AirCoding/AirPlan/docs/architecture/branch-deepcode-cli

DeepCode CLI Branch

Date: 2026-05-28 Status: AirCoding mainline branch, pre-design research phase

DeepCode CLI is a mainline AirCoding branch (not a downstream simplification like VibeBox). It explores integrating DeepSeek's model architecture innovations and research capabilities into the AirCoding agent runtime.

Reference Documents

• DeepSeek Research Survey — Complete inventory of DeepSeek papers, models, architectures, and technical innovations (36 repos, ~395K stars)

Key DeepSeek Innovations Relevant to DeepCode CLI

1. Multi-head Latent Attention (MLA) — 93% KV cache reduction for long code contexts
2. DeepSeekMoE — Fine-grained experts + shared experts, 37B activated from 671B total
3. GRPO — RL without critic model, applicable to code/debug agent training
4. DeepSeek Sparse Attention — Efficient long-context processing
5. DeepSeek-Coder/V2 — Fill-in-Middle, SWE-bench, Codeforces competitive
6. DeepSeek-R1 — Emergent chain-of-thought reasoning via pure RL
7. DeepSeek-Prover-V2 — Formal Lean 4 theorem proving
8. DeepSeek-Math-V2 — Self-verifiable reasoning, IMO 2025 gold
9. Engram — Conditional memory sparsity as new axis beyond MoE
10. DeepSeek-OCR/OCR-2 — Visual understanding for GUI evidence
11. 3FS/FlashMLA/DeepGEMM/DeepEP/DualPipe — Full inference infrastructure stack

Open Design Questions

1. How does DeepCode CLI relate to AirCoding mainline — does it add DeepSeek as a provider, or fork architecture?
2. Which DeepSeek innovations should DeepCode CLI adopt at the runtime level vs. treat as provider capabilities?
3. Does DeepCode CLI target DeepSeek models as primary, or remain provider-agnostic with DeepSeek optimizations?
4. Should MLA/MoE-aware context assembly be part of the runtime, or handled by provider adapter?
5. How should formal proving (Prover-V2) integrate with the existing review/debug workflow?
6. Should self-verifiable reasoning (Math-V2 style) influence the verification architecture?
7. Does DeepCode CLI need its own GRPO-trained code agent, or reuse general DeepSeek models?
8. How does the Engram conditional memory concept map to AirCoding's memory/skills system?
9. Should infrastructure tools (3FS, smallpond) be optional capabilities for large-codebase workflows?