v0.3.1 - Scrollable Evolve Mode Menu

This patch release improves the self-evolution mode menu in constrained viewports.\n\n- Keeps the menu at a stable 420px design height when space permits.\n- Automatically places the menu above or below the trigger.\n- Reduces the menu height when the viewport has less available space.\n- Enables internal vertical scrolling and prevents scroll chaining.\n- Publishes the full Chinese README as the npm package README.\n- Includes no behavior or storage migration changes.

v0.3.0 - dsh-evolve-modes Self-Evolution

Install

Install the GitHub release tarball:

dsh plugin --profile web add https://github.com/GraySilver/dsh-evolve-modes/releases/download/v0.3.0/graysilver-dsh-evolve-modes-0.3.0.tgz

Restart the Web profile after installation.

Renamed Identity

• Repository: GraySilver/dsh-evolve-modes
• Package: @graysilver/dsh-evolve-modes
• Cordis/plugin ID: dsh-evolve-modes
• Commands: /evolve-mode and /evolve-mode-review
• Remote namespace: evolveModes
• Storage domains: graysilver_dsh_evolve_modes and graysilver_dsh_evolve_modes_evolution
• Learned-instruction marker: <dsh-evolve-modes-learned-instructions>

Existing session settings, proposals, approved rules, backups, and learning runs are copied from the former storage domains on first startup under the new identity. Existing new-domain data is never overwritten.

Highlights

• Adds global, human-reviewed self-evolution settings and pending-proposal management.
• Self-evolution defaults to Propose, with a default learning batch of 3 completed replies.
• Learning uses one dedicated system prompt and one explicit JSON data message, without inherited parent history, work context, tools, a learning subagent, or source-workspace AGENTS.md.
• Keeps up to 100 source messages per session and truncates long assistant context to 2000 characters.
• Includes evidence validation, one repair retry, failed-run retention, rename migration tests, and isolated LLM request tests.

Full details: https://github.com/GraySilver/dsh-evolve-modes/blob/v0.3.0/README.md

v0.2.0 - Composable Task Modes

Install

Install the public npm bundle:

npx -y @deepseek-ai/dsh plugin --profile web add @graysilver/dsh-task-modes@0.2.0

If dsh is installed globally:

dsh plugin --profile web add @graysilver/dsh-task-modes@0.2.0

Restart the Web profile after installation.

One Control, Three Decisions

v0.2.0 replaces the old single mode selector with one compact task-mode control. Compose each session from three independent dimensions:

• Working state: Execute or Plan. Plan reuses the official DSH Plan service, persisted plan/mode event, and exit_plan_mode approval workflow.
• Reasoning strategy: Standard or First principles. First-principles guidance is persisted in request/header.system and visible in Trajectory as evidence of the model input.
• Quality gate: Off, Adversarial review, or Acceptance review.

The control summarizes the active combination, for example Execute · Standard · Off.

Independent Quality Reviews

• Adversarial review runs a forked reviewer after each completed parent response to find unmet requirements, unsupported claims, omissions, regressions, counterexamples, and security risks.
• Acceptance review compares the task, candidate answer, and an approved plan when available. Its Markdown report separates Met, Gap, Unverified, Evidence, and Concrete follow-up.
• Reports render as collapsed Markdown below the corresponding assistant reply, with a fixed-height scrollable body.

Reviews are advisory. They never silently rewrite, retry, or fix the parent result.

Compatibility and Limits

• Existing normal, first-principles, and adversarial-review selections migrate automatically to the new reasoning and quality settings.
• The command value general-review remains supported for compatibility; the visible product name is Adversarial review.
• Plan mode permits read, glob, grep, read_image, the configured shell, and exit_plan_mode. The shell remains fully capable, so Plan mode is not an operating-system sandbox.
• Quality reviews add one model call and latency per completed parent response.

For source auditing, install the pinned release revision:

dsh plugin --profile web add github:GraySilver/dsh-task-modes#9163f0fab972aef43d897ce4bb4781aafc50cad6

See the English README or 中文 README for configuration, persistence, and compatibility details.

v0.1.9 - Task Modes for DeepSeek Harness Web

Install

Install the public npm bundle:

npx -y @deepseek-ai/dsh plugin --profile web add @graysilver/dsh-task-modes@0.1.9

If dsh is installed globally:

dsh plugin --profile web add @graysilver/dsh-task-modes@0.1.9

Restart the Web profile after installation.

First Principles Trajectory

• First Principles remains a real System Prompt section for the next model request.
• Trajectory now shows the exact injected prompt as a context-style row, derived from the persisted request/header.system that was actually sent.
• Switching to another mode removes that section from the next request; the historical Trajectory evidence remains.
• The visualization does not append a Session event, user message, or any extra model-visible input.

Modes

• Normal mode leaves requests unchanged.
• First Principles injects a structured system-prompt section.
• Adversarial Review forks an independent child agent after each completed text answer and renders a compact Markdown report beneath it.

Limits

Adversarial Review adds one model call and latency per parent answer. Reports are advisory and do not rewrite parent answers. Prompt-level reviewer tool restrictions are not an operating-system sandbox; the plugin shares Harness process privileges.

For source auditing, install the pinned Git revision:

dsh plugin --profile web add github:GraySilver/dsh-task-modes#a093bd23d1ae7e4beb009ed058360d8a7d39a3b1

See the English README or 中文 README for configuration, persistence, and compatibility details.

v0.1.8 - Task Modes for DeepSeek Harness Web

Install

Install the public npm bundle without a global DSH installation:

npx -y @deepseek-ai/dsh plugin --profile web add @graysilver/dsh-task-modes@0.1.8

If dsh is installed globally:

dsh plugin --profile web add @graysilver/dsh-task-modes@0.1.8

Restart the Web profile after installation. Adversarial Review requires its fork/subagent capability.

npm release

0.1.8 is the first public npm release. The package ships its prebuilt DSH entries, TypeScript declarations, bundle patch, bilingual documentation, and real Web UI screenshots.

Modes

• Normal mode leaves requests unchanged.
• First Principles injects a structured system-prompt section.
• Adversarial Review forks an independent child agent after each completed text answer and renders a compact Markdown report beneath it.

Limits

Adversarial Review adds one model call and latency per parent answer. Reports are advisory and do not rewrite parent answers. Prompt-level reviewer tool restrictions are not an operating-system sandbox; the plugin shares Harness process privileges.

For source auditing, install the pinned Git revision:

dsh plugin --profile web add github:GraySilver/dsh-task-modes#4ec9f5f63679784ef6ce248aae42e373d7a8d049

See the English README or 中文 README for configuration, persistence, and compatibility details.

v0.1.7 - Task Modes for DeepSeek Harness Web

Install

Install the pinned Web bundle revision:

dsh plugin --profile web add github:GraySilver/dsh-task-modes#4ec9f5f63679784ef6ce248aae42e373d7a8d049

Restart the Web profile after installation. Adversarial Review requires the fork/subagent capability in the profile.

What is included

• Normal mode for everyday requests.
• First Principles mode, which injects a structured system-prompt section.
• Adversarial Review mode, which forks an independent child agent after each completed parent answer and renders a compact Markdown report beneath it.
• Reviewer inspection tools: read, glob, grep, read_image, and platform shell (bash/pwsh).

Limits

Adversarial Review adds one model call and latency per parent answer. Reports are advisory and do not rewrite the parent answer. Prompt-level tool restrictions are not an operating-system sandbox; the plugin shares Harness process privileges.

See the English README or 中文 README for commands, configuration, and compatibility details.

v0.1.6

Hide internal automatic review lookups from the chat transcript.\n\nInstall this release instead of v0.1.4 or v0.1.5.

v0.1.5

Fix the review panel's chain-slot registration.\n\nInstall this release instead of v0.1.4.

v0.1.4

Render adversarial reviews beneath their corresponding AI replies.\n\nReviews are collapsed by default; expanded Markdown is contained in a fixed-height scrollable panel.

v0.1.3

Fix the Web task-mode selector by binding the commands Remote explicitly.\n\nThe first-principles and adversarial-review options now execute /task-mode commands, and command failures are visible in the composer.