💡 Tips & Tricks

High-payoff omp habits, ordered roughly by leverage.

Route by role, not by hope

• Cheap model → smol (fan-out, titles, mechanical work); strong model → slow and plan.
• Stop running everything on the biggest model “to be safe”.
• Alt+M assigns roles; Ctrl+P cycles; --smol / --slow / --plan override at launch.

/omfg turns annoyance into policy

• Agent reached for a banned API or ignored a convention? Type /omfg immediately after.
• Forges a TTSR rule (time-traveling stream rule) from your complaint.
• Rule sleeps until a regex matches the live output stream, then aborts mid-token, injects the correction, retries from that point.
• Zero context cost until it fires; permanent once written. Inspect with omp ttsr.

Plan mode

• /plan (or Alt+Shift+P) makes the agent hard read-only and writes an execution spec to local://<slug>-plan.md.
• Quality bar: a competent implementer who never saw the conversation executes the file top to bottom with zero design decisions.
• Read the plan before approving.
• Trap: --plan sets the plan role’s model; it does not enter plan mode. See Traps.

ultrathink

• Bare ultrathink in a prompt requests maximum reasoning for that turn only.
• Cheaper than raising defaultThinkingLevel globally.
• Shift+Tab cycles the level live for a stretch of work.

Know your context exits

Not interchangeable:

Command	What happens	Cost
/compact	LLM summarizes older history	a summarization call
/shake	heavy tool results become recoverable artifact:// refs	free; nothing lost, just parked
/handoff	fresh session carrying a handoff document	live context is gone
/fresh	re-keys a wedged provider stream, context intact	free; use before /clear
/context	shows what is actually filling the window	free; run it before choosing

/btw and /tan protect your context

• /btw — side question against current context without polluting it.
• /tan — full background agent for tangential work. See Multi-agent.

Memory is off until you flip it

• memory.backend ships off; nothing persists between sessions.
• omp config set memory.backend mnemopi → local SQLite bank, project-scoped.
• /memory stats shows what it kept; /memory enqueue forces full retention now.

Read the meter before tuning

omp stats -s          # where the model-minutes went
omp usage             # remaining quota on every authed account
omp bench             # race your models: TTFT, tokens/s

• ~/.omp/stats.db is plain SQLite: tables messages, tool_calls, user_messages.
• One query answers e.g. “how many requests does my advisor fire per main turn”.
• Set advisor.syncBacklog, compaction, and fallback chains from numbers, not vibes.

Edits are anchored, not guessed

• Default edit mode hashline anchors patches to a content hash minted by the last read.
• Stale edit is rejected with a diff of what changed — never misapplied.
• MismatchError means the file moved under the agent; that is the system working.

eval beats bash: python -c

• eval runs persistent Python and JS cells: shared state, structured display, cancellation.
• Can call back into the agent’s own tools — tool.read(...), agent(...), parallel(...) — turning “spawn some subagents” into a loop with a budget.
• python -c loses all of that.

Everything is a path

Read this	Gets
read pr://1428	a pull request
read issue://42	an issue
read agent://Scout	a subagent’s output
read ssh://prod/var/log/app.log	a remote file
read xd://	list of deferred tools

Point at content instead of pasting it into the prompt.

Queue follow-ups instead of interrupting

• Ctrl+Q queues your next message while the agent works; Alt+Up pulls it back.
• Steering messages interrupt by default (interruptMode: immediate); set wait to let the agent finish its thought.

Pin secrets down

• secrets.enabled: true obfuscates configured secrets and credential-shaped tokens before anything leaves for a provider.
• /share runs a redactor over session snapshots (share.redactSecrets, on by default).
• Redaction targets credentials, not client data — review plain confidential data yourself.

Let long commands background themselves

• bash.autoBackground (on by default) moves slow commands into a managed background job instead of blocking the turn.
• The hub tool then greps its logs, follows them, restarts it, or sends keystrokes.
• Dev servers and watchers stop being a reason the agent “hangs”.

Import your history from other tools

• omp --from-claude / omp --from-codex import a Claude Code or Codex session with history intact.
• omp natively reads CLAUDE.md, AGENTS.md, Cursor MDC, .clinerules, Copilot instructions.

omp cleanse for the diagnostics pile

• omp cleanse fans out weighted parallel subagents to fix accumulated project type errors and lint noise.

omp git when you want hands on the wheel

• omp git — fullscreen git UI: split diffs, staging sidebar, commit composer.
• omp commit writes the message and updates changelogs.

See also: Cheat sheet, Settings, Getting started, Guide.

Condensed from the omp guide by Hugo Lopes (MIT). Verify against your install: omp config list.