v2026.4.26: Gateway busy-loops on bundled openai SDK directory walk; stops accepting connections

Summary

After upgrading to v2026.4.26, the gateway boots, briefly serves /health, and within minutes becomes unresponsive while pegging one CPU core. strace shows an inescapable directory walk over /usr/lib/node_modules/openclaw/node_modules/openai/resources/{audio,beta,chat,conversations,containers,evals,fine-tuning,graders,realtime,responses,skills,uploads,vector-stores,webhooks}, repeating in the same alphabetical order indefinitely. The plugin runtime cache size freezes — no forward progress. Listening socket Recv-Q fills with queued connections that are never accept()ed.

Rolling back to v2026.4.24 restores normal operation.

Environment

• OpenClaw: 2026.4.26 (be8c246), installed via sudo npm install -g [email protected]
• Prior known-good: 2026.4.24 (cbcfdf6)
• Node: 22.22.2
• OS: Linux 6.8.0-106-generic (Debian-family, systemd-managed gateway)
• Filesystem: ext4 on /dev/sda1, no exotic mounts
• Plugins enabled: memory-core, memory-wiki, telegram, plus bundled defaults (acpx, bonjour, browser, device-pair, phone-control, talk-voice). qqbot pinned enabled: false. Feishu/whatsapp inherit gating from prior config (require >=2026.4.25).
• One custom provider: models.providers.custom-opencode-go-extras proxying through https://opencode.ai/zen/go/v1 (OpenAI-compat endpoint).
• 5 agents, 3 cron jobs, no MCP servers, no TTS configured.

Steps to reproduce

1. From a known-good v2026.4.24 install, take a config backup: cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak.
2. sudo npm install -g [email protected]
3. openclaw doctor --fix — minor changes only on this host: removed retired agents.defaults.llm.idleTimeoutSeconds block and archived 1 orphan transcript. Pre-staged ~157 MB of plugin runtime deps under ~/.openclaw/plugin-runtime-deps/openclaw-2026.4.26-<hash>/.
4. sudo systemctl restart openclaw
5. Wait until log shows [gateway] http server listening (3 plugins: memory-core, memory-wiki, telegram; 160.4s) and [gateway] ready. curl http://127.0.0.1:18789/health returns 200 in ~30ms.
6. Continue waiting 1–3 minutes. /health starts timing out; openclaw cron list returns gateway timeout after 30000ms.

Expected behavior

Gateway remains responsive after the [gateway] ready log line, accepts incoming WebSocket connections from the local CLI, and continues to background-stage plugin runtime dependencies without stalling the event loop.

Actual behavior

• /health and CLI WS calls intermittently fail, then permanently time out.
• ss -ltnp shows Recv-Q > 0 on the listening socket — connections queued, never accepted.
• ps: gateway PID at ~55–112% CPU, growing CPU-time, climbing RSS (peaked at 1.3 GB during the incident).
• Plugin runtime cache size frozen at 369 MB across 10+ second observations — write activity occurs but appears to overwrite the same files.

Diagnostic evidence

strace -p <pid> -c (4-second sample)

% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 21.98    0.117403         598       196           statx
 19.93    0.106469         335       317           read
 16.80    0.089729         509       176           openat
 16.15    0.086274          89       969           access
 12.22    0.065258         372       175           close
  2.62    0.013981        1165        12        12 link
  2.40    0.012822         320        40           fstat
  1.94    0.010346         862        12           unlink
  1.73    0.009221         614        15        14 mkdir
  1.19    0.006361         530        12        12 fchown
  0.90    0.004833         402        12           chmod
  0.75    0.003990         332        12           fchmod
  0.71    0.003809         317        12           copy_file_range
  0.56    0.002972         247        12           ftruncate
...
total    0.534178         263      2030        38

Heavy directory traversal (statx, openat, read, access) plus per-iteration write ops (link, unlink, mkdir, chmod, copy_file_range, ftruncate). The write counts repeat on each sample — not draining.

strace -p <pid> -e trace=openat (representative slice)

openat(AT_FDCWD, "node_modules/openai/resources/beta/threads/runs", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
openat(AT_FDCWD, "node_modules/openai/resources/beta/realtime", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
openat(AT_FDCWD, "node_modules/openai/resources/beta/chatkit", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
openat(AT_FDCWD, "node_modules/openai/resources/audio", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
openat(AT_FDCWD, "node_modules/openai/resources", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
openat(AT_FDCWD, "node_modules/openai/resources/webhooks", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
openat(AT_FDCWD, "node_modules/openai/resources/vector-stores", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
... [continues with skills, responses, realtime, graders, fine-tuning, evals, conversations, containers, chat, beta]
... then re-enters from the top:
openat(AT_FDCWD, "node_modules/openai/resources/beta/threads/runs", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
openat(AT_FDCWD, "node_modules/openai/resources/beta/realtime", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 36
...

The walk traverses the same 14 top-level subdirectories of openai/resources/ (plus their nested beta/*, chat/*, containers/*, evals/*, fine-tuning/*) in identical order, repeating without termination.

Gateway log (relevant)

2026-04-28T12:26:41.871+07:00 [gateway] loading configuration…
2026-04-28T12:26:42.235+07:00 [gateway] starting...
2026-04-28T12:29:20.193+07:00 [gateway] starting HTTP server...
2026-04-28T12:29:22.682+07:00 [gateway] agent model: custom-opencode-go-extras/deepseek-v4-flash
2026-04-28T12:29:22.925+07:00 [gateway] http server listening (3 plugins: memory-core, memory-wiki, telegram; 160.4s)
2026-04-28T12:29:23.752+07:00 [gateway] ready
2026-04-28T12:29:24.080+07:00 [telegram] [default] starting provider (@AntonioCOCBot)
2026-04-28T12:29:25.694+07:00 [telegram] menu text exceeded the conservative 5700-character payload budget; shortening descriptions to keep 57 commands visible.
[no further log entries — gateway hangs while telegram polling is in flight]

No error or warning is emitted while the gateway hangs. Health-monitor's 60s startup grace and 300s interval pass silently.

Suspected root cause

The release notes for v2026.4.26 include several plugin-discovery / manifest-cache changes that could plausibly cause this. Candidates:

• "Follow symlinked plugin directories in global and workspace plugin roots" (Plugins/Discovery)
• "Reuse manifest records already loaded for bundled web provider candidate discovery" (Plugins/Web)
• "Resolve runtime manifest-contract plugin owners from one plugin index" (Plugins/Contracts)
• "Reuse one manifest registry pass while resolving bundled document and web-content extractor plugins" (Plugins/Extractors)
• "Stage bundled plugin runtime dependencies before Gateway startup" (Plugins/Install)

The OpenAI SDK's resources/ tree has many cross-importing barrel files (index.js files that re-export sibling and child resources). A discovery walker that enters this tree from node_modules/openai and naively follows each barrel's exports would re-discover the same files via different paths, never converging unless the walker dedupes by realpath or by (dev, inode).

The write-side syscalls (link, copy_file_range, mkdir, chmod) suggest the walker is also mirroring entries into dist/extensions/node_modules/openclaw/plugin-sdk/ (or a similar staging path). That mirror operation appears to create or overwrite files that are then re-visited on the next pass, sustaining the loop.

Workaround

Rollback to v2026.4.24:

sudo systemctl stop openclaw
sudo npm install -g [email protected]
# v4.26's install partially overwrites shared paths under
# ~/.openclaw/plugin-runtime-deps/openclaw-2026.4.24-<hash>/dist/extensions/node_modules/openclaw/plugin-sdk/
# so the v4.24 cache must be rebuilt:
rm -rf ~/.openclaw/plugin-runtime-deps/openclaw-2026.4.24-*
rm -rf ~/.openclaw/plugin-runtime-deps/openclaw-2026.4.26-*
sudo systemctl start openclaw
# Gateway runs `npm install` for bundled plugin deps and is ready in ~90–130s.

Notes

• The [gateway] agent model: custom-opencode-go-extras/deepseek-v4-flash log line is unrelated — it appears at startup on both v4.24 and v4.26 simply because that custom provider is registered. It is not a model-resolution error.
• Doctor's --fix change to meta.lastTouchedVersion: 2026.4.26 survives a v4.24 rollback; restoring the pre-upgrade config silences the subsequent Config was last written by a newer OpenClaw warnings.
• I can attach a longer strace log or full gateway log if useful — held back to keep this report scannable.

Severity

High for any deployment that auto-upgrades to v2026.4.26. The gateway becomes unresponsive without emitting an error, which makes the failure mode hard to diagnose without strace.