* Skip completed Codex launch preparation when inputs match
* Stop validation work once preparation cannot be reused
* Combine file observation imports for code quality checks
* Use annotations for preparation test controls
* Read preparation snapshots asynchronously and refresh context after waits
* Reuse Codex configuration preparation at the actual launch boundary
* perf(cli): use packaged Node for ordinary Mac commands
* test(cli): couple Mac shim fixture to packaged helper layout
* fix(cli): narrow optional native process replacement API
* test(cli): isolate native replacement API across test platforms
* test(cli): allow spying on the isolated replacement API
* test(cli): use NodeNext-compatible dynamic import specifiers
* Keep desktop lifecycle commands on the direct Electron route
* fix(macos): FDA toast copy claims only observed file prompts, not agent interruptions
* fix(macos): keep FDA toast actions inline on roomy cards and ring the dismiss link on focus
* fix(native-chat): offer Submit on the last question step after skipping
The last step's footer label came from the current question only, so a
skipped last question showed an outlined Skip even when earlier answers
would be submitted. Label the last step Submit whenever any question is
answered; an all-skipped prompt keeps Skip, which still dismisses.
* test(native-chat): add skipped-last-step question card fixture
agent-browser's `cookies clear` ignores --name/--domain and wipes the
whole profile cookie jar, so `orca cookie delete` signed users out of
every site. Delete through the guest debugger instead: list cookies with
Network.getAllCookies, match by name and host (host-only, dot-prefixed,
and parent-domain cookies for --url / tab URL), and call
Network.deleteCookies per match, passing CHIPS partition keys back.
Report deleted:false when nothing matched, and reject tab URLs without a
cookie host (about:, file:, data:) asking for --domain or --url. The
external Chromium provider no longer maps cookie delete to the
jar-wiping agent-browser command.
Fixes#42229
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Fix worktree deletion on server hosts
* Finish workspace removal cancellation and final saves
* Wait for WSL workspace deletion to finish
* Capture CLI deletion errors in the server integration test
* test: canonicalize removal fixture and expose serve stop phases
* test: align serve fixture versions and report workspace paths
* test: isolate removal lifetimes and resolve native fixture paths
* Type-check cancellation assertions on the shared removal mock
Pi's launch catalog now opts into launch-time model/effort preferences, so
`orca agent launch --agent pi` and `orca orchestration worker-start --agent pi`
accept `--model <provider/id>` and `--effort <level>` instead of refusing.
Pi models without a provider/id shape are refused up front, before a
worktree or worker is created. A new host capability,
`agent.launch.pi-model.v1`, lets the CLI refuse a Pi model/effort on older
execution hosts that would otherwise start Pi on its default model.
Co-authored-by: Neil <neil@stably.ai>
On macOS and Linux a backslash is an ordinary filename character, but
absolutePathToFileUrl converted every backslash into a path separator, so
typing /tmp/site\draft.html opened /tmp/site/draft.html instead.
Keep backslashes (encoded as %5C) for rooted paths when the local host is not
Windows. On Windows every path still has backslashes normalized to slashes,
so drive, UNC and rooted /foo\bar inputs behave as before.
Fixes#42256
Co-authored-by: Neil <neil@stably.ai>
* chore(test): load Vitest config as native ESM to drop configLoader warning
* chore(test): use rewriteRelativeImportExtensions so type-aware lint accepts .ts imports
macOS Local Network caches one executable UUID per bundle id, and the terminal host shares the app's bundle id. With a UUID of its own the cache flips between the two, which blocked some connections and prompted again after an update. Give the host exactly the main executable's UUID per CPU type, and check that in packaging and after notarization.
The markdown grammar's frontMatter rule embeds source.yaml at the start
of the document, but Orca loaded YAML only when a ```yaml fence label
was seen, so frontmatter stayed plain text. Load the YAML grammar
together with markdown.
Fixes#41658
Co-authored-by: Garrett Peppers <garrett@pepperstech.net>
* fix(native-chat): say /fork is refused while the agent replies, and to run it again
The old notice read as a promise to fork later. Like the common pattern, a running turn is not forkable and nothing is deferred, so the copy now names the agent and tells the user to run /fork again.
* chore(fixtures): show the /fork refusal notice at a narrow composer width
* Revert "chore(fixtures): show the /fork refusal notice at a narrow composer width"
This reverts commit 4ae7756dfd.
* fix(native-chat): /fork during a reply forks through the newest finished response
Like /btw and the common pattern for a fork that names no response, a running reply no longer refuses /fork when a finished response exists; only a first reply or an agent that forks only its latest turn still refuses, with the retry wording.
* perf(dashboard): window agent board lanes instead of mounting every card
Each dashboard lane now renders its cards through a measured per-lane
virtualizer in the existing scroller, so opening a board with thousands of
cards mounts only the viewport, overscan and a constant set of pinned cards
(lane first/last, last focused, dialog opener, pending Tab target).
Search, filters, grouping, counts, dialog routing and acknowledgments still
run on the full card arrays. Subagent disclosure moves to lane-owned state so
it survives scrolling away and resets when the card leaves the lane, as the
old unmount did. Tab and Shift+Tab bridge to the logical neighbour card when
native order would skip to a distant pinned card.
The workspace board's single-index range extractor is generalized into a
shared pinned-index extractor with the same behavior for its caller.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* perf(dashboard): keep the lane card list compiler-compatible
Move the lane virtualizer into its own hook so the card list compiles, and
prune lane disclosure/focus keys during render instead of in an effect.
Give the dashboard destination integration test a lane viewport, since
cards now mount only inside the measured window.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore(dashboard): record isolated virtualizer compiler diagnostic
* perf(dashboard): preserve layout identity and retire departed measurements
* fix(dashboard): reveal the focused edge of tall cards
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The pane-to-subagent map only dropped an entry on going back to the main
agent, so a tab closed with a subagent open kept its entry for the
renderer's life. The removed-chat-tab seam that retires draft positions
now retires those entries too.
Orca terminals run under the signed helper, whose executable was stock Node,
sharing its LC_UUID with every install of that Node build. macOS Local Network
privacy maps a responsible process to an app by that UUID, so terminals could be
attributed to another program and blocked from the LAN while Orca showed allowed.
Rewrite the helper executable's UUID during assembly (same scheme as #41676),
and fail packaging and the release check if it drifts.
Fixes#21988
The per-removal safety re-check listed and projected every workspace on the
host before filtering to one id, so M removals rebuilt the listing M times.
getWorktreePs now takes an optional id and projects only that workspace's
rows; PTY liveness still refreshes against every worktree.
Refs STA-15293
A run whose agent terminal or chat session had opened, but which then failed (prompt delivery failed, or the launcher threw after a partial start), lost its Started time once dead-terminal retention cleared its pane key.
Runs now store an optional launchedAt: the dispatch-start instant, recorded only once the launch produced a terminal or chat session. The persistence layer keeps the first value through later rewrites, and the Started time prefers it, falling back to the existing rules for older records. A launch that fails before opening anything records nothing.
Tests cover delivery failure followed by retention, a partial start, and the no-evidence case; each fails without the change. The recent-runs rail test now asserts hardcoded date patterns instead of calling the production formatter.
* fix(computer): press, set and scroll Windows native controls without activating the window
Measured on Windows CI: UI Automation's Invoke, Toggle, ValuePattern and
ScrollPattern focus a background WinForms control first, which activates its
window. Native controls now go through their own messages instead, each
owner-checked: BM_CLICK for buttons, check boxes and radio buttons, select-all
plus EM_REPLACESEL for set-value on edits, CB_SETCURSEL with CBN_SELCHANGE for
combo boxes (exact-text match, read back after the app's handler runs) and
LB_SETTOPINDEX for list box pages.
* fix(computer): refuse read-only Windows documents before typing keys, document the caret fallback
* fix(computer): keep Windows native routes to measured controls and refuse partial writes
BM_CLICK only for WinForms buttons, multi-column lists stay on ScrollPattern, a
list already at its end reports success, edits refuse text past their limit,
the combo's owner is checked right before the selection, a cleared selection
reads as a mismatch and a disabled control is element_not_enabled.
* fix(computer): hold Windows edit limits without a readback, require a complete combo search
* fix(computer): read a long Windows edit's length from the field for its limit check
* test(computer): share the Windows fixture action helpers
* wip
* wip
* wip
* wip
* Improve experimental worktree cards with compact agent summaries
- Add an optional details row for workspace metadata and indicators
- Roll up descendant activity timestamps to parent cards
* Reduce worktree card metadata indentation
* wip
* Expand sidebar agent lines with subagents and remote host status
- Show subagent activity, send targets, and unread signals in slot cards
- Identify remote hosts and provide SSH reconnection controls
* Show child workspace status counts in sidebar cards
Distinguish monitoring agents and refine lineage alignment so workspace
activity is easier to scan.
* Enable navigation from worktree sidebar activity indicators
- Open the matching agent or workspace directly, or choose among multiple sources
- Use host-qualified identities for nested workspace activity
- Show pull requests in 2.0 card details and refine compact card layout
* Show every status in the sidebar card corner and let cards open and close
The title-row corner now shows one glyph per status the card speaks for
(needs, failed, unread, working), most urgent first, followed by the time,
instead of only the most urgent one. Each glyph opens its own sources.
Collapsed child workspaces report all their statuses through a flat,
shallow-comparable record.
Experimental 2.0 cards can also be opened and closed; a closed card's
title slot speaks for its hidden agents and children.
* final worktree design
* wip
* rm
* wip
* wip
* fix style
* fix commit
* Improve worktree card interactions and spacing
- Let users toggle cards by clicking their headers.
- Delay details until the pointer rests to avoid accidental popups.
- Add breathing room around card content.
* wip
* fix issues
* wip
* Improve workspace sidebar icons and navigation
- Batch AI icon generation and add settings to customize and regenerate icons.
- Fix agent navigation and unread handling across execution hosts.
- Keep linked reviews visible while loading and reduce unnecessary refresh work.
* Fix sidebar agent ownership and targeted title updates
Resolve unstamped agent sessions to their workspace runtime so sidebar
summaries and unread state follow the correct host. Limit title-change
invalidation to affected workspaces.
* fix build
* Simplify worktree sidebar icons and agent rows
- Generate monochrome icons that match the sidebar foreground
- Move agent model names into row tooltips to reduce clutter
* Keep agent timestamps visible alongside dismiss controls
* Fix worktree icon regeneration and bound the cache
Cancel stale requests when clearing icons so older results cannot
restore the previous cache. Evict unused renderer entries to limit
memory growth.
* Unify experimental workspace cards around title status
- Surface hidden agent and child workspace activity in the card title.
- Retire the second experimental layout and AI-generated icons.
- Preserve existing experimental card selections and add a details row option.
* wip
* wip
* wip
* fix bugs
* improve perf
* Fix sidebar lineage rollup: unread dot on children, running status on collapsed parent
- A card's unread dot and bold title now include its own visible agent rows'
unread results, so expanding a parent reveals which child the unread came from.
- A collapsed parent's status lane shows the most urgent descendant state
(needs you > failed > working) when it beats the parent's own status.
- Host-scoped workspace unread markers and related attention fixes.
* Improve sidebar activity rollup performance
- Refresh only stale descendant slots and scope unread updates to their workspace.
- Share agent row projections and index unread markers to reduce repeated work.
- Preserve host-specific agent activation and show the latest activity time.
* wip
* Fix sidebar attention state and optimize slot invalidation
- Preserve unread flags raised by users or earlier sessions when markers clear.
- Scope agent list expansion to each host.
- Track displayed panes and freshness deadlines to limit slot updates.
* wip
* Fix sidebar workspace status and host ownership
- Keep agent navigation and unread state scoped to the correct host,
including paired SSH workspaces and runtimes.
- Correct collapsed lineage status, stale review display, and settings
synchronization while reducing unnecessary sidebar refreshes.
* Show unread agent completions and scope workspace reads by host
* fix compile issue
* wip fix static checks
* wip
* Fix workspace sidebar agent ownership and PR display
- Update agent rows when terminal host bindings arrive and preserve
local ownership for unstamped retained completions.
- Keep newer hosted PRs visible when an older merged PR is cached.