Launching an agent

On a pane with nothing on its workbench, press ⌁ LAUNCH AGENT, or just ↵. A panel opens with your projects listed, newest first, and the one you're in already picked.

Type to filter, then choose Claude Code or Codex, a model, an effort level, and how far the agent may go without asking: this repo, this machine or anywhere. The bottom of the panel shows the whole command it will run. ↵ runs it.

If the pane you started from is sitting at a shell prompt, the agent starts right there. If something is already running in it, you get a new tab instead, next to the one you're in. The last line of the panel tells you which before you press ↵.

Technical has the keys, the exact commands and where the defaults live.

The panel

Opening it#

• The ⌁ LAUNCH AGENT button under NOTHING ON THE WORKBENCH.
• ↵ on a workbench with no agent and no card open.
• When a pane's agent has ended, the verb at the end of its bench strip changes from END SESSION to LAUNCH AGENT.

Keys#

Projects#

Directories one level deep under a fixed set of folders, your home directory and ~/Work among them, newest-modified first, with hidden directories skipped. The set can't be changed yet, so projects two levels down, such as ~/code/app, don't appear (issue 802). ◆ marks a git repository. Matches rank by exact name, then prefix, then anywhere in the name, then anywhere in the path; a repository beats a plain directory at the same rank. Up to 40 rows show; the filter reaches the rest. With nothing matching: nothing matches — ⌫ to widen.

Choices#

Switching harness moves the effort to the nearest level the other one has, so max becomes xhigh on Codex. The panel always opens on anywhere.

The command#

claude --model <model> --effort <effort> [reach] --append-system-prompt "$(cat '<briefing>')"
codex --model <model> [reach] -c model_reasoning_effort=<effort>

The briefing is the workbench's instructions to the agent, written owner-only to briefing.txt in the session's state folder. Codex takes no briefing flag and reads AGENTS.md instead.

Where it lands#

If the pane you opened the panel from is known to be at an idle prompt, Terminal Delight types into it: Ctrl+U to clear anything half-typed, cd to the project if the pane is somewhere else, then the command. If the pane is busy, or hasn't been read yet, a new tab opens in the same branch of the left bar. The hint ends in ↵ starts in this pane or ↵ opens a new tab. The pane keeps whichever face it had.

Each launch is recorded in launches.jsonl in the session's state folder: harness, model, effort, reach, directory, command and where it landed.

Defaults#

Open Σ usage, unfold LAUNCH AGENT CONFIG and press a chip. Each row reads chosen or unset; pressing a lit chip clears it. That writes ~/.config/terminal-delight/launch.toml at once:

harness = "codex"
model = "gpt-5"
effort = "high"

Every key is optional. A missing file, a word neither harness knows, or a model from the other harness all read as unset, and unset means Claude, its first model and its default effort. Reach is never stored.

After it starts#

The bench strip has model and effort dials. They type the harness's own /model and /effort commands and keep whatever you had in the composer. The launcher only starts agents; bringing one back is the job of session restore and the graveyard.