Summary

After upgrading from 2026.4.15 → 2026.4.27, the gateway exhibits ~7-second synchronous event-loop stalls during normal agent activity, blocking Telegram polling and forcing watchdog-triggered container restarts. 24h post-upgrade window: 27 gateway crashes, 25 Tier 2 (full container restart). Pre-upgrade had been operating without this pattern.

.cpuprofile evidence (6 stalls aggregated, 230s combined CPU) shows the dominant hot path is createOpenClawTools / createOpenClawCodingTools invoking loadPluginRegistrySnapshotWithMetadata and resolveProviderAuthEnvVarCandidates synchronously on every tool-creation pass. Top self-time frames are lstat, readFileSync, path.resolve — classic uncached fs scan.

This may share root cause with #73532 (plugin loader hot loop on 4.25/4.26) but differs in entry point and version: that report's hot stack is prepareBundledPluginRuntimeDistMirror (mirror-init); mine is createOpenClawTools (tool-creation per agent invocation). Filing separately so the maintainer can dedup or split as appropriate.

This is distinct from #74345 (ACP session leak / 480s task-registry-maintenance event-loop holds, also on 4.27) — different fingerprint (~7s stalls, no ACP cleanup loop, no session-write-lock warnings).

Environment

• OpenClaw 2026.4.27 (c27ae31043c7) — installed via global npm
• Node v22.22.1
• Linux x86_64, host-mode container (Docker), gateway port 54401
• Channels enabled: Telegram (multi-bot), WhatsApp
• ~10 user agents bound to lanes via bindings[]; many tool-call-heavy

Symptom fingerprint

/data/.openclaw/logs/gateway-eventloop.jsonl — continuous ~7s stalls during normal load:

ELU peaked at 0.81 across two consecutive samples, with maxMs=6941–7650ms. CPU profiles auto-saved per stall.

gateway-monitor.log — 27 gateway crashes / 24h, 25 Tier 2 (full container restart), concentrated in the post-upgrade window. Watchdog kills gateway, container restarts, load resumes, stalls return.

While stalled: Telegram long-poll stays alive (update-offset-*.json mtime advances), getChat/getChatMember confirm bot-side state is fine, but channel-post-routed agent invocations don't dispatch — the agent isn't getting CPU.

CPU profile analysis (6 stalls aggregated, 230s combined)

Top 5 by inclusive time (caller + descendants):

Top 5 by self time (where CPU actually burns):

The hot-path pattern is unambiguous: createOpenClawTools + createOpenClawCodingTools running synchronous plugin-registry rebuild + provider-auth-env-var resolution on every tool-creation pass, dominated by lstat / readFileSync / existsSync / readdir.

loadInstalledPluginIndex (22,638ms inclusive) and loadPluginManifest (905ms self) confirm the plugin index is being rebuilt, not cache-served.

resolveProviderAuthEnvVarCandidates → resolveManifestProviderAuthEnvVarCandidates → isProviderApiKeyConfigured → hasAuthForProvider indicates per-tool-call provider-auth resolution that walks process.env against every model provider's possible env var name.

Inferred bug class

In 2026.4.27, the tool-creation path appears to:

1. Reload the entire plugin registry from disk synchronously on every invocation rather than serving from the cache that loadOpenClawPlugins is supposed to provide.
2. Re-resolve provider-auth env-var candidates per tool synchronously without caching — this scales with (number of tools × number of providers × number of env-var aliases per provider).

Each pass takes ~7s of CPU time, blocking the event loop. Under any non-trivial agent activity (multiple tool calls per minute), the gateway never recovers and watchdog kills it.

This blocks: Telegram long-poll dispatch, agent response generation, all PM2 worker IPC.

Likely overlaps with #73532's "cache key varies between callers" / "cache invalidated too aggressively" hypotheses, but expressed via the tool-creation entry point rather than the mirror-prep one.

Reproduction

1. Install [email protected] (or pull c27ae31043c7)
2. Configure ≥2 enabled user plugins and ≥2 model providers with auth profiles
3. Bind ≥1 Telegram bot to an agent and send agent any tool-call-heavy prompt under steady load (~1–2 invocations/minute)
4. Within minutes, gateway-eventloop.jsonl will show 6000–7700ms stalls per sample window
5. Within ~1h, watchdog Tier 2 restart will fire

CPU profiles auto-write to <DATA>/.openclaw/logs/gateway-stall-profiles/<pid>-<ts>/*.cpuprofile. Loading any in Chrome DevTools → Performance shows the flame graph matching the table above.

Suggested next steps

• Instrument getCachedPluginRegistry / setCachedPluginRegistry (per Plugin loader hot loop: prepareBundledPluginRuntimeDistMirror + JSON5 manifest parsing saturate gateway and starve event loop #73532's suggestion) to log cache key + hit/miss; verify whether createOpenClawTools is cache-missing every call.
• Audit resolveProviderAuthEnvVarCandidates for memoization — if there's no per-process or per-tool cache, the per-invocation env walk is the easy half of this regression to fix.
• Consider whether the 4.15 → 4.27 plugin-loader refactor moved registry-load from startup-time to tool-factory-time, and whether memoization survived the refactor.

Artifacts available

• 6 .cpuprofile files (10MB total) — happy to attach a redacted bundle on request
• gateway-eventloop.jsonl excerpt
• gateway-monitor.log crash log
• Internal finding doc with full top-10 inclusive/self frame analysis

Related

• Plugin loader hot loop: prepareBundledPluginRuntimeDistMirror + JSON5 manifest parsing saturate gateway and starve event loop #73532 — plugin loader hot loop (prepareBundledPluginRuntimeDistMirror + JSON5 parse). Adjacent code path, possibly same root cause.
• Event-loop saturation and ACP session leak on 2026.4.27 (9d8de70) #74345 — ACP session leak / 480s task-registry-maintenance holds on 4.27. Distinct fingerprint.
• [2026.4.26] Gateway main thread stalls under orchestration load; /readyz timeouts and reload deferred behind active runs #73467 — main thread stalls under orchestration load on 4.26. Possibly upstream of the same regression.