dsh-autoresearch: DeepSeek Harness Persistent Auto-Research Loop
`dsh-autoresearch` is a plugin designed for DeepSeek Harness (DSH) to address the inefficiency of manually managing iterative loops in agent development. It supports persistent automated research experimentation loops, enabling agents to continuously modify code, measure results, and iterate within a session. Core features include Web GUI management, flexible experiment control (editing, rolling back, and preserving state), and real-time result visualization. Users can create a research experiment, check status, or clear the ledger with simple commands. Data is stored in the local `.auto/` directory, with Git or private snapshots protecting key files. This plugin lowers the barrier to agent development and is well suited to experimental development workflows that require long-running, multi-round iteration.
Read Moredsh-autoresearch: DeepSeek Harness Bounded Metrics Auto-research Plugin
`dsh-autoresearch` is a plugin for DeepSeek Harness designed to achieve automatic evaluation and screening of candidate modifications. Its core mechanism involves establishing a baseline under fixed metrics and determining whether to keep or revert candidate modifications using bounded keep/reject strategies. The plugin executes the optimization loop within an isolated Git worktree, utilizes SQLite to persist run, experiment, and trial records, and supports a failure shutdown recovery mechanism. Its key features include: using a single scalar metric to strictly determine minimization or maximization; relying on the trusted evaluator identity provided by the Host (such as commands, datasets, etc.) to prevent the proposal model from tampering with the evaluation environment; limiting the scope of writable paths via mutable_globs; and defaulting to using dsh-job...
Read More