dsh-evolve-modes: A Composable Workflow and Self-Evolution DSH plugin for DeepSeek Harness
graysilver/dsh-evolve-modes
A DSH plugin adding composable task controls and human-reviewed self-evolution for agents.
Make DeepSeek Harness Agent working styles composable, reviewable, and continuously improvable, enabling Agent Self Evolving.
dsh plugin --profile web add github:graysilver/dsh-evolve-modes- Category
- Skills & Agents
- Platform
- DSH-Plugin
- Author
- graysilver
- Distribution
- Plugin
dsh-evolve-modes Key Features
dsh-evolve-modes Repository
dsh-evolve-modes Repo Summary
dsh-evolve-modes is a Web plugin (DSH plugin) for DeepSeek Harness, maintained by GraySilver under the MIT license, last updated in 2026-08. It addresses the problem that Agent working styles are not composable, not reviewable, and hard to improve continuously. It provides a compact workflow control in the input area to combine work state, thinking strategy, quality gates, and self-evolution behavior, ultimately enabling Agent Self Evolving. The plugin does not fork DeepSeek Harness, does not copy the Agent loop, and does not modify core code. After installation, the current task's combination is always shown next to the input area, while the global "Self-Evolution Mode" setting manages cross-session learning proposals and approved rules.
What core features does the dsh-evolve-modes DSH plugin support?
The plugin offers four independently adjustable work dimensions: work state (Normal or Plan), thinking strategy (Standard or First Principles), quality gate (Off, Adversarial Review, or Acceptance Review), and self-evolution (Off or On). These are not mutually exclusive "persona modes" but a set of work decisions that can be recombined for each task. For example, for quick daily work choose Normal · Standard · Off, for high-impact decisions choose Plan · First Principles · Off, for confident delivery choose Normal · Standard · Acceptance Review, for challenging high-risk answers choose Normal · First Principles · Adversarial Review, and for consolidating stable preferences choose Normal · Standard · Off · Evolution On.
How does the self-evolution mode of dsh-evolve-modes work?
Self-evolution defaults to "Propose" mode, which only suggests candidate rules without automatically changing behavior. A learning batch starts after every 3 completed parent Agent replies (adjustable from 1 to 100 in global settings). Learning requests are isolated: they do not inherit parent session history, do not carry tools, and do not load AGENTS.md or CLAUDE.md from the source session's working directory; they only analyze potentially long-term identity, preferences, and work requirements. Proposals must be human-confirmed; they can be applied or ignored. Approved rules are written into a system prompt section tagged with <dsh-evolve-modes-learned-instructions> and mirrored in the Trajectory. The plugin uses only its own persistent storage and never writes to any project files.
How do the quality gates and plan mode of dsh-evolve-modes work?
Adversarial review launches an independent review Agent after the parent Agent's reply, checking for unmet requirements, unsupported conclusions, omissions, regressions, counterexamples, and security risks. It only reports evidence and gaps, without silently rewriting or retrying. Acceptance review checks against the task, candidate answer, and approved plan if present, with a fixed report distinction: Met, Gap, Unverified, Evidence, Concrete follow-up. First Principles writes goals, facts, assumptions, constraints, derivations, and verification into request/header.system; Plan mode delegates to the official @deepseek-ai/dsh-plan-mode service, reusing DSH's plan persistence and exit_plan_mode approval flow. Tool strategy is controlled via DSH's tools/pre-execute pipeline, allowing read, glob, grep, read_image, and configured platform shell by default.
How do I install dsh-evolve-modes?
It is recommended to install a fixed version via npm into the DeepSeek Harness Web profile. After restarting, the self-evolution mode control appears next to the input area tools. Open the top-level "Self-Evolution Mode" settings to manage global learning rules. The installation command is:
dsh plugin --profile web add github:graysilver/dsh-evolve-modes
For source audit or development, you can install a fixed Git revision, but Git installs execute install-time code, so only install trusted revisions.
Which DeepSeek Harness versions and platforms does dsh-evolve-modes support?
As an official DSH plugin, it depends on official packages such as @deepseek-ai/dsh-api-remotes, @deepseek-ai/dsh-agent, @deepseek-ai/dsh-plan-mode, and @deepseek-ai/dsh-tools, integrating deeply with DeepSeek Harness's Web profile. It supports planning, review, and self-evolution capabilities. The plugin does not modify core code, ensuring good compatibility for DSH users who need composable workflows and human review.