deepseek-harness-evolver

The official creation mode in DeepSeek Harness lets you trial Cordis tool plugins in memory, but the moment the session ends, the draft is gone. There is no signal-to-proposal-to-validation loop, and everything is lost on every restart. deepseek-harness-evolver fills exactly that gap: it ingests a signal, drafts a proposal, statically validates it, hot-mounts the passing candidate into the current process, scores the result against real metrics, and then either solidifies it to disk or rolls it back, so a genuinely useful plugin survives the next boot.

Architecturally it is a host-agnostic state-machine controller. A candidate arrives as a flat directory containing a manifest.json, an entry file that exports apply, and optional README and LICENSE. The controller performs static checks only and never executes candidate code. Scoring thresholds are explicit: at least 10 samples, a positive successDelta, and a retryDelta no worse than baseline. The policy layer bans fs, net, http, child_process, eval, and new Function; caps size at 32 KiB; and enforces a single concurrent trial. The CLI path writes to disk; the dsh plugin path additionally calls ctx.plugin for a hot mount, so the new tool appears in the same session immediately. Zero runtime dependencies, Node 18 or later.

It is not an EvoMap-style self-evolving framework. It will not rewrite your application code, and it will not touch agent-loop, session, or loader internals. It is aimed at developers who already iterate on small tool plugins in creation mode and want the ones that genuinely pass validation to become reusable, rollback-safe, restart-persistent assets with a clear audit trail.