devswha/ovis v0.4.21

Ovis v0.4.21

on GitHub

Added

• The agent picker of New workspace, New tab and New worktree ends with "Add another agent…": it shows
  the agents the PC does not have yet as the setup guide's tiles, installs the one you tap right
  there, then picks it.
  (#2413)

• Settings → Agent integrations lists what the PC open there has, with versions, and the agents it
  can install as the same tiles; herdr's integrations stay below. "Open setup guide" at its top opens
  the first-run guide again.
  (#2413)

• An agent can be installed and signed in to from Ovis: the server opens a terminal tab on that PC
  and types the vendor's own install command for it, or the agent's own sign-in, so every line
  shows and Ctrl+C stops it (POST /api/agent-catalog/install and /sign-in, also for remote
  PCs). A setup card follows the install, finds the agent once it is in place, and offers its
  sign-in.
  (#2328)

• Claude Code and Codex sign in from the browser. Sign in on an agent's row in Settings →
  Agents, Sign in again, Add account → On this PC and the setup guide's Sign in to … run
  the agent's own login on the PC as before, and a small dialog follows it: the sign-in page to open
  (on the provider's own address), Codex's one-time code to enter there, and a box for the code
  Claude's page shows, typed into the login for you. It says how the sign-in ended; Open terminal
  stops the private login before starting a visible terminal tab. The private prompt belongs to the
  device that started it, and closing it frees an unused account name. The agent writes its own
  credentials; Ovis keeps no code after the login ends.
  (#2428)

• The user guide shows how to run Ovis on one PC that stays on and add your other PCs to it as
  remote PCs, instead of an Ovis on every PC kept on the same version. Suggested by @WOULDU-pres
  in #2056.
  (#2347)

• Chat: select text in an agent's reply to comment on it. Comments wait in one chip over the message
  box and go first in your next message, each as a > quote and then the comment. A slash command or
  an answer to the agent's question leaves them waiting, and they only ever go to an agent: the bridge
  refuses them where no agent is in front. Requested by @aNNdii (#511).
  (#2350)

• A Claude Code pane's message box can show an Artifacts chip beside the background tasks chip
  once the session publishes a page with Claude Code's Artifact tool. It lists each page once, with
  the title and time of its latest publish, newest first, and opens the page on claude.ai in a new
  tab. It is off until turned on in Settings → Chat → Composer → Artifacts chip, so the
  message box stays minimal. Ovis reads the pages from the session's transcript; a remote PC needs
  this version to list them.
  (#2405 by @jpcarranza94)

• Settings → Subscription usage has Read Claude's limits from its status line for networks where
  Anthropic's usage endpoint is refused (a 403 that showed as an expired sign-in). Turned on, the
  server's PC wraps the status line set in Claude's settings.json (under CLAUDE_CONFIG_DIR too)
  and runs that command unchanged, keeping only the five-hour and weekly windows and the account,
  owner-only, in ~/.herdr/statusline-last.json; turned off, it puts settings.json back as it was
  and deletes what it added, never a file there it did not write. A Claude account refused that way
  offers Read from Claude's status line instead in the usage popover. A reading counts only for
  the account that drew it and only until its window resets; the endpoint stays the fallback. A
  watching device cannot turn it on. By @CharleoYang.
  (#2269 by @CharleoYang)

• omp's context ring shows a percent, read from the context figure in omp's own status bar while
  the pane is scrolled to the bottom. By @CharleoYang.
  (#2269 by @CharleoYang)

• The user guide shows how to reach Ovis through a Cloudflare Tunnel, and what to keep on behind
  Cloudflare Access, which cannot hand the app its token.
  (#2412)

• Letta Code panes show Letta's own mark in the sidebar, the agent picker and the composer, in place
  of the initial in a disc.
  (#2414 by @homebodify)

• A setup guide on a fresh install: Ovis asks whether this is your first time, then checks which
  coding agents this PC has and shows every one as a tile, the same whatever it found. Pick an
  installed one and Start opens your first project with it (Claude Code, else Codex, else the
  first found); tap Install on a missing one and Ovis runs its official installer right there, a
  ring turning on its tile until it is installed. Plan usage turns on for accounts already signed
  in, and Connect your phone beside Start shows a QR code and the address to open on the phone (on a fresh
  install it pairs this browser first, so the phone does not shut it out). It
  opens by itself only once, when herdr has no workspaces and Ovis has no saved settings or paired
  devices; "I've used Ovis before" goes straight to the app.
  (#2332)

• Settings → Shortcuts has Previous tab, Next tab, Previous workspace and Next workspace, with no
  key until you choose one: they step through the open workspace's tabs in the tab strip's order,
  or the workspaces in the sidebar's order, and wrap around at the ends.
  Requested by @odfalik (#2038).
  (#2349)

• A bare file name in the terminal (design.md, no folder in it) opens in the file viewer on
  Cmd+click (Ctrl+click off a Mac), underlined while the key is held, as in VS Code's terminal.
  It is looked for under the pane's folder, and several files of that name are listed to choose
  from. A plain click or a tap on one still lands on the terminal.
  (#2348)

• An OpenCode pane's background commands show in the Background tasks list with its subagents,
  as a Claude Code pane's do: running first, then the ones that ended. Reported by @sulthonzh
  (#2340).
  (#2341 by @sulthonzh)

Changed

• Ovis checks agent versions in the background when its server starts, so opening the agent catalog
  after those checks finish no longer waits for slow version commands.
  (#2401)
• The sidebar's Agents section remembers whether you folded it, on each device, across reloads
  and launches. With nothing remembered it still starts folded in the phone drawer and open on a
  desktop.
  Reported by @Ploppy3 (#2236).
  (#2351)
• Settings → Subscription usage and Settings → Agent integrations are one page, Settings →
  Agents: one row per coding agent on the PC with its version, whether it is signed in (as far
  as its plan accounts show), and, opened, its accounts and their limits, Add account, Sign in
  again and herdr's session resume for it; New agents keeps how New workspace starts Claude
  Code and Codex. Accounts no agent there signs in to (Cursor, an agent not installed) and the hub
  are under Other accounts; the agents to install are under Install another agent, and the
  plan-limit display settings close the page. Links and the command palette that named either old
  page open Agents.
  (#2427)
• The daily anonymous usage message now also says which kinds of screen opened the app that day
  (the phone and Mac apps, an installed web app or a browser tab), how they reached the PC, and how
  many prompt answers were sent and pane alerts delivered and tapped. Counts only: never which pane,
  agent, device or time. The one-time notice is shown again, and nothing of the usage is counted or
  sent until it has been.
  (#2392)

Removed

• The sidebar no longer shows Install app; install from Settings → Phone & devices → Install,
  which offers it when the browser can install and says how otherwise.
  (#2391)

Fixed

• Pi's Earlier sessions list retries a busy reader automatically and offers Retry if it stays busy,
  rather than showing an error or an empty list. A bounded metadata cache avoids rereading
  unchanged transcripts while picking up new prompts and renamed or cleared session names.
  (#2377 by @IdrisGit)
• Opening the setup guide again from Settings no longer turns plan usage back on.
  (#2413)
• An agent whose installer adds its folder to the login shell's PATH is offered in the agent picker
  once that PATH has it: the server reads it again for the install, and the dialog waits for up to
  a minute before saying it is not offered yet.
  (#2413)
• Opening Ovis on a slow link, such as a phone on mobile data or the app's first connect, shows a
  loading spinner while the app downloads instead of an empty screen that looks frozen. Thanks to
  @Shunamxiao for the report (#2329).
  (#2367)
• An installed agent that does not start (a wrapper whose target was removed, for example) is no
  longer shown with a version read from its error message: the agent catalog marks it as not
  running, and the setup card does not call it ready.
  (#2375)
• The agent catalog runs a version check with the PATH the agent was found on, so an agent run by
  an interpreter only the login shell knows (nvm's Node, for one) is not taken for broken.
  (#2375)
• The install catalog says what installs on a real PC showed: Oh My Pi's installer stops while an
  older Bun is on PATH, Devin's may wait at its own sign-in until you press Ctrl+C, and OpenCode's
  installer puts its folder first on PATH ahead of an opencode of your own. Copilot gets its
  copilot login sign-in, and Hermes, whose version check is slow, is no longer shown without a
  version.
  (#2382)
• Local Markdown images in Chat answers render inline using the pane's PC and working directory,
  with the file viewer's size limit and link fallback. Click or keyboard activation opens the
  existing image viewer. URL-encoded filename characters resolve correctly.
  (#2435 by @cheol-95)
• Relative Chat images follow the foreground process folder. Failed local lookups offer Try again
  so a delayed file or temporary server failure can recover without reconnecting.
  (#2435 by @cheol-95)
• A message sent while the chat was still loading no longer shows twice, as its turn and its faded
  waiting bubble, for up to a minute when the agent writes whole-second times and nothing after
  that turn: the turn alone settles the bubble.
  (#2365)
• A Chat message to Claude Code is no longer refused as "input box may not be empty" because of
  Claude's grey suggested prompt while the pane's terminal is scrolled up into its history.
  (#2406)
• A Codex pane's chat no longer shows another session's conversation from weaker screen evidence:
  it reads the screen only while a Codex process runs in the pane, a long answer must match the
  one shown with its punctuation (a < b is not a > b, feature-one is not feature_one), and
  a command Codex shows it or you ran, or is running, and a tool's output, are not taken for an
  answer. When nothing proves the conversation, the pane shows its terminal.
  (#2430)
• A Codex pane's chat no longer shows another session's conversation: short answers count only
  below the prompts they answered, a long answer quoted in a prompt or the input box is no match,
  and the thread herdr reports for the pane (which can be stale, or another pane's through Codex's
  shared daemon) is shown before its answer is on screen only when it is the pane's brand-new
  conversation. When ownership cannot be established, the pane shows its terminal.
  (#2426)
• Codex Chat can identify a conversation whose answer links an absolute local file path
  containing spaces, which the terminal displays relative to its working directory.
  (#2417 by @cheol-95)
• Large OmO and GJC session replays yield between read chunks so the web server keeps answering
  requests while it reconstructs agent status.
  (#2361 by @eddieparc)
• A trackpad wheel held with Ctrl scrolls a terminal application at its normal speed again, as
  before the xterm.js 6 upgrade, instead of faster; only Alt speeds the wheel up.
  (#2357)
• Recently viewed terminal tabs keep their screen current while this window is in use, so switching
  back shows the pane as it is now at once: the three recent tabs, or the most recent one on a
  phone. A hidden tab still lets go of the pane and its input; the existing alongside-herdr setting
  still releases hidden panes.
  (#2418 by @eddieparc)
• On an iPhone, tapping the mic records with the browser's speech recognition instead of saying the
  microphone permission was denied: iOS starts it only from a tap, so a touch starts and finishes a
  take by tapping, where a mouse still holds to talk. In a home-screen app, turning Voice input on
  no longer asks for the microphone at once, the request reported to close or reload the app; the
  first tap of the mic asks. Speech recognition that is off on the device (Siri & Dictation) says so
  instead of blaming the permission. Thanks to @osamahbeig for the report and the diagnosis.
  (#2370)
• A sentence dictated into the terminal grid with an iPhone keyboard that types it all at once
  (FluidVoice, other dictation keyboards) now reaches the pane whole, not only its first letter.
  Thanks to @blaker3115-Agri for the report (#1488).
  (#2386)
• The key bar's Show/Hide keyboard control on a touch screen no longer fights itself: its label,
  pressed state and its tap's action are one word from the tap, held until the keyboard measured
  after that tap confirms it or the settle window runs out — a tap can never do the opposite of what
  the control just said, and a quick reversal is not undone by the last animation. The word is the
  tapping pane's own, so another pane's control is never stuck on it. Show puts a keyboard back
  after Android's back gesture (blur and refocus inside the tap; a focus alone raises nothing
  there), and the blur reads as the keyboard going where the measurement is blind (a zoomed page, a
  floating keyboard), the word standing through the hiding animation the measurement still reports.
  (#2333 by @IdrisGit)
• Switching the key bar's Input line and direct typing on a touch screen no longer raises a
  keyboard the user put away (including after Android's back gesture), and with the keyboard up
  the new mode's field takes it, focus and all. A keyboard the page cannot measure (an iPad's
  floating keyboard, a zoomed page) counts as down there: Show brings it back after the switch.
  (#2333 by @IdrisGit)
• In the installed desktop app with the title bar hidden, the image and diagram previews' top bar
  now sits below the title bar area: its buttons no longer sit under the browser's own window
  controls (a click on the close button could close the whole app) or start a window drag.
  Thanks to @Hyd3-14 for the report and the root cause (#2026).
  (#2381)
• One slow snapshot no longer marks a connected local PC as a connection error. A second
  consecutive timeout, a startup failure, or a definite transport failure still reports the error.
  (#2418 by @eddieparc)
• An OmO pane keeps its chat when OmO's runtime started a few seconds late on a busy machine,
  instead of falling back to the terminal with no conversation. Reported by @Noobear (#2281).
  (#2424)
• In an OpenCode pane, a drag in the desktop terminal goes to OpenCode, as a click already did, so
  OpenCode selects inside the panel the drag starts in instead of Ovis selecting whole rows across
  its side panel. Shift+drag (Option+drag on a Mac) still selects and copies in Ovis. Reported by
  @matiemer in #2362.
  (#2368)
• OpenCode's shared event connection now closes when a snapshot confirms its last pane stopped
  running OpenCode, even when no status event arrived, without dropping it during server handover.
  (#2389)
• Pi's Earlier sessions keeps available sessions visible when a large transcript takes too long to
  scan. Unfinished labels show the session ID; later opens resume the scan and fill in the labels.
  (#2410)
• Pi's Earlier sessions list now shows names and first prompts recorded deep in long sessions,
  follows later renames, and restores the first prompt when a name is cleared.
  (#2377 by @IdrisGit)
• With Windows high contrast on, the selected workspace and agent rows in the sidebar keep a
  Highlight bar on their leading edge, so the selection no longer disappears there.
  (#2352)
• Turning anonymous usage counts off now holds when two Ovis servers of this version or later share
  one state folder (an installed app and a source checkout, say): the other server recording a visit
  or a count at that moment could turn them back on, and the next day's message would go out.
  (#2408)
• A screenshot pasted into the phone's terminal input line is uploaded and its path put in the line,
  as in the Chat composer and the terminal itself; before, the paste did nothing (#801). The key
  bar's Paste pastes an image the same way where the browser can read one, and otherwise says to
  press and hold the line to paste it instead of calling the clipboard empty.
  (#2390)
• Terminal gutters follow the active theme instead of showing a black frame after the xterm.js 6
  upgrade.
  (#2355 by @imurfutur3)

Maintenance

• The installer test times the unattended question beside its real 20-second read, so a slow reader
  of the terminal no longer measures zero.
  (#2359 by @YuuuXie)
• CI's parallel browser lane starts its longest scripts first, so the lane no longer ends with one
  slow script running alone for about three minutes; on the self-hosted PC the slowest CI job
  takes about 2.5 minutes less.
  (#2339)
• CI's browser-other job uploads its ~2,000 screenshots only when it fails or is cancelled; a green
  run keeps the report and every script's log, and ends about 45 s sooner.
  (#2376)
• The chat tool-paths demo check keeps opening folded tool rows until every file shows, instead of
  one pass over the rows it counted first, which left the last row folded on a busy run.
  (#2394)
• The Copilot integration tests wait for a transcript answer through a switch's in-between replies
  (scrollback, or a 409 when ownership changes mid-read) instead of reading every reply as one,
  and print each reply they wait past.
  (#2385)
• The input line's image-paste check waits for the frame that puts the caret after a pasted path
  before it rewrites the line, so on a busy PC that frame no longer collapses the next rewrite's
  selection and makes it add to the line instead of replacing it.
  (#2399)
• The integration test for a launched agent replaced before its page attaches starts the replacement
  before the first agent ends, so no status read lands while the shell alone is in front and
  rightly finds the launch ended.
  (#2380)
• Two take-over integration tests lengthen the attach hold, so a busy runner never passes on a
  failing or displaced attach's last line before its exit arrives.
  (#2373)
• The key bar focus check proves the shell took the bar's Enter by the output of a command it ran,
  not by a count of prompts on screen, which a prompt pushed off the top can leave unchanged.
  (#2395)
• Markdown parser regression tests bound closer searches by count and scanned characters instead of
  CPU-time ratios, so shared-runner load cannot fail the two search bounds.
  (#2393)
• The output contract tests for a terminal attach that cannot spawn now make the spawn fail on the
  Rust PTY host too, so they pass in a checkout with the Rust host built, not only with the Node
  sidecar. Thanks to @kilhyeonjun for the report and the diagnosis (#2344).
  (#2364)
• The terminal keyboard demo check lets the recorded shell finish playing before it writes its
  focus-reporting marker, on a line of its own, so a late recorded frame cannot overwrite it.
  (#2379)
• The terminal viewport browser check waits for the jump-to-bottom button to go once native paging
  reaches the bottom, instead of sampling it before the page has drawn the scroll answer.
  (#2378)
• The Codex process tests read their fixtures' whole READY line, which a busy pipe can hand over in
  two pieces, instead of taking the first read for the line.
  (#2383)
• The omp picture test that reads two 20 MiB records gets the 30 s budget its 40 MiB neighbour
  has, instead of the default 5 s, which its passing Windows runs came within half a second of
  and three runs went past.
  (#2372)
• The two Pi session listing tests that read 50 long sessions (the reader's and the history API's)
  check what is listed with the listing deadline set past their own timeout, instead of racing the
  two-second production deadline on a busy runner.
  (#2402)
• The plugin start test counts a stopped server that has exited but is not yet reaped as gone, as
  stop itself does, instead of failing while signal 0 still reaches it.
  (#2388)
• The two Portal unit tests that spawn an unrelated process (a shell, a Bun script) give it 20 s to
  print READY and to exit, instead of 2 s, which a loaded runner's Bun start exceeded.
  (#2409)
• The usage browser check reloads only once the shared preference has saved usage off, and reads
  the reloaded page's requests after it has applied the server's preferences.
  (#2387)