Ovis v0.4.20

New features

• Chat reads Letta Code conversations.
• Start Claude Code with Ultracode on, at your chosen effort level.
• Sign in again to an expired Claude or Codex account from usage.
• Give Ovis a pure black background with the new Black theme.
• Send Claude Code's existing draft without losing your Chat message.

Improvements

• The terminal draws whole screens without mixing old and new output.
• A sent Chat message appears at once while the agent takes it.
• Chat tabs reopen faster, with a choice of history page size.
• Long OMO and GJC chats refresh faster while the agent works.
• Terminal text uses the same coding font on Mac, Safari and Chrome.

Bug fixes

• Korean typing in the iOS app is fixed by the PC update, without a new app build.
• Chat keeps following new output when work blocks fold.
• Claude Code Chat follows the conversation after branching.
• Slow file previews no longer freeze every terminal.
• Windows app updates recover from an old server left running.

Full changelog

Added

• In New workspace and New tab, Claude Code's Ultracode switch fits on one line under Model and
  Effort, with its minimum version visible on a phone too.
  (#2282 by @fivetaku)
• The server lists every agent Ovis can install, with each vendor's own install command for the
  PC's operating system and whether the agent is installed there, at which version
  (GET /api/agent-catalog, also for remote PCs). Nothing is installed yet: the install flow
  builds on it.
  (#2326)
• Add Black, with pure black backgrounds and the Tokyo Night accent colors.
  (#1911 by @jjsteffen)
• When Chat does not send a message because Claude Code's input box already holds a draft, the
  note now has a Send it button beside Clear it and send. It sends the draft as Claude
  Code holds it, and your Chat message stays in the box.
  (#2301 by @jpcarranza94)
• Chat shows a Claude Code bash-mode command (!) as one card: the command, then what it
  printed, with the output of a failed command in red. Before, Chat showed the raw
  <bash-input> and <bash-stdout> tags in two messages.
  (#2176 by @jpcarranza94)
• New workspace and New tab have an Ultracode switch for Claude Code, on one line under Model and
  Effort, so a session can start with Ultracode on at any effort level
  (--settings '{"ultracode":true}', which a Windows PC passes as a small settings file in
  %LOCALAPPDATA%\ovis\claude-settings). It needs Claude Code 2.1.284 or later, a model with xhigh
  and dynamic workflows on. Reported by @fivetaku, whose #2282 suggested the compact line under the
  effort level.
  (#2266)
• Claude Code's /effort card has a Turn Ultracode on or Turn Ultracode off row where Tab, the
  default key, flips the switch: it switches Ultracode for this session and keeps the effort level.
  (#2266)
• Chat shows a Letta Code pane's conversation, read from Letta's own transcript, when herdr's Letta
  integration reports it and Letta's own record names the same conversation; otherwise the
  terminal's text stays.
  (#2312)
  (#2338)
• The ask_user form of the @eko24ive/pi-ask pi package shows as a chat card and is answered
  from the chat, one question per step, once the web UI's pi extension is installed with
  bun scripts/install-pi-integration.ts. The answer goes to pi-ask through its own events,
  never as typed keys; without the extension the form stays in Terminal (#1202). Reported by
  @chobijaeyu.
  (#2218)
• An expired Claude or Codex account in Subscription usage has a Sign in again button, in Settings and in the
  usage popover: it opens the official sign-in in a terminal tab for that same account, replacing
  its expired sign-in instead of adding a new account.
  (#2309)

Changed

• A message sent from Chat shows at once as your own bubble, faded, until the agent writes it as a
  turn, so the chat is never blank between Send and the agent's turn. A small clock joins it only
  when the agent has not taken it after 5 seconds. A command (/cost, ! ls) gets no bubble, since
  it never comes back as that message, and a bubble whose turn has not come within a minute goes.
  (#2270 by @CharleoYang)
• New workspace and New tab no longer list ultracode among Claude Code's effort levels; the PC
  still accepts --effort ultracode from other clients.
  (#2266)
• Switching tabs is faster: a tab's terminal takes input about 100 ms sooner. The bridge passes an
  attach's screen on as soon as herdr draws its text, instead of holding every attach 100 ms.
  The hold stays for an attach that shows no text yet, one that was resized while it started, and
  a pane another web bridge held, so herdr's refusal is still never shown.
  (#2255)
• A long OMO or GJC chat refreshes faster while its agent works: each update now parses only the
  records the agent added since the last one, as OMP, Claude Code and Codex chats already do,
  instead of the whole running task again. Based on
  #1634 by @Yoonwoo-Ha.
  (#2291)
• The Beta page on ovis.sh tells Android testers, in all four of its languages, to open the
  opt-in link in Chrome on the phone rather than inside another app, that You are a tester
  means they are in, to check that the Play Store app's profile is the same account, and that
  Google Play takes a few hours after they join the group, so an app that is not found right
  after joining is worth trying again later.
  (#2076)
• The terminal draws its Latin text in the bundled JetBrains Mono (regular and bold) on every device,
  so Ovis for Mac, Safari (the installed app) and Chrome on the same Mac show the same font, row height
  and number of rows. A font chosen in Settings still takes its place.
  (#2272)

Fixed

• A message to Claude Code is sent at once when its own turn has ended and only background
  agents are still working, instead of waiting for those agents to finish.
  (#2271 by @CharleoYang)
• Switching Chat tabs loads a smaller page of history and keeps more tabs' pages ready, instead
  of fetching the whole page again at each switch. Settings lets you choose the history page size.
  (#2274 by @CharleoYang)
• A terminal resized while it attaches no longer shows its first screen at the old size, or takes
  input before the program has the new size. The bridge now waits for herdr's redraw at the
  terminal's current size, also after a resize back to the size it started at, for up to a
  second.
  (#2324)
• A folded work block no longer opens again when hidden reasoning before it joins the same run of
  work (#2111).
  (#2240)
• A chat message sent to a PC whose server cannot confirm messages and does not name its herdr
  now stays in Recover unsent text after a reload, instead of being lost with the page (#2172).
  (#2240)
• A diagonal trackpad swipe over an overflowing tab strip is left to the browser, so the strip
  no longer moves against the swipe's sideways motion (#2103).
  (#2240)
• Chat keeps following new output to the end when part of the transcript folds on its own: a turn
  settling at the end, or a block above the view folding while the view still glides to the end.
  Before, the list holding the rows in place read as the reader scrolling up: following stopped
  mid-transcript and New messages showed.
  (#2117)
• The numbers of a numbered list in the chat are no longer cut off on the left with a wide Chat font
  such as a monospace one, or once a list reaches 10 or 100: the indent grows with the Chat font
  and the widest number.
  (#2096)
• A chat work block you folded no longer opens again when the words before its tools join it, as
  when a later Codex record marks them commentary or a blank commentary is filled in.
  (#2098)
• A chat copy that reaches past the chat, such as Select All, keeps a tool card's whole file paths
  instead of the short labels read from the pane's folder.
  (#2195)
• A pane on a UNC folder written with forward slashes (//server/share/app) shows its chat tool
  card paths as written, as it already did for \\server\share\app.
  (#2195)
• A message queued for a claude attach pane waits for the conversation of the job it opened, as it
  does for any Claude Code pane, instead of going wherever the pane ends up.
  (#2119)
• Chat follows a claude attach <name> pane whose name is followed by debug options or
  --setting-sources, as Claude Code reads it.
  (#2119)
• A /resume of a background session no longer reads that job's state without a size limit, nor
  waits on a FIFO put in its place.
  (#2119)
• A Claude Code pane whose turn has ended no longer reads as working while a background subagent it
  started works on: Claude Code keeps its busy title spinner up for the subagent, which herdr reads
  as work. The pane now waits on that work (BG), as it does when the subagent is quiet, so the chat
  no longer shows "Working…" with a Stop button under the finished answer. (#2069)
  (#2072 by @Noobear)
• After Claude Code's /branch, Chat takes a transcript as the branch only from its own first message,
  and keeps to the branch file it checked when another project holds a file of the same id.
  (#2126)
• After Claude Code's /branch, Chat follows the branch Claude went on in and shows its replies,
  also when Herdr still reports the session the branch was copied from. Reported by @Hyd3-14 in
  #2020.
  (#2099)
• A Claude skill or command sent from Chat that asks a question while Claude's fullscreen diff
  panel is shown now waits on that question in its receipt, as it does without the panel.
  (#2199)
• After /model or /effort in a Claude Code pane, the composer's model pill shows the new model
  and effort as soon as Claude Code confirms the change, not after the next answer. A model name
  printed in a form that does not map back to one model id keeps the old pill until that answer.
  Thanks to @jpcarranza94 for the report and the list of forms.
  (#2305)
• In a Claude Code pane, the composer's model pill and effort no longer go back to an older model
  when many records with no answer (bash-mode commands, a long tool output) follow the answer or
  the /model / /effort change that set them.
  (#2319)
• Large files opened from Claude Code's scratch folder keep their size (and download progress)
  when many are fetched at once, instead of some arriving without a length.
  (#2317)
• Chat no longer shows an OmO tool's output as a big message from you when OmO runs Claude Code
  through its SDK (a Tool result (mcp__custom-tools__eval, id=…): bubble between work blocks).
  The output now folds into the work block as that call's result, and the call no longer counts as
  failed while OmO runs it. A message you sent along while the tools ran still shows as yours.
  (#2310)
• Chat no longer shows OmO's own notices as messages from you when OmO runs Claude Code through its
  SDK: reminders, monitor events, recalled memory notes, goal continuations and the history a new
  session replays stay hidden, like Claude Code's own reminders. A background task that ended shows
  as its task card, and a team member's message as a message from another session.
  (#2322)
• A message queued for a Claude Code conversation is no longer typed into another one when
  /clear, /resume or a claude attach job switches the pane's session while the message is
  being sent: the session is checked again right before the paste, and once the paste is in, an
  Enter is pressed only while the conversation can still be told. Otherwise the message is held
  before the paste, or left marked uncertain without its Enter.
  (#2318)
• After Claude Code's /branch, Chat follows the branch's continuation in the branch's own project
  instead of a file of the same id in another project, and reads a transcript's records however
  their keys are spelled or spaced.
  (#2241)
• A claude attach <name> pane with a debug value given apart from its option (-d api,hooks,
  --debug api) keeps its terminal: Claude Code then looks no name up.
  (#2241)
• A message queued for a Claude Code conversation is held, not sent, once the conversation in the
  pane can no longer be told (its PID record or its job's state cannot be read).
  (#2241)
• A Chat or queued message to Claude Code that ends in a backslash is now sent. Before, Claude
  Code read the backslash and the Enter after it as a new line, and the message stayed in its
  input box.
  (#2278 by @jpcarranza94)
• Claude Code's /effort card now shows while Ultracode is on, at max where Claude Code offers no
  Ultracode, and in panes narrower than 70 columns down to about 45, where the slider was left to
  the terminal before. A pane of about 44 columns or fewer still leaves it there, as Claude Code
  cuts the level names.
  (#2266)
• Chat shows what Claude wrote right before a question card, read from the terminal, while the card
  waits. Claude Code 2.1.311 saves those words to its session file only with the answer, so the card
  lacked the context the terminal showed above the question. Words whose start is above the
  terminal's screen are left to the terminal, and once Claude's own record of the turn arrives the
  chat shows that instead. Reported by @Xelanor.
  (#2217)
• The message box no longer names an older model before an agent's first message when its terminal
  is scrolled up, or was cleared, above an earlier Codex or Claude frame: the model is read from the
  live screen only, and none is named while the terminal is scrolled into its history.
  (#2163)
• A Codex pane that resumed another folder's conversation through the picker is no longer bound to
  it while another loaded conversation could be the one it shows: on a busy Linux PC one read of
  the kernel's lock table could leave a loaded conversation out.
  (#2101)
• Listing a pane's slash commands no longer stops the whole server while its skill and command
  folders sit on a slow filesystem (a Windows folder mounted into a container, a network share):
  the scan runs on the file threads, so terminals, typed input and other requests keep answering
  while it reads. Thanks to @PayrollFuchs for the report and the measurements (#2254).
  (#2311)
• Copilot Chat retries a failed process lookup once and revalidates log continuity after temporary
  descriptor read errors, without restoring stale or revoked session evidence.
  (#2283 by @YuuuXie)
• Copilot Chat keeps its messages while retrying a failed transcript page read when the same
  conversation is still verified, instead of briefly showing the terminal fallback.
  (#2068 by @YuuuXie)
• A failed read of the gap between Chat pages no longer discards the messages already displayed.
  (#2068 by @YuuuXie)
• Copilot conversation ownership uses the kernel's stable process birth, so clock corrections no
  longer reset a verified binding or make a previously denied fallback eligible again.
  (#2293 by @YuuuXie)
• Find in a diff now brings the match itself into view on a wrapped line taller than the screen,
  such as a long line on a phone, instead of the line's middle, which could leave the match below.
  (#2156)
• A Codex pane's DONE dot now clears when you read the answer in Chat even if Codex recorded the
  turn's completion after the chat first loaded the answer. Before, it stayed until the next finish.
  (#2136)
• A message sent from Chat right after the page reconnects (after a server restart, or while the PC is
  busy) no longer fails with "this PC has not said yet what it supports": the server now says what it
  supports as soon as the connection opens, instead of only with the pane list, which can take seconds
  to read from herdr or fail. Older servers keep the previous behaviour.
  (#2268)
• Opening, previewing or looking up a file on a slow or stalled disk (a network share, a large
  rendered page) no longer freezes every terminal and request: the file routes read the disk on a
  few background threads, and answer 503 to try again when those are busy or the disk does not
  answer in time.
  (#2256)
• On Windows, Hermes skills show in the command menu again when %LOCALAPPDATA% is redirected: Ovis
  cannot read a Windows pane's environment, and now looks in the server's %LOCALAPPDATA%\hermes, as
  it already did for Hermes transcripts, instead of <profile>\AppData\Local\hermes.
  (#2316)
• The installer no longer links the Node it installs into ~/.local/bin, where it went before the
  Node of your own projects. Once Ovis runs this release, running the installer again removes a link
  it made earlier and says so: if that link was the only node your shells had, install your own
  Node to keep using node outside Ovis.
  Terminals that run on the Node fallback, and updates from Settings, use the right runtime even
  when herdr's PATH starts with an older Node (reported by @wafe).
  (#2100)
• Sending an @skill to OpenCode no longer types a second space after the skill when the pane redraws
  blank while Ovis waits for Tab's result.
  (#2239)
• On Windows, Hermes transcripts, session history and skills are found under %LOCALAPPDATA%\hermes,
  where Hermes 0.21.5 keeps them (a store an earlier Hermes left in %USERPROFILE%\.hermes is still
  read), and agent stores are looked up under USERPROFILE even when HOME is set for Git or Emacs.
  (#2239)
• Korean typed in the terminal in the iOS app now reaches the pane as whole syllables (안녕).
  Before, each jamo went out on its own (ㅇㅏㄴㄴㅕㅇ) because the app's web view does not say
  "Safari" in its user agent. Safari and the Home Screen app already worked.
  This comes with the PC update; no new iOS app build is needed.
  (#2307)
• The iOS app no longer shows WebKit's form bar (the up and down arrows and the check mark) over
  the keyboard, so only Ovis's own key bar sits above it and the page keeps that row of the
  screen. This needs the new iOS app build; updating the PC alone does not remove the bar.
  (#2109)
• A message sent to a PC whose Ovis is too old to confirm messages is no longer treated as delivered:
  the composer, Send now on a held copy and the held queue's Send now keep its text as a
  "Not confirmed" row to copy or discard, instead of dropping it.
  (#2164)
• A terminal command sent from Chat right after a pane opens now opens the terminal view even when
  the terminal had not loaded the pane's command list yet, and a chat shown again draws its
  "Ran in terminal" rows at once instead of as plain messages first.
  (#2164)
• Ovis for Mac takes the installer only over HTTPS from ovis.sh or herdrweb.dev, reads the phone
  address only from this Mac's own server, and draws the QR code only for Tailscale's .ts.net
  HTTPS address.
  (#2116)
• Ovis for Mac makes the phone address with the Ovis release that runs now, not the one first
  installed, and shows Tailscale's answer when it fails.
  (#2116)
• Remove in Ovis for Mac stops at a step that fails (herdr's plugin list included) and shows why,
  and takes the phone address off Tailscale even when the server is down.
  (#2116)
• Start in Ovis for Mac no longer adds a login item for herdr set up from a terminal; it runs herdr
  for this login only.
  (#2116)
• Ovis for Mac moves herdr's log to herdr-server.log.1 once it passes 10 MB, when the app opens
  or starts herdr, and its refresh button has a VoiceOver label.
  (#2116)
• The sidebar no longer goes stale for about three seconds when many agents change status at
  once. A burst of more than five status changes closed the browser's event stream, which then
  reconnected and read the whole roster again, and a finish in between rang no sound and showed
  no drop-in on the page. The stream now lets a browser fall up to 4 MiB behind, beside the one
  large roster it may be reading, before ending it, and every browser's stream together holds at
  most 32 MiB for those that fell behind, so stalled devices can no longer fill the server's memory.
  (#2260)
• A PC set up at another SSH address no longer shows or saves the roster of the address it left once
  it moves there. When the new address's roster waits for that server to prove it is the one
  answering, the PC's card has no panes until it does, and then shows that address's own.
  (#2181)
• A command sent to a custom OmO (one whose launcher does not name omo) keeps its row in Chat after
  the turn ends: the row follows the OmO process that holds the session, as a queued message does,
  instead of the bundled Claude SDK process that exits with each turn.
  (#2186)
• A question OmO 5.1.48 asks without waiting for the answer shows in the chat as its question card
  with its options again, instead of a "Waiting for input" card offering only Enter and Esc.
  (#2090)
• omp's tool approval card now also shows when omp's composer is not the default status band: a
  box (composer.shape: box), the Claude Code, Pi, borderless, rule, field and rail shapes, any
  status-line preset, and the ascii symbols. omp's Ask card shows under the rule, field and rail
  shapes too.
  (#2245)
• An omp card no longer shows, and an answer sends no key, while omp is stopped behind its shell
  (Ctrl+Z, or a suspended job): herdr still names the pane omp, but the shell holds the input.
  (#2245)
• An omp Ask card now shows the question and its options in the agent's own Markdown (bold, code,
  links, lists, a fenced block), and a question omp cuts at four rows in full. The card takes the
  text from the session's open ask call only when it matches the menu on screen; otherwise it
  shows the screen's text as before, and the answer keys still come from the terminal's menu.
  (#2220)
• Chat shows omp's pictures from its blob store on macOS and Windows too, not only on Linux: the
  bridge opens a blob through the store directory it holds open, so a link or a directory swapped
  in above it can never stand in for the store.
  (#2114)
• An omp started from its npm package (bunx or npx run dist/cli.js with bun) shows its approval
  card in Chat again; the card stayed hidden and the answer had to be given in the terminal (#2280).
  (#2320)
• An omp Ask question with a Mermaid diagram, or with options that would read alike on the card,
  shows as omp draws it, so the card never shows a diagram in place of the code or two identical
  rows (#2233).
  (#2320)
• A prompt card's typed answer no longer picks one of two rows that look alike across scripts
  (Cyrillic о beside Latin o), and a typed row number picks its row even when its name looks
  like another row's. Commas or spaces alone on a multiple choice are the typed text, not an empty
  answer (#2233).
  (#2320)
• Chat messages sent to a starting omp now also wait for herdr's integration when omp loads it
  from the project's .omp/extensions, an -e/--extension/--hook file, or the agent directory
  PI_CODING_AGENT_DIR, PI_PROFILE or PI_CONFIG_DIR names. An omp started with
  --no-extensions or --trusted-extension is sent to at once instead of after 30 seconds, and a
  wait that runs out without a report is now noted in the server log.
  (#2159)
• omp's tool approval ("Allow tool: …") shows as an approval card again, with the call as its body
  and Approve and Deny sent with the keys omp names, instead of a fallback card (omp 18.10 and
  later; 18.13's "approve for this session" and grant rows are offered too). Reported by @Damin-Lee.
  (#2106)
• An OpenCode @skill sent from Chat is no longer reported as not attached when the terminal is
  resized while it is accepted, or when a narrow terminal's status row shows only the model.
  (#2185)
• An OpenCode 2 pane started with --standalone or --server now offers its skills as @skill
  before its first session, even when no OpenCode background service has written its service file,
  and a Chat @skill send to it shows a state once Chat has listed the pane's commands.
  (#2315)
• A PC whose own Ovis asks for its access token (OVIS_TOKEN) now shows the port it really answers
  on and its "Open that PC's Ovis" https link on its card and in the dialog, instead of the
  bridge's registered port and no link, once that Ovis runs this version too.
  (#2248)
• The token another PC's app uses to update this PC's own Ovis now opens only the update status read
  and the install request, each by its own method and with no query string.
  (#2193)
• A pairing link is tried only once per page load, including when the sign-in screen returns.
  If pairing succeeds but the browser does not stay signed in, the screen asks you to check
  cookies instead of reporting that the already-used code is wrong or expired.
  (#2321 by @kilhyeonjun)
• A command sent from Chat to a pi session now shows its result when the session moved to another
  branch and /tree came back before the result was written, instead of ending as unknown.
  (#2146)
• A message sent to Pi from Ovis Chat no longer reads as delivered when the same text, typed into
  Pi's terminal before it, reaches Pi's transcript first.
  (#2323)
• Pi sends retry a briefly incomplete conversation record before giving up transcript confirmation,
  including when the writer paused before Ovis started reading the file.
  (#2231 by @philipprass-prog)
• Two plugin starts at the same moment, such as herdr's startup hook and the Start action, no
  longer leave stop saying Ovis is not running while a server keeps going: starts and stops take
  turns, one that waits says so in the plugin's log, and a start no longer opens a second server
  while one an earlier start left running (herdr away, say) still runs.
  (#2166)
• An OmO question answered from a Chat card no longer goes to another open question when the
  pane is too small to show the form's title: when the form left in view fits more than one open
  question, the card is refused and nothing is sent.
  (#2237)
• A Claude command sent from Chat beside the fullscreen diff panel no longer takes the question of
  the same command typed in the terminal after an earlier send's echoes were counted without the
  panel.
  (#2237)
• A command sent to an OmO pane with two OmO processes holding sessions keeps its outcome on the
  OmO it was sent to, also when it waited in the queue and herdr listed the other one first by
  then, and never on a later process that got the OmO's process number.
  (#2237)
• When Turn alerts on again cannot make a push subscription, the alerts notice and Settings now
  say why: the browser would not make one (in Brave, its push service is off until Use Google
  services for push messaging is on), the page's service worker is not running, or the PC did not
  take it. Before, the button only flickered and the notice came back unchanged.
  (#2145 by @mattmarcin)
• The alerts notice can be dismissed. It stays closed on that device for the same problem, and
  shows again for a different one or after alerts work again. Settings still shows the problem.
  (#2145 by @mattmarcin)
• A push endpoint its push service retired stays refused, however long ago that was, until its
  owner's device is revoked or a thousand later retirements push its mark out. Before, after 30 days
  another device that knew the endpoint could register it as its own, and so could another spelling
  of it (a #fragment, an uppercase host or :443) at any time.
  (#2140)
• A subscription its push service retired is removed even when push-retired.json cannot be
  written (a full disk, a read-only state directory). Before, the failed write left it registered
  and the next alert went to the dead endpoint again.
  (#2140)
• When a crash left the saved permission to connect to the PCs saved in herdr unreadable, Settings
  now says to choose Connect on those PCs again, instead of telling you to add them again.
  (#2192)
• A file opened from a Claude Code session's scratch folder is served, described and drawn as a page
  only from the file actually opened inside that folder: a link put in its place after it was found
  no longer serves a file from outside the folder.
  (#2265)
• Read DONE dots no longer come back after herdr restarts on a PC: a late read of the old herdr run
  could make the server drop the new run's read finishes and stop saving new ones.
  (#2249)
• The server no longer keeps or hands out the read finishes of a herdr run it knows has ended, so a
  browser that reconnects stops writing them back into its storage, and the next change drops them
  from seen.json. The records of a PC this server has not heard from since it started are kept.
  (#2197)
• On a phone, a jump of several entries back through the browser's history menu from a Settings
  opened again over a preview now goes to the pane it lands on. Settings no longer keeps the mark an
  earlier opening from the command palette left behind, which made the jump count as closing it.
  (#2182)
• Closing a sidebar filter that the command palette opened no longer leaves the focus on nothing when
  what opened it is gone; it goes to a workspace row. Opening or clicking a row's menu no longer closes
  that empty filter, and words typed into it right after it lost the focus are no longer cleared.
  (#2141)
• The session read, the /ws snapshot and a PC's relayed or proxied snapshots no longer carry the
  token values herdr's sidebar rows hide, so no browser, watch-only ones included, receives them in
  any snapshot. Another Ovis server reading a PC still gets them, to apply its own rows.
  (#2171)
• The agent rows' machine events and PC roster no longer carry the literal text and numbers of the
  hide rules in herdr's sidebar rows, nor the token values those rules hide, so watch-only browsers
  no longer receive them there, and a hidden value no longer shows for a moment while the page
  loads.
  (#2139)
• A vertical mouse wheel over a tab strip whose tabs do not fit scrolls it sideways, so tabs past
  its end can be reached with a mouse.
  (#2095)
• Korean lines in the terminal no longer nearly touch: Hangul is widened to its two cells
  instead of enlarged, so it keeps its own height, and a row is made tall enough for Hangul and a
  gap below it where the Latin font's line is shorter (Menlo, Consolas), keeping every row where it
  is not.
  (#2243)
• Korean text on a background colour (Claude Code's prompt, a highlighted or inverse line) no longer
  shows a thin light line after each word in the terminal, in Ovis for Mac, Safari and Chrome.
  (#2263)
• Powerline separators in a prompt no longer make the terminal redraw every row when the icon font
  arrives: they are painted as boxes, not drawn from the font. A pane that prints icons still
  redraws once their font arrives, so they are sized from it.
  (#2158)
• Settings → Phone & devices keeps the "This device" badge on one line beside a device's name on a
  phone, moving it whole to the next line when the name fills the row, instead of splitting it.
  (#2187)
• When the agent's next todo update folds the update you were on into a Plan row, the keyboard focus
  stays on that update instead of dropping to the page, and an update opened inside the Plan row
  now shows the input it sent and the output it got back, not only the list again.
  (#2161)
• An Ovis restart while a Sign in again or Add account login is still open in its terminal tab no
  longer lets a second login start in the same credential home: the restarted server finds the tab
  still running and answers that a sign-in is already open, until that login ends.
  (#2325)
• Voice input no longer fails in a pane with many slash commands when the transcription server
  caps a request at 32 fields (codex-lb does): only as many command hints are sent as fit.
  (#2235 by @vuongnmhvn)
• A device paired to watch only can no longer read a pane's command rows (custom commands' first
  lines and file paths) through /api/pane/commands, on this PC or another: they are refused as a
  folder's commands already were, and its chat no longer keeps asking for them.
  (#2246)
• A device paired to watch only can no longer list the file paths under a pane's folder through
  /api/pane/files, or read a plugin action run's output through /api/plugin/action, on this PC
  or another: they are refused as a folder listing already was.
  (#2252)
• A device paired to watch only can no longer list a repository's checkouts (their folders and
  branches, those not open in herdr too) through /api/worktree/list, on this PC or another: it is
  refused as a folder listing already was. Its sidebar no longer asks for them and shows no branch
  names.
  (#2304)
• Ovis for Windows' Update for a server under the old id installs the newest release even when
  the PC sets an older OVIS_REF, and when the installer or herdr's reinstall fails, the server
  that ran before starts again, with or without herdr running. A running server that reports no
  version and runs without the update supervisor, such as one just started whose supervisor has
  not reported yet, is no longer offered Update, and the
  version check waits once more for a server slow to answer and takes a token set only in the
  environment.
  (#2113)
• Ovis for Windows no longer offers Start for an Ovis server older than the app, or one installed
  under the old devswha.herdr-web-ui id: it shows Update, which updates it in place (herdr, Bun,
  settings and token stay), starts it, sets herdr to start at sign-in and ends at the phone's QR
  code. It says so when it cannot read the server's version, and names a herdr that is too old.
  The plugin's status now prints the version that runs, which can be newer than herdr's list.
  (#2091)
• Ovis for Windows and install.ps1 now stop an Ovis server from before the rename (0.4.12 and older)
  that kept running after herdr closed, so Update no longer fails with "the file is in use" and then
  keeps showing the same Update screen.
  (#2275)
• Try again after a failed Update in Ovis for Windows reads the PC again instead of running the old
  server's removed files ("Module not found").
  (#2275)
• Ovis for Windows shows the phone QR code for a server that asks for a token, instead of telling a
  PC that already has Tailscale to install it.
  (#2275)
• The log in Ovis for Windows shows the installer's messages as plain lines, not raw PowerShell
  CLIXML.
  (#2275)
• On native Windows without HOME, chat transcripts and session lists find each agent's store
  under USERPROFILE too, as skill discovery already does.
  (#2157)
• Windows remote PCs wait for Windows sign-in before starting herdr, so Codex can use its
  private-desktop sandbox. Startup options offer an explicit start without sign-in for headless
  PCs, including when Task Scheduler cannot start herdr, with the error and sandbox and desktop-tool
  limitations shown before that one start. Existing herdr servers are reused without restarting
  their panes. Signing out of Windows ends a desktop-started herdr and its panes; lock the PC
  instead. SSH disconnects and bridge updates leave them running.
  (#2087 by @JJLiebig)
• Slow Windows launchers keep their startup options through the launch deadline. Failed startup
  reviews restore saved PCs' reconnection attempts; Task Scheduler failures show their own error
  instead of a sign-in instruction and stop background retries until startup options are reviewed.
  (#2087 by @JJLiebig)
• The terminal no longer shows half of an agent's new screen over half of its old one while the
  agent streams, which happened most often with Korean text: the browser terminal is now xterm.js
  6, which draws a frame only once it is whole (#1617). Trackpad scrolling, Option+arrow word moves
  and Korean typing keep working as before.
  (#2335)

Maintenance

• Diagrams use updated DOMPurify and KaTeX versions without their published security advisories.
  (#2290 by @DravenDou)
• A stopped server also closes its OpenCode event stream and forgets the OpenCode panes it verified, as
  it already did for Codex's app-server connections. The session list's Resume window and Hermes's
  result readers drop their duplicate code paths.
  (#2200)
• A local browser lane run (no CHECK_DIR, no CI) takes the empty evidence folder it hands each
  check as none in every check: the touch-burst and terminal-autoswitch checks no longer write
  their captures into the checkout, and with BROWSER_PARALLEL=1 the question-wheel, CJK
  punctuation and late-faces checks no longer fail at once creating an empty path (#1554).
  (#2196)
• CI's parallel browser lane stops a script that runs past BROWSER_SCRIPT_TIMEOUT_SECONDS
  (600 s by default) with every process it started, and names it with its last log lines, so a
  hang fails fast instead of running into the job's timeout.
  (#2287)
• The xterm source build compiles the exact files its cache key was made from, and the key now
  covers xterm's package manifest, so a source edited mid-build or a changed typings entry can no
  longer reuse an old build. A killed build's leftover staging directory is cleared by the next
  build, and the published build is readable by other users of the checkout.
  (#2242)
• The Windows-native tests wait for the desktop child's whole result file instead of parsing it
  half written, and every Windows-only case has a 30 s budget for busy hosted runners.
  (#2242)
• bun run check now ends a failed run with the reasons it failed, such as a source that changed
  while it ran, and the report path, instead of only ok steps and an exit code.
  (#2135)
• Self-hosted CI runners keep a git mirror and the default branch's Bun packages on the host, read-only
  to jobs, so a job no longer downloads about 150-200 MB of checkout, tags and packages.
  (#2215)
• The unit test run (bun run test:unit, CI's Fast checks) now also runs the component tests
  named *.test.tsx, which it had never picked up, and the agent sidebar's test passes when it
  runs on its own.
  (#2094)
• Command discovery keeps following a linked command folder that points outside the project, as
  Command Code's own loader does, so the menu lists the commands Command Code runs; a test now pins
  it and the walker says why.
  (#2134)
• Conversation fallbacks and recoveries now emit sanitized, rate-limited diagnostic episodes,
  including the failed Copilot identity stage without logging session IDs or transcript paths.
  (#2292 by @YuuuXie)
• The file-thread test that holds a thread inside a read can no longer hang the unit run: its
  stand-in thread opens the FIFO it waits on before saying it is reading, and the test releases it
  with a write that never waits, so a thread ended in between no longer blocks the test forever.
  (#2297)
• The sidebar no longer walks every stored read finish on each roster update for a PC that has
  none, and code the finish index kept for no current use is gone.
  (#2183)
• The Copilot switch-back integration test now waits for the resumed transcript itself, not for the
  moment between the log's unregistering and registering lines, when a read rightly fails closed.
  (#2224)
• The integration test that holds a first message once a same-kind agent replaces its own now starts
  the replacement before the first agent ends. Before, the shell alone was in front between the two,
  and a read in that gap rightly found the launched agent gone, which failed the test now and then.
  (#2222)
• The OpenCode Tab agent-switch integration tests now start from a closed autocomplete. The test
  before them left the skill row selected on screen, and a send that read that old screen could not
  tell the agent before Tab, which failed the tests now and then.
  (#2223)
• scripts/roster-bench.ts measures what a status burst costs the server and the browser with
  20 to 300 agent panes in an isolated herdr (#772).
  (#2260)
• The Unicode case folding and confusables tables are generated only from source files whose SHA-256
  matches the pinned one (#2233).
  (#2320)
• The browser check that a focused window takes its pane back no longer times out waiting for a
  resize after the first attach: since a new terminal attaches once its font is set, that attach
  already carries the final grid, so the check now waits for the shell to report the grid each
  window last asked for, by attach or by resize.
  (#2295)
• The Release run now refuses while the remote PC bundle it names is only tagged: it waits until
  that bundle's manifest and archives can be downloaded, as remote PCs download them.
  (#2212)
• The self-hosted runner launcher's tests of a hung first cache refresh no longer fail on a loaded
  host: they wait for the refresh to hang through a FIFO the stub writes to (no polling), allow
  30 s for that and for the launcher to end instead of Bun's 5 s test timeout, and let no runner
  start before the refresh has hung; only the stop on SIGTERM keeps a tight bound (4 s) (#2261).
  (#2267)
• The terminal arrows and clicks browser check no longer fails now and then on a busy CI runner.
  It reads each letter before pressing the arrow that follows it, waits for xterm to report the
  mouse before clicking, and proves that nothing arrives by typing a mark instead of waiting a set
  time. A new unit test checks that the server holds an arrow back after the typing before it.
  (#2219)
• The browser lane follows synchronized output (CSI ? 2026 h … l) from a real pane to the
  rendered rows: the bridge passes herdr's frames intact, herdr holds a pane's frame until it
  closes, and the run counts the renders xterm 5.5 draws from half a frame (#1617).
  (#2331)
• Three tests that failed on busy hosts with nothing wrong in the code now measure what the code
  controls: the browser lane's stubbed runs get a 60 s hang guard, which ends the lane's process
  tree and fails with what it printed, instead of a 5 s limit (#2110); the Codex matcher's
  bounded-time cases compare the CPU time of an input with that of one as long that the code
  reads in one pass, which leaves the work both share out of the ratio while each still fails the
  quadratic scan it guards (#2142); and the OpenCode first-key gap is measured from the last thing that send itself
  waited on (#2143).
  (#2150)
• Native TypeScript checks use less CPU with GOGC=400 and at most 32 Go scheduler threads,
  retaining GC for concurrent CI runners. Caller GOGC, GOMAXPROCS, and GOMEMLIMIT
  settings are preserved.
  (#2276)
• The Windows plugin test that stops the server's process tree no longer fails now and then on
  CI: it took a process whose recorded parent number Windows had handed on to a server process
  for part of the tree, and waited for unrelated processes, its own test runner among them, to
  exit. It now walks the tree the way stop does, skipping a process older than its parent.
  (#2202)
• The native Windows terminal check now tests a pull request's merge with main, as the other CI
  checks do. On a branch cut before the check's script was fixed, it ran that branch's old script
  and timed out after the wheel pan, which made the check look flaky.
  (#2253)
• The native Windows terminal check (a mirrored PowerShell pane's wheel, touch, panning and IME
  input in Chromium) now runs in CI on a Windows runner, windows-terminal-qa.yml, for pull
  requests that touch the mirror, the terminal's viewport code or the xterm patch, and weekly. It
  now picks Original width for the pane first: since wide mirrors fold on a phone by default, its
  panning steps could not pass.
  (#2211)
• Builds running at once in one checkout, like the browser lane's parallel workers, no longer fail
  now and then with a missing xterm export: xterm's source is built once into its own cache
  directory and never rewritten while another build reads it.
  (#2138)