[GH-ISSUE #128] [Proposal] DeepCoder: A complete multi-agent orchestration design for DeepSeek-V4 (Pro/Flash) – design ready, seeking feedback or co-implementation #48

New issue

Open

opened 2026-10-26 09:44:17 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:44:17 +00:00

Owner

Originally created by @Coolmajintao on GitHub (May 21, 2026).
Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/128

Hi DeepSeek team,

I have designed a complete multi-agent framework specifically for DeepSeek-V4 (Pro/Flash). The design is implementation-agnostic and focuses on maximizing the Pro/Flash cost-performance advantage.

Project: DeepCoder

GitHub repo (design document only): https://github.com/coolmajintao/DeepCode

What's included in the design (20k+ words):

Role matrix: Pro as chief engineer (decision-making), Flash as front desk (intent classification), Flash workers (parallel code generation), Pro arbitrator (deadlock resolution)
Full pipeline: intent classification → requirement clarification → task decomposition (WBS) → parallel execution with file checkout locks → final judgment with predefined error codes (A-1~D-10, saving ~90% token in final verdict)
9 production-ready system prompts in English, optimized for DeepSeek-V4 (with injection timing and caching strategy)
Extensible ecosystem: installable Agents / MCP Servers / Skills
Technology stack: Python + Typer + Rich + Docker sandbox (recommended)
CLI design and terminal UI specification

Current status:

✅ Complete design document (MIT licensed)
🚧 Code implementation is NOT yet available – I am unable to finish it alone due to capacity. That's why I'm open-sourcing the design first, hoping that DeepSeek team or the community could take it forward.

What I'm asking from you:

Take a look at the design document (it's in docs/DESIGN.md in the repo)
If you find any part valuable (e.g., the Pro/Flash scheduling strategy, error code system, prompt injection timing), feel free to adopt, reference, or improve it.
If the design aligns with your vision, I would be honored if DeepSeek team would consider co-hosting or recommending this as a blueprint for multi-agent applications on your models.

I believe this design can showcase the best of DeepSeek Pro (reasoning) and Flash (speed & cost) working together in a real-world coding assistant.

Thank you for your amazing work. I look forward to any feedback – even a simple "we've seen it" would mean a lot.

Best regards,
JinAni (coolmajintao)

Originally created by @Coolmajintao on GitHub (May 21, 2026). Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/128 Hi DeepSeek team, I have designed a complete multi-agent framework specifically for DeepSeek-V4 (Pro/Flash). The design is **implementation-agnostic** and focuses on maximizing the Pro/Flash cost-performance advantage. ### Project: DeepCoder **GitHub repo (design document only)**: https://github.com/coolmajintao/DeepCode ### What's included in the design (20k+ words): - Role matrix: Pro as chief engineer (decision-making), Flash as front desk (intent classification), Flash workers (parallel code generation), Pro arbitrator (deadlock resolution) - Full pipeline: intent classification → requirement clarification → task decomposition (WBS) → parallel execution with file checkout locks → final judgment with predefined error codes (A-1~D-10, saving ~90% token in final verdict) - 9 production-ready system prompts in English, optimized for DeepSeek-V4 (with injection timing and caching strategy) - Extensible ecosystem: installable Agents / MCP Servers / Skills - Technology stack: Python + Typer + Rich + Docker sandbox (recommended) - CLI design and terminal UI specification ### Current status: ✅ Complete design document (MIT licensed) 🚧 **Code implementation is NOT yet available** – I am unable to finish it alone due to capacity. That's why I'm open-sourcing the design first, hoping that DeepSeek team or the community could take it forward. ### What I'm asking from you: - **Take a look at the design document** (it's in `docs/DESIGN.md` in the repo) - If you find any part valuable (e.g., the Pro/Flash scheduling strategy, error code system, prompt injection timing), feel free to adopt, reference, or improve it. - If the design aligns with your vision, I would be honored if DeepSeek team would consider co-hosting or recommending this as a blueprint for multi-agent applications on your models. I believe this design can showcase the best of DeepSeek Pro (reasoning) and Flash (speed & cost) working together in a real-world coding assistant. Thank you for your amazing work. I look forward to any feedback – even a simple "we've seen it" would mean a lot. Best regards, JinAni (coolmajintao)

bsakaguchi commented

2026-10-26 09:44:18 +00:00

Author

Owner

@yiqiaowang-arch commented on GitHub (Aug 16, 2026):

Hi @Coolmajintao, I have just implemented deepseek harness plugin called "orchestra" with a simpler but similar idea: https://github.com/yiqiaowang-arch/orchestra, where a coordinator can monitor and coordinate other chats/sessions with automatic handoff function whenever the context window is growing too large.

Most coding-harnesses can easily generate subagents, but they are often forced by harness architecture to share the same context (workspace). This plugin enables cross-workspace coordination, which might suit your proposal for keeping the workspace tidy for parallel task completion.

Could you please also check it out? Thanks in advance for feedback!

@yiqiaowang-arch commented on GitHub (Aug 16, 2026): Hi @Coolmajintao, I have just implemented deepseek harness plugin called "orchestra" with a simpler but similar idea: https://github.com/yiqiaowang-arch/orchestra, where a coordinator can monitor and coordinate other chats/sessions with automatic handoff function whenever the context window is growing too large. Most coding-harnesses can easily generate subagents, but they are often forced by harness architecture to share the same context (workspace). This plugin enables cross-workspace coordination, which might suit your proposal for keeping the workspace tidy for parallel task completion. Could you please also check it out? Thanks in advance for feedback!

bsakaguchi referenced this issue

2026-10-26 09:46:26 +00:00

[PR #48] [CLOSED] docs: add MateClaw integration guide #169