openclaw - 💡(How to fix) Fix [Bug]: snapshot reuse misses when reader workspace differs from startup workspace (follow-up to #77519 fix) [1 comments, 2 participants]

Bug type

Regression follow-up

Summary

The v2026.5.4 fix for #77519 (48ff390953 + a7263de258) covers readers that omit workspaceDir and hot paths that previously dropped it. There is a third path it does not cover: when a reader passes a concrete agent workspace that is not the same string as the startup snapshot's stored workspace, the snapshot is rejected on every call.

For a deployment with multiple agents that each have an explicit workspace: entry pointing to its own directory, this is the ordinary state — every non-default-agent dispatch goes through loadPluginMetadataSnapshot and runs the per-plugin stat sweep on the main thread.

Steps to reproduce

1. Configure several agents in agents.list[] with explicit workspace: paths that differ from the default workspace (e.g. one workspace dir per agent).
2. Drive dispatch on the non-default agents (channel inbound, webchat, etc.).
3. strace -p $GATEWAY_PID -c for 5 seconds during dispatch.

Expected

Snapshot reuse on dispatch hot paths, similar to single-workspace setups.

Actual

% time     seconds  usecs/call     calls    errors syscall
 70.35    0.194137           3     48914           statx
  9.95    0.027445           3      6895       917 access
  8.24    0.022730           4      4564           openat
  ...
  ~67k syscalls / 5s, 70% in statx

Pattern: alphabetical traversal of dist/extensions/<plugin>/{package.json,openclaw.plugin.json} for all bundled plugins, repeating while dispatch is active.

Effect on the same workload, v2026.5.3-1 vs v2026.4.27:

• sessions.list: 156 s vs ~150 ms
• chat.history: 96 s vs ~155 ms
• eventLoopUtilization: 1.0 vs ~0.05
• Discord WS heartbeat misses + simultaneous close 1000 storms
• v2026.4.27 strace: 0 statx, ~671 total syscalls / 5s

(Numbers from the original v2026.5.3-1 reproduction; we have not yet captured a v2026.5.4 strace from the same setup — see note below.)

Where it shows up in source

Startup stores the snapshot with defaultWorkspaceDir:

// gateway startup
setCurrentPluginMetadataSnapshot(pluginLookUpTable, { config });
// snapshot.workspaceDir ends up = ~/.openclaw/workspace

Dispatch reader passes the dispatching agent's resolved workspace:

const workspaceDir = resolveAgentWorkspaceDir(cfg, agentId);
// e.g. ~/.openclaw/workspace-<agent-id>
const snapshot = getCurrentPluginMetadataSnapshot({ config, workspaceDir });

Equality check at current-plugin-metadata-snapshot.ts:80-87:

if (
  params.workspaceDir !== undefined &&
  (snapshot.workspaceDir ?? "") !== (params.workspaceDir ?? "")
) {
  return undefined;
}

Two different non-empty strings → cache miss → fallback runs safeFileSignature for every record in index.plugins.

The v2026.5.4 allowWorkspaceScopedSnapshot opt-in does not apply here because the reader has a real workspaceDir to pass.

Environment

• v2026.5.3-1 (rolled back to v2026.4.27 to restore service)
• Linux x86_64, Node v22.x supported / observed on v25
• ~10+ agents with explicit per-agent workspace: config, ~96 bundled plugins, several inbound channels

Note

We have not yet run v2026.5.4 in production from this setup, so the regression is reasoned from source rather than verified against the new release. Happy to capture a v2026.5.4 strace in a controlled window if that would help.

No fix proposal here — leaving design discretion to maintainers; the safety boundary around workspace-plugin discovery is non-trivial.

Related

#77519, #77532, #77529 (closed setter-only PR), #76182, #73353