DeepSeek Harness Evolve Modes

English｜中文

Make the Agent's way of working composable, reviewable, and continuously improvable, ultimately achieving Agent Self Evoling.

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

The plugin does not fork Harness, does not copy the Agent loop, and does not modify core code. After installing, the current task's combination is always shown next to the input area; the global "Self-Evolution Mode" 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 Installation

We recommend installing 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 you have the DSH CLI installed globally, you can use the short form:

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

After restarting the Web profile, the self-evolution mode control appears next to the input-area tools. Open the top-level Self-Evolution Mode setting to manage global learning rules.

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

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

The Git install package executes code at install time; only install trusted revisions.

Feature Overview

输入区会显示当前组合，例如：

正常 · 标准 · 关 · 进化 开

Open the controls to adjust the four dimensions:

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 work through first principles, or stress-test requirements and decisions through round-by-round follow-up questions.
Quality gate	Off · Adversarial review · Acceptance review	No extra review, independently look for risks, or verify results against the task and any approved plan.
Self-evolution	Off · On	Automatically analyze sessions, generate rule proposals for human review, and auto-optimize AGENTS.md (but never modifies AGENTS.md)

这不是互斥的“人格模式”，而是每个任务都可以重新组合的一组工作决策。

Self-Evolution Mode

Self-evolution identifies stable user identity, preferences, and work requirements from completed Agent collaborations. It defaults to Propose, which only proposes candidate rules without automatically changing subsequent behavior.

Default Settings

Setting	Default	Description
Self-evolution	Propose	On by default for new sessions and for old sessions without an explicit saved self-evolution choice.
Learning batch	3	Starts one learning run every 3 completed parent Agent replies accumulated. Adjustable to 1..100 in global settings.
Pending proposal cap	100	Proposals stop piling up once the cap is reached. Adjustable to 1..1000 in global settings.
Learning scope	Source sessions where self-evolution is on	Project scope is no longer configured; if self-evolution is open for the current session, it is included in the learning scope.
Rule scope	Global	Rules that are applied take effect across sessions and are not bound to a project directory.

How Learning Data Is Organized

• Each source session takes at most the 100 most recent learning messages.
• Keeps the complete user message for each round, plus the last visible assistant message of that round.
• Assistant messages are context only; when exceeding 2000 characters, the first 1000 and last 1000 are kept with ... in between.
• Proposal evidence must be quoted verbatim from user messages. Assistant inferences, one-off task details, implementation results, silence, or "not mentioned again" cannot serve as sole rule evidence.

Learning Requests Are Isolated

Each learning run uses a dedicated single learning persona/system prompt for the plugin, passing the current batch as a structured JSON user message. A learning request:

• Does not inherit the parent session's history;
• Does not inherit the parent Agent's working context;
• Does not create learning sub-Agents;
• Does not carry tools;
• Does not load AGENTS.md or CLAUDE.md from the source session's working directory;
• Only analyzes identity information, preferences, and work requirements that may be valid long-term.

学习失败会记录在设置页，未完成的批次会保留，方便之后重试；不会阻断父 Agent 的正常回复。

Proposals Require Human 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 the learning batch size and the pending proposal cap;
• View each proposal's category, inference type, and original user evidence;
• Apply or ignore proposals;
• Manually add, modify, and delete global rules;
• View learning run logs and failure reasons;
• Restore from the backup auto-created 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 file.

为任务选择组合

You need to…	Recommended combo	Why it fits
Get routine work done fast	正常 · 标准 · 关	Keeps execution pace without extra process.
Make high-impact decisions	计划 · 第一性原理 · 关	Research first, surface assumptions, then enter DSH plan-approval flow.
Deliver an implementation with confidence	正常 · 标准 · 验收审查	After implementation, an independent review Agent checks results against task goals.
Challenge a high-risk answer	正常 · 第一性原理 · 对抗性审查	Unfold reasoning explicitly, then look for gaps, counterexamples, regressions, unsupported conclusions.
Distill stable personal preferences	正常 · 标准 · 关 · 进化 开	Identify long-term rules in batches of 3 replies by default; proposals only, no auto-enable.

Quality Gates

Adversarial Review

在父 Agent 回复完成后启动独立审查 Agent，检查未满足要求、缺少依据的结论、遗漏、回归、反例和安全风险。它只报告证据、缺口和后续行动，不会静默改写、重试或修复父回复。

Acceptance Review

Checks the task, candidate answer, and any approved plan. The report is fixed to distinguish:

Met
Gap
Unverified
Evidence
Concrete follow-up

Review reports appear below the corresponding assistant reply. Each quality review adds one model call and latency per completed parent reply, but never auto-runs project 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 as auditable evidence. Disabling it only affects subsequent requests; historical evidence is not deleted.
• Grilling: First establishes decisions and their dependencies; each round raises only questions whose preconditions are already clear; questions use a numbered list with a recommended answer. The Agent investigates discoverable facts on its own, summarizes the shared understanding once all branches are clear, and waits for user confirmation before implementing. After confirmation, it continues in the current "Normal/Plan" working mode.
• Plan mode: Delegates to the official @deepseek-ai/dsh-plan-mode service, reusing DSH's plan persistence and exit_plan_mode approval flow instead of reimplementing another planning system.
• Tool policy: Planning and quality review control tools 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 layer, not an OS-level sandbox.

Commands

Available in the Web input area or 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

The 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 per the old mode; 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 new value tool unbundling, verified across both versions
0.4.0	0.1.1-rc.2	Legacy compatible version; adds 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-verified	No longer supported; upgrade recommended

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

0.3.2 uses the new command attachment parameter and strict storage domain types added 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 skips these two legacy UI extension points via capability detection; the review command and persistent records remain available, but the per-turn inline review is only displayed in the 0.1.1-rc.2 UI.

The bundle auto-selects the platform shell. It is only overridden when the target profile has registered that tool:

- 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 survives service restarts and session reloads. 0.3.0 auto-migrates session settings, proposals, approved rules, backups, and learning records from older versions, without overwriting data that already exists under the new names.

Feedback

Please file bugs and feature suggestions in 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 license.