[Perf] Linux: Grok sessions walk /proc every 25 ms on the server event loop, although the cgroup lease already handles teardown
Maintainers usually reply within 1 day
Nobody has claimed this yet.
Assessment
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Newbie friendliness
- 78/100
- Issue type
- Bug
- Clarity
- Clearly specified
- Activity status
- Active
- Tech stack
- linux, typescript
- Domain
- backend, performance
Research direction
Start in packages/provider-acp/src/server/AcpSessionRuntime.ts at the cgroup lease setup and continuous ownership observer. Check the teardown branches to confirm the ownership ledger is only used without a lease, then add or update the counting-controller test described in the issue. Done means the observer captures the root once under a lease, idle polling stops, and closing the session empties the lease.
Written by the indexing model from the issue text.
Description
The problem
On Linux, when a Grok session has a cgroup v2 lease (t3-acp-*, the normal case under systemd),
AcpSessionRuntime still runs observePosixOwnershipLedgerContinuously for the whole life of the
session. Every 25 ms it walks the agent's process tree in /proc on the server's main thread. Its
result (the ownership ledger) is read only by terminatePosixProcessTree, which runs only when
there is no lease. Under a lease, teardown is cgroup.kill plus the root identity captured
once at spawn. The walk is unused work on the event loop.
Claude and Codex sessions have no equivalent. They are spawned detached and torn down with one
kill(-pgid) (ClaudeAdapterV2.ts:320, provider-core/src/server/processGroup.ts). Of the
providers in this repo, only provider-grok sets ownDescendantProcessGroups.
Measured impact (T3 v0.1.1, fe27e8e3, headless t3 serve, Ubuntu 26.04, cgroup v2)
Controlled test on one server process. Each Grok session was a task whose only action was
sleep 300. Per session: grok agent --always-approve stdio (46 threads) and
t3 acp-mcp-bridge (11 threads) in their own lease. Main thread measured with
/proc/<pid>/task/<pid>/stat and bpftrace raw_syscalls filtered to the main thread.
| Grok sessions | main thread CPU | main-thread /proc opens/s |
getdents64/s |
kernel time in these calls |
|---|---|---|---|---|
| 0 (Claude and Codex sessions live) | 6.8% of a core | 0 | 0 | — |
| 1, idle | 15.6% | 2,415 | 284 | 5.6% of a core |
| 3, idle | 24.6% | 6,390 | 759 | 12.9% of a core |
| 3, turns finished, sessions still open | 21.0% | 5,858 | 395 | 11.6% of a core |
| 0, after the 30 min idle release | 10.2% | 0 | 0 | 0.03% of a core |
- An idle session runs the walk at the full rate: about 40 passes/s × (one
task/<tid>/children
read per agent thread +statandreadlink exeper process). The cost scales with sessions and
with the agent's thread count. - It lasts beyond the turn. Idle provider sessions are released after
DEFAULT_IDLE_TIMEOUT_MS
(30 min, up to 4 h when pinned;ProviderSessionManager.ts:56-57). The test sessions were
released 30 min after their last activity, and the/proctraffic stopped with them. - Under load the walk runs slower (we saw about 5 passes/s on a busy server), because it shares
the main thread. Then it shows up as event-loop delay instead of CPU.
This is the measurement requested when #19255 was closed.
Root cause
packages/provider-acp/src/server/AcpSessionRuntime.ts @ fe27e8e3:
- Grok enables descendant ownership on Linux:
provider-grok/src/server/acpSupport.ts#L90-L101. - That creates a cgroup v2 lease and joins the agent to it before
exec:
L1402-L1421. - Without checking for the lease, it then forks the continuous observer for the session's
lifetime:
L1493-L1534,
loop at
L739-L758
(25 ms target, insideEffect.sync, so each pass blocks the event loop). - With a lease, graceful teardown
(L1596-L1609)
and forced teardown
(L1624-L1641)
use only the root identity from the initial capture andterminateLinuxCgroupLease. The ledger
is consumed only byterminatePosixProcessTree, reached whenlinuxCgroupLease === undefined.
Proposed fix
Start the continuous observer only when there is no lease. The initial root capture stays, so
signalLinuxCgroupRootTerm still targets the right pid:starttime. Sessions without a lease
(cgroup v1, no delegated cgroup) behave exactly as today.
- yield* observePosixOwnershipLedgerContinuously({
- childQueues: posixOwnershipChildQueues,
- controller: posixController,
- frontier: posixOwnershipFrontier,
- ledger: posixOwnershipLedger,
- root: posixOwnershipRoot,
- rootPid: Number(child.pid),
- }).pipe(Effect.forkIn(runtimeScope));
+ // Under a cgroup lease, teardown is cgroup.kill and reads only the root captured above.
+ if (linuxCgroupLease === undefined) {
+ yield* observePosixOwnershipLedgerContinuously({
+ childQueues: posixOwnershipChildQueues,
+ controller: posixController,
+ frontier: posixOwnershipFrontier,
+ ledger: posixOwnershipLedger,
+ root: posixOwnershipRoot,
+ rootPid: Number(child.pid),
+ }).pipe(Effect.forkIn(runtimeScope));
+ }
Test: with a counting fake posixProcessTreeController and a temp-directory lease, assert that
childPidsOf runs once (the initial capture) over an idle window of 10+ intervals, and that closing
the session still empties the lease. #19255 had this test (8 reads instead of 1 on main).
Expected result
Under a lease: 0 /proc opens from idle Grok sessions, the same as Claude and Codex, and about
5–9% of a core of main-thread time back per open Grok session.
- Dominant language
- TypeScript
- Stars
- 24.8k
- Forks
- 6.4k
- Avg merge
- 1h 22m
- Merged PRs (30d)
- 18
Getting set up
Starts the project's dev container in your browser, under your own GitHub account.
- No Dockerfile or Docker Compose file
- Has a pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from pingdotgg/t3code
-
MCP endpoint matches the Authorization scheme case-sensitively and logs the same reason for "absent" and "wrong scheme" (ChatGPT connector gets missing_bearer_token)Possibly taken A pull request linked to this issue is open or already merged. Openbug needs-triage
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 77/100
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 77/100
Maintainers usually reply within 1 day
-
bug needs-triage
Difficulty 2/5 1-3 hours Newbie friendliness 74/100
pingdotgg/t3code#23175 · 2 comments ·
Maintainers usually reply within 1 day
-
[Bug]: preview_evaluate accepts returnByValue:false but returns no usable object referencePossibly taken @juliusmarminge claimed this 1 day ago. Openbug needs-triage
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
Maintainers usually reply within 1 day
All issues in pingdotgg/t3code
Similar issues
-
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
Maintainers usually reply within 1 day
-
bug
Difficulty 1/5 Under an hour Newbie friendliness 76/100
logto-io/logto#9837 · 1 comment ·
Maintainers usually reply within 1 day
-
community first-timers-only good first issue hacktoberfest help wanted low hanging fruit up-for-grabs
Difficulty 1/5 Under an hour Newbie friendliness 76/100
lingdojo/kana-dojo#32964 · 1 comment · 5 reactions ·
Maintainers usually reply within 1 day
-
Documentation status: needs triage
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
mastra-ai/mastra#28194 · 1 comment ·
Maintainers usually reply within 1 day
-
priority-low refactor-3
Difficulty 2/5 1-3 hours Newbie friendliness 84/100
Maintainers usually reply within 3 days