Docs
Workflows
Chain models with > and +. Steer a turn with /frontier, /team, /repeat, /oss, /btw, and the director itself with /captain. Captain still routes bare prompts and owns handoff and memory.
Control words (steer without leaving the thread)
The brain parses these before routing without a model call. Tasks, workflow compilation and memory distillation still use models.
Pick who runs it
| Word | Effect |
|---|---|
/claude /codex /codex-cli /luna /cursor /grok /grok-max /gemini /deepseek /ds-flash /ds4-flash /kimi /glm /minimax /qwen /step /gpt-oss /free /openshell | Force a specific leg for this turn (captain legs lists yours) |
/frontier <task> | Frontier lane at max effort (Claude and Codex CLI rotate); /frontier /codex … runs that leg's frontier model |
/team <task> | Director plans an ensemble, one reviewed answer. /team /frontier … binds a member; /team /openshell … is sandbox-only |
/quality /speed /save | Lanes: /quality rotates the two best legs at high effort; /save rotates qualifying open-weight legs; the director picks inside /speed only (aliases /best, /fast, /cheap) |
/oss | Open-weight models only; composes with other modifiers in either order |
/deterministic | ADI-green serving tuple only among registered legs; pins the stack for the run. See ADI |
/noslop | Plain-writing rules for every worker (on by default; restores when disabled). OpenShell included |
/private | Curate the private-names list used by captain leakcheck (add, dismiss, remove) |
Steer the director
These persist across turns and print the result.
| Word | Effect |
|---|---|
/captain director | Who directs, why, and how much each judge has worked lately |
/captain <leg> / frontier / quality / auto / reset | Switch the director live: one leg, the best-ranked judge, the best of tier 2, the least-used capable one, or the default |
/captain more oss / /captain less cheap | Raise or lower one axis of the routing mix by 5 points; repeats compound |
/captain more oss 20% | 20% more than its current share (20% becomes 24%); the other axes fund it |
/captain oss=20% frontier=30% … | Set the balance directly; unnamed axes fill the rest |
/captain targets / /captain mix reset | Print the mix, or go back to the default |
Run things in the background
| Word | Effect |
|---|---|
/repeat N <task> | Rounds until N or until stopped; /repeat status, show, watch, stop, abort |
/parallel <task> | A second task beside this chat; show, status, stop |
/btw <note> | Steer the worker already running (mid-turn on claude/opencode legs) |
/interrupt [reason] | Stop the running worker; keep a handoff note or marked partial |
/wf <english> | Compile English into a workflow, preview, then /run wf_id. /wf save|run|list for named workflows |
Settings and memory
| Word | Effect |
|---|---|
/init | Regenerate worker permissions (web, edits, bash; .env stays denied) |
/context status / publish on|off / consume all|none|<path> | Cross-project context sharing, off by default |
/euclid status / /euclid distill [apply] | Euclid memory: which brains this folder reads, and register edits proposed from the run journal |
/captain or /help | The cheat sheet, in the TUI |
The mix axes are frontier, quality, cheap, fast, oss and deterministic. Unset, the targets are 20% each with deterministic at 0, and that default does not steer. Once saved, bare prompts lean toward legs that close the gap; a modifier on the turn still wins.
While a turn streams, the TUI queues what you type until the turn ends (Esc frees the input). Unknown /words are refused before send, with nearby suggestions. Press Esc, or use the shell: captain kill, captain stop, captain status. The sidebar shows harness, provider, model, and effort; step narration appears on the activity feed. See the CLI cheat sheet and OpenShell.
Workflow chaining syntax
One line. > (or then, ->) advances a stage; + (or and, &) runs legs in parallel inside a stage. A connector counts only when a leg prefix follows it, so prose is never misread.
/grok analyse @src/queue.ts > /cursor review it for missed cases > /codex red-team it + /claude red-team it
/grok draft the migration > /frontier harden it + /codex write the docs
/claude fix the flaky test gate: go test ./... > /codex review the fix
- A bare leg (
> /codex) requests a refinement of the previous output - One task per line also fans out in parallel
gate: <cmd>must exit 0 for that worker’s result to count as successful; one repair retry. Other successful workers can still feed the next stage- Preference words (
/cheap,/quality) are rejected inside a chain: name the legs - Max 4 stages, 4 legs per stage, 8 declared worker slots. Retries and the final director review add executions. The review combines the findings into one answer
Loop compiler
State intent ("until tests pass", "while coverage < 80%"). Use /repeat N to bound rounds and gate: <command> in a workflow to check a result. A natural-language target alone is not a machine-verified completion condition.