Summary

Gateway-owned plugin metadata snapshots are not always reused during model alias / manifest model-id normalization. In a workspace-scoped gateway runtime, resolveDefaultModelForAgent() can fall back to loadPluginMetadataSnapshot() during every embedded agent attempt, causing repeated plugin index / manifest discovery in the Node gateway process.

Observed impact

On a live gateway running OpenClaw 2026.5.4 (70d92b5), embedded agent prep repeatedly logged:

prep stages ... totalMs=15513
system-prompt:openclaw-reference-paths:8788ms

An inspector CPU profile showed the time was not actually reference-path lookup; the marker included the preceding model resolution path:

runEmbeddedAttempt
resolveDefaultModelForAgent
resolveConfiguredModelRef
buildModelAliasIndex
parseModelRefWithCompatAlias
parseModelRef
normalizeProviderModelId
loadManifestModelIdNormalizationPolicies
resolveMetadataSnapshotForPolicies
loadPluginMetadataSnapshot
loadPluginRegistrySnapshotWithMetadata
loadInstalledPluginIndex
discoverOpenClawPlugins

This made every agent turn pay a large fixed cost before the stream was ready, and contributed to gateway event-loop pressure.

Root cause

The gateway sets the current plugin metadata snapshot without the gateway workspace:

setCurrentPluginMetadataSnapshot(pluginLookUpTable, { config: gatewayPluginConfigAtStart });

But manifest model-id normalization resolves the active registry workspace:

const workspaceDir = params.workspaceDir ?? getActivePluginRegistryWorkspaceDirFromState();
const current = getCurrentPluginMetadataSnapshot({ config, env, workspaceDir });

That can make the current snapshot lookup miss even though the gateway already has a valid metadata snapshot for the running plugin set. When it misses, the normalization path reloads plugin metadata instead of reusing the gateway-owned snapshot.

Suggested fix

Two small changes fixed the live repro locally:

1. Include defaultWorkspaceDir when handing off the current metadata snapshot at gateway startup and reload:

setCurrentPluginMetadataSnapshot(pluginLookUpTable, {
  config: gatewayPluginConfigAtStart,
  workspaceDir: defaultWorkspaceDir,
});

2. In manifest model-id normalization, when the caller did not explicitly request a workspace, allow reuse of a workspace-scoped current snapshot:

const workspaceScopedCurrent = getCurrentPluginMetadataSnapshot({
  config: params.config,
  env,
  allowWorkspaceScopedSnapshot: true,
});

Validation from local patch

After applying the above locally:

• The per-turn prep stages totalMs=15513-16467 warnings stopped appearing.
• A shira embedded agent smoke completed successfully with OK.
• Targeted tests passed:

node scripts/test-projects.mjs \
  src/plugins/current-plugin-metadata-snapshot.test.ts \
  src/plugins/manifest-model-id-normalization.test.ts \
  src/agents/pi-embedded-runner/run/attempt-stage-timing.test.ts

Test Files 3 passed
Tests 18 passed

Also passed:

pnpm tsgo:core
pnpm build

Version notes

I also inspected the published [email protected] tarball. The dist bundle still appears to call:

setCurrentPluginMetadataSnapshot(pluginLookUpTable, { config: gatewayPluginConfigAtStart });
setCurrentPluginMetadataSnapshot(nextPluginLookUpTable, { config: params.nextConfig });

So this does not appear fixed in 2026.5.5.