Move make-widget/run-server skills to root .claude/skills/

Nested .claude/skills/ dirs (previously under server/) are only
auto-discovered on-demand once a file under that subdirectory is
touched, so /make-widget and /run-server weren't invocable from a
fresh session. Root .claude/skills/ is scanned at session start.

Fixes setup.sh/start-server.sh's relative cd-depth math (was hardcoded
for the old server/.claude/skills/run-server/ depth) to instead
resolve the repo root via git and cd into server/ explicitly, and
updates SKILL.md/driver.py's path references to match the new layout.

.claude/skills/make-widget/SKILL.md

@@ -0,0 +1,178 @@
+---
+name: make-widget
+description: Scaffold a new widget type for the espresso_frame server (the ~13-file checklist a widget type touches -- config table, grid footprint, render module, registry, migration, config-save + endpoints, dialog template + JS, script tag, WIDGET_LABELS, docs, tests). Use when asked to add a new widget type to a frame's panel (e.g. "add a text widget", "add an RSS widget", "add a weather-only widget").
+---
+
+Adding a widget type is a very consistent, repeated pattern in this
+codebase (`photos`/`calendar`/`whiteboard`/`tasks`/`static`) -- see
+`docs/widgets.md` for the system's actual data model/rendering/dialog
+architecture (read that first if you haven't). This skill is the
+checklist of every file that pattern touches, so nothing gets silently
+dropped (the static-image widget shipped without a `docs/widgets.md`
+update; this skill exists so that doesn't keep happening).
+
+**For a complete worked example touching every item below**, `git show
+35e80c6 --stat` (the static-image widget's commit) in this repo.
+
+All paths below are relative to `server/`.
+
+## Before writing any code: shape decisions
+
+Answer these first -- they determine which existing widget type is the
+closest template to copy from:
+
+- **Live upstream to poll, or self-contained/user-authored?** Calendar/
+  whiteboard/tasks fetch from somewhere external on a throttle
+  (`checked_at` + `get_or_refresh_*` in `routers/common.py`). Photos'
+  queue and the static image widget don't -- their content is set once
+  via the dialog (an upload, a pick) and just sits there until changed.
+  A text widget is almost certainly this second shape.
+- **Single source, or multi-source merge?** Calendar/tasks merge
+  several *people's* data (`FrameCalendar`/`FrameTaskList`, owner-adds/
+  anyone-mutes). Only reach for that shape if the new type genuinely
+  needs to combine several linked users' own data -- most new widget
+  types are single-owner/single-config and don't need it.
+- **Any button actions**, or is `ACTIONS = {}` correct (nothing to
+  advance/back/force)? Tasks and static image are both `{}`.
+- **Minimum sane grid footprint** -- how small can this widget be
+  before its content is illegible/pointless?
+
+Pick your template accordingly:
+
+| New widget shape | Copy from |
+|---|---|
+| Self-contained, user-authored/uploaded, no fetch, no actions | `app/widgets/static_image.py` |
+| Single external source, throttled fetch, one "check_now" action | `app/widgets/whiteboard.py` |
+| Multi-source merge, owner-adds/anyone-mutes permissions | `app/widgets/tasks.py` (simpler) or `calendar.py` (also has size-tier rendering) |
+| Stateful queue/rotation with advance/back | `app/widgets/photos.py` |
+
+## The checklist
+
+1. **`app/models.py`** -- new `<Type>WidgetConfig` table, `widget_id`
+   `Mapped[int]` primary key `ForeignKey("widgets.id", ondelete="CASCADE")`,
+   plus whatever fields the type needs. Add it to the `WIDGET_CONFIG_MODELS`
+   dict at the bottom of the file.
+2. **`app/grid.py`** -- add an entry to `MIN_FOOTPRINT`.
+3. **`app/widgets/<type>.py`** -- new module exposing:
+   - `render(db, frame, widget, target_w, target_h, is_normal_wake=True) -> Image.Image`
+     -- RGB, exactly `target_w x target_h`, **never raises** for a
+     foreseeable failure (missing config, fetch error) -- fall back to
+     `._shared.placeholder_image(target_w, target_h, [lines])` instead.
+   - `ACTIONS: dict[str, Callable[[Session, Frame, Widget], None]]`
+   - `ACTION_LABELS: dict[str, str]`
+4. **`app/widgets/__init__.py`** -- import the new module, add it to
+   `WIDGET_TYPES`.
+5. **`app/migration.py`** -- new `_migration_N`. A brand-new table with
+   no legacy data to carry forward is just
+   `Base.metadata.create_all(bind=conn)` (see `_migration_20`) -- it
+   only creates the one new table, existing ones are untouched. Register
+   `(N, _migration_N)` as the new last entry in `MIGRATIONS`.
+6. **`app/routers/api_widgets.py`**:
+   - Add any new `Form(...)` fields to `api_widget_config_save`'s
+     signature, and a new `elif widget.widget_type == "<type>":` branch
+     inside its body. Reuse an existing field name (e.g. `display_mode`)
+     where the semantics genuinely match -- fields are namespaced by
+     which widget type actually reads them, not by name collision, so
+     this is safe (see the comment above `display_mode` in that
+     function).
+   - Add type-specific endpoints as needed (upload/source-select/etc.).
+     Use `require_widget_control` for widget-wide settings a dialog Save
+     button changes; use `require_widget_view` (not control) for the
+     owner-adds/anyone-mutes multi-source pattern, matching
+     `api_widget_calendar_select`/`api_widget_task_list_select`.
+   - Add a `GET .../preview/<type>` endpoint mirroring the others --
+     `render_preview_png` (the full palette/dither pipeline) for
+     image-like content, or a dedicated `render_<type>_preview_png` in a
+     rendering module for text/graphics content (see
+     `calendar_render.render_tasks_preview_png`).
+7. **`app/routers/frame_pages.py`** -- import the new config model, add
+   an `if widget.widget_type == "<type>":` branch in `widget_dialog()`
+   returning `templates.TemplateResponse("_widget_dialog_<type>.html", {...})`.
+8. **`app/templates/_widget_dialog_<type>.html`** -- the dialog
+   fragment: settings card(s) + `<img class="preview-img"
+   id="<type>-preview">` + a refresh button, using the existing
+   `.card`/`.card-title`/`.sub`/`.checkbox-row` classes from
+   `theme.css` rather than inventing new ones.
+9. **`app/static/widget_dialog_<type>.js`** -- an `init<Type>Dialog()`/
+   `close<Type>Dialog()` pair (not a page-load script -- see any
+   existing `widget_dialog_*.js`'s header comment for the contract).
+   `window.fetch` already CSRF-injects (see `common.js`), so POSTs don't
+   need a manual header. **Never build user-supplied text into the DOM
+   via `innerHTML` string interpolation** -- use `textContent`/
+   `createElement` (a filename, a task summary, anything another linked
+   user's account could have set is a stored-XSS vector otherwise).
+10. **`app/templates/frame_layout.html`** -- add
+    `<script src="/static/widget_dialog_<type>.js"></script>` next to
+    the other widget dialog scripts.
+11. **`app/static/frame_layout.js`** -- add the type to both
+    `DIALOG_INIT` and `DIALOG_CLOSE`.
+12. **`app/static/common.js`** -- add a `WIDGET_LABELS` entry (the
+    human label shown in the add-widget button, the widget box, and the
+    button-assignment picker in `frame_config.js`).
+13. **`docs/widgets.md`** -- update every place that enumerates widget
+    types: the intro sentence, the `widget_type` column-value list, the
+    `MIN_FOOTPRINT` prose line, the `app/widgets/` module list. This is
+    the project's own "start here" doc per `CLAUDE.md` -- don't ship a
+    widget without it staying accurate.
+
+## Tests (`server/tests/`)
+
+- `test_widgets_<type>.py` -- unit-level `render()` tests, no HTTP:
+  correct size/mode with no config, with config, at
+  `grid.MIN_FOOTPRINT`'s smallest box, `ACTIONS == {}` if passive.
+  Mirror `test_widgets_static.py` (self-contained) or
+  `test_widgets_tasks.py` (fetch-backed, monkeypatches the fetch call).
+- `test_widget_config_and_queue_endpoints.py` -- add a
+  `_add_<type>_widget` helper plus an HTTP-level
+  `test_config_save_updates_a_<type>_widget` test, and tests for any new
+  endpoints (upload/select/preview: 400 before configured, 200 after,
+  400 for the wrong widget type via `_require_widget_type`).
+- `test_migrations.py` -- add the new table to
+  `test_expected_columns_exist_on_current_schema`'s spot-check
+  (`inspector.get_table_names()` or `inspector.get_columns(...)`).
+- Owner-adds/anyone-mutes multi-source table? Add cases to
+  `test_permission_boundaries.py` following its existing
+  calendar-select/task-list-select pattern (owner can add, non-owner
+  can mute but not add, 404 for an unrelated widget id, 400 for the
+  wrong widget type).
+- Any pure-logic helper module (decoding, parsing -- like
+  `app/image_upload.py`) gets its own `test_<module>.py`: no HTTP, no
+  DB, just the function.
+
+Run the full suite before calling it done:
+
+```bash
+cd server && .venv/bin/pytest -q
+```
+
+Comfortably under 30s for the whole suite (~200+ tests) -- there's no
+reason to skip this or run a subset.
+
+## Browser verification (required, not optional)
+
+Per `CLAUDE.md`, reading the JS is not enough -- this project has
+shipped UI bugs (mobile viewport CSS collapse, a dialog's status message
+landing behind its own backdrop, a JSON/form body mismatch) that only
+showed up live. Use the `run-server` skill:
+
+- Clear existing widgets and add one of the new type
+  (`POST /api/frames/1/widgets`), resize it (`PATCH`), open its dialog
+  (`click .widget-box-settings`), exercise its actual settings/upload
+  flow through the real UI controls (not just a raw `fetch` in `eval` --
+  that only proves the endpoint works, not that the button is wired to
+  it), and check `console-errors` for anything beyond the expected
+  favicon 404.
+- Check the full composited panel preview
+  (`#frame-preview-thumb` on `/frames/{id}/config`) actually shows the
+  new widget's content -- not just its own dialog's `preview/<type>`
+  image, which only proves the render function works in isolation.
+- Screenshot **both** desktop (`viewport 1280 900`) and mobile
+  (the driver's default) widths -- the layout genuinely forks at the
+  860px breakpoint in `theme.css`.
+
+## Commit
+
+One commit for the whole widget (models + migration + render + router +
+UI + tests + docs) -- this project's convention is one feature per
+commit, not split by layer. No `Co-Authored-By: Claude` trailer (see
+root `CLAUDE.md`).

.claude/skills/run-server/.gitignore

@@ -0,0 +1,2 @@
+# Generated by setup.sh -- bakes in this host's /tmp paths, not portable.
+env.sh

.claude/skills/run-server/SKILL.md

@@ -0,0 +1,234 @@
+---
+name: run-server
+description: Build, run, and drive the espresso_frame FastAPI server (server/) -- start it against a scratch DB, browser-test its UI, run its pytest suite. Use when asked to start the server, take a screenshot of a frame page, click through the web UI, or verify a server/UI change actually works.
+---
+
+The espresso_frame server is a FastAPI app (`app.main:app`) with a
+server-rendered Jinja UI. For agent/automated use it's driven by a
+Playwright REPL at `.claude/skills/run-server/driver.py`, run under
+tmux -- `chromium-cli` isn't available in this container, so this
+driver replaces it (same command vocabulary: nav/wait-for/click/fill/
+screenshot/eval/console-errors).
+
+This container ships with **no Python, Node, Docker, or browser, and
+no sudo**. `setup.sh` bootstraps everything non-root; it's the bulk of
+what makes this skill non-obvious. Run it once per fresh container.
+
+Commands below (`setup.sh`, `start-server.sh`, `stop-server.sh`,
+`env.sh`, `driver.py`) are invoked from the **repo root** via their
+`.claude/skills/run-server/` path -- the scripts `cd` into `server/`
+themselves. Anything under `.venv/` (the venv itself, `pytest`,
+`uvicorn`) lives inside `server/`, so those commands need `server/`
+prefixed or `cd server` first.
+
+## Prerequisites
+
+None to install manually -- `setup.sh` does it all without root, using
+only `curl`/`apt-get download`/`dpkg-deb -x` (never `apt-get install`,
+which needs root). It downloads ~450MB total (Python, Chromium, tmux,
+shared libs) on first run.
+
+```bash
+bash .claude/skills/run-server/setup.sh
+```
+
+Re-running is safe and fast -- every step checks whether it already
+happened (venv exists? chromium downloaded? libs extracted? tmux
+present?) before doing any work.
+
+This creates:
+- `.venv/` -- Python 3.12 + `requirements.txt` + `playwright`, via `uv`
+  (a static Rust binary that fetches its own Python -- no compiler
+  needed, install via `curl -LsSf https://astral.sh/uv/install.sh | sh`)
+- `~/.cache/ms-playwright/` -- Chromium (full `chrome` + headless-shell)
+- `/tmp/run-server-chromium-deps/` -- Chromium's + tmux's shared libs
+  and fonts, extracted (not installed) from `.deb` files
+- `.claude/skills/run-server/env.sh` -- the `PATH`/`LD_LIBRARY_PATH`/
+  `FONTCONFIG_PATH`/`RUN_SERVER_CHROME_BIN` exports the driver needs
+  (gitignored -- host-specific `/tmp` paths, regenerated by setup.sh)
+
+## Run (agent path)
+
+**1. Start the server** against a scratch DB (never the real
+deployment's data -- see `CLAUDE.md`):
+
+```bash
+bash .claude/skills/run-server/start-server.sh
+# -> server PID <pid> up on http://127.0.0.1:8420 (log: /tmp/run-server-scratch/server.log)
+```
+
+Optional args: `start-server.sh [scratch-dir] [port]` (defaults
+`/tmp/run-server-scratch` / `8420`).
+
+**2. Drive it**, wrapped in tmux so you can send one command at a time
+and read the response before sending the next. `tmux` itself was
+extracted the same non-root way as Chromium's libs (see Prerequisites)
+and needs `LD_LIBRARY_PATH` set in *this* shell too, not just inside
+the pane -- source `env.sh` before the first `tmux` call:
+
+```bash
+source .claude/skills/run-server/env.sh
+tmux new-session -d -s runserver -x 200 -y 50
+tmux send-keys -t runserver \
+  'source .claude/skills/run-server/env.sh && server/.venv/bin/python .claude/skills/run-server/driver.py' Enter
+timeout 20 bash -c 'until tmux capture-pane -t runserver -p | grep -q "driver>"; do sleep 0.3; done'
+
+# Every fresh scratch DB starts with no users -- bootstrap-admin
+# completes first-run /setup and logs in (see Commands table):
+tmux send-keys -t runserver 'bootstrap-admin' Enter
+timeout 15 bash -c 'until tmux capture-pane -t runserver -p | grep -q "bootstrapped admin"; do sleep 0.3; done'
+
+tmux send-keys -t runserver 'nav /frames/1/config' Enter
+tmux send-keys -t runserver 'wait-for #frame-preview-thumb' Enter
+tmux send-keys -t runserver 'screenshot before' Enter
+tmux capture-pane -t runserver -p
+```
+
+Poll for a specific marker between `send-keys` and `capture-pane`
+(`driver>`, `bootstrapped admin`, `screenshot:`, ...) rather than a
+fixed `sleep` -- it's faster and fails loudly instead of capturing a
+half-rendered screen. Give each poll its own `timeout` (~15-20s); don't
+chain many polling loops inside one shell invocation -- see Gotchas.
+
+Screenshots land in `/tmp/run-server-shots/` (override:
+`SCREENSHOT_DIR`). **Actually look at them** -- a blank or error-page
+screenshot is a failure to launch, not success.
+
+**Test every UI change at both a desktop and a mobile viewport.** The
+driver defaults to a desktop size (1280x900); switch with `viewport`.
+The app's mobile breakpoint is 860px (`theme.css`) -- below that the
+sidebar goes off-canvas behind a hamburger (`.mobile-bar`). A page that
+looks right at 1280px can still overflow, overlap the mobile bar, or
+mis-center a `<dialog>` at phone widths -- screenshot both:
+
+```bash
+tmux send-keys -t runserver 'viewport 1280 900' Enter   # desktop (also the default)
+tmux send-keys -t runserver 'screenshot desktop-x' Enter
+tmux send-keys -t runserver 'viewport 390 844' Enter     # iPhone-ish mobile width
+tmux send-keys -t runserver 'screenshot mobile-x' Enter
+```
+
+### Driver commands
+
+| command | what it does |
+|---|---|
+| `nav <path-or-url>` | navigate (relative paths resolve against `http://127.0.0.1:8420`, override with `RUN_SERVER_BASE_URL`) |
+| `wait-for <selector>` | wait up to 10s for a selector (plain CSS -- see Gotchas for attribute selectors) |
+| `click <selector>` | click an element |
+| `fill <selector> <value>` | fill an input |
+| `press <key>` | keyboard press (e.g. `Enter`) |
+| `screenshot [name]` | → `/tmp/run-server-shots/<name>.png` |
+| `eval <js>` | evaluate JS in the page, prints JSON |
+| `console-errors` | prints all captured console/page errors as a JSON array |
+| `viewport [w] [h]` | resize the viewport, default `390 844` -- use `1280 900` for desktop (see Gotchas re: real touch input) |
+| `is-open <dialog-selector>` | prints `true`/`false` for a `<dialog>` element's `.open` -- use this instead of `wait-for sel[open]` |
+| `bootstrap-admin [user] [pass]` | completes first-run `/setup` (defaults `admin`/`testpassword123`); links frame #1 and logs in |
+| `quit` | closes the browser, exits the driver |
+
+**3. Stop the server** when done:
+
+```bash
+bash .claude/skills/run-server/stop-server.sh
+tmux kill-session -t runserver
+```
+
+## Run (human path)
+
+```bash
+cd server
+DATABASE_URL="sqlite:////tmp/dev.db" CONFIG_PATH="/tmp/dev-config.json" \
+  .venv/bin/uvicorn app.main:app --reload --port 8420
+```
+
+Open `http://localhost:8420` in a real browser. `start.sh` (the Docker
+entrypoint) is not this -- it also launches the Node whiteboard
+render-service sidecar, which needs Node (not installed here) and
+isn't needed for most UI testing.
+
+## Test
+
+```bash
+cd server && .venv/bin/pytest
+```
+
+Uses its own tempfile SQLite per run (`tests/conftest.py`) -- no setup
+needed beyond the venv.
+
+---
+
+## Gotchas
+
+- **`viewport` only resizes the window -- it does not emulate touch
+  input.** `click` still dispatches a mouse click, not a tap; there's
+  no touch-delay, no `:hover`-stickiness-after-tap, no `hasTouch`
+  context. It catches real bugs (layout overflow, off-canvas sidebar,
+  a `<dialog>` mis-centering at phone widths) but won't catch anything
+  that's specifically a touch-vs-mouse event difference. Good enough
+  for CSS/layout verification; not a substitute for testing on an
+  actual phone if a change touches touch-specific interaction.
+
+- **`chrome-headless-shell` (Playwright's default headless target)
+  crashes on basic calls in this container**, e.g. `page.set_content()`
+  returns `TargetClosedError`, even after every `ldd`-reported missing
+  library is resolved. The full `chrome` binary (`chromium-*/chrome-linux64/chrome`)
+  + `--no-sandbox` is stable; `driver.py` and `setup.sh` both use it,
+  not the headless-shell default.
+
+- **Missing fonts silently break `fill()`, not just rendering.** Before
+  `fontconfig`/`libfontconfig1` were extracted and `fonts.conf` pointed
+  at the extracted font dir, `page.fill()` ran with no error but left
+  inputs empty (`input_value()` returned `""`), and all text rendered
+  invisible in screenshots. It looks like a scripting bug, not a
+  missing-lib problem -- if `fill` silently no-ops, suspect fonts
+  before suspecting the selector or a race.
+
+- **No `apt-get install` / `playwright install-deps` (no root) and no
+  `apt-get update` into the real `/var/lib/apt/lists` (root-owned).**
+  Worked around by redirecting apt's state dirs to a scratch,
+  user-writable path (`-o Dir::State::Lists=... -o Dir::Cache=...`),
+  which makes plain `update` and `install --download-only --print-uris`
+  work as a non-root user; then `dpkg-deb -x <deb> <root>` (extract,
+  not install) needs no root either. `setup.sh` does this for
+  Chromium's deps *and* for `tmux` itself, which also isn't
+  preinstalled.
+
+- **`wait-for` with an attribute selector like `#frame-preview-dialog[open]`
+  is unreliable through `tmux send-keys`** -- shell/tmux escaping of
+  `[`/`]` easily mangles it (seen: a real 10s Playwright timeout from a
+  garbled selector, not a fast failure). Use the app-specific `is-open
+  <selector>` command instead of `wait-for sel[open]` to check a
+  `<dialog>`'s open state.
+
+- **Don't chain many `tmux send-keys` + polling-`timeout` loops inside
+  one shell invocation.** Each poll can legitimately take up to its own
+  timeout (e.g. 15s) if a selector is wrong; five or six chained in one
+  command can add up past this tool's own command timeout even though
+  each individual step is fine. Send one or two commands per shell
+  call and check the pane before continuing.
+
+- **A crashed `chrome-headless-shell` process can leave a large core
+  dump file** (`server/core`, ~170MB, from the crash described above)
+  if core dumps are enabled. It's not part of the app -- delete it, and
+  use full `chrome` (as `driver.py` does) to avoid triggering it again.
+
+## Troubleshooting
+
+- **`error while loading shared libraries: libglib-2.0.so.0` (or similar)
+  when launching Chromium directly**: `LD_LIBRARY_PATH` isn't set --
+  source `.claude/skills/run-server/env.sh` first, or run through
+  `driver.py`, which reads `RUN_SERVER_CHROME_BIN` from it.
+- **`fc-list` prints nothing after extracting fonts**: `fonts.conf`'s
+  `<dir>` entries still point at the real (unpopulated) `/usr/share/fonts`.
+  `setup.sh` patches this with `sed`; if you extracted packages by hand,
+  do the same.
+- **`E: Could not open lock file ... Permission denied` from `apt-get`**:
+  you're missing the `-o Dir::State::Lists=... -o Dir::Cache=...`
+  overrides -- plain `apt-get update`/`install` always needs root here.
+- **`tmux: command not found`**: not preinstalled and no sudo; run
+  `setup.sh`, which fetches it the same non-root way as Chromium's libs.
+- **`tmux: error while loading shared libraries: libutempter.so.0`**:
+  you sourced `env.sh` inside the driver's tmux pane but not in the
+  shell that *invokes* `tmux` itself -- `tmux` was extracted from the
+  same non-root `.deb` set as Chromium and needs `LD_LIBRARY_PATH` too.
+  `source .claude/skills/run-server/env.sh` before the first `tmux`
+  command, not just inside `send-keys`.

.claude/skills/run-server/driver.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python3
+"""REPL driver for the espresso_frame server's web UI.
+
+Playwright-based since chromium-cli isn't available in this container.
+Reads one command per line from stdin, prints a result line -- built
+for tmux send-keys/capture-pane use by an agent. Vocabulary mirrors
+chromium-cli where it overlaps (nav/wait-for/click/fill/screenshot/
+eval/console-errors).
+
+Requires setup.sh to have run first (Python venv + Playwright Chromium
++ the non-root shared-lib/font extraction). Run via:
+
+    .claude/skills/run-server/env.sh sourced, then
+    server/.venv/bin/python .claude/skills/run-server/driver.py
+
+See SKILL.md for the full agent-path invocation (tmux wrapping etc).
+"""
+import glob
+import json
+import os
+import sys
+
+from playwright.sync_api import sync_playwright
+
+SHOT_DIR = os.environ.get("SCREENSHOT_DIR", "/tmp/run-server-shots")
+os.makedirs(SHOT_DIR, exist_ok=True)
+BASE = os.environ.get("RUN_SERVER_BASE_URL", "http://127.0.0.1:8420")
+
+
+def find_chrome() -> str:
+    override = os.environ.get("RUN_SERVER_CHROME_BIN")
+    if override and os.path.exists(override):
+        return override
+    matches = glob.glob(os.path.expanduser("~/.cache/ms-playwright/chromium-*/chrome-linux64/chrome"))
+    if not matches:
+        sys.exit("chrome binary not found -- run setup.sh first")
+    return matches[0]
+
+
+pw = sync_playwright().start()
+# The FULL `chrome` binary, not chrome-headless-shell (Playwright's
+# default headless target): chrome-headless-shell crashed on basic
+# calls like set_content() in this container even once every
+# ldd-reported missing lib was resolved. Full chrome + --no-sandbox is
+# stable here.
+browser = pw.chromium.launch(executable_path=find_chrome(), args=["--no-sandbox"])
+page = browser.new_page(viewport={"width": 1280, "height": 900})
+console_errors: list[str] = []
+page.on("console", lambda msg: console_errors.append(msg.text) if msg.type == "error" else None)
+page.on("pageerror", lambda exc: console_errors.append(str(exc)))
+# Playwright auto-DISMISSES native confirm()/alert() dialogs by default
+# (returns false) -- several destructive actions in this app (remove
+# widget, clear all widgets) gate on `confirm()`, so without this a
+# `click` on one of those buttons would silently no-op. Auto-accept
+# instead, since a driver testing a "yes, do the destructive thing"
+# flow needs the confirm to actually go through.
+page.on("dialog", lambda dialog: dialog.accept())
+
+
+def cmd_nav(arg):
+    url = arg if arg.startswith("http") else BASE + arg
+    page.goto(url)
+    print("nav ->", page.url)
+
+
+def cmd_wait_for(arg):
+    page.wait_for_selector(arg, timeout=10000)
+    print("found:", arg)
+
+
+def cmd_click(arg):
+    page.click(arg)
+    print("clicked:", arg)
+
+
+def cmd_fill(arg):
+    sel, _, value = arg.partition(" ")
+    page.fill(sel, value)
+    print("filled:", sel, "=", value)
+
+
+def cmd_press(arg):
+    page.keyboard.press(arg)
+    print("pressed:", arg)
+
+
+def cmd_screenshot(arg):
+    name = arg or f"ss-{len(os.listdir(SHOT_DIR))}"
+    path = os.path.join(SHOT_DIR, name + ".png")
+    page.screenshot(path=path)
+    print("screenshot:", path)
+
+
+def cmd_eval(arg):
+    try:
+        print(json.dumps(page.evaluate(arg)))
+    except Exception as e:
+        print("ERROR:", e)
+
+
+def cmd_console_errors(_arg):
+    print(json.dumps(console_errors))
+
+
+def cmd_viewport(arg):
+    """Resize the viewport. No args -> 390x844 (iPhone-ish mobile
+    width); the page itself starts at 1280x900 (desktop) on launch, so
+    `viewport 1280 900` gets back to that. The app's mobile breakpoint
+    is 860px (see theme.css) -- anything under that exercises the
+    off-canvas sidebar/mobile-bar layout."""
+    parts = arg.split()
+    width = int(parts[0]) if len(parts) > 0 else 390
+    height = int(parts[1]) if len(parts) > 1 else 844
+    page.set_viewport_size({"width": width, "height": height})
+    print("viewport:", width, "x", height)
+
+
+def cmd_is_open(arg):
+    """App-specific: print whether a <dialog> element is open (true/false)."""
+    print(json.dumps(page.eval_on_selector(arg, "el => el.open")))
+
+
+def cmd_bootstrap_admin(arg):
+    """App-specific: complete first-run /setup (username/password args,
+    default admin/testpassword123). Every scratch DB starts with no
+    users, and /setup is the only way in -- it also auto-links the
+    pre-existing frame #1 (created by migrations) to the new admin, so
+    /frames/1/... is reachable right after this."""
+    parts = arg.split()
+    username = parts[0] if len(parts) > 0 else "admin"
+    password = parts[1] if len(parts) > 1 else "testpassword123"
+    page.goto(BASE + "/setup")
+    page.fill("input[name=username]", username)
+    page.fill("input[name=password]", password)
+    page.click("button[type=submit]")
+    page.wait_for_load_state("networkidle")
+    print("bootstrapped admin, now at:", page.url)
+
+
+def cmd_quit(_arg):
+    browser.close()
+    pw.stop()
+    sys.exit(0)
+
+
+COMMANDS = {
+    "nav": cmd_nav,
+    "wait-for": cmd_wait_for,
+    "click": cmd_click,
+    "fill": cmd_fill,
+    "press": cmd_press,
+    "screenshot": cmd_screenshot,
+    "eval": cmd_eval,
+    "console-errors": cmd_console_errors,
+    "viewport": cmd_viewport,
+    "is-open": cmd_is_open,
+    "bootstrap-admin": cmd_bootstrap_admin,
+    "quit": cmd_quit,
+}
+
+print("run-server driver -- commands:", ", ".join(COMMANDS), flush=True)
+print("driver>", end=" ", flush=True)
+for line in sys.stdin:
+    line = line.strip()
+    if not line:
+        print("driver>", end=" ", flush=True)
+        continue
+    cmd, _, rest = line.partition(" ")
+    fn = COMMANDS.get(cmd)
+    if fn is None:
+        print("unknown command:", cmd, "-- try one of:", ", ".join(COMMANDS))
+    else:
+        try:
+            fn(rest)
+        except Exception as e:
+            print("ERROR:", e)
+    print("driver>", end=" ", flush=True)

.claude/skills/run-server/setup.sh

@@ -0,0 +1,110 @@
+#!/usr/bin/env bash
+# One-time (idempotent) environment bootstrap for running the
+# espresso_frame FastAPI server and browser-driving its UI, in a
+# container that ships with NO Python/Node/Docker/browser and NO sudo.
+# Re-run any time; every step checks whether it already happened.
+set -euo pipefail
+cd "$(git -C "$(dirname "${BASH_SOURCE[0]}")" rev-parse --show-toplevel)/server"
+
+UV_BIN="$HOME/.local/bin/uv"
+DEPS_ROOT="/tmp/run-server-chromium-deps"
+APT_WORK="/tmp/apt-work-run-server"
+SKILL_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ENV_FILE="$SKILL_DIR/env.sh"
+
+# 1. uv: a static Rust binary that can fetch its own Python build with
+# no C compiler needed (this container has none).
+if [ ! -x "$UV_BIN" ]; then
+  echo "installing uv..."
+  curl -LsSf https://astral.sh/uv/install.sh | sh
+fi
+
+# 2. Python 3.12 + venv + server deps
+if [ ! -x .venv/bin/uvicorn ]; then
+  echo "creating venv + installing server deps..."
+  "$UV_BIN" python install 3.12
+  "$UV_BIN" venv --python 3.12 .venv
+  "$UV_BIN" pip install -r requirements.txt
+fi
+
+# 3. Playwright (Python) + its Chromium download (~280MB: full chrome +
+# chrome-headless-shell + ffmpeg)
+if ! .venv/bin/python -c "import playwright" 2>/dev/null; then
+  echo "installing playwright..."
+  "$UV_BIN" pip install playwright
+fi
+if ! ls "$HOME"/.cache/ms-playwright/chromium-*/chrome-linux64/chrome >/dev/null 2>&1; then
+  echo "downloading chromium..."
+  .venv/bin/playwright install chromium
+fi
+
+# 4. Chromium's shared libs + fonts. `playwright install-deps` and
+# `apt-get install` both need root; neither is available. Instead:
+# download the .deb files directly (apt-get download works read-only
+# without root once given a user-writable state dir) and extract
+# (not install) them with dpkg-deb -x, which needs no root either.
+if [ ! -f "$DEPS_ROOT/usr/lib/x86_64-linux-gnu/libglib-2.0.so.0" ]; then
+  echo "fetching chromium's shared libs + fonts (non-root)..."
+  mkdir -p "$APT_WORK/lists" "$APT_WORK/cache/archives/partial" "$APT_WORK/debs" "$DEPS_ROOT"
+
+  apt-get -o Dir::State::Lists="$APT_WORK/lists" -o Dir::Cache="$APT_WORK/cache" \
+    -o Dir::Etc::SourceParts=/dev/null update
+
+  PKGS="libglib2.0-0t64 libnspr4 libnss3 libatk1.0-0t64 libatk-bridge2.0-0t64
+        libdbus-1-3 libxcomposite1 libxdamage1 libxfixes3 libxrandr2 libgbm1
+        libxkbcommon0 libasound2t64 libatspi2.0-0t64 libcups2t64 libcairo2
+        libpango-1.0-0 libpangocairo-1.0-0 libx11-6 libxcb1 libxext6
+        fonts-liberation fontconfig libfontconfig1"
+  # ^ first row: what chrome-headless-shell's ldd reported missing.
+  # second row: what the FULL chrome binary additionally needed (we use
+  # full chrome, not headless-shell -- see Gotchas in SKILL.md).
+
+  apt-get -o Dir::State::Lists="$APT_WORK/lists" -o Dir::Cache="$APT_WORK/cache" \
+    -o Dir::Etc::SourceParts=/dev/null install --download-only --reinstall -y \
+    --print-uris $PKGS | grep -oP "^'[^']+'" | tr -d "'" > "$APT_WORK/urls.txt"
+
+  (cd "$APT_WORK/debs" && xargs -n1 -P8 curl -sS -O --max-time 30) < "$APT_WORK/urls.txt"
+  for f in "$APT_WORK"/debs/*.deb; do dpkg-deb -x "$f" "$DEPS_ROOT"; done
+
+  # fonts.conf as shipped points at the real /usr/share/fonts, which is
+  # root-owned and has nothing extracted into it. Point it at our
+  # extracted copy instead, and give it a writable cache dir.
+  mkdir -p /tmp/run-server-fontcache
+  sed -i "s#<dir>/usr/share/fonts</dir>#<dir>$DEPS_ROOT/usr/share/fonts</dir>#" \
+    "$DEPS_ROOT/etc/fonts/fonts.conf"
+  sed -i "s#<cachedir>.*</cachedir>#<cachedir>/tmp/run-server-fontcache</cachedir>#" \
+    "$DEPS_ROOT/etc/fonts/fonts.conf"
+
+  PATH="$DEPS_ROOT/usr/bin:$PATH" \
+    LD_LIBRARY_PATH="$DEPS_ROOT/usr/lib/x86_64-linux-gnu:$DEPS_ROOT/lib/x86_64-linux-gnu" \
+    FONTCONFIG_PATH="$DEPS_ROOT/etc/fonts" \
+    "$DEPS_ROOT/usr/bin/fc-cache" -f
+fi
+
+# 5. tmux -- also missing, also no apt/sudo. Same non-root download +
+# dpkg-deb -x trick, into the same extracted root (so its `usr/bin` is
+# already on PATH via env.sh).
+if [ ! -f "$DEPS_ROOT/usr/bin/tmux" ]; then
+  echo "fetching tmux (non-root)..."
+  mkdir -p "$APT_WORK/lists" "$APT_WORK/cache/archives/partial" "$APT_WORK/debs" "$DEPS_ROOT"
+  apt-get -o Dir::State::Lists="$APT_WORK/lists" -o Dir::Cache="$APT_WORK/cache" \
+    -o Dir::Etc::SourceParts=/dev/null install --download-only --reinstall -y \
+    --print-uris tmux | grep -oP "^'[^']+'" | tr -d "'" > "$APT_WORK/tmux_urls.txt"
+  (cd "$APT_WORK/debs" && xargs -n1 -P3 curl -sS -O --max-time 30) < "$APT_WORK/tmux_urls.txt"
+  for f in $(sed -E 's#.*/##' "$APT_WORK/tmux_urls.txt"); do dpkg-deb -x "$APT_WORK/debs/$f" "$DEPS_ROOT"; done
+fi
+
+CHROME_BIN="$(ls "$HOME"/.cache/ms-playwright/chromium-*/chrome-linux64/chrome | head -1)"
+
+cat > "$ENV_FILE" <<EOF
+# Generated by setup.sh. Source this before running driver.py (it
+# needs LD_LIBRARY_PATH/FONTCONFIG_PATH set before the Chromium
+# subprocess launches -- driver.py does not source it for you).
+# start-server.sh does NOT need this file -- uvicorn has no such deps.
+export PATH="\$HOME/.local/bin:$DEPS_ROOT/usr/bin:\$PATH"
+export LD_LIBRARY_PATH="$DEPS_ROOT/usr/lib/x86_64-linux-gnu:$DEPS_ROOT/lib/x86_64-linux-gnu"
+export FONTCONFIG_PATH="$DEPS_ROOT/etc/fonts"
+export RUN_SERVER_CHROME_BIN="$CHROME_BIN"
+EOF
+
+echo "setup complete -> $ENV_FILE"

.claude/skills/run-server/start-server.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+# Background-launch the server against a scratch DB/config -- never the
+# real deployment's data (see CLAUDE.md). Waits for readiness, prints
+# the PID and log path. Usage: ./start-server.sh [scratch-dir] [port]
+set -euo pipefail
+cd "$(git -C "$(dirname "${BASH_SOURCE[0]}")" rev-parse --show-toplevel)/server"
+
+SCRATCH="${1:-/tmp/run-server-scratch}"
+PORT="${2:-8420}"
+mkdir -p "$SCRATCH"
+
+DATABASE_URL="sqlite:///$SCRATCH/test.db" \
+CONFIG_PATH="$SCRATCH/config.json" \
+.venv/bin/uvicorn app.main:app --host 127.0.0.1 --port "$PORT" \
+  > "$SCRATCH/server.log" 2>&1 &
+PID=$!
+echo "$PID" > "$SCRATCH/server.pid"
+
+for _ in $(seq 1 30); do
+  curl -sf -o /dev/null "http://127.0.0.1:$PORT/" && break
+  sleep 0.5
+done
+
+if ! curl -sf -o /dev/null "http://127.0.0.1:$PORT/"; then
+  echo "server did not become ready -- check $SCRATCH/server.log" >&2
+  exit 1
+fi
+
+echo "server PID $PID up on http://127.0.0.1:$PORT (log: $SCRATCH/server.log, db: $SCRATCH/test.db)"

.claude/skills/run-server/stop-server.sh

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+# Usage: ./stop-server.sh [scratch-dir]
+SCRATCH="${1:-/tmp/run-server-scratch}"
+if [ -f "$SCRATCH/server.pid" ]; then
+  kill "$(cat "$SCRATCH/server.pid")" 2>/dev/null || true
+  rm -f "$SCRATCH/server.pid"
+  echo "stopped"
+else
+  echo "no $SCRATCH/server.pid -- nothing to stop"
+fi