Kimi Code

Kimi Code runs whichever model you configured with /login or /provider, so Saggar offers one launch row and leaves the model to Kimi. Saggar doesn't install Kimi hooks yet, so it follows Kimi from its process and its screen.

Command	kimi
Install	curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash
Hook configuration	None yet
Skill location	~/.kimi-code/skills/saggar-cli/SKILL.md
Resume	kimi --continue

Launching

The launch row runs kimi with the model Kimi is configured to use. Kimi takes no reasoning-effort flag, and preset context doesn't reach it.

Starting with a task

Kimi can't open its terminal interface with a message: kimi -p runs one turn and exits. So Saggar launches kimi bare, holds the task, and types it into Kimi's input box as soon as the empty box appears. That works from every place that hands an agent a task, including the new-session sheet, @kimi fix the failing test, presets, follow-ons, and scheduled runs.

While it waits, the task shows as a queued prompt on the terminal, where you can send it now or cancel it. It doesn't wait for other sessions in the project, and it isn't typed while Kimi's folder-trust question is on screen.

Folder trust

Kimi asks whether to trust a folder the first time it runs there. Saggar answers yes for a folder you added to the project menu, or a worktree of one, when the dialog names that folder. Anything else is left for you.

Prompts and automatic approval

Without hooks, a waiting prompt is noticed from the screen and turns the session Needs you. Saggar has no dialog vocabulary for Kimi, so automatic approval never types into a Kimi session. Launching Kimi with --yolo, --auto, or -y turns its approvals off, so Saggar holds that command for you like any other unsupervised agent launch.

Resume

Restore runs kimi --continue, which reopens the most recent Kimi conversation in that folder. With two Kimi sessions in one folder, both reopen the latest.