oil-oil/codex-deepseek-subagent

codex‑deepseek‑subagent – Make DeepSeek a native sub‑Agent for the Codex desktop app

What it is

• A Skill (plugin) for the Codex desktop application (the UI that ships with OpenAI’s ChatGPT/Codex product). The skill registers DeepSeek’s large‑language‑model API as a native sub‑Agent that the Codex app can call directly.
• It does only configuration, health‑checking and repair of the DeepSeek integration – it does not perform regular coding or generation tasks itself.

Why you might need it

• Codex can spawn sub‑Agents that run on other LLM providers. Without this skill you would have to manually edit configuration files, manage API keys, and verify that the DeepSeek model is correctly hooked up.
• The skill automates the whole workflow: install, select a DeepSeek model (Flash or Pro), supply the API key via the OS credential store, run a verification session, and then let Codex call spawn_agent(agent_type="DeepSeek") in normal tasks.

Key capabilities

Feature	What it does
Global install	npx skills add oil-oil/codex-deepseek-subagent -g -y registers the skill for every Codex user on the machine.
Model selection	During the first configuration you pick either deepseek‑v4‑flash (cheaper, faster) or deepseek‑v4‑pro (more capable).
Credential handling	On macOS the API key is stored in the Keychain, on Windows in Credential Manager. The skill never sends the key in chat.
Status / test / repair commands	CLI scripts (`codex_deepseek.py status
Automatic backup & rollback	Before writing any config or model files a backup is created; on failure the previous state is restored.
Compatibility checks	The skill validates that the parent Codex model matches the selected DeepSeek model and that the returned token NATIVE_DEEPSEEK_OK is present.
No visual input	DeepSeek handles only text; any image/video must be pre‑processed by the main Codex Agent.

Installation & first‑time setup

1. Prerequisites – macOS or Windows, Python 3.11+, the Codex desktop app installed, and a DeepSeek API key.
2. Run the global install command shown above.
3. Restart Codex, start a new chat, and type:

   帮我把 DeepSeek 配置成 Codex 的原生子 Agent。
   

4. Choose the desired DeepSeek model when prompted.
5. If the skill needs the key, a small UI (Node 22+ required) opens; you paste the key and it is saved to the system credential store.
6. After the skill reports status: ready, restart Codex again. From then on you can simply say:

   用 DeepSeek 子 Agent 检查这个项目。
   

   and Codex will invoke the DeepSeek sub‑Agent.

Managing the integration

• Status: python3 scripts/codex_deepseek.py status --json
• Test (runs an isolated acceptance session): ... test --json
• Repair / switch model: ... repair --model deepseek-v4-pro --json
• Disable: ... disable --json
• Uninstall: ... uninstall --json

On Windows you can use the equivalent py -3 commands; if Codex’s executable isn’t found automatically, set CODEX_DESKTOP_BIN to the path of codex.exe.

Safety & rollback

• API keys are never echoed to the console or sent over chat; they are stored only in the OS‑provided secure store.
• All writes are transactional – a backup is kept and restored automatically on failure.
• The skill never changes the main Codex model or login method; it only touches the DeepSeek sub‑Agent configuration.

Licensing

• Distributed under the MIT License.

Where to learn more

• Compatibility & routing details: codex-deepseek-subagent/references/compatibility.md
• Skill execution rules: SKILL.md
• API‑key UI guide: references/api-key-setup.md
• Official DeepSeek model list & Codex install script links are provided in the README for reference.

---

This explanation is based solely on the repository’s README; no additional features have been inferred.