⚡ dsh-agent-conductor

Let your DeepSeek Harness agent dispatch tasks to 11 external agent CLIs — Codex, Claude Code, TraeCode, OpenCode, Gemini, Cursor, Kimi, Qwen, Copilot, WorkBuddy, Grok — headlessly, and bring results back into the conversation.

English · 简体中文

DeepSeek Harness is a great reasoning engine — but sometimes the job is better done by another coding agent: a Codex translation, a Claude Code investigation, a Cursor refactor. This plugin lets your DSH agent recognize when to delegate (by skill-description matching), dispatch a self-contained task to one of 11 external agent CLIs in headless mode, and bring the stdout result back as the answer.

> Inspired by Multica — the "agent squad" idea as a zero-install DSH skill.

🎯 When to use Conductor

• Independent second opinion: ask Codex, Claude Code, or Gemini to review the same design and compare their reasoning.
• Legacy-code investigation: have one agent trace a call chain while another checks risks, regressions, and migration options.
• Cross-model implementation: use an agent whose toolchain or model fits the job, then bring its result back to DSH.

✨ What you get

🧭 Why a Skill (and not just a plugin)

One SKILL.md + a zero-dependency dispatch.py (Python stdlib only).

⚡ Quick start (Skill)

Copy skills/conductor/ to any skill root (project-level .dsh/skills/ or global ~/.dsh/skills/):

mkdir -p .dsh/skills/conductor
cp -R skills/conductor/. .dsh/skills/conductor/

No restart needed — from the next message on, just say:

• "派 codex 把这份 README 翻译成繁体中文" (have Codex translate this README)
• "让 Claude Code 查一下这个报错的成因" (ask Claude Code to investigate this error)
• "用 Codex 独立实现一个 XXX" (have Codex implement XXX independently)

The agent auto-recognizes the need (SKILL.md description matching) → runs dispatch.py → returns the result.

Before an unfamiliar or potentially write-producing dispatch, preview it without consuming the target agent's quota:

python3 skills/conductor/scripts/dispatch.py --dry-run claude-code "Read src/auth.ts and explain the token refresh flow. Do not edit files."

🧭 Agent capability matrix

The matrix is a routing starting point, not a hard rule. Prefer an installed, logged-in CLI that has access to the correct workspace.

📋 Copy-ready dispatch recipes

1. Safe diagnosis:

You are investigating a bug in the current repository. Do not edit files.
Reproduce the issue from this description: .
Read only the minimum relevant files. Return:
1) root cause, 2) evidence with file:line, 3) three ranked fix options,
4) tests that should be added or updated.

2. Independent architecture review:

Review this design as an outside staff engineer. Do not assume hidden context.
Design summary: .
Challenge scalability, failure modes, security boundaries, and migration risk.
Return disagreements first; for each, include impact, evidence, and a fallback.

3. Test-backed implementation:

Implement this change in the current repository: .
Work in small steps. Add or update tests first where practical.
Before finishing, run the relevant test command and report:
files changed, tests run, remaining risks, and follow-up work.

🛠️ Prereqs: install the CLIs you want to dispatch to

# Codex / Claude Code / OpenCode / Gemini / Qwen
npm i -g @openai/codex
npm i -g @anthropic-ai/claude-code
npm i -g opencode-ai
npm i -g @google/gemini-cli
npm i -g @qwen-code/qwen-code
# Kimi / Grok / Copilot (headless flags verified against each CLI's own --help)
npm i -g @moonshot-ai/kimi-code
npm i -g @xai-official/grok
npm i -g @github/copilot
# TraeCode CLI: https://docs.trae.cn/cli_command-line-parameters
# (already have codex-cli? symlink instead: ln -s ~/.codex/plugins/.plugin-appserver/codex ~/.local/bin/codex)

> Headless note: Copilot non-interactive mode must auto-approve tools, so the registry runs copilot -p "" --allow-all-tools. The other CLIs run their standard print/headless flag.

Then check what's actually dispatchable on this machine:

python3 skills/conductor/scripts/dispatch.py doctor
# ✅ Codex  …/codex — codex-cli 0.151.0 …
# ❌ Gemini … 未找到 `gemini`  →  安装：npm i -g @google/gemini-cli

> Working directory is auto-detected. The bundle tool runs each CLI in the current session's workspace (exec.agent.session.header.cwd), then CONDUCTOR_CWD, then the harness cwd — no hardcoded paths. Codex still requires a trusted git repo: point CONDUCTOR_CWD=/path/to/git/repo (env or ~/.dsh/secrets/media-tools.env) at one if the auto-detected folder isn't trusted. The skill script honors CONDUCTOR_CWD and falls back to the current directory.
> To let the dispatched agent write files: add sandbox_mode = "workspace-write" to Codex's ~/.codex/config.toml.
> Dispatching consumes the target CLI's login quota.

✅ Verified vs ⏳ pending

> Install packages were verified against the npm registry: Kimi is @moonshot-ai/kimi-code (bin kimi), Grok is @xai-official/grok (bin grok), Copilot is @github/copilot (bin copilot), Codex is @openai/codex. The kimi-cli npm package is an unrelated placeholder with no binary — do not use it.

📦 Optional: bundle install (host-only tool)

This repo is also a host-only dsh bundle (declares dsh.bundle, zero client code — the Web UI is untouched). One command installs the conductor_dispatch tool into a profile:

dsh plugin --profile web add github:MJorgin/dsh-agent-conductor

• The tool and the skill share the same CLI registry (index.js ⇄ conductor-dynamic.js ⇄ dispatch.py — keep the three in sync when adding CLIs);
• No client half, so the Web UI is never affected (the early panel-carrying client version was removed — see git log);
• Hardened host execution: supports dry-run previews, explicitly declares the subprocess dependency, enforces a real 10-minute timeout that terminates the whole process tree (also on cancel — no orphaned CLIs), and clips over-long output (head + tail, ~20k chars) so a chatty agent can't blow up the context;
• Panels / task-board recycling are on the roadmap.

📂 Repo layout

index.js                      # bundle host half: conductor_dispatch tool (host-only)
cordis.patch.yml              # bundle layer (one row, no client)
skills/conductor/SKILL.md      # skill definition: trigger description + dispatch rules + privacy
skills/conductor/scripts/dispatch.py  # dispatch engine (Python stdlib, zero deps)
conductor-dynamic.js           # alternative: dynamic-plugin edition (cordis_define route)

🗺️ Roadmap

• [x] doctor health check and dry-run preview
• [ ] Routing by capability, speed, and cost
• [ ] Panel UI (optional, dynamic-plugin client half)
• [ ] Task-board recycling: dispatch results written back to dsh-task-board cards
• [ ] Squad orchestration: one task fanned out to several agents and merged (Multica squads shape)

🔑 Keys & privacy

• Keys are never stored in this repo. The skill script reads env vars, then ~/.dsh/secrets/media-tools.env (same convention as dsh-media-skills).
• Task text is sent to the target CLI's provider — never put secrets or internal data into a task.
• Results belong to the target CLI's terms of service; mark deliverables as "done by ".

License

MIT

Related

• 🎨 dsh-media-skills — free vision & image generation for DSH (GLM-4V-Flash + Qwen3-VL + Gemini failover)
• 📬 Community listing PR: awesome-dsh-plugin#664
• 📋 awesome-dsh-plugin · ecosystem index: dsh-plugin topic
• 🎯 Benchmarks: Multica (multi-agent workbench) · dsh-web-ui (task board)