Karabiner converter: AI-assisted Tier 2+ conversion for complex rules

[malpern]

Summary

After the MVP Karabiner converter ships (Tier 1: deterministic conversion of simple remaps, tap-hold, chords, macros, app rules, launchers), explore using AI to improve conversion fidelity for complex rules that resist mechanical translation.

Tier 2+ Candidates for AI-Assisted Conversion

Complex variable state machines

Karabiner's set_variable/variable_if can model multi-state systems beyond simple on/off layers. AI could analyze the intent (e.g., "this is a Vim normal/insert/visual mode system") and produce optimal Kanata layer structures.

Ambiguous rule intent

Some Karabiner rules could map to multiple KeyPath patterns. AI could choose the best-fit pattern based on surrounding context (e.g., is this a navigation layer or a modifier layer?).

Post-conversion validation & repair

Use existing AnthropicConfigRepairService infrastructure to validate converted Kanata output and auto-fix issues.

Natural language descriptions

Generate helpful notes and description fields for converted rules so users understand what each imported rule does.

Approximate conversion of Tier 3 features

Features like to_delayed_action, complex expression_if conditions, and multi-variable logic could be approximated with AI-suggested Kanata patterns (fork, switch, macro with delays).

Context

• Feature matrix: docs/karabiner-converter-feature-matrix.md
• MVP converter: Karabiner JSON → Kanata config converter / visual diff tool #202
• KeyPath already has AnthropicConfigRepairService for AI-powered config repair

Notes

This is explicitly post-MVP. The deterministic Tier 1 converter should ship first and work well for ~75-80% of real configs. AI assistance is for closing the gap on the remaining ~15-20%.