Project Settings
Project Settings configure a single repository. With a project selected, open them via:
- The Settings icon in the chat header, or
- ⋯ More options → Settings (Project section), or
- Settings → Project → Settings in the settings overlay
Click Save when done — this closes settings and returns you to the chat view.
Settings Reference
Section titled “Settings Reference”The display name shown on the projects home grid and project landing page. Changing the name does not affect the repository or any branch directories.
Setup Script
Section titled “Setup Script”A shell command that runs automatically in a new branch’s directory immediately after the branch is created (i.e. after the clone completes).
Examples:
npm installbun installpip install -r requirements.txtcargo buildbundle installThe script runs in the branch’s working directory. If it fails, the branch is still created but a warning is shown. Leave this blank if no initial setup is needed.
Validation Script
Section titled “Validation Script”An optional shell command that runs automatically after each agent response completes. Use this to verify the agent’s work — for example, running linters, type checks, or test suites.
If the validation script fails, Tarsk can attempt to fix issues automatically up to the retry limit. The agent’s changes are not reverted automatically.
Examples:
npm run lintbun run typecheckpytestRun Command
Section titled “Run Command”The command used by the Run/Stop button in the chat header to start your development server. Click Suggest to auto-detect a dev script from package.json.
Examples:
bun devnpm run devpython manage.py runserverrails serverEach branch has its own independent instance of this process. See Dev Server for details.
The project landing page also shows Start App when a run command is configured and the server is not running.
Commit Method
Section titled “Commit Method”Controls how the Git Ops button handles finished work:
| Method | Behaviour |
|---|---|
| Pull Request | Feature branches and pull request workflow via the gh CLI |
| Direct | Commits are pushed directly to the branch’s git branch |
The gh CLI must be installed and authenticated for pull request mode. See cli.github.com.
Confirm Commands
Section titled “Confirm Commands”When on, the agent asks for approval before running bash commands, skill scripts, and browser shell calls. When off (default), commands run without per-command prompts.
When confirm commands is enabled, you can define allowed command patterns — commands matching these patterns (e.g. npm test, git status) run automatically without prompting.
You can also toggle Confirm commands temporarily from the chat mode selector menu. See Permissions.
Built-in Browser
Section titled “Built-in Browser”| Setting | Behaviour |
|---|---|
| On | Shows the Browser panel in the chat dock; detected dev URLs open inside Tarsk (default) |
| Off | Hides the Browser panel; detected dev URLs open in your system browser |
See Browser Preview.
Environment Variables
Section titled “Environment Variables”Under Project → Environment Variables in settings:
- Add key/value pairs injected into the agent’s shell environment for all branches
- Values are stored encrypted locally
- Variable names are normalized to uppercase
- Existing values are masked; leave the value field blank to keep the current value
Advanced Settings
Section titled “Advanced Settings”System Prompt
Section titled “System Prompt”Under Advanced → System Prompt, toggle between the generated system prompt (built from AGENTS.md, rules, skills) and a custom prompt you write yourself. When custom is enabled, the agent uses your text instead of the auto-generated one.
Model Aliases
Section titled “Model Aliases”Under Advanced → Model Aliases, map short names to enabled models:
| Alias | Typical use |
|---|---|
fast | Quick responses for light tasks |
local | On-device or local inference |
cheap | Lower-cost models for routine work |
smart | Most capable models for hard tasks |
Enable the target model in Providers first. Set Cheap to enable Orchestrate mode. Auto mode needs Fast or Cheap.
Under Advanced → Tools, enable or disable built-in agent tools per project. Disabled tools are hidden from the agent in every branch. Disabling MCP blocks every MCP server tool for the project.
Smart JS (off by default) switches the agent to exclusive JS orchestration: the model’s direct tools are Run JS, Ask, Write, and Subagent. Other tools stay available through callTool inside the Run JS Web Worker, with a compact tool index on Run JS and listTools / describeTool helpers when the mode is on. Prefer the direct Write tool for large new files so content is not embedded in Run JS code.
See Chat for the full tool list.
Deleting a Project
Section titled “Deleting a Project”On the projects home screen, hover a project icon and click the trash button. A confirmation dialog lists all branches that will be removed. Deletion removes the project record and all branch directories from disk. This cannot be undone.
Related
Section titled “Related”- Settings Overview — project vs general scope
- Permissions — command approval details
- AGENTS.md — branch-level instructions