v1.49.0

Termpolis now asks first

Your agents' configs. Termpolis connects to Claude Code, Codex CLI and Gemini CLI only with your consent.

• The first step of the first-run tour lists exactly what it writes, with a Connect agents toggle. Skip tour keeps what the step shows.
• Updating from an earlier version? A one-time dialog asks you. Your current setup keeps working until you answer.
• Settings ▸ Agent Integration has an on/off toggle and a Disconnect button that removes everything Termpolis wrote. Uninstalling on Windows does the same; on any platform, Termpolis --disconnect-agents does it from the command line.
• The mcp__termpolis__* wildcard permission is gone, from new and existing configs. An explicit list of read-only and memory tools replaces it; every other Termpolis tool asks first.
• Termpolis never trusts your home folder or a drive root on an agent's behalf, and leaves the Codex approval settings you chose alone.
• It no longer writes AGENTS.md into your repos, and the duplicate Claude Code plugin is removed.
• Folders an earlier version trusted stay trusted, because nothing recorded which ones Termpolis trusted. The exceptions are your home folder, the folders above it and drive roots: v1.49 withdraws trust from those.

Approval prompts are yours. Termpolis no longer answers permission prompts, MCP-server approvals, [Y/n] questions or numbered choices. It confirms an agent's folder-trust prompt only when your agents are connected and neither the folder nor its git root is your home folder or a drive root. On first-run screens it still presses Enter and takes the default: Claude Code's intro splash, its theme and login-method pickers, and Gemini CLI's terms and sign-in screens.

Telemetry is off unless you turn it on. Settings ▸ General ▸ Privacy has two separate switches: crash reports, and a daily anonymous ping that carries only the app version. Skipping the tour leaves both off. A yes given to an earlier version doesn't carry over: Termpolis sends nothing until you answer once more, and both switches start off. Crash minidumps are no longer uploaded.

The token-saving proxy has an off switch: Settings ▸ Token Savings ▸ Claude Code compression proxy. It stays on by default, steps aside by itself when you already route Anthropic traffic (ANTHROPIC_BASE_URL, an HTTP(S) proxy, Bedrock, Vertex or Foundry), and no longer rewrites tool inputs.

Second Opinion is read-only. It no longer runs the reviewing agent with --dangerously-skip-permissions.

Full Changelog: v1.48.0...v1.49.0