openclaw - 💡(How to fix) Fix liveness warning fires during cold-start blocking (no startup grace window)

Problem

The diagnostic liveness sampler starts immediately after gateway boot with zero startup-grace window. During normal cold-start operations (model resolution ~18s, auth ~14s, session-resource-loader ~26s, plugin registration ~4s), the Node.js event loop is legitimately blocked for 50-60s total. This triggers false WARN-level liveness warnings:

liveness warning: reasons=event_loop_delay interval=35s eventLoopDelayP99Ms=14503.9 eventLoopDelayMaxMs=24912.1 eventLoopUtilization=0.975 cpuCoreRatio=1.047

The channel-health-monitor already has a startupGraceMs guard (default 60s), but the diagnostic heartbeat has no equivalent.

Environment

• OpenClaw 2026.5.7 (eeef486)
• Node.js v24.13.0
• Windows 10 x64
• 5 model providers, 4 enabled plugins, 12 in allowlist

Existing Work

This is a known issue with multiple related reports and attempted fixes:

• #75378 - Gateway event loop saturation during parallel subagent spawn (OPEN)
• #78352 - 16 Telegram bots on Windows cause event loop starvation up to 65s (OPEN)
• #79838 - event-loop saturation under heavy streaming + multi-channel load (OPEN)
• #79264 - post-attach.update-sentinel ~9s sync block per model_call (OPEN)

PRs

• #76401 - suppress liveness warnings during startup-grace window by @hclsys — Added \startupGraceMs: 60_000\ to \startDiagnosticHeartbeat, 23 tests passing. Closed by author after 20h with no maintainer signal. This is the correct fix.
• #79137 - une diagnostic liveness and idle queue depth — Raises max delay threshold to 10s while keeping P99 at 1s. Still OPEN.
• #79647 - do not mark event-loop degraded from ELU/CPU without delay evidence — Closed as already-implemented on main.

Proposed Fix

Re-open and merge #76401 (or a fresh PR with the same approach):

1. Add \startupGraceMs\ option to \StartDiagnosticHeartbeatOptions\
2. Skip liveness warning emission while
   ow < graceUntil\
3. Keep the sampler running during grace (for metric resets and baselines), only suppress emission
4. Default to 60-100s grace window

The key insight from PR #76401: \livenessSample = inStartupGrace ? null : rawLivenessSample\ — sampler still runs every tick, but warnings are gated until the grace window expires.

Additional Request

Consider making the grace window configurable via \OPENCLAW_STARTUP_GRACE_MS\ env var (as a future enhancement), so users with heavier startup loads can tune it.

Workaround (local)

Until an official fix ships, I patched the dist files locally with the #76401 approach:

• \diagnostic-DbrcADVi.js: Added \diagnosticStartupGraceUntil\ with 100s default
• \server.impl-DAWM0jI6.js: Same grace for event-loop health snapshot

This suppresses false warnings during cold start without affecting steady-state monitoring.