openclaw - 💡(How to fix) Fix [Feature]: Expose `installBundledRuntimeDeps: false` + make `isolatedExecutionRoot` default for bundled plugin runtime deps [3 comments, 3 participants]

Summary

Two adjacent config-surface changes to give operators an escape hatch from the bundled runtime deps lazy-stage cycle, which has been a recurring source of event-loop stalls and disk bloat (#74949, #73532, #56733, #74860 are all downstream symptoms).

What's already in source

Reading dist/bundled-runtime-deps-*.js and dist/loader-*.js in 2026.4.24 shows two relevant code paths that exist but are gated:

1. loader-NucjcOgv.js:2240 references installBundledRuntimeDeps: false as an option — the code branch is there but isn't surfaced through openclaw.json config schema or CLI
2. bundled-runtime-deps-BdEAdjwi.js#resolveBundledRuntimeDepsExternalBaseDir already implements an isolatedExecutionRoot path (separate staging dir + atomic replace) — but the default installExecutionRoot is the legacy shared-root path

Operators currently have to either patch package.json manually (per #74949) or apply OPENCLAW_PLUGIN_STAGE_DIR via systemd drop-in to get the isolated path. Neither is documented.

Proposed changes

1. Surface installBundledRuntimeDeps: false as a config field

# openclaw.json
{
  "plugins": {
    "installBundledRuntimeDeps": false   // default: true
  }
}

When false, the plugin loader should:

• Skip the lazy-stage spawnSync entirely
• Assume the operator has provided the runtime deps externally (e.g., via the host's package.json + npm install already done)
• Fail fast with a clear error if a required plugin dep is missing, rather than silently re-installing

This is the proper fix for operators who:

• Run on disk-constrained hosts (lazy-stage cache reaches 30+ GB)
• Run latency-sensitive workloads where the 12s+ event-loop block is unacceptable
• Have already declared the 39 plugin deps in their host package.json (the #74949 workaround)

2. Make isolatedExecutionRoot the default

The current default puts plugin deps installs at the same root the host openclaw uses. For multi-plugin deployments this causes the destructive prune problem #74949 describes (different plugins' lazy-stage cycles mutually pruning each other's deps).

Changing the default to the isolated path that already exists in source would:

• Make a single-host, multi-plugin deployment safe out of the box
• Remove the need for users to discover and apply the OPENCLAW_PLUGIN_STAGE_DIR env var as a workaround
• Not change behavior for anyone explicitly opting into the legacy shared root

3. Document the resulting matrix

Why this matters

• #74949 fix is necessary but not sufficient — it stops the destructive prune but lazy-stage still fires every 3-12 min, blocking the event loop with each spawnSync
• #73532 documents the hot-loop residual after #74949
• The issues listed above are all distinct symptoms of the same root cause: a lazy-stage system that's mandatory and shares state across plugins

Giving operators these two knobs would let production deployments turn lazy-stage off entirely, while keeping the convenience for single-plugin / dev setups that benefit from the auto-install behavior.

Environment of the deployment that motivated this request

• OpenClaw: 2026.4.24
• Node: v22.22.0, npm: v10
• Plugins enabled: discord, feishu, telegram, qqbot, wecom, ddingtalk, openai, kimi (8 channels)
• Disk impact: ~38 GB cache bloat from repeated re-installs across ~14 days, 91% disk usage on a 120 GB volume before manual cleanup
• Channel impact: Discord WebSocket invalid-session loops during npm storms; Feishu HTTP timeouts (axios 30s)

Happy to test patches against this deployment.