Skip to content

Chat UI: shadcn/ui layout, a truthful loading state, config-by-URL, and a slot-filled settings dialog - #15

Merged
mkotelnikov merged 20 commits into
mainfrom
llm-chat-ui
Sep 27, 2026
Merged

Chat UI: shadcn/ui layout, a truthful loading state, config-by-URL, and a slot-filled settings dialog#15
mkotelnikov merged 20 commits into
mainfrom
llm-chat-ui

Conversation

@mkotelnikov

@mkotelnikov mkotelnikov commented Sep 22, 2026

Copy link
Copy Markdown
Contributor

Rebuilds apps/llm-chat's UI on shadcn/ui: a layout that works on a phone, a loading state that
tells the truth while a local model thinks, configuration by fetched URL with a trust rule, and a
settings dialog filled entirely through slots — while index.html still works with no mesh at
all.

Implements statewalker/umbrella#36. Spec and plan:
docs/superpowers/specs/2026-09-20-llm-chat-ui-design.md and the matching plan (umbrella repo,
statewalker/umbrella#39). Independent of the appliance PR #14 — different files, no overlap.

What changed

  • shadcn/ui scaffold + nine primitives. The old Modal.tsx had no focus trap and no Escape
    handling; the Dialog now gets both from Radix, and a test asserts it.
  • Three run phases — idle → waiting → streaming. isRunning is now derived from
    phase, so the two cannot disagree about whether a request is in flight. Against a local
    llama.cpp model waiting legitimately lasts 30 s+, and one spinner cannot tell that apart from a
    hung request; waiting shows a placeholder with an elapsed counter and Stop.
  • Sidebar on desktop, Sheet drawer below md, sticky composer with safe-area padding and
    h-dvh. The sidebar appears in the accessibility tree exactly once.
  • ?config=<url> — the app fetches a config document. Resolution: defaults < document < what
    the user saved.
  • A trust rule for that URL, applying discover.ts's "TRUST ONLY THE HUB" reasoning: a URL the
    app obeys is a redirect of every message and any key in it.
  • A settings dialog filled through @statewalker/shared-slots. It imports no panel. The chat
    contributes Connection and Models; mesh.html contributes Sharing and Keys. index.html shows
    two tabs and mesh.html four, with no branch anywhere deciding it.

Measured

  • Suite 1373 passed / 8 skipped / 0 failed, against a pre-branch baseline of 1181/8/0.
    apps/llm-chat went 124 → 238.
  • tests/e2e/standalone.spec.mjs 7/7, tests/e2e/responsive.spec.mjs 10/10 at 1280×800
    and 390×844, tests/e2e/mesh.spec.mjs passing live against a real appliance with a streamed
    reply from a local model.
  • index.html's bundle contains zero httpeers/libp2p content matches (grepped by content, not
    chunk name): ~770 kB raw / ~239 kB gzip.

Security notes worth reading before merging

  • On mesh.html, a ?config= cannot override the endpoint discovered from the hub. That is
    deliberate: otherwise a crafted link would redirect the chat to a foreign endpoint while the user
    believed they were talking to the mesh they had just joined, carrying the mesh's key header. The
    spec originally advertised the opposite and was amended. A test pins it, and inverting the
    precedence makes exactly that test fail.
  • Same-origin trust is skipped entirely once an edge base is set, because the mesh edge host
    serves every peer's mount side by side — so "same-origin" would otherwise re-admit sibling peers
    and .. escapes.
  • An untrusted config document is fetched before it is confirmed, because the dialog must
    name the baseUrl it would use. Nothing is applied without consent, and the fetch carries no
    credentials — a crafted link is a ping, not a leak. Spec §6 was amended to say this plainly.

Decisions reversed on evidence

  • assistant-ui's shadcn-registry components were tried, not assumed. They install cleanly
    against React 19.3.0 and @assistant-ui/react 0.15.19, and were rejected for four concrete
    conflicts: they rewrite three files other tasks own, install radix-ui as a meta-package
    alongside the individual packages, pull 17 files of attachment/tool-call/branch-picker surface
    unused by a text-only chat, and break the controller-free Thread interface. Recorded in
    Thread.tsx's header with what it costs.

Known gaps

  • The elapsed-counter ticking test flaked once in seven runs under heavy parallel load — likely a
    fake-timer race, not a logic bug. Worth watching in CI.
  • ModelPicker.tsx has no tests (pre-existing, untouched here).
  • styles.css carries a global !important overriding Radix ScrollArea's inline display:table,
    which defeated truncate and clipped a long session title's delete button out of the mobile
    drawer — with no page-level scroll to show for it. One consumer today; the risk of a second is
    noted beside the rule.

🤖 Generated with Claude Code

mkotelnikov and others added 20 commits September 22, 2026 15:06
@mkotelnikov @claude
test(llm-chat): a jsdom + Testing Library substrate for component tests
5d1e233 
Nothing in this app rendered a component under test before. The UI work needs
a red/green loop that does not launch a browser.

Global jsdom environment (vs per-file docblocks): switching environment from
"node" to "jsdom" globally did not break idb-stores.test.ts (fake-indexeddb)
or any of the pure-Node tests — all 124 pre-existing tests plus the new one
pass. No per-file override was needed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): distinguish waiting from streaming in the run state
f874b6f 
Against a local model, prompt processing can make 'sent, nothing back yet'
last 30s. One spinner cannot tell that apart from a hung request, and neither
can the person watching it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): the shadcn/ui scaffold and nine primitives
4984e12 
Adopted for the accessibility the hand-rolled Modal never had: the dialog now
traps focus and closes on Escape, and the test asserts it.

The shadcn CLI (v4.21, Nova preset) works against Tailwind 4 + Vite + React 19,
but its `init` pulls in a whole opinionated design system (Geist font via
@fontsource, tw-animate-css, `shadcn` itself as a runtime dependency, a
third-party `cn` package) and writes components to src/components/ui, not the
src/ui/primitives the brief calls for. Reverted that init and hand-installed
the nine primitives from the canonical shadcn/Radix source instead: real
@radix-ui/* packages, clsx + tailwind-merge for cn(), class-variance-authority
for variants, lucide-react for icons — matching what the brief named. A `@/*`
path alias was required for shadcn-style imports (tsconfig paths, plus a
resolve.alias in vite.config.ts and vitest.config.ts); TypeScript here has
dropped `baseUrl`, so `paths` alone does the job. theme.css carries the light/
dark CSS variables in Tailwind 4's CSS-first form, imported from styles.css,
which stays the single Vite entry. A package-local biome.json (extends the
root config by relative path; the `//` monorepo-root shorthand didn't resolve
here) enables the CSS parser's tailwindDirectives for @theme/@apply.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): a React binding for shared-slots, local for now
f80ecc6 
useSyncExternalStore over observe/getSnapshot, so a panel registered after the
dialog mounts appears without a remount. getSnapshot on a keyed slot returns
a ReadonlyMap; the derived T[] is cached per Map reference so the array
identity only changes when the bus actually notifies, not on every render.
Extracted to a package when a second app needs it, not before.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): one settings dialog, filled through a slot
1bc54c5 
The dialog imports no panel and nothing from src/mesh -- a test enforces it.
That is what lets index.html show two tabs and mesh.html four, with no branch
anywhere deciding it.

ConnectionPanel and ModelsPanel take the old SettingsDialog/ModelDialog
content and are registered into settingsPanelsSlot by ChatApp for now (Task 6
moves the registration up to an app shell). Deleting Modal.tsx/ModelDialog.tsx
touched five more importers: Thread.tsx and pages/mesh.tsx now source their
button/input classes from the new ui/button-styles.ts, and mesh/member-key.tsx
moved its own dialog onto the shadcn Dialog primitive instead of keeping a
second implementation around.

startupStep's non-dismissible first-run gating is preserved (now driving one
dialog instead of two) and covered by a new chat-app.test.tsx, since no prior
test exercised ChatApp's render tree; that render also needed a ResizeObserver
polyfill in tests/setup.ts that assistant-ui's Thread requires under jsdom.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
fix(llm-chat): open the settings dialog on the right tab for the step
f9fdae2 
startupStep("models") kept the dialog non-dismissible but SettingsDialog
always defaulted to the first panel (Connection) -- a user who just saved
their base URL landed back on the form they finished, not on Models.

SettingsDialog takes an optional initialPanelId (a plain panel id, never a
concept like "models", so the dialog still imports no panel): resolved fresh
on every render against the currently registered panels, falling back to the
first panel when absent or when the id names a panel that isn't registered
(yet, or ever) -- a panel can register after the dialog has mounted. ChatApp
passes MODELS_PANEL_ID when startupStep === "models", using the same id
constant it registers the panel under.

Both new settings-dialog.test.tsx cases (opens on a registered id; falls back
for an unregistered one) and the extended chat-app.test.tsx "models" case
(asserting the Models tab is aria-selected) failed against the prior code
before this change.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): a sidebar/Sheet shell, with the sidebar in the a11y t…
3ca086f 
…ree once

AppShell takes no chat concept, so the mesh page reuses it unchanged. The
breakpoint itself is Playwright's to prove, not jsdom's.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): render waiting and streaming differently, with an ela…
c5d907e 
…psed counter

Thread no longer takes a ChatController; it takes ChatState, now, and plain
callbacks, so it renders and asserts standalone. Three phases replace the old
ThreadPrimitive.If running split: idle (nothing), waiting (a placeholder
bubble, data-testid="waiting-indicator", aria-live="polite", elapsed whole
seconds, Stop), streaming (partial text, Stop stays). thread-adapter.ts
derives isRunning from phase (isRunningFor) so phase is the one thing passed
through to assistant-ui's runtime, not a duplicate field.

Composer, previously inline in Thread's own tree, is now its own component
wired through AppShell's composer slot -- the sticky, safe-area-padded
footer (Task 6) finally reaches the real app instead of sitting unused.

Watched against scripts/fake-llm.mjs with an opt-in FIRST_TOKEN_DELAY_MS: a
6s wait shows a ticking 0s..5s counter and a live Stop button, not a hang;
streaming replaces the placeholder the instant the first token lands; Stop
mid-wait aborts immediately and clears the indicator.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
docs(llm-chat): record the attempted assistant-ui Thread registry route
7071445 
Fix round 1/5 for Task 7. Actually ran `pnpm dlx shadcn@latest add
"https://r.assistant-ui.com/thread" --overwrite` against React 19.3.0 and
@assistant-ui/react 0.15.19 (both unforced) instead of reasoning about it
from outside. It installs, but its output conflicts on every other axis:
resurrects the `cn`-npm-package and `radix-ui`-meta-package choices Task 3
deliberately rejected (radix-ui lands alongside the individual
@radix-ui/react-* packages, not instead of them), rewrites button.tsx/
dialog.tsx/tooltip.tsx which this task doesn't own, and pulls unrelated
dependencies and 17 files for attachment/reasoning/tool-call/branch-picker
surface this text-only OpenAI-compatible chat has no use for.

Reverted the experiment in full (working tree confirmed clean, node_modules
resynced), kept the hand-rolled Thread/Composer, and recorded the attempt
and the concrete conflicts in Thread.tsx's header comment so the next
reader doesn't re-litigate it from scratch. No dependency or version
changed; 180/180 still green, typecheck and build clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): fetchable connection configuration, with the saved va…
939835e 
…lue winning

A user who typed an endpoint does not have it replaced on the next load. A
static host's 200-with-an-HTML-index is rejected where it happens, not three
hops later.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): decide which config URLs may be obeyed without asking
8826e4a 
Same rule discover.ts already applies to the hub's service document: a URL the
app obeys redirects every message and every key. Origin comparison, never a
string prefix -- the @-userinfo and suffix cases are in the suite. Once a mesh
edge is known, plain same-origin is not enough either: the edge's host serves
every peer's mount side by side, so only the specific edge path is trusted,
matching discover.ts's "another peer is not your hub" rule. An RFC 3986 §4.2
guard (colon in a relative reference's first segment) keeps "::::" distrusted
without ever needing to throw.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): accept a ?config= URL, confirming any foreign origin
173de6a 
On load each page reads ?config=: judgeConfigUrl (Task 9) decides trust
against the page URL and, on mesh.html, the discovered edge base --
mesh.html only mounts ChatApp once discovery has succeeded, so a
?config= arriving earlier is naturally held rather than fetched.
Trusted URLs fetch, parse, resolveConfig and save without asking.
Untrusted ones go through the new ConfirmConfigDialog; accept does the
same as trusted, reject leaves the config untouched and says so. Any
fetch/parse failure is a visible, non-fatal notice naming the URL,
falling back to the saved config or the settings screen. Cancel is the
dialog's default focused action and Escape rejects, because its job is
to stop something, not to be dismissed out of the way.

ExternalConfig.defaultModel, parsed since Task 8 but dropped by
applyEndpoint, is now fed as the candidate default to refreshModels
right after GET /models -- the same function the model picker's
"Refresh" already uses. The dialog never names the WHATWG opaque
origin "null"; callers substitute the raw URL. config-trust.ts also
now refuses %2f/%5c in a mesh-edge candidate path, since RFC 3986
means new URL() never collapses them the way a downstream decoder
might.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
test(llm-chat): pin that a mesh ?config= document cannot override the…
2f5c609 
… hub's endpoint

Fix round 1/5 for Task 10's review. mesh.tsx's discover() always saves
the hub-discovered endpoint before ChatApp mounts, so resolveConfig's
"saved wins" rule already made ?config= inert on the mesh page --
confirmed deliberate, not accidental: a document that could redirect
the chat away from the discovered hub is exactly the attack
discover.ts's "trust only the hub" rule and the edge-base
same-origin-skip exist to prevent. Add a test pinning that precedence
(pre-populated configStore + edgeBase set, a same-edge document with a
different baseUrl, endpoint unchanged), and a comment at the
resolveConfig call site explaining why, not just that.

?config= remains a standalone-page mechanism; the mesh page always
prefers what it discovered.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
feat(llm-chat): mesh sharing and keys arrive as slot contributions
62005be 
index.html shows two tabs and mesh.html four, with no branch deciding it and
no mesh code reachable from the standalone entry point -- asserted by closure.

ChatApp gains an optional `slots` prop so mesh.tsx's registerMeshPanels can
contribute into the same dialog Connection and Models already fill, instead
of the mesh page's own headerExtra render.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
test(llm-chat): the standalone page reaches for nothing but its endpoint
1732dd7 
"Works without a mesh" asserted as no ServiceWorker and no request to any
other host -- not merely that the page renders. Seven assertions against a
real Chromium driving dist/index.html: a non-dismissible first-run settings
screen, the fake endpoint's models, a delayed-first-token waiting indicator
with an elapsed counter, Stop-during-wait leaving the session intact, the
two-tab (Connection/Models) settings dialog, no ServiceWorker/no stray
requests, and zero page errors or unhandled rejections.

fake-llm.mjs's CLI entry point gains --first-token-delay-ms, reconciled with
Task 7's FIRST_TOKEN_DELAY_MS env var rather than adding a second mechanism
-- both feed the same startFakeLlm({ firstTokenDelayMs }) option the new
spec calls directly.

scripts/smoke.mjs had known stale drift from Task 5's slot-filled tabbed
Settings dialog (it still targeted a dead two-dialog "Connection
settings"/"Choose a model" flow with a "Change connection" button); updated
its selectors to the current single dialog so it stays a useful, passing
regression check alongside this new acceptance spec rather than bitrotting
further.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
test(llm-chat): real-viewport acceptance for the sidebar/Sheet layout…
ed6536a 
…, and a live-mesh spec

Ten assertions at 1280x800 and 390x844 (responsive.spec.mjs), including no horizontal
scroll and a composer that survives focus. jsdom asserts none of this and never claimed
to. The first real run failed in the test harness itself (waiting on "New chat", which
is legitimately hidden at a phone viewport) rather than in AppShell; fixing that surfaced
a genuine bug while stress-testing assertion 5 with a long session title: ScrollArea's
Viewport wraps children in an inline `display:table` div that defeats `truncate`/`min-w-0`
below it, so a long title pushed its delete button out of the mobile Sheet drawer with no
page-level horizontal scroll to show for it. Fixed with `min-w-0` on the row (ThreadList.tsx)
and a targeted `!important` override for Radix's inline style (styles.css), and pinned by
strengthening assertion 5 to check that a long-titled row's own delete button stays on
screen, not just that the thread-list landmark is visible.

mesh.spec.mjs promotes scripts/mesh-smoke.mjs, which had drifted onto the dead
"Choose a model" dialog and a `?join=` link. The new spec joins by pasting a blob into
the join form (never `?join=`), confirms the settings dialog's four tabs (Connection,
Models, Sharing, Keys), mints a key for a member from the Keys panel, and waits for a
genuinely streamed reply -- run and passing against the live local appliance at
worktrees/llm-appliance-local, minting its own admin invitation via that appliance's own
bin/invite.sh rather than a hardcoded door URL (the old script's HUB_DOOR_URL defaulted
to :8080, which on this box is someone else's SSH tunnel to production, not this
appliance). Skips with a clear message when no appliance is reachable.
mesh-smoke.mjs is now a one-line forward so the two can't drift apart again.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
docs(llm-chat): the two pages, ?config=, slots, and the three run phases
3306eff 
Audited against the code before merge: README.md written from scratch (none
existed), covering index.html vs mesh.html, the ?config= document shape /
resolution order / trust rule (foreign origin confirmed not refused; the
mesh page's discovered endpoint always wins by design), the settings-dialog
slot contract (SettingsPanel shape, register(decl, id, value) is 3-arg,
getSnapshot returns a ReadonlyMap, useSlot returns an array), the three run
phases and why `waiting` is its own phase, and how to run each test layer
including the mesh spec's live-appliance requirement and clean skip.

Doc-comment sweep against `git diff origin/main --stat`: two stale comments
found and fixed. ChatApp.tsx's `headerExtra` prop still claimed to be "the
mesh page's link status and links", but mesh.tsx stopped passing it once
Sharing/Keys became settings-dialog tabs and the pre-chat header grew its
own links -- neither shipped page passes it today. button-styles.ts claimed
Thread.tsx's composer as a consumer; Thread's composer moved onto shadcn
primitives and no longer imports it -- mesh.tsx's pre-chat forms are the
only consumer left. `config.ts`'s "one config per page" and `discover.ts`'s
trust-rule header were checked and are still accurate; left alone.

Three of the four deferred findings closed rather than recorded: added
tests/member-key.test.tsx (accessible name + Escape close path for the
shadcn Dialog migration that only typecheck/build covered before); added a
config-trust.test.ts case pinning judgeConfigUrl's opaque origin "null" and
a chat-app.test.tsx case proving ChatApp substitutes the raw URL for
display, never "null"; annotated styles.css's global `!important` override
with a note for whoever adds ScrollArea's second consumer. The fourth
(theme.css/ThreadList's XIcon/mesh-session.ts's configRef idiom) was left
alone, as instructed, since none of those files were touched.

Bundles (dist/, confirmed by content grep, not just chunk names):
index.html loads styles-*.js (759.6 kB/231.7 kB gzip, React+assistant-ui+
app code, zero @statewalker/httpeers-* and zero libp2p/multiaddr matches)
plus a 0.28 kB entry, a 0.77 kB runtime chunk and 28.1 kB/6.4 kB gzip of
CSS -- about 770 kB raw / 239 kB gzip total, no mesh code. mesh.html adds
mesh-*.js (690.6 kB/208.0 kB gzip, confirmed to contain libp2p) plus two
QR-scan chunks used only by the join widget (151.4 kB/55.1 kB gzip,
369.4 kB/108.4 kB gzip) -- about 1.91 MB raw / 610 kB gzip total.

Suite: 1365 passed / 8 skipped / 0 failed assembly-wide (was 1259/8/0),
apps/llm-chat at 230/20 files (was 124; +4 tests/+1 file from this task,
226/19 unchanged from Task 13). No new failures anywhere. typecheck and
lint:check clean. All three e2e specs pass: standalone (7/7), responsive
(10/10, no failures to diagnose this run), mesh (ran to completion against
a live appliance rather than skipping -- join, four tabs, minted key,
genuinely streamed reply all observed).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
fix(llm-chat): final-review fix wave -- five minor findings
2e2ef06 
- SettingsDialog: reset the manually-picked tab when the dialog closes, so a
  later `initialPanelId` (e.g. routing a user with no default model to
  Models) is never permanently defeated by an earlier, unrelated tab click.
- Thread `fetchImpl` through to all four listModels() call sites
  (ModelPicker refresh, ConnectionPanel's Test, ModelsPanel's load) so the
  doc comment's claim actually holds.
- Add unit coverage for ConnectionPanel and ModelsPanel, including that the
  resolved default model is marked (spec §9), previously asserted only
  incidentally by an e2e spec.
- Pin useNow's setInterval tick in vitest with fake timers, so deleting it
  can no longer ship silently past a vitest-only CI run.
- Stop offering ConfirmConfigDialog for an untrusted ?config= on the mesh
  page: its "Use this configuration" button could never do what it said,
  since the hub's discovered endpoint always wins there. Ignore the
  document with a plain notice instead; behavior is unchanged, only the
  promise the dialog made is fixed.

Each fix has a regression test that was confirmed to fail before the fix.
apps/llm-chat: 238/238 (was 230). Assembly-wide: 1373 passed / 8 skipped.
standalone.spec.mjs 7/7, responsive.spec.mjs 10/10; mesh.spec.mjs skipped --
the live appliance's traefik is bound to 127.0.0.1:8080, the port this pass
was instructed never to touch.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@mkotelnikov @claude
fix(llm-chat): stream the reply through the injected fetchImpl
875dbf3 
ChatApp's fetchImpl reached the config and model calls but not the chat
stream: endpointClient called streamChat without it, so the stream always
used the global fetch. The elapsed-time test's never-resolving stub was
therefore a real DNS lookup of llm.test, and the test failed whenever that
lookup failed before the fake clock advanced 3 s -- 2 in 40 runs locally,
and the one red check on PR #15.

endpointClient now takes a fetch getter, which ChatApp feeds from a ref as
it does the config. Production is unaffected: both pages pass the bound
global fetch. A new test pins the wiring deterministically; the 40-run
reproduction now passes 120/120.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@mkotelnikov
Merge remote-tracking branch 'origin/main' into llm-chat-ui
5b94964
@mkotelnikov
mkotelnikov merged commit a887f6d into main Sep 27, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@mkotelnikov