dsh-task-modes

中文文档

An independent, installable DeepSeek Harness Web bundle that adds three task modes: normal execution, first-principles prompting, and an independent forked adversarial review. It works through the DSH plugin layer and does not modify DSH core.

Install

Install the pinned npm release into a Web profile without requiring a global DSH installation:

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

If dsh is already installed globally, use the shorter command:

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

Restart the Web profile after installation. The mode selector appears beside the composer tools.

For source auditing or development, install the pinned Git revision instead:

dsh plugin --profile web add github:GraySilver/dsh-task-modes#ca239e55f382816d0d1444f4eb9742e22945018e

Git-hosted plugins execute install-time code, so install only revisions you trust.

Modes

Mode	Behavior	Best for
Normal mode	Sends the request without extra mode instructions.	Everyday work.
First Principles	Injects a system-prompt section that asks the model to state objectives, separate facts from assumptions, identify constraints, derive the approach, and verify the result. Trajectory projects that exact section from the persisted request header as a context-style inspection row.	Ambiguous or high-leverage decisions.
Adversarial Review	After the parent text answer completes, starts a forked child agent to inspect the current task and answer, then renders a Markdown report below that reply.	Catching omissions and unsupported assumptions.

The reviewer can inspect with read, glob, grep, read_image, and the configured platform shell (bash on macOS/Linux or pwsh on Windows). Its prompt requests non-mutating inspection and no background processes. A reviewer failure is recorded as unavailable and never blocks the parent answer.

Cost and limits

• Adversarial Review adds one model call and corresponding latency per completed parent answer.
• The report is advisory. It does not automatically rewrite or retry the parent answer.
• The child uses the tools registered in the profile; tool restrictions in the prompt are not an operating-system sandbox.
• The plugin shares the Harness process privileges. Treat it as trusted code and install pinned revisions.

Persistence and commands

The bundle stores its records in the graysilver_task_modes storage domain. It does not add custom DSH session event types, so it remains compatible with released DSH persistence while mode selections and review reports survive service restarts and session reloads.

Use these commands in the Web composer:

• /task-mode shows the current mode.
• /task-mode <normal|first-principles|adversarial-review> switches modes.
• /task-mode review <turn> opens one review report.
• /task-mode reviews lists saved reports.

The Web client renders each report as collapsed Markdown beneath its corresponding AI reply. Expanding it opens a fixed-height, scrollable panel.

The first-principles Trajectory row is inspection-only. It is derived from a persisted request/header.system that contains the exact guidance; it does not append a user message or change the model request. Switching to another mode removes the section from the next model request, while earlier Trajectory rows remain as evidence of what those historical requests used.

Configuration

The bundle selects the normal platform shell automatically. Override it only when the target profile registers that tool:

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

Adversarial review requires the fork/subagent capability in the selected profile.

Compatibility

Requires a DeepSeek Harness release that provides the Web plugin loader, client UI slots, storage domains, and forked subagents. npm is the recommended stable distribution channel; pinned Git revisions remain available for source auditing and development.

Release

See v0.1.9 for the current release.

Feedback

Please use GitHub Issues for bugs and feature requests. Showcase and integration feedback is welcome in the DeepSeek Harness Discussions.

License

MIT