* fix(mobile): make file links in phone plan approvals actionable
The phone chat view now provides its file opener through one per-chat
context. Chat markdown renders through a shared adapter that reads it,
and tool, diff and attachment targets read it directly, replacing the
onOpenFile props threaded by hand through rows and cards. The plan
approval card was the surface that never got it.
Tracked in STA-13722.
* fix(mobile): open attached files while editing a queued message
Composer attachment chips read the chat's file opener from the per-chat
context instead of the structured composer config, which the queued
message editor built without it.
* test(mobile): provide the chat file opener to the attachment preview test
* fix(agents): tell Codex subagent rows apart by task name instead of role
A Codex subagent's task name only reached the sidebar when one early read of its rollout succeeded; otherwise every child fell back to its role, so two workers both read "worker". The child poll now re-reads an unnamed child's own rollout, the first-line read allows the existing 256 KiB line cap, and a child with no task name or nickname carries the tail of its id, as Codex's own agent list does. Labels keep the path below the root so nested children with the same leaf stay distinct.
* fix(agents): retry an unnamed Codex child's own rollout without a parent transcript
A relay restart can leave a child hook as the only evidence, sent before the child's rollout was readable. Keep that path as pending identity work and retry it through the existing transcript poll, which now stays scheduled while a child still owes its name.
* fix(agents): carry the id tail to every child-row surface and bound identity retries
- Full-mode sidebar rows and dashboard subagent cards use the same suffixed label as compact rows.
- A role a host publishes as the name (Claude's subagent_type) no longer counts as a name.
- A final identity miss no longer re-arms polling; retries take turns, eight children per poll.
- Replace an error cast with a type guard.
* fix(agents): keep every identity read inside the per-poll budget
The parent-rollout poll now queues an unnamed child for the bounded, rotating identity retries instead of reading it directly, and full-mode rows keep the id tail for a child with no name or role.
* fix(agents): keep plain names on task-only rosters and guard the queued rollout path
A host that publishes only a task roster carries no role, so its names can't be told from roles and get no id tail. Update the sidebar lineage test for a role-only child's id tail.
* fix(agents): tell role-named tasks apart on task-only rosters
A host that publishes only a task roster sends Claude's subagent_type as the task name, so two such tasks read identically. A named agent task with no description now carries its id tail; an unnamed one keeps its kind label.
* fix(agents): a task path found meanwhile wins over a queued header read
Placeholder labels (e.g. unknown) count as unnamed when queueing, and a child the parent rollout named before its turn drops its pending read instead of being overwritten by the header's nickname.
* fix(acp): permission prompts from an agent's own work wait for you
A permission or question that an ACP agent (OMP, Grok, OpenCode, others)
asks during work it started itself, such as picking up a finished
background job, was answered "cancelled" at once and never shown. It now
shows as a card, raises the needs-input alert, and waits for the answer,
unless Stop, a steer or teardown cuts it short.
While OMP refuses a sent message as busy, a request from its own turn now
appears at thread level instead of being cancelled, so it never lands in
the held message's turn.
Tracked in STA-13625.
* fix(acp): keep later requests askable after a Stop, and honor the agent's withdrawal
A Stop or a card's Cancel left the chat refusing every later permission
request until the session ended; the stop now covers only the work it
interrupted. A request the agent withdraws with $/cancel_request now
withdraws its card. Cancel on a card from the agent's own work reaches the
agent even when no message is running.
* test(acp): narrow approval rows before reading their resolution
* fix(acp): a live request's card outlives its turn instead of being cancelled with it
Requests group with the active turn again, as before. A permission or question
the agent is still waiting on now stays pending and answerable when its turn
ends, is replaced, or is settled by the host; only the person's answer, the
agent's withdrawal, Stop, a steer or the session ending settles it. Settled
requests are pruned at turn end.
Stop covers only Orca's running message, so later requests from the agent's
own work can still ask. Cancel on a card with no running turn now reaches the
host from desktop and phone for every agent, and a question the agent
withdraws before its card opens no longer leaves a stuck card.
* fix(acp): never strand a live request without a card, and scope Cancel to it
A request the agent still waits on now always gets an answerable card, even
when its turn settled before the card was written, and answering it releases
the timeline's hold. Cancel on a card from an agent's own turn refuses further
requests from that turn until it ends. Desktop never turns Cancel on a card
into a whole-chat Stop on hosts that can't take a card cancel.
* fix(acp): a new message is never covered by a Stop of the agent's earlier turn
* fix(native-chat): a card's Cancel with no running turn passes the RPC schema
The cancel params required a turn whenever a card was named, so Cancel on a
card from an agent's own work (and Pi's turnless card Cancel) was rejected
before it reached the host. A background-task cancel still names its turn.
* test(rpc): a card with no running turn cancels by itself
* refactor: organize runtime semantic operations by concrete owner
* refactor: consolidate terminal tab close type imports
* refactor: finish runtime contract owners and mobile catalog snapshot ownership
Double-clicking a file or folder name no longer starts inline rename.
The second click of a fast pair on a folder now toggles it back, and a
file name double-click pins the preview like the rest of the row.
Rename stays on Enter and the context menu.
Update, Rollback, Recover and Restart of a managed server already reset
its crash-loop brake as Connect does, but left the user's Disconnect of
the server and its SSH host in place, so the action was refused. They
now lift the holds Connect lifts (server and host Disconnect, cancelled
sign-in, re-pair cap) without dialing the host first, since Connect's
server-update check would run before the action (a Rollback would then
roll back the update it just made). Status resumes once the action ends.
A host with "Reconnect on launch" off holds its servers disconnected at
every launch. Their status now says so ("Reconnect on launch is off for
<host>. Connect to resume.") instead of claiming a manual disconnect the
user never made. The error code is unchanged.
Refs #38347
* perf(plugins): read catalog images in batched Git calls
Each catalog refresh ran one ls-tree and one cat-file per image or
manifest read. Batch every file an extraction round asks for into one
ls-tree listing and one cat-file --batch read, then rerun extraction
over the loaded files until nothing is missing.
* fix(plugins): bound speculative catalog blob prefetch
---------
Co-authored-by: m4air <m4air@Mac.localdomain>
Co-authored-by: Neil <neil@stably.ai>
* perf(tab-bar): reuse each tab's activity status across unrelated store writes
* fix(tab-bar): retain only each tab's activity memo inputs
---------
Co-authored-by: m4air <m4air@Mac.localdomain>
Co-authored-by: Neil <neil@stably.ai>
Opening a branch in Source Control already compares it. The first status
and upstream loads then each looked like a HEAD move or a push and
queued one or two more full compares. Only treat known-to-different
values as a change.
Co-authored-by: m4air <m4air@Mac.localdomain>
Forge detection asks GitLab, GitHub, GitHub Enterprise, Bitbucket, Azure
DevOps and Gitea in turn, and each probe ran its own git remote get-url.
Share one read per remote across a single detection pass; the scope
closes when detection returns, so later reads see current config.
Co-authored-by: m4air <m4air@Mac.localdomain>
Each local terminal spawn with a Pi agent home rewrote three Orca-owned
extension files (~54 KB) synchronously on the main thread even when their
contents were already current. Compare the read we already do for the
ownership check and write only when the content differs.
Co-authored-by: m4air <m4air@Mac.localdomain>
The first agent-availability derive forced a second interactive login shell
moments after the Claude profile router's capture; settings changes did the
same per update. Recapture now accepts a capture started within the 10s
shell-environment TTL, and startedAfter lookups answer from a finished run
that started after the request.
Co-authored-by: m4air <m4air@Mac.localdomain>
The saved-password detector re-scanned every credential form on each
mutating page twice a second (forcing layout per input and walking the
page's buttons), but main only reads those snapshots for 2.5 s after a
login navigates. Main now tells each document at enable whether it wants
them.
Co-authored-by: m4air <m4air@Mac.localdomain>