using-flywheel
GitHub指导用户显式选择并配置 Agentic Coding Flywheel 多智能体工厂，涵盖环境搭建、技能注入及工作流执行。强调将其作为执行适配器而非结果验证源，避免将机器收敛误判为语义正确性。
Trigger Scenarios
Install
npx skills add boshu2/agentops --skill using-flywheel -g -y
SKILL.md
Frontmatter
{
"name": "using-flywheel",
"consumes": [
"explicit-packets"
],
"metadata": {
"tier": "execution",
"effects": [],
"stability": "experimental",
"disposition": "keep_optional_adapter",
"capabilities": [
"route_to_native_flywheel_workflow",
"expose_agentops_skills",
"observe_flywheel_runtime"
],
"dependencies": [],
"canonical_status": "canonical"
},
"produces": [
"flywheel-runtime-evidence"
],
"practices": [
"team-topologies",
"design-by-contract"
],
"context_rel": [
{
"kind": "partnership",
"with": "using-gc"
}
],
"description": "Operate the Agentic Coding Flywheel as a caller-selected software factory; keep its runtime state out of AgentOps verdicts. Triggers: \"using flywheel\", \"agent flywheel\".",
"hexagonal_role": "driving-adapter",
"user-invocable": true,
"output_contract": "runtime evidence pointers for processed beads, candidate commits or worktrees, and invoked AgentOps skills; never an AgentOps verdict",
"skill_api_version": 1
}
Using the Agentic Coding Flywheel
Use the Flywheel only when the caller explicitly selects it. Treat it as a replaceable execution adapter, not a correctness or completion boundary. The adapter cannot select AgentOps semantics, issue a binding verdict, or turn factory completion into delivery or validation proof.
Insight: a factory's own completion signals — closed beads, converged agents, a quiet swarm — measure that its machinery finished, not that the result is semantically correct. This skill exists to prevent the failure mode of reporting Flywheel convergence as an AgentOps PASS.
Choose the factory first
AgentOps supports two external software-factory runtimes: Gas City (using-gc) and the Agentic Coding Flywheel. Use this skill only for the Flywheel. AgentOps supplies skills and evidence contracts to either factory; it does not wrap one factory in the other, and it owns no formulas, roles, or orchestration inside either.
What the Flywheel is
Jeffrey Emanuel's free, open-source stack that turns a dedicated VPS into a supervised multi-agent factory: Claude Code, Codex CLI, and Antigravity CLI as worker runtimes, coordinated through NTM (orchestration), Agent Mail (coordination and file reservations), Beads/BV (task graph), and the wider Flywheel toolset. Its methodology is planning-first: decompose work into beads, run agent swarms against them, and detect convergence between agent outputs.
Not for: single-session local work (use the default one-agent loop), Gas City cities (use using-gc), or as a verdict source.
Inputs
- An explicit caller selection of the Flywheel as the factory.
- A provisioned Flywheel host, or authorization to provision one via the upstream wizard.
- The caller-owned work intent (beads or a decomposable goal).
Procedure
-
Provision with the upstream wizard at agent-flywheel.com (OS selection through final verification; a single-curl install on a dedicated Ubuntu VPS), then run its
onboardtutorial once. AgentOps does not fork, pin, or mirror the Flywheel stack; upstream owns its installer and versions. -
Make AgentOps skills visible to each worker runtime the Flywheel drives (Claude Code, Codex CLI, Antigravity CLI), using the install paths in the repository README. Verify per runtime:
ls ~/.claude/skills/validate ~/.codex/skills/validate 2>/dev/nullA runtime that cannot list the skill will never invoke it.
-
Run the Flywheel's native workflow — bead decomposition, swarm dispatch, convergence — unchanged. Skill presence and skill invocation are different facts: when an AgentOps skill's use is an acceptance condition, name it on the work item or worker prompt, and check the transcript for its use.
-
Read Flywheel runtime state (bead graph, Agent Mail threads, NTM pane truth) as evidence pointers only. When agents disagree with tracker state, trust the more direct observation: pane truth over roster claims, bead state over prose.
-
Two consecutive non-converging swarm rounds on the same intent is the stop condition: stop, report the divergence to the caller, and do not dispatch a third round on your own authority.
Output
Runtime evidence pointers for the caller: which beads the Flywheel processed,
where the candidate commits and worktrees live, and which AgentOps skills its
agents actually invoked. Done when the caller holds those pointers and — if
proof was requested — a fresh validate context has judged the exact
candidate. Factory state alone is never the done signal.
Checks
- Every runtime the Flywheel drives can list the AgentOps skills it is expected to use (step 2 command exits 0).
- No Flywheel bead close, convergence signal, or agent self-report was translated into an AgentOps PASS, FAIL, or verdict.
- Provenance: the factory boundary this skill applies is declared in README.md ("Choose a software factory") and docs/operations/gas-city-reliability.md.
Failure behavior
Report the concrete failure — unreachable host, missing skill visibility, non-convergence, or an upstream installer change — and stop. Upstream Flywheel defects belong to its maintainer; the caller owns any revision, retry, or factory switch.
Version History
- 7b07a7d Current 2026-08-19 22:01
