dsh-autoresearch

Optimization research loops, where you propose a change, run an evaluation, compare a scalar metric against a baseline, and keep or reject the result, are exactly the kind of repetitive, error-prone workflow that dsh-autoresearch automates for DeepSeek Harness agents. You declare one immutable objective, a narrow set of mutable file globs, and a Host-registered evaluator; the plugin then drives a bounded keep-or-reject search inside an isolated Git worktree. The proposal agent may edit only the declared paths, while a fixed, deployment-pinned metric acts as the sole source of truth. Because the proposal model cannot alter the evaluator command, metric direction, dataset identity, or runtime environment, the self-scoring pitfall that plagues naive auto-research loops is eliminated by design.

Three design pillars set this plugin apart. First, boundedness and safety: candidate caps, per-attempt wall-clock watchdogs, and a finite target threshold all bound the run; any recoverable failure consumes one ordinal, restores the accepted worktree, and proceeds to the next candidate, while policy, persistence, or provenance violations halt the run entirely. Second, durability and fail-closed recovery: every state transition is transition-checked and hash-bound in a SQLite tracker; cancellation and resume reconcile durable evidence before any mutation or evaluator spawn, so a crashed controller never leaves a half-applied candidate behind. Third, a minimal trust surface: the evaluator runs through the managed subprocess provider with a closed environment and bounded output capture, and the isolated worktree protects repository state. The plugin explicitly does not provide OS-level sandboxing, so deployments with adversarial-code requirements must supply an external isolation layer themselves.

This plugin is best suited for developers and teams who have a reproducible, algorithm-qualified evaluation command and want to explore dozens of code or configuration candidates while automatically retaining the best one. Whether you are tuning prompts, searching hyper-parameters, or incrementally improving an algorithm module, as long as the metric is well-defined and the evaluation is cheap enough to run repeatedly under a timeout, dsh-autoresearch turns the entire trial, measure, and compare cycle into a single background job with full auditability and resumability.