Run modes

Most of omp’s behavior is shaped by one-shot slash commands. Three of those flip a persistent or single-turn mode instead of just dispatching a side action: /loop, /force, and /fast. They compose — /loop with a duration cap is the standard “let it cook overnight” recipe.

Loop mode — /loop [count|duration]

/loop toggles a state where the prompt you send next is re-submitted after every yield, automatically, until a limit hits or you stop it. Useful when the work is iterative and the model already knows the acceptance check: “fix the next failing test”, “process the next file in the queue”, “review the next PR in the list”.

/loop          # unlimited; runs until you cancel
/loop 10       # cap at 10 iterations
/loop 30m      # wall-clock cap
/loop 2h

Accepted units: s, m/min, h/hr, and their plurals. Mixing forms (/loop 10 5m) is rejected. The first prompt you send after toggling becomes the looped prompt; subsequent yields re-fire that same text. Esc cancels the current iteration without disabling loop mode; running /loop again disables it; the duration form ends on its own when the deadline passes.

Loop status appears in the mode segment of the status line alongside any active plan-mode indicator.

Force mode — /force <tool> [prompt]

/force pins the next turn to a specific tool. Scope is exactly one turn — after that turn returns, tool choice is unpinned. Use it when the model keeps reaching for the wrong tool: calling edit on a file that doesn’t exist yet, refusing write on a fresh scaffold, talking instead of dispatching a subagent.

/force write src/server/auth.ts: stub a JWT verifier
/force task                       # pins next message you send

If you provide a prompt after the tool name, it’s submitted in the same turn. With just /force <tool>, the pin attaches to whatever you type next. The form /force:<tool> is also accepted.

Fast mode — /fast

/fast toggles the priority service tier on outgoing requests: OpenAI models get service_tier: "priority", Anthropic models get the fast-mode beta (speed: "fast" on supported Opus models). Priority traffic is routed ahead of default-tier requests at a higher per-token cost; it’s the same dial you’d set in the API call yourself. omp counts each priority request against the premiumRequests budget — the same bucket GitHub Copilot premium requests land in — so it surfaces in /usage and the omp stats premium card.

/fast          # toggle
/fast on
/fast off
/fast status

The mode is session-persisted (it survives reloads via service_tier_change entries in the session log) and shows as a small badge in the status line. On providers without a priority tier the flag is dropped before the request leaves, so toggling it is a no-op there.

Which one for which goal

• Same prompt against a queue or until a check passes → /loop with a count or duration cap.

• Locking the next turn to a specific tool → /force <tool> (with optional prompt).ng run you want to walk away from** → /loop with a duration cap, in a terminal pane or multiplexer you can leave unattended.

• Model keeps picking the wrong tool → /force <tool> for one turn, then let it go.

• A turn is latency-critical and you’re on OpenAI or Anthropic → /fast on, accept the higher cost.

Force is per-turn and unpins itself. Fast is per-session until you toggle it off. None of them replace a clear prompt — they just remove the manual re-submit, the wrong tool choice, or the queueing latency.

Related

• Slash commands — full reference for the commands above.

• Goal mode — persistent autonomous objective that pairs with /loop.

• Plan mode — draft-before-execute mode for harder changes.\n