Skill v1.6.0
currentAutomated scan100/100name: karina-orchestration description: >- Use Karina orchestration for structured multi-agent coordination on Windows, or to work as a dispatched worker (heartbeats, ask, worker_done): persistent threaded messages between agent terminals, blocking ask/reply flows, task DAGs with dependencies, dispatching work to supervised worker tabs, waiting for worker_done or escalation, decision gates, and coordinator loops. Use whenever work is split across agent terminals and someone must track which task runs where and who reported what. Use plain terminal commands for ordinary shell work that needs no coordination state. Use karina-workspace for a one-shot handoff that needs no task tracking (spawn an agent in a worktree, then read or wait on its terminal). version: 1.6.0
Inter-Agent Orchestration (Windows)
This file is a discovery stub, not the usage guide. The full, version-matched orchestration reference is served by the karina-cli binary itself — kept out of this file on purpose so it can never drift from the binary that will actually run your commands.
Engage Karina's orchestration surface whenever coordination state matters — when someone has to know which task is running where, who reported what, and whether a message was acknowledged. It covers persistent threaded messages between agent terminals, blocking ask/reply flows, task DAGs with dependencies, dispatching work to supervised worker tabs, waiting for worker_done or escalation, decision gates, and coordinator loops.
Use plain terminal commands for ordinary shell work that needs no coordination state.
This provider is Windows-only.
This surface talks to the running app
Unlike karina-computer-use, these commands do not act alone — Runs, tasks, dispatches, and the message inbox live inside the Karina app, and the CLI reaches it over a named pipe. The app must be running. If it is not, every command answers with a JSON error saying so; the CLI will not launch the app for you, because that would tie the app's lifetime to a single command.
Resolve the CLI for this session
Choose the executable once and reuse it for every later command:
- If the
KARINA_CLI_COMMANDenvironment variable is set, use its value. - Otherwise, use the one already on
PATH:karina-clifor the release build,
karina-dev-cli for the dev build (Settings → AI 기능 → Karina CLI puts whichever build you are running there, and each name reaches its own app).
- Otherwise, use the full path of the installed binary.
It is the CLI binary, not the app binary. Karina ships two executables that are built together: the app opens the workspace window and has no console, so it does not answer these commands; the CLI (karina-cli.exe, or karina-dev-cli.exe in a dev build) is the one that reads them and writes JSON to stdout.
Below, KARINA is a placeholder for the executable you resolved. Substitute it before running anything; do not create a shell variable or run KARINA literally. This works the same way in PowerShell, cmd.exe, and POSIX shells.
If the selected executable cannot run, report its exact error and stop. Do not fall through to another executable, which could silently target a different Karina build.
Load the full guide before running Karina commands
KARINA skills get karina-orchestration
That prints the complete, version-matched guide for the exact binary that will handle your next commands — the six domains, every command and flag, the message contract, and the error codes. Read it first, then run the specific command you need.
Don't guess subcommands or flags from memory or from a cached copy of this stub. They may change between Karina releases, and this file deliberately does not list them. Prefer --json for agent-driven calls; every response is JSON, including errors.
Drain stdout while the command runs — never wait for exit and read afterwards. This very command prints about 30 KB, more than a pipe's default buffer holds, so a caller that blocks on process exit before reading will deadlock: the CLI is blocked on a write nobody is reading, and neither side moves. The same applies to stderr if you capture it — check --wait writes a keepalive line there every 15 seconds for up to an hour. Use a call that reads and waits together (subprocess.run(..., capture_output=True), Command::output(), execFile), not wait() followed by read(). This applies to every Karina command, not just this one.
Orientation commands
These three are read-only and safe to run before you have read the guide:
KARINA orchestration run-current --jsonKARINA orchestration run-list --jsonKARINA orchestration worker-list --json
Beyond these, read the guide rather than guessing a command surface. In particular, do not invent a polling loop: the guide documents a blocking wait, and a timeout there is a checkpoint, not a failure.
Why the name starts with karina-
~/.agents/skills/ is a shared directory that several tools install into, so a skill whose name collides with another tool's gets overwritten by whichever was installed last. That is why this skill is named karina-orchestration and drives karina-cli orchestration ....
Karina coordinates on this PC only. There is no remote-host or relay-routing axis here (--to <host> and friends); do not carry those flags over from elsewhere.