sandbox_child_nproc_limit` (PR #1497) merged 2026-05-27 but absent from all release binaries through v0.0.82

[strongchang]

Agent Diagnostic

• Investigated with Claude (agent-driven diagnosis on a 19-sandbox Docker-driver fleet)
• Tested OpenShell v0.0.72 (CLI/gateway/sandbox from GitHub release assets, x86_64-unknown-linux-gnu)
• Checked latest releases: downloaded release assets v0.0.74/76/78/80/82 and scanned both
  openshell-gateway and openshell-sandbox binaries with strings — no
  sandbox_child_nproc_limit / nproc_limit / RLIMIT_NPROC-related config token in any of them
  (only Z3 "inprocess" noise in the gateway)
• Cross-checked GitHub: PR Make sandbox child nproc limit configurable #1497 merged into main 2026-05-27 (commit 9bfcad4); the
  branch_commits/tag-containment view lists it as reachable from v0.0.51+, contradicting the
  binary contents
• Functional test on v0.0.72: added sandbox_child_nproc_limit = 4096 under
  [openshell.drivers.docker] and [openshell.gateway] — gateway starts either way, but a bogus
  test field is also silently accepted in both tables, so acceptance proves nothing
• Recreated a sandbox from scratch after the config change: in-sandbox sessions still show
  Max processes 512 in /proc//limits
• Searched existing issues for sandbox_child_nproc_limit / nproc: found fix(sandbox): use cgroup pids controller instead of RLIMIT_NPROC on K8s #1390 / fix(sandbox): use cgroup pids controller instead of RLIMIT_NPROC on K8s #1325
  (shared-UID fleet reports) and PR Make sandbox child nproc limit configurable #1497 itself, but no issue about the feature missing from
  release binaries
• Cannot upgrade to a fixed version because no published release contains the feature (verified
  through v0.0.82)

Description

Summary

PR #1497 ("Make sandbox child nproc limit configurable", merged into main on 2026-05-27 as commit 9bfcad4) adds the sandbox_child_nproc_limit setting to lift the hardcoded RLIMIT_NPROC=512 for shared-UID fleet deployments. GitHub's tag-containment view lists the commit as reachable from tags v0.0.51 onward, but the published release binaries do not contain the feature.

Evidence

1. String scan: strings openshell-gateway | grep -i nproc (and the same for openshell-sandbox) on release assets v0.0.72, v0.0.74, v0.0.76, v0.0.78, v0.0.80, v0.0.82 (*-x86_64-unknown-linux-gnu.tar.gz) shows no nproc_limit / RLIMIT_NPROC / sandbox_child_nproc token — only Z3 inprocess noise in the gateway.
2. Functional test on v0.0.72 (Docker driver): adding sandbox_child_nproc_limit = 4096 under [openshell.drivers.docker] or [openshell.gateway] is silently accepted (note: a bogus test field is also silently accepted in both tables, so unknown TOML fields appear to be ignored), and a freshly recreated sandbox still shows Max processes 512 for all in-sandbox sessions.

Expected

A published release ships PR #1497, so sandbox_child_nproc_limit = 4096 (or 0 to defer to the container cgroup pids limit) takes effect for newly created sandbox sessions.

Questions

1. Which release is expected to ship PR Make sandbox child nproc limit configurable #1497?
2. Are the GitHub release assets built from a branch other than the tagged main history? (Tag-containment view vs. binary content disagree.)
3. Is the silent acceptance of unknown fields in [openshell.gateway] / [openshell.drivers.docker] intended? v0.0.54 notes say unknown fields in nested gateway config tables are rejected.

Context / impact

19-sandbox single-host fleet (Docker driver); all sandbox inner sessions share one UID. At ~30–50 threads per sandbox the fleet saturates the hardcoded 512 and any new fork fails with EAGAIN (see also #1390 / #1325). We currently mitigate with a periodic prlimit --nproc=8192 systemd timer and would like to retire it.

Environment: openshell 0.0.72 (CLI/gateway/sandbox from GitHub release assets), Docker driver, Ubuntu host, NemoClaw v0.0.81+ (source main).

Reproduction Steps

Summary

PR #1497 ("Make sandbox child nproc limit configurable", merged into main on 2026-05-27 as commit 9bfcad4) adds the sandbox_child_nproc_limit setting to lift the hardcoded RLIMIT_NPROC=512 for shared-UID fleet deployments. GitHub's tag-containment view lists the commit as reachable from tags v0.0.51 onward, but the published release binaries do not contain the feature.

Evidence

1. String scan: strings openshell-gateway | grep -i nproc and same for openshell-sandbox from release assets v0.0.72, v0.0.74, v0.0.76, v0.0.78, v0.0.80, v0.0.82 (*-x86_64-unknown-linux-gnu.tar.gz) show no nproc_limit / RLIMIT_NPROC / sandbox_child_nproc token (only Z3 inprocess noise in the gateway).
2. Functional test on v0.0.72 (Docker driver): adding sandbox_child_nproc_limit = 4096 under [openshell.drivers.docker] or [openshell.gateway] is silently accepted (note: unknown TOML fields appear to be silently ignored in both tables — a bogus test field is also accepted), and a freshly recreated sandbox still shows Max processes 512 for all in-sandbox sessions.

Questions

1. Which release is expected to ship PR Make sandbox child nproc limit configurable #1497?
2. Are the GitHub release assets built from a branch other than the tagged main history? (Tag-containment view vs. binary content disagree.)
3. Is the silent acceptance of unknown fields in [openshell.gateway] / [openshell.drivers.docker] intended? v0.0.54 notes say unknown fields in nested gateway config tables are rejected; that made our config validation attempts inconclusive.

Context / impact

19-sandbox single-host fleet, all sandbox inner sessions share one UID; at ~30–50 threads per sandbox the fleet saturates the hardcoded 512 and any new fork fails with EAGAIN (see also #1390 / #1325). We currently mitigate with a periodic prlimit --nproc=8192 systemd timer and would like to retire it.

Environment: openshell 0.0.72 (CLI/gateway/sandbox from GitHub release assets), Docker driver, Ubuntu host, NemoClaw v0.0.81+ (source main).

Environment

• OS: Ubuntu 24.04 (Nutanix VM)
• Docker: Docker Engine 29.1.3 (Docker driver)
• OpenShell: v0.0.72 (openshell 0.0.72; CLI/gateway/sandbox binaries from GitHub release assets)
• Latest release checked: yes — downloaded and string-scanned release assets up to v0.0.82; feature absent in all of them
• Possible duplicates checked: yes — found fix(sandbox): use cgroup pids controller instead of RLIMIT_NPROC on K8s #1390 / fix(sandbox): use cgroup pids controller instead of RLIMIT_NPROC on K8s #1325 (shared-UID nproc saturation reports) and PR Make sandbox child nproc limit configurable #1497 itself; no existing issue about the feature missing from release binaries
• Agent stack: NemoClaw v0.0.81+ (source main) with Hermes Agent v0.18.0, 19-sandbox fleet

Logs

# String scan of v0.0.82 release assets (same result for v0.0.74–v0.0.80 and v0.0.72):
$ strings openshell-gateway | grep -ioE "[a-z_]*nproc[a-z_]*" | sort -u
inprocess
inprocessing
(only Z3 solver noise; no config token)
$ strings openshell-sandbox | grep -ioE "[a-z_]*nproc[a-z_]*" | sort -u
(no matches)
$ grep -ac "nproc_limit" openshell openshell-gateway openshell-sandbox
openshell:0
openshell-gateway:0
openshell-sandbox:0

# Functional test on v0.0.72: config accepted, no effect
$ grep -n nproc openshell-gateway.toml
32:sandbox_child_nproc_limit = 4096
$ pgrep -af openshell-gateway   # gateway starts fine with the field present
2592034 openshell-gateway[nemoclaw=nemoclaw;port=8080]

# Freshly recreated sandbox after the config change:
$ docker exec <sandbox> sh -c 'for p in $(pgrep -f "hermes|socat"|head -3); do grep -i processes /proc/$p/limits; done'
Max processes             512                  512                  processes
Max processes             512                  512                  processes
Max processes             512                  512                  processes

# Bogus-field control test (shows unknown fields are silently ignored, both tables):
$ grep -n bogus openshell-gateway.toml
33:bogus_test_field_xyz = 1
$ pgrep -af openshell-gateway   # still starts
2592100 openshell-gateway[nemoclaw=nemoclaw;port=8080]

Agent-First Checklist

• I pointed my agent at the repo and had it investigate this issue
• I loaded relevant skills (e.g., debug-openshell-cluster, debug-inference, openshell-cli)
• I checked the latest OpenShell release and either reproduced the issue there or explained why I cannot upgrade/test it
• I searched existing issues for possible duplicates or explained why I could not
• My agent could not resolve this — the diagnostic above explains why