dsh-agent-conductor — in-session cross-agent dispatch for DeepSeek Harness # ⚡ 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.* [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE) [![CI](https://github.com/MJorgin/dsh-agent-conductor/actions/workflows/ci.yml/badge.svg)](https://github.com/MJorgin/dsh-agent-conductor/actions/workflows/ci.yml) [![DeepSeek Harness](https://img.shields.io/badge/DeepSeek%20Harness-plugin-4D6BFE)](https://github.com/topics/dsh-plugin) [![Zero deps](https://img.shields.io/badge/deps-zero-2EA44F)](skills/conductor/scripts/dispatch.py) [![Agents](https://img.shields.io/badge/agents-11-blue)](README.md#-agents) **English** · [**简体中文**](docs/lang/README_ZH.md)
--- 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](https://github.com/multica-ai/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 | Capability | What it does | Cost | |---|---|---| | 🧠 Auto-triggered dispatch | Say "have Codex translate this README" — the model matches the skill, runs the dispatch script, and answers from the result | Free (uses the target CLI's quota) | | 🔧 `conductor_dispatch` tool (optional bundle) | The same registry as a first-class DSH tool, installed into a profile with one command | Free | | 🩺 `doctor` self-check | `python3 dispatch.py doctor` probes which CLIs are installed (resolves PATH + tries `--version`) before you dispatch | Free | | 👀 Dry-run preview | Preview the target, working directory, argv, and shell-equivalent command without starting an external agent | Free | | 👥 11 agent CLIs | Codex, Claude Code, TraeCode, OpenCode, Gemini, Cursor, Kimi, Qwen, Copilot, WorkBuddy, Grok | Their login quotas | | 🔒 Privacy | Task text goes only to the CLI's own provider; keys stay local | — | ## 🧭 Why a Skill (and not just a plugin) | | profile plugin / bundle | dynamic plugin | **Skill (this repo)** | |---|---|---|---| | Install | write profile + restart | define in-session | **copy a folder** | | Trigger | manual | model calls a tool | **description matching, model auto-recognizes** | | Risk | touches the host | session-scoped, gone on restart | read-only script, host-agnostic | | Result | tool result | tool result | **stdout directly becomes the answer** | 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/`): ```sh 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: ```sh 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 | Task | First choices | Why | |---|---|---| | Architecture review / second opinion | Codex, Claude Code, Gemini | Different reasoning styles are useful for catching design blind spots | | Legacy-code call-chain investigation | Claude Code, Codex, Copilot | Good for tracing repository-local dependencies and attaching evidence | | GitHub-centric issue or Actions analysis | Copilot, Codex | Natural fit for repositories, pull requests, and CI diagnostics | | Editor-coupled refactor | Cursor, Codex | Useful when the workflow is tightly connected to an existing project tree | | Chinese localization / Chinese docs | TraeCode, Qwen, Kimi, Codex | Strong Chinese-language options; keep terminology consistent in the task | | Fast alternative implementation | OpenCode, Qwen, Kimi | Useful for bounded spikes where a separate login/toolchain is available | 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: ```text 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: ```text 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: ```text 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 ```sh # 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: ```sh 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 | CLI | Headless command | Status | |---|---|---| | Codex | `codex exec "{task}"` | ✅ field-tested (translation task delivered) | | Claude Code | `claude -p "{task}" --output-format text` | ✅ installed; per official docs | | TraeCode | `traecli exec "{task}"` | ✅ per official docs | | OpenCode | `opencode run "{task}"` | ✅ per official docs | | Gemini CLI | `gemini -p "{task}"` | ✅ per official docs | | Qwen Code | `qwen --prompt "{task}"` | ✅ per official docs (Gemini-CLI fork) | | Kimi CLI | `kimi --prompt "{task}"` | ✅ flag confirmed from the CLI's own help (`-p, --prompt` = non-interactive) | | Copilot CLI | `copilot -p "{task}" --allow-all-tools` | ✅ flag confirmed from `copilot --help` (bin is `copilot`, not `github-copilot`; `--allow-all-tools` is required headless) | | Grok CLI | `grok -p "{task}"` | ✅ per official README (`grok -p "..."` = run one task) | | Cursor CLI | `cursor-agent -p "{task}"` | ⏳ command shape pending field test (install via cursor.com; the npm `cursor-agent` package is unrelated) | | WorkBuddy | `workbuddy -p "{task}"` | ⏳ command shape pending field test | > 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: ```sh 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](LICENSE) ## Related - 🎨 [dsh-media-skills](https://github.com/MJorgin/dsh-media-skills) — free vision & image generation for DSH (GLM-4V-Flash + Qwen3-VL + Gemini failover) - 📬 Community listing PR: [awesome-dsh-plugin#664](https://github.com/awesome-dsh-plugin/awesome-dsh-plugin/pull/664) - 📋 [awesome-dsh-plugin](https://github.com/awesome-dsh-plugin/awesome-dsh-plugin) · ecosystem index: [dsh-plugin topic](https://github.com/topics/dsh-plugin) - 🎯 Benchmarks: [Multica](https://github.com/multica-ai/multica) (multi-agent workbench) · [dsh-web-ui](https://github.com/zhu1090093659/dsh-web-ui) (task board)