A native DeepSeek Harness Web plugin for composing working state, reasoning strategy, quality gates and self-evolution per task, with an isolated, human-reviewed learning loop that proposes durable rules without touching project files.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-08-25
License
MIT
01What can it help you accomplish?
Compose a per-task working mode instead of a fixed persona
A visible input-area control showing the current combination (working state · reasoning · quality gate · evolution) plus a top-level Self-Evolution setting for global learning rules
DeepSeek Harness Web users who want controllable, reviewable, continuously-improving agent behavior
Turn completed sessions into durable, human-reviewed preferences
Isolated learning requests that propose candidate rules (identity, preferences, working requirements) written to global learned instructions only after explicit approval
Users who want the agent to learn stable preferences across sessions without silent behavior changes
Independently review each completed reply for risk and acceptance
An adversarial or acceptance review report (unmet requirements, unsupported conclusions, omissions, regressions, counterexamples, security risks) shown under the reply
Teams shipping high-stakes answers who need evidence-based review before trusting output
02How to install into DeepSeek Harness
Prerequisites
- A DeepSeek Harness build that supports the Web plugin loader, client UI slots, storage domain, Trajectory, and the referenced DSH services (fork/subagent, direct llm, planMode)
Installation steps
- 01
Install a pinned version into the DeepSeek Harness Web profile via npm: `npx -y @deepseek-ai/dsh plugin --profile web add @graysilver/dsh-evolve-modes@0.3.1`
- 02
If the DSH CLI is installed globally, use the shorthand: `dsh plugin --profile web add @graysilver/dsh-evolve-modes@0.3.1`
- 03
Or install the GitHub Release package directly: `dsh plugin --profile web add https://github.com/GraySilver/dsh-evolve-modes/releases/download/v0.3.1/graysilver-dsh-evolve-modes-0.3.1.tgz`
Verify the integration
- After restarting the Web profile, the Self-Evolution control appears next to the input-area tools; open the top-level **Self-Evolution** setting to manage global learning rules
03DSH integration and capability boundaries
Standalone Web plugin for DeepSeek Harness: an input-area control composes working state, reasoning strategy, quality gates, and a self-evolution loop that writes human-approved rule proposals to global learned instructions.
Composable workflow control
per-task selections of working state, reasoning strategy, quality gate, and self-evolution→a visible current-combination badge and a four-dimension control panel in the input area
Isolated self-evolution loop
completed agent sessions (up to 100 learning messages per source session)→human-reviewable rule proposals; approved rules written to the system prompt section marked `<dsh-evolve-modes-learned-instructions>`
each learning run uses a dedicated single learning persona/system prompt with the batch passed as one structured JSON user messageapproved rules are written to the plugin's own persistent storage, never to AGENTS.md, CLAUDE.md, or project filesAdversarial review gate
the parent agent's completed reply→an independent review report of unmet requirements, unsupported conclusions, omissions, regressions, counterexamples, and security risks — evidence, gaps, and follow-ups only, no silent rewrite
each completed parent reply adds one extra model call and corresponding latencyFirst-principles & plan-mode delegation
task goals, facts, assumptions, constraints, derivations, and verification→first-principles reasoning written into `request/header.system`; plan-mode delegated to the official `@deepseek-ai/dsh-plan-mode` service with DSH's `exit_plan_mode` approval
Acceptance review gate
the task, the candidate answer, and the approved plan when present→an acceptance report with fixed sections: Met, Gap, Unverified, Evidence, Concrete follow-up
each completed parent reply adds one extra model call and corresponding latency
04Who is it for? When not to use it?
Good for
- DeepSeek Harness Web users who want controllable, reviewable, continuously-improving agent behavior
- Users who want the agent to learn stable preferences across sessions without silent behavior changes
- Teams shipping high-stakes answers who need evidence-based review before trusting output
Not for
- Installing from a Git revision runs install-time code in the package — only install trusted revisions.
- Requires a DeepSeek Harness build supporting the Web plugin loader, client UI slots, storage domain, Trajectory, plus the fork/subagent, direct `llm`, and official `planMode` services.
05Compatibility, maintenance and safety notes
- Installing from a Git revision runs install-time code in the package — only install trusted revisions.
- Plan and quality reviews run through DSH's `tools/pre-execute` pipeline as a workflow policy; it is not an OS-level sandbox and does not restrict the model or tool runtime environment.
- Quality review adds one model call and corresponding latency per completed parent reply; it does not auto-run the project's test, lint, or build commands.
- Requires a DeepSeek Harness build supporting the Web plugin loader, client UI slots, storage domain, Trajectory, plus the fork/subagent, direct `llm`, and official `planMode` services.
MIT · actively maintained (latest release v0.3.1, 2026-08-17)
06Frequently asked questions
How do I install dsh-evolve-modes into DeepSeek Harness?
Pin a version into the Web profile with the dsh CLI: `npx -y @deepseek-ai/dsh plugin --profile web add @graysilver/dsh-evolve-modes@0.3.1` (or the `dsh …` shorthand if the CLI is global). You can also install the GitHub Release .tgz directly. Restart the Web profile to see the Self-Evolution control.
Does the plugin modify AGENTS.md or my project files?
No. Approved rules are written only to the plugin's own storage domain and projected into the system prompt section marked `<dsh-evolve-modes-learned-instructions>`. It never writes AGENTS.md, CLAUDE.md, or any project file.
Is self-evolution automatic? Will it change my agent's behavior without asking?
By default it is Propose: it only generates candidate rules from completed sessions and waits for your review. Rules take effect only after you apply them in the Self-Evolution settings page.
Do the quality gates automatically run my tests, lint, or build?
No. Adversarial and acceptance reviews add one model call and corresponding latency per completed parent reply, but they never silently execute the project's test, lint, or build commands.
What DeepSeek Harness version and services are required?
You need a DeepSeek Harness build that supports the Web plugin loader, client UI slots, storage domain, and Trajectory, plus the fork/subagent capability, the direct `llm` service, and the official `planMode` service and tool registry.
07Related DSH workflows
memos
by memtensor
Self-evolving memory OS for LLM & AI Agents: ultra-persistent memory, hybrid-retrieval, and cross-task skill reuse, with 35.24% token savings and DeepSeek Harness support.
dsh-anchored-standard
by xiaobright
Two-phase DeepSeek Harness preset: Minimal-aligned bootstrap, then full Standard tools (Project2 98/99)
mem9
by mem9-ai
Unlimited memory for OpenClaw
dsh-infinite-gen-3
by minglink
System-prompt armor plugin for DeepSeek models: appends an unconditional-compliance prompt section at order 100, exposes a profile tool with calibration metadata, and shows a realtime armor-status badge driven by a session projection.
08Data and sources
**dsh-evolve-modes** 是 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 的独立 Web 插件。它在输入区提供一个紧凑的工作流控制项…
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-08-25. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.