82f3140847
Complete architecture document set with multi-model review remediation: - Frozen interface contracts, runtime semantics, DB schemas - Event/tool/error/provider registries - Scheduler and main agent state machines - C4 module/code views, solution architecture, baseline V1 - Multi-model review reports and joint assessment - Phase-gate remediation complete (P0/P1/P2/UX resolved) - Implementation plan with T-000A through T-045 - Reference folders kept as placeholders only
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
- Multi-head Latent Attention (MLA) — 93% KV cache reduction for long code contexts
- DeepSeekMoE — Fine-grained experts + shared experts, 37B activated from 671B total
- GRPO — RL without critic model, applicable to code/debug agent training
- DeepSeek Sparse Attention — Efficient long-context processing
- DeepSeek-Coder/V2 — Fill-in-Middle, SWE-bench, Codeforces competitive
- DeepSeek-R1 — Emergent chain-of-thought reasoning via pure RL
- DeepSeek-Prover-V2 — Formal Lean 4 theorem proving
- DeepSeek-Math-V2 — Self-verifiable reasoning, IMO 2025 gold
- Engram — Conditional memory sparsity as new axis beyond MoE
- DeepSeek-OCR/OCR-2 — Visual understanding for GUI evidence
- 3FS/FlashMLA/DeepGEMM/DeepEP/DualPipe — Full inference infrastructure stack
Open Design Questions
- How does DeepCode CLI relate to AirCoding mainline — does it add DeepSeek as a provider, or fork architecture?
- Which DeepSeek innovations should DeepCode CLI adopt at the runtime level vs. treat as provider capabilities?
- Does DeepCode CLI target DeepSeek models as primary, or remain provider-agnostic with DeepSeek optimizations?
- Should MLA/MoE-aware context assembly be part of the runtime, or handled by provider adapter?
- How should formal proving (Prover-V2) integrate with the existing review/debug workflow?
- Should self-verifiable reasoning (Math-V2 style) influence the verification architecture?
- Does DeepCode CLI need its own GRPO-trained code agent, or reuse general DeepSeek models?
- How does the Engram conditional memory concept map to AirCoding's memory/skills system?
- Should infrastructure tools (3FS, smallpond) be optional capabilities for large-codebase workflows?