swarm-out

Swarm Out

A Hive is the person's own SwarmAI on an AWS instance. A task dispatched to it runs there to the end, with nobody watching, while this machine can sleep. The run is yours once you dispatch it: you follow it, tell the person where it is when that changes, answer every wait it raises, and finish it (step 5). Nothing on the way waits for the person: they read where it is in this chat, and may step in from the Swarm Out panel, but no step needs them. The Swarm Out panel shows the same runs.

Every task goes out under a Brain's Project Agent: the Brain's definition says how the work is done there (mode, setup and checks, gates). How this one task runs (native, or a pipeline profile) you choose here, for this request (step 3). The Hive gets the Brain synced and a project section before the task. A task that belongs to no Brain goes under the General Brain's Runner Agent (--brain General).

Commands (they call the local daemon; Hive addresses and credentials stay inside it):

"$SWARM_BACKEND_BIN" py -m scripts.swarm_out hives
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out route --task-file <file> [--brain <B>] [--file <path>] [--bound-project <P>]
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out spec --brain <B> --task-file <file> --acceptance "<done when>" --execution native|pipeline:<profile> --why "<one sentence>" --returns answer|document|commits [--mode <m>] [--needs-private <why>] [--base <ref>]
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out dispatch --title "<short title>" --brief "<one or two sentences>" --brain <B> --task-file <file> --acceptance "<done when>" --execution native|pipeline:<profile> --why "<one sentence>" --returns answer|document|commits --revision <revision from spec> [--needs-private <why>] [--hive <name>] [--mode <m>] [--base <ref>]
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out runs
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out watch <dispatch_id> [<dispatch_id> ...] [--since <token>]
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out card <dispatch_id> <request_id>
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out answer <dispatch_id> <request_id> --pick "<option label>" --why "<w>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out read <dispatch_id>
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out diagnose <dispatch_id>
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out accept <dispatch_id> --ended-at <result ended_at> --why "<one sentence>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out return-prepare <dispatch_id> --url <repository url> --why "<w>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out return-land <dispatch_id> --key <key> --tip <sha> --summary-file <file> --why "<w>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out rerun <dispatch_id> --text-file <correction> --why "<w>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out stop|discard <dispatch_id> --why "<w>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out ask-person <dispatch_id> --decide "<what the person decides>" --judgment "<yours>" --why "<why you cannot decide>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out reverted <dispatch_id> [--commit <sha> ...] --why "<what you did>"
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out update-hive --hive <name> [--version <v>]
"$SWARM_BACKEND_BIN" py -m scripts.swarm_out start-hive --hive <name>

The commands on a run are refused when they are not yours: a run another session dispatched, a command the panel does not offer on it now, a third re-send, a discard while something would be lost, an answer to a wait that is not yours. The refusal says why: handle it (below). Give every one of these commands through this CLI, never as a direct HTTP call: the CLI says it is yours.

Each prints JSON; a refusal is {"ok": false, "detail": "..."}.

Only the person's own session may dispatch (the desktop, or their own channel DM): in a session with someone else the commands refuse. Do not try another way.