fabric
Master entry for the fabric family. Use for /fabric or when you must pick a multi-agent mode: reads the task, picks one mode, says why in one line, then follows that skill.
SKILL.mdskills/basic/fabric/SKILL.md
GitHub/fabric — pick the mode, then run it
fabric has no pipeline of its own. It reads the task, chooses one mode from the table, prints one line (fabric: <mode> — <reason>), and then follows that skill's contract. Ask the human only if two modes fit equally well.
| Mode | Pick it when | Skill |
|---|---|---|
fabric-graph |
An approved plan should be built and correctness beats speed: durable production_artifacts/state.json, Reviewer repair loop, automated quality gate, human escalation. |
fabric-graph |
fabric-multigraph |
A small throwaway fan-out (2-4 nodes, nothing persistent) in any project. | fabric-multigraph |
fabric-teamwork |
The brief is still vague and needs 9-step prompt crafting before delegation to a team. | fabric-teamwork |
fabric-swarm |
Many similar workers taking items from one shared pool (production_artifacts/swarm.json, 4 workers, cap 6). |
fabric-swarm |
fabric-loop |
A repeating cycle until an exit check is green (CI green, review rounds), bounded by round limits. | fabric-loop |
Choosing
- The task needs resumable state, a review loop or a quality gate before anything ships:
fabric-graph. - The task splits into a few independent slices and nothing has to survive the session:
fabric-multigraph. - The goal is unclear and a team will receive it:
fabric-teamworkfirst, then one of the above. - Repeat-until-done with an exit check and round limits:
fabric-loop. Many near-identical slices:fabric-swarm. - A small task needs no mode at all: do it yourself.
Pass the goal text through verbatim; flags are described below.
Flags and building blocks
--skill=<name>: handled byfabric-graphandfabric-multigraph.fabric-swarmandfabric-looppass it through to each worker brief.fabric-teamworkdoes not support it.--profile=<name>: domain profile from.agents/profiles/<name>.json; handled by the workflow script offabric-graph. An unknown profile escalates.--workstreams: workflow script offabric-graphonly; the Architect plans 1-6 dynamic workstreams (target 4, cap 6). Off by default.- Teams:
aos-team list,aos-team show <id>,aos-team new <id> [--from-playbook <name>]; add--jsonfor machine output. Team JSON lives in~/.aos/teamsor.aos/teams(project wins); shipped teams are inlib/roles/teams. - Roles: a registry of role JSON files (types plan, approve, build, review, gate, specialist). Precedence: project
.aos/rolesover user~/.aos/rolesover shipped. Role flags beyond the registry: unverified.
One loop for every harness. fabric-graph (and a fabric-loop round) runs on the aos-fabric CLI: start, then next and report by hand with your own executor, or drive for a headless run. It needs no workflow runtime and works without AO.
Agents never invoke each other. The dispatcher (the main session or the workflow script) decides every next step. Independent sub-tasks go to subagents; only work whose members must react to each other needs a team.
Safety
git push, npm publish and npm version stay behind the go-gate. Choosing a mode never grants a GO.