback. processRunner.test.ts 19/19.
• Linux, real linux-x64 single executable from this branch, parent holding 3.1 GB: main-thread time per spawn 63.12 ms direct → 0.06 ms through the helper; worst event-loop stall 82 ms → 11 ms. The helper held 136 MB of anonymous memory (1.6 MB of page tables) and plateaued at about 213 MB RSS over 4,000 spawns.
• tsc-rs passes for apps/server; vp fmt --check and vp lint are clean on the changed files. The lockfile is checked with pnpm install --frozen-lockfile.

Not covered: spawns that bypass ChildProcessSpawner still fork the server. That's the Claude SDK's CLI launch, node-pty for opening terminals, and some CLI-only commands, together about 4 of 222 measured spawns. macOS and Windows behave as before. Related: #99 (Cursor agent-store memory growth) and #100 (shared in-flight git reads).

🤖 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: 54f74c8 · Source CI: failure

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.