dsh-evolve-modes: A Composable Workflow and Self-Evolution Plugin for DeepSeek Harness

What Does It Do?

dsh-evolve-modes is a standalone web plugin for DeepSeek Harness (DSH) that makes agent workflows composable, reviewable, and continuously improvable, ultimately enabling Agent Self Evolving. It solves the problem of rigid, opaque, and hard-to-improve agent behavior by providing a compact workflow control in the input area, letting you combine working state, reasoning strategy, quality gates, and self-evolution behavior. Core capabilities include plan mode, first-principles reasoning, adversarial review, acceptance review, and human-reviewed self-evolution learning, without modifying DSH core code.

Core Features

• Composable workflow control: displays the current combination (e.g., "Normal · Standard · Off · Evolution On") and lets you adjust four dimensions: working state, reasoning strategy, quality gate, and self-evolution.
• Plan mode: delegates to the official DSH plan service, reusing plan persistence and approval flow without reimplementing.
• First-principles reasoning: writes goals, facts, assumptions, constraints, derivations, and verification into the request header; Trajectory preserves the evidence.
• Quality gates: adversarial review independently finds risks; acceptance review checks results against the task and plan, reporting Met, Gap, Unverified, Evidence, and Concrete follow-up.
• Self-evolution mode: defaults to propose mode, triggers an isolated learning batch every 3 parent replies, generates human-reviewable rule proposals, and applies approved rules to global learned instructions without touching project files.
• Command support: provides /evolve-mode commands usable in the web input area or command API.

How to Use This Plugin?

After enabling it in DSH, the self-evolution control appears next to the input area; click it to combine the four dimensions. For example, choose "Normal · Standard · Off" for quick daily tasks, "Plan · First-Principles · Off" for high-impact decisions, "Normal · Standard · Acceptance Review" for confident delivery, "Normal · First-Principles · Adversarial Review" for high-risk answers, and "Normal · Standard · Off · Evolution On" to accumulate stable personal preferences. You can also type /evolve-mode commands directly. The global "Self-Evolution Mode" settings page lets you adjust batch size (default 3) and max pending proposals (default 100), view proposal evidence, apply or ignore proposals, manually manage rules, and restore from backups.