Introduction

The DSH plugin ecosystem emphasizes “everything is a plugin.” For long tasks that require breakdown, execution, checking, and rework, different models can be placed in different roles: Codex is responsible for planning, scoring, and reviewing, DSH is responsible for execution, and Kimi is responsible for summarizing, cross-validation, and visual acceptance.

Below is an introduction to the positioning, core capabilities, installation methods, typical usage, and configuration notes for dsh-longtask-orchestrator, including Codex, Kimi, and state directory.

Positioning

dsh-longtask-orchestrator is a DSH long-task orchestration plugin maintained by rocker2018-droid.

Its one-line positioning is:

Long task orchestration: Codex plan/score/review + DSH execution + Kimi supplement (visual acceptance/summary/cross-validation)

The license is:

MIT

Core Capabilities

This plugin organizes long tasks into a set of callable steps, rather than relying on a single round of prompts to complete the task at once.

The main capabilities include:

  1. Long task initialization: longtask_begin is planned by Codex and written to state.json.
  2. Plan decomposition: codex_plan generates a JSON subtask list.
  3. Execute subtasks: Executed by DSH’s agent tools or subagent.
  4. Result scoring: codex_score scores 0-10 and provides modification suggestions.
  5. Result supplementation: kimi_summarize does summary compression, kimi_review does cross-validation, and kimi_call does visual acceptance.
  6. Final review: codex_review performs the final audit.
  7. State read/write: orchestrate_state reads and writes the long task state.
  8. Auto-trigger: The skill longtask-orchestrator provides loop templates to automatically trigger long tasks.
  9. Model groups: “DeepSeek + Long Task Orchestration” is a delegating adapter; once selected, it injects orchestration prompts every round.

The plugin registers a total of 9 tools. The tool names are as follows:

codex_call
kimi_call
codex_plan
codex_score
codex_review
kimi_summarize
kimi_review
orchestrate_state
longtask_begin

Notably, codex_call is the underlying Codex bridge, utilizing a read-only sandbox, auto-agent, and serial queue.

Installation and Activation

The installation command provided in the documentation uses a local link path; no remote GitHub installation command is provided.

Installation command:

dsh plugin --profile web add link:/绝对路径/dsh-longtask-orchestrator

After installation, you need to restart dsh web for it to take effect:

dsh web

Typical Usage

The following is a flow that you can follow.

  1. In a new session, switch to the “DeepSeek + Long Task Orchestration” model group; you can also directly propose a long task, and the skill longtask-orchestrator will trigger automatically.
  2. Initialize the long task, for example:
longtask_begin(goal)

This step is planned by Codex and written to state.json.
3. Execute the subtasks one by one based on the generated list.
4. Call codex_score to score the result of each subtask, and continue iterating based on the modification suggestions.
5. After all pass, call codex_review for the final review.
6. Output the final report.

Configuration Guide

Codex Binary

Default Codex binary path:

/Applications/ChatGPT.app/Contents/Resources/codex

If this path does not exist, it will fall back to codex in PATH.

Available override:

ORCHESTRATE_CODEX_BIN

Proxy

The plugin automatically detects the macOS system proxy:

scutil

The subprocesses will carry:

HTTPS_PROXY
HTTP_PROXY

Kimi

Kimi configuration reuse:

~/.dsh/vision.env

Related configuration items include:

VISION_API_KEY
BASE_URL
MODEL

Available override:

ORCHESTRATE_KIMI_*

State Directory

Default state directory:

<工作区>/.orchestrate

Available override:

ORCHESTRATE_STATE_DIR

Applicable Scenarios and Notes

It is suitable for handling long tasks in DSH that require planning, execution, scoring, and review. Before use, it is recommended to pay attention to the following points:

  • The plugin runs with the permissions of the current dsh process; you should check the source code and license before installing. The license for this project is MIT.
  • Codex uses ChatGPT quota and shares the login state with Codex++.
  • codex calls are serialized; try to use Codex++ sparingly during long tasks.
  • If Codex is unavailable, the task will pause with a prompt and will not use Kimi to replace planning/review.
  • codex_call serves as the underlying Codex bridge, featuring a read-only sandbox, auto-agent, and serial queue.

Links

Directory page:

https://www.skillhub.cn/plugins/rocker2018-droid/dsh-longtask-orchestrator

GitHub:

https://github.com/rocker2018-droid/dsh-longtask-orchestrator

The directory page here is an independent community directory, not an official app store; it has no official affiliation with DeepSeek or Huansuan.