CLI — run and chat
The okami CLI is organized by domain. With no argument, okami shows the grouped overview; okami <command> --help lists the flags. All interactive config is arrow-ke…
Conventions and global flags
-p, --providerand-m, --modelappear onrun/task/chat— they override the okami.yaml default.-a, --agentruns AS an agent (agents/<id>); without it, uses the default agent from setup.-w, --workspacesets the working directory (default varies per command).-y, --yoloauto-approves sensitive actions for that session only.okami --version/-Vprints the version and exits.
Exit codes
| code | means |
|---|---|
| 0 | success (task: COMPLETE state) |
| 1 | execution error / invalid config |
| 2 | task not COMPLETE (BLOCKED/NEEDS_INPUT/FAILED) or gate/scan failed |
okami setup
Configuration wizard (providers, login, memory, identity, channel). Forks Quick (detects what you already have) vs Full on the first prompt. okami setup <section> jumps straight in.
params [section] arg Jump to a section: provider · memory · agent · channel · voice · approvals · learning · persona. --memory fts5|holographic|holographic+honcho Non-interactive shortcut: writes only the memory block. --honcho-url / --honcho-key string Honcho endpoint/credential (with --memory holographic+honcho). --embedder-url / --embedder-model string Embedding server and model (OpenAI-compat).
okami run
A RAW round-trip to the provider — no session, no harness. For testing a prompt or piping.
params <prompt> arg The prompt for the model (required). -p, --provider string Provider to use (default: from okami.yaml). -m, --model string Override the provider's model. -s, --system string Optional system prompt. --no-stream flag Disable streaming (return the full response at once).
okami task
Runs the harness until COMPLETE/BLOCKED/NEEDS_INPUT/FAILED. The verifiable entry point — ideal for CI, cron and scripts. With no goal, it asks.
params [goal] arg The task goal. Empty → asks (or errors when piped). -e, --exit crit (repeatable) Verified exit criterion: file_exists, file_contains, shell_ok, ui_gate. --max-steps int Harness step ceiling. 24 --escalate provider Strong provider for the cascade if it stalls. --mode manual|smart|off Approval mode for this run only. -y, --yes / --yolo flag Auto-approve everything this session. -a / -p / -m / -w — Agent · provider · model · workspace (default workspaces/default).
okami task "add a test for parse_date" \
-e "shell_ok:pytest -q tests/test_dates.py"okami chat
Chat in the terminal. By default opens the full-screen TUI (fixed regions, mouse, scroll, button approval); the session persists and resumes on reopen. With a message as argument, it does a one-shot (script/pipe).
params [message] arg Single message (one-shot). Empty → interactive REPL/TUI. --tui / --no-tui flag Full-screen TUI (default) or line REPL. tui --new flag Start fresh (archives the terminal's previous conversation). -y, --yolo flag Auto-approve sensitive actions this session. -a / -p / -m / -w — Agent · provider · model · workspace.
okami room
Multi-agent brainstorm: a cheap moderator picks who speaks (or no one), no stampede. Groups come from groups: in okami.yaml.
params <message> arg Message to the group (use @id to mention and force a reply). -g, --group int Group index in okami.yaml (groups). 0 --moderator provider Cheap provider for the moderator to pick the speaker.