Repository navigation
Conversation
…another machine or boot (stablyai#29952) On a shared home, a peer lock's PID is evidence only on the machine that issued it. Cache retention judged the instance lock and daemon PID records by PID alone, so a second machine evicted the slot another machine's live serve or daemon runs from.
…efault flip (stablyai#29949) * fix(ssh): keep existing hosts' orca CLI control switch off after the on-by-default flip Builds before the flip saved an explicit off as an absent value, so absent=on silently turned on hosts the user had turned off. On the first load after this upgrade, pin every saved host's absent switch to an explicit false and record a one-shot marker in settings so hosts added later keep the new default. * test(ssh): check the saved profile without a type assertion * fix(ssh): say the CLI control switch is on by default for new hosts at the call site too
…ts, not only the desktop (stablyai#29913) useWorkspaceServer keyed the server's status read on the desktop's connection state, so a server restart or update under a connected desktop never re-read what it supports. It now passes the server's computed state. Co-authored-by: Claude <noreply@anthropic.com>
…ath twin's shells, tabs and edits (stablyai#29954) * fix(projects): removing one host's project keeps its same-path twin's shells, tabs and edits Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): close the removed host's browsers, chats and late tabs under a twin id Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): attribute a twin's terminals by every PTY id it keeps, SSH included Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): kill a removed local/SSH twin's shells directly and keep its files out of reopen history Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): prune the removed twin's closed editors from the reopen stack Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): recheck twin survivors after the host call and keep reopen order in step Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): close a removed twin's simulators and prune its closed browsers Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): attribute an unstamped twin browser by its pages' host Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): plan a twin's terminal teardown so PTYs other tabs bind survive Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): treat a null browser host as local and close stale stamped editor tabs Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): a host stamp yields to a PTY naming another host; sweep tabless sleeping records Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): cancel a twin's pending launches, clear its PTY records, prune its local editors Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): an ownerless remote pane blocks a stamp; keep the focused surface on background twin removal Co-Authored-By: Claude <noreply@anthropic.com> * chore(projects): leave closeFile's active-surface choice to closeFile Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): keep the selection fields consistent after twin teardown Co-Authored-By: Claude <noreply@anthropic.com> * fix(editor): closing a background workspace's file keeps the focused surface Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): a file's own host outranks its tab stamp in twin attribution Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): kill only a twin tab's live PTYs and honor a browser's own host stamp Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): skip twin files the removal refresh already closed Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): kill a batch's live PTYs whichever tab the planner chose; background file close keeps focus Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): resolve a removed SSH copy's paired-runtime alias in twin attribution Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): stop the removed host's pane in a tab another host's pane keeps Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): judge a still-present twin surface on its latest evidence Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): apply the removed copy's runtime alias to launches and history; layout-only hints prove nothing Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): repair a selection naming the removed copy's runtime alias Co-Authored-By: Claude <noreply@anthropic.com> * fix(projects): leave single-host selection to the bulk purge; test background file close Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
… as stale (stablyai#29957) The preparation lock reason named only the owner's pid, so a host sharing the repo's disk (an NFS home) read another machine's in-progress preparation as abandoned and discarded it. The reason now carries the writer's machine identity (optional, appended after the fields older builds read); a different or unknown machine is unverifiable, and a reason without one is judged by pid only on a local disk.
…pen terminals (stablyai#29958) * fix(orcad): replace a retired terminal daemon when a client returns during idle shutdown * fix(daemon): drop idle-retired legacy generations when replacing the daemon * fix(orcad): treat a lost or late daemon retirement reply as committed * fix(orcad): settle a lost retirement reply with the daemon itself before replacing it * fix(orcad): bind the replacement to the retired daemon and honor the router fence * test(daemon): type the pending census * fix(daemon): skip the owner check only for a proven-dead endpoint
…say when a conversation isn't found (stablyai#29105)
… still starting (stablyai#29911) The + agent menu waited 5 s for the new tab's terminal and otherwise showed "Couldn't launch X — the terminal did not start." Any slow spawn (Windows, a cold or busy terminal host, a queued spawn) passed through that state and got the false toast before the agent came up. The pane's own spawn already reports a real failure in the pane, so the guess is removed.
…er (stablyai#29946) * fix(sidebar): order sidebar text the same on every device Sidebar names, project labels, project-group names and folder workspace names were ordered with localeCompare, whose result depends on the ICU data bundled with Electron, the phone WebView or Node. Order them with one comparator in src/shared/sidebar-list/text-order.ts instead: case-folded and NFC-normalized, digit runs by value, the rest by code point, and equal text broken by host-qualified id. Co-Authored-By: Claude <noreply@anthropic.com> * perf(sidebar): fold each sort label once per sort; reuse the phone row identity helper Co-Authored-By: Claude <noreply@anthropic.com> * fix(sidebar): ignore accents when ordering sidebar names Accents and case now only break exact ties, so "Éclair" sorts beside "eclair". Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
…tablyai#29124) * fix(native-chat): identify the owner when resume create is refused * test(native-chat): consolidate transcript refusal cases * test(native-chat): keep ownership coverage at the launch boundary * test(native-chat): update adoption fixture for early ownership admission * test(native-chat): align resume admission assertions with host ownership * test(native-chat): keep RPC coverage focused on ownership proof * test(native-chat): assert blank request presence before optional source * fix(native-chat): admit the actual CLI source during chat creation * fix(native-chat): satisfy source admission lint checks * test(native-chat): provide the waiting hook prompt * test(native-chat): isolate source admission fixtures from host homes * test(native-chat): canonicalize adoption fixture directories * test(native-chat): preserve early ownership admission in source stack * fix(native-chat): preserve resume refusal behavior and honest transcript feedback * fix(native-chat): retain import read evidence in typed result * fix(native-chat): preserve admitted source and prepared execution home * fix(chat): retain source evidence for dispatched adoption * fix(chat): keep source capture within module limits * test(chat): allow host evidence in create routing * fix(native-chat): admit an unsaved transcript under a linked account folder A missing transcript stayed unresolved while its account root resolved through symlinks, so containment failed and a not-yet-saved conversation showed a load error instead of the unsaved notice. * refactor(native-chat): move the newest-item import window into its own module Keeps journal-legacy-import.ts under the line limit after the source identity observer landed on top of the streaming import. * fix(native-chat): restore the refusal error import lost in the main merge * fix(native-chat): admit Pi transcripts against the root Session History walks A PI_CODING_AGENT_DIR that already ends in sessions, or any custom folder, made admission expect <dir>/sessions while the lookup searched <dir>, so a found conversation was refused. * refactor(native-chat): admit and look up CLI transcripts in one store per account Each agent's import described where an account keeps its conversations twice, once for the path check and once for the lookup, and the two could disagree (the linked-folder and Pi custom-root refusals). The import now names one conversation store per account; the create path resolves it once and both the hook-path check and the fallback lookup judge that same store. OpenCode's store is its exact database.
…e like a sidebar delete (stablyai#23593) * fix(workspace-cleanup): let select-all include the workspace you are in Deleting the current workspace is allowed like any other, so select-all now includes it; a live agent stays excluded. Cleanup removal now hands focus to a surviving sibling after the batch, as the sidebar's batch delete does. Linear: STA-4686 * test(workspace-cleanup): cover focus handoff after the dialog closes and after Delete anyway * fix(workspace-cleanup): move focus off a deleted active workspace as soon as its row is gone * fix(workspace-cleanup): don't reopen a workspace the user left while a cleanup batch runs The per-row focus handoff fired whenever no workspace was selected, so closing the successor's last tab mid-batch re-activated it (and spawned a new terminal) on the next row. Move focus only when the workspace the batch left the user on is actually gone and nothing else took focus. * refactor(workspace-cleanup): move focus off the active workspace when a cleanup delete starts Switch to the successor once, right after the batch rows are marked deleting, instead of repairing an empty selection after each row settles. The active row is no longer active when it is removed, so the store never clears the selection and no later handoff is needed. * fix(workspace-cleanup): never hand focus to a workspace in the same cleanup batch The scan qualifies local rows as host `local`, so their delete state lives under a host-qualified key, while a sidebar row with no stored host only matches an unqualified key. The successor picker therefore saw batch rows as not deleting and could focus one about to be removed, leaving an empty screen once it went. Exclude the batch's own ids directly. * test(workspace-cleanup): use scan-shaped hosts in the active-workspace focus tests Candidates now carry the `local` host the scan publishes, which is what exposed the picker gap. The other-host case also stamps the viewed row with its host so it resolves; before, the lookup missed and the test passed with the host check removed. * feat(workspace-cleanup): confirm before a cleanup delete stops a running agent A cleanup delete (bulk, single row, or "Delete anyway") that includes a workspace with a running agent now shows a step naming those workspaces before anything is queued. The check reads live agent state at click time, attributed like the sidebar so structured chats count too, and re-asks if another agent starts while the step is open. The approved candidate records which agents the user agreed to stop; the per-row preflight that runs right before each delete refuses a row whose agent started after confirmation. With the step in place, live-agent rows join select-all like the active workspace. * refactor(workspace-cleanup): name the agent-stop same-action check * fix(workspace-cleanup): drop an open agent-stop step when the dialog reopens mid-batch * style(workspace-cleanup): match the agent-stop step's heading to the delete step * test(workspace-cleanup): type the cache-hydration mocks in the reopen test * style(workspace-cleanup): let the agent-stop step's heading use the dialog's own typography The changed-lines design-system gate rejects restyling DialogTitle and DialogDescription; the delete step's overrides predate that gate. * fix(worktree-delete): key delete state by the identity the sidebar card reads A cleanup delete names local rows with the 'local' host, so their queued and deleting state landed under 'local|<id>' while a card stored without a host reads the bare id and never showed it. Every delete-state write now derives its key from the store's own worktree record. * refactor(workspace-cleanup): reuse the sidebar delete's focus hand-off Drops the move-focus-before-delete path. Each confirmed row prepares the same prepareActiveWorktreeFocusAfterDelete committer the sidebar delete uses and runs it when that row's own removal succeeds, so the user stays on the workspace while it is queued, lands on the most recently visited surviving sibling once it is gone, and stays put if its delete fails. * refactor(workspace-cleanup): reuse the close-terminal stop-agent wording The stop step's single-agent title is the existing "Stop this agent?" string, and its description follows the close-terminal prompt's wording. * test(worktree-delete): cover a local-qualified removal failure on an unhosted card * docs(workspace-cleanup): describe the after-delete focus hand-off in the bulk-select note * fix(workspace-cleanup): name one agent in the stop step's button and keep one agent term per locale * fix(terminal): keep starting terminals in a workspace that is only queued for deletion A cleanup batch now marks its local rows with the same delete state the sidebar card reads, so the terminal spawn guard sees them. The guard exists because removal just killed the workspace's terminals; a queued row still has them and can fail its own delete and stay, so a new terminal opened there (often the current workspace) stayed blank. * refactor(terminal): name the teardown-started check the spawn guards share * chore(workspace-cleanup): keep one agent term in ko/zh and refresh a spawn-guard comment * fix(workspace-cleanup): refresh consent and focus per removal * fix(workspace-cleanup): invalidate detected host evidence * fix(workspace-cleanup): admit removal at transport send * test(workspace-cleanup): expect the send-time admission callback and a published chat tab
…ames (stablyai#29966) Records the post-stablyai#29946 text order for workspace rows (Name, Recent and project-name sort), project group names and folder workspace names, ungrouped and grouped by project, so a future ordering change shows up as a golden diff.
…t reported, not the capped status store (stablyai#29905) * fix(native-chat): decide owed helper work from what the provider reported, not the capped store Idle cleanup, replacing an outgrown agent, /clear, /compact, rewind, command handover and the teardown resume marker decided whether a chat still had helpers running by reading the agent-status store. That store is a bounded display projection: it refuses records when it is full or for per-record reasons, and a refused helper is still running. The host now keeps an uncapped per-session set of the child work its provider reported running and has not reported over, fed from every child-work edge before the store sees it. Every owed-work decision reads that set. An entry leaves on the child's own end (a late end of an earlier run leaves the current run open, as the store's run verdict does), on a Stop the store settled, or on the session's end. The strip's "incomplete" note and Stop all's unlisted route are now derived: open work that no live current-run record answers to. This replaces the refused-for-room side list, so every refusal reason is covered and nothing needs separate bookkeeping. * test(native-chat): the real-store test sink reads recorded helper handles, as production's does * fix(native-chat): keep helper run identity through delayed events and Stop * fix(native-chat): name the host helper lifetime token * test(native-chat): explain the Stop adapter fixture boundary * fix(native-chat): derive helper coverage from running store records * docs(native-chat): clarify running helper coverage * test(native-chat): use an explicit helper end event * Scope background Stop to captured child work * Exercise refused helper Stop race through status store
…#29224) * fix(mobile): remove the dead space around the chat composer The Tools toggle floats over the transcript and shows only when the chat has tool calls; the composer is one row (attach, message, mic, send) until focused or drafted; Stop takes Send's slot while there is nothing to send; the bridge lane's working dots move to the transcript's tail. * fix(mobile): name the composer's content facts for their role * fix(mobile): preserve composer controls and picker focus * fix(mobile): typecheck working indicator test lookup * test(mobile): find the floating Tools pill when checking a chat switch keeps it expanded
…, for Settings and chat (stablyai#29049) * feat(acp): one install row shows and drives a managed runtime install, for Settings and chat The row reads the shared host state: Install, byte progress with Cancel, extract and verify lines, Retry after a failure or cancel, and in Settings Reinstall or Update once installed. A failed status read says it could not check, never that nothing is installed. Strings in all six locales. * fix(settings): the install row says when a copy is damaged, when a PATH copy is used, and when the host needs unzip - A damaged copy reads "This copy of {agent} is damaged." with Install, which repairs it. - A copy found on this host outside Orca's reads "Uses the copy of {agent} found on this host." with no Install, since a launch would never run the download. - A host without the program to unpack the archive says so, with no Install. * fix(settings): say why a host running its own copy is offered no download * fix(settings): the install row says Cancelling only where Cancel was clicked, offers Orca's copy beside one found on the host, and keeps its status line after a failed install The host no longer publishes a cancelling phase or a missing unzip, so the row drops both: the row whose Cancel is still out says Cancelling and holds its button until the host's cancel settles. A copy found on PATH or in an install folder still gets Install of Orca's copy in Settings (the chat keeps its composer). An older installed release reads Installed. with Update. A failed first install keeps Not installed. beside the failure and Retry. The status line is a polite live region. * fix(settings): the install row says it is retrying a failed status check, and names the step after the download * fix(settings): a Cancel that fails for no stated reason says the cancel failed, not the install
… the saved layout (stablyai#29961)
…ect groups (stablyai#29962) Relinking projects the server already held still re-imported this desktop's retained copy of their project groups, so a group renamed on the server while it was managed failed the reconnect with orcad_migration_project_group_id_conflict. A relinked move now leaves to the server every group its projects sit in (read from repo.list's projectGroupId) and those groups' ancestors unless a moving row needs one, journals those ids, and keeps them out of the exported rows while still reading them for folder-workspace host ownership. Later delta moves subtract them as they do moved groups.
…self (stablyai#29964) * fix(runtime): the host stops a removed project's terminals itself Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): a runtime-stamped project still stops the shells that run here Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): refresh the PTY inventory and sweep again after a project's removal Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): bound the project inventory refresh and recheck the WSL move fence Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): fence a project's removal sweep by its SSH connection and wait out in-flight spawns Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): bound the project stop's lease wait by the removal deadline Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): refuse spawns into a removed project and skip a stop a same-id sibling shares Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): bound each project stop by the removal deadline Co-Authored-By: Claude <noreply@anthropic.com> * fix(runtime): pass the removal deadline to the lease so a timed-out waiter is dequeued Co-Authored-By: Claude <noreply@anthropic.com> * test(runtime): type the project-removal fixtures instead of casting Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
…s again (stablyai#29984) Partially reverts stablyai#27540 and the plain-http part of stablyai#27552, which broke web pairing on upgrade for 1.4.227 users who open Orca at a LAN or Tailscale address: - The host no longer refuses a first pairing from a plain-http page on another machine (insecure_origin, ORCA_ALLOW_INSECURE_WEB_PAIRING). - Such a page keeps its pairing code in local storage and signs in with it, as in 1.4.227, so reloads, new tabs and bookmarks stay paired; no notice. - Session-only browser keys (pairing.device-key-ephemeral.v1, the idle sweep) stay for browsers without Ed25519. - A web client whose credentials are refused shows the host's reason instead of an empty workspace and "Session restore failed".
…-owner build only) (stablyai#29988) * fix(browser): keep slept workspace browser tabs off the reopen-closed list Sleeping a workspace closed its browser tabs as if the user had closed them, so Reopen Closed Tab could bring back a tab that was only slept. Sleep now closes them with a 'sleep' reason that skips the reopen stack while still releasing host pages, doc grants and webviews as before. * feat(layout): send a browser tab's window timestamps with its open (dark) A layout-owner build's openBrowserTab carried no createdAt, so main stamped the tab-bar entry and the browser record with its own apply time and the window's next save diverged on createdAt for both. The window now sends both times, as createTerminalTab does; only switch-on builds send them. * Revert "fix(browser): keep slept workspace browser tabs off the reopen-closed list" This reverts commit 6798c3c. * test: drop the slept-tab close test that went with the held sleep fix
… (dark, layout-owner build only) (stablyai#29967) * feat(layout): main-made terminal tabs drawn with main's facts; resume records reach main record-only (dark) In layout-owner builds only: - A terminal tab main made itself (drag-out, CLI or host terminal, live-surface adoption) is drawn with main's createdAt and creation fields from the frames, and the window sends no create for it. - An editor tab main records takes the time the window drew it. - Every change of the window's agent resume records (live and quit captures, dropped records) reaches main through record-only sleep and wake commands, which stop or start nothing. - A terminal row's remount generation is declared the window's. * fix(layout): real wake reaches main before the record drop; drag-out keeps a frame-drawn tab (dark) In layout-owner builds only: - A resumed pane's wake is sent before the window drops its record, so main's wake still finds the pane sleeping. - A drag-out whose new tab main's frame already drew shows the pane in that tab instead of throwing "Terminal tab already exists". - A main-made tab keeps the window's shell, cwd, host-runtime and launch agent where main's copy names none (a runtime create records a shell only when its launch named one). - No record-only wake for a tab the window closed; main dropped its records with the tab. - Tests for the resume-record follower. * fix(layout): a dragged-out chat pane's new tab opens in chat view in main (dark) Main's movePaneToNewTab now gives the new tab viewMode 'chat' when the moved pane is its source tab's chat pane, as the window's own drag-out does. In a layout-owner build the window adopts the row main's frame drew, so without this the tab opened in terminal view. * fix(layout): a bound pane with a resume record starts as a no-op; resume records only for held tabs (dark) - startPane checks the pane's live terminal before its resume record: a running agent's live or quit record does not make its pane asleep. A pane with no terminal and a record is still refused. - The window sends record-only sleeps, like record-only wakes, only for tabs it still holds; a capture from a stale agent status no longer names a tab main already closed.
…ark, layout-owner build only) (stablyai#29987) * Agree with main on a workspace's first group and the tab order, and stop writing a preview flag main never stores (dark, layout-owner build only) Switch-on builds only; switch-off behaviour is unchanged. Group ids: a window save made before main's frame with a workspace's first tab carries none of that workspace's tab-bar records, so the Loader rebuilt the tab in the group it names for the workspace while main had minted a random one: every workspace's first save replaced main's group. Main now names a workspace's first group as the window's root group and the Loader do (`workspaceRootGroupId`), or takes the sender's when the window drew the tab in a group main lacks (the window keeps its emptied group); the window's `ensureGroup` names it the same way. The detector reads a lone group with no stored tree as that group's leaf, as main's own save writes it. Empty groups: the window never saves a group it holds empty, so a save cannot bring back one main closed (locked by a test). Tab order: a reorder the window makes itself (an anchored CLI or agent create, a reopened tab's old slot) is recorded in main as `moveTab` steps at the store boundary; callers that already asked main pass `layoutOwnerHandled`. The reopened-editor record this replaces is deleted. A paired client's move (`ui:moveSessionTab`) goes to main first through the tab-bar drop path. Preview: activating a tab removed its preview flag by writing `isPreview: false` on every tab; main stores no flag, so every activation changed the tab in main's model. The window now drops the flag instead. * Record a reopened tab's place in main every time, leave a move into a tab's own group alone, and keep a preserved preview untouched (dark, layout-owner build only) Review fixes, switch-on only: - A reopened tab's place is recorded as one moveTab whenever its old position is known, as the deleted editor-only record did, now for every kind; the reorder it may make is not recorded twice. - A move into the tab's own group (a paired client can send one) goes to today's no-op instead of a moveTab that reordered the tab. - Activation that preserves a preview returns the tab unchanged. Tests: switch-off first group stays minted; tabOrderMoves refuses a duplicate order. * Send a paired client's tab reorder to main as its whole order (dark, layout-owner build only) A client reorder names the group's full order, but it went to main as one moveTab of the named tab, so [a,b,c] -> [b,a,c] naming c moved nothing anywhere. It now takes the store's reorder path, whose record turns the whole order into main's moves (tabOrderMoves), the one rule for every window-side reorder. Moves and splits still go main-first.
…on suite (stablyai#29991) Seeds local+ssh, local+runtime and folder twins (one id on two hosts) with live shells, a dormant tab, dirty docs, drafts, sleeping-agent records and module-level stores, then runs a hand-declared action table (remove, folder delete, purge, kill, sleep, select) on one twin and asserts the other twin is untouched. Today's leaks are listed cell by cell and asserted to still leak, so a fix turns them red until the entry is removed.
…all in place of its composer (stablyai#29051) * feat(native-chat): a chat whose host has no program to run shows Install in place of its composer The chat's one input slot is chosen in one place: a final read refusal, then an owned prompt card, then Install while the host has nothing runnable (its install included), then the composer. Until the first status answer the slot stays empty instead of flashing either. Send and slash commands are refused while Install holds the slot, so nothing hidden reaches the host. The draft stays in its store and is there when the composer returns. A failed status read leaves the composer; the host remains the final refusal. * fix(native-chat): keep the composer slot reserved until the host's agent list says, and say nothing about sending while Install shows - While the host's capabilities or agent list are still loading for a chat's agent, the slot stays quiet instead of flashing the composer. - While the install row holds the composer's place, a failed start's notice ("Send your message to try again", Retry) is left out: nothing starts until the install lands. - The chat install card's title is semibold like the other cards in that slot. * fix(native-chat): keep the chat session view within its line limit after the install slot * fix(native-chat): the composer stays while the host's first install answer is out, its sends held, so no agent's chat loses its input box * fix(native-chat): the install slot holds Send through the composer's existing held-send reason instead of a flag of its own * fix(native-chat): the install slot hands the chat its held-Send reason, keeping the chat session view under its line limit
…ainst the window in dev and test (stablyai#29980) * feat(sidebar): main produces per-workspace activity facts in shadow mode * perf(sidebar): agent pings keep last-read surfaces; shadow counts comparisons * fix(sidebar): settle review findings in the activity-facts producer and shadow check * fix(sidebar): refresh on inventory verdict clears; resync agents only when tabs move * fix(sidebar): scope the shadow comparison by each row's owner, not the focused host * refactor(sidebar): name the shadow scope after what it keeps * fix(sidebar): classify sidebarInputs as host-owned in the operation ownership census
…tablyai#30137) Two merged changes left lines past the formatter's 100-column width in structured-session-create-support.ts, structured-agent-session-create-adoption.ts and transcript-opencode-sqlite-query.ts. Any branch that merges main gets them reformatted by the pre-commit hook, which pushes two of them past the 300-line limit and fails CI. Format them, and keep them under the limit by moving the bounded row readers and the resume adopt payload into the modules that own them. No behavior change.
stablyai#30356) (cherry picked from commit df460b8)
…ce and Project views (stablyai#29977)" This reverts commit 407bf03.
…stablyai#30436) * fix(native-chat): a Claude chat keeps a model its CLI stopped listing Claude Code 2.1.302 stopped listing opus[1m], which still runs, and chats on it moved to the list's default on the next send, effort remapped. The picker no longer derives a replacement from the list: the chat's own model shows as a selected-only entry, its effort read-only, and every send carries both. A real model_not_found reply still recovers: the options answer shows what the reset child runs, so the next send records it. A reopened chat's catalog read lists its own model so older clients that replace one keep it too. Tracked in STA-10947. * test(mobile): drop the send-options capability the options hook no longer reads * test: an unlisted model's saved effort shows read-only at rest, as live, for every agent * fix(native-chat): round-1 review fixes; a new chat's unlisted remembered model falls back as on main - The read-only value predicate keys on the effort row, so the phone's legacy host Permissions row opens as before. - The desktop read-only effort row honours a row kept from the previous model. - A live child with no listing of its own answers with the saved listing's ids, so the read-only effort reads the same live and at rest. - The read-only effort borrows listed rows' effort label and level names. - The host seeds a new Claude chat from a remembered default the saved listing no longer names on that listing's default, effort nearest, as main did; a chat's own saved or running model is never replaced. Tracked in STA-10947. * test: pin the new-chat seed fallback on its typed helper, with no casts * fix(native-chat): round-2: a carried effort moves with a fallen-back model; one guard - An at-rest create carrying an effort but no model, on a remembered model the saved listing no longer names, now moves that effort to the nearest level the fallback model offers, as main did. - A create whose own options name no model, and whose merged options carry none, reads no listing. - The seed fallback keeps one guard and no unreachable branch. - The read-only effort finds its level's name in any listed model's choices. Tracked in STA-10947. * fix(native-chat): a live options read with no record borrows no saved ids and never throws (cherry picked from commit 0322185)
…hile the agent holds the send (stablyai#30134)" This reverts commit fc925c3.
…ai-orca-ru-20261018T1400 v1.4.230 # Conflicts: # config/scripts/electron-builder-mac-channel-config.test.mjs # config/scripts/node-pty-windows-pty-teardown-patch.test.mjs # config/scripts/win32-test-lane-registration.test.mjs # mobile/src/session/MobileNativeChatView.test.ts # resources/skills/release-mapping.json # resources/skills/snapshot-registry.json # src/cli/specs/search.test.ts # src/main/ipc/ssh-relay-session-callbacks.ts # src/main/ipc/ssh.ts # src/main/ipc/worktrees/removal/remove-unregistered-worktree-host-home.test.ts # src/main/native-chat/agent-session-journal/journal-row-writer.test.ts # src/main/native-chat/agent-session-journal/journal-subagent-liveness.test.ts # src/main/native-chat/structured-agent-session-create-support.test.ts # src/main/providers/ssh-git-provider-diff.test.ts # src/main/providers/ssh-git-read-provider.ts # src/main/providers/ssh-git-remote-sync-provider.ts # src/main/providers/ssh-pty-provider-rpc-operations.ts # src/main/providers/ssh-pty-provider.ts # src/main/runtime/rpc/methods/structured-agent-session-optional-model-capability.test.ts # src/main/runtime/rpc/rpc-streaming-dispatcher.ts # src/main/ssh/ssh-relay-session.ts # src/main/ssh/ssh-remote-orca-cli.ts # src/main/stats/stats-snapshot-writer.ts # src/relay/git-exec-validator.ts # src/relay/git-handler-diff-read-coalescing.test.ts # src/relay/git-handler-registration.ts # src/relay/git-handler-submodule-ops.test.ts # src/relay/git-handler-sync-operations.ts # src/relay/relay-runtime-services.ts # src/relay/remote-artifact-cli-input.ts # src/renderer/src/components/native-chat/NativeChatComposerActions.test.tsx # src/renderer/src/components/native-chat/native-chat-composer-prompt-recall.test.tsx # src/renderer/src/components/native-chat/native-chat-observed-send.ts # src/renderer/src/components/native-chat/native-chat-queue-send-confirm.test.tsx # src/renderer/src/components/settings/setting-ownership.test.ts # src/renderer/src/components/sidebar/SidebarWorkspaceFilterSection.tsx # src/renderer/src/components/sidebar/visible-worktree-kinds.ts # src/renderer/src/components/sidebar/worktree-list/grouping/build-rows.ts # src/renderer/src/components/sidebar/worktree-list/grouping/worktree-grouping.ts # src/renderer/src/runtime/web-session-tabs-sync.ts # src/shared/git-exec-mutation.test.ts # src/shared/git-exec-mutation.ts # tests/e2e/structured-attention-relay-settlement.unit.test.ts
Resolve every path to its manifest owner after merging v1.4.230 onto the v1.4.227 sync (PR #120), re-home fork seams onto upstream's module splits, retire the fork's SSH-relay halves (upstream replaced the relay with orcad), and adopt v1.4.230's owner-routing and layout-writer ratchets in fork code.
There was a problem hiding this comment.
Your trial has ended. Reactivate Greptile to resume code reviews.
…lls check run-node-server-tests.mjs pinned the bundled CLI skills source to stablyai/orca, but the fork's skills-repository feature points installs at zpyoung/orca, so the Headless Node server persistence job failed on every fork PR.
v1.4.230 renamed the changed-code-quality and React Doctor steps, so the fork's >500-file skip was not replayed onto them. On a sync PR every upstream line counts as changed, and both scans report upstream code that upstream's own CI never sees. Baseline lint and typecheck still run.
… fork code Heimdall fork code built worktree ids by hand; it now uses composeWorktreeId and getRepoMainWorktreeId. The fork worktree-groups e2e spec now passes the fetch options and host id that upstream's store actions require.
…lysis The rebuilt pr.yml took upstream's two localization steps verbatim, which call verify-localization-extraction.mjs and verify-localization-catalogs.mjs directly and skip the fork wrapper that overlays fork-*/locales catalogs, so every fork key read as missing. Both steps now run fork-localization-catalog-check.mjs, matching package.json's verify:localization-* scripts, and the upstream scope test that pins the extraction command follows through a one-line seam.
…ware v1.4.230's workspace-attention slice reaches visible-worktree-options-from-state, whose fork seam imported the activity/review filter hook modules and through them @/store, so the store was built before that slice finished loading and every suite constructing the store crashed. The pure context readers now live in store-free modules that the hooks re-export. The two advisory benchmark workflows now report 'no baseline' when the benchmarked script is new in the PR, build the mobile-web bundle before the orcad smoke arm, run the fork localization wrapper, and give the compilers arm a timeout it does not pass only by luck.
…ests Upstream's new censuses (host-operation ownership, WSL-move storage) now require the fork's RPC namespaces and stores to be classified; fork seams no longer read state upstream test doubles omit (automation launch settings, group menu repos, activity option row); fork tests follow upstream's new shapes (settings-free composer sends, upstream code-fence coloring, owner routing mocks, store mock shape) and the fork's repository URL and app id. The sandbox runner now matches CI with an init process and ORCA_BACKGROUND_LAUNCH=1.
…f upstream's highlighter v1.4.230 colors native-chat code fences itself, so the fork's hljs path no longer rendered. Per the owner's decision the fork drops it (CommentMarkdown / message-row seams, .native-chat-code rules, their tests) and keeps the tool-glyph coloring. Inline code in chat prose loses the fork's accent chip, which rode on the same highlightCode prop.
- artifact-passwords: read the safeStorage backend through guardedSafeStorage so the store honors the keychain withhold (safe-storage import ratchet) - terminal-dock: move the remote-ConPTY dataset key into a UI-free module so pane-created no longer drags the editor tree into launch-parity tests - cross-version harness: map a module a pre-v1.4.228 baseline release lacks to its earlier location, so upstream's suite runs against the fork's 225-based release - fork tests follow upstream's new device-credential minting, attention 'viewed' semantics, nested session-info locales and fork fix-checks module - PR benchmarks: plan the native compiler with heap estimates that fit a 16 GB runner and retry the release-JavaScript jobs snapshot
…y mistake The benchmark run's local output under .build/ was swept into 3e4ed86; it trips the root-directory guard and the ownership guard's coverage check.
…k-with-upstream-stablyai-orca-ru-20261018T1400
… route Upstream's release-branch commit 5974258 copied main's pr.yml, adding a required dependency_isolation job, without the rest of stablyai#30254: the scope script never emits the selection and the job's test file does not exist in this release, so verify fails every PR with 'missing or invalid selection'.
zpyoung
deleted the
zpyoung/auto-sync-fork-with-upstream-stablyai-orca-ru-20261018T1400
branch
October 19, 2026 19:02
This was referenced Oct 19, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ELI5
This brings the fork up to upstream's newest stable release, v1.4.230. It's built on top of the still-open v1.4.227 sync (#120), so merging this lands both, and #120 then closes as merged.
v1.4.230 is much bigger than a normal sync. It was cut from upstream
mainafter ~1,400 commits, and it deletes the SSH relay the fork's SSH features ran on. As decided by the repo owner, the fork drops its SSH-relay code paths and keeps every feature's local behavior.What Changed
v1.4.230atb2879bd5aa, merged onto sync: absorb upstream v1.4.227 #120's head (d3dfc479b8, which carries v1.4.227). Resolution:auto-ours+tree(43 delete/modify tree conflicts, resolved by the two mechanical rules, then by the manifest).origin/backup/main-20261019-102904Zatc923f8e30a.SSH relay retired (owner decision). Upstream replaced the relay with a managed
orcadon every SSH host. The fork deletedsrc/relay/**and the SSH git/pty providers, and plainssh:hosts now report "not served".nativeChat.*RPC.orca ledgerdeferred item to re-home onto orcad RPC later.Module splits re-homed. Upstream dissolved the
orca-runtime-*.tsmodules, the sidebar grouping files,web-session-tabs-sync,ipc/runtime-environment-*and others. Every fork seam moved to the new home. Stale fork copies of deleted files are gone. Where upstream already did the same thing, the fork seam was dropped.Native chat. The fork's input-quarantine delta (bug-1) is re-applied on upstream's settings-free send path (stablyai#27279). All 15 tier-2 forked copies now point at v1.4.230. Hidden-pane read gating, which the fork had dropped only for relay reconnects, is restored.
New upstream ratchets adopted in fork code, with no baseline lines added:
headless-tab-properties.ts.text-[11px]→text-2xsin the plugin-attribution picker seam, and oneimport()type annotation.CI.
pr.yml,release-cut.ymlandrelease-mac-build.ymlare rebuilt as v1.4.230 plus the fork's own delta (fork ownership guard, macOS-only release, no Windows PR lanes, file-based changed-paths).verify.needs. Upstream's aggregate-gate tests learn about it through a newconfig/scripts/fork-ownership-guard/module plus two seams.Release-branch workflows. Upstream's
597425845eswapped 9 release-branch workflows formain's without updating the tag's tests, so 8 of the tag's own config tests are red against it. As in #120, the 8 upstream-owned ones take597425845e^. The next tag resets them.Retired exceptions:
ssh-remote-commands.test.ts: upstream now passes or matches;verify-skill-update-roundtrip.mjsandmax-lines-baseline.txt: identical to upstream;relay-windows-process-tree.ymland a staledeletedentry.Sandbox runner. It now ships
src/packagesand the native workspace manifests, because upstream movedsrc/shared/child-processinto@orca/process-host.Why
Taking v1.4.230 means taking the relay deletion. Keeping a fork-owned copy of the whole relay (480 files plus the SSH providers) would be permanent upstream-shaped ownership. Re-homing onto orcad in the same sync would have doubled its size and risk. Retiring the SSH halves keeps every local behavior and stays reviewable. The re-home is tracked in the ledger.
Linked Issue
N/A. This is an upstream sync.
Visual Proof
N/A. This is an upstream merge, and the fork's UI changes are routing-only. Nothing has been checked in a running app yet.
Testing
Local gate: all of these pass.
pnpm install --frozen-lockfile--verify-seams(1,211 lines / 605 files) and--verify-residuals v1.4.230(548 files)check-fork-ownership.mjson the committed headtypecheck:node,typecheck:webandtypecheck:cli, run one after anotherpnpm lintconfig/scripts+.githubtests on the remote sandbox: 4,763 passed and 5 failed.--ignore-scripts), installs a pnpm older than the pin, and ships git 2.43 without reftable. None can fail on GitHub-hosted runners.The full unit suite is left to this PR's CI.
I manually tested these changes locally
Automated tests added/updated, or explained why not below
Review
Agent skill upstream boundary
Notes
projectGroup.deleteandrepo.rmstay in the uncatalogued RPC list, because project-ledger extends them with main-side schemas.unwrittenin upstream's dark layout-owner store, so shipped builds are unaffected. A deferred item is filed for before that build switch ships.fork-heimdall/orchestration/federation-sync.tsuses a free-form copy header, so tier-2 replay can't see it.upstream/mainpast this head as of the fetch.Checklist
N/Awith reasonpnpm lint,pnpm typecheck,pnpm test, andpnpm buildpass (or CI will cover; local preferred)