diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json new file mode 100644 index 0000000..6483120 --- /dev/null +++ b/.claude-plugin/marketplace.json @@ -0,0 +1,15 @@ +{ + "name": "html-kit", + "description": "Self-hosted marketplace for the html-kit plugin.", + "owner": { + "name": "KN", + "url": "https://github.com/nyashkn" + }, + "plugins": [ + { + "name": "html-kit", + "source": "./", + "description": "Turn any plan, spec, decision doc, synth, or dataset into a polished, self-contained single-page HTML artifact, styled in the Anthropic palette, with one /html-kit command inside Claude Code." + } + ] +} diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json new file mode 100644 index 0000000..e631fd8 --- /dev/null +++ b/.claude-plugin/plugin.json @@ -0,0 +1,19 @@ +{ + "name": "html-kit", + "version": "0.5.0", + "description": "Turn any plan, spec, decision doc, synth, or dataset into a polished, self-contained single-page HTML artifact in the Anthropic palette. Skills: html-kit, html-kit-drain, html-kit-add-to-patterns.", + "author": { + "name": "KN", + "url": "https://github.com/nyashkn" + }, + "homepage": "https://github.com/nyashkn/html-kit", + "repository": "https://github.com/nyashkn/html-kit", + "license": "MIT", + "keywords": [ + "html", + "artifacts", + "skills", + "visualization", + "claude-code" + ] +} diff --git a/.github/CONTRIBUTING.md b/.github/CONTRIBUTING.md index ce3dda1..81da72a 100644 --- a/.github/CONTRIBUTING.md +++ b/.github/CONTRIBUTING.md @@ -6,7 +6,7 @@ adding a new archetype. ## Dev setup -Prerequisites: [Bun](https://bun.sh) and Claude Code. +Prerequisites: [Bun](https://bun.sh) >= 1.1 and Claude Code. ```bash bun install # dev deps: Playwright, Pagefind, @types/bun @@ -21,9 +21,9 @@ re-run after pulling updates or adding a skill. The suite is [Playwright](https://playwright.dev): ```bash -bun test # headless run -bun test:headed # visible browser -bun test:ui # Playwright UI mode +bun run test # headless run +bun run test:headed # visible browser +bun run test:ui # Playwright UI mode ``` If this is your first run, install the browser binaries once: @@ -39,20 +39,28 @@ An archetype is just a directory under `archetypes/`. To add one: 1. Create `archetypes//exemplar.html` — a complete, self-contained page (inline CSS / SVG / JS, no external assets) using the Anthropic palette tokens from `tokens/anthropic-palette.css`. If it's vendored from a - third-party source, keep the inline `` - comment at the top and add it to `ATTRIBUTION.md`. -2. Create `archetypes//recipe.md` — *when to use* this archetype, the - *expected data shape*, and a `## Gotchas` section. The router skill reads - these to choose between archetypes, so be specific. -3. Add a row to the "Archetype selection cheatsheet" in - `skill/html-kit/SKILL.md`. -4. If the README's archetype count changes, update it. + third-party source, keep its license header/provenance comment at the top + and add it to `ATTRIBUTION.md` (note the exact license — upstream sources + can change license over time, don't assume MIT). +2. Create `archetypes//recipe.md` starting with frontmatter: + ```yaml + --- + name: + when: + --- + ``` + followed by *when to use* this archetype, the *expected data shape*, and a + `## Gotchas` section. +3. Run `bun run cheatsheet` — it regenerates the "Archetype selection + cheatsheet" in `skills/html-kit/SKILL.md` and the archetype count in + `README.md` from every `recipe.md`'s frontmatter. Don't hand-edit either + table. CI runs `bun run cheatsheet --check` and fails if you forgot. No installer edit is needed — archetypes are discovered at render time. ## Adding a skill -Drop a directory under `skill//` with a `SKILL.md` (YAML frontmatter + +Drop a directory under `skills//` with a `SKILL.md` (YAML frontmatter + body) and re-run `./install.sh`. Skills are auto-discovered; no edits to the installer are required. diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml new file mode 100644 index 0000000..3d823b8 --- /dev/null +++ b/.github/workflows/ci.yml @@ -0,0 +1,170 @@ +name: CI + +on: + push: + branches: [main] + pull_request: + +permissions: + contents: read + +jobs: + test: + name: Playwright suite + runs-on: ubuntu-latest + timeout-minutes: 15 + steps: + - uses: actions/checkout@v4 + + - uses: oven-sh/setup-bun@v2 + with: + bun-version: latest + + - name: Install dependencies + run: bun install + + - name: Check generated cheatsheet is up to date + run: bun run cheatsheet --check + + - name: Install Playwright browsers + run: bunx playwright install --with-deps chromium + + # playwright.config.ts targets a singleton daemon on :63839 rather than + # a webServer block (the daemon is PID-locked and shouldn't be + # respawned per test run), so start and health-check it ourselves. + # + # Health-check hits /api/repos rather than "/": "/" 404s until + # ~/.html-kit/_daemon/index.html has been built (never true yet here — + # HOME is the runner's empty home), while /api/repos always returns + # 200 JSON regardless of build state. Playwright's globalSetup + # (tests/global-setup.ts) registers this repo and builds the index. + - name: Start html-kit daemon + run: | + bun scripts/html-kit-daemon.ts > daemon.log 2>&1 & + for i in $(seq 1 30); do + if curl -sf http://localhost:63839/api/repos > /dev/null; then + echo "daemon is up" + exit 0 + fi + sleep 1 + done + echo "daemon did not become healthy in time" + cat daemon.log + exit 1 + + - name: Run Playwright suite + run: bun run test + + plugin-eval-free: + name: Plugin eval (free graders) + needs: [] + runs-on: ubuntu-latest + timeout-minutes: 30 + if: github.event_name == 'push' || github.event.pull_request.head.repo.full_name == github.repository + steps: + - uses: actions/checkout@v4 + + - uses: oven-sh/setup-bun@v2 + with: + bun-version: latest + + - name: Install dependencies + run: bun install + + - name: Install sandbox backend (bubblewrap + socat) + run: | + sudo apt-get update -qq && sudo apt-get install -y -qq bubblewrap socat + if [ "$(sysctl -n kernel.apparmor_restrict_unprivileged_userns 2>/dev/null)" = "1" ]; then + printf 'abi ,\ninclude \nprofile bwrap /usr/bin/bwrap flags=(unconfined) {\n userns,\n include if exists \n}\n' | sudo tee /etc/apparmor.d/bwrap >/dev/null + sudo systemctl reload apparmor + fi + + - name: Install Claude Code CLI + run: | + curl -fsSL https://claude.ai/install.sh | bash -s 2.1.273 + echo "$HOME/.local/bin" >> "$GITHUB_PATH" + + - name: Run free-grader eval cases + env: + CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }} + run: | + claude plugin eval . \ + --tag free \ + --trust-plugin \ + --scaffold \ + --allow-tools Bash Write Edit \ + --ablation none \ + --model claude-haiku-4-5 \ + --threshold 0.8 \ + --max-cost-usd 5 \ + --no-publish \ + --json evals/results/free-result.json + + - name: Summarize free eval results + if: always() + run: | + jq -r '.cases[] | .name as $n | .arms.with[] | "\($n): score=\(.score) error=\(.error // "none")", (.graders[]? | " \(.name): \(.passed)")' evals/results/free-result.json || true + + - name: Upload free eval report + if: always() + uses: actions/upload-artifact@v4 + with: + name: eval-free-report + path: evals/results/ + if-no-files-found: ignore + + plugin-eval-paid: + name: Plugin eval (llm/baseline graders) + runs-on: ubuntu-latest + timeout-minutes: 30 + if: github.event_name == 'push' && github.ref == 'refs/heads/main' + steps: + - uses: actions/checkout@v4 + + - uses: oven-sh/setup-bun@v2 + with: + bun-version: latest + + - name: Install dependencies + run: bun install + + - name: Install sandbox backend (bubblewrap + socat) + run: | + sudo apt-get update -qq && sudo apt-get install -y -qq bubblewrap socat + if [ "$(sysctl -n kernel.apparmor_restrict_unprivileged_userns 2>/dev/null)" = "1" ]; then + printf 'abi ,\ninclude \nprofile bwrap /usr/bin/bwrap flags=(unconfined) {\n userns,\n include if exists \n}\n' | sudo tee /etc/apparmor.d/bwrap >/dev/null + sudo systemctl reload apparmor + fi + + - name: Install Claude Code CLI + run: | + curl -fsSL https://claude.ai/install.sh | bash -s 2.1.273 + echo "$HOME/.local/bin" >> "$GITHUB_PATH" + + - name: Run paid-grader eval cases + env: + CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }} + run: | + claude plugin eval . \ + --tag paid \ + --trust-plugin \ + --allow-tools Bash Write Edit \ + --model claude-haiku-4-5 \ + --judge-model claude-haiku-4-5 \ + --threshold 0.8 \ + --max-cost-usd 15 \ + --no-publish \ + --json evals/results/paid-result.json + + - name: Summarize paid eval results + if: always() + run: | + jq -r '.cases[] | .name as $n | .arms.with[] | "\($n): score=\(.score) error=\(.error // "none")", (.graders[]? | " \(.name): \(.passed)")' evals/results/paid-result.json || true + + - name: Upload paid eval report + if: always() + uses: actions/upload-artifact@v4 + with: + name: eval-paid-report + path: evals/results/ + if-no-files-found: ignore diff --git a/.gitignore b/.gitignore index 6e6cd8e..87ef0d0 100644 --- a/.gitignore +++ b/.gitignore @@ -13,3 +13,4 @@ playwright-report/ !.html-kit/.config.json !.html-kit/01_v0.3.0-before-after-eval.html !.html-kit/02_v0.4.0-interaction-loop.html +.html-kit/_archive/ diff --git a/ATTRIBUTION.md b/ATTRIBUTION.md index 51f6649..c1789e7 100644 --- a/ATTRIBUTION.md +++ b/ATTRIBUTION.md @@ -1,8 +1,14 @@ # Third-party attribution -## Vendored exemplars — `thariqs/html-effectiveness` +All HTML exemplars under `archetypes//exemplar*.html` are vendored from +[ThariqS/html-effectiveness](https://github.com/ThariqS/html-effectiveness). +That upstream repo relicensed from MIT to Apache-2.0 on 2026-05-18; which +license applies to a given exemplar depends on when it was vendored (see the +two sections below and each file's own header comment). -The HTML exemplars under `archetypes//exemplar*.html` are vendored from: +## Vendored exemplars — MIT (`thariqs/html-effectiveness`, root-level) + +The exemplars vendored on 2026-05-14 (before the upstream relicense) are: - **Source:** [ThariqS/html-effectiveness](https://github.com/ThariqS/html-effectiveness) - **License:** MIT @@ -19,6 +25,38 @@ The exemplars have been adapted (Anthropic palette tokens, `data-component` annotations, structural trims) for use as html-kit archetype templates; the original MIT terms continue to apply to the upstream-derived portions. +## Vendored exemplars — Apache-2.0 (`thariqs/html-effectiveness`, `unknowns/`) + +Upstream relicensed from MIT to Apache-2.0 on 2026-05-18. The 11 files below +were vendored on 2026-09-05 (commit `82fa96f`), after the relicense, from the +`unknowns/` sub-page — they carry Apache-2.0 in their own file headers and are +**unmodified** copies of the upstream source (no palette/structural edits): + +| html-kit file | Upstream source | +|---|---| +| `archetypes/blindspot-pass/exemplar.html` | `unknowns/01-blindspot-pass.html` | +| `archetypes/vocab-teacher/exemplar.html` | `unknowns/02-color-grading-explainer.html` | +| `archetypes/design-reaction-review/exemplar.html` | `unknowns/03-design-directions.html` | +| `archetypes/throwaway-mock/exemplar.html` | `unknowns/04-toolbar-mock.html` | +| `archetypes/intervention-brainstorm/exemplar.html` | `unknowns/05-churn-brainstorm.html` | +| `archetypes/requirements-interview/exemplar.html` | `unknowns/06-interview.html` | +| `archetypes/reference-port-verify/exemplar.html` | `unknowns/07-reference-port.html` | +| `archetypes/tweakable-plan/exemplar.html` | `unknowns/08-implementation-plan.html` | +| `archetypes/implementation-notes/exemplar.html` | `unknowns/09-implementation-notes.html` | +| `archetypes/buy-in-pitch/exemplar.html` | `unknowns/10-pitch-doc.html` | +| `archetypes/merge-quiz/exemplar.html` | `unknowns/11-change-quiz.html` | + +Each file's header comment reads: + +```html + +``` + +Upstream copyright: **Copyright 2026 Anthropic PBC**. Full license text: +[`LICENSES/Apache-2.0.txt`](LICENSES/Apache-2.0.txt). Upstream carries no +`NOTICE` file (checked at repo root), so there is nothing to reproduce under +Apache-2.0 §4(d). + ## Other dependencies Runtime/dev dependencies are declared in `package.json` and carry their own diff --git a/CHANGELOG.md b/CHANGELOG.md index 26ffe66..ec32146 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,52 @@ are pinned in `package.json` / skill manifests; the heading dates use ISO-8601. ## [Unreleased] +## [0.5.0] - Unreleased + +Plugin packaging: html-kit installs via `/plugin` alongside the existing +`install.sh` path. + +### Added + +- `.claude-plugin/plugin.json` + `.claude-plugin/marketplace.json` — plugin + manifest and self-hosted single-plugin marketplace (`source: "./"`). +- `hooks/hooks.json` — `PostToolUse` hook runs render-audit on every Write/Edit + and feeds audit failures back to Claude after each write (the write itself + already happened — `PostToolUse` doesn't block it). `render-audit.ts` also + gained external-asset checks (fails a render that references a CDN + script/style/font/image — artifacts must be self-contained) and an opt-out + for the palette-token check via + ``. +- `scripts/build-cheatsheet.ts` (`bun run cheatsheet`, `--check` for CI) — + regenerates the archetype cheatsheet in `skills/html-kit/SKILL.md` and the + archetype count in `README.md` from each `archetypes//recipe.md` + frontmatter (`name` + `when`), replacing hand-edited rows. +- `recipe.md` frontmatter — `name:` and `when:` (a one-line routing phrase) + are now required at the top of every archetype's `recipe.md`, consumed by + `build-cheatsheet.ts`. +- `evals/` — plugin eval cases (`claude plugin eval`). +- `.github/workflows/ci.yml` — CI: Playwright suite + plugin evals. +- Licence attribution fix — upstream `ThariqS/html-effectiveness` relicensed + MIT → Apache-2.0 on 2026-05-18; the 11 exemplars vendored after that date + (commit `82fa96f`) are Apache-2.0, not MIT. `ATTRIBUTION.md` now splits + vendored exemplars by license and per-file source mapping, and + `LICENSES/Apache-2.0.txt` carries the full license text. + +### Changed + +- `skill/` renamed to `skills/` for plugin auto-discovery. +- All 3 `SKILL.md` files call bundled scripts via + `bun ${CLAUDE_SKILL_DIR}/scripts/ + + + diff --git a/docs/specs/2026-09-16-plugin-packaging.md b/docs/specs/2026-09-16-plugin-packaging.md new file mode 100644 index 0000000..7a67dea --- /dev/null +++ b/docs/specs/2026-09-16-plugin-packaging.md @@ -0,0 +1,154 @@ +# Spec: html-kit as a 2026-standard plugin + +Date: 2026-09-16 · Status: proposed · Base commit: `82fa96f` + +Research notes: +- `~/.search-cache/synth/20260916-2015_claude-skill-plugin-packaging-best-practices.md` +- `~/.search-cache/synth/20260916-2015_html-artifact-skill-comparables.md` + +## 1. Verdict + +The core design is right. The exemplar + recipe pairs, the dated gotchas in each recipe, the searchable gallery of past renders and the annotation loop are all things no comparable repo has together. + +The packaging is behind 2026 norms. html-kit is a symlinked skill bundle, not a plugin. It cannot be installed with `/plugin`, has no evals and no CI, and its script calls only work when the working directory is the html-kit repo. + +## 2. Findings (verified 2026-09-16) + +### 2.1 Where we diverge from best practice + +| # | Finding | Evidence | Rule / source | +|---|---------|----------|---------------| +| F1 | No plugin manifest or marketplace | no `.claude-plugin/` | plugins-reference: `.claude-plugin/plugin.json` + `marketplace.json`; component dirs at plugin root | +| F2 | Skills live in `skill/`, not `skills/` | repo tree | plugin auto-discovery expects `skills//SKILL.md` | +| F3 | Script calls depend on cwd | `skill/html-kit/SKILL.md:56,63,128`: `bun scripts/resolve-out-path.ts`; drain skill uses `` and `/Users/.../` placeholders | skills doc: use `${CLAUDE_SKILL_DIR}` in body and in `allowed-tools` so bundled scripts run without a prompt | +| F4 | No `allowed-tools` | SKILL.md frontmatter has only `name` and `description` | same as F3 | +| F5 | Cheatsheet has drifted | 28 cheatsheet rows vs 30 archetype dirs; README says 19 | hand-maintained index | +| F6 | Personal defaults are hardcoded | `scripts/discover-roots.ts:32`, `scripts/build-index.ts:317,324` assume one user's dev folder | no hardcoded or user-specific paths | +| F7 | Client and project names in public recipe gotchas | `archetypes/{annotated-flowchart,explainer,component-variants,clickable-flow}/recipe.md` name real clients, vendors and internal projects | open-source hygiene | +| F8 | No evals, no CI | no `evals/`, no `.github/workflows/` | `claude plugin eval` (6 grader types, `--threshold` CI gate) | +| F9 | Daemon started manually; annotations need a manual `/html-kit-drain` | README; SSE emits only `index:updated` (`scripts/html-kit-daemon.ts:162,169`) | plugin `monitors/monitors.json`: a session-lifetime command whose stdout lines reach Claude as notifications | +| F10 | Self-containment check is written as an instruction, not enforced | SKILL.md step 7 asks the agent to run `render-audit.ts` | hooks over prose: a `PostToolUse` hook is deterministic | + +### 2.2 Already compliant + +- `SKILL.md` is 151 lines, under the 500-line limit. +- The description is written in third person and includes trigger phrases. +- Deep detail lives one level down, in `references/`. +- Scripts exit non-zero with stderr messages instead of passing errors to the model. +- The skill-dir symlinks (`skill/html-kit/archetypes -> ../../archetypes`) resolve inside the plugin root. Per plugins-reference, those symlinks are **preserved** when the plugin is copied into the cache. The current layout survives plugin packaging. + +## 3. Comparables (verified stars / last push) + +| Repo | Stars | Has | Lacks | Take from it | +|------|-------|-----|-------|--------------| +| [plannotator/effective-html](https://github.com/plannotator/effective-html) | 3.2k · 09-14 | Same source (Thariq's gallery); 6 skills; `.claude-plugin` + `.codex-plugin` + `skills.sh.json` | No gallery index, no feedback loop | Its multi-harness packaging layout | +| [backnotprop/plannotator](https://github.com/backnotprop/plannotator) | 8.7k · 09-16 | Browser annotation that feeds back into the agent loop | Scoped to plans and diffs only | Feedback payload shape; hook-driven round trip | +| [nicobailon/visual-explainer](https://github.com/nicobailon/visual-explainer) | 9.8k · 08-28 | `plugins/` marketplace layout; 11 themes that can be swapped at runtime | No archetype method, no index | Theme presets (`tokens/.css`) | +| [dogum/html-artifacts](https://github.com/dogum/html-artifacts) | 147 · 09-14 | plugin.json; CI-built `.skill` release; gallery issue template | Static examples only | Issue template for community archetype submissions | +| [subsy/skill-cabinet](https://github.com/subsy/skill-cabinet) | 413 · 09-02 | Local daemon + browser catalog | Catalogs skills, not outputs | Daemon lifecycle UX | +| [anthropics/skills](https://github.com/anthropics/skills) | 177k · 09-10 | `theme-factory`, `frontend-design`, `web-artifacts-builder` | Not archetype-based | One theme showcase page | + +**What sets html-kit apart:** the other projects cover at most two of these three: (a) an archetype library with recipes, (b) a searchable history of past artifacts, (c) annotation feedback into the agent. html-kit already has all three. The job is packaging and wiring, not new capability. + +## 4. Target layout + +``` +html-kit/ +├── .claude-plugin/ +│ ├── plugin.json # name, version, description, author, license, keywords +│ └── marketplace.json # self-hosted marketplace, single plugin, source "./" +├── skills/ # renamed from skill/ +│ ├── html-kit/ # SKILL.md + references/ + symlinks (kept) +│ ├── html-kit-drain/ +│ └── html-kit-add-to-patterns/ +├── monitors/monitors.json # P2 +├── hooks/hooks.json # P3: render-audit on Write +├── evals/ # P3 +├── archetypes/ patterns/ tokens/ assets/ scripts/ # unchanged, shared +├── skills.sh.json # P4: npx skills grouping +└── install.sh # kept: dev / non-plugin install +``` + +## 5. Plan + +### P0 — Hygiene before any public push (≈1 h) + +1. Replace client and project names in recipe gotchas with generic descriptions ("a fintech money-flow map", "an email campaign preview"). Keep the lesson, drop the identity. (F7) +2. Remove the user-specific scan-root default: an empty `scanRoots` means no scan until configured; generalize the project-name stripping at `build-index.ts:324`. (F6) +3. Fix recipe formatting: `clickable-flow` line 21; order the `annotated-flowchart` entries by date, above the footer. +4. Commit the pending daemon and index fixes, then the recipe notes. + +**Done when:** a grep for the known client, project and username strings over tracked files returns nothing. + +### P1 — Plugin packaging (≈3 h) + +1. `git mv skill skills`; update `install.sh` to glob `skills/*/SKILL.md`. (F2) +2. Add `.claude-plugin/plugin.json` and `marketplace.json`. Version from CHANGELOG (bump to `0.5.0`). (F1) +3. In all 3 SKILL.md files, replace `bun scripts/X.ts` with `bun "${CLAUDE_SKILL_DIR}/scripts/X.ts"`. Add `allowed-tools: Bash(bun ${CLAUDE_SKILL_DIR}/scripts/*)`. Remove the `` and `/Users/.../` placeholders. (F3, F4) +4. Generate the cheatsheet: give each `recipe.md` a frontmatter `when:` line, and have `scripts/build-cheatsheet.ts` rewrite the table between markers in SKILL.md and the count in the README. (F5) +5. State the runtime requirement: Bun ≥ 1.x. Scripts exit with a clear message if `bun` is missing. + +**Done when:** `claude --plugin-dir .` loads 3 skills; `/plugin marketplace add nyashkn/html-kit` then `/plugin install html-kit@html-kit` renders an artifact from an unrelated repo's cwd without permission prompts for bundled scripts. + +**Check:** confirm `${CLAUDE_SKILL_DIR}/scripts` resolves through the preserved symlink after cache copy (it should, per plugins-reference "within the plugin's own directory"). + +### P2 — Live feedback loop via monitor (≈3 h) — highest value + +1. Daemon: on annotation POST, `broadcastSSE("annotation:new", {artifact, item_id, type, value})`. (F9) +2. `scripts/annotation-monitor.ts`: + - If port 63839 is closed, start the daemon detached (PID lock already prevents duplicates). + - Subscribe to `/events`. + - Print one line per `annotation:new` for artifacts under `$CLAUDE_PROJECT_DIR` or `~/.html-kit/`. + - Format: `html-kit annotation: # : `. +3. `monitors/monitors.json`: + ```json + [{ + "name": "html-kit-annotations", + "command": "bun \"${CLAUDE_PLUGIN_ROOT}/scripts/annotation-monitor.ts\"", + "description": "Annotations left on html-kit artifacts in the browser", + "when": "on-skill-invoke:html-kit" + }] + ``` +4. Keep `/html-kit-drain` as the fallback for non-interactive sessions and other harnesses, where monitors don't run. The drain cursor stays the single source of "already handled". + +**Done when:** render an artifact, click accept or comment in the browser, and Claude receives a notification in the same session without a manual drain. + +**Constraints (from docs):** interactive CLI only; runs unsandboxed at hook trust level; an update needs a session restart; cannot read `${user_config.*}`. + +### P3 — Enforcement + evals + CI (≈3 h) + +1. `hooks/hooks.json`: `PostToolUse` on `Write|Edit`. When the file path matches `.html-kit/*.html`, run `render-audit.ts`. Exit 2 with the missing tokens so the agent must fix them. Remove the prose step 7 audit instruction. (F10) +2. `evals/` — 6 cases: + + | Case | Graders | + |------|---------| + | route-plan: "render this plan as html" | `tool_used: Skill` (html-kit), `file_exists: .html-kit/*.html` | + | route-negative: "summarize this in markdown" | `tool_used` absent Skill html-kit | + | self-contained | `regex` output has no ` bodies dropped, tags kept (so