scaffold

Scaffold the engineering standards

Set up a repository so the engineering standards are enforced by tooling, not by memory. This is goal-driven and stack-agnostic: detect the stack and machine first, then realize each goal in that stack's idioms. Generate artifacts tailored to what you detect and have the human confirm before writing — never copy a fixed bootstrap or assume a language. For the standards themselves, defer to the engineering-standards guardrail skill; this skill operationalizes them.

1. Detect first

Before generating anything, detect:

• Language(s) and package manager / build tool (manifests, lockfiles).
• Test runner, and whether an acceptance/BDD layer already exists.
• Whether nix is available (and direnv); the OS / platform.
• Existing toolchain pins, lint config, CI config, ADRs — extend these, never clobber them.

Summarize what you found and the proposed plan, then confirm with the human before writing files.

2. Realize each goal in the detected stack

For each area: hit the goal, understand the why, then realize it with the detected stack's idioms. Skip an area only when the ecosystem genuinely cannot support it — and record that gap as an ADR.

a. Reproducible dev environment

• Goal: anyone — and CI — gets the identical toolchain from one command, and "global" package-manager installs never leak into $HOME.
• Why: reproducibility and isolation; machine state can't silently drift.
• Realize: if nix is available, generate a flake devShell that pins the toolchain and redirects "global" installs into a git-ignored, project-local sandbox (e.g. a .dependencies/ dir, via the package manager's prefix/cache env vars); wire direnv (use flake) if present. Otherwise use the stack's nearest equivalent — a version manager (e.g. mise, asdf), a per-runtime pin (e.g. .tool-versions, .nvmrc/volta, a toolchain file, .python-version), or a devcontainer — and still redirect global installs into a git-ignored project-local dir. Add the sandbox dir to .gitignore.

b. Pinned toolchain

• Goal: exact compiler/runtime/tool versions are declared in-repo and used everywhere.
• Why: eliminate "works on my machine"; make upgrades explicit, reviewable changes.
• Realize: pin via whatever the environment from (a) reads (flake inputs, a tool-versions file, a toolchain file, an engines field, …). Manage dependencies through the package-manager CLI so versions and feature flags are checked at the time of change — never hand-edit the manifest.

c. Strict lint allowlist

• Goal: the strictest practical static analysis is on by default, warnings fail the gate, and every relaxation is a deliberate, documented decision.
• Why: the friction is the point — confronting each lint forces an intentional choice about what correct code looks like here. Only ever ratchet stricter.
• Realize: turn on every lint group/level the toolchain offers as groups, treat warnings as errors in the gate, and forbid panic-prone constructs on production paths. Then relax individual lints only with a narrowly-scoped, reason-carrying suppression recorded as a project-policy decision — never a blanket allow, never to save time. Wire format + lint into the local gate.

d. Mutation testing

• Goal: the tests are proven to detect injected faults — a 100% mutant kill target where a tool exists.
• Why: coverage shows a line ran; mutation shows the tests would actually catch a regression.
• Realize: if the ecosystem has a mutation-testing tool, wire it, target 100% kill, and gate it in CI (release-gated if it is slow). If none exists, record the gap in an ADR and compensate with the strongest property-based and coverage discipline the stack offers.

e. Black-box BDD / acceptance harness

• Goal: externally-observable behavior (including edge cases) is covered by executable specs that exercise only the public surface, built one step at a time.
• Why: black-box specs survive refactors; one-step-at-a-time keeps every commit green and reviewable.
• Realize: set up the stack's acceptance/BDD tooling (Gherkin-style or plain specs that drive the CLI / API / UI from outside). Specs must never touch internal modules or private types. Establish the rhythm: get one Given/When/Then step green with all gates passing, commit, then the next step. For cross-target or cross-harness behavior, parametrize examples so parity is part of each slice's definition of done.

f. Decisions + guardrail docs

• Goal: every architectural decision is recorded, and the standards live in one harness-agnostic place.
• Why: future readers — human and agent — need the why and the rules without a tool-specific scavenger hunt.
• Realize: create docs/adr/ with a MADR-lite template and an NNNN-*.md numbering scheme; seed ADR 0001 with the overall architecture. Put the guardrails in AGENTS.md + docs/rules/, and make harness-specific files (e.g. CLAUDE.md) thin pointers to AGENTS.md.

g. CI/CD themes

• Goal: changes land via reviewed PRs and ship through managed, automated releases — on whatever forge and runner the project already uses.
• Why: the gate must be enforced off the author's machine, and releases must be repeatable.
• Realize: configure CI to mirror the local gate exactly (format, lint as errors, tests, mutation, dependency audit). Require PR-based changes with ≥1 approval and automated code review contributing to it. Add managed/automated releases (versioning, changelog, publish). Stay platform-agnostic — detect the forge and runner from the repo; never hard-code one provider, registry, or service as a requirement.

3. How to apply

1. Detect the stack + machine, read existing config, summarize the plan, and confirm with the human.
2. Realize each applicable goal, generating artifacts tailored to the stack; adapt or skip what the ecosystem can't support and record the gap as an ADR.
3. Wire one local gate command that runs format + lint + tests (+ mutation), so the gate is a single command locally and in CI.
4. Make the first commit establish the harness and a green gate; then build behavior one BDD step at a time, committing between steps.
5. Validate: the gate command passes, the dev environment builds from a clean checkout, and CI mirrors the local gate.

For per-ecosystem tool choices and copy-paste templates (env-redirect snippets, strict-lint configs, the ADR template, the guardrail-docs layout, and a provider-agnostic CI gate outline), see references/playbook.md — load it on demand when you need the concrete artifact for the detected stack.