iii-hq/skills-and-validation

최근 기록된 소스 활동: 2026년 5월 28일 19:21

SkillsMP 카탈로그 업데이트: 2026년 11월 2일 18:02

수집된 skills: 5

GitHub 스타: 6
GitHub 포크: 1

AI 어시스턴트로 설치

이 프롬프트를 복사해 사용 중인 AI 어시스턴트에 입력하세요.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/iii-hq/skills-and-validation.

저장소 README

README 원본 파일: README.md

README 원본 보기

README 전체 펼치기

# skills-and-validation Render and validate iii worker docs against the project's voice, structure, and Diataxis rules. You write short markdown partials under `<worker>/docs/`; this project renders them into the worker's `README.md` and a single self-contained `skill.md`, then verifies the result with Vale and an AI pass on every commit and every PR. It also supports a **docs mode** for standalone (Mintlify, Fumadocs, etc.) `.md` / `.mdx` documentation. --- ## Install There are a few prerequisites to install; the pre-commit hook is optional. ### 1. Vale Vale is used to validate prose against static rules. ```bash brew install vale # macOS or Linux Homebrew vale --version # confirm it's on PATH ``` Other installs: [vale.sh](https://vale.sh) or the [release page](https://github.com/errata-ai/vale/releases). The composite GitHub Action installs Vale on the CI runner for you, so only local setups need this step. ### 2. The `iii-skill-render` and `iii-skill-check` binaries ```bash curl -fsSL https://raw.githubusercontent.com/iii-hq/skills-and-validation/latest/scripts/install.sh | bash ``` The installer places binaries at `~/.local/bin/iii-skill-{render,check}` and the bundle (rules, Vale styles, skill bundles, templates, scripts) at `~/.local/share/skill-check/current/`. Add `~/.local/bin` to your `PATH` if it isn't already. ### 3. The `iii-skill-authoring` skill bundle This is the bundle agents (and you) read to know how to write worker partials. Pick whichever surface fits your tooling. **`skillkit`**: installs both `iii-skill-authoring` and `iii-doc-authoring`: ```bash cd $HOME && npx skillkit add iii-hq/skills-and-validation/content/skills ``` ### 4. The pre-commit hook (optional; per repo) `cd` into the repo whose workers you're validating, then run the installer. The script refuses to install into this repo (it's only meaningful in consumer repos). ```bash cd /path/to/your/consumer-repo ~/.local/share/skill-check/current/scripts/install-hook.sh ``` The hook symlinks into `.git/hooks/pre-commit`. On every commit it: 1. Detects staged paths under any worker dir (`<worker>/iii.worker.yaml` + `<worker>/docs/`). 2. Re-renders each affected worker with `iii-skill-render --write`. 3. Re-stages the rendered `README.md` and `skill.md`. 4. Runs `iii-skill-check verify-rendered` + `iii-skill-check verify --layers structure,vale`. 5. Blocks the commit on remaining violations. The hook deliberately skips the AI layer (slow, costs tokens). CI runs the AI layer on every PR. ### Upgrading Re-run install steps 2 and 3. --- ## Configure your repo Two files per consumer repo: `.skill-check.yaml` at the root, and a GitHub Actions workflow. ### `.skill-check.yaml` Copy from `~/.local/share/skill-check/current/templates/.skill-check.yaml` and edit. Worker-mode minimum: ```yaml version: 1 ai_check: provider: anthropic model: claude-opus-4-7 api_key_env_var: ANTHROPIC_API_KEY max_tokens: 6000 ``` | Field | Purpose | | -------------------------- | ------------------------------------------------------------------------------------------------ | | `version` | Schema version. `1` = worker mode (implicit). `2` requires the `mode` field below. | | `mode` | v2 only. `worker` or `docs`. Leave unset on v1 for worker mode. | | `ai_check.provider` | Currently only `anthropic`. | | `ai_check.model` | Anthropic model id, e.g. `claude-opus-4-7`. | | `ai_check.api_key_env_var` | Env var holding the API key. Read by the validator, the action, and `scripts/verify-workers.sh`. | | `ai_check.max_tokens` | Output token budget per AI call. | ### GitHub Actions workflow Copy `~/.local/share/skill-check/current/content/github_workflows_example.yml` to `.github/workflows/skill-check.yml` and pin the action ref. The minimum: ```yaml name: skill-check on: pull_request: push: branches: [main] permissions: contents: read # report-only by default; set to write to opt into auto-commit pull-requests: write # opt-in: enables the sticky PR comment actions: write # opt-in: persistent AI skip-cache jobs: skill-check: runs-on: ubuntu-latest steps: # No explicit ref: — checkout defaults to refs/pull/N/merge on # pull_request, which works for same-repo and fork PRs. Setting # `ref: github.head_ref` breaks fork PRs (head branch isn't here). - uses: actions/checkout@v5 - uses: iii-hq/skills-and-validation@v0.4 with: anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }} ``` Set `ANTHROPIC_API_KEY` as a repo secret so the AI layer runs in CI. Without it, CI runs structure + Vale only and emits a workflow warning. `workers-glob` defaults to `*/iii.worker.yaml`. Override it (or set `config-path`) for non-default layouts. Full input reference under [Action inputs](#action-inputs). ### Re-render on demand (checkbox in the PR comment) When `scope: pr-diff` is in use, the verify step can pass while rendered artifacts drift from sources (drift is detected globally; verify only checks files in scope). On a `write: false` run, the sticky comment ends with a checkbox: ```markdown - [ ] **Re-render this branch and commit rendered artifacts** ``` Adopt the opt-in listener so the box triggers a re-render. Add this **separate** workflow file: ```yaml # .github/workflows/skill-check-recheck.yml name: skill-check recheck on: issue_comment: types: [edited] # Reusable workflows do NOT inherit the callee's `permissions:` — # the caller has to grant them, or GitHub fails the job with a # startup_failure before any step runs. permissions: contents: write # push the auto-render commit pull-requests: write # update the sticky comment actions: write # AI skip-cache (actions/cache) jobs: recheck: uses: iii-hq/skills-and-validation/.github/workflows/recheck-on-comment.yml@v0.4 with: config-path: .skill-check.yaml # match the main workflow's config-path action-ref: v0.4 # match the main workflow's pin secrets: anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }} ``` Two critical constraints: 1. **The file must live on the default branch.** GitHub runs `issue_comment` workflows from `main` only (security feature). Adding it on a feature branch is a no-op for that PR; merge to `main` first, then the box works on all future PRs. 2. **You must grant `permissions:` in this consumer file.** Reusable workflows don't inherit the callee's permission declaration; the caller does. Gating: the user who edited the comment (`github.event.sender`) must have `admin` / `maintain` / `write` access on the repo, checked via the collaborators API at run time. The bot that posted the comment is irrelevant to authorization. PR authors who lack write access can't trigger the auto-commit; they're expected to re-render locally and push themselves. Fork PRs short-circuit with an explanatory comment, because `GITHUB_TOKEN` cannot push to forks. --- ## Author a worker's README and SKILLs A worker that uses this validator keeps its SKILL.md and README.md source material under `docs/`. You edit short partials; the renderer produces three artifacts. Never edit the rendered files by hand. ### Use the iii-skill-authoring skill The bundle in `content/skills/iii-skill-authoring/` is the canonical reference. After step 3 of [Install](#install), read individual topics: ```bash skillkit read iii-skill-authoring/quickstart # what makes a good ## Quickstart skillkit read iii-skill-authoring/structure # the slot order in each rendered artifact skillkit read iii-skill-authoring/skeleton # copy-paste starter for a new worker skillkit read iii-skill-authoring/voice # voice rules (slop lists, banned phrasing) skillkit read iii-skill-authoring/leaves # per-function `docs/leaves/<leaf>.md` bodies skillkit read iii-skill-authoring/llm-only-blocks # spans visible to agents, hidden from humans skillkit read iii-skill-authoring/check # running the validator locally ``` The same files are on disk under `~/.local/share/skill-check/current/content/skills/iii-skill-authoring/` and the engine surfaces them through `iii://` once the `skills:` glob is set. Coding agents in the consumer repo should read these topics before editing partials. ### What to write (inputs) ``` <worker>/ ├── iii.worker.yaml # name (required) + description + tags (recommended) feed the searchable frontmatter; plus build characteristics ├── config.yaml # worker runtime config; inlined verbatim under ## Configuration └── docs/ ├── intro.md # intro paragraph(s) inserted after the title (README + skill.md) ├── quickstart.md # body of ## Quickstart (README only) ├── companions.md # appended inside "## Install" and covers common workers that a user # or agent may want to use alongside this worker (optional). ├── migration.md # body of ## Migration notes (optional, README only) # Used to document any breaking changes that require migration steps. └── leaves/ └── <leaf>.md # inlined under ## Additional HOWTOs in both README and skill.md; write any number # they cover specifics that individual commands that an agent needs to # know how to do (ex: enqueue.md, chmod.md). # They DO NOT cover function signatures, outputs, or other API-level documentation. # They ARE (like the skills) intended to be a how-to do something. # They CAN be multi-step common actions like `check-queue-success.md`. ``` Notes on the inputs: - **`docs/intro.md`**: one or two short paragraphs covering what the worker does, who calls it, and the single most important thing it gives you. Can contain ` ... ` blocks for agent-only routing hints. - **`docs/quickstart.md`**: the meat of the README. Show one fenced code block per SDK language (Rust `iii_sdk`, TypeScript `iii-sdk`, Python `iii`), ≤ 30 lines each, with a realistic payload and the expected output/result. One to three functions, chosen for introductory value, not breadth. - **`docs/leaves/<leaf>.md`**: the leaf name is the suffix of the function id after the last `::`. `textstats::analyze` → `docs/leaves/analyze.md`. H1 is a topical phrase (`# Sizing text before provider calls`), never the function id. Canonical sections: `## When to use` (three to five bullets of realistic call sites) and `## Notes` (gotchas, edge cases an agent will trip on). - **`config.yaml`**: runtime config for the worker itself. The renderer inlines it verbatim under `## Configuration`. It preserves comments and all `config.yaml` entries should contain accompanying comments on usage. Function signatures, payload schemas, and `RegisterFunction::new("…").description("…")` text are generated by a separate auto-gen system (iii-directory worker). Don't duplicate them in the partials. ### Outputs: what gets rendered ``` <worker>/ ├── README.md # published on iii.dev (human-facing) └── skill.md # single agent-facing file for iii://<worker> (leaves inlined) ``` Two files, no `skills/` directory: each leaf is inlined into both artifacts under `## Additional HOWTOs`. Both files open with an identical YAML frontmatter block sourced from `iii.worker.yaml` (`name` is always present; `description` and `tags` when set) that makes the worker searchable in the registry. A missing `description` or `tags` is omitted from the block and reported as a structure-layer warning, not an error. Below the frontmatter, each artifact starts with a generated banner comment; everything after that is derived from your partials. Rendered files should never be hand-edited (the structure layer will flag drift and re-renders blow away your edits anyway). **README.md slot order:** 1. Frontmatter block (`name` / `description` / `tags`). 2. Generated banner. 3. `# <name>` (from `iii.worker.yaml.name`). 4. `intro.md`. 5. `## Install` + `iii worker add <name>` boilerplate, optionally followed by `companions.md`. 6. `## Quickstart` + `quickstart.md`. 7. `## Configuration` + code blocked `config.yaml`. 8. `## Migration notes` + `migration.md` (only if present). 9. `## Additional HOWTOs`: each leaf inlined, its `# Title` demoted to `### Title` (omitted when `docs/leaves/` is empty). **skill.md slot order** (llm-only blocks unwrapped, human-only dropped): 1. Frontmatter block (identical to README's). 2. Generated banner. 3. `# <name>`. 4. `intro.md`. 5. `companions.md`. 6. `## Additional HOWTOs`: same inlined leaves as the README. **Inlined leaves:** each `docs/leaves/<leaf>.md` is appended under `## Additional HOWTOs` with its headings demoted by two levels (`#` → `###`, `##` → `####`). A top-level H1 is required per leaf (the structure layer flags its absence). ### Render and verify locally ```bash iii-skill-render <worker> --write # produces README.md + skill.md (removes any stale skills/ dir) iii-skill-check verify <worker> # all three layers (structure + Vale + AI) iii-skill-check verify <worker> --layers structure,vale # offline; no API key needed iii-skill-check verify <worker> --layers ai # AI only — fastest signal after a tweak iii-skill-check verify-rendered <worker> # confirm rendered artifacts match partials ``` `verify-rendered` re-renders in memory and diffs against the on-disk files. Non-zero exit means an artifact drifted; re-run `iii-skill-render <worker> --write`. The pre-commit hook does this for you on every commit. For the AI layer, export the key named in your `.skill-check.yaml`: ```bash export ANTHROPIC_API_KEY=sk-ant-… ``` --- ## `llm-only` and `human-only` blocks A single source partial can carry three audiences cleanly: plain prose (everyone sees it), `llm-only` spans (only the agent-facing artifact gets them), and `human-only` spans (only the human-facing render gets them). Both block (`:start`/`:end`) and inline forms are implemented for both directives. | Shape | HTML form (worker `.md`) | MDX form (`.mdx`) | | ------ | ------------------------------------------------------- | ----------------------------------------------------- | | Block | `` … `` | `{/* llm-only:start */}` … `{/* llm-only:end */}` | | | `` … `` | `{/* human-only:start */}` … `{/* human-only:end */}` | | Inline | `` | `{/* llm-only: short note */}` | | | `` | `{/* human-only: short note */}` | Marker lines must each sit on their own line. The structure layer balance-checks both block types per file; an unclosed `:start` fails verify. HTML and MDX forms count together, so a block opened in one form and closed in the other still balances. ### Worker `.md` partials (every form works) Worker partials are never rendered to humans directly. The renderer produces `README.md` (humans) and `skill.md` (agents) from the same source, so block and inline forms both work fully for both `llm-only` and `human-only`. Use whichever fits the payload. ```markdown ## Quickstart This worker does X. Use `worker::get` and `worker::set`.  Prefer `get` over `set` for read-only flows; `set` invalidates the cache.   **Heads-up for maintainers:** the legacy `foo` cli wrapper is under `tools/legacy/foo.sh`. It's deprecated.  ``` In the rendered README the `llm-only` block is invisible and the `human-only` block reads as a normal paragraph. In `skill.md` the inverse: `llm-only` body is expanded, `human-only` block is dropped entirely. Inline form: each comment must be on a line of its own. The renderer expands it to its payload for the visible side and drops it on the hidden side. Embedded mid-paragraph inline comments are not parsed and pass through verbatim. ```markdown The worker exposes `worker::set_token`.  Add the worker.  ``` ### `.mdx` docs sources exception: Do not use mutli-line tags. Mintlify treats MDX comments as line-by-line MDX files in docs mode are rendered to humans directly by Mintlify. Mintlify treats `{/* … */}` as an invisible single-line comment per line; it does not collapse a multi-line `{/* … */}` span into one comment. That asymmetry decides which form works for each directive: - **`llm-only` in `.mdx`**: only the inline form hides from humans, and the comment must be on a line of its own (mid-paragraph inline comments are not parsed): ```mdx The worker exposes `set_token`. {/* llm-only: call this before any other op; tokens cache for 60s */} ``` The block form `{/* llm-only:start */}` … `{/* llm-only:end */}` is still parsed by the renderer (so `skill.md` is correct), but the prose between the markers is regular MDX content that Mintlify renders to readers. The `:start`/`:end` lines themselves are invisible; the body between them leaks. There is no way to hide a multi-line `llm-only` span in MDX today; keep the payload to a short inline comment, or move it to a worker `.md` partial. - **`human-only` in `.mdx`**: use the block form when you want humans to see the payload: ```mdx {/* human-only:start */} Maintainers: this image is rebuilt nightly. Bump the tag in `iii.lock` when you upgrade. {/* human-only:end */} ``` Mintlify treats both marker lines as invisible comments and renders the body in between as ordinary prose, which is exactly what `human-only` wants for the human-facing side. The renderer then strips the entire block from `<source>.skill.md` so the agent never sees it. The inline form `{/* human-only: … */}` does _not_ work in `.mdx`; Mintlify drops the whole comment, so the payload is invisible to readers too. No render pass runs between the source and Mintlify to expand it. Reserve inline `human-only` for worker `.md` partials and for maintainer notes the agent should never see (the dropped-from-`skill.md` behavior, with no human surface). Full authoring guides ship under `content/skills/iii-skill-authoring/llm-only-blocks.md` (worker mode) and `content/skills/iii-doc-authoring/llm-only-blocks.md` (docs mode); `human-only-blocks.md` exists in both bundles. Read via `skillkit read iii-skill-authoring/llm-only-blocks` etc. ### Relationship to `skill:...` markers The `` markers (docs mode only) decide _whether_ a doc or section enters the skill set at all. `llm-only` and `human-only` decide _which spans_ of an in-scope source go to which audience. They compose: a marker block inside an `` section is dropped regardless, because the whole section is excluded. --- ## Fix errors the validator finds The validator emits one line per violation in this format: ``` <file>:<line>:<severity> — <message> ``` `<severity>` is `error` (fails the run) or `warning` (surfaces but doesn't fail). AI failures appear under `[AI] <path>` blocks at the end with the model's full violation list. In CI the same violations surface three ways: inline annotations on the Files diff, a markdown table in the Checks tab run summary, and (with `pull-requests: write`) a sticky PR comment that updates in place on each push. ### Common error categories and how to fix them **Vale slop / marketing language** (`Terminology.SlopMarketing`, `Terminology.SlopMagic`, `Terminology.SlopEase`, `Terminology.SlopConnection`, `Terminology.SlopFlow`) ``` textstats/docs/intro.md:3:error — Avoid marketing/anthropomorphic phrasing 'blazing fast'. Use concrete, technical language. ``` Rewrite with concrete technical language. `blazing fast` → `sub-millisecond at p99`. `effortless` → drop the adjective and show the call. `wire up X to Y` → `register X; Y invokes it`. The full token lists live in `~/.local/share/skill-check/current/content/styles/Terminology/`. **Em dashes** (`Terminology.EmDash`) Rewrite with commas, parentheses, periods, or colons. The rule is at error level. **Forbidden terms** (`Terminology.ForbiddenTerms`, `Terminology.BackendSoftware`) Includes the bare term `telemetry` (disambiguate: `OpenTelemetry` / `observability` for traces/metrics/logs, or `iii-telemetry` for usage analytics) and a handful of backend-software terms with project-specific alternates. **Diataxis voice drift in a how-to** (`Diataxis.HowTo`, `Diataxis.CrossContamination`) Worker partials are how-tos. Phrases like `in this guide you will learn` or `step 1`, `step 2` are tutorial framing and get flagged. Rewrite as direct instructions: `Run X.`, `Then Y.`. `CrossContamination` warnings are advisory; the others are errors. **Structure violations** (structure layer) Missing or out-of-order sections in `README.md`, install command doesn't match `iii.worker.yaml.name`, source-build instructions in the rendered output, an `iii://<name>/<leaf>` link that doesn't resolve, or an unbalanced llm-only marker. Most of these mean a rendered file was hand-edited or a partial drifted from the slot order in [Outputs](#outputs-what-gets-rendered). Re-render and fix the source partial. **Unbalanced llm-only block** (structure layer) Every `llm-only:start` needs a matching `llm-only:end` on its own line. HTML and MDX forms count together, so you can open with one form and close with the other. **AI layer failure** (`[AI] <path>`) The AI layer reads the project rules in `content/project-rules/` (voice, general, workers, sdks, …) and reviews each rendered artifact against them. Failures cite specific passages. Treat the model's feedback as a peer review: the surfaced issue is usually real, but the specific rewrite it suggests is a starting point, not the only fix. **Render drift** (`verify-rendered` non-zero exit) The on-disk artifact doesn't match what the partials would produce. Run `iii-skill-render <worker> --write`, stage the result, retry. ### When auto-fix mode is on With `write: true` plus `contents: write` on the action, the bot commits rendered output back to the PR branch (only after `verify` passed). If the bot's `chore: auto-render skill artifacts` commit lands before your next local push, rebase preferring the upstream side (the bot's render is the authoritative one): ```bash git fetch origin git rebase -X ours origin/<branch> ``` `-X ours` during a rebase resolves conflicts in favour of the side being rebased _onto_. Double-check non-artifact files with `git diff @{u}..` before pushing. ### Bypassing - A single commit: `git commit --no-verify`. - Re-running only the AI layer after a small tweak: `iii-skill-check verify <worker> --layers ai`. - Letting an old binary keep running while you update: pass `--allow-old-version`. - Suppressing the update check (offline, CI, batch scripts): `export SKV_NO_UPDATE_CHECK=1`. --- ## Docs mode The same render → verify → optional auto-commit pipeline also runs against standalone Mintlify-shaped `.md` / `.mdx` documentation. Opt in with `version: 2` and `mode: docs` in `.skill-check.yaml`; each in-scope doc renders into a sibling `<source>.skill.md`. Diataxis ruleset is selected per doc by the frontmatter `type:` field (`tutorial`, `how-to`, `reference`, `explanation`). Full guide: ```bash skillkit read iii-doc-authoring/quickstart skillkit read iii-doc-authoring/frontmatter skillkit read iii-doc-authoring/types skillkit read iii-doc-authoring/markers #  etc. skillkit read iii-doc-authoring/llm-only-blocks # MDX caveats ``` Worker and docs modes share the validator binary; what differs is the unit and the rendered artifacts: | | Worker mode | Docs mode | | -------------------------- | ----------------------------------------------------------------- | ---------------------------------------------------- | | Unit | A worker dir (one `iii.worker.yaml`) | One `.md` / `.mdx` doc | | Sources | `<worker>/docs/*.md`, `iii.worker.yaml`, `config.yaml` | The doc itself, plus its YAML frontmatter | | Rendered artifacts | `<worker>/README.md`, `<worker>/skill.md` (leaves inlined) | `<source>.skill.md` sibling next to each doc | | Action input scope | `workers-glob` (default `*/iii.worker.yaml`) | `docs-glob` (default `**/*.md **/*.mdx`) | | `.skill-check.yaml` schema | v1 (no `mode`) or v2 with `mode: worker` | v2 with `mode: docs` + `docs.include`/`docs.exclude` | --- ## Action inputs | Input | Default | Description | | ------------------- | ------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `version` | from `.skill-check.yaml` | Pinned validator version, without the `v` prefix | | `config-path` | `.skill-check.yaml` | Path to the `.skill-check.yaml` controlling this run. Override per matrix entry to validate multiple modes in one repo. | | `workers-glob` | `*/iii.worker.yaml` | Worker mode: glob of worker manifests to verify. Ignored in docs mode. | | `docs-glob` | `**/*.md **/*.mdx` | Docs mode: glob(s) of doc files to verify (space-separated). Filtered per-file against `docs.include`/`docs.exclude`. | | `layers` | `structure,vale,ai` | Comma-separated subset of layers to run | | `vale-version` | `3.14.1` | Pinned Vale version | | `anthropic-api-key` | (none) | API key for the AI layer; AI is auto-skipped when unset | | `write` | `false` | Auto-render and commit the diff back to the PR branch when sources drift from rendered output. Requires `contents: write`. | | `scope` | `all` | `all` validates every artifact matching the glob; `pr-diff` (PR events only) restricts to files changed against the merge base. A diff that touches any `.skill-check.yaml` falls back to full scan. In worker mode, a changed file under `<worker-dir>/` validates the whole worker since rendering is per-worker. | The action auto-detects mode by reading the `.skill-check.yaml` named by `config-path`. Repos that mix worker dirs and docs run the action multiple times via a matrix strategy with one entry per controlling config; the sticky PR comment is keyed off `config-path` so each entry gets its own comment. --- ## Troubleshooting **`vale: command not found`.** Vale is required for local runs; see [Install → 1. Vale](#1-vale). **`verify-rendered` fails after editing a `docs/` partial.** Re-render the worker: `iii-skill-render <worker> --write`. The hook does this automatically on commit. **Bot keeps auto-committing on top of my pushes.** See [When auto-fix mode is on](#when-auto-fix-mode-is-on). If you'd rather re-render locally than trust the bot, disable `write: true` in your workflow. **Bundle lookup misses on a moved binary.** `bundle::find_content_root` walks up from the binary looking for `content/` with `project-rules/` and `.vale.ini`. Pass `--rules-dir` and `--vale-config` explicitly to `iii-skill-check verify` if you've relocated the layout. **Anonymous download returns 404 on a public repo.** Confirm the asset name matches `skills-and-validation-{version}-{target}.tar.gz` exactly. The git tag has the `v` prefix; the asset filename does not. **`cargo test` fails on path-dependent tests after renaming the repo dir.** `CARGO_MANIFEST_DIR` is baked into the test binary at compile time. Run `cargo clean` to force a rebuild. --- ## Layout (for contributors to this repo) ``` crates/iii-skill-core — shared lib (render, structure, vale, ai, config, bundle) crates/iii-skill-render — render-only binary (no network deps) crates/iii-skill-check — verify + verify-rendered binary (Vale + AI) content/ — project-rules, styles, skills/, .vale.ini templates/ — .skill-check.yaml + example-worker the consumer copies fixtures/ — intentionally broken/targeted workers used by tests scripts/ — shared between the composite action and pre-commit hook action.yml — composite action consumed via `uses: iii-hq/skills-and-validation@v0.3` ``` Local end-to-end check: ```bash ./scripts/test-e2e.sh # offline phases (build, fixtures, scripts) ANTHROPIC_API_KEY=sk-ant-… ./scripts/test-e2e.sh # also exercises the live AI layer ``` Pass `--clean` if you've renamed the repo directory (`CARGO_MANIFEST_DIR` is baked into test binaries at compile time).

저장소 README를 불러오는 중…