Expo Skill Eval

Expo Skill Eval

Evaluates skills in plugins/expo/skills/ for trigger accuracy, generated code quality, and/or runtime rendering in Expo Go.

Requirements: macOS with Xcode (iOS simulators), Android SDK with at least one AVD, and bun. No other device tooling is assumed.

Workspace root: /private/tmp/expo-skill-eval-<skill-name>/iteration-N/ (e.g. /private/tmp/expo-skill-eval-expo-ui/iteration-4/).

Before starting — clarify scope

Confirm all of the following up front, before any pipeline work — don't skip any (only skip a given item if the request already states that choice). Batch them into AskUserQuestion calls of ≤4 questions each, in this order:

1. Which skill to eval (if not clear from the request).
2. Prompts — which prompts drive the eval. Built-in prompts (from the skill's eval cases) are pre-selected all; drop any, add a custom text prompt, or build from an uploaded screenshot (a target UI the skill must reproduce). See Prompts below.
3. What to verify — one multi-select of three options: Runtime + screenshots / Trigger accuracy / Code checks (no device). See What to verify below.
4. Expo SDK — latest (default, auto-detected) or a pinned version.
5. Runner — Expo Go (default) or development build.
6. Platforms — iOS / Android / web (always offer all three).
7. Permission flag for claude -p — skip-permissions (default) or accept-edits.
8. Viewer delivery — local only (default) or publish a shareable Artifact.
9. If trigger accuracy is selected — confirm the published expo plugin is disabled (or not installed).

Each is detailed below. Items 4–6 (SDK, runner, platforms) fit naturally in one AskUserQuestion call.

If the skill to eval is not clear from the request, list available skills from plugins/expo/skills/ and ask which one to evaluate.

How the skill under test is loaded — two mechanisms, one per phase (don't pick one globally): executor runs reference it by file path (SKILL_PATH = plugins/expo/skills/<skill>/SKILL.md, read explicitly), while the trigger eval loads it as a plugin (--plugin-dir plugins/expo, so the model can auto-select it from its description). Both point at the local, in-repo version — that's what you're evaluating. You do not need any special flag to launch the harness session itself (the harness finds the skill by repo path); the mechanisms apply to the claude -p subprocesses it spawns. See steps 1 and 3 for why each phase differs. One pre-run check (required when the trigger eval is in scope): if the published expo plugin is installed/enabled, disable it (via /plugin) before launching the harness and re-enable after. A single disable is a global-config change that both this session and the spawned claude -p subprocesses inherit. Why it's required for the trigger eval: that phase loads the local skill via --plugin-dir, and a second installed expo collides with it — the model may trigger the published expo:expo-ui, and since detection only sees the tool-call name you'd silently score the published description instead of your local edits (the collision could also just error). The executor / runtime / static phases are not affected — they read the skill under test by its local SKILL_PATH with no --plugin-dir — so a run with no trigger eval can skip the disable. Disabling expo does not disable expo-skill-eval (a standalone project skill, not part of the expo plugin), so the harness stays available.

Surface this to the user as an explicit up-front confirmation — the same way you confirm which skill to eval. When the trigger eval is in scope, ask the user to confirm the published expo plugin is disabled (or not installed) before you start step 1; if it's still enabled, pause and have them disable it via /plugin. Don't run the trigger eval until they confirm — the harness can't reliably detect installed plugins on its own (reading the global plugin config or claude plugin list would prompt), so this is a manual confirmation, not an auto-check.

Pick the prompts — built-in, custom, or a target screenshot. The prompts are the inputs that drive the executor (with-skill and without-skill); they are separate from what you verify. Confirm them with AskUserQuestion (skip if the request already names a prompt):

• Built-in prompts — representative prompts you generate by reading the skill under test (its SKILL.md + references/) and references/runtime-matrix.md, covering the skill's standard use cases. (If the skill already ships eval cases under evals/evals.json, fold their prompt fields in too — but most skills don't, so you usually derive them.) Pre-select all so the default run exercises the skill's standard cases; let the user deselect any.
• Custom text prompt — a one-off prompt the user types. Don't spend a dedicated option slot on this: AskUserQuestion auto-adds a "Type something" / Other entry, and anything typed there becomes a custom text case.
• Build from an uploaded screenshot — the user gives the path to a target screenshot (a UI to reproduce). The executor is told to open it — claude -p reads PNGs with its Read tool — and build an app matching it; the case records the path as reference_image, and grading compares the generated app to that target (step 6). This is the strongest visual test for a UI skill: "build this."

Respect AskUserQuestion's 4-option-per-question cap with this priority (the bug to avoid: the upload option silently dropped once the four slots fill up):

1. Always reserve a slot for "Build from an uploaded screenshot." It's the whole point of the visual eval and must never be the option that gets dropped.
2. Don't add an explicit "Custom text prompt" option — the auto "Type something" / Other entry already covers it.
3. Fill the remaining ≤3 slots with the built-in/representative prompts, pre-selected. If there are more than 3, collapse them into one pre-selected "All built-in prompts (default)" option and offer subset-picking in a short follow-up, so the upload option still fits.

Present it as a multi-select. When "Build from an uploaded screenshot" is picked, ask for the target image path in a follow-up. Each selected prompt (built-in, typed, or image) becomes one eval case (run with-skill and without-skill).

Always confirm what to verify unless the request makes it unambiguous. Present these options and let the user pick one or more (defaults in bold based on the skill's references/runtime-matrix.md entry):

Read references/runtime-matrix.md to find the skill's default mode before suggesting. If the request already specifies a mode (e.g. "just check if it triggers", "run it on device"), skip the question and proceed.

Pick the Expo SDK version — once, up front. Detect the latest with bash /abs/path/expo-skill-eval/scripts/latest-sdk.sh (it prints the major, e.g. 56; internally it uses bun to run npm view expo dist-tags --json and read the major via JSON.parse/semver, and it's covered by the bash-scripts rule — so don't run the registry query inline yourself, which would prompt). Then confirm with AskUserQuestion: default to that latest SDK, or let the user pin an older one (e.g. to reproduce a version-specific issue). Use the chosen version everywhere the fixture is built — pass it as the <sdk> arg to make-fixture.sh and write it into each eval case's runtime.sdk. If the request already names a version ("eval on SDK 54"), skip detection and use it.

Default to the latest — it stays compatible with the Expo Go that expo start installs on the device. Pinning an SDK older than the device's installed Expo Go makes expo start try to prompt "Install the recommended Expo Go version?"; with no TTY (the snapshot scripts read stdin from /dev/null) it dies with Input is required, but 'npx expo' is in non-interactive mode and every snapshot fails. So only pin an older SDK when you also pre-install a matching Expo Go on the simulator/emulator — otherwise stick with latest.

Pick the runner — Expo Go (default) or a development build. Ask with AskUserQuestion (skip if the request already says which):

• Expo Go (default) — the snapshot scripts run the app with expo start --ios / expo start --android as-is. Fast (no native compile), and it runs anything Expo Go bundles (including @expo/ui on SDK 56+). Cannot run custom native code (expo-modules, config plugins, native deps not in Expo Go).
• Development build — the snapshot scripts run expo run:ios / expo run:android instead, compiling a native dev client per fixture. Use this for skills whose output needs custom native code (the cases that would otherwise be static-only). Much slower — expo run prebuilds and natively compiles each fixture (minutes, especially the first), and needs the full iOS/Android build toolchain — so only choose it when the skill actually requires native code. Disk-heavy: each fixture's native build is multi-GB. The snapshot phase runs clean-fixture.sh after each fixture to keep peak usage to ~one build, but still prefer fewer eval cases and a single platform for dev-build runs, and keep a few GB free. clean-fixture.sh removes the per-fixture build output (node_modules, ios, android, .expo, dist, and the fixture's iOS DerivedData) and keeps the app source + git. The lever for dev-build disk is fewer eval cases + one platform — it only reclaims per-fixture build output and never touches shared dependency caches, so nothing gets re-downloaded.

Pass the choice to the snapshot scripts via the EXPO_SKILL_EVAL_RUNNER env var (expo-go default, or dev-build), and reflect it in each eval case's runtime.mode (expo-go or dev-build). See step 5.

Pick the platforms — always ask, regardless of skill. Offer iOS / Android / web (multi-select) with AskUserQuestion; default to iOS + Android, but always present web as an option — don't pre-filter by skill. Web is a valid choice for most skills: @expo/ui's universal components (Host, Row, Column, Button, List, …) render on web, as do expo-dom, NativeWind/Tailwind, API routes, and plain React Native. The only thing that won't show on web is a platform-specific native tree (@expo/ui/swift-ui or @expo/ui/jetpack-compose), which renders blank there — and that blank is itself a useful signal, so it's still the user's call. Web runs via snapshot-web.sh (expo start --web + Playwright/Chromium) regardless of the runner (expo run is native-only; there's no web dev build), and it's the least-exercised path. Write the chosen set into each eval case's runtime.platforms and have run_snapshots.py loop them.

Confirm how claude -p subprocesses run — once, before starting. Ask with AskUserQuestion whether they may run with --dangerously-skip-permissions, then apply the same answer to every subprocess this run (never re-prompt mid-run):

• Skip permissions (recommended) — pass --dangerously-skip-permissions. Each subprocess runs unattended inside a throwaway fixture under /private/tmp/expo-skill-eval-* and can write files and run setup commands without prompting.
• Accept edits only — pass --permission-mode acceptEdits instead. Bash/installs are auto-denied (no TTY), so some evals may produce partial output.

A bare claude -p with neither flag can't write files at all. If the request already states a preference ("skip permissions", "don't use the dangerous flag"), skip the question.

Confirm how to deliver the results viewer — once, up front. Publishing to claude.ai is outward-facing, so never do it mid-run by surprise; ask in the same up-front AskUserQuestion (alongside the permission flag):

• Local only (default) — generate_viewer.py writes viewer.html and opens it in the local browser. Nothing leaves the machine.
• Publish a shareable Artifact — additionally render the viewer to a claude.ai Artifact (a default-private web page the user can share with teammates) at the very end. Only do this if the user opts in here.

If the request already says whether to share/publish, skip the question. See the Viewer section for the publish mechanics.

Eval case schema

You generate the run's eval cases — one per chosen prompt — and write them to <workspace>/iteration-N/evals.json (the viewer reads them from there). Each case extends the standard skill-creator eval-case shape with a runtime block and visual expectations:

{
  "id": 1,
  "prompt": "Build me a settings screen with a dark mode toggle and a list of options",
  "expected_output": "Working Expo Router screen",
  "expectations": [
    "Uses Expo Router file-based routing",
    "TypeScript compiles with no errors"
  ],
  "runtime": {
    "mode": "expo-go",
    "platforms": ["ios", "android"],
    "sdk": "56"
  },
  "visual_expectations": [
    "No red error screen or Expo Go error overlay on any platform",
    "A settings screen with a visible toggle control is rendered"
  ]
}

• runtime.mode: how the eval runs after the static gate —

  • "expo-go": run in Expo Go (expo start --<platform>) and screenshot. Fast, JS-only. Default.
  • "dev-build": build a native dev client (expo run:<platform>) and screenshot. For skills whose output uses custom native code; much slower (native compile per fixture).
  • "static-only": stop after the static gate — for skills that produce no UI, or when you don't want to run a device at all (CI).

  Consult references/runtime-matrix.md for which repo skills support which mode. (dev-build lets you actually run skills that previously had to be static-only for needing native code.)

• runtime.platforms: subset of ios, android, web — chosen up front (always offered, not gated on the skill; see Before starting). Defaults to ["ios", "android"].

• runtime.sdk: Expo SDK major for the fixture app — set it to the version chosen up front (see Before starting — clarify scope). Omit to use the latest template.

• reference_image (optional — image prompt): absolute path to a target screenshot the skill must reproduce. When set, the executor is told to open it (via its Read tool) and build a matching app, and the grader scores how closely the generated app reproduces it (step 6) on top of the usual expectations. Set in the Prompts phase via "build from an uploaded screenshot."

An image-prompt case is a normal case with reference_image set; enable "Runtime + screenshots" so the harness captures the result to compare against the target:

{
  "prompt": "Build an app whose UI matches the attached reference screenshot.",
  "reference_image": "/abs/path/to/target.png",
  "runtime": { "mode": "expo-go", "platforms": ["ios"], "sdk": "56" },
  "visual_expectations": ["Matches the reference's layout, components, and color treatment"]
}

Pipeline per eval case

Orchestration model — on the main thread you run python3 <orchestrator> and almost nothing else. Every phase is driven by a small Python orchestrator you Write into the workspace and run with python3 /private/tmp/expo-skill-eval-<skill>/<phase>.py (covered by the python3 rule). The orchestrators are the only place the scripts/*.sh files are invoked — always via subprocess.run(["bash", "<scripts>/<name>.sh", …]), which runs as a child of python3 and needs no rule of its own — and the only place parallelism, logging, and directory creation live. So on the main thread you only ever: Write orchestrators, run them with python3, inspect outputs with the Read/Glob/Grep tools, and spawn the grader subagent. Never put a command inside a chained/backgrounded/piped shell construct, and never run ad-hoc mkdir/ls/cat/tail/echo — that is what prompts. (A single standalone bash …/scripts/<name>.sh … is fine for one-off manual debugging, e.g. re-running one flaky snapshot, but the pipeline itself goes through the orchestrators.) Run each orchestrator in the foreground — let the tool call block until it finishes; the orchestrators already parallelize within a phase, so you don't need to overlap phases. Do not shell-background a phase with … & echo "$!" / wait (the &, echo, and wait segments have no rule and prompt). If you genuinely must run a phase while continuing other work, use the Bash tool's run_in_background parameter on a plain python3 <orchestrator> 2>&1 | tee <ws>/…log call — never hand-rolled shell &. Expect exactly one permission prompt at the very start: the first Write into the workspace. allowed-tools can suppress Bash/Read but not Write/Edit, so choose "allow all edits in this directory for the session" on that first prompt — it covers every orchestrator, evals.json, and viewer file for the whole run.

0. Workspace setup

Create the run's directory tree once, with the workspace script — never with ad-hoc mkdir (a raw mkdir prompts: there is no mkdir rule, and a "$WORKSPACE/…" variable can't match a path glob anyway):

bash /abs/path/expo-skill-eval/scripts/make-workspace.sh /private/tmp/expo-skill-eval-<skill> iteration-N <num-evals>

This creates trigger-evals/scratch and iteration-N/eval-<i>/{with_skill,without_skill}/outputs for every eval. It is covered by Bash(bash *expo-skill-eval/scripts/*), and the mkdirs inside run as children of the script (no rule of their own). After this, every other directory is made by the scripts/orchestrators that need it (make-fixture.sh, the executor orchestrator's os.makedirs, the snapshot scripts) or by the Write tool auto-creating parents — so you never need another mkdir.

1. Trigger eval (should-trigger only)

Write a run_trigger_eval_real.py script under the workspace's trigger-evals/ directory. Use only "should_trigger": true queries — the expo plugin is a family of complementary skills, so multiple skills triggering on the same prompt is not a failure. Measure recall only: realistic prompts that should use the skill, scored by trigger rate.

The script should run claude -p <query> per query (with --output-format=stream-json --verbose --include-partial-messages, CLAUDECODE stripped from the env, and the permission flag confirmed up front in Before starting — clarify scope) and detect whether the target skill was triggered by watching for its Skill or Read tool call in the stream. Note: --include-partial-messages requires both --output-format=stream-json and --verbose — omitting either causes an immediate CLI error.

Load the skill under test — pass --plugin-dir to every trigger subprocess. The trigger eval measures whether the skill's description makes the model reach for it, so the subprocess must have the local skill (the version with your edits) loaded. A claude -p subprocess does not inherit the parent session's --plugin-dir, so add it explicitly: --plugin-dir <plugin-root>, where <plugin-root> is the absolute path to the plugin directory that owns the skill — the plugins/expo ancestor containing .claude-plugin/plugin.json (e.g. --plugin-dir /Users/.../skills/plugins/expo). It must be absolute: the subprocess runs from the throwaway scratch/ cwd, so a relative plugins/expo won't resolve — and a missing plugin dir silently loads nothing, which masquerades as a 0% trigger rate. Then watch for the skill triggering under its plugin-qualified name (<plugin>:<skill>, e.g. expo:expo-ui). Two caveats: (1) if the published expo plugin is also installed globally, disable it (via /plugin) for the run and re-enable after — otherwise two copies of expo collide in the subprocess and the model may trigger the published expo:expo-ui, silently scoring its description instead of your local edits (trigger detection only sees the tool-call name, so it can't tell the copies apart; dev checkouts usually don't have it installed). (2) Never make a synthetic duplicate of the skill — a real loaded copy always wins, so the synthetic harness scores 0%. (Executors are unaffected by an installed plugin: they read the local SKILL_PATH directly and pass no --plugin-dir.)

Run each query's subprocess from an empty throwaway cwd (e.g. trigger-evals/scratch/), not the repo root. A should-trigger prompt like "build me a settings screen" can make the subprocess write files, and with --dangerously-skip-permissions those writes would otherwise land in the skills repo. Trigger detection only needs the skill's Skill/Read call to appear in the stream — it doesn't need a fixture — so any incidental writes are throwaway.

Set a per-query subprocess timeout of at least 300 seconds. A 180s limit is too short — some queries cause the model to start generating code before triggering the skill, which pushes total runtime past 3 minutes.

Run trigger evals once per skill, not per code eval case.

2. Fixture

Each executor run gets a fresh Expo app, created by scripts/make-fixture.sh <app-path> <sdk> [clean|full]:

scripts/make-fixture.sh <workspace>/iteration-N/eval-X/<config>/app <sdk>          # blank app (default)
scripts/make-fixture.sh <workspace>/iteration-N/eval-X/<config>/app <sdk> full     # keep example tabs

The script creates the app with bunx create-expo-app -t default@sdk-<version> (or the latest template when no version is given) once per SDK version + variant, caches it under ~/.cache/expo-skill-eval/fixtures/, and clones the cache with APFS copy-on-write — so the first run per variant pays the install cost and every later run is near-instant. The default clean variant runs the template's reset-project script, so executors start from a blank app and every screen in the output is theirs — a much cleaner grading signal. Use full only when the eval prompt assumes an existing app (e.g. "I have an app with two tabs..."). The script also resets git inside the clone, so git diff in the app shows exactly what the executor changed (useful evidence for the grader).

Build fixtures sequentially, then fan out executors — never create fixtures concurrently. make-fixture.sh shares a cache under ~/.cache/expo-skill-eval/fixtures/ keyed by SDK+variant. If two runs both find the cache cold and call bunx create-expo-app at the same time, bun's link step collides and one fails with EEXIST / "could not determine executable to run for package create-expo-app". So in the executor orchestrator (step 3), create all fixtures one at a time first — a plain Python loop calling subprocess.run(["bash", "<scripts>/make-fixture.sh", app, sdk, variant]) (where sdk is the version chosen up front) — then fan out the claude -p executors with a ThreadPoolExecutor. Sequential creation is cheap: only the first fixture per SDK+variant pays the install cost; the rest are ~1s APFS clones. (And never fan fixtures out with ad-hoc shell like make-fixture.sh A & make-fixture.sh B & wait — the &/wait segments prompt; the sequential Python loop avoids both the race and the prompt.)

3. Generate (executor subagents)

Run executors as claude -p subprocess calls from a Python script, not via the Agent tool. The Agent tool spawns subagents with their own permission context — file edits inside the fixture app will prompt the user. A claude -p subprocess is a separate process outside the permission system entirely (the same pattern the trigger eval harness uses).

Write a Python script to /private/tmp/expo-skill-eval-<skill>/run_executors.py. First create every run's fixture in a sequential loop — subprocess.run(["bash", "<scripts>/make-fixture.sh", app, sdk, variant], …) one at a time (concurrent creation races the shared bun cache — see step 2). Then run the with-skill and without-skill claude -p calls in parallel via a ThreadPoolExecutor. Both phases live inside Python (covered by the python3 rule), so nothing runs as ad-hoc shell on the main thread. Each executor prompt must include:

• The skill path (with-skill runs only) and the eval prompt.
• Image-prompt cases (reference_image set): the absolute path to the target screenshot plus an instruction like "Open the reference screenshot at <path> with your Read tool and build an app whose UI matches it as closely as you can — layout, components, spacing, and colors." (claude -p renders PNGs read this way, so the executor can actually see the target.)
• The fixture app path: "Make your changes inside <app-path>. The project already exists and has dependencies installed. Use absolute paths for all file operations."
• "Before writing any files, inspect the project layout — run ls, read package.json and app.json — to find the correct routes directory. Recent SDK default templates place Expo Router routes in src/app/; older ones use app/ at the project root — inspect to confirm which this fixture uses."
• "Do NOT start the dev server, boot simulators, or take screenshots — the harness does that after you finish."
• Where to save a short summary of what was built.

Flags for the claude -p subprocess:

• Strip CLAUDECODE from the environment (env = {k: v for k, v in os.environ.items() if k != "CLAUDECODE"}) — otherwise claude -p hangs silently when nested inside a running Claude Code session.
• A permission flag, confirmed with the user up front (see Before starting — clarify scope): either --dangerously-skip-permissions or --permission-mode acceptEdits. Bake the chosen flag into the generated script. A bare claude -p with neither flag can't write files — it has no TTY to approve the edit and emits code as text instead.
• Do NOT pass --plugin-dir to executors (unlike the trigger eval). The with-skill run already reads the skill by its absolute SKILL_PATH, so it tests the local content directly; and the without-skill run must have no skill available at all — loading the plugin would let the skill auto-trigger and contaminate the baseline. Keeping executors path-based also cleanly separates the two questions: the executor measures content quality (is the skill useful once read?), the trigger eval measures triggering (does the description get it picked?).

Capture stdout/stderr per run to a log file next to the fixture for grading evidence. Set timeout to 900s per executor — with-skill runs read multiple reference files before coding and regularly take 5–10 minutes.

4. Static gate

Write run_static.py and run it with python3. For each eval/config app it calls subprocess.run(["bash", "<scripts>/check-static.sh", app, "ios,android"], capture_output=True, …) across a ThreadPoolExecutor (static gates are independent — run them concurrently inside Python, never with shell &/wait), and writes each result to eval-<i>/<config>/static.json (exit code + captured output) for the grader.

check-static.sh runs tsc --noEmit, expo lint, and expo export for the listed platforms. A passing export catches most import/syntax/missing-module failures without touching a device; a failing export short-circuits step 5 with a clean FAIL — record it and have the snapshot orchestrator skip that app.

5. Run + screenshot (serial across evals)

Write run_snapshots.py and run it with python3. Simulators and emulators are shared resources, so this orchestrator runs serially (no thread pool): for each app that passed the static gate, and each platform, it os.makedirs the outputs/ dir and calls subprocess.run(["bash", "<scripts>/snapshot-<platform>.sh", app, f"{outputs}/<platform>.png", port], env={**os.environ, "EXPO_SKILL_EVAL_RUNNER": runner}, …). Pass the port as a positional argument: use 8081 for iOS and 8082 for Android — expo run:ios/android --port N is supported and using separate ports lets you run both platforms without port collisions if you ever parallelize. Screenshots land in the run's outputs/ directory so the viewer renders them inline.

Reclaim disk after each fixture — essential for dev-build runs. Once all selected platforms for an app are captured (and before the next fixture builds), call subprocess.run(["bash", "<scripts>/clean-fixture.sh", app]). Each expo run:<platform> leaves multi-GB native build output (iOS Pods + DerivedData, Android Gradle build); without this, evals × configs × iterations pile up and fill the disk mid-run (the instability you'll see is the disk filling). clean-fixture.sh removes the heavy regenerable dirs (node_modules, ios, android, .expo, dist) and the fixture's iOS DerivedData, keeping the app source + git so the grader's git diff still works. With serial snapshots + per-fixture cleanup, peak disk stays at ~one fixture's build instead of all of them. (Harmless for expo-go runs too — they just have little to reclaim.)

runner is the up-front choice (expo-go default, or dev-build). The snapshot scripts honor EXPO_SKILL_EVAL_RUNNER: expo-go launches with expo start --<platform> (and the Expo Go install/deep-link dance); dev-build launches with expo run:<platform> --port <port>, which compiles+installs a native dev client and skips the Expo Go steps. The scripts already default the dev-build timeout to 900s, but bump EXPO_SKILL_EVAL_BUNDLE_TIMEOUT higher if the first native compile needs it. make-fixture.sh pre-installs expo-dev-client in every fixture so the dev-client URL scheme is registered before expo run tries to deep-link the app open.

Snapshot scripts always capture the initial route /. They open the app via a deep link and take one screenshot — they cannot tap or navigate. Design eval prompts so the feature under test renders at the root route. If the executor places the main UI behind a navigation action (e.g. an "Open Settings" button on the index), the snapshot will miss the feature entirely and all visual expectations will fail.

Each snapshot-<platform>.sh frees its Metro port on startup (kills any stale process left on it by a crashed prior run) and tears Metro down on exit — so you never need to run lsof/kill/pkill yourself to clear ports (that would prompt, and it's already handled). It then starts Metro, waits for the "Bundled" line in the Metro log, settles, captures a screenshot, and tears Metro down. iOS boots the newest available iPhone simulator if none is booted; Android boots the first AVD if no device is attached (the slow path — boot once and reuse across the whole iteration). Android first recycles a wedged/offline emulator (graceful adb emu kill, then force-kill + adb reset) so a half-dead instance can't poison the run, and boots with hardware GPU (-gpu host, Metal-accelerated on Apple Silicon). If host self-aborts the emulator on a given machine (qemu SIGABRT deep in gfxstream/Metal — possible on Apple Silicon under load), edit GPU_MODE in snapshot-android.sh to a software mode (guest renders reliably but slowly — bump the settle; avoid swiftshader_indirect, which hangs at boot on arm64). snapshot-web.sh runs only when platforms includes web. Each script writes a Metro log next to the screenshot (<name>.metro.log) — include it in the grader's inputs. If a script exits non-zero it still attempts a best-effort screenshot (an error screen is evidence too). dev-build relaunch: after Metro is up, the scripts relaunch the app via xcrun simctl launch (iOS) and adb shell am start -n <pkg>/.MainActivity (Android) — both avoid the "Open in X?" system dialog that a URL-scheme deep link triggers on first launch.

After all screenshots for the iteration are captured, always generate the viewer — pass the workspace root to the checked-in script:

python3 /abs/path/expo-skill-eval/scripts/generate_viewer.py /private/tmp/expo-skill-eval-<skill>

It writes viewer.html into the workspace root (one level above iteration-N/) and opens it in the browser itself (via webbrowser.open) — so no separate open command (and no Bash(open:*) rule) is needed. See the Viewer section below.

6. Grade

Spawn a grader subagent in the foreground. Its prompt must include:

• The eval prompt, expectations list, and visual_expectations from the eval case.
• The instructions in agents/visual-grader.md (screenshot grading, redbox detection).
• The screenshot files, Metro logs, and the step-4 static.json as inputs.
• Image-prompt cases (case has reference_image): also include the target screenshot (reference_image), references/design-rubric.md, and the fixture's git diff. Tell the grader to compare the generated screenshot(s) to the target and emit the reference_match + quality blocks below.

The grader writes grading.json next to the outputs with this shape:

{
  "score": 8.5,
  "max_score": 9,
  "expectations": [
    {"text": "...", "passed": true, "evidence": "..."}
  ],
  "reference_match": {
    "score": 7, "max": 10,
    "evidence": "ios.png vs target.png: same two-section grouped list + toggle; accent color differs (blue vs target's green); row spacing tighter than target"
  },
  "quality": {
    "dimensions": [
      {"name": "Layout & hierarchy", "score": 2, "max": 3, "evidence": "ios.png: …"}
    ],
    "subtotal": 17,
    "max": 24,
    "summary": "…"
  },
  "user_notes_summary": {"needs_review": false, "notes": ""}
}

Visual expectations go into the same expectations array with evidence naming the screenshot file and describing what is visible. The reference_match block (how closely the generated app reproduces the target screenshot) and the quality block (design-rubric scores from references/design-rubric.md) are emitted only for image-prompt cases — or when a quality grade is explicitly requested. Omit both for plain text-prompt runs.

Rollout phases

Build out and debug the pipeline in this order — each phase is independently useful:

1. Static: steps 1–4 only (runtime.mode: "static-only" for everything). No devices needed; CI-friendly.
2. iOS: add snapshot-ios.sh to the loop. simctl is the most scriptable target.
3. Android: add snapshot-android.sh. Emulator boot is the slowest part — keep one emulator running for the whole session.
4. Web: add snapshot-web.sh for skills that target web (uses Playwright via bunx; first run downloads Chromium).

Practical notes

• Temp locations: all eval workspaces go under /private/tmp/expo-skill-eval-<skill-name>/iteration-N/. Everything in this run — Read, Write, Edit, and Bash — is covered by the allowed-tools frontmatter, so a correctly-loaded skill runs prompt-free.
• Permission rule forms (why this skill stays prompt-free): the rule syntax matters and the two tool families behave differently:
  • Bash(...) rules — path-scoped to the skill's own code (no broad interpreters). Bash(python3 /private/tmp/expo-skill-eval-*) (plus the /tmp alias) runs the Python orchestrators you generate under the workspace; Bash(python3 *expo-skill-eval/scripts/*) runs the checked-in scripts/generate_viewer.py; Bash(tee /private/tmp/expo-skill-eval-*) (+ /tmp) lets python3 … 2>&1 | tee <workspace>/…log write a log without prompting; Bash(bash *expo-skill-eval/scripts/*) runs only this skill's scripts/*.sh. Because every path is pinned, the escape hatches stay denied: python3 -c …, bash -c …, tee /etc/…, and running code anywhere else do not match (verified empirically — a scoped rule allows bash <dir>/run.sh but blocks bash -c … and any other path). Commands the scripts call internally — bunx, xcrun simctl, adb, git, mkdir, expo — are children of the script, not Bash tool calls, so they need no rule. Do not run ad-hoc mkdir/ls/find/cat/grep from the main thread (they have no rule and prompt — and a raw mkdir "$WORKSPACE/…" can't match a path glob because the path is an unexpanded variable): create the directory tree with make-workspace.sh (step 0), let orchestrators create their own dirs (os.makedirs), and inspect results with the Read/Glob/Grep tools (no Bash rule needed).
  • Bash rule matching (tested, non-obvious): a Bash rule is a gitignore-style glob over the command string. * matches any run of characters including / and spaces and works mid-pattern — so Bash(python3 /private/tmp/expo-skill-eval-*) matches python3 /private/tmp/expo-skill-eval-x/run.py 2>&1, and Bash(bash *expo-skill-eval/scripts/*) matches bash /any/abs/path/expo-skill-eval/scripts/foo.sh args. Two gotchas that burned earlier attempts: ** is matched literally (never use it in a Bash rule), and the :* suffix only works right after the command token (Bash(python3:*)) — not after a partial path (Bash(python3 /path-:*) does not match). Compound commands split on |, &&, ||, ;, & and each segment needs its own matching rule.
  • Read rules suppress prompts; Write/Edit rules do not. This is a Claude Code asymmetry (not a pattern bug, and not reload — in a session where the Bash/Read rules from this same frontmatter are clearly working, Write still prompts): file creation/editing always goes through Claude Code's edit-approval flow regardless of allowed-tools. The frontmatter still scopes Read/Write/Edit to …/expo-skill-eval-*/** (both the /tmp and /private/tmp forms, since macOS doesn't auto-resolve the symlink) as documentation and a guardrail, but those Write/Edit entries won't silence the prompt on their own. Practical consequence: at the start of a run you get one Write prompt for the workspace — choose "Yes, allow all edits in this directory for the session" and every later orchestrator / evals.json / viewer write under that workspace goes through silently. That single directory approval, not a rule, is what makes file-writing prompt-free.
  • Reload after editing frontmatter — a full restart, not /reload-skills. allowed-tools is read once when the skill loads at session start; /reload-skills reloads the skill body but does not reliably refresh the permission rules. After editing this file, quit Claude Code entirely and start a new session, then re-run the skill — otherwise a stale (cached) ruleset keeps prompting even though the file on

…(truncated)