Summary

Elevate aprender-shell from solid baseline to world-class AI shell completion.

Current State (v0.2.0)

• N-gram language model (trigram default)
• Sub-10ms latency, 2.7MB binary
• Local-only, privacy-preserving
• Basic augmentation and AutoML tuning

Gap Analysis

Roadmap to World Class

Phase 1: Context Awareness

• Current directory context
• Git repository state (branch, status)
• Environment variables
• Recent command success/failure
• Time-of-day patterns

Phase 2: Semantic Understanding

• Command embedding space (Code2Vec style)
• Intent classification (file ops, git, docker, etc.)
• Argument type inference
• Error prediction (will this command fail?)

Phase 3: Transformer Integration

• Small transformer encoder (distilled)
• Pretrain on shell corpus (bash/zsh scripts)
• Fine-tune on user history
• Quantized for <50ms latency

Phase 4: Natural Language

• NL → command translation
• "find large files" → find . -size +100M
• Explanation mode ("what does this do?")
• Safety warnings for destructive commands

Phase 5: Cross-Session Learning

• Persistent learning across sessions
• Workspace-specific models
• Team model sharing (opt-in)
• Federated learning (privacy-preserving)

Technical Requirements

• Latency: <50ms (transformer), <10ms (fallback n-gram)
• Binary size: <20MB (with embedded model)
• Memory: <100MB runtime
• Privacy: Local-first, opt-in cloud features
• Platforms: Linux, macOS, Windows

Dependencies from aprender

From more-learning-specs.md:

• §3: Contrastive learning (command embeddings)
• §31: Code-specific ML (AST tokenizer)
• §32: Embedding & retrieval (semantic search)
• §33: Transfer learning (cross-user knowledge)

Success Metrics

References

• Fig - Commercial leader (acquired by AWS)
• Warp - Modern terminal with AI
• GitHub Copilot CLI - LLM-powered
• zsh-autosuggestions - OSS baseline