v0.3.9

Fixed

• The readable-chat mode is now a standing choice, not a button that
  appears in one state only. Removing the last allowed chat correctly
  left an empty allowlist — agents able to read nothing — but the way back
  to "every chat" was a plain button at the foot of an empty list, easy to
  miss at exactly the moment it mattered. Both modes are now always on
  screen with the current one marked, so leaving the empty state never
  depends on finding a control inside it, and the warning banner names the
  way out rather than only the problem.

v0.3.8

Added

• Limit which chats agents may read. Connecting an MCP client used to
  hand it every chat on the account. setup now asks what connected
  agents should see — every chat, only your own self-chat, or nothing
  until you say so — and the choice is editable afterwards from the
  dashboard's clients tab or the new scope command
  (--allow/--deny/--all/--list).

  Once a limit is on, chat-addressed reads outside it are refused and say
  the chat is out of scope rather than pretending it does not exist, so a
  model stops asking instead of retrying. Questions asked across all chats
  — global search, poll_new_messages, call history, contact search —
  come back filtered; contact search included, because an agent that can
  still enumerate your address book leaks exactly what the limit exists to
  withhold. Enforcement wraps the store once rather than at each call
  site, so a read tool added later cannot forget it.

  The limit is one setting for the whole server, not one per client: every
  client authenticates with the same bearer token, so the server cannot
  tell them apart. Only you can change it — no MCP tool writes
  config.json, the same rule the trust list already follows. And an
  empty allowlist means nothing is readable rather than everything,
  which is why the mode is stored separately from the list; check names
  that state, since every read tool then returning empty looks exactly
  like a broken install. The dashboard is deliberately exempt: it is your
  own window onto your own messages.

• A clients tab in the dashboard. Every MCP client detected on this
  machine, whether it is installed, whether this server is added to it,
  and where its entry points — with one click to add or remove that entry.
  A config the doctor calls broken can now be fixed without dropping to a
  terminal. The broken flag repeats the doctor's own verdict rather than
  inventing a second opinion.

Fixed

• Pairing a second number no longer shows the first number's data.
  Everything that belongs to a WhatsApp account — its messages, downloaded
  media, backups, trust list, readable-chat allowlist and pending
  scheduled sends — now lives under accounts/<number>/, and only the
  machine's own settings (rate limits, outbox roots, the HTTP token) stay
  shared. Switching numbers is clean in both directions: pair the first
  one back and its history and settings are exactly as they were.

  The worst case this closes was not the visible one. A pending scheduled
  send sat in a single shared file, so a message scheduled under one
  number would have fired from whichever number happened to be paired when
  its time came — a real message, to a real person, from an account they
  did not expect.

  Separation is by file rather than by a column every query has to
  remember: two accounts cannot see each other because their rows are not
  in the same database, including from code written long after this. The
  upgrade moves an existing install into the new shape by renaming files,
  never rewriting rows — deliberately, because adding an account column to
  the existing tables would mean rebuilding them, and that renumbers
  SQLite rowids, which here are both the full-text index's link to its
  content and the cursor poll_new_messages hands out. Renumbering them
  would have silently corrupted search results and made agents skip
  messages they never received.

Changed

• trust and scope take a name or a phone number, not just a JID.
  Nobody knows their friend as 15551234567@s.whatsapp.net, and the old
  failure mode was silent: a mistyped JID was accepted, stored, and simply
  never matched anything. Both now accept a contact or group name, a phone
  number in any of the shapes people paste, or a JID. A name matching
  several people is never guessed at — the CLI lists the matches and asks,
  the dashboard shows them as a chooser, and nothing is written until one
  is picked. The same goes for the dashboard's trust and readable-chat
  boxes.

v0.3.7

Fixed

• Reactions now sit on the message they react to, as a small pill on the
  bubble's lower edge, instead of floating as their own [reaction]
  bubbles — the way the real app shows them. One pill per person (a newer
  reaction from the same person replaces their older one, and an un-react
  removes it), with a count when several people used the same emoji.
• Images opened from the chat pane now scale to fit the screen. The
  lightbox sized the image against a content-derived box, so a tall or
  large picture overflowed the viewport; it is now bounded to the viewport
  in both dimensions and always fits.

v0.3.6

Fixed

• The chat pane now shows a chat's newest messages again. It paginated by
  database rowid (insertion order), which assumed rowid order matched time
  order — but a chat can hold a history-sync backfill whose rows were
  written long after the moments they record, and the LID fold in 0.3.4
  merged chats whose rows interleave in time. The result was an active
  chat opening on months-old backfilled messages instead of today's. The
  pane (initial load and in-place refresh) is now ordered by timestamp.
  No data was lost — only the view's ordering was wrong. The
  poll_new_messages MCP tool keeps its rowid cursor, which is correct
  for an agent reacting to newly ingested messages.

Added

• Scroll up in a chat to load older messages — the pane fetches further
  back a page at a time and keeps your place as it grows, instead of only
  ever showing the newest ~50.
• An "↑ older" control in the chat header asks your paired phone for
  messages from before the oldest one stored — the way to pull a chat's
  history in from scratch, one page at a time. It requests your own
  history from your own phone (it messages nobody) and is rate-limited by
  two cooldowns — one per chat, one global — so a held button or a script
  can't turn into the burst of history requests that risks a number being
  flagged; a cooldown shows as a short "try again in Ns" note, not an
  error.

v0.3.5

Added

• Message search in the dashboard now works the way a chat app's does:
  results arrive as you type (no button), listed in the chat list's
  place with the chat's name, who said it, a snippet with the matches
  highlighted, and when. Clicking a result opens that chat parked on the
  message — scrolled to it and ringed — with a "latest" control to return
  to the live tail. With a chat open, results can be narrowed to it.
• Pictures in the chat pane open full-size on click, brought forward over
  the blurred page; click outside, the close control, or Escape returns.

Fixed

• The pair tab no longer shows the "Start pairing" button alongside the
  linked panel when this server is already paired (clicking it only ever
  produced "already paired").

v0.3.4

Fixed

• One person is one chat again. WhatsApp increasingly addresses people by
  a privacy LID (...@lid) instead of their phone number, and the same
  contact could show up twice in the chat list — a phone-number chat with
  the history and a nameless LID chat with the newest messages — with the
  LID half reading as user …12345. Every LID the server can pair with a
  phone number (from the message itself, or from the pairing table
  WhatsApp syncs to the device) is now keyed on the phone number as it is
  stored: chats, senders, contacts, read receipts and calls alike. What was
  stored split by earlier versions is folded together on the next start,
  and again after every history sync, so pairings that arrive late still
  take effect. LIDs WhatsApp never pairs stay as they are.
• History-sync stubs — the status lines WhatsApp keeps in a chat with no
  message behind them — no longer become empty [other] rows, or the
  phantom LID chats made of nothing else. Existing ones are dropped on the
  next start.

v0.3.3

Fixed

• serve --http no longer gives up after one failed connect at startup.
  A service started at login often beat the network to it: with DNS not
  yet answering, the single connect attempt failed, and since WhatsApp's
  auto-reconnect only covers a connection that once existed, the process
  sat paired-but-offline until someone restarted it — with the dashboard
  reading "connecting" the entire time. The startup connect is now
  retried with a doubling backoff (5s up to a minute) until it lands, each
  failure logged with the wait before the next attempt, and a failed
  attempt now reports offline rather than a permanent connecting.

v0.3.2

Added

• Refresh buttons on the dashboard's chat screen: one fetches only the
  messages that arrived since the pane was loaded (appending them in
  place, following the scroll only if you were already at the bottom)
  and one reloads the chat list. No background polling — updates load
  on demand.
• Profile pictures in the dashboard's chat list, served through the
  server itself (the page still makes no external requests). Lookups are
  cached for a day — including the "no picture" answer — and live
  fetches are rate-capped, so rendering the list never turns into a
  burst of WhatsApp queries. The initial-letter avatar remains the
  fallback whenever no picture is available.
• Media in the dashboard's chat screen: images and stickers preview
  inline (lazily, so only what scrolls into view is fetched), and every
  other attachment gets a download link. Files are served from the same
  per-chat media directory download_media uses, so each file is
  fetched from WhatsApp at most once, however many times it is viewed.
  Content types are verified server-side: only raster images ever
  render inline, and everything else — including SVG and HTML — comes
  down as an opaque attachment, so WhatsApp-supplied content can never
  run script on the dashboard's origin.

v0.3.1

Fixed

• The dashboard UI is no longer cached by the browser: after an upgrade,
  a previously opened dashboard could keep showing the old interface
  (e.g. the pairing controls while already paired) until a hard refresh.
  The Pair tab also renders the correct linked/unpaired panel instantly
  from the last known state, and "Paired." only announces a pairing you
  actually watched complete.

v0.3.0

The personal-robustness release: the server heals itself after a
WhatsApp-side logout, Windows gets service support, messages get a backup
command — and a local web dashboard arrives, with QR pairing in the
browser, WhatsApp-style chat browsing, trust/schedule/draft management,
and one-click unlink.

Added

• New logout command and an Unlink button on the dashboard's Pair tab:
  both sign this device out on WhatsApp's servers (the phone stops listing
  it under Linked devices), keeping local messages and settings, and reset
  to unpaired so a new QR pairing can start immediately. remove remains
  the local-only variant that never contacts WhatsApp.
• Chats with no stored name now resolve a display name through the
  contacts table and lid mapping — in the dashboard, in list_chats/
  get_chat, and in send previews — and the dashboard falls back to a
  readable "+number" instead of a raw JID. Empty-bodied messages (system
  events, removed reactions) render as dim kind placeholders instead of
  blank bubbles, and a signed-out dashboard now says so in its header
  instead of freezing on the last live state.
• The dashboard now shows designed signed-out and not-found pages that say
  how to get back in (run whatsapp-connect-mcp dashboard for a fresh
  login link), unknown API paths answer with a JSON 404, and a browser
  opening the bare server address is redirected to the dashboard.
• The dashboard's chat window now looks and scrolls like WhatsApp Web's
  dark theme: independently scrolling chat list and message pane, message
  bubbles with in-bubble times, day dividers, and per-sender colors in
  groups. The whole dashboard adopts WhatsApp's dark palette, with
  loading skeletons and clearer action feedback throughout.
• New backup command: writes a consistent snapshot of the message
  database to <data-dir>/backups/ (or --dest), safe to run while the
  server is up. Message history is the one thing re-pairing cannot
  recover.
• service install|uninstall|restart now works on Windows, using a Task
  Scheduler logon task. The server runs as a minimized console window;
  closing it stops the server. Creating the task may require an
  Administrator terminal.
• Connection-health diagnostics: dropped connections, keepalive timeouts,
  refused connects, temporary bans, and an outdated client now each print
  one clear line to the server log instead of passing silently.
• Failed writes of incoming events to the message store are now counted
  and reported (previously silent).
• check (and the doctor tool) gains an ingest finding that fails when
  incoming events are not reaching the message store, and the session
  finding now says when the server was logged out by WhatsApp mid-run.
• status now shows connection state, last event time, reconnect count,
  and ingest-failure count.
• A local web dashboard on the shared HTTP server: open it with the new
  dashboard command (prints a one-time login link). This first slice
  shows connection health, store counts, and doctor findings; pairing and
  management arrive in following changes. Localhost-only, protected by
  the same bearer token as the HTTP transport.
• QR pairing from the dashboard: when the session is unpaired, the Pair
  tab shows the QR code in the browser — no terminal needed, including
  re-pairing after a WhatsApp-side logout.
• Dashboard: read-only chat browsing and full-text search. Media messages
  show as placeholders; the dashboard never fetches media.
• Dashboard: trust-list management (applies to a running server
  immediately), scheduled-send list/cancel, and a one-click backup.
• Dashboard: pending send drafts can be reviewed, approved, or discarded.
  Approving delivers the exact content the agent drafted (rate-limited as
  always); a draft can be committed at most once no matter how the
  dashboard and the agent race.

Changed

• serve --http now starts its listener before waiting for pairing, so
  the server is reachable while unpaired: store-backed read tools serve
  local data, and live/send tools report "no longer paired" until pairing
  completes. stdio serve still waits for pairing before starting.

Fixed

• A WhatsApp-side logout no longer leaves a running server permanently
  broken: the server now says clearly that it was logged out, waits, and
  reconnects on its own as soon as the device is paired again — no restart
  needed. Tool calls while unpaired fail with "no longer paired — run
  setup" instead of a generic connection error.
• A session-store failure while waiting for pairing (at startup, or after
  a logout) is no longer mistaken for a clean shutdown: it now prints a
  diagnostic and exits non-zero at startup, so a service manager restarts
  the server, and after a logout it says the re-check failed instead of
  going silent.
• Dashboard: a failed pair start, trust change, schedule cancel, backup,
  or draft approve/discard now shows a message in that tab instead of
  failing silently.
• Dashboard: starting pairing only accepts POST, matching every other
  mutating dashboard route.