v0.7.0

Settings inside the launcher, Ask AI with any model, and the list that
learns what you use.

Added

• Used recently, and Stats. The root list starts with the five things
  opened most recently, marked. Every open is one line in
  ~/.local/state/swoop/usage.jsonl; a Stats row shows what you open
  most, how often, when last, and a fortnight by day. A setting turns the
  group off; ctrl-k in Stats clears the log.
• Start Screensaver, a row beside Sleep and Lock Screen.
• ctrl-k on New conversation changes the model. The same list as
  Settings, Enter picks one and returns to the pane; the New row says
  which model will answer.
• Ask AI with any model. Presets under AI model in Settings: claude,
  codex, each ollama model, each model LM Studio has loaded, and
  openai:<model> for OpenAI or any server with that API (OpenRouter,
  Groq, vLLM) with the address under API URL and the key under API key.
  Two transports: a command with the conversation on stdin, or a small
  HTTP client in the OpenAI chat shape, streamed, with a model's thinking
  shown as "thinking" under the dots rather than in the answer. Web
  search, when on, reaches claude, codex and OpenAI; the line under the
  dots says when the model you picked cannot search.
• Settings, inside the launcher. Type settings, or cmd+, in the
  frame: one row per setting, Enter to change it. The hotkey, the preview
  width, the AI model (claude, or ollama with any model you have), web
  search on or off, the transcript renderer, and a door to the config
  folder. Every value is one line in ~/.config/swoop/config. The frame
  watches that file, so a new hotkey works within a second.
• A Homebrew tap.

Changed

• Web search is off by default. claude may search and fetch only when
  the Web search setting is on. Turn it on in Settings, or web = on in
  the config file. brew install beatzball/tap/swoop, and on a Mac
  brew services start swoop runs the frame at login. The formula is
  written by scripts/tap-formula from a release's checksums, and the
  release workflow pushes it to the tap when it has a token for it.

Fixed

• ollama answers are clean. ollama run wrapped words with cursor
  moves even through a pipe, and a thinking model's reasoning came out
  as the answer. The ollama line now runs with --nowordwrap,
  --hidethinking and --think=false, escape codes are stripped from
  what any command prints, and with web search on and an ollama model
  the line under the dots says the model cannot search.