Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

After upgrading to 2026.5.18, every cold CLI session takes 27-45s for a trivial response due to models.authStatus latency of 10-24s; previously much faster.

Steps to reproduce

1. Install OpenClaw 2026.5.18 with Codex plugin on Windows 11.
2. Run: openclaw agent -t test -m "respond only: test" --timeout 90
3. Observe 27-45s total time for a one-word response.
4. Check logs: models.authStatus shows 10-24s on cold connection, ~300ms when cached=true.

Expected behavior

Cold CLI session responds in under 10s, as observed in earlier versions.

Actual behavior

Every cold CLI session takes 27-45s. Logs show models.authStatus taking 10-24s without cache, and ~300ms with cached=true. The 60s cache TTL expires between separate CLI invocations, so every new session starts cold.

OpenClaw version

2026.5.18 (50a2481)

Operating system

Windows 11

Install method

npm global

Model

openai/gpt-5.5

Provider / routing chain

openclaw -> @openclaw/codex plugin -> codex app-server (stdio spawn) -> OpenAI

Additional provider/model setup details

Gateway runs as Windows Scheduled Task on loopback (127.0.0.1:18789).
Plugin: @openclaw/codex with codexDynamicToolsLoading=direct.
Auth: OAuth via ChatGPT/Codex browser login.
codex app-server daemon not supported on Windows, so no persistent process available.

Logs, screenshots, and evidence

Startup stages trace (from gateway log):
runtime-plugins: 3549ms
model-resolution: 4310ms
auth: 2271ms
Total startup: ~10-11s

models.authStatus timings observed:
- Cold (no cache): 10237ms, 11469ms, 11588ms, 11962ms, 12008ms, 12113ms, 12502ms, 14685ms, 17595ms, 18928ms, 19118ms, 19227ms, 24263ms
- Warm (cached=true): 267ms, 276ms, 292ms, 311ms, 1476ms

Cache TTL: 60s hardcoded in server-methods-BAy3NbTS.js (CACHE_TTL_MS = 6e4)
Each new CLI invocation opens a fresh WS connection, always hits cold path.

Total measured latency:
- Cold session: 27-45s
- codex exec direct: ~3s (Codex itself is not the bottleneck)

Impact and severity

Affected: All Windows users using openclaw agent CLI with Codex plugin.
Severity: High (blocks practical use — 45s for a trivial response).
Frequency: Always, every cold CLI session.
Consequence: Agent unusable for interactive workflows; each conversation requires waiting 27-45s before first response.

Additional information

Last known good version: NOT_ENOUGH_INFO (issue noticed after upgrading to 2026.5.18).
codex app-server daemon workaround not available on Windows.
Switching codexDynamicToolsLoading from "searchable" to "direct" reduced latency from ~78s to ~45s, but cold path remains slow.
Node.js version: 24.14.1. Codex CLI: 0.131.0.