Introduction¶
DeepSeek Harness (DSH) extends conversational capabilities in a plugin-based manner, and the community directory is a standalone site, not indicating any official affiliation with DeepSeek / High-Flyer.
In agent development, a common scenario is that DSH conversations can handle reasoning and orchestration, but certain tasks are better suited to external coding agent CLIs that have already integrated with the corresponding provider. For example, having Codex translate a README or using Claude Code to troubleshoot an error message.
MJorgin/dsh-agent-conductor is designed for this scenario: it dispatches self-contained tasks to 11 external agent CLIs within a DSH session and brings the target CLI’s stdout results back to the current conversation.
What This Is¶
dsh-agent-conductor is maintained by MJorgin and licensed under MIT.
It provides two integration methods:
- Zero-dependency skill: Copy
skills/conductor/to the project-level or global skill directory, and useSKILL.mdto describe matching triggers for delegation. - Optional host-only bundle: Install
conductor_dispatchinto the DSH profile, invoking it as a DSH tool.
The target CLIs include Codex, Claude Code, TraeCode, OpenCode, Gemini, Cursor, Kimi, Qwen, Copilot, WorkBuddy, and Grok—a total of 11.
Core Features¶
In-Session Delegation¶
When a task in the conversation matches the description in SKILL.md, the model runs dispatch.py to delegate the task to the target CLI. The target CLI’s stdout results are then brought back to the DSH conversation.
Multiple Target CLIs¶
The repository lists 11 external agent CLIs as delegation targets:
- Codex
- Claude Code
- TraeCode
- OpenCode
- Gemini
- Cursor
- Kimi
- Qwen
- Copilot
- WorkBuddy
- Grok
Optional Host-Only Tool¶
In addition to the skill, the repository provides a conductor_dispatch tool that can be installed into the DSH profile. This is suitable for scenarios where you want to invoke the same CLI delegation logic at the profile level.
Task Text and Keys¶
Task text is only sent to the target CLI’s own provider; keys remain local. Avoid including keys or internal data in the task text when using this tool.
Codex Working Directory¶
Codex requires a trusted git repo as the working directory. You can specify the target directory via CONDUCTOR_CWD.
Installation and Enabling¶
Installing the Skill¶
Copy skills/conductor/ to the project-level .dsh/skills/ or global ~/.dsh/skills/:
mkdir -p .dsh/skills/conductor
cp -R skills/conductor/. .dsh/skills/conductor/
Once done, you can trigger delegation using natural language in a DSH session.
Installing the Host-Only Bundle¶
If you want to install conductor_dispatch as a DSH profile tool, use:
dsh plugin --profile web add github:MJorgin/dsh-agent-conductor
Typical Usage¶
Preparing the Target CLI¶
Before dispatching tasks, ensure the target CLI is installed and available. Using Claude Code and OpenCode as examples:
npm i -g @anthropic-ai/claude-code
npm i -g opencode-ai
Other CLIs should be prepared according to their respective installation methods.
Triggering with Natural Language¶
In a DSH session, you can write:
have Codex translate this README- “派 codex 把这份 README 翻译成繁体中文”
Once triggered, the process is: the model matches the task based on the SKILL.md description, runs dispatch.py, and then brings the target CLI’s stdout results back to the conversation.
Setting the Working Directory for Codex¶
If dispatching a task to Codex, you need to specify a trusted git repo as the working directory. You can add:
# ~/.dsh/secrets/media-tools.env
CONDUCTOR_CWD=/path/to/git/repo
Allowing Codex to Write Files¶
If you want to allow the delegated agent to write files, you need to configure Codex as follows:
sandbox_mode = "workspace-write"
Use Cases and Notes¶
This tool is suitable for developers who are already using multiple external agent CLIs and want to trigger tasks uniformly from within a DSH session. Before use, note:
- The plugin runs with the current DSH process permissions; inspect the source code and license before installation. This project is licensed under MIT.
- Dispatching tasks consumes the target CLI’s login quota.
- Task text is sent to the target CLI’s provider and should not contain keys or internal data.
- Results are subject to the target CLI’s terms of service, and deliverables should be marked as completed by the corresponding agent.
- The headless command shapes for some CLIs (Gemini, Cursor, Kimi, Qwen, Copilot, WorkBuddy, Grok) are still marked as pending verification.
Conclusion¶
dsh-agent-conductor connects DSH conversations with the execution results of external agent CLIs: DSH handles conversation and task identification, while the target CLI handles execution, with stdout results returned to the current conversation.
It provides both a zero-dependency skill and an optional host-only bundle, allowing integration at the project-level, global level, or profile level.
The community directory page URL was not provided in the verified materials. If you need to locate it, search for the plugin name dsh-agent-conductor in the DeepSeek Harness community directory. The GitHub repository address is:
https://github.com/MJorgin/dsh-agent-conductor