Repository navigation
herdr web ui v0.3.50
Changed
-
The app brings its own typefaces instead of counting on the ones a device has: Pretendard for
the interface and the chat, and JetBrains Mono for code in the chat and the interface, so a
phone or a PC without them draws Latin text and Korean in the same letters as one with them.
Japanese and Chinese are not the same everywhere: Pretendard has kana and only a few hundred
ideographs, so kanji and Chinese text are still drawn in the device's own font, and a Japanese
line mixes the two. Nothing is fetched from the internet: the files come from your own PC
with the app, a page downloads only the pieces of
Pretendard its text needs (about 90 KB for a chat in English, about 200 KB more once Korean
is on the page), and each piece is kept for offline use after that. The terminal is untouched: its
font, its Terminal font setting and its grid are as before, and a Chat font you chose
still comes first. In the chat, answers, your own messages and what the agent says while it
works are a step larger (15px at the default Chat font size, which still scales
everything) on a looser line; bold is a semibold; inline code sits a little under its line and
code blocks are a step larger with more room; the message box is typed at the same size with
a mouse and stays 16px on touch; the empty chat's question is a larger line. Tool rows, times
and other small labels keep their size.
(#473) -
In the chat's message box the agent's mark, the model, the reasoning level and the context
ring sit together in one quiet pill, and the model is shown by its name where the id is a
regular one:claude-opus-5-5reads Opus 5.5,claude-sonnet-5Sonnet 5,gpt-5.6
GPT-5.6,gpt-5.6-solGPT-5.6-Sol,glm-5.3GLM-5.3, and the level follows as High. No name is ever
guessed: any other id (a dated snapshot, a suffix such asgpt-5.6-sol-max, another provider's
prefix, another vendor, a spelling that is not the vendor's own such asclaude-sonnet-5-05) is
shown exactly as received, in the code font. Behind a name the id
as received is in the name's tooltip and is read by a screen reader; on a touch screen it is
not shown yet. A pane that records no reasoning level shows the name alone. The pill only shows the model, it is not a button yet. Where the
pill does not fit beside Queue (seen on a 390px phone) it steps aside whole while Queue is
showing, and the context ring stays.
(#467) -
In the chat, the line with the model sits inside the message box, as its last row, instead of
above it, on a phone as on a desktop. It shows the agent's mark, background tasks, the model,
the reasoning level as one word (high) and the context ring. The agent's written name and the
READY / RUN / INPUT word are no longer drawn there (a screen reader still reads them): the
header names the pane, and Stop, the working row and the approval card say the state. DONE is
still drawn, since nothing else in the chat says a turn ended and has not been seen. Where
the box is narrow, beside the sidebar for one, the background-task chip shows its count
before the model's name is cut. In dark themes Queue loses its outline.
(#456) -
The chat's message box is laid out like a chat app's: a rounder card with the message on top
at the card's full width, and one row of controls under it. On the left a + button
attaches files (it was a paperclip) and the background-task chip follows it; on the right are
the model, the reasoning level and the context ring, then one round button. Send is an arrow
pointing up, and Stop is the same circle in the same place, so only the glyph changes when a
turn starts and ends. While the agent works, Queue appears once the box holds text or a
file that is still uploading; with an empty box Stop is the only button, and Queue still holds a message only
when you press it, after which the keyboard's focus is back in the message box. Where the two
do not fit side by side (a phone, or a long model name beside the sidebar) the model's name
steps aside while Queue is showing, so neither is cut; without Queue the reasoning level
steps aside before the name loses a letter. While reconnecting, "Reconnecting… message held here, never queued" is said once: in
the empty box, and under the model once there is a draft, on a phone too, where it used to be
hidden. It is no longer cut with an ellipsis, and neither is "Uploading file…"; Queue is not
shown while not connected, and a disabled Stop loses its colour. With a mouse, the bar that
resizes the box shows when the pointer is over the box, while dragging, on keyboard focus
and while a height is set by hand; on a touch screen it is always shown.
(#465) -
The chat is wider on a large screen: the conversation and the message box follow the pane,
at least 820px and at most 960px wide, 71% of the pane between. A laptop window keeps the
820px it had and a large monitor grows to 960px. Settings → Chat → Chat width chooses
Narrow (820px, as before), Default, Wide (1152px) or Full, the whole pane.
The held messages, the approval card and the background-task list keep the same column.
Phones and narrow panes look the same as before.
(#463) -
Held messages above the message box are quiet rows instead of a tinted box: one line above
them, a clock with the sentence the box already showed ("Held until the agent is ready", and
"· 2 messages" from two), then each message on its own row with Send now and Discard
beside it. A message is still edited in place and still goes out only when you press
Send now. The rows are as wide as the message box at every window width, and the list
scrolls after two and a half rows. When the agent is ready, the clock and Send now take
the accent colour. While an approval card is open, or in a phone window 600px tall or less (an
Android phone with its keyboard up), the rows fold into that one line; tap it to open them. On
a phone, Discard is an X, and the message box keeps the same side margin as the
conversation.
(#466) -
In the chat, an agent's question, approval or plan no longer sits at the end of the
conversation: its card is docked directly above the message box, on the same column, so what
is asked and where you answer are one block and the conversation scrolls freely behind it.
Held messages stay above it, folded to their one line. The card is plainer: no shadow, its
title is the one thing in red (the "INPUT NEEDED" badge is still read by a screen reader),
and the options are flat rows with their number as a key, all of equal weight. A command or
plan to approve stays in the card's code box and is the only part that shrinks when room is
short (six lines, down to two); the card itself is at most 60% of the window, or six rows on a
phone with its keyboard up, and scrolls beyond that with Confirm always in sight. The message
box says "Type 1–3 to choose…" (or "Type 1–3 or your own reply…") in place of "Answer above:
type 1–3 to choose…", which was cut on a phone. While the agent waits for you, the open work
block reads "Needs you" with a still red dot in place of "Working…". Answering works as
before: a press on an option, a typed number followed by Confirm, or your own reply; Enter in
an empty box answers nothing. After a press with a mouse or keyboard, focus goes to the
message box, on a tablet or a touch-screen laptop too; a tap never moves it there. The card keeps following Settings → Chat font size and Chat font, and on a phone
with its keyboard up a tap on the card's text, or a drag down it, puts the keyboard away as
on the conversation.
(#468,
#475) -
In chat, a tool row reads as what the agent did: "Read src/metrics.ts", "Edited
src/pages/Reports.tsx", "Ran pnpm test", in place of an icon, the tool's id and a slash. The
rows start on the same left edge as the block's header and the prose. The tool's own id
(exec,Bash,apply_patch) is the row's tooltip and the first line of the row once
opened; a tool without a verb keeps its id in the row. While the agent works, what it says
between tool calls is in the answer's size and colour instead of small and dim.
(#457) -
The chat transcript is quieter. A finished turn folds its work under one dim "Worked for" row
and only the running turn stays open; a block you open, fold or work inside stays as you left
it, and a turn that ended without an answer keeps its words in view. The line
above your messages is gone, answers sit closer to their prompt, and your bubble has even
corners (no edge in the dark amber, report and charcoal palettes). File paths and links are
underlined quietly and take the accent color on hover or focus. With a mouse, a code block's
language and copy button appear in its corner on hover, a message's time and copy sit beside
the bubble, and an answer has one copy button with "Plain text" beside it; touch keeps the
code strip and the MD and TXT buttons.
(#458) -
A chat whose conversation was read and holds no messages yet asks "What should Omo do in
my-project?" over the message box, with the PC and the full path under it, instead of a dim
"No conversation yet" line in the middle of an empty pane. In a desktop window the question
and the message box sit in the middle of the pane until the first message is sent; on a phone
the box stays at the bottom. The question is not asked while the agent is working or waiting
for an answer, or while a message is held for it. A pane without an agent, or one whose
conversation could not be read, looks as before. That still includes a new Claude Code or
Codex pane: neither writes its conversation before the first message, so the chat shows the
terminal output until then.
(#460) -
The header is one line at every width: the pane's title, then PC › workspace › folder beside
it. The folder shows as its last name, and only when the title, the PC or the workspace does
not already say it; where the line has no room for the whole of it, it is left out rather than
cut. The full path is in the header's tooltip and at the top of the new ⋯ menu.
(#461) -
New tab, Browse files and Alerts are the items of one ⋯ menu in the header, in
place of three buttons; on a phone it also opens the command palette, whose own button left
the phone header. Keyboard shortcuts are unchanged. The Alerts item says this device's state
in words (On in the app, On in this tab, On, pushed to this device, Off on this
device), and the ⋯ button carries a dot while alerts are off on this device. This
reverses the old bell, whose outline lit while alerts were on.
(#461) -
The connection chip shows only while the app is reconnecting or disconnected; live is no
longer written out. In a window narrower than 900px the chip is its pulsing dot.
(#461) -
From 769px wide the sidebar has its own top row, with its toggle and the search, and the chat
has no bar above it: the conversation scrolls under a short fade. The header is 46px there in
the comfortable density too. Phones keep their bar. The update notice and a PC's "update the
bridge" line are drawn over the pane, not across the window.
(#461) -
The sidebar's footer ends at Settings and the plan meters: the "herdr web ui v…" line and the
herdr version beside it are gone from under them. Both versions are read in Settings, in every
state: Updates always opens with the running app version, and herdr shows the running
herdr version also where herdr cannot be updated from the app, such as on Windows. A tab that
has not been reloaded since the server updated says which version it still runs.
(#462)
Fixed
- An answer tapped in the chat for a row the cursor is not on looks at the screen again before
it presses Enter (or the first key that does more than move the cursor). A menu answered in
the terminal while the answer was moving the cursor could be replaced by another menu, which
then took the Enter; the answer now stops with "the prompt changed" unless the screen still
shows the card's menu with the cursor on its row, having pressed only ↑ and ↓. It also stops
once the agent is back at work under it. And a question asked again with the same text is a
card of its own where the app can see that the first asking ended (it was answered from the
app, the agent went back to work, or the prompt left the screen), so a typed pick waiting for
Confirm, or the card still open on another device, no longer answers the second asking. Not
covered: the keys after the first one that does more than move (the text and Enter of a typed
answer, the later ticks of a multiple choice), and a prompt answered in a terminal and asked
again word for word with nothing the app can see in between. Remote PCs get this with the
nextremote-vNruntime.
(#470) - The installed app opens without a network right after an update that renewed its offline
store, as this one does. The old store used to be emptied the moment the new version took
over, so until the app had been opened online once more, a reload with no connection showed
the browser's error page. The old store is now kept until the new one holds the app.
(#473) - Ctrl+Enter typed into the live terminal reaches Claude Code as Ctrl+Enter, so a message typed
while it works is sent at once instead of waiting in its queue. The terminal sent plain Enter; it
now sends the key the way the pane's program asked for (modifyOtherKeys), also to a device that
opens the pane later, and keeps Enter for programs that asked for nothing. A phone's input line
and a Windows mirror pane send Enter as before.
(#471 by @WOULDU-pres) - A draft in the message box keeps its full height when the window or the pane is resized, or
the chat width changes: the box used to keep the height of its old line breaks until the next
key press. (#463) - With a browser font size other than 16px (from 11.4px up), the Default chat width is never
wider than Wide.
Its widest was a fixed 960px while Wide follows the font size, so at a 13px font Wide (936px)
made the chat narrower than Default on a large monitor. Default's widest now follows the font
size too (960px at 16px, 1200px at 20px), and it is never narrower than Narrow's 820px: under
an 11.4px font that 820px is itself wider than Wide.
(#474) - An update asked for in the first moments after the app starts, or right after another update,
waits until the start is over. It used to stop the app in the middle of its start-up check,
and the app then fell back to the source checkout or did not come up.
(#455) - A PDF on a remote PC opens from the file viewer on Android. Chrome there shows it as an
Open button, and opening it answeredinvalid_origin("Use PC controls from this app")
instead of the file; a PDF on the app's own PC already opened.
(#448 by @nahwan-kim) - In one tab of the app, the alert sound chimes once for alerts that come together from several
panes, instead of sounding over itself. A pane that needs input right after one that finished
still chimes, once the first chime ends. Two open tabs each chime, as before.
(#439 by @WOULDU-pres) - In the terminal on macOS, Cmd+Left and Cmd+Right move to the beginning and end of the
input line, using the same terminal keys as Ctrl+A and Ctrl+E.
(#437 by @WOULDU-pres) - The DAG viewer pane omo-herdr-dag opens beside an OmO pane no longer appears in the
sidebar or the tab strip: an OmO workspace with its viewer shows as a single pane, as it does
without one. The viewer can still be opened from the command palette.
(#447 by @nahwan-kim) - Attaching a file over 8 MB says so at once, with its size and the limit, instead of uploading
it first and answering/api/pane/image failed (413). Past about 96 MB that message gave no
reason at all.
(#446) - Switching to another pane no longer shows the previous pane's chat for a moment before the
new one loads.
(#422 by @Haeminway1) - Remote bridge updates use Windows' native tar even when Git's tar comes first on PATH.
Independently managed web servers are directed to their own update controls before any
remote bundle is installed, instead of repeatedly attempting an update that cannot own them.
(#424 by @islee23520) - Chat reads a running GJC session in a pane herdr names no agent for, also when the pane still
carries an earlier Codex session report. It used to follow that report, ask the wrong agent
and fail to load.
(#423 by @Kinetic27)