跳到正文
cccskills
免费在 GitHub 公开

wiring-contract

Derive this project's own integration conventions from what the code already does, draft them as a wiring contract, and hold new work to it. Closes the built-but-not-wired gap at L3 (walking skeleton) — a mechanism that ships and is never called reads green in every other check.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(1)

  • SKILL.md5.7 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Wiring Contract

Answer one question for this project and then keep answering it automatically: when something is built here, what does it have to be connected to before it counts as done?

The failure is not carelessness, it is structural. A model's context makes it unable to see whether the element it just generated is referenced anywhere else, so plausible-but-unreferenced code is the expected output. Across 304,362 verified AI-authored commits in 6,275 repositories, "unused variables or parameters" is the second most frequent issue class, and "undefined variable or reference" — writer and reader disagreeing about a name or path — is the most common runtime bug category. It does not wash out: AI-authored code survives longer than human-written code (53.9% vs 69.3% line death rate), so an orphan is less likely to be cleaned up than a human's.

Every ordinary signal misses it. Tests pass, coverage rises, lint is clean, review approves — because each of those verifies a thing against itself, and wiring is a property of a pair.

When to Use

  • At L3, before implementation — the walking-skeleton obligation (G-V13). A design is not complete until its joins are named.
  • On an existing project, once, to discover what conventions it already has.
  • After a structural change (new directory, new layer, a move) — regenerate, never hand-edit.
  • When check_wiring_contract.py reports UNGOVERNED files you did not expect.

Workflow

1. Detect — read the repo's own majority convention

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_wiring_contract.py" --root . --detect

This proposes a rule only where the repo already behaves that way, and records the evidence:

- id: wc-002
  pattern: "scripts/*.py"
  obliges:
    - referenced_by: ["**/*"]
  detected_from: "18 of 20 files already satisfy this"
  confidence: 0.90

A rule without observed support is a preference, not a contract. Two obligation types exist and that is deliberate — referenced_by (something outside the file references it, by import, invocation, path, or a glob that matches it) and sibling_reference (a same-directory barrel or registry lists it). Nine specific obligations rot; two general ones that always run do not.

2. Review with the human — this is a draft, not a finding

Present the proposal and ask, per rule, in plain language: "18 of 20 scripts here have a caller. Should that be required?" Watch for two failure shapes:

  • A majority that is an accident. Four files in a directory that happen to be referenced does not make referencing them a rule. Ask what the directory is for.
  • A convention the detector cannot see. Files invoked by reflection, by a registry built at runtime, or by a config string are wired and will look orphaned. These need an exemption with a reason — and "it is hard to check" is not one.

Write the confirmed contract to .claude/harness/wiring-contract.yml. Per the evidence-writes rule, the contract is a human-confirmed draft, never written unattended.

3. Enforce — and put it where the agent will feel it

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_wiring_contract.py" --root .

Per G-V14, add this to the command run after every change, not only to CI. Agents respond to automated signals far more reliably than to documentation: a failing command gets fixed, a guardrail paragraph in the context window may not be read at all.

4. Read the UNGOVERNED count — it is the load-bearing number

Files matching no rule are listed with a count. Zero violations across a governed subset is not zero violations. A clean report beside 300 ungoverned files describes the health of a fraction while implying the whole — the precise false green this skill exists to end. Either extend coverage or state why those files are outside it.

Output

  • Creates/updates: .claude/harness/wiring-contract.yml (human-confirmed)
  • Reports: violations, and the ungoverned count
  • Feeds: /mycelium:definition-of-done (G-V14 verification block), /mycelium:bvssh-check (CALMS Automation wiring_integrity)

Theory Citations

  • Cockburn — walking skeleton: "a tiny implementation of the system that performs a small end-to-end function… it should link together the main architectural components." Wiring exists before features.
  • Hunt & Thomas — tracer bullets: the same instinct, framed as end-to-end feedback first.
  • Ford, Parsons, Kua — architecture fitness functions: objective, continuously-executed measures of properties the design must preserve. This check is one, and belongs in the pipeline rather than in a review checklist.

Deliberately NOT a catalogue

Mycelium does not ship a list of integration points to check. A sibling project hand-listed 4 component and 9 plugin integration points; such a list is stale the moment someone adds the tenth, and hand-enumerated scope is the origin of this entire failure family. The rules here are derived from the project in front of you, and regenerated when it changes.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Accessibility audit, scoped to the surfaces a product actually has. Detects web / rendered_markdown / terminal / native_app / video_audio / document / headless, then applies only the criteria that bind. WCAG 2.1 AA in full for web; not at all for headless.

中文简介正在准备中。当前显示原文说明。

haabe/mycelium472026年10月18日 更新

adopt

免费

Bring Mycelium into a project that already has code. Detects that the repo predates the framework, asks before touching anything, then reads the codebase to draft what it CAN establish (delivery, solution shape) and — the actual point — names what it cannot (purpose, strategy, real user evidence). The output is a discovery backlog with a head start, never a filled canvas.

中文简介正在准备中。当前显示原文说明。

haabe/mycelium472026年10月18日 更新

Design the smallest viable test to validate or invalidate a critical assumption. Based on Torres's assumption testing framework, organized by Gilad's AFTER model (Assessment → Fact-Finding → Tests → Experiments → Release Results).

中文简介正在准备中。当前显示原文说明。

haabe/mycelium472026年10月18日 更新

Use before any research activity or significant decision. Reviews cognitive biases relevant to the current stage.

中文简介正在准备中。当前显示原文说明。

haabe/mycelium472026年10月18日 更新

Use to evaluate whether current work aligns with Better Value Sooner Safer Happier. Run at diamond completion and periodically.

中文简介正在准备中。当前显示原文说明。

haabe/mycelium472026年10月18日 更新

Lint canvas files for staleness, missing fields, inconsistent evidence types, and orphaned references. Run periodically or before major transitions.

中文简介正在准备中。当前显示原文说明。

haabe/mycelium472026年10月18日 更新

查看 haabe 的全部技能

报告此技能的问题