v0.5.0

Highlights

• Memory — ai --remember / --forget / --memory persist reusable notes (preferred tools, machine quirks) into ~/.ai-sh/memory.md, sent with every prompt (#5)
• Session history — opt-in per-terminal follow-ups on cloud backends (ai --new, --status), scoped to the shell and directory, with secret-shaped values redacted before writing to disk
• Sharper prose fallback — tightened shell system prompt so non-shell questions are answered as prose instead of forced into a command
• Installer safety net — install.sh pins to a known-good binary version instead of latest, and checks for --setup support before relying on it
• README — badges, quick-install callout, and a table of contents; project is now MIT licensed

Full changelog

v0.4.0...v0.5.0

v0.4.0 - cloud providers

Adds cloud LLM backends (Mistral, Groq, OpenRouter, Cerebras, or any OpenAI-compatible API) alongside the local llama.cpp model.

• ai --setup picks a backend and saves it to ~/.ai-sh/config.json
• install.sh asks local vs. cloud and hands off to ai --setup
• Config honors AI_SH_* env overrides and conventional provider keys (MISTRAL_API_KEY, ...)

Note: the v0.3.0 binary predates --setup, so the cloud path in the current installer failed against it. This release fixes that.

v0.3.0

• 'e' now refines the command: give extra context and re-runs inference in a loop
• Current working directory sent to model for context-aware suggestions
• POSIX sh compatibility: system prompt updated, commands run via /bin/sh

v0.2.0

• Replace DeepSeek with Qwen2.5-Coder 3B as option 3
• Interactive model selection now works during curl | bash installs
• Single-keypress UX: Enter runs, e edits in $EDITOR, any other key cancels
• Fixed llama-cli output capture (Setsid, stop tokens, spinner stripping)

v0.1.0

Initial release. Converts natural language to bash commands using a local LLM (llama.cpp).