macOS PTYs split unframed writes into 1,022-byte reads, so UNFRAMED_AGENT_INPUT_MAX_BYTES = 1024 left a gap where single-line chat inputs could split across multiple unframed paste reads in Claude Code and other agent TUIs. Lower UNFRAMED_AGENT_INPUT_MAX_BYTES to 512 UTF-8 bytes so long single-line inputs are bracket-framed.
Fixes#32071
The orchestration skill guide and reference examples hardcoded --agent codex, leading coordinators on hosts without Codex installed to fail at agent_readiness. Replace hardcoded worker-start agent ids with <agent>, document how to pick an agent id across shared vs remote/WSL/--on execution hosts, and regenerate bundled skill guides.
Fixes#30495
Disable commit/tag GPG signing via a cached wrapper GIT_CONFIG_GLOBAL when
the developer's global gitconfig enables signing, clear inherited GIT_CONFIG_GLOBAL
in GitBinaryTestFixture, strip inherited TERM_PROGRAM* in Vitest setup, and
run ssh-posix-command-wrapper shell checks against an empty HOME/ZDOTDIR/XDG_CONFIG_HOME.
Fixes#27210
Co-authored-by: outof-place <contact@outofplace.space>
Callers such as getRepoUpstream and the tracked-upstream branch fallback called getGitHubApiRepositoryForRemote without checking the cached remote listing, causing repos without an upstream remote to run failing git remote get-url upstream execs on cold lookups. Gate non-origin remotes on shouldProbeGitRemote at the shared entry point while keeping origin unconditional.
Refs #22684
Co-authored-by: h5k <31630310+h5kk@users.noreply.github.com>
Escape or an item that skips refocus unmounts the focused menu onto <body>;
refocus the menu pane then, but leave focus where an outside click put it.
Fixes#21015
Co-authored-by: ckdwns9121 <40492343+ckdwns9121@users.noreply.github.com>
Decorated Claude task titles such as '◐ Alternativas ao Gemini Live' or '. Gemini rate limits' matched the bare gemini token, causing normalizeTerminalTitle to rewrite them to '✦ Gemini CLI'. Behind a spinner or '. '/'* ' prefix, require a Gemini label frame in any wrapper segment while keeping native Gemini glyphs decisive, and reuse the shared detector in terminal-title-agent-type.
Refs #29266
Co-authored-by: Mateus <mateuslannes100@gmail.com>
* feat(native-chat): run GitHub Copilot in Chat over its ACP server
Copilot chats launch `copilot --acp` with Copilot's own sign-in (or gh's),
so enterprise hosts and managed policy carry over. A Copilot dialect turns
its tool results into plain output and exit codes, and withholds the
commands that would switch Copilot to approving everything behind the
chat's permission mode.
The CLI version reader now accepts a version that ends a sentence, as
Copilot prints it.
Co-authored-by: Morten Trydal <trydis@hotmail.com>
* fix(native-chat): show Auto when a session lists no models instead of hiding the model pill
* fix(native-chat): don't name Copilot chats with a Copilot run that stays in its history
---------
Co-authored-by: Morten Trydal <trydis@hotmail.com>
* perf(orchestration): an unreadable item holds back only its own conversation, and remote-owned queries use their indexes
* fix(orchestration): a held conversation stays skipped for its batch, and pulls past the hold-back cap wait locally
* Fix Codex account status for Bedrock and other built-in providers
Derive system-default authentication from the configured provider and honor the declared credential mode before reading ChatGPT identity.
Co-authored-by: Christopher Ryu <cryu17@gmail.com>
* Reuse provider classification until canonical config changes
* Bound provider config work and refresh after file removal
* Use portable module counters for provider config regression checks
* Complete the host target in the account warning fixture
* Include the account warning module in Node test typechecking
---------
Co-authored-by: Christopher Ryu <cryu17@gmail.com>
* fix(jcode): show a swarm helper's permission request on the main row
Since helper hooks are fenced off the coordinator's row, a helper blocked on
request_permission left the pane reading working while the swarm waited on a
user who never saw the question. A helper's request now restates the
coordinator's row as waiting with the helper's question, and its answer (or
the helper's turn ending) puts the saved row back. A rested row comes back as
a session boundary so the answer does not read as a second finished turn. The
root's own next hook drops the saved row.
* fix(jcode): restore helper permission waits onto the pane's current row
Restore the saved status onto the row the pane has now, so a pane move
between request and answer does not restore onto the old pane key. Ignore
helpers from an earlier launch, keep the wait until the main agent's row
actually replaces it rather than on any of its hooks, and show the next open
question when one of several helpers settles.
* Keep browser chrome metadata current without synchronous guest reads
* Provide the guest Escape subscription in browser pane fixtures
* Provide native navigation state in browser guest fixtures
* Preserve browser navigation and loading across asynchronous metadata reads
* Retire browser metadata when its actual guest is removed
* Explain native identity casts in browser disposal fixtures
* Preserve URL titles for untitled client browser pages
* Preserve current browser navigation and untitled page labels
* Keep merged browser failure type within lint limit
* Keep merged browser failure type within lint limit
* Bound published browser titles to the existing wire limit
* Use template literals in browser title regression
OpenCode, OpenCode 2 and Z Code keep every conversation in one shared
database, so Session History could only hand off the row's preview
messages. The session's own host now copies that one conversation into
the per-chat hand-off snapshot the chat Hand Off already uses, and the
dialog points the new agent at it. Older hosts, SSH rows and failed
reads keep the preview.
* fix(orchestration): tell the coordinator when a worker is lost
A worker that could no longer run its task sent no mail: a failed start,
a person closing its pane, tab or chat, or a terminal gone after a restart.
Each now sends one "Worker lost without reporting" alert through the
turn-stop report, with its cause, last output and any leftover workspace.
An unexpected exit and a lost Codex app-server use the same path. An exit
or loss says the Dispatch is closed as failed and gives the retry command.
Failure text in the alert is redacted like terminal output.
The coordinator's own worker-stop stays silent, locally and on a paired
server. A finished turn is not reported when that turn's worker_done was
rejected; later turns still are. worker-read and worker-show say
"never started" for a start that never got a terminal.
* fix(orchestration): a lost worker's restart command names its Task id and suits its Task's state
- The command named the Task's label, not its id.
- Restart recovery puts the Task back to ready, which --retry-of refuses:
the command is a plain worker-start then.
- worker-show does not call a paired server's failed worker "never
started": its record carries no terminal details.
* test(orchestration): import the turn-stop report after main's test cleanup
Jcode's daemon fires an inline swarm worker's hooks with the coordinator
pane's key, so the worker's turn_end settled the main agent's turn and its
model and session replaced the main agent's. Jcode saves each worker's
session file with parent_id set before the worker runs, while the pane's own
session has none. The hook owner preflight reads that field once per session
(a 4 KB prefix, async, cached) and normalization drops hooks from sessions
with a parent.
Co-authored-by: TheophileDiot <tdiot@bunkerity.com>
* fix(emulator): keep wheel scrolling moving across simulator screen edges
* test(emulator): keep wheel tests to the behavior contracts
* fix(emulator): drop wheel input when the control socket is closed
Replaying each edge segment through the gesture RPC opened a socket per
segment, could overlap them on one finger, and turned small scrolls into taps.
* fix(emulator): stop wheel playback on rotation and ignore retired control sockets
A replaced socket's late close reset the live socket's connected state and
keyboard timers, cancelling scrolls on the new socket.
---------
Co-authored-by: Kien Le <122910950+kien-ship-it@users.noreply.github.com>
Co-authored-by: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com>
* fix(jcode): keep the agent row for Jcode panes and read its working title
On Linux the Jcode release runs jcode-linux-<arch>.bin behind a jcode wrapper, so
the foreground read never identified the pane, and Jcode's own title
(<emoji> jcode/<server> <Codename> · work ~2s) was not classified. Once Jcode
replaced its startup title, the title-derived row lost its only identity and
disappeared. Map the platform binary to jcode, accept the server-qualified
identity title, and read the work/last tail as working/idle.
* fix(jcode): read the bare title tail only in a Jcode pane
The activity tail on conversation and SSH titles is too generic to prove an
agent, so title-only readers must not treat it as one. Gate it on the pane's
agent, as Hermes markers are, keep identity titles self-describing, narrow the
server segment, and keep the title strip linear.
* fix(jcode): keep the turn tail out of title-only readers
Title-only readers (completion tracking) saw a turn from Jcode's identity title
and then idle when Jcode switched to a conversation title naming another agent.
Read the tail only through the pane-scoped reader, check it before generic
title status in the sidebar row, and stop treating a bare jcode/<dir> as Jcode.
* fix(jcode): escape the title separator so the orcad bundle stays ASCII
Usage attribution only knew live worktrees, so deleting one moved its
sessions out of the Orca scope on the next scan and Stats & Usage totals
went down. Remember every worktree Orca has had in a small usage-owned
history and attribute against live plus remembered worktrees, with live
rows claiming reused paths first. The history loads off the main thread,
scans wait for it, and an unreadable or newer file is never overwritten.
Neil
Brennan Benson
Kelvin AmoabaandMorten Trydal
OrcaWinandOrcaWin
Noah
Jinwoo Hong
itigoppoanditigoppo
Wooseong Kim
gum798and서정화
Nawapat Buakoet
Dat
Aashish Mahato
mousepotato03
registered2nd
