You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
An OmO pane shows how many background tasks it still has running, as a small count beside its
state in the sidebar. A turn can be DONE while tasks it started still work, and they wake the
session by themselves; the count tells that pane from one with nothing left to do.
(#287)
Settings → Appearance has a wheel scroll speed for the terminal, from 1× (as it was) to 10×.
xterm sends herdr at most one wheel report per wheel event however far the wheel turned, so a
long history took a lot of turning; at 3× the same turn scrolls three times as far. A trackpad
is scaled the same way; a pinch and touch scrolling are unchanged.
(#274 by @taehwan08)
A file path or a file:/// address in the terminal opens in the file viewer with a click or
a tap, as one in the chat does. A path counts when it names a folder (src/App.tsx, server/index.ts:120); a bare name such as README.md stays text, since a terminal is full
of those and a tap to focus the pane would open the viewer. In the chat a file:/// address
opens too, written plain, as code or as a link, and one that names a folder shows the folder's
files. process.env, Math.random and the like no longer read as files in the chat.
(#266 by @beomq)
A pi or omp pane says so when you type /tree into the chat. The command runs, in the terminal,
and opens a tree the chat cannot show — so the composer says that while you are typing it, and
where to go instead. Wording checked at 390 wide: the lens switch is an icon there (its word is
hidden below 560px, and a title is a tooltip a phone never shows), so the hint names where the
button is rather than a word no phone displays
(#262 by @diegolhambi)
A pi pane says when the conversation is bigger than what the chat shows. /tree moves pi's leaf
pointer and writes nothing, so the turns a /tree left behind simply disappeared from the chat
with no sign they had been there. The chat now names how many it holds back and across how many
branches, and where pi summarized the abandoned path — answering /tree's Summarize branch?
— it shows that summary. Going back to those turns still means /tree in the terminal: the chat
reads them but cannot move pi's pointer, and a chat that could would offer to undo a branch by a tap
(#262 by @diegolhambi)
A pi pane is a chat. Its own session file under ~/.pi/agent/sessions is read through herdr,
so prompts, thinking, tool calls with their results, the model and the reasoning level show
the way they do for omp. /new, /resume, /fork and /clone move the chat to the session
that replaced the old one, and after a /tree it shows the branch in play: the paths pi
navigated away from stay in the file, unread, as they do in pi. Where pi folded old context
into a summary — on /compact, or by itself — the chat says so and opens the summary. A pi
dialog answers from the chat too: an extension's question, a Yes/No, and one that wants text
typed, though not /tree itself, which would move the session's branch. pi calls itself idle
while it waits on a dialog, never blocked, so the card comes from its screen. A session
directory chosen with --session-dir or the sessionDir setting is not found; PI_CODING_AGENT_SESSION_DIR is.
(#262 by @diegolhambi)
A picture a pi tool opened shows in the chat. pi keeps the image a read returned in the
session file beside the text it answers with, so a screenshot the agent looked at was in the
transcript and invisible here; it now shows inside that tool's row, and opens full size on a
tap. Only the address travels in the conversation, so the bytes arrive when the row is opened
and not on every poll, and an image left behind by a /tree stays out of reach, the way its
output does.
(#262 by @diegolhambi)
A pi pane's context ring shows. What a request filled was already recorded; the window it
filled was not, so the ring had nothing to divide by and stayed hidden. It now reads the
window from the same ~/.pi/agent/models.json pi reads its own providers from, and the
figure matches what pi's own footer says. A model that file does not state keeps no ring:
pi answers those from its built-in catalogue or from a running llama.cpp server, neither of
which a reader of the transcript can ask, and a window invented would show a percentage no
different from the real one.
(#262 by @diegolhambi)
A pi pane's message box lists pi's own slash commands as you type /, the way it does for
Claude Code, Codex and omp: pi's built-ins in pi's words, the templates in ~/.pi/agent/prompts
and the skills under ~/.pi/agent/skills and ~/.agents/skills. A project's own skills and
prompt templates stay out of the list, because pi loads those only once you have trusted the
folder, and a command the agent will not run is worse than one missing from the menu.
(#262 by @diegolhambi)
The browser demo runs a pi pane, backfilled into the recorded snapshot when the fixtures
predate it: site/demo/fixtures.ts is the only file a new demo agent needs.
(#262 by @diegolhambi)
Changed
Remote PCs use the remote-v10 runtime, which carries the server-side fixes since remote-v9
to them, Windows PCs included: New session with Gajae Code, the chat of a Gajae Code or omp
pane, pasted and sent lines that stay in an agent's message box, typing that no longer outlives
its connection, and OmO detection. A PC connected with the remote-v9 bridge updates as it did
for earlier runtimes. (#282)
Removed
/tree is no longer suggested in a pi pane's chat. pi offers it, and typing it works, but the
chat reads the tree browser it opens as nothing at all — no card, and the pane still looks idle
while the terminal waits for arrow keys — so a reader who picked it from the menu landed in a
state only the terminal lens can leave. Navigating was never the chat's to do; omp curates its own
list the same way. What the chat does say is where a /tree left the conversation: the marker
above the turns names how many it holds back and across how many branches
(#262 by @diegolhambi)
Fixed
A Codex pane whose answers are all short shows its chat. The chat finds a Codex conversation by
an answer it can see on the pane's screen, and only one of 64 letters or more counted, so a
session of one-line answers stayed empty until a longer answer came. When nothing else tells
(herdr names no session, Codex was not resumed, and no long answer found the conversation
before), the newest two or more answers of a conversation, shown one after another on the
screen, now count together. That is weaker evidence than one long answer, so it is used only
when exactly one of the folder's conversations shows that way, no other one said those answers
as far as its text tells, every one of them was read whole (none missing its file, forked from or continuing
another, or too long to read at once), and the folder has no more than 32. An answer with a
link in it does not count, since the screen shows a link its own way. It is never
remembered: once the answers scroll away the chat is empty again until a long one shows. Text
pasted into the pane or printed by a command can still look like another conversation's
answers and show that conversation instead.
(#284)
The chat of a Gajae Code pane on a Windows PC follows /new and /resume. It stayed on the
previous conversation until an answer of the new one was long enough to recognise, which a
one-word answer never is. It now goes by the session title gjc shows in its status line: when
that names one session of the folder, the chat shows that one, and when two sessions share
the title it shows none rather than the wrong one. A later gjc that Windows gave the same
process number no longer inherits the old session, and a process list that could not be read
keeps the chat as it was instead of dropping it.
(#285)
An OmO pane reads RUN while OmO works and DONE when it finishes. herdr reports nothing for such
a pane (it read READY or DONE whatever OmO did), so a message sent meanwhile was not held, Stop
was not offered, and no done alert came, in the browser or by web push. The status is now read
from OmO's own session file, also for a turn a finished background task, a monitor or a goal
starts by itself. Nothing of OmO's or Claude's is installed or changed for it. A remote PC gets
this with the next remote bundle.
(#287)
On a PC whose Tailscale node is tagged, your own devices are no longer refused as "another
Tailscale user". A tagged node has no person's login, and the server took the node's own name
for one, which no device could match. Such a PC now asks every device to pair, yours included,
and HERDR_WEB_TAILSCALE_OWNER names the login that gets in without pairing.
(#280)
Several lines sent to an agent from the terminal's input line on a phone stay in its message
box until sent, on a Windows PC too. The mirrored terminal there never learns the agent's paste
mode, so the lines went as typed and the agent sent the first one alone. They now go as one
paste, as the same lines pasted into that terminal do since 0.3.40.
(#281)
The chat of an OmO, omp or Gajae Code session shows the reasoning level the session runs at
now. It read the level from the first 64 KB of the transcript and the newest page, so a level
changed in between (/thinking, or a model switch) was never seen: a session started at high and switched to medium still showed high. A model or thinking-level change between
them now counts, in order, without reading more of the file.
(#279 by @ddotz)
The chat shows the conversation of OmO installed with bun add -g omo-ai again, and the
sidebar marks its pane as OmO. A global bun install puts OmO's engine next to omo-ai rather
than inside it, so the pane runs bun …/@code-yeongyu/senpi/dist/bundle/cli.js --extension …/omo-ai/plugin, and 0.3.40 no longer took it for OmO: the chat found no conversation and
New session did not see OmO start. That engine counts as OmO again when omo-ai's plugin is one
of its extensions.
(#278 by @ddotz)
Claude usage on a Mac no longer shows expired while Claude Code is signed in. Claude Code
started outside the desktop session (over SSH, or by a background service) cannot write its
keychain item, so it refreshes only ~/.claude/.credentials.json and the item keeps a token
that expired hours ago. The app took the keychain item whenever it could read it; it now
takes whichever of the two expires later, the way it already does for Cursor. Both are still
only read.
(#277 by @ddotz)
Claude's subscription usage no longer shows a locked keychain on a Mac whose server was
started outside the desktop session (over SSH, or by a detached multiplexer), where security cannot open the login keychain. The item is read through a one-shot launchd job
in the desktop session instead.
(#255 by @ddotz)
A pi pane no longer shows a context window pi never used, and now shows the one it does. A
provider that states contextWindow beside a model that names none reported that number as the
model's denominator; pi does not inherit it, and runs the model at its own default instead, so the
ring is now quiet there rather than showing a figure pi never had. And modelOverrides is applied
the way pi applies it — an override that raises a model to 256k was read at 128k, which drew twice
the utilisation pi shows. Both checked against pi 0.87.1 by pointing PI_CODING_AGENT_DIR at a
crafted models.json and reading pi --list-models.
(#262 by @diegolhambi)
A malformed entry in ~/.pi/agent/models.json no longer takes the chat down with it. One model
that was not an object — a stray null from a hand edit — threw while the window was being read.
(#262 by @diegolhambi)
A prompt template reached through a symbolic link is offered again. A .md linked from a dotfiles
repo into the prompt directory was hidden from the menu while pi loaded it and ran it.
(#262 by @diegolhambi)
A command name longer than the menu no longer runs out from under it, on a phone or a desktop. A
template is named after its file, and the name kept its full width past the row's edge.
(#262 by @diegolhambi)
A model list caught mid-redraw is refused instead of offered in part. A screen written while pi is
still drawing it ends a row inside its provider bracket; the rows above it were kept and offered
as a catalogue, so counting down walked into models pi had not drawn.
(#262 by @diegolhambi)
A path abandoned at the very root of a pi session is counted. Navigating pi's tree to its first
entry clears where the next entry attaches, and starting a second path beside the first was read
as no navigation at all, hiding what was left behind.
(#262 by @diegolhambi)
A pi dialog shows on a phone. A phone leaves pi a pane barely wide enough for its own hint, so
the hint wraps, and pi keeps its footer underneath: the wrapped hint then sat further from the
bottom than the window that tells a live dialog from an answered one reached, so no card came up
at all. /model failed twice over — the wrapped hint, and a model's own name wrapping so that
the provider's bracket, which is what tells a row from pi's notes, fell onto a line of its own
among them. The wrap is joined back before the screen is read, and a row still missing its
provider voids the whole reading, because a name cut in half would switch pi to a model that
does not exist
(#262 by @diegolhambi)