Ai

Hermes Agent v0.15.1 Patches Dashboard Loop and Worker Issues

Hermes Agent v0.15.1 Patches Dashboard Loop and Worker Issues

Hermes Agent dropped a same-day hotfix for v0.15.0 on May 29, 2026 — v0.15.1 is a patch release that kills the dreaded dashboard infinite-reload loop. Anyone running v0.15.0 in loopback mode (Docker, hosted Hermes, fresh installs) hit a wall where the dashboard just kept spinning. That's fixed. But the patch isn't a one-trick pony. It bundles several other follow-ups: the kanban worker SIGTERM handling, a unified /model picker, and a /yolo session bypass. This is what a responsible hotfix looks like — narrow scope, big impact.

What Changed

The headline fix targets the dashboard infinite-reload loop specifically in loopback mode. The bug was a race condition in the websocket reconnection logic. When the agent ran in loopback (i.e., self-hosted with local networking), the dashboard client would attempt to reconnect to the backend before the server fully initialized. Each failed attempt triggered a full page reload, creating an unbreakable cycle. The fix introduces a retry backoff with a 3-second initial delay and a max retry count of 10. After that, it shows a static error page instead of looping.

The kanban worker SIGTERM fix is a subtle but critical one. Prior to v0.15.1, sending a SIGTERM to the kanban worker process during a task execution could orphan the task — it would remain locked in the database. The patch adds a signal handler that gracefully cancels the current task and marks it as failed with a reason code. This prevents manual cleanup and ensures queue integrity.

The /model picker unification is a UX consistency fix. In v0.15.0, the drop-down for selecting a model behaved differently on the agent configuration page versus the chat interface. One used a modal, the other a direct select. Now both use the same React component. It's a small change — 42 lines of code — but it removes a source of confusion.

Finally, the /yolo session bypass addresses a bug where the session timeout logic was too aggressive for the /yolo endpoint. /yolo is meant for quick, one-off requests that don't need a persistent session. In v0.15.0, a /yolo call would still create a session object that expired after 5 minutes, causing unnecessary garbage collection. Now it bypasses session creation entirely, reducing memory overhead by about 8% per request.

Why It Matters

Hotfixes are routine but this one matters because the infinite-reload loop was a showstopper. If you run Hermes Agent in production with Docker — and plenty of teams do — you couldn't even open the dashboard. That's a critical failure mode for monitoring and debugging. The fix restores observability, which is the backbone of any AI deployment.

The SIGTERM handling is equally important for reliability. Orphaned tasks in the kanban worker can pile up, especially during rolling updates. Without the patch, an operator would have to manually query the database and clean up stuck tasks. That's not just an hour of lost time — it's a potential source of data corruption if the task held a resource lock.

The /model picker unification may seem cosmetic, but I've seen teams waste days debugging model selection inconsistencies. Standardizing the component reduces cognitive load and support tickets. And the /yolo session bypass? It's a performance win. In high-throughput environments, those 8% memory savings per request add up. Less GC pressure means lower latency tails.

This patch doesn't add features — it fixes the cracks that appeared hours after v0.15.0 shipped. That's exactly what a hotfix should do. No fluff, just execution. The 21 merged PRs in 28 commits show a team that moves fast but responsibly. If you're on v0.15.0, upgrade yesterday. If you skipped it, v0.15.1 is the stable version you should aim for.

Official Source: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.5.29