Summary

On OpenClaw 2026.4.15, Anthropic auth works in the interactive main session but fails in isolated cron runs that should be able to resolve the same auth profile state.

Environment

• OpenClaw: 2026.4.15
• Host: macOS (local gateway on Mac mini)
• Working interactive model path: anthropic/claude-opus-4-7
• Failing isolated path: cron runs / isolated sessions

Expected behavior

A cron run that uses the same configured Anthropic model/auth profile as the main session should resolve auth successfully, or fail with a clearly scoped config difference that explains why.

Actual behavior

Interactive checks show Anthropic auth is available, but isolated cron runs fail auth resolution and fall back or error.

From local memory/logging, the issue was narrowed to this pattern:

• interactive openclaw models status resolves the auth profile correctly
• isolated cron runs do not resolve it the same way
• workaround currently is to keep cron jobs on openai-codex/gpt-5.4 even though the main session can use Anthropic successfully

Why this matters

This creates a confusing split-brain auth state:

• main session appears healthy
• automation silently cannot use the same provider
• users may wrongly conclude Anthropic auth is broken globally when it is actually a cron/isolation resolution problem

Repro outline

1. Configure Anthropic auth so the main session works normally
2. Verify interactive model/auth status succeeds
3. Run an isolated cron job or equivalent isolated run using the same Anthropic model
4. Observe auth failure or fallback in the isolated run

Current workaround

Keep cron jobs on openai-codex/gpt-5.4 until the isolated Anthropic auth path is fixed.

Notes

I originally filed this from memory after diagnosing an internal cron-auth mismatch on a real agent setup. If needed I can provide more precise local config details and logs from the affected machine.