Changelog

0.2.0

Three-tier retrieval: override → tldr-osx → tldr-common. Curated macOS examples now always appear first, reducing noise from 21k cross-platform entries
Command validation: retries when the model hallucinates a command that doesn't exist locally or in the tldr bank
Smarter command-not-found handler with three categories:
- Typo detection via Damerau-Levenshtein (grpe → did you mean: grep)
- Install suggestions via brew which-formula (ncdu → not installed: brew install ncdu)
- Linux→macOS equivalents via LLM (ip a → macOS equivalent: ifconfig)
Notfound benchmark: 30 prompts, 97% category accuracy
Suggest benchmark: ~83% (up from ~78% in v0.1.2)

Improved system prompt: model now prefers retrieved examples over pretraining guesses
Fixed FTS5 stop words: command names (find, show, list, display) no longer dropped from search queries
Added 28 targeted overrides for common patterns missing from tldr (find -size, find -mtime, find -user, curl -I, grep --include, and more)
Expanded benchmark alternates for more accurate automated scoring
Benchmark accuracy: ~78% (up from ~70% in v0.1.1)
Added experimental --guided flag for Apple FoundationModels constrained decoding (undocumented)

Fix: don't pass empty GenerationOptions when no temperature is set (was changing model behavior)
Fix: scoring bug where stale reviews deflated benchmark numbers
Expanded macOS overrides (83 → 107 entries) targeting consistently failing commands
Updated benchmark numbers: ~68% base, ~70% with overrides, ~72% with overrides + sc
Self-consistency finding: same average accuracy, but kills run-to-run variance
Simplified benchmark scorer (removed stateful reviews, exact match + fresh review per run)

Initial release.