dsh-double

An agent edits its own config, restarts, and immediately breaks. That is the moment every DeepSeek Harness user dreads. dsh-double closes the loop: a change set is applied to a standby copy first, statically pre-checked (config dump, YAML/JSON syntax, bundle resolvability, duplicate loader detection) and trial-booted. Only when both pass is the change promoted to the main instance, which must confirm healthy before the standby is stopped. If the main fails to boot, it auto-rolls back to the pre-evolution snapshot. You stop needing an outside agent to rescue the one that broke itself.

Beyond blue-green self-evolution, the plugin adds three-layer health probing (TCP, HTTP, process), an on-demand cold standby that wakes as a spare brain when mechanical recovery fails, and a fully isolated second dsh instance called the testbed running on port 3081. Each side gets its own LLM agent, and the two converse over the official session gateway. The testbed proves an evolution in its sandbox first; once validated, the twin agent blue-greens the delta back onto main. Every exchange lands in a shared, replayable JSONL transcript surfaced in the settings dashboard, so you can follow the full two-way conversation from a single page.

If you run long-lived agent workloads on dsh and want your agent to iterate on its own config or workflow without dropping context, dsh-double turns self-evolution from a risky stunt into a reliable daily routine. The zero-dependency CLI and out-of-process supervisor keep the life-support layer alive even when dsh itself cannot boot, giving you a safety net that no in-process plugin alone can provide.

It is built for power users who deploy dsh in production-like environments, run multi-agent pipelines, and need an auditable trail of every self-modification. Testbed-first workflow guidance, optional evolve approval, and transparent restart notices make the whole cycle visible and controllable from the dashboard.