Launch presets

A launch preset is the set of arguments Coldtea passes to an agent CLI when you start a pane. You set them once per agent under Settings, then Agents, in the "Launch Presets" group. Coldtea flattens your choices into the launch arguments and remembers them for every new pane of that agent.

Each agent has its own row. A checkmark means Coldtea found the CLI; a dashed circle means it did not (see Connect agents). Drag rows to reorder them; the order sets the quick-launch shortcut each agent gets.

Autonomy

The "Autonomy" preset controls how much the agent does without asking. The options differ per agent because each CLI has its own permission model. Coldtea maps each preset to that CLI's own flags.

Claude:

• Supervised: asks before edits and commands.
• Plan: read-only, proposes a plan without making changes.
• Auto-edit: applies edits without asking, still prompts for other actions.
• Locked-down: never prompts, for trusted scripted runs.
• Full autonomy: skips all permission checks. Marked as dangerous.

Codex:

• Supervised: workspace-write sandbox, asks on request.
• Plan / read-only: read-only sandbox, only trusted commands run unprompted.
• Locked-down: read-only, never asks, for trusted scripted runs.
• Full access: skips approvals. Marked as dangerous.

Gemini:

• Supervised: prompts for approval.
• Plan: read-only mode.
• Auto-edit: auto-approves edit tools.
• Full autonomy: auto-approves all tools. Marked as dangerous.

OpenCode:

• Supervised: runs OpenCode's native interface, where OpenCode owns setup, model switching, and permission prompts.
• Plan: asks before edits, using OpenCode's built-in plan agent.

OpenCode has no full-autonomy preset. Its unattended-run flag is not available on every OpenCode version, so Coldtea does not ship it as a preset; use OpenCode's own permission config for that.

The preset's description shows under the row so you can see what it does before you launch.

Model and reasoning effort

The "Model" field sets which model the agent launches with. Leave it empty to use the agent's own default. OpenCode takes a provider/model value here.

The "Reasoning effort" field appears for agents that support it (Claude and Codex). Leave it on "Default" to inherit the CLI's setting. Gemini and OpenCode have no reasoning-effort control.

Advanced arguments

The advanced field holds raw launch arguments Coldtea does not model as a control. Anything you put here is preserved: when Coldtea reads your saved arguments back into the controls, anything it does not recognize stays in the advanced field rather than being dropped. Each row links to that CLI's argument reference.

Coldtea keeps the last-wins behavior of the CLI, so a duplicate flag you add by hand stays after the preset's flags and still takes effect.

Agent behavior and custom prompts

Two more groups on the same screen shape how agents work, separate from launch arguments.

"Agent behavior" holds the "Auto pane labeling" switch. With it on, agents can name their own panes; you can still rename a pane by hand at any time.

"Custom prompts" adds standing instructions for specific workflows. The "Tasks" prompt is a message Coldtea gives agents when they work on a Coldtea task, so you can set house rules (for example, summarize the task first and log progress at each milestone) without repeating them. Edit the text and select "Save prompt".

Where to go next

• Connect agents if an agent shows as not installed here.
• Cloud agent credentials for the cloud-run equivalent, launch defaults.
• Agent matrix for what each agent supports.