v1.2.0 beta

• Browser work is up to 10× cheaper on long sessions: the repeated part of each request comes from the provider's cache. The agent sees and does exactly what it did before.
• A light theme.
• Usage in the settings: tokens spent by day, week, month and all time, per provider and model, with your ChatGPT plan limits and DeepSeek balance.
• Clear approvals: in Ask mode the agent says in plain words what it wants to do, and the card shows what the step does and where.
• Rename chats right in the list. Effort shows its level name.
• API keys are encrypted by the operating system instead of kept in plain text.
• Cmd+A, C, V and Z work on Mac. On Linux, paths respect letter case, and approvals understand bash and zsh.
• A long paste turns into a card, and one click turns it back into text. Esc stops the agent from anywhere, instantly.

Thanks to @kodachromez, @Bruno8R, @anupamme and @mike13kohler for the pull requests and reports that made this release more secure.