openclaw - 💡(How to fix) Fix [Bug]: github-copilot model-resolution takes ~5s on every agent run (token exchange not cached) [2 comments, 2 participants]

Bug type

Performance regression / missing token cache

Summary

Every agent run with the github-copilot provider incurs ~5 seconds of model-resolution latency. Based on the docs, this is the GitHub token → Copilot API token exchange step. It appears to happen on every run with no caching between turns.

Steps to reproduce

1. Configure OpenClaw with github-copilot/claude-sonnet-4.6 or any github-copilot/* model
2. Send any message
3. Observe [trace:embedded-run] startup stages in gateway.log

Expected behavior

Token exchange should be cached for the duration of the token's TTL (or at minimum for the gateway process lifetime), making model-resolution sub-second on warm runs.

Actual behavior

model-resolution consistently takes ~5,200–5,380ms on every single run, even back-to-back turns in the same session.

Startup trace evidence (after adding NODE_OPTIONS=--dns-result-order=ipv4first)

Before fix (DNS issue compounding it — 11s model-resolution): model-resolution:11508ms model-resolution:11507ms model-resolution:11635ms

After DNS fix — model-resolution improved but still consistently ~5.2s: model-resolution:5380ms model-resolution:5230ms model-resolution:5212ms

auth stage drops to ~4-6ms after DNS fix, confirming the DNS workaround works for auth but model-resolution remains slow regardless.

Environment

• OpenClaw: 2026.5.2
• OS: Windows 11 (x64)
• Node: v24.14.1
• Install: npm
• Provider: github-copilot (built-in, device-login flow)
• Model: github-copilot/claude-sonnet-4.6

Additional context

The copilot-proxy plugin config schema exposes zero configurable options, so there's no user-side workaround available. A token cache with TTL-aware refresh (similar to how most OAuth clients handle short-lived tokens) would fix this entirely.