Introduction

In DeepSeek Harness (DSH), running Agents typically involves modifying the system prompt, switching personas, or hand-writing review scripts. When tasks change, these configurations often require full replacement, making it difficult to mix and match “planning,” “reasoning strategies,” “quality checks,” and “long-term preference learning.” Writing review and self-evolution logic directly into the main session can also pollute the context or automatically modify project files.

dsh-evolve-modes is an independent plugin maintained by community contributor GraySilver for DSH Web (SkillHub directory page, GitHub). It does not fork DSH, replicate the Agent loop, or alter core code; instead, it provides composable workflow controls in the input area and restricts self-evolution to isolated learning, manual confirmation, and plugin-owned storage. This article explains what problems it solves, how to use it, and what to watch out for.

What This Is

dsh-evolve-modes (npm: @graysilver/dsh-evolve-modes) is designed for the DeepSeek Harness Web profile. After installation, a four-dimensional configuration for the current task appears alongside the input area; a global “self-evolution mode” setting manages cross-session learning proposals and approved rules.

The plugin’s positioning can be summarized as: making Agent workflows composable, reviewable, and continuously improvable, ultimately enabling Agent Self-Evolving with human gates. It is MIT-licensed; the README specifies compatibility with DeepSeek Harness 0.1.0-rc.6 and versions supporting the Web plugin loader, client UI slots, storage domain, Trajectory, and related services.

Core Features

Four Composable Dimensions

The input area displays the current combination, for example:

Normal · Standard · Off · Evolution On

Clicking the controls allows separate adjustments:

Dimension Options Effect
Working State Normal · Plan Execute the task immediately or enter the official DSH planning workflow.
Reasoning Strategy Standard · First Principles Provide a regular response or explicitly outline goals, facts, assumptions, constraints, deductions, and validations.
Quality Gate Off · Adversarial Review · Acceptance Review No added review; independently search for risks; or validate results against the task and approved plan.
Self-Evolution Off · On Analyze the session and generate rule proposals for human review, optimizing logic automatically without modifying project files like AGENTS.md.

This is not a mutually exclusive “personality mode” but a set of workflow decisions that can be recombined for each task.

Quality Gates

Adversarial Review: After the parent Agent responds, an independent review Agent checks for unmet requirements, unsupported conclusions, omissions, regressions, counterexamples, and security risks. It only reports evidence and follow-up actions, without silently rewriting the parent response.

Acceptance Review: Validates against the task, candidate answer, and approved plan. Reports distinguish Met, Gap, Unverified, Evidence, and Concrete follow-up. The review report is displayed below the corresponding assistant response. Each completed parent response incurs an additional model call and latency; the plugin does not automatically execute the project’s tests, lints, or builds.

First Principles & Plan Modes

  • First Principles: Writes goals, facts, assumptions, constraints, deductions, and validations into request/header.system; the Trajectory retains the same instruction as inspectable evidence. Turning it off only affects subsequent requests.
  • Plan Mode: Delegates to the official @deepseek-ai/dsh-plan-mode service, reusing DSH planning persistence and exit_plan_mode approval without building a separate planning system.
  • Tool Strategy: Planning and quality reviews are controlled via DSH’s tools/pre-execute pipeline; by default, read, glob, grep, read_image, the configured platform shell, and exit_plan_mode are allowed. This is a workflow policy, not an OS-level sandbox.

Self-Evolution Mode

Self-evolution identifies stable user identities, preferences, and work requirements from multiple completed Agent interactions. The default is Propose: it only suggests candidate rules and does not automatically change subsequent behavior.

Default settings (adjustable in global settings):

Setting Default Description
Self-Evolution Propose New sessions default to proposing.
Learning Batch Size 3 Learning triggers every accumulated 3 parent Agent responses (adjustable 1..100).
Pending Proposals Limit 100 Stops accumulating beyond this limit (adjustable 1..1000).
Learning Scope Source sessions with self-evolution enabled The current session is included once self-evolution is enabled.
Rule Scope Global Applied rules take effect across sessions, not tied to the project directory.

Learning data organization rules:

  • Each source session takes up to the latest 100 learning messages.
  • Retains each complete user message and the last visible assistant message for that turn; when the assistant message exceeds 2000 characters, it keeps the first and last 1000 characters, replacing the middle with ....
  • Proposal evidence must be verbatim from user messages; assistant inferences, one-off task details, etc., cannot be used alone as rule evidence.

Learning requests are isolated from the parent session: they do not inherit the parent session’s history and work context, do not create learning sub-Agents, do not carry tools, and do not load the source directory’s AGENTS.md or CLAUDE.md. They only use the plugin’s dedicated learning persona to analyze long-term identities, preferences, and work requirements. Failures are logged on the settings page, incomplete batches are retained for retry, and parent Agent execution is not blocked.

Workflow:

Complete 3 parent Agent responses
        ↓
Isolated learning request
        ↓
Pending review proposal
        ├─ Apply → Write to global learned instructions
        └─ Ignore → No change to subsequent behavior

Approved rules are written to a system prompt section marked with <dsh-evolve-modes-learned-instructions> and projected in the Trajectory; the plugin uses only its own persistent storage and does not write to AGENTS.md, CLAUDE.md, or any project files. The settings page supports adjusting batch size and limits, reviewing/applying/ignoring proposals, manually adding/editing/deleting global rules, viewing learning records and failure reasons, and restoring from automatic backups made before changes.

Installation & Enabling

It’s recommended to install a pinned version to the DSH Web profile via npm:

npx -y @deepseek-ai/dsh plugin --profile web add @graysilver/dsh-evolve-modes@0.3.1

If DSH CLI is installed globally:

dsh plugin --profile web add @graysilver/dsh-evolve-modes@0.3.1

Alternatively, install from a GitHub Release package:

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

For auditing or development, install a pinned Git revision (code runs during installation; use only trusted commits):

dsh plugin --profile web add github:GraySilver/dsh-evolve-modes#<trusted-commit>

After restarting the Web profile, the self-evolution mode controls appear next to the input area tools; open the top-level Self-Evolution Mode settings to manage global learning rules.

Optional configuration (the bundle automatically selects the platform shell; override only if the profile has registered that tool):

- id: dsh-evolve-modes
  config:
    shellTool: bash

Typical Usage

Combinations by Scenario

The README provides recommended combinations:

If you need to… Recommended Combination Reason for Fit
Quickly complete daily work Normal · Standard · Off Maintains execution pace without extra processes.
Make high-impact decisions Plan · First Principles · Off Research and expose assumptions first, then enter DSH planning approval.
Confidently deliver implementations Normal · Standard · Acceptance Review Independent review checks results against task goals.
Challenge high-risk answers Normal · First Principles · Adversarial Review Expand reasoning then find omissions and unsupported conclusions.
Consolidate stable personal preferences Normal · Standard · Off · Evolution On By default, identify long-term rules every 3 responses and generate proposals only.

Command-Line Switching

Use in the Web input area or via command API:

/evolve-mode
/evolve-mode working execute
/evolve-mode working plan
/evolve-mode reasoning standard
/evolve-mode reasoning first-principles
/evolve-mode quality off
/evolve-mode quality general-review
/evolve-mode quality acceptance-review
/evolve-mode evolution off
/evolve-mode evolution propose
/evolve-mode evolution batch-size <1..100>
/evolve-mode evolution max-pending-proposals <1..1000>
/evolve-mode review <turn>
/evolve-mode reviews

Legacy aliases normal, first-principles, adversarial-review are still supported for migration: the working state maps to execute, and reasoning and quality gates follow legacy mappings; current self-evolution settings are preserved.

Use Cases & Considerations

Who it’s for: Users who need to switch “execution/plan,” “standard/first-principles reasoning,” “no review/adversarial/acceptance” per task in DSH Web and want to consolidate long-term preferences into global rules via manual review. Categorized as a workflow plugin; GitHub repo has about 69 stars and 3 forks (check the repo page for current stats).

Dependencies & Compatibility:

  • Quality review requires DSH’s fork/subagent capability.
  • Self-evolution analysis requires DSH’s direct llm service.
  • Plan mode requires the official planMode service and tool registry.
  • From 0.3.0 onward, it automatically migrates legacy session settings, proposals, rules, backups, and learning records.

Security & Permissions: The plugin runs with the current DSH process permissions; read the source code and MIT license before installation. Git-based installation executes code during setup—only install trusted revisions. SkillHub is a community directory site with no official affiliation to DeepSeek / High-Flyer; do not treat it as an official app store.

Issues and suggestions can be submitted to GitHub Issues; integration feedback can also be posted to DeepSeek Harness Discussions.

Conclusion

dsh-evolve-modes decomposes working state, reasoning strategy, quality gate, and self-evolution into composable controls, isolates review and learning from the main session, and requires human confirmation before writing proposals to global rules. If you’re already using DSH Web, you can install 0.3.1 using the commands above and start trying it out with the input area combination controls or the /evolve-mode command.