openclaw - 💡(How to fix) Fix gateway: prewarmConfiguredPrimaryModel blocks channel startup for ~4 minutes on every restart [1 comments, 2 participants]

Summary

After every gateway restart, there is a ~4 minute gap before channels (Feishu, Telegram, etc.) become ready, and the first message after restart takes 3–6 minutes to process. During message processing, all WebSocket/HTTP connections drop simultaneously.

After thorough investigation (tracing code paths, benchmarking all components, testing APIs directly), the root cause has been identified.

Environment

• OpenClaw: 2026.4.26
• OS: Windows 11 (x64)
• Node.js: v24.13.1
• Primary model: github-copilot/claude-sonnet-4.6

Root Cause: prewarmConfiguredPrimaryModel blocks channel startup

File: dist/server.impl-hNr66nDN.js

In startGatewaySidecars(), channels are started after prewarmConfiguredPrimaryModel() completes:

// dist/server.impl-hNr66nDN.js ~line 8554
await prewarmConfiguredPrimaryModel({ cfg: params.cfg, log: params.log });
await params.startChannels();   // ← blocked until above resolves

Inside prewarmConfiguredPrimaryModel, ensureOpenClawModelsJson() is called:

await ensureOpenClawModelsJson(params.cfg, agentDir);

ensureOpenClawModelsJson calls planOpenClawModelsJson which fetches the latest model list from the provider API (GitHub Copilot /models endpoint) to rebuild models.json. This network call has no short timeout, and on this setup consistently takes ~4 minutes before resolving (possibly due to retry/backoff behavior or a slow response for enterprise Copilot endpoints).

Evidence:

• Every restart shows the same ~4 minute gap between [client ready] and channel startup
• models.json write timestamp (12:44) matches exactly when channels finally started (12:43:06 + ~4 min)
• Disabling store.vector.enabled had zero effect on the gap, ruling out memory-core
• All individual API calls tested directly are fast: Copilot token ~517ms, embedding ~1s, chat ~1.7s

Startup timeline (every restart)

08:50:56  starting channels and sidecars...
08:50:58  hooks loaded
08:51:43  [client ready]
           ← ~4 minutes: prewarmConfiguredPrimaryModel / ensureOpenClawModelsJson →
08:55:38  feishu / browser / telegram all start simultaneously

Secondary Issue: First message slow + simultaneous channel disconnects

After channels start, the first message triggers memory_search which calls ensureProviderInitialized(). Additionally, the first model API call (with full system prompt) takes longer than the Feishu WebSocket idle timeout (~60s), causing:

• write ECONNABORTED on Feishu WS + reconnect at the 60s mark
• Telegram Polling stall detected at the ~120s mark
• Both disconnect at exactly the same moment — confirming event loop / connection starvation

Expected Behavior

1. prewarmConfiguredPrimaryModel should not block channel startup. Either run it concurrently with channel initialization, or add a short timeout (e.g. 5–10s) so channels are not held waiting.
2. ensureOpenClawModelsJson should use a short timeout for the startup fetch, falling back to the existing cached models.json if the API call times out.
3. Channels should be ready to accept messages while model warmup proceeds in the background.

What was ruled out

• ❌ Session context size (149K tokens) — new sessions exhibit the same behavior
• ❌ memory-core local embedding provider blocking — actual provider is github-copilot (remote)
• ❌ GitHub Copilot API latency — all endpoints respond in <2s when called directly
• ❌ SQLite vector search blocking — FTS queries take <1ms, full scan 77ms
• ❌ store.vector.enabled = false — no effect on the 4-minute gap

Workaround

None available currently. The 4-minute delay on every restart is unavoidable with the current code path.