skills.rest
Zedong-Liu
Zedong-LiuCOMMUNITY
Explore all skills by this author →
GitHub

deepseek-codex-subagent

Delegate bounded coding tasks to DeepSeek-backed Codex CLI subagents with persistent sessions.

21|3|Updated Jun 1, 2026
Search Tags:#code review#codex#session management#subagent#task delegation#deepseek#api proxy
One-click install
npx skills add https://github.com/Zedong-Liu/codex-deepseek-sidecar --skill deepseek-codex-subagent-zedong-liu
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: deepseek-codex-subagent
Source: https://github.com/Zedong-Liu/codex-deepseek-sidecar
Command: npx skills add https://github.com/Zedong-Liu/codex-deepseek-sidecar --skill deepseek-codex-subagent-zedong-liu

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) components.

What problem does it solve? Running long tests, log analysis, broad code exploration, and independent reviews on an expensive frontier model consumes large token budgets. This Skill lets the main Codex agent offload those bounded, token-heavy worker tasks to cheaper DeepSeek models while keeping planning and synthesis on the primary model. ## Core Features & Use Cases - Sidecar task delegation: Launch DeepSeek V4 Flash or Pro workers through a stable wrapper script with named task IDs, status checks, blocking waits, and resumable sessions stored under ~/.codex/state. - Official native API transport: The Flash profile connects directly to DeepSeek's native Responses API with a Keychain-backed credential; Pro/Beta profiles use a small built-in local Python proxy that bridges function tools and DSML tool calls. - Interactive Terminal monitor: Each run opens a Terminal window showing the verified model, effort, profile, live output, and a follow-up prompt after completion. - Use Case: Ask Codex to run a slow test suite in a DeepSeek sidecar while it reviews the code in parallel, then collect the sidecar's findings and synthesize a final fix plan at a fraction of the token cost. ## Quick Start Ask Codex to install this repository as a skill, store your DeepSeek API key in the macOS Keychain item codex-deepseek-official, run scripts/codex-deepseek-subagent --configure, and then launch a DeepSeek sidecar for a bounded task in the current repository.

Frequently Asked Questions about deepseek-codex-subagent

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I delegate a coding task to a DeepSeek subagent from Codex?▼

Run scripts/codex-deepseek-subagent with --cd pointing at the project and a bounded task prompt. The wrapper launches a DeepSeek worker session, opens a Terminal monitor, and supports --status, --wait, and --resume for follow-ups.

How do I configure the DeepSeek API key for Codex sidecar profiles?▼

Store the key in the macOS Keychain item codex-deepseek-official, or set DEEPSEEK_API_KEY or DEEPSEEK_API_KEY_FILE on other platforms. Then run scripts/codex-deepseek-subagent --configure, which writes sidecar profiles without touching your main Codex login state.

What is the difference between the Flash and Pro sidecar profiles?▼

The default ds-sidecar-flash profile uses deepseek-v4-flash over the official native Responses API with max thinking effort for cheap worker tasks. The ds-sidecar-local profile uses deepseek-v4-pro through the local proxy for correctness-sensitive or long-horizon work.

Does this skill work with Claude Code or OpenCode?▼

Yes, separate adapters are included. The .claude-plugin directory and skills/claude-deepseek-sidecar provide a Claude Code adapter, while .opencode contains an OpenCode adapter that talks to an OpenAI-compatible endpoint.

Why should I not run DeepSeek's official one-click setup script?▼

The official script rewrites the global model, model_provider, and auth keys in ~/.codex/config.toml and hides the ChatGPT login session group. This skill implements the same official provider settings inside isolated sidecar profile files instead.

What are the limitations of the local DeepSeek responses proxy?▼

The proxy supports text, function tools, streaming, and cache usage, but intentionally drops image data URIs, so image findings must be described in text. Built-in hosted Responses tools are unsupported; only function tools are bridged.

Related Skills

Explore More Software Engineering Skills→
benjaminshafiibenjaminshafii
10

openwork-notary

Store Apple notarization credentials in macOS Keychain for OpenWork notarization workflows.

Community
Advanced
kenn-iokenn-io
1.6k

roborev-review

Request background code reviews for commits and present structured verdicts.

Official
Intermediate
kitaekattkitaekatt
3

cache-report

Parse cache_report.py output to display prompt-cache hit rates and token consumption.

Community
Intermediate
ruvnetruvnet
70.1k

deepseek-reason

Runs reasoning completions against DeepSeek's deepseek-reasoner model and returns chain-of-thought separately.

Community
Intermediate