- Configure server write loop in hub.rs with max_server_message so outgoing ServerMsg payloads above 64 MiB (e.g. document snapshots in Welcome messages) are delivered without socket termination.
- Set core max_message_bytes in local live session HostConfig to max_server_message.
- Map tungstenite Capacity errors in guest read loop to Outcome::Failed with accurate size limits rather than Outcome::Refused which falsely reported 'The host refused'.
- Add loopback tests verifying Welcome delivery above client message limits and accurate diagnostics on client capacity overrun.
Found while testing crates/collab-net/src/invite.rs:242 (percent_decode)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Found while testing crates/collab-net/src/invite.rs:valid_host (around line 205)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Unrelated to the security changes in this branch; `main` already failed
`clippy --workspace --all-targets -- -D warnings` without them. Eight
clippy::nonminimal_bool sites across seven files, each taking clippy's own
suggested rewrite.
Every rewrite is pure boolean algebra with the negation kept outside, so the
float comparisons and their NaN behaviour are untouched: !A && !B becomes
!(A || B), and !opt.is_some_and(f) becomes opt.is_none_or(!f). Two sites needed
explicit parentheses the suggestion omitted, where the rewritten term is the left
operand of an && and || binds looser.
designcraft cleared the same lint in bc840130; this is figcraft's equivalent.
The hub kept the addresses removed from every file it ever had open. The
entry now goes when the file is closed (everyone left) or removed from the
server, so a long-running public server doesn't keep collecting them.
Right-click a person's avatar ▸ "Remove <name> from session", or
`collab.removePerson {client}`, when you host the session or own the file
on a server. The hub closes their connection as the end of the session for
them (close 1001: they keep their copy and don't reconnect) and bars their
address from that file while it runs, so the code or link they hold no
longer gets them in; everyone else stays connected.
Protocol addition within version 2: `Admin::RemovePeer {file, client}` on
an open hosting connection with the file's owner token (or the server
key). Older servers say they can't.
Tests: loopback (removed, can't rejoin, others stay), engine (host can,
guest can't, bad ids refused), codec round trip; the collab junk sweep
covers the new command.
The sessions test waited for a new layer to arrive and then asserted on its name, which the host sets in a second batch; it now waits for the name. Test only.
The recovery copy of one of your online files is no longer discarded when
the file reopens from its server. It is kept until the outcome is known:
- the server has everything in it: the copy goes;
- the copy has changes that never reached the server and nothing else
changed there: they are applied on top as one edit and sent to everyone;
- anything else (others changed the file, the server forgot the client):
the copy opens beside the file as "<name> (recovered copy)", with a toast;
- the server can't be reached: the copy opens as an unsaved document.
The note beside a copy is written after it and records who the app was in
the file and the server's last change it had. The app joins again as that
client (the first hello resumes; a new sync client numbers its batches
after the ones the server processed), so the layers it made are its own to
send. The old copy is removed only once this run has copied what it became.
collab.recoverFile; a tooltip on the tab mark; docs.
Online, the file is its owner's: others get links and codes of their own, and
Stop sharing (collab.stop) revokes those on the server. Whoever joined with
them is disconnected at once and told sharing was stopped (the session's end,
with their copy); the owner stays in the file, which stays on the server, and
Start sharing gives the same file new links and codes. Removing the file is a
separate, confirmed step under Advanced.
The join field is the shared text input: as tall as the Join button beside it.
The Share popover is one flow: Online (the file is published to the server in
Preferences and the server names its links with short codes) or Local network
(the app hosts; the code carries its address and the session's secret), one
Start sharing button, the code large with Copy code and Copy link, who is here,
can edit / can view, and Stop sharing. Join with a code is one field in the
popover, on the home screen and in the Join dialog, for a code of either kind
or an invite link.
Protocol (within version 2, new variants last): Resolve / Resolved and Admin
code. The server keeps a stretched hash and the wrapped token, never the code,
and limits wrong codes per address (with a lockout) and for everyone. Older
servers leave the links in place of a code; older apps join by link.
collab.share, collab.join {code}, code and share in collab.peers; the server
preference defaults to the built-in server and can be reset or turned off.
Live sessions are per document instead of one for the whole app.
- **Engine** (`crates/engine/src/live.rs`): `Session.live` is now `Lives`: the app-wide parts (the
installed link, the invite-link registration) plus one `Live` per session. One place resolves
"the session of this document": `Session::live()` (the active document's; status off when it
has none), `live_mut()`, `live_of(uid)`. Sync, undo, read-only checks, follow, cursor chat,
spotlight and presence all go through the session of the document they are about, so edits and
presence never cross. Sessions behind the active tab keep syncing; your presence there is empty.
- **Link** (`LiveLink`): every call takes the `SessionId` the engine picked and `poll` returns
`(SessionId, Event)`, so one link carries several sessions. The desktop link and the browser
link keep them in a shared table with one mailbox per session
(`crates/collab-net/src/sessions.rs`); each hosted document listens on its own port.
- **Commands**: `collab.*` act on the active document's session; no parameters changed.
`collab.peers` keeps its shape and gains `session` (its id, null when the document isn't in
one) and `sessions` (every session: `{session, document, status, reason, role, fileName, link,
peers}`). `collab.join` returns the new session and opens its file in a tab of its own without
touching other sessions; until then (or if it fails) it shows on the tab it was started from.
Starting a session for a document replaces only that document's own session. Closing a document
ends its session only.
- **UI** (`crates/ui-egui/src/live.rs`): the Share popover, avatars, cursors, chat, spotlight
banner, follow frame and read-only reason read the session of the tab in front, per window. What
belongs to one session in the window (cursors shown, your chat box, a report being written) is
dropped when another tab comes to the front. `ui.inspect`'s `live` gains `session` and
`sessions`. No new strings or command ids.
- **Docs**: `docs/collab-protocol.md`, `docs/mcp.md`, `docs/control-protocol.md`,
`docs/collaboration-parity.md`, `docs/gaps.md` (gap 2), `ROADMAP.md`.
Third review:
- a broadcast counts once per recipient block, welcomes don't count
against block caps, and presence, cursor chat and pings have their own
per-address budget, so an office behind one address works together
- when everyone's reading budget runs out, the biggest readers are cut
off until it fits (their bytes freed at once), no cliff
- block counters are forgotten once only the table holds them
- install.sh: root's downloads, clone, swap file and built binary in a
root-only directory; the build user builds its own copy; the binary is
copied out with the build user's rights
- recommended open-mode settings in install.sh's config and the docs
Second security review:
- a broadcast counts once against the queued-bytes budget; each block of
addresses gets at most an eighth; when the budget runs out the biggest
consumer is shed, not the receiver
- forced compactions at most once a minute per file
- plugin downloads check the per-address share before the global budget
and before reading
- install.sh never replaces a Caddyfile with other sites (one import line,
once; install-test.sh in packaging lint); the built binary runs only as
the server's account
- bytes being read count against per-block and global budgets, the
biggest readers dropped first
- block multipliers x2 (/56) and x4 (/48), connection caps at most a
sixteenth/eighth of all; global publish and download budgets eight times
a /48's share
- problems capped at 200; writes refused below max(1 GiB, two files) free;
the private-server file cap documented
Fixes from the security review of the public server:
- team libraries only in key-published files on open servers; unpacking
bounded by the file cap (zip bombs refused)
- byte-bounded send queues per connection and in all, a bound on bytes
waiting for the hub, per-address and global download budgets
- batches checked against the file, per-address and total caps as files
grow (refused with a Reject the app rolls back); the core's recent-batch
memory bounded in bytes; nothing confirmed when a log can't be written
- usage reports kept in memory; per-kind message caps checked while read
- layered IPv6 limits (/64, /56, /48), a global publish rate, connections
reserved for sessions, short idle timeout for one-shot requests
- revoked links and removed files cut library connections
- data directory 0700, server key 0600 from creation
- install.sh: SSH ports from sshd, ufw left off when other services
listen, unprivileged builds of the latest tag, settings checked before
use, own Caddy snippet; systemd memory and network caps; backups check
free space and say they hold the key
- linger reads raw bytes after an oversized frame (tungstenite asserted)
Config-gated (open_publish): anyone publishes without the server key and
gets an owner token (128-bit, shown once, stored hashed) that manages only
that file. Per-address limits (X-Forwarded-For only from trusted_proxy,
bounded tables), global caps, idle expiry, reports, persisted blocks,
files list --sort, check, GET /healthz on the same listener, startup
warnings. The app asks the server first, publishes without a key when it
is open, shows the owner token once, the host/terms line, the expiry date
and a report button; collab.serverInfo, collab.manageFile, collab.report
and pref.publishServer for MCP/CLI. Protocol additions stay within
version 2, gated like signals.
- **Installed plugins survive a crash.** Importing, installing, updating or removing a plugin, and a
plugin changing its settings, now sets `Runtime::unsaved`; at the end of each frame the app hands
the preferences to a new `Services::save_prefs`, and the desktop app writes `state.json` at once
(temporary file + rename, the same atomic write the quit path uses, now shared). Nothing waits for
quit any more.
- **The UI never waits on the network.** New `Session::start_background(id, params)`: the command
runs as usual until it makes a server request; the request goes out on the link's own thread
(`LiveLink::background`, implemented by the desktop `NetLink` with one short-lived thread per
request) and the command stops; when the answer arrives (`Event::Net`, taken by `live_poll`, which
the network's waker already triggers) the command runs again from the start on the file it
started on, answered from the answers so far, until it needs no new request. All server requests
go through one funnel (`Session::net`), so `plugin.directory.*`, `library.publishToServer`,
`library.add` (links), `library.usage` and library fetches work this way without being split.
A re-run of `library.publishToServer` caches the library it actually sent (a fresh copy differs
in its time and, on a first publish, its id). One job per command; `cancel_background` drops a
late answer.
- **UI:** `FigApp::run_background(ctx, id, params, done)` with a boxed completion. Manage plugins >
Browse shows "Fetching from the server..." with Cancel instead of Browse; the install confirmation
shows "Installing..." with Cancel and stays open until the download is in (an error leaves it open
to retry); row buttons wait during an install; closing the window cancels. Assets > Libraries:
Publish to server shows "Publishing..." + Cancel, Show usage and pasting a team library link show
"Fetching from the server..." + Cancel. No polling repaints: the window draws again when the
answer's waker fires; pending state is a scan of a short Vec, no per-frame allocation.
- **Programmatic callers are unchanged:** MCP, CLI, control channel and scripts still run the same
commands synchronously (and the browser link, which has no background requests, too).
- New string "Installing..." (en/es); docs: plugins.md (immediate saves, Browse/Install never wait),
server.md (team library requests never wait).
A self-hosted plugin directory, with no third-party service: `figcraft-server` keeps plugin bundles,
the app browses, installs and updates from it, and `figcraft-cli` publishes to it.
- **Bundles** (`figcraft_plugin::bundle`): a plugin's manifest, module and up to 32 plain-named
assets in one uncompressed, length-prefixed binary form (nothing can expand beyond the bytes
sent). Its SHA-256 names it. `Bundle::decode` never allocates beyond its input and refuses
trailing bytes; `Bundle::check` parses the manifest (refusing a plugin API this build doesn't
run), requires a version of 1-4 numbers, compiles the module without running it (only the
sandbox's imports allowed) and checks asset names (no paths, no dot files, no Windows device
names, unique ignoring case). Manifests gain an optional `author`.
- **Protocol**: a new connection kind next to managing and team-library connections:
`OpenPlugins {key}` then `Plugins {list | fetch | publish | unpublish}`. Browsing needs no
credential; publishing and unpublishing need the server key (a wrong key is refused). Older
servers refuse the connection and the app says the server has no plugin directory. The protocol
version is unchanged; old clients are unaffected.
- **Server** (`apps/figcraft-server/src/directory.rs`): `<data>/plugins/index.json` plus
`<data>/plugins/<sha256>.bin`, so no publisher-chosen name reaches the file system. Every publish
is checked in full and must be a newer version; caps `max_plugins` (versions, 0 turns it off),
`max_plugin_mb` (<= 32), `max_plugins_mb`. Fetches re-hash the file before sending (a damaged
bundle is never sent) and share a download budget (~32 MiB/s) so a flood is refused, not queued.
Damaged index entries are skipped and reported, never deleted; an unreadable index lists nothing
and refuses writes until the operator fixes it. Operator moderation:
`figcraft-server plugins [list]` and `plugins remove ID [--version V]` (offline or through the
running server). Logs name ids and versions, never keys.
- **App**: Plugins > Manage plugins gets Installed / Browse tabs. Browse lists a server's plugins
(version, author as typed, size, access) with a search; nothing is fetched until Browse is
pressed. Install / Update open a confirmation listing the permissions asked for, the source and
the SHA-256; only its button downloads. The download must match the listed SHA-256 (and the one
shown), then is checked like an imported plugin before anything is written. Installed plugins
live per user under `<settings>/FigCraft/Plugins/<hash of id>/` and run in the same sandbox;
Remove deletes the folder.
- **Commands**: `plugin.directory.list {server?, query?}`, `plugin.directory.install {id, version?,
confirm, server?, sha256?}` (refused without `confirm: true`; never opens a dialog),
`plugin.directory.publish {path | plugin, server}`, `plugin.directory.unpublish {id, version?,
server}`. `ui.set {pluginsTab, pluginConfirm}` on the control channel.
- **CLI**: `figcraft-cli plugin publish FOLDER --server ws://host:port/` with the key from
`FIGCRAFT_SERVER_KEY` (or `?key=`), never printed.
- Docs: plugins.md (directories, publishing, commands, manifest), server.md (operating the
directory: access, checks, caps, moderation, backups, security), collab-protocol.md, mcp.md,
cli.md, control-protocol.md, target-app-parity area 18 and Ecosystem, ROADMAP, roadmap.md M11,
gaps.md gap 6. Plugins & widgets 55%/32% -> 60%/35%, 10-30 h -> 8-26 h.
Two presence features for live sessions, with no accounts needed:
- **Cursor chat.** In a live session, `/` on the canvas opens a bubble in your colour at your pointer. What you type goes to everyone as you type, and others see it next to your cursor in place of your name tag. Enter (or clicking away) sends it and closes the box. On every screen it fades a few seconds after its last change (5 s, then a 1 s fade; with reduced motion it vanishes without fading). Escape takes it back at once and doesn't deselect. A box left open without typing closes by itself. Limits: 100 characters, control characters become spaces. Agents use `collab.cursorChat {text}`.
- **Spotlight.** Your avatar (top of the right panel) now has a menu with Spotlight me / Stop spotlighting (`collab.spotlight {on?}`). Everyone else follows your view and page, using the existing follow. People who join later are told by the server and follow too. Anyone who pans, zooms or presses Escape stops following, and a canvas banner ("Hana is spotlighting", Follow) lets them follow again. A newer spotlight takes over the old one. The presenter leaving or stopping ends it. The presenter sees a banner with Stop. After a reconnect, your spotlight goes out again.
- **Protocol: signals.** New `ClientMsg::Signal {kind, text}` and `ServerMsg::Signal {client_id, signal}` carry ephemeral presence. Signals never go into the document, the log or a saved state; the server keeps only who is spotlighting, for newcomers. The server cleans each signal: unknown kinds, malformed spotlights, a client's `ready` and anything over 1 KiB are dropped, and chat is capped and has control characters replaced. The app checks what arrives again before showing it. The hub rate-limits signals per connection (20 a second, bursts of 40) and drops the excess without closing the connection.
- **Compatibility, both directions.** A client that reads signals announces it with the word `signals` in its hello's `client_version`, and the server sends signals only to those clients, so older clients never see a message they can't decode. A server that takes signals answers with a `ready` signal after the welcome, and the client sends signals only after that, so older servers never get one; the app says the host runs an older version. Signals are kind plus text, so clients skip kinds added later and servers drop them.
- Engine: `Live` keeps `signals`, `chats`, `chat`, `spotlight` and `spotlighting`, all reported by `collab.peers`. UI: chat bubbles, the chat box, the spotlight banner (the existing canvas banner widget), the avatar menu and `ui.inspect` fields. Keymap: `/` maps to `collab.cursorChat` (Essential). Spanish labels for both commands, plus en/es UI strings.
- Docs: collab-protocol (messages, compatibility, Presence, security limits, agents), mcp.md, collaboration-parity rows, target-app-parity, ROADMAP, gaps, ui-parity.
The desktop app joins wss:// sessions (rustls with webpki roots, certificates always checked, clear errors; FIGCRAFT_CA_FILE for private authorities), with one TLS session shared by the reader and writer threads. figcraft-server stops gracefully on SIGTERM/SIGINT/SIGHUP/Ctrl+C, ships in Linux x86_64/aarch64 release archives, and docs/server.md gains a Deploy on a Linux server guide with a hardened systemd unit, Caddyfile, nginx config and Dockerfile in packaging/server.
A file's library can be published to figcraft-server and enabled in other files by a share link (view access is enough); placing, styles, variables and library updates with overrides kept work as for local libraries, with a background check while the Assets panel is open. library.swap re-points instances, styles and variables to another library by name (one undo step, unmatched items listed). The server keeps per-file usage reports (library.usage).
figcraft-server hosts many files over the same hub the app uses for sessions, persists each as a checksummed snapshot plus an append-only log (synced before replies, compacted, damaged files skipped and reported) so restarts lose nothing confirmed, and gives per-file share links with edit or view access (hashed 128-bit tokens, revocable). The app can publish a file to a server and join server-hosted files; view-only tabs refuse edits. Protocol version 2.
- Browser client (step 4): the web build joins sessions a desktop app hosts
through the browser's WebSocket API (web-sys, binary frames, the same
message encoding). Its connection logic is a target-independent state
machine (collab-net driver.rs): hello on open, changes and presence held
until the welcome, throttled presence, keep-alive pings, a silent host
counted as dropped, reconnect with backoff, final refusals. Tested natively
against a scripted socket and a real host over loopback. Hosting stays
desktop-only (LiveLink::can_host; collab.peers canHost). https pages are
told up front that they can only reach ws:// sessions on this computer.
- Shared join logic (join.rs: GuestConfig, outcomes, taking in host messages)
and transport limits (limits.rs) used by the native guest and the browser.
- Invite links: `figcraft figcraft://join?...` joins on start (a bad link
opens the Join dialog with it); file:// arguments are paths. Windows
registers a per-user figcraft:// opener (HKCU\Software\Classes\figcraft)
behind a Registry trait (tests use a fake): collab.registerLinks {on?} and a
Preferences checkbox, default on for installed builds, never for builds
from source. macOS Info.plist declares CFBundleURLTypes; the Linux desktop
entry declares x-scheme-handler/figcraft and passes %U.
- Join dialog: Escape closes it and no longer reaches the canvas (selection
kept); its field has the keyboard when it opens.
- Docs: collab-protocol (browser client, links, mixed content), mcp,
control-protocol, README, web hosting notes, collaboration wording in the
progress docs (numbers unchanged).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* M9: collaboration: per-user undo, rejections, reconnect
Step 3 of docs/collab-protocol.md, the parts that don't need the transport:
- Per-user undo and redo (figcraft_collab::Client::{undo, redo}): each undoable batch is a step
(values before/after by target, nodes created/deleted). Others' batches ordered after the step
take its targets for good; creates invert to deletes (not when others built inside), deletes
re-create the subtree. An undo or redo waits for every earlier batch and is remade from the
confirmed state when it goes out; derived data follows as a follow-up batch.
- Rejections: refused batches leave the history (with an undo waiting on them) and queue a
Notice; the engine shows it through Session::message (the UI's toast).
- Offline queue and reconnect: Hello.resume {session, client_id, seq}; the server keeps a
bounded log and answers with the missed batches or a snapshot, plus what it did with the
client's batches (processed {last_batch, refused}); queued batches then go out.
Server::state / restore persist the session (client-id counter, log, session id).
- Presence carries editing_text {node, caret, selection}.
- Engine: documents in a live session (LiveDoc, Session::collab_*): commands, gestures and
groups sync as one step each; edit.undo / edit.redo route to the per-user history; remote
edits show under a gesture in progress. Outside a session snapshot undo is unchanged.
- Sim: dropped connections, undo/redo, and a checker that no undo reverts a change by someone
else its author had seen. New tests in undo_tests, reconnect_tests, collab_live_tests,
collab_undo_tests.
- CLAUDE.md / AGENTS.md: collab in the L2 layer list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M9: collaboration wording in the progress docs after per-user undo and reconnect
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M9: protocol doc: what the engine's live-session tests cover
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* M9: real-time collaboration protocol (design)
docs/collab-protocol.md: WebSocket sessions with a server-ordered,
per-property last-writer-wins document sync, client-scoped ids,
fractional positions for children, ephemeral presence (cursors,
selections, follow), per-user undo, hosting from the app first and a
self-hosted server next, and the implementation plan. ROADMAP gains M9.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M9: collaboration core: ids, change sets, server ordering, convergence
Step 1 of docs/collab-protocol.md, no networking:
- Client-scoped node ids: Document::set_client_id makes alloc_id hand out
client << 32 | counter; single-user files allocate as before; the id
counter is no longer part of document equality.
- crates/collab (layer 2): the synced object/property model (every saved
Node field, plus document-level styles, comments, variables, images,
dev snapshots, library data), Change batches, an O(changed) diff, strict
and lenient apply, fractional sibling positions, snapshots.
- Server (ordering, validation, rejects with fixes) and Client (optimistic
view, replay under remote batches, author-side repair of derived data)
as pure state machines; binary (version byte + postcard, JSON values) and
JSON message forms, fuzzed decoders.
- Engine: normalize as the recompute hook, collab.diff.
- Tests: round trips over random real commands, an N-client convergence
fuzz test over delayed in-order links, junk batches.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M9: fmt after merge
* M9: live sessions: host from the app, invite links, live cursors and presence
Step 2 of docs/collab-protocol.md, networking and presence half:
- crates/collab-net (layer 6): WebSocket transport (sync tungstenite, a
reader and a writer thread per connection, no async runtime), the
session host the app embeds (runs figcraft_collab::Server as the
authority, carries its messages), invite links, presence throttling
(~30 Hz, latest wins), limits (sockets, hello/idle timeouts, message
size, rate, slow clients), keep-alive, clean close, reconnect with
backoff; a browser stand-in until the web-sys client (step 4).
- engine: live session state (status, peers, presence, follow) and the
collab.host/stop/join/leave/peers/follow/cursor commands; the network
is reached only through the LiveLink the app installs.
- ui-egui: remote cursors (pointer drawn in code, name tag, gliding),
selection outlines (once documents sync), avatars in the right panel
with follow, the Share popover (start, invite link + URL with Copy,
stop), File > Join live session.
- desktop app and MCP headless session install the link; xtask layers
keeps networking crates out of L0-L5.
- Documents don't sync between apps yet (follow-up).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M9: live sessions: document sync between apps
- The host's session runs figcraft_collab::Server seeded with the shared
document; every app (the host's too, through the hub without a socket)
is a collab Client. Guests open the host's file from the welcome as a
new tab; leaving keeps it as a local copy.
- Engine: after each local action (a gesture once it ends; nothing during
version or text previews) the session document is diffed into one
Changes batch; server batches go through Client::receive with
EngineRecompute and are installed with Session::apply_remote, which
never touches the undo history. Batches arriving mid-gesture wait.
- Transport: Changes/Reject/Welcome reach the engine as Event::Sync;
LiveLink::send carries batches in order; hellos are limited to 64 KiB,
joined clients to the core's message size.
- Tests: engine sync against a real Server (ids in the client's space,
one batch per gesture, remote edits skip undo, deferral, closing the
file leaves), three apps over loopback converging (create, move,
recolour, delete, rename) and a late joiner getting the server's state.
- Docs: collab-protocol, mcp, README, roadmap docs; the UI's
"cursors only" note is gone, replaced by the per-person undo caveat.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
bflatastic