openclaw update leaves plugin-runtime-deps/<version>/.../plugin-sdk in ENOTEMPTY state — every channel plugin silently fails to load

Summary

After running openclaw update from 2026.4.23 → 2026.4.24, every bundled channel plugin failed to load with ENOTEMPTY, Directory not empty: .../plugin-sdk. The gateway process stayed up and openclaw gateway status reported healthy (probe ok, admin-capable, listening on the configured port), but all messaging channels were dead, so from a user-facing perspective the agent was completely offline (no Telegram, no WhatsApp, no iMessage, no Signal, etc.).

Recovery required several invocations of openclaw doctor --non-interactive --fix and took ~10–13 minutes of downtime. There was no clear surface signal about what had failed — only gateway status (which was misleadingly green) and the JSON log file.

Environment

• OS: macOS Darwin 25.3.0 (arm64), Apple Silicon Mac mini
• Node: v25.6.1 (homebrew)
• OpenClaw: upgraded 2026.4.23 → 2026.4.24 (cbcfdf6)
• Install path: /opt/homebrew/lib/node_modules/openclaw
• Plugin runtime deps path: ~/.openclaw/plugin-runtime-deps/

Reproduction

1. Have a working OpenClaw install on 2026.4.23 with multiple channel plugins active (telegram, whatsapp, imessage, signal, etc.).
2. Run openclaw update to 2026.4.24.
3. Observe gateway start up cleanly per openclaw gateway status (running, probe ok), but no messages flow on any channel.

Symptoms

/tmp/openclaw/openclaw-2026-04-25.log showed 19 occurrences of:

[channels] failed to load bundled channel <NAME>: ENOTEMPTY, Directory not empty:
/Users/<user>/.openclaw/plugin-runtime-deps/openclaw-2026.4.24-da6bdffc3d96/dist/extensions/node_modules/openclaw/plugin-sdk
'/Users/<user>/.openclaw/plugin-runtime-deps/openclaw-2026.4.24-da6bdffc3d96/dist/extensions/node_modules/openclaw/plugin-sdk'

…across at least these bundled channels: telegram, whatsapp, imessage, signal, line, msteams, mattermost, nextcloud-talk, qqbot, tlon, twitch, zalo, synology-chat.

Concurrent secondary noise (likely cascading from the failed plugin loads):

• 16 occurrences of Unhandled promise rejection: CIAO ANNOUNCEMENT CANCELLED (Bonjour/CIAO mDNS advertiser).
• Bonjour repeatedly: bonjour: restarting advertiser (service stuck in announcing for ~Xms ...).
• 15 stability bundles written to ~/.openclaw/logs/stability/openclaw-stability-2026-04-25T22-...-unhandled_rejection.json between 15:43 and 15:50 PDT.
• WebSocket handshake timeouts: cause=handshake-timeout handshake=failed durationMs=15761.
• During reconnects: chat.history unavailable during gateway startup.

Root cause (best guess)

openclaw update extracted plugin-runtime-deps/openclaw-2026.4.24-da6bdffc3d96/dist/extensions/node_modules/openclaw/plugin-sdk non-atomically. The plugin-sdk directory contained ~308 files of partial/leftover content (likely hardlinks from npm/pnpm sharing with the previous version's tree or from the older openclaw-unknown-d9b993c1ba70 tree that still lives in the same parent dir). On plugin load, the channel plugin loader tried to clean/relink that directory and got ENOTEMPTY.

The system on this machine had two plugin-runtime-deps trees side by side:

~/.openclaw/plugin-runtime-deps/
├── openclaw-2026.4.24-da6bdffc3d96/   # current
└── openclaw-unknown-d9b993c1ba70/      # orphan from a prior version

It's possible the orphan tree contributed shared hardlinks that confused the cleanup step.

Fix that worked

openclaw doctor --non-interactive --fix
# (rerun several times; total recovery time ~10 minutes)

After enough retries, doctor --fix was able to remove the offending plugin-sdk directory and let the new tree extract cleanly. All channels then loaded and the gateway accepted traffic again.

Suggested fixes / improvements

1. Atomic extraction of plugin-runtime-deps/<version>/. Extract into a sibling temp dir and rename(2) into place, so no consumer ever observes a partial plugin-sdk directory.
2. openclaw update should run the equivalent of doctor --fix automatically on completion, or at minimum verify that all bundled channels load before declaring the update successful.
3. Surface plugin-load failures in openclaw gateway status. Today, status reports running / probe ok / admin-capable even when every channel plugin has failed to load. A "channels: N healthy / M failed" line (or a non-zero exit code on degraded state) would have made this trivially diagnosable.
4. Garbage-collect orphan plugin-runtime-deps/openclaw-<old> trees during update, or warn when more than one exists.
5. Make the CIAO/Bonjour Unhandled promise rejection: CIAO ANNOUNCEMENT CANCELLED a logged warning, not an unhandled rejection. Each one writes a stability bundle (15 in 7 minutes here), inflating the stability dir and noise.

Logs / evidence

Available on request from the affected machine (paths shown for reference):

• /tmp/openclaw/openclaw-2026-04-25.log
• ~/.openclaw/logs/stability/openclaw-stability-2026-04-25T22-*-unhandled_rejection.json (15 files)
• ls ~/.openclaw/plugin-runtime-deps/ showing both openclaw-2026.4.24-da6bdffc3d96 and openclaw-unknown-d9b993c1ba70.

Happy to share scrubbed log excerpts if helpful.