For Just $0.03, DSH Brought a Broken Hermes Agent Back to Life

In the previous post, we tried getting two agents to work together inside the same cloud environment: Hermes Agent as the navigator - understanding requirements, designing the plan, and reviewing results - and DeepSeek Harness (DSH) as the driver - entering the actual workspace to write and verify code.

A new problem gave this combo a new purpose.

Recently, several members of our community ran into errors like these from their Hermes Agent:

Type one: No reply: the turn was stopped because session storage could not be written (the transcript would have been lost on restart). Check the state database health (hermes doctor), then send your message again.

Type two: Session storage was temporarily unavailable, so this turn was stopped to protect your conversation history. Your message should already be saved. Please send it again in a moment.

Both trace back to the same root cause: a Hermes Agent file, state.db, becoming unwritable or unreadable for various reasons. To protect existing data, the agent then refuses further conversations. The really painful part: rebuilding the agent from LightVela's Diagnostic Center didn't fix it either.

This used to be a genuinely tricky situation, until there was a DeepSeek Harness running in the same environment.

Why Can DSH Fix It?

The repair itself uses Hermes Agent's native recovery mechanism. The tricky part is that this repair has to be initiated from inside the Hermes Agent's runtime environment.

DSH runs in that same environment. It can reach Hermes' processes, logs, configuration, and files, and it can execute commands. So we can simply hand the repair job to it.

How, Exactly?

First, install and configure DSH in LightVela. See the DeepSeek Harness documentation.

Important: when DSH works on this problem, make sure it's operating in the /home/agentuser/ directory.

There's also an optional but recommended enhancement: give DSH a Hermes Agent operations manual first, so it doesn't have to spend time discovering how Hermes is configured on its own. To install it, send this in DSH's chat box:

given https://skillhub.cn/install/skillhub.md, install @user_70854421/hermes-agent-skill-for-dsh.

Once installed, give it the first instruction:

There is a Hermes Agent running in your current environment; please locate its installation directory. I need your help troubleshooting and resolving an issue by referencing the skill that was just installed.

If you'd rather play it safe, phrase it this way:

There is a Hermes Agent running in your current environment. Please locate its installation directory. I need you to help me troubleshoot and resolve my issue, referencing the skill that was just installed, regarding:
{Describe your issue here}.

Please first identify the cause of the problem and design a fix, but do not execute it yet.

When it comes back with a repair plan and you're not sure about it, ask in the community group first. The QR code is at the end.

In our case it went smoothly: right after locating the Hermes Agent, our DSH had already identified the state.db corruption, planned out the repair steps, and got to work:

Within a few minutes, the repair was complete:

Token consumption for the whole run: about 2.6M input tokens and 31.3K output tokens, with a 97% cache hit rate, a very high level. We were running DeepSeek-V4 Flash, so the total cost came to roughly ¥0.25, about four US cents. Compared to hiring an expert or grinding through it yourself, that's about as cheap as fixing things gets.

One last step, of course: go back to the Hermes Agent chat window and confirm it's actually alive again:

Final Words

Adventurers welcome: join the LightVela community group and let's explore what cloud agents can do together.