Summary

Addresses the Codex P2 review on router-for-me#3726 ("Preserve cooldown handling for status-bearing EOFs").

The unexpected EOF fast path in shouldRetryAfterError now only fires for statusless errors:

-if isUnexpectedEOFError(err) {
+if status == 0 && isUnexpectedEOFError(err) {
     if !m.retryAllowed(attempt, providers) { return 0, false }
     return 0, true // immediate retry, no cooldown wait
 }

Why: isUnexpectedEOFError matches the substring "unexpected eof", so a status-bearing error (e.g. a 5xx whose body text contains "unexpected EOF") could match. For such an error MarkResult has already applied a 60s model/auth cooldown. The unconditional fast path returned an immediate retry that selection then rejects with a model_cooldown error — and when that cooldown exceeds max-retry-interval (default 30s) the cooldown error is surfaced to the client without ever retrying upstream. Gating on status == 0 lets status-bearing errors fall through to the existing cooldown-aware path instead.

This does not change the real-world behavior the original fix targeted: genuine unexpected EOFs in this codebase are statusless — both io.ReadAll(decodedBody) (claude_executor.go:310-313) and scanner.Err() (claude_executor.go:512-519) return plain errors with no StatusCode().

Tests

• New TestManager_ShouldRetryAfterError_StatusBearingEOFUsesCooldownPath — a 500 error whose message contains "unexpected EOF", with a single cooled-down credential (cooldown > maxWait), returns shouldRetry=false (no immediate fast path).
• Existing statusless EOF tests still pass (immediate retry, retry-count/disabled limits honored).
• gofmt, go vet ./sdk/cliproxy/auth/, go build ./..., go test ./... all pass.

Link to Devin session: https://app.devin.ai/sessions/eb1be9597f8d44c9af3d671e9fda7c03
Requested by: @hex-ci