DeepSeek Harness Evolve Modes

English｜中文

Make an Agent's working methods composable, auditable, and continuously improvable, ultimately achieving Agent Self Evolving.

dsh-evolve-modes is a standalone Web plugin for DeepSeek Harness. It provides a compact workflow control in the input area, letting you combine an Agent's working state, thinking strategy, quality gates, and self-evolving behavior.

The plugin does not fork DeepSeek Harness, does not replicate the Agent loop, and does not modify core code. After install, the active combination shows next to the input area; the global Self-Evolving Modes setting manages cross-session learning proposals and approved rules.

The current version 0.4.0 supports DeepSeek Harness 0.1.1-rc.2. If you are still using Harness 0.1.0-rc.6, please install plugin 0.3.1.

Quick Install

The recommended approach is to 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@0.4.0

If the DSH CLI is already installed globally, you can use the shorthand:

dsh plugin --profile web add @graysilver/dsh-evolve-modes@0.4.0

After restarting the Web profile, the evolve-modes control appears beside the input tools. Open the top-level Self-Evolving Modes setting to manage global learning rules.

If you need to audit the source code or do development work, you can install a pinned Git revision:

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

Git installs execute code at install time; only install trusted revisions.

Feature Overview

The input area displays the active combo:

正常 · 标准 · 关 · 进化 开

Expand the control to tune each dimension:

Dimension	Options	Effect
Working State	Normal · Plan	Complete the task immediately, or enter the official DSH plan workflow.
Thinking Strategy	Standard · First Principles · Grilling	Answer normally, explicitly reason from first principles, or stress-test requirements and decisions through rounds of targeted follow-up questions.
Quality Gate	Off · Adversarial Review · Acceptance Review	No additional review, independently hunt for risks, or verify results against the task and approved plan.
Self-Evolving	Off · On	Automatically analyze sessions, generate rule proposals pending human review, and auto-optimize AGENTS.md (but without modifying AGENTS.md).

These are not exclusive "personality modes" but recombine-able working decisions per task.

Self-Evolving Mode

Self-evolving identifies stable user identity, preferences, and working requirements across completed Agent sessions. It defaults to Propose mode — suggesting candidate rules only, never auto-applying them.

Default Settings

Setting	Default	Description
Self-Evolving	Propose	Enabled by default for new sessions and older sessions without an explicit self-evolve choice.
Learning Batch	3	Triggers a learning run after every 3 completed parent Agent replies. Adjustable to 1..100 in global settings.
Pending Proposal Cap	100	Stops accumulating proposals once the cap is reached. Adjustable to 1..1000 in global settings.
Learning Scope	Source sessions with self-evolve enabled	No project scoping; if the current session has self-evolve on, it is included.
Rule Scope	Global	Approved rules take effect across sessions and are not bound to a project directory.

How learning data is organized

• Each source conversation contributes at most its most recent 100 learning messages.
• Every turn's full user message is retained, along with the last visible assistant message of that turn.
• Assistant messages serve as context only; when they exceed 2000 characters, the first 1000 and last 1000 characters are kept, with the middle replaced by ....
• Proposed evidence must come verbatim from user messages. Assistant inferences, one-off task details, implementation outcomes, silence, or "not mentioned again" cannot stand alone as rule evidence.

Learning requests are isolated

Every learning run uses a single, plugin-dedicated learning persona/system prompt, passing the current batch as one structured JSON user message. Learning requests:

• Do not inherit the parent conversation history;
• Do not inherit the parent Agent's working context;
• Do not create learning sub-Agents;
• Do not carry tools;
• Do not load AGENTS.md or CLAUDE.md from the source conversation's working directory;
• Analyze only identity information, preferences, and work requirements that may hold long-term.

Learning failures are logged on the settings page; unfinished batches are kept for retry. They never block the parent Agent.

Proposals require manual confirmation

完成 3 次父 Agent 回复
        ↓
隔离的学习请求
        ↓
待审阅提议
        ├─ 应用 → 写入全局 learned instructions
        └─ 忽略 → 不改变后续行为

On the top-level Self-Evolution Mode settings page you can:

• Adjust batch size and pending-proposal cap;
• View each proposal's category, inference type, and user evidence;
• Apply or ignore proposals;
• Add, edit, or delete global rules manually;
• View learning run logs and failure reasons;
• Restore from the auto-created backup before each change.

Approved rules are written into a system prompt section marked with <dsh-evolve-modes-learned-instructions>, and the same content is projected into the Trajectory. The plugin uses only its own persistent storage and never writes to AGENTS.md, CLAUDE.md, or any project files.

Task combination picks

You need to…	Recommended combo	Why it fits
Quick everyday tasks	正常 · 标准 · 关	Keeps execution pace, no extra process.
High-impact decisions	计划 · 第一性原理 · 关	Research and surface assumptions first, then DSH plan approval.
Confident delivery	正常 · 标准 · 验收审查	After implementation, an independent Agent checks results vs. goals.
Challenge high-risk answers	正常 · 第一性原理 · 对抗性审查	Make reasoning explicit, then hunt for gaps, counterexamples, regressions, and unsupported claims.
Codify stable preferences	正常 · 标准 · 关 · 进化 开	Finds long-term rules in batches of 3 replies by default; proposals only, never auto-enabled.

Quality gates

Adversarial review

Spawns a review agent post-reply to flag unmet requirements, unsupported conclusions, omissions, regressions, counterexamples, and security risks. Reports only; never rewrites, retries, or fixes the parent's reply.

Acceptance review

对照任务、候选答案以及存在时的已批准计划进行验收。报告固定区分：

Met
Gap
Unverified
Evidence
Concrete follow-up

Review reports appear under the matching assistant reply. Each quality review adds one model call and its latency to a completed parent reply, but never auto-runs the project's test, lint, or build commands.

Thinking & Planning

• First Principles: Goals, facts, assumptions, constraints, derivations, and verification are written into request/header.system; the Trajectory retains the same instruction block as auditable evidence. Once disabled, only subsequent requests are affected; historical evidence is not deleted.
• Grilling: Decision items and their dependencies are established first, and each round raises only questions whose prerequisites are already clear; questions use a numbered list with a recommended answer. The Agent investigates discoverable facts on its own, and after all branches are clear, summarizes shared understanding and waits for user confirmation; nothing is implemented before confirmation. After confirmation, it continues under the current "Normal/Plan" work mode.
• Plan mode: Delegates to the official @deepseek-ai/dsh-plan-mode service, reusing DSH's plan persistence and exit_plan_mode approval flow, rather than reimplementing a second plan system.
• Tool policy: Plans and quality reviews control tools through 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 layer, not an OS-level sandbox.

Commands

Available in the Web input area or via the command API:

/evolve-mode
/evolve-mode working execute
/evolve-mode working plan
/evolve-mode reasoning standard
/evolve-mode reasoning first-principles
/evolve-mode reasoning grilling
/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 single-mode aliases can still be migrated: normal, first-principles, and adversarial-review. They convert the working state to execution and map reasoning and quality gates according to the legacy modes; the current self-evolution settings are preserved.

Config & compatibility

Plugin version	DeepSeek Harness version	Status
0.4.1	0.1.1-rc.2, 0.1.2-rc.1	Currently supported; compatible with the new value tool unbundling, validated on both versions
0.4.0	0.1.1-rc.2	Legacy compatible version; added Grilling
0.3.2	0.1.1-rc.2	Legacy compatible version
0.3.1	0.1.0-rc.6	Legacy compatible version
0.3.0 and earlier	Not re-validated	No longer supported; upgrade recommended

Since 0.3.2, this table is updated in sync with every plugin release, and a machine-readable Harness compatibility range is declared via peerDependencies.

0.3.2 uses the command attachment parameters and strict storage domain types introduced in Harness 0.1.1-rc.2, so it is not backward-compatible with 0.1.0-rc.6. In production, pin both the plugin and Harness versions; do not rely on floating tags.

Harness 0.1.2-rc.1 removed the legacy per-turn tail and command view slots. The plugin detects capabilities and skips these two legacy UI extension points; review commands and persistent records remain available, but the per-turn inline review is only shown in the 0.1.1-rc.2 UI.

The bundle automatically selects the platform shell. Overrides apply only when the target profile has the tool registered:

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

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. The plugin requires a DeepSeek Harness version that supports the Web plugin loader, client UI slots, storage domain, Trajectory, and the DSH services above.

Plugin state is stored in its own storage domain and can be reloaded across service restarts and session reloads. 0.3.0 automatically migrates legacy session settings, proposals, approved rules, backups, and learning records, without overwriting newer data that already exists.

Feedback

Submit bugs and feature requests to GitHub Issues. Share integration and usage feedback in DeepSeek Harness Discussions.

License

MIT. The Grilling thinking strategy in this project is adapted from Matt Pocock's Grilling skill; see THIRD_PARTY_NOTICES.md for third-party copyright and licensing.