What's New in Tarsk: June 2026

The past week in Tarsk shipped more than bug fixes. Settings got a full navigation pass. You can see exactly what fills your context window. Chat drafts survive when you jump into settings. Markdown files preview in the explorer. And your preferred editor opens with one click.

Settings you can actually find

Settings used to scatter options across modals and sidebar shortcuts. Now you get a left-nav layout with clear sections: General, Models, Providers, Subagents, Skills, History, and Account.

General is the default landing page. A font-size slider scales text across the whole app, which helps on a laptop screen or a 4K monitor. Project settings moved out of a cramped dialog and into the main settings view, so you edit repo config without losing your place.

Provider search filters the list when you run more than a handful of API keys. Provider-specific environment variables (beyond the main API key) now have dedicated fields in onboarding and settings, so Azure resource names and similar values do not live in shell exports alone.

Context window, broken down

Hitting the context limit mid-task is frustrating when you cannot tell what ate the tokens. The usage ring in the branch bar now opens a panel that splits context by category: system prompts, messages, tool output, and the rest.

You see a percentage, a token count against the model's window size, and a color bar for each category. The numbers update while the agent streams, so you can watch tool calls push usage up in real time. That visibility matters when you pick a 200K window model and still run out of room after a long grep session.

Chat input that remembers

You draft a long prompt, open settings to check a model, come back, and the text is gone. Tarsk now stores a draft per thread in app state. Navigate to settings, the explorer, or another tab and your unsent message waits where you left it.

Scroll position in the chat persists across those same hops. Streaming token usage shows while the model responds. If you queue a follow-up while a thread is still running, the input clears only after you send, not when you switch views.

Markdown preview in the explorer

Agent-generated README files and skill docs live as .md files in your repo. The explorer now opens them in rendered view first. A View / Edit toggle switches between the formatted page and the Monaco editor. You read the output without raw syntax, then flip to edit when you need a fix.

Open in your editor

The Open With button detects editors installed on your Mac: VS Code, Cursor, Windsurf, Antigravity, Xcode, Android Studio, Kiro, and others. Pick one, open the project or thread workspace, and Tarsk remembers your choice for the next session. No more selecting Cursor from a dropdown every time you bounce between the agent and your IDE.

Subagents from a marketplace

The new Subagents settings tab lists agents already in your project, plus a searchable catalog of popular agents you can install with one click. Install pulls the agent folder from GitHub into .agents/agents/ on the active thread. You can also create a custom agent or delete one you no longer need. Open any installed agent in the explorer to edit its AGENT.md directly.

Streaming across threads

Long agent runs no longer block the rest of the UI. Each thread tracks its own streaming state. Start a refactor in one thread, switch to another to ask a quick question, and both show accurate progress indicators. Tool calls in the chat stream got clearer execution badges so you can tell a running bash command from a finished file read.

Model picker improvements

Model names now normalize across providers, so the same endpoint does not show up three times under slightly different strings. Sorting and a contextWindow field on each entry make it easier to pick the right model before you start a long session. Open Settings → Models to browse coding and image models in separate tabs.

Key takeaways