dsh-autoprompt

August 21, 2026 · View on GitHub

English | 中文

Upstream License: MIT

Autoprompt for DeepSeek Harness (DSH) — explicit-only, useful-first orchestration as a native DSH plugin. Invoke /autoprompt <mission> to turn a mission into one independently-approved executable roadmap, build dependency-safe lanes with 25 fixed-persona subagents, and verify the result with blind fresh review.

原作者声明 / Original work attribution

本插件改编自 Spielewoy/autoprompt-skill(MIT 许可,Copyright © 2026 Spielewoy)。Autoprompt 的全部核心内容——编排方法论、SKILL.md 指挥提示、GATES/MODES/PLAYBOOKS 执行契约、25 个 ap-* 角色定义、17 个 task frameworks——均为原作者的成果。

This plugin is an adaptation of Spielewoy/autoprompt-skill (MIT, Copyright © 2026 Spielewoy). All core Autoprompt content — the orchestration methodology, the SKILL.md conductor prompt, GATES/MODES/PLAYBOOKS execution contracts, the 25 ap-* persona definitions, and the 17 task frameworks — belongs to the original author.

本仓库仅做了打包形态的改变:将上游以散装 patch 文件(headless.patch.yml / agent.cordis.yml)交付的 DeepSeek Harness 集成,重打包为原生 cordis 插件(程序化注册子代理 + skill 自动同步)。上游仓库支持 9+ 种 coding agent,本插件只取其 DeepSeek Harness 变体。

This repo only changes the packaging form: upstream's DeepSeek Harness integration shipped as loose patch overlays is repackaged as a native cordis plugin (programmatic subagent registration + automatic skill sync). Upstream supports 9+ coding agents; this plugin carries only its DeepSeek Harness variant.

What it does

When loaded into a DSH profile, the plugin:

  1. Registers the 25 ap-* subagent tools by reusing @deepseek-ai/dsh-tool-subagent's own apply() with each persona config (provider: spawn, maxDepth: 4, per-role tool deny-lists) — equivalent to upstream's patch overlays, but always-on for enabled profiles:
    • L1 coordinators: ap_scope_coordinator, ap_feature_coordinator, ap_sweep_coordinator
    • L2 manager: ap_manager
    • L3 executors: ap_scoper, ap_researcher, ap_synthesizer, ap_planner, ap_implementer, ap_reviewer, ap_verifier, ap_sweeper, ap_execharness_resolver, ap_framework_generator
    • L4 terminal leaves: ap_fresh_verifier, ap_depth_prober, ap_framework_validator, ap_juror, ap_goal_checker, ap_arbiter, ap_re_anchor, ap_scribe, ap_janitor
  2. Syncs the autoprompt skill tree (SKILL.md conductor prompt + GATES.md / MODES.md / PLAYBOOKS.md contracts + agents/ + frameworks/) into the agent skills directory (~/.agents/skills/autoprompt by default), so /autoprompt <mission> is invocable. Idempotent and content-aware.

Usage

/autoprompt <mission>
  • Attended session: the conductor asks undefined knobs (concurrency tokensaver/wide/custom, agents off) in one question before starting.
  • Model/effort routing is inherited-only: every role inherits the session's parent model.

Installation

  1. Clone to any location (e.g. E:\Deepseek Harness\dsh-autoprompt).
  2. Junction it into a DSH profile's node_modules:
    mklink /J "C:\Users\<you>\.dsh\profiles\desktop\node_modules\dsh-autoprompt" "E:\Deepseek Harness\dsh-autoprompt"
    
  3. Register in the profile's cordis.patch.yml:
    - insert:
        - id: dsh-autoprompt
          name: 'dsh-autoprompt'
    
  4. Restart DSH Desktop.

The plugin's own node_modules must resolve @deepseek-ai/dsh-tool-subagent, @deepseek-ai/dsh-tools, @deepseek-ai/schemastery and cordis — junction it to the DSH app's unpacked dependencies:

mklink /J "E:\Deepseek Harness\dsh-autoprompt\node_modules" "C:\Program Files\DSH Desktop\resources\app.asar.unpacked\node_modules"

Configuration

OptionDefaultDescription
enabledtrueMaster switch (false skips both subagent registration and skill sync)
skillDir~/.agents/skillsSkills directory override (env: DSH_SKILL_DIR)

Requirements

  • DeepSeek Harness ≥ 0.1.0-rc.7 (tested against 0.1.1-rc.1)
  • The spawn subagent provider (bundled with dsh-base)

License

MIT — see LICENSE. Original Autoprompt content © 2026 Spielewoy; plugin packaging © 2026 Asaka-RUM.