14 amazing things to test in oh-my-pi.

Three lowercase words typed in normal prose inject hidden, expert-level instructions for that turn. The TUI even highlights them with animated gradients as you type — the visual hint that something is wired up.

• —ultrathink — requests careful multi-step reasoning and cranks thinking to the highest effort the current model supports
• —orchestrate — scope the full task, delegate substantial independent work to parallel subagents, verify each phase
• —workflowz — deterministic multi-subagent workflow through the eval kernel's agent()/parallel()/pipeline() helpers (needs eval + task tools active)

ultrathink about the failure modes before changing this API
 
orchestrate the migration described in docs/plan.md
 
workflowz an adversarial review of the authentication changes

docs/magic-keywords.md

omp bench runs the same prompt against N models and reports time-to-first-token and generation throughput — averaged over repeated runs, optionally with a cold/warm prompt-cache test. Fuzzy selectors work, so you can race the models you actually pay for.

# race two models, 3 runs each
omp bench opus sonnet --runs 3
 
# include prompt-cache cold/warm pairs
omp bench anthropic/claude-sonnet-4-5 openai-codex/gpt-5.5 --cache
 
# machine-readable
omp bench opus gpt-5.2 --runs 3 --json

verified via omp bench --help on a live install

omp usage shows quota bars for every authenticated account across providers — plan meters with reset timers. Then learn the TUI chords; this is where omp stops feeling like a chatbot.

omp usage            # bars per account, per plan, with reset timers

docs/keybindings.md · usage verified against live accounts

/vibe turns your session into a director: your own toolset drops to read-only, and you drive persistent background worker sessions through five tools. Workers come in two tiers — fast (bundled sonic agent, cheap model, drafts & mechanical work) and good (bundled task agent, judgment calls, reviews the fast output).

• —vibe_spawn — start a worker with a self-contained brief (workers never see your conversation)
• —vibe_send — steer a running turn mid-stream, or queue the next one
• —vibe_wait — block until the first watched turn settles
• —vibe_list / vibe_kill — roster and cleanup

# inside an omp session:
/vibe fix the flaky test in packages/tui
# ...the director spawns workers; you steer with messages...
/vibe        # run again to exit — kills every worker in scope

docs/vibe-mode.md

While subagents run, hit Alt+A and get a live roster: status, model, cost, tokens, request and tool-call counts per agent — with an inspector showing the current tool call, context usage, and lineage. You can open any subagent's live transcript and type into it: steering a running turn, prompting an idle one, reviving a parked one, or killing a stuck one with x — all without touching the parent session.

# ask omp to fan out first, then:
press Alt+A          # roster — press t to toggle parent/child tree
press Enter on a row # focus that agent's transcript, type to steer
press Esc / ←←       # back to the main session

docs/agent-hub.md

Three commands that turn the messy parts of git into agent work: a P0–P3 ranked review with a ship/don't-ship verdict, atomic multi-commit splitting with dependency ordering, and a diagnostic cleanse by weighted parallel subagents.

# 1 · code review with priorities and a verdict (inside a session)
/review                    # or /review <branch> / single commit / uncommitted
 
# 2 · atomic commits: reads the tree, splits unrelated changes,
#     orders by dependency, rejects cycles, scores source > tests > docs
omp commit --dry-run
omp commit --push
 
# 3 · detect + fix project diagnostics with parallel subagents
omp cleanse -n 4           # 4 file-disjoint workers
omp cleanse -t             # also run the project test suites

README §10, §16 · subcommand help verified locally

The signature feature. A rule sits dormant until a regex (or ast-grep pattern) matches the live output stream — then omp aborts the generation mid-token, injects the rule as a system reminder, and retries from the same point. Course-correction with zero context cost on every other turn. Injections survive compaction.

---
description: ban explicit any in TypeScript
condition: "const \w+: any\b|as any"
scope: [text, tool]
interruptMode: always        # never | prose-only | tool-only | always
---
Don't use explicit `any`. Use a concrete type, `unknown` with a
narrowing check, or a generic. If truly untyped, say why.

---
astCondition: "console.log($$$)"
scope: [tool:edit, tool:write]
---
No console.log in production paths; use the logger.

omp ttsr list
omp ttsr test 'const x: any = 1'
omp ttsr test --rule .omp/rules/no-any.md --source tool --path src/foo.ts 'const x: any = 1'
omp ttsr scan          # which rules would fire across the repo

docs/ttsr-injection-lifecycle.md · frontmatter schema confirmed in src/capability/rule.ts of the installed package · omp ttsr --help verified locally

Pair a reviewer model to the advisor role. It reads every turn the main agent takes — reasoning, tool calls, results — on its own context and model, then injects notes with severity: nit (batched aside), concern (interrupting), blocker (interrupting even at a terminal answer). It catches what the doer rushed past.

omp -p --advisor "Review this task."

modelRoles:
  advisor: anthropic/claude-sonnet-4-5:medium
advisor:
  enabled: true

/advisor status    # runtime state, model, context + token usage, cost
/advisor dump      # copy its transcript to the clipboard
/advisor configure # TUI editor for WATCHDOG.yml rosters (multi-advisor)

docs/advisor-watchdog.md

Off by default. Enable it and a background pipeline reads your past sessions, extracts durable signal (decisions, constraints, resolved failures, workflows), and consolidates it into MEMORY.md + a compact summary injected at every session start — plus generated skill playbooks. Lessons you capture with learn land in learned.md, secret-redacted, capped, deduped.

memory:
  backend: local      # off | local | hindsight | mnemopi
autolearn:
  enabled: true       # unlocks the learn tool

# inside a session:
/memory view          # what will be injected
/memory stats
/memory enqueue       # force consolidation now (picked up next startup)
 
# memory is readable as URLs:
memory://root                  # the injected summary
memory://root/MEMORY.md        # full long-term document
memory://root/learned.md       # explicit lessons

docs/memory.md · docs/mnemosyne-memory-backend.md

Sixteen :// schemes resolve inside every file-shaped tool. read pr://1428 returns the same shape as read src/foo.ts. grep walks a diff like a directory. Merge conflicts become writable URLs: write @theirs to conflict://1 and the file resolves.

omp read pr://can1357/oh-my-pi/1063     # a PR as a file
omp read issue://123                    # an issue
omp read skill://tdd                    # a skill's instructions
omp read ssh://host/etc/hosts           # remote file over ssh
omp read "agent://<id>/findings.0.path" # a field out of a subagent's output

README §12, §17, §18 · omp read --help verified locally

A C binary segfaults: the agent attaches lldb, steps to the bad pointer, reads the frame. A Go service hangs: it attaches dlv and walks the goroutines. A Python process is wedged: debugpy, pause, inspect, evaluate. The debug tool speaks DAP: launch/attach, breakpoints, stepping, threads, stack, scopes, variables, memory, disassembly. omp setup installs what's missing.

omp setup                          # guided install for optional features
 
pip install debugpy                # python
go install github.com/go-delve/delve/cmd/dlv@latest   # go
# lldb-dap: ships with Xcode/LLVM on macOS (used in the README demo)
# js/ts: js-debug-adapter via Mason, tarball, or JS_DEBUG_DAP_SERVER

"the demo binary segfaults — attach the debugger and find where"
"this request handler hangs under load — attach dlv and walk the goroutines"

README §03 · docs cover adapters incl. the 17.2.11 js-debug-adapter path notes

/collab puts your live session on a relay and hands back a link and a QR code. A teammate joins from another terminal with omp join or just opens the link in a browser — no omp install on the guest side. Guests render your session natively: streaming text, tool cards, footer state, even Agent Hub control. Every frame is sealed with AES-256-GCM client-side; the relay sees only ciphertext.

/collab            # full control link (read + prompt + interrupt + hub)
/collab view       # read-only link — watch but never steer
/collab status     # link + participants
omp join "<link>"  # from any machine, any directory
/collab stop

docs/collab.md

The browser tool drives real Chromium with stealth on by default — or, with the relay extension, it adopts the Chrome tabs you already have open without stealing focus. The same API drives any Electron app in place: point it at Slack and it reads your DMs. And computer reaches the desktop itself: windows, screenshots, native input, the accessibility tree, the clipboard.

omp setup               # Chromium + optional pieces
omp browser-relay       # local CDP relay for driving your own tabs
# install the "OMP Browser Relay" Chrome extension for the relay path

"open hacker news and summarize the front page"          # headless
"read the tab I have open on stripe.com"                  # relay
"screenshot the current screen and tell me what's focused" # computer

README §20, §21 · docs/computer-use.md

An extension is a TypeScript module with the same tool API, slash-command registry, hotkey table, and TUI primitives the built-ins use — nothing is reserved. Ask omp to write the piece you're missing, then /reload-plugins. 17.2.11 added the Agent Plugins 1.0.0 standard for signed, validated packages.

ls ~/.bun/install/global/node_modules/@oh-my-pi/pi-coding-agent/examples
# → custom-tools/  extensions/  hooks/  sdk/
 
# load one ad-hoc:
omp -e ./my-extension.ts
omp --hook ./my-hook.ts

import { createAgentSession, SessionManager, ModelRegistry,
         discoverAuthStorage } from "@oh-my-pi/pi-coding-agent"
 
// ModelRegistry → createAgentSession → session.prompt("...")
// plus --mode rpc / --mode rpc-ui for non-Node embedders, and `omp acp`
// to run omp inside Zed as an editor agent

README §extensibility · docs/extensions.md · docs/sdk.md · examples dir verified in a global install