Skip to content

README.en.md

Latest commit

 

History

History
117 lines (80 loc) · 5.48 KB

README.en.md

File metadata and controls

117 lines (80 loc) · 5.48 KB
Raw

English | 中文

Give Codex a persistent DeepSeek subagent for parallel coding tasks.

🔥 Now supports Codex, Claude Code, and OpenCode

Delegate code reviews, debugging, research, and other bounded tasks without interrupting your main Codex workflow. The expensive GPT stays the brain — planning, judging, synthesizing — while cheaper DeepSeek workers handle long tests, log analysis, broad exploration, and other token-heavy execution. Each task runs in its own Terminal with live output and keeps its full session context.

When the task finishes, the Terminal becomes an interactive deepseek> prompt for follow-ups. You can close it, check status by task ID, and resume the same session later.

No third-party proxy required. Bring your own DeepSeek key; the included local proxy handles Codex Responses traffic. 🚀

codex-deepseek-sidecar

🚀 You only need these prompts

This skill is meant to be read and executed by Codex, not memorized by humans. On a machine with Codex CLI, Python 3, and a DeepSeek key, Codex should have everything running in about five minutes.

Install — give this repo to Codex

Install and configure https://github.com/Zedong-Liu/codex-deepseek-sidecar.
I have a DeepSeek API key — ask me for it if it's not configured on this machine yet.
If a local proxy or profile needs to be set up, handle that too.
If using the included local proxy, start it and keep it running for sidecar tasks.
Then launch a DeepSeek sidecar for this repo to handle long or log-heavy tasks.

Delegate a task

Use a DeepSeek sidecar to run the slow tests while you review the code.

Use a DeepSeek sidecar to analyze this CI log and find the real failure.

Use your own judgment on when to split work across DeepSeek sidecars.
Auto-dispatch for tests, log analysis, and broad exploration, then synthesize results.

✨ Why use it

  • 💸 Dramatically lower worker token cost — shift repeated file reads, log inspection, test runs, and broad exploration from expensive GPT tokens to DeepSeek worker tokens. Many workflows target an 80–90% lower token cost.
  • No third-party proxy needed — a small built-in Python proxy bridges Codex Responses API to DeepSeek Chat Completions. When using the built-in profile, keep the local proxy process running.
  • GPT stays the brain — the expensive model handles planning, judgment, and synthesis; DeepSeek handles bounded worker tasks.
  • Still Codex harness — sidecars retain Codex file access, command execution, sessions, and evidence reporting.

💸 Cost estimate

Prices per 1M tokens, based on OpenAI GPT-5.5 model page and DeepSeek pricing docs. DeepSeek Pro input price uses cache-miss for a conservative estimate.

Model Best role Input Output
GPT-5.5 Brain: planning, judgment, synthesis $5.00 $30.00
DeepSeek V4 Pro Strong worker: review, debug, implement $0.435 $0.87
DeepSeek V4 Flash Fast worker: logs, exploration, cheap parallel passes $0.14 $0.28

At 1M input + 200K output:

Routing Approx cost Token cost reduction
All GPT-5.5 $11.00 baseline
DeepSeek V4 Pro worker tokens $0.61 ~94%
DeepSeek V4 Flash worker tokens $0.20 ~98%

In real use, GPT still spends tokens on coordination and review. That is the point: spend expensive tokens on judgment, move repetitive worker-token budgets to DeepSeek. For many agent workflows, 80–90% token cost save is a realistic target.

🧠 What Codex does behind the scenes

When you ask Codex to use this skill, it can:

  • Install the repo as a Codex skill.
  • Start the built-in lightweight proxy if no local DeepSeek provider is available.
  • Configure a Codex profile once so future tasks skip cold-start setup.
  • Launch DeepSeek sidecars for bounded tasks.
  • Track task IDs and sessions so follow-ups resume to the correct worker.
  • Collect sidecar conclusions and synthesize them into the main response.

These operational details belong in SKILL.md, not in front of human readers.

🔌 Built-in proxy

The bundled deepseek-responses-proxy is intentionally minimal: Python stdlib only, localhost by default, designed for Codex's large request bodies. It bridges function tools and ignores Responses built-in tools that DeepSeek Chat does not support, returning a clear error if one is explicitly required. For reusable startup, it can read the key from a private file with --api-key-file.

If you already use VibeAround or another compatible provider, Codex can keep using that instead.

🧩 Framework adapters

Codex remains the main, stable entrypoint. Other framework adapters live in their own install surfaces so they do not change the Codex skill behavior:

  • .opencode/ contains the OpenCode adapter and OpenCode skill.
  • .claude-plugin/ plus skills/claude-deepseek-sidecar/ contains the Claude Code adapter.

📦 Repo layout

.
├── .claude-plugin/
├── .opencode/
├── SKILL.md
├── agents/openai.yaml
├── skills/claude-deepseek-sidecar/
├── scripts/codex-deepseek-sidecar
├── scripts/codex-deepseek-subagent
├── scripts/deepseek-responses-proxy
└── scripts/terminal-chat

License

Apache 2.0 — see LICENSE.