Ai

OpenAgent v4.5.0 Hardens Runtime Safety and Refactors Prompt System

OpenAgent v4.5.0 Hardens Runtime Safety and Refactors Prompt System

OpenAgent just shipped v4.5.0, and it's a release that quietly fixes some dangerous edge cases while overhauling the entire prompt system under the hood. If you're running agents in production, this one matters.

What Changed

Three safety guards now close pre-publish gaps that could silently break workflows. The biggest: the runtime fallback no longer blindly trusts isRetryable: true from providers. If a provider mislabels a 401, 403, or 404 as retryable — something that happens more often than you'd think — every configured fallback used to fire in an endless loop. Now those codes are treated as non-retryable. Period.

Another fix targets parent-wake stalls. When a child agent completed but its parent never got the signal, the entire chain hung indefinitely. v4.5.0 adds a timeout and re-wake mechanism that cuts that deadlock.

The third guard validates publish payloads before they hit the wire. Empty payloads or malformed JSON now get caught early, not at the broker.

Why It Matters

These aren't theoretical problems. In a real deployment with five fallback providers, a mis-tagged 403 could cause hundreds of retry requests per minute — burning quota and delaying responses. The fix is surgical but essential. And the prompts-core refactor? It's the unsexy work that makes future changes safer. Splitting every agent and mode prompt into versioned markdown files means teams can now diff prompt changes in code reviews. That's a huge win for compliance and debugging.

I've seen too many teams treat prompt management as a black box. This release pulls back the curtain. It doesn't add flashy features, but it makes the engine room less likely to explode. That's the kind of release that earns trust.

Official Source: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.5.0