集群摄入量
- 作者仓库星标 2
- 作者仓库 anton
- 领域
- 人工智能
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @wcygan · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 运行环境要求
- 未检出明确要求
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
在一个段落中重述,通过询问用户来填补任何空白:这是什么?名称、项目 URL、helm 图表源。什么意图?其中之一:具体需求——“我需要这个,因为现在我不能___。”诚实学习——“我想学习___;我知道它可能会被删除;我的时间范围是___;我的退出计划是___。”必须填写所有四个空白。
首先,检查该组件或其类别是否已在 anton 的删除历史记录中。移除墓地已迁移到 ADR 系统中: 然后查询 ADR 索引以获取先前的 `Reverted` 决策: `Glob('context/adrs/0[0-9][0-9][0-9]-*.md')` — 列出所有 ADR 对于每个 ADR,读取前文并筛选 `status: Reverted` 列出任何 `affects:` 与候选人类别匹配的内容,*或*其标题明确引用相同组件的内容
七门。每个都是通过/失败,测量时间≤2 分钟。完整的测量程序 → 标题。在学习意图下,仅评估门 1-5(遏制门 - 无论意图如何,它们都会保护集群的其余部分)。门 6 和门 7 被标题 → 包含的学习变体中描述的特定于学习的检查所取代。许可和退出成本——许可的 OSS;卸载是 `flux suspend` + `kubectl delete ns`,没有您仍然需要的孤立终结器或 CRD。 *(适用于所有意图。)*
项目 GitHub 上的五分钟预算。危险信号不会自动失败,而是会流入软分数中。完整信号列表 → 可持续性信号。最低浏览量: 过去 30 天内的提交 过去 6 个月的发布 超过一名活跃维护者
7 个属性的得分 `2` / `1` / `0`,加权。如果总数 < 22/34,则拒绝。完整矩阵 → 标题。在学习意图下,跳过此步骤,而是填写包含学习清单 → 标题 → 包含学习变体。重点是遏制(命名空间隔离、Tier-0 不干扰、可预测的爆炸半径),而不是可持续性——如果项目有两位明星和一位维护者,只要用户睁大眼睛并有时间限制就可以参与其中。
在同一矩阵上得分“不执行任何操作”。如果候选人的“无所作为”分数在 4 分之内,则拒绝——摄入税本身就价值约 4 分。在学习意图下,决胜局不适用。学习具有无所作为所没有的价值——重点是用户想要接触它。
与著名的家庭实验室地雷进行交叉检查。 → 已知的不良模式。任何未声明学习摄入量的比赛都将被自动拒绝。几种模式(3 个节点上的 HA Postgres、完整的 LGTM 可观察性、自托管电子邮件)具有明确的“学习摄入变体如果……”则可以排除;在自动拒绝之前阅读该模式的学习笔记。
每个判决——添加、推迟或拒绝——都会通过移交给`adr`技能(`/adr new`)来记录。该技能拥有规范的 anton ADR 模板(`.Codex/skills/adr/references/template.md`),分配下一个 NNNN,并在`context/adrs/`下写入 ADR 文件。不要停留在“ADR 就绪”——判决必须在`context/adrs/` 中得出。四种判决形状:添加(具体需要)-所有 7 个硬门均通过,软分数 ≥22,明确什么都不做,没有已知的不良匹配。将`adr`技能交给`status: Accepted`、`intent: concrete-need`。然后交给`add-flux-app`到脚手架清单。
# Cluster intake
Read-only decision gate for "should this component join the anton cluster?" Walks hard vetoes first, then a weighted soft score, then returns an add / defer / reject recommendation with an ADR-ready summary. Never scaffolds files, never applies manifests, never commits — that is what `add-flux-app` is for, *after* this skill says yes.
## Why this skill exists
Anton has paid real tuition on loose intake — Rook-Ceph, the full LGTM observability stack, CloudNativePG, Dragonfly, TiDB, Scylla, Redpanda, and Harbor were all installed and later removed. But not every removal was a mistake. Some of those components (TiDB, Scylla, Redpanda) were honest learning experiments that served their purpose: tried it, learned it, moved on. Anton is partly a learning cluster, and "things that don't scale" are explicitly welcome when that's the point.
This skill is a gate, not a wall. It catches the *wrong* kind of intake — completionism dressed up as need, speculative production patterns, and adds that will quietly accrete into maintenance debt — while making space for honest learning intake with a declared timebox and an exit plan.
## Three intents, three rubrics
Every intake decision starts with **what is this for?** The user must pick one:
| Intent | Example sentence | Rubric applied |
|---|---|---|
| **Concrete need** | "I need this because right now I cannot ___." | Full production rubric — hard gates + sustainability + full soft score |
| **Honest learning** | "I want to learn ___; I know it may be removed; my timebox is ___; my exit plan is ___." | Contained-learning rubric — blast-radius / reversibility / integration / secrets gates only + timebox + exit plan |
| **Completionism-as-need** | "Every real cluster has ___." / "My stack feels incomplete without ___." | Reject. The pattern is what produced the LGTM removal. |
The skill's job is to accept the first two honestly and reject the third honestly. It does **not** lecture the user on whether a learning experiment is "worth it" — that's their call. The gate's job is to make sure learning intake is *contained* (can't take down the rest of the cluster) and has a *known exit* (won't accidentally become permanent).
## Scope
| Concern | Owner | Output |
|---|---|---|
| Evaluate a candidate component | **this skill** | add / defer / reject + ADR summary |
| Scaffold a new Flux app | **`add-flux-app`** | 3-file manifests |
| Expose a service | **`expose-service`** | HTTPRoute + DNSEndpoint |
| Secret store choice (SOPS vs ESO) | **`anton-repo-conventions`** | reference |
| Cluster health before/after add | **`anton-cluster-health`** | triage |
| Remove / retire an existing component | out of scope | — |
## Hard rules
- **Never scaffold files** during intake. The rubric runs *before* any `add-flux-app` invocation. If intake passes, hand off explicitly.
- **Never run mutating commands** — no `kubectl apply`, `flux reconcile`, `sops -e -i`, `git commit`, `helm install`, `gh pr merge`.
- **Never recommend adding anything already in the removal graveyard** without the user stating an explicit delta: what changed about the component, what changed about the cluster's need, OR that the intent this time is declared learning (which is itself a valid delta from a prior concrete-need attempt).
- **Never skip the containment gates.** Gates 1–5 (license, blast radius, reversibility, integration fit, secrets model) apply under every intent. A learning experiment that forks the secrets story or takes out Cilium is still a reject.
- **Never lecture the user on whether their learning experiment is worthwhile.** The gate's job is containment and honest intent declaration, not judging the educational value of the user's choices. If the user declares honest learning intent with a timebox and an exit plan, your job is to help them run it safely, not to argue them out of it.
- **Never accept "concrete need" from a completionism sentence.** If the user's "I cannot ___" resolves to "my stack feels incomplete" or "every real cluster has this," ask once whether they want to reframe as a learning intake. If they insist it's concrete need when it isn't, reject under Gate 7.
## Green-path: the 30-second heuristics
Apply the short-circuit filter first. If any fire, stop — do not run the full rubric, do not fill a matrix. Details → [heuristics](references/heuristics.md).
## Full workflow
### Step 1 — Frame the candidate and declare intent
Restate in one paragraph, filling any gap by asking the user:
- **What is it?** Name, project URL, helm chart source.
- **What intent?** One of:
- **Concrete need** — "I need this because right now I cannot ___."
- **Honest learning** — "I want to learn ___; I know it may be removed; my timebox is ___; my exit plan is ___." All four blanks must be filled.
- **Both** — there's a real need AND a learning angle. Apply the concrete-need rubric; the learning angle is a bonus, not a softener.
- **What is the current alternative?** "Do nothing" or an existing component or a SaaS — name it explicitly.
- **What state does it own?** Stateless, ephemeral PVC, persistent PVC, cluster-wide CRDs, mutating webhooks.
- **Who uses it?** You alone, family, public internet.
If the user cannot pick an intent, or picks "concrete need" but can't complete the sentence, ask once: **is this actually a learning intake?** It's legitimate to say yes — honest learning intake is an accepted path on anton. What is **not** legitimate is smuggling completionism in as concrete need; if the "I cannot ___" sentence turns out to be "my stack feels naked without it," it's completionism — reject under Gate 7.
Intent is the axis that decides which rubric applies:
- **Concrete need or both** → full production rubric (all 7 hard gates + sustainability + full soft score)
- **Honest learning** → contained-learning rubric (gates 1, 2, 3, 4, 5 only; skip sustainability gating; require a timebox and exit plan)
- **Completionism-as-need** → reject, cite pattern #10 in [known-bad-patterns](known-bad-patterns.md)
### Step 2 — Check the removal graveyard (now in `context/adrs/`)
Before anything else, check if this component or its category is already in anton's removal history. The removal graveyard has been migrated into the ADR system:
```sh
git log --oneline --all --grep='remove\|rip\|drop\|abandon\|descope' -i | head -50
```
Then query the ADR index for prior `Reverted` decisions:
- `Glob('context/adrs/0[0-9][0-9][0-9]-*.md')` — list all ADRs
- For each, read the frontmatter and filter for `status: Reverted`
- Surface any whose `affects:` matches the candidate's category, *or* whose title clearly references the same component
- Read each matched ADR's "Re-adoption guidance" section — that holds the conditions under which a re-attempt would be valid
- Read `context/adrs/RE-ADOPTION-RUBRIC.md` for the policy on which categories require an explicit intent declaration before re-adopting
If the candidate or its category matches a `Reverted` ADR, demand an explicit delta from the user (concrete-need reframing) **or** an explicit learning-intake declaration with timebox + exit plan (learning-intake reframing). The matched ADR's Re-adoption guidance section spells out the acceptable forms. If neither path applies, reject under the matched ADR's lesson.
### Step 3 — Run the hard vetoes (any red = reject, no override)
Seven gates. Each is pass/fail, measured in ≤2 minutes. Full measurement procedure → [rubric](references/rubric.md#hard-gates).
Under **learning intent**, only gates 1–5 are evaluated (the containment gates — they protect the rest of the cluster regardless of intent). Gates 6 and 7 are replaced by the learning-specific checks described in [rubric → Contained learning variant](references/rubric.md#contained-learning-variant).
1. **License & exit cost** — permissive OSS; uninstall is `flux suspend` + `kubectl delete ns` without orphaned finalizers or CRDs you still need. *(Applies to all intents.)*
2. **Blast radius** — failure degrades *only* the component's own app, not CNI/DNS/storage/auth/ingress. *(Applies to all intents.)*
3. **Reversibility** — removable in <30 min with no data-migration ritual. *(Applies to all intents.)*
4. **Integration fit** — speaks Gateway API + ESO + Prometheus + SOPS. No parallel ingress / secret / cert / monitoring stack. *(Applies to all intents — a learning experiment that forks the secrets story is still a mess.)*
5. **Secrets model** — pulls from 1Password via ESO. No hardcoded admin passwords. No sealed-secrets or its-own-secret-CRD. *(Applies to all intents.)*
6. **Tested restore runbook (stateful, concrete-need only)** — written *before* install. Untested backups are folklore. Under learning intent, a stateful component is allowed without a restore runbook **if and only if** the user explicitly acknowledges the data is throwaway — see the contained-learning variant.
7. **Honest intent (concrete-need only)** — the user's "I need this because right now I cannot ___" sentence resolves to a real present-day problem, not completionism in disguise. Under learning intent, this gate is replaced by the timebox + exit-plan check.
### Step 4 — Skim sustainability signals
Five-minute budget on the project's GitHub. Red flags do not auto-fail but flow into the soft score. Full signal list → [sustainability-signals](references/sustainability-signals.md).
Minimum skim:
- Commits in last 30 days
- Release in last 6 months
- More than one active maintainer
- Official helm chart in-repo (not community; k8s-at-home/charts is **deprecated** — treat community charts as yellow at best)
- CNCF sandbox / incubating / graduated status (prior, not a gate)
- Median time-to-first-response on open issues
### Step 5 — Fill the weighted soft score *(concrete-need intent only)*
Score `2` / `1` / `0` across 7 attributes, weighted. Reject if total < 22/34. Full matrix → [rubric](references/rubric.md#soft-scoring).
Under **learning intent**, skip this step and instead fill the contained-learning checklist → [rubric → Contained learning variant](references/rubric.md#contained-learning-variant). The focus is containment (namespace isolation, Tier-0 non-interference, predictable blast radius), not sustainability — it's fine if the project has two stars and one maintainer, as long as the user is going into it with eyes open and a timebox.
### Step 6 — Do-nothing tiebreaker *(concrete-need intent only)*
Score "do nothing" on the same matrix. If the candidate is within 4 points of do-nothing, reject — the intake tax is worth ~4 points by itself.
Under learning intent, the tiebreaker does not apply. Learning has value that do-nothing doesn't — the whole point is the user wants to touch it.
### Step 7 — Known-bad pattern check
Cross-check against well-known homelab landmines. → [known-bad-patterns](references/known-bad-patterns.md). Any match that is **not** a declared learning intake is an auto-reject. Several patterns (HA Postgres on 3 nodes, full LGTM observability, self-hosted email) have an explicit "learning-intake variant is okay if…" carve-out; read the pattern's learning note before auto-rejecting.
### Step 8 — Return a structured recommendation, then hand off to `adr`
Every verdict — Add, Defer, or Reject — is recorded by handing off to the **`adr` skill** (`/adr new`). That skill owns the canonical anton ADR template (`.Codex/skills/adr/references/template.md`), allocates the next NNNN, and writes the ADR file under `context/adrs/`. Do not stop at "ADR-ready" — the verdict must land in `context/adrs/`.
The four verdict shapes:
- **Add (concrete need)** — all 7 hard gates pass, soft score ≥22, do-nothing clear, no known-bad match. Hand off to `adr` skill with `status: Accepted`, `intent: concrete-need`. Then hand off to `add-flux-app` to scaffold manifests.
- **Add (learning)** — gates 1–5 pass, the contained-learning checklist is complete (timebox, exit plan, containment verified), the user has consciously accepted the sustainability / resource trade-offs. Hand off to `adr` skill with `status: Accepted`, `intent: learning`, and `review-by:` set to the declared review date (a learning intake without a review date is just permanent intake with extra steps — if missing, downgrade to Defer). Then hand off to `add-flux-app`.
- **Defer** — a blocking prerequisite is missing (e.g., no timebox declared for a learning intake, no tested restore runbook for stateful concrete-need intake, marginal soft score under concrete need). Hand off to `adr` skill with `status: Deferred` — the ADR's body holds the exact conditions to unblock. No further hand-off; the ADR *is* the durable record.
- **Reject** — any containment gate (1–5) fails, completionism-as-need detected, known-bad match without a learning-intake carve-out, or historical-removal match without stated delta. Hand off to `adr` skill with `status: Rejected` — the ADR's body explains which gate and why, and — if appropriate — whether reframing as a declared learning intake would change the answer. No further hand-off.
## What this skill does NOT do
- Does not write manifests — that is `add-flux-app`
- Does not expose services — that is `expose-service`
- Does not pick between SOPS and ESO — that is `anton-repo-conventions`
- Does not verify cluster health — that is `anton-cluster-health`
- Does not install the thing, ever
- Does not make unsolicited recommendations for components the user has not named
## Related skills
- `add-flux-app` — run *after* this skill returns **Add**
- `anton-repo-conventions` — for SOPS vs ESO decisions once a candidate is accepted
- `anton-cluster-health` — verify before and after adding anything non-trivial
- `anton-upgrade-audit` — every new component raises the future Renovate-PR tax; pair intake with upgrade cadence
## Anti-patterns
- **Rejecting honest learning intake.** If the user has declared learning intent with a timebox and exit plan, and the candidate passes gates 1–5, do not veto it for being "speculative" or "doesn't solve a present-day problem" — that's the whole point of a learning cluster.
- **Accepting completionism disguised as need.** The user says "I need this" but the sentence that completes "I cannot ___" is really "my stack feels naked." Ask once to reframe; reject if they don't.
- **Running Step 5 before Step 3.** Soft-scoring a candidate that will fail a containment gate is wasted work.
- **Making exceptions to containment gates (1–5).** The word for that is "regret." These protect the rest of the cluster regardless of intent.
- **Skipping Step 2 (the removal graveyard).** History is the cheapest teacher anton has. But note that a graveyard hit under declared learning intent with a stated delta is valid — "I want to actually learn Rook-Ceph this time, timebox 30 days, exit plan is delete the namespace" is an honest learning intake.
- **Lecturing the user on whether their learning is "worth it."** Not your call.
- **Evaluating components the user did not ask about.** This skill is reactive, not speculative.
按说明文件中的顺序排列。
为什么存在这个技能
三个意图,三个标准
适用范围
硬性规定
绿色路径:30 秒启发法
文件/命令 -> add-flux-app · expose-service · anton-repo-conventions · anton-cluster-health · kubectl apply · flux reconcile · sops -e -i · git commit
内容校验值 -> 82210dadea39
当前原文的自动摘译,尚未人工审校。
为什么存在这个技能
Anton 在松散的摄入量上支付了真正的学费——Rook-Ceph、完整的 LGTM 可观测性堆栈、CloudNativePG、Dragonfly、TiDB、Scylla、Redpanda 和 Harbor 都已安装,后来又被删除。但并非每次移除都是错误。其中一些组件(TiDB、Scylla、Redpanda)是诚实的学习实验,达到了它们的目的:尝试它,学习它,继续前进。安东在某种程度上是一个学习集群,当这就是重点时,“无法扩展的东西”是明确欢迎的。
三个意图,三个标准
每个录取决定都从这个目的是什么开始?用户必须选择一个: 意图 · 例句 · 应用的标题 具体需求 · “我需要这个,因为现在我不能 ___。” · 完整的生产标题 — 硬门 + 可持续性 + 完整的软分数 诚实学习 · “我想学习 ___;我知道它可能会被删除;我的时间范围是 ___;我的退出计划是 ___。” · 封闭式学习标题 — 爆炸半径/可逆性/集成/仅秘密门 + 时间盒 + 退出计划
适用范围
关注点 · 所有者 · 输出 评估候选组件 · 此技能 · 添加/推迟/拒绝 + ADR 摘要 构建新的 Flux 应用程序 ·
add-flux-app· 3 个文件清单 公开服务 ·expose-service· HTTPRoute + DNSEndpoint 秘密存储选择(SOPS 与 ESO) ·anton-repo-conventions· 参考硬性规定
切勿在摄入期间搭建文件架。该标题在任何
add-flux-app调用之前运行。如果摄入通过,请明确交接。切勿运行变异命令 - 禁止kubectl apply、flux reconcile、sops -e -i、git commit、helm install、gh pr merge。切勿在用户没有声明明确的增量的情况下添加移除墓地中已有的任何内容:组件发生了什么变化,集群需求发生了什么变化,或者这次的意图被声明为学习(这本身就是之前具体需求尝试的有效增量)。绿色路径:30 秒启发法
首先应用短路滤波器。如果发生任何火灾,请停止 - 不要运行完整的标题,不要填充矩阵。细节→启发法。