Demo-data screenshots of the desktop app

Produces docs/screenshots/light/*.png — the desktop counterparts of the images main's README used — from the real Tauri webview, over a corpus that is invented end to end. The map of which file shows what, and what a designer may still want to touch, is docs/screenshots/README.md; this skill is how the set is made. It builds on ui-verify (driving the webview) and adds the data side.

The idea in one paragraph

A debug build derives everything from HOME: the data dir is $HOME/.agento-desktop-dev and the scanner walks $HOME/.claude. So the app is launched with a fake HOME (~/.agento-demo by convention) holding a generated Claude Code corpus, and a fresh database is created there. Nothing reads or writes the real install. The only thing that has to be real is the set of project directories: the scanner decodes -home-you-Projects-auth into /home/you/Projects/auth only if that directory exists, and the UI only collapses /home/<x>/… to ~/… — so ten empty dirs are created under ~/Projects for the duration of the shoot and removed afterwards.

Run it

K=.claude/skills/app-demo-ui-screenshot
DEMO=$HOME/.agento-demo

# 1. corpus + the dirs it decodes against
python3 $K/gen_corpus.py "$DEMO"          # ~300 sessions, 10 projects, sub-agents
$K/make_dirs.sh                           # empty ~/Projects/<name> dirs (records what it created)

# 2. the app, isolated, under XWayland so the window can be sized
.claude/skills/ui-verify/app.sh --stop
setsid nohup $K/launch.sh "$DEMO" > /tmp/agento-app.log 2>&1 < /dev/null &
#    wait for :9224, then size the window to the README frame:
W=$(xwininfo -root -tree | grep '("agento" "Agento")' | grep -v 10x10 | awk '{print $1}')
wmctrl -i -r $W -b remove,maximized_vert,maximized_horz; xdotool windowsize $W 1440 900

# 3. agents / integrations / tasks through the app's own API, chats + job
#    history straight into the scratch DB (no create API for those)
python3 $K/seed_api.py "$DEMO"
.claude/skills/ui-verify/app.sh --stop      # one writer at a time
python3 $K/seed_db.py "$DEMO/.agento-desktop-dev/agento.db"
#    + mark the integrations authenticated and fill service tool lists — see
#      "Hand edits" below
setsid nohup $K/launch.sh "$DEMO" > /tmp/agento-app.log 2>&1 < /dev/null &   # relaunch, resize again
#    session_insights fills itself within a minute of the relaunch — see below

# 4. shoot (light theme — set it in the status bar / ⌘K first)
$K/shoot.sh docs/screenshots/light

# 5. clean up
.claude/skills/ui-verify/app.sh --stop
$K/make_dirs.sh --clean

gen_corpus.py is seeded and deterministic in everything except the session UUIDs, so a regenerated corpus is a new corpus to the scanner: rescan (POST /api/claude-sessions/refresh) and the app reconciles the orphaned session_insights rows and recomputes the new ones itself.

The parts that are not obvious

• session_insights fills itself now, and this step used to be manual. Until #408 the app had the nine insight processors as pure functions and the summary endpoint that reads the table, but nothing that wrote one — so Insights was empty on every fresh install, and this skill could only photograph it by building a Go writer out of git history (backfill.sh, pinned at 3b54e41) and running it against the demo data dir with Agento stopped. backfill.sh is deleted with the bug.

  What replaces it is native/insights/worker.rs, which sweeps at boot and every five minutes. Practically: relaunch and give it a minute, then check before shooting, because an Insights shot taken too early is a page of zeros that looks like a rendering bug rather than a timing one:

  sqlite3 -readonly "$DEMO/.agento-desktop-dev/agento.db" \
    "select (select count(*) from session_insights),
            (select count(*) from claude_session_cache)"
  

• Hand edits after seed_api.py, in the scratch DB (sqlite3):

  • UPDATE integrations SET auth='{"validated":true,"username":"acme-platform-bot"}' WHERE type='github' (and team_name / display_name / bot_username for slack / jira / telegram) — the real auth/validate dials the provider and 400s on a fake token, so the rows would otherwise read "Not connected".
  • fill every services.<svc>.tools list with the provider's tool names (they are in src/views/integrations/catalog.ts) — /available-tools iterates the explicit lists, so an empty list means an empty agent builder. Agent capabilities.mcp must use the real tool names (list_pulls, get_pull, not list_pull_requests).

• Project names must not share a prefix with a real ~/Projects entry. decode_project_path resolves greedily: docs-site beside a real docs/ decodes to …/docs/site, fails the existence check and falls back to the raw -home-you-Projects-docs-site. make_dirs.sh refuses to clobber an existing dir for the same reason.

• Settle before every shot. shoot.sh waits until a predicate is true on four consecutive polls before each shot, because Page.snapshotRect can race React's re-render — the first pass painted the previous task's toggle and schedule type under the new task's name. Wait on the new view's content, never on the navigation, and beware that text=Sessions in the sidebar does not pop an open session detail (shoot.sh clicks the toolbar's back button first).

• The do grammar splits on |, so no || inside an eval/wait line; write cond ? a : b or a nested ternary instead.

• Timestamps are UTC in the generator (NOW there); keep the newest synthetic session a couple of hours behind the real clock or the list shows "in 8 m".

• Windows sizing needs X11. launch.sh exports GDK_BACKEND=x11; on Wayland nothing outside the app can resize it, and window-state may restore it maximised (hence the wmctrl line).

What it leaves behind

~/.agento-demo (corpus + scratch DB, ~30 MB) — keep it to re-shoot, delete it whenever. make_dirs.sh --clean removes the ~/Projects stubs it created. Nothing under the real ~/.agento or ~/.claude is touched; verify with an mtime check if in doubt.