Each project can pick a browser profile (Default or a named one) for its new tabs, overriding the Browser setting's scope. Also settable with `orca repo set --browser-profile`.
Co-authored-by: Dunz Koi <258568505+dunzkoi@users.noreply.github.com>
* feat(computer): list macOS apps without a Dock icon while they show a window
list-apps on macOS only listed regular apps, so menu bar and other accessory apps could not be found or targeted by name. Accessory apps are now listed while they own an on-screen window between the normal and status bar levels, which keeps the dozens of windowless system agents out, and each listed app says whether it is an accessory app.
* fix(computer): list only normal-level accessory windows and resolve hidden accessory apps
* fix(computer): prefer a same-named regular app when resolving, not in the list order
The appearance methods first shipped in v1.4.251 and editor.readTypography
in v1.4.257, so a plugin requiring the old floors installed on Orca that
lacks them.
* feat(plugins): check host API calls against released Orca at validation
orca plugin validate and pnpm validate:plugin now read the literal host
method names a plugin's worker, panel and editor code calls, and report
methods this Orca does not have, proposed methods, wrong-surface calls,
and methods newer than engines.orca (unless guarded by orca.host.has).
Deprecated methods warn with a stable 'deprecated host method <name>'
prefix. Adds the Host API minor to first-Orca-release table.
* fix(plugins): avoid host-call validation false positives
Guarded proposed and wrong-surface calls warn instead of failing, and an orca.api.minor comparison counts as a floor guard. Only host.call/has receivers count, comments are stripped, files are stat-capped and read through their real path inside the plugin, root-relative script sources resolve, and each file is read once.
* fix(plugins): keep has-guarded host calls as errors before orca.host.has existed
* fix(plugins): fail only certain host-call problems and bound the scan
Floor and wrong-surface findings are warnings, since guards and per-surface reachability are not visible to a text scan; unknown and proposed methods without a literal has guard stay errors. Read attempts, references per file and queued files are capped, has?.() counts as a guard, and HTML script and comment scanning is linear.
* fix(plugins): treat any orca.host.has call as feature detection for unknown methods
* build(cli): ship the host-call validation module with the packaged CLI
Add `orca fill --value-stdin` so passwords and other secrets stay out of
argv and shell history. Fill drops one trailing newline from stdin (secret
readers end with one) and rejects empty stdin so a failed reader cannot
silently clear the field; `--value ''` still clears.
Share one flag-or-stdin reader with `orca computer type`/`set-value` and
the `--body-file -` reader. Computer commands keep their existing
semantics: no newline trimming, set-value still accepts empty stdin, and
`--text-stdin=true` style keeps reading stdin.
Fixes#29849
Co-authored-by: Neil <neil@stably.ai>
* feat(plugins): list any workspace's agents with agents.list({ workspaceId })
* feat(plugins): version agents.list workspaceId as host API 1.10
An added param is new surface a plugin must be able to require, so specs
gain paramsSince and api.minor counts it. Regenerates the declarative
typings the first commit missed.
* fix(plugins): report runtime failure codes for agents.list workspaceId
The param is 1.10, after the runtime failure codes, so a call naming a
workspace reports unavailable for a disconnected host; calls without it
keep action_failed.
---------
Co-authored-by: Neil <neil@stably.ai>
* fix(computer-use): put the macOS pointer back after --restore-window clicks, scrolls and drags
HID gestures left the user's pointer at the target. Save it first, and after a
200 ms settle warp it back unless it no longer sits where the last posted event
left it, which means the user moved it. The HID idle counter can't tell: posting
to the HID tap resets it too.
* fix(computer-use): keep the pointer where the user moved it during a macOS HID gesture
* fix(computer-use): save the macOS pointer right before the first HID event
* 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
* feat(computer): add launch-app to start apps without taking the foreground
* fix(computer): leave running apps alone and keep the launch env clean
* fix(computer): recheck aliases, keep focused launches visible, refine launch hints
* chore(computer): document the launch result casts
* test(computer): check launch-app reports where the app ended up on Windows CI
* docs(computer): point app_not_found at launch-app
* test(computer): format the merged Windows background e2e
* fix(computer): match launch paths to the running file and word background launch as an attempt
* fix(computer): type into Windows text fields without raising the window
Adds type-text --element-index on Windows through EM_REPLACESEL for user32,
RichEdit and WinForms edits (owner-checked, read back with WM_GETTEXT), or a
ValuePattern write merged with the TextPattern selection. Keyboard verbs on a
background window now refuse with background_unavailable before sending
anything and point at type-text --element-index. Foreground typing uses one
SendInput KEYEVENTF_UNICODE batch instead of WM_CHAR posted to the top-level
window, and get-app-state shows multiline edit text again.
* fix(computer): refuse typing into Windows list items, keep the caret on a confirmed prefix
window.print() opens the system print dialog and, on Windows, a passkey request
opens Windows Hello. Both are modal windows that take the foreground, and an
agent click could set either off over the user's work. Main now tells each
page's frames whether an agent drove it since the desktop user last gave it
input, and the session frame preload refuses print() on such a page (reported
to the agent as effects.blockedPrint, pointing at pdf) and, on Windows, holds
its passkey requests unsent, still reported as passkeyRequest. The user's own
input releases the page and cancels held requests so the page can retry.
* fix(computer-use): observe minimized, hidden and other-Space windows on macOS without a restore
macOS only enumerated on-screen windows and dropped minimized AX windows,
so an app whose window was minimized, hidden or on another Space (any
full-screen app) could not be observed, and agents fell back to
--restore-window, which activated every window of the app and switched
the user's Space.
Enumerate every window, keep the off-screen ones accessibility can still
read (other-Space windows by remote token, closed windows dropped by
their Space membership), and report their real minimized and off-screen
state. Background pointer and key input on such a window refuses before
sending anything. --restore-window now raises only the target window,
once per request, and delivery is measured from whether the target
became frontmost.
* fix(computer-use): pin the restored window and prefer visible windows by default
* fix(computer-use): keep window-index cache keys and stale-element checks across a restore
* fix(computer-use): restore only this app's resolved window and refuse a closed target before raising
* docs(computer-use): say macOS reads minimized, hidden and other-Space windows
* feat(computer-use): scroll and select rows in background Linux windows, and aim GTK 4 pointer input
Scroll moves the nearest scroll bar's value by whole viewports, handing
the scroll outward at a view's end, or pages a web document with
Component.ScrollTo, before falling back to the pointer. A click on a
list row with no action selects it through the list's Selection when
the index is proven to name that row. GTK 4 rectangles come from
WINDOW extents plus the X11 client origin, since its SCREEN extents
read [0, 0, w, h].
* fix(computer-use): keep Linux background scroll and select honest under partial and slow results
Wait for ScrollTo to move before planning the next page, report pages
already scrolled instead of falling back, let refused scroll bars fall
through, clear a multi-select list before a plain-click select, accept
only an InvalidArgs refusal as index proof, bound child walks, and
read GTK 4 origins lazily.
* fix(computer-use): scroll the innermost Linux view that overflows, and bound the scroll search
ScrollTo now runs on each ancestor that overflows its own box, innermost
first, so an inner scroller moves before the page; coordinate scrolls
start from the deepest node under the point. Cap the scroll bar search
to small containers, report a document at its end as a no-op, keep
looking past an unusable spanning child, and look up GTK 4 window names
once per operation.
* fix(computer): ask the user before retrying with --restore-window
Error hints, recovery steps, the skill guide and the CLI follow-up command
told agents to retry with --restore-window on their own, which raises the
target app and takes the user's screen. They now say to ask the user first,
stop recommending a restore just to observe, and fix stale per-platform
background claims.
* docs(computer): name the Linux keyboard gate and the Wayland limit
* fix(computer-use): keep Linux actions in the background unless the user allows a raise
Pointer fallbacks, scroll and drag no longer raise the target window and
warp the pointer without --restore-window; they refuse with
background_unavailable before sending anything. With the flag, the
target point is validated before the raise, a window on another
workspace is refused instead of switching the user's view, and the
pointer is put back afterwards.
Also refuse background clicks that open menus, keyboard-grabbing popups
or dialogs (ported from agent-desktop's foreground classifier), refuse
synthetic input on Wayland where it never lands, press Return through a
GTK entry's activate action, re-check focus and held modifiers while
typing, and report action.delivery from what happened.
* fix(computer-use): check the workspace before every Linux raise and release held buttons
Run the other-workspace check before GrabFocus too, raise before a
user-approved menu or dialog opener, validate a pointer target before
asking for the flag, count newly active app windows as foreground,
release synthetic buttons when input is interrupted, and stop typing
without an exception exit so the user's held keys are left alone.
* fix(computer-use): keep held modifiers from other X screens and settle remapped keys on early stop
* fix(computer-use): refuse a Linux raise when the target's workspace can't be read
* fix(computer-use): refuse macOS background actions that open a menu or dialog over the user
* fix(computer-use): let web comboboxes click in the background and stop probing a hung app
* fix(computer-use): refuse background drags that start on a menu owner
* docs(computer-use): note background drags on menu owners are refused
* fix(computer-use): classify the element under a drag's start and check AXConfirm/AXOpen secondary actions
* fix(computer-use): refuse Windows pointer fallbacks without --restore-window
Coordinate, right, middle, modified and multi-clicks, clicks on elements with no
primary pattern, unscrollable wheel scrolls and drags raised the window and
moved the real cursor without being asked. They now refuse with
background_unavailable before anything is sent. Coordinate and fractional
scrolls go through ScrollPattern first. With --restore-window only the target
window is raised (never a cloaked one), with a UIA activation fallback and a
second attempt, points are validated first and the cursor is put back behind
the input. A failing UIA pattern or disabled element is an error instead of a
real click, and delivery is measured after the post-action observe.
* fix(computer-use): settle Windows clicks before restoring the cursor and keep scroll writes honest
* fix(computer-use): restore the Windows cursor with SetCursorPos and hit-test coordinate scrolls
* test(computer-use): assert Windows raise outcomes, not attempt order
* fix(computer-use): report a Windows page scroll the provider ignores as not scrolled
* feat(computer): type into a named text field without focusing it on Linux and macOS
* fix(computer): report unconfirmed Linux text readback and cover element typing
* test(computer): send an element with benchmark text only for targeted typing
* test(computer): pick gedit's editor, not its hidden overview map, in Linux e2e
* docs(computer): confirm focus before the unindexed type-text fallback
* fix(computer): report unreadable element text readback as readback_unsupported
* fix(computer-use): scroll, expand and set sliders on Windows without raising the window
* fix(computer-use): wait for Windows scroll and slider writes to settle before verifying
* fix(computer-use): reject NaN for Windows slider set-value
* test(computer-use): wait on the Windows fixture state with vi.waitFor and a schema
* test(computer-use): check activation per action against a holder window in the Windows fixture
* test(computer-use): log fixture window events for activation failures
* fix(computer-use): report foreground when a Windows app raises itself for a UI Automation action
WPF activates its own window for Scroll, ExpandCollapse, RangeValue and SelectionItem calls from any UIA client, so the runtime now compares the foreground window before and after a pattern action instead of assuming background. The Windows e2e checks delivery against the fixture's own activation count, with a holder window in another process.
* test(computer-use): skip an unstarted holder when closing the Windows fixture
* fix(computer-use): bound Windows pattern scroll paging and tighten slider read-back
* fix(computer-use): cap Windows slider read-back tolerance and run the UIA e2e only on CI
* fix(computer): verify macOS type-text and paste-text sent as keys by reading the field back
When text can't be written through AXValue (Chromium/Electron fields, fields
that refuse the write), type-text and paste-text send keys or Cmd+V and always
reported unverified. Now the helper predicts the focused field's value from its
value and selection before sending, then reads it back until it settles:
verified on a match, unverified (text_mismatch) otherwise. Secure fields and
fields without a readable value and selection stay unverified as before, and
the result never repeats the typed text or the field's contents.
* fix(computer): keep the caret put after a refused macOS text write and skip unprovable readbacks
A refused AX value write no longer moves the caret before falling back to keys, so the keys and the prediction both use the user's selection. Text with control characters, a value the field already holds, and a field still filling at the 2 s cap stay unverified rather than verified or text_mismatch. The quiet window starts after the first read answers.
* fix(computer): bound macOS text readback size and say a mismatch can mean keys went elsewhere
* fix(computer): treat a macOS text readback past the cap as undetermined
* fix(computer): refuse quit, log out and lock shortcuts unless explicitly allowed
* fix(computer): close Linux keysym bypasses in the system shortcut guard
* fix(computer): refuse Linux keycode, L1/L2 and VT-switch keysym bypasses
* fix(computer): catch control characters in Linux keycode tokens and point recovery at the stated effect
* feat(computer): show where snapshots were cut and read one element's subtree
* fix(computer): validate subtree targets against the fresh walk and keep scope for focused input
* fix(computer): return whole-window action results after subtree reads
* fix(computer): count kept records toward the macOS limit and index subtree reads against the latest cache
* fix(computer): index Linux/Windows subtree reads against the snapshot they started from
* feat(computer-use): add orca computer wait for text, enabled controls and window titles
* fix(computer-use): bound wait reads by the deadline and wait out missing Linux/Windows windows
* fix(computer-use): read disabled traits after role text with parentheses and bound the capability probe
* fix(computer): check macOS click targets are enabled, in view and on top before sending input
* fix(computer): scope macOS click cover checks to the target window
* fix(computer): scroll before trusting container hits and stop hit walks at the shared container
* refactor(computer): share the one-scroll rule in the macOS click check
* Hide projects from the sidebar without stopping them
Adds Repo.hiddenAt, a per-host, view-only flag. A hidden project's
workspaces, empty header and project-filter entry leave the desktop
sidebar, the paired phone's list, the new-workspace picker and the
palette's project filter. Its terminals, agents and automations keep
running, and repo.list / worktree.ps still list it.
- Hide Project in the project menu (toast with Undo), a collapsed
"N hidden projects" list under the workspace list with Show buttons,
and a Hide from sidebar switch in Settings > Projects.
- repo.setHidden RPC ({repo, repoHostId?, hidden} -> {repo, hostId,
hidden}) and `orca repo hide|unhide [--host]`; `repo list` marks
hidden rows.
- Opening a hidden project's workspace (notification, CLI, deep link)
focuses it without unhiding it, lifting sidebar filters, or leaving a
pending sidebar reveal.
- repo.update echo check: a host that drops hiddenAt reports failure
instead of reading as hidden.
* Address hidden-projects review: multi-host hide, picker scope, mute, stale filters, shared predicate, restore list styling
* Hidden projects re-review: keep saved filters, Agents filter, composer seed, per-toast Undo, shared default host
* Hidden projects: address review threads (pre-feed, palette search, notices, dialog default, seed remap, row keys, a11y)
* Hidden projects: no implicit dialog default when every project is hidden
* feat(plugins): host API 1.9: agent.listSessions (history:read), resume past sessions, agents.list task titles
* test(plugins): host API 1.9 coverage; regenerate plugin artifacts
* fix(plugins): host API 1.9 review fixes
- agent.resume starts a past session where the desktop's Resume would: the same preparation
request and answer handling (now shared), and the transcript in the home it pins
- AI Vault list cache keeps four scopes, so plugin reads never evict the panel's list
- list fingerprints ignore scan id and issues, so an unchanged history revalidates
- workspace mapping resolves symlinked workspaces and prefers a project over a folder
- agents.list task only for open dispatches, read at most once a second
- running matches OpenCode versions; history reads reset when the vault list is invalidated
- truncation, consent copy and limitations documented
* fix(plugins): forced history refresh supersedes cached scopes; resume legacy Codex sessions from their recorded transcript
* fix(plugins): address host API 1.9 review threads
- a scan that began before a forced history refresh no longer caches its older result
- list fingerprints count which issues there are (not their wording)
- symlinked workspaces rank by their resolved folder
- a session-only status row is no live pane: resume prepares the session
- a failed task-title read leaves agents.list rows intact
- the symlink test uses a junction on Windows
* test(ai-vault): order the stale scan before the forced refresh
* feat(plugins): proposed tier for host methods not yet frozen
A method spec may set `proposed: true`: only development plugins may call
it (others get not_permitted), and the minor, frame/worker method lists,
typings and skill references omit it. The reference renders such methods
in a separate Proposed section; dev workers' host.has answers for them.
* feat(plugins): param capabilities, host confirmations, automation provenance
- PluginHostMethodSpec.paramCapabilities: a passed param can require a
second capability; executePluginHostCall and the mock host refuse it with
capability_denied before any audit or effect, and the reference shows it.
- confirmPluginAction generalizes the external-open prompt into a host-owned
confirmation, exposed as PluginHostServices.confirm (desktop only; absent
in headless serve). It re-checks the plugin's approval after the answer.
- Automations record createdBy { kind: 'plugin', pluginKey, name }, set by
the runtime from the in-process plugin caller (RpcContext.plugin), never
from params; the Automations detail shows "Created by <plugin>".
* feat(plugins): workspace:create, automation:read and automation:write capabilities
New closed-enum kinds (older Orcas refuse such manifests with the engines
message), worker-only, with consent copy that names what each allows;
release floor PLUGIN_HOST_API_1_8_SINCE (1.4.259) joins
PLUGIN_SCHEMA_NEWEST_SINCE. agent:launch copy now says it also resumes.
* feat(plugins): host API 1.8 methods: workspace.create, agent.resume, automation.*
- workspace.create (workspace:create; agent also needs agent:launch): a
worktree in a Git project, as orca worktree create does; with agent, the
CLI's agent.launch create-worktree path, so the agent starts after setup.
Setup is skipped or follows the project's settings; never run, an
approved hash or a startup command.
- agent.resume (agent:launch): terminal.ensureAgentSession by session id;
started is false when a terminal already ran it. agents.list rows and
agent.status.changed carry sessionId.
- automation.list/create/update/delete (automation:read/write): create
forwards only name, prompt, agent, target, schedule, time zone and
enabled; update/delete of automations another plugin or the person made
ask the person first and report cancelled or consent-unavailable.
* test(plugins): host API 1.8 coverage; regenerate plugin artifacts
Covers param capabilities (host and mock host), workspace.create payloads
and refusals, agent.resume keys and adoption, sessionId on agents.list,
automation projection, forbidden create fields, confirmation outcomes,
runtime-set provenance and capability access growth. Regenerates the
reference, schemas, typings, declarative runtime and skill references.
* fix(plugins): host API 1.8 review fixes
- Confirmations spend the frame dialog budget, allow one question per
plugin at a time and pause after a refusal (rate_limited); a worker's
question shows in the focused window, else the first visible one.
- Changing what an automation with extra agent settings runs (prompt,
schedule, resuming) asks first even for its creator, showing the prompt.
- Removing, disabling or blocking a plugin pauses its automations and ends
its ownership (createdBy.endedAt), so neither re-enabling nor a reinstall
under the key resumes or edits them silently; the remove dialog says so.
- workspace.create and agent.resume record the terminals they return as
the plugin's own; a launch that fails after its worktree exists names it.
- Setup-decision refusals are told by the runtime's error class (kept as
the in-process RPC error's cause); generated names see every worktree
of the project; resume transcripts match the workspace and host.
- Created by shows the plugin id beside its name.
* fix(plugins): end automation ownership only on uninstall, disable or block
- The ownership sweep no longer reads discovery: only an uninstall
(announced by the removal path), the disabled list, the kill list or
the plugin-system switch end it, and only once the service is ready,
so an unreadable plugins folder pauses nothing. An install under the
same key after a removal keeps what it creates.
- A new time zone on an automation with extra agent settings asks first.
- A failed launch names only a new worktree made from the requested name.
- Settings' plugins-off state and the CLI's plugin system off and plugin
disable say that plugin-created automations are paused and stay so.
* fix(plugins): address host API 1.8 review threads
- Confirmations close when the worker's call ends, freeing the plugin's
turn; they name the automation (short id, project, workspace,
schedule), show any new prompt, and are refused as `changed` when the
automation was edited while the person was asked.
- automation.create takes hostId (through the project's setup on that
host) and a creationKey scoped to the plugin; the runtime rechecks the
plugin's approval right before the write.
- agent.resume records only terminals it started; both it and
workspace.create recheck approval before recording.
- automation.list never cuts a schedule (null when too long) or fails on
an id too long to show.
- The Automations detail says when a creator plugin was removed or
disabled.
* feat(plugins): host API 1.7 — workspace ids, project/terminal list, terminal focus/close
Plugins that track workspaces and terminals across Orca had to read the private runtime
token or shell out to the CLI. Host API 1.7 gives them the same through consented methods:
- workspace.readContext (and workspace.changed) carry `workspaceId`, the id workspace.list
gives; workspace.created/removed and agent.status.changed add `workspaceListId` beside
their existing raw `workspaceId`, which keeps its meaning.
- workspace.list rows name their `projectId`; opt-in `includeStatus` joins each row's agent
state, unread flag, live terminal count and last activity from worktree.ps.
- project.list (workspace:read), terminal.list (workspace:read), terminal.focus
(workspace:write, desktop only, steers only this window) and terminal.close
(terminal:send, reports whether the process is known to have stopped).
- Runtime refusals (unknown or ambiguous workspace, stale or exited terminal) reach plugins
as fixed messages instead of runtime wording.
- Consent copy for workspace:read, workspace:write and terminal:send says what they now
allow; terminal.focus and terminal.close are audited by terminal id.
- PLUGIN_SCHEMA_NEWEST_SINCE covers the editor typography and theme video floors it missed,
and the provisional 1.7 floor, so development builds accept plugins requiring them.
* fix(plugins): narrow terminal.close and harden host API 1.7 from review
- terminal.close closes only terminals of the focused workspace (the check terminal.sendText
uses) or ones this plugin opened with terminal.create or agent.launch, recorded in a bounded
per-plugin registry that is forgotten when its worker stops or it is disabled or removed.
terminal.focus stays any-terminal; summaries and terminal:send copy say so.
- Lifecycle events carry the worktree's identity and host: from the new row on create, and read
host-qualified before the purge on every removal path. agent.status.changed looks up the
identity on the reporting host only. The global removed-identity map is gone.
- includeStatus shares one in-flight worktree.ps and a 1 s cache across plugins, and reports
statusTruncated when that listing stopped short.
- Stored remote contexts get their identity host-qualified; panels and editors never get a raw,
path-bearing workspace id from readContext or workspace.changed.
- Runtime refusals map to fixed messages (tab pinned, no window, tab gone, host disconnected);
any other runtime error or malformed answer is a fixed action_failed with the detail logged.
- hostId is `local`, `ssh:` or `runtime:` in both workspace.list and project.list; rows past the
id or path bounds are left out with truncated set instead of failing the list.
- The mock testing host answers terminal.focus and terminal.close by default.
* fix(plugins): keep shipped host API values and codes; honest terminal results
- workspace.list hostId keeps the values 1.3 shipped (recorded host, else null; folder rows
their executionHostId); project.list always names the host. Both summaries say which.
- Methods that shipped before 1.7 report every runtime refusal as action_failed, as before;
the unavailable/conflict mapping applies to 1.7 methods only.
- includeStatus shares a pending worktree.ps however long it runs; the 1 s TTL starts when it
lands.
- The opened-terminal registry records only for a call still approved and not cancelled, and is
pruned on disable or removal only, so event-driven workers keep close rights across restarts.
- terminal.close answers { closed: true }: a local close cannot tell whether the process exited.
terminal.focus answers focused: false when the request was superseded or not navigated.
- workspace.list and project.list cut labels to 512 instead of failing the list.
- Removal events from the runtime job and the IPC retry are tested to carry the identity read
before the purge; folder deletion carries its host.
* fix(plugins): name the status-listing owner type
* feat(orchestration): let coordinators approve worker permission prompts when the user turns it on
One setting on the orchestration page, off by default. When on, the
escalation offers orchestration approve and deny, both with a reason, and
the user still sees Allow/Deny: the first answer wins. A prompt the
worker's agent lets only a person approve is refused and stays with the
user. The machine running the worker enforces its own setting.
* fix(native-chat): a worker approval's receipt shows the command, not its tool input as JSON
* test(orchestration): list every worker-side case in the coordinator approval table
* chore(i18n): ship the coordinator approval setting's search keywords in the runtime catalog
* fix(orchestration): a reasonless allow keeps the person's refusal, and terminal checks no hook may approve are person-only
* fix(orchestration): address review on coordinator approvals
- A person-only chat prompt answered meanwhile reports as closed.
- A terminal allow reads Claude's warning from the screen it checks, so a
late-painted person-only warning still refuses before the hook is used.
- The receipt keeps a chat worker's whole command.
- orchestration approve prints the reason it sent.
* fix(native-chat): read the receipt's command through a parsed shape, not Reflect
* docs(skills): drop worker permission prompt guidance; the escalation carries it
The numeric reasoning_effort a model sees is not on a 0-100 scale
(claude-opus-5-5: low=5, medium=10, high=15, xhigh=40). Workers launched
with --effort high read 15 as low and misreported their effort.
* Revert "Make chat wait check the chat's work state instead of reading its history every second (#41129)"
This reverts commit 9a10464bec.
* Revert "feat(cli): read, message and wait on agent chats with orca chat list/read/send/wait (#40836)"
This reverts commit c2e6ebbf0b.
* fix(browser): stop agent clicks from opening the system file picker
An agent click on a file input opened the macOS Open panel, which brought
Orca's window forward over the app the user was in. Page actions now catch
the picker the way upload does and tell the agent to use upload.
* fix(browser): enable Page, cancel the picker, and clean up outside host scope
* fix(browser): limit picker blocking to input delivery and bound its setup
* perf(browser): skip picker blocking for actions without user activation
* fix(browser): bound picker interception cleanup on a hung page
* fix(browser): bound the picker interception setup as a whole
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
* 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
* feat(orchestration): deny a worker's permission prompt, and say when mail is held
A chat worker that stops at a permission prompt now escalates with the prompt
itself as structured data (approval id, state, worker, summary, allowed
actions). The coordinator can deny it with a reason through
`orchestration deny`; the worker carries on and gets the reason as mail.
Only a person can allow a prompt: the answer method decides "person" from
the transport, never from params. The stop report keys one escalation per
approval and sends one closing line when it is answered, and no longer tells
the coordinator that `orchestration send` answers a prompt.
`orchestration send` now warns when the delivery gate would hold the mail
right now (an open prompt, a blocked status, typed text), and says delivery
state is unknown for a worker on a paired server.
* fix(orchestration): only the desktop app counts as a person for worker approvals
A paired device can carry a remote CLI, so it is treated as an agent: it can deny with a reason but never allow.
* fix(orchestration): approval answers and held receipts follow the real gates
- A same-option answer counts as Orca's only when Orca recorded it; a retry returns the same
receipt with no second reason mail.
- An unreadable chat is not a closed prompt: the answer refuses, and the closer keeps watching.
- A closing line tells a reason only when it went to the worker as mail.
- The terminal held receipt prints the delivery gate's own refusal; the chat one prints what the
pointer's admission and queue gate would do (not attached, refused, or behind an open prompt).
* fix(orchestration): read the held receipt only after the mail is stored
An await before the insert let a worker settle meanwhile, so the mail could land on a settled
Dispatch after its follow-up scan. The closer also removes only the approval map it read.
* fix(orchestration): tell the coordinator not to block on a worker's permission prompt
The approval escalation now says the worker is waiting for permission, tells a chat
coordinator the user has been shown Allow/Deny (a terminal coordinator: a person must
answer it in Orca or the worker), asks it to carry on with other work and not message
the worker until the answer arrives, and offers deny only as the way out when the
prompt is holding up the task. The guide and CLI docs say the same.
* fix(orchestration): a paired coordinator without an Orca session id mails a deny reason from its Run
* feat(orchestration): answer a Claude terminal worker's permission prompt
A worker pane's PermissionRequest hook now waits for Orca's answer, so the coordinator can deny
it with a reason (the worker continues) and a person can allow it. One record per waiting hook,
keyed by Claude's prompt and tool call; it is answerable only while the hook waits and the dialog
is on screen, and anything else gets `{}` as before.
* test: annotate the fake screen instead of casting
* test: wait on the hang-up instead of sleeping
* fix(orchestration): claim a terminal answer only from what Claude shows
An answer is written only while a caught-up screen still shows the dialog
in the same process run, and claimed only once Claude shows it applied: the
hook's deny line for a deny, no other close for an allow. A Yes, a No or the
2:00 auto-block that wins the race is refused and named from the screen.
A hook that hangs up is not an answer: with the dialog still up the
coordinator is told a person must answer it in the worker's terminal.
A PermissionRequest entry the user made async stays async: it prints {}
first and never waits for a decision.
* fix(orchestration): name a terminal prompt's close from its tool_result
The screen could show an older prompt's close, so who answered is now read
from the tool_result Claude's transcript records for that tool call: Orca's
own deny text credits Orca, the user's No and the 2:00 safety-check block
name themselves, and a run names an allow. Hook events cannot say this: a
deny or a No fires no PostToolUse. The screen only says whether the dialog
is still up before a write, and a PostToolUse seen mid-answer is kept.
* fix(orchestration): read a prompt's tool_result without blocking main
The transcript read is async, reads only its tail, and skips any path that
is not a regular file, so a FIFO or device cannot stall the main process.
A closed prompt still awaiting its result gets one last read when the lead's
turn stops, so a result written late still names the close.
* feat(native-chat): answer workers' approvals from the coordinator chat
A worker's open permission prompt appears as a card at the end of its coordinator's chat, drawn
from the escalation payload as soon as it is stored. The person allows or denies it there; the
card's buttons answer the worker's own prompt over the desktop app's IPC, so only a person can
allow. Once answered, the card turns into a receipt saying who answered.
* fix(orchestration): read a terminal prompt's countdown when it is first escalated
* refactor(orchestration): give a terminal prompt's hold state its own module
* fix(native-chat): let a worker's approval card answer, and drop it once it lapses
The window's runtime calls never carried the orchestration contract
version, so the card's Allow and Deny were refused as an obsolete client.
The renderer ships with its main process, so desktop IPC states the current
contract. An open card whose agent's own countdown has passed, with no
closing message (the app restarted while it was open), is no longer shown.
* test: say why the desktop fence test's runtime stub is enough
* fix(orchestration): a chat worker's approval asks to allow its tool, like a terminal worker's
* fix(native-chat): pause the approval countdown while the window is hidden
* chore(i18n): protect the approval card's worker line copy
* fix(orchestration): close a terminal prompt's card as soon as Orca can't answer it
When the dialog leaves without an answer through Orca, the closing line is
published at once, so the coordinator's card becomes a receipt; a tool
result that arrives later adds a second line naming how it closed, without
waking the coordinator. On startup, open terminal prompts left by an earlier
run are closed the same way, since their holds lived in memory. A Claude
terminal card's reason is the warning Claude prints under the command.
* fix(orchestration): keep the dialog-warning patterns ASCII for the orcad bundle
* fix(native-chat): say when a deny reason reached the worker as a separate message
A coordinator's deny reason for a chat worker whose prompt takes no message is sent
as mail. The receipt showed it the same as a reason that went with the denial, so it
now labels it 'Reason, sent as a message'.
* chore(native-chat): follow main's moved window-activity and agent-session modules
* fix(orchestration): only Orca's own reports draw approval cards, and every card closes
- Cards and the startup sweep read approvals only from rows whose id is the stop report's own
stable id for that approval, through one shared parser; a worker's send cannot draw a card.
- A chat worker's card closes as soon as its own chat answers it (a journal watch replaces the
turn-edge check), when an answer finds the prompt already gone, and after a restart.
- A terminal coordinator is told nobody was shown the prompt and to tell the user; the
terminal-only close asks the same and goes out as an escalation.
- A chat coordinator's line names only the answers its card offers; a prompt with none is a
plain prompt with no card.
- A person's Deny on a chat worker mails no extra reason.
- A record a quit or budget end closed stays closed.
- One closed-approval builder; `waitingOn` becomes `awaitsReply`.
* fix(orchestration): a failing read cannot hang an answer, fail a stored send, or fail the database open
* refactor(orchestration): add the held-send line in send-delivery-hold, keeping send-methods under the size cap
* test(orchestration): read the worker's mail with a schema, not a cast
* fix(orchestration): an open approval stays visible and closes even behind many closed ones or a missing database
* fix(recipes): warn which credentials recipe scripts inherit, add opt-in passEnv allowlist
Recipe lifecycle scripts run locally with Orca's full environment. Keep that
default (provider CLIs and git auth rely on it) but name the credential-like
variables in Settings and let a recipe opt into passEnv so only listed
variables plus a safe base set reach its scripts.
* fix(recipes): fail closed on a malformed passEnv instead of dropping the recipe
Dropping the recipe let a stored snapshot or a same-id plugin recipe run with
the full environment. Keep it with an empty allowlist and a diagnostic, match
base variable names exactly outside Windows, and make the warning copy apply
to any recipe source.
* fix(recipes): import user-agent-platform from its new app-shell location
New profiles default artifactSharingEnabled to true. Existing users are
unchanged: stored settings win over defaults on load, and a stored profile
saved before the key existed loads as false instead of inheriting the new
default. The gate still requires an explicit true, sign-in is still
required, and web clients still mirror the host (their local fallback no
longer inherits the desktop default). showArtifactsButton is unchanged.
* feat(cli): read and message agent chats with orca chat list/read/send
Carried over from #33115 (squashed onto current main; conflicts with chat rename resolved).
* feat(cli): wait for a chat to settle, read text from stdin, refuse paired hosts
chat send --wait and chat wait poll the history's working flag until the
chat is idle or waits on a question, keeping a queued message pending until
it runs. chat read is one history fetch and returns a cursor. Unattached
and paired-host refusals say what they mean.
* fix(cli): settle a held queued message, skip subagent rows, resolve chat list --worktree
* fix(cli): show chat errors, keep waiting through queue handoffs, return the whole settled turn
* fix(cli): settle any queued message the host will not send, show grouped questions and only answerable prompts
* test(cli): drive chat tests through a real client and fake timers
* fix(cli): describe chat --session, --wait and --state in command help
* fix(cli): report chat wait in chat list's states, reject unknown --state, flag truncated reads
* docs(cli): say chat wait settles on the chat's own turn; no truncated flag on held
* fix(cli): chat rename help asks for a session id, not a handle
---------
Co-authored-by: YangNuoCheng <42427439+YangNuoCheng@users.noreply.github.com>
Prints the notes the user left on a browser tab and has not sent, the same text the tab's Copy button gives (--json returns them). Reads the saved session, so it works for closed-and-restored tabs; SSH hosts reach it through the desktop browser bridge. A page drawn by a paired desktop reports that its notes stay there.
* feat(plugins): add editor.readTypography and editor.typography.changed
Plugins that render code next to Orca's editor had to open the private
profile database and reimplement Orca's font rules to match it. Host API 1.6
adds editor.readTypography (no capability; worker, panel and editor frames):
the font family and CSS stack, size with editor zoom, weight, line height in
px, ligatures, word wrap and minimap that the file editor resolves now.
editor.typography.changed carries the same values after a settings write or
an editor zoom changes them; it is the first event that needs no capability.
The pure resolution (editor typography, editor zoom, monospace font stack)
moves to src/shared so main and the renderer use one implementation.
* fix(plugins): keep editor typography resolution from throwing in main
A hand-edited profile with a non-string font family made the shared resolver
throw, which the main-process watcher would surface at startup and on every
settings write. Font values are now type-checked, the watcher catches and
logs, and it reads live settings so a write without listeners cannot leave the
event payload stale. Over-long font families are cut to the schema bound
instead of failing every read. Regenerates the testing runtime.
* docs: load Orca browser guidance for ordinary web tasks
Extend the Orca CLI skill description to include local and external page opening, reading, testing, and captures while working in Orca. Preserve explicitly chosen browser tools and the Computer Use boundary. Regenerate discovery/install bytes and append the new digest.
Fixes#39880
* fix: route existing Edge windows to desktop browser control
* docs: preserve configured desktop-control tool precedence
* docs(skills): scope orca-cli browser trigger to page work and keep the Computer Use rule
Drop "reading" so doc lookups stay on fetch tools, say plain HTTP/API calls don't need the skill, and restore the original Computer Use sentence and guide body, which the issue didn't ask to change.
* docs(skills): list Edge among existing browser windows in the orca-cli guide
The wider page trigger makes "screenshot my open Edge tab" load orca-cli, and the guide named only Chrome and Safari as windows that go to desktop control.
---------
Co-authored-by: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com>
* feat(orchestration): mail agents on a paired server with send --on, replies come back by themselves
* fix(orchestration): gate --on on the carrying Orca, ack retried deliveries, refuse payloads and federated workers
* chore(orchestration): annotate peer-mail row casts and drop a wall-clock wait in its test
* chore(orchestration): name mail-between-servers codes server_mail_unsupported and mail_requires_server_pairing
* docs(cli): list --on in the orchestration reply usage
* fix(orchestration): name --environment, the flag on main, in agent mail refusals and the guide
* chore(orchestration): mark the two deliberate agent-kind branches mail between servers needs
* docs(cli): leave --on out of the reply usage; it is accepted only to refuse it
* fix(orchestration): keep peer conversations to peer ids, survive resets and re-pairs, keep pulling on a registry error
- A server takes only peerconv_ conversation ids, so a paired computer cannot name another Orca's dispatch in the relay.
- A replayed reply after a mail reset is acknowledged, not retried forever; reset tasks keeps replies waiting to be pulled.
- A re-paired server starts a new conversation instead of refusing every later send and reply.
- A failed registry read keeps the pull loop retrying.
- Drop the two ast-grep-ignore lines that only #40193's rule needed.
* fix(orchestration): worker-start --on runs on the caller's own Orca, like send --on
An inherited ORCA_ENVIRONMENT sent worker-start --on to that server, so the
--on command the agent refusal suggests did not work. One rule now keeps
send, reply and worker-start with --on on the caller's own Orca.
* fix(orchestration): refuse reply --on on both sides of a conversation between servers
The side that opened the conversation still took --on when it named the right
server; the reply usage says --on is accepted only to refuse it.
* feat(plugins): add workspace.rename and project names in workspace.list
Plugins that name workspaces had to read Orca's private runtime token and call
worktree.set over the runtime socket, bypassing capability consent. workspace.rename
(workspace:write, worker only, host API 1.5) does the same update as a sidebar
rename. workspace.list rows gain the project name so plugins can prefix names
without reading the runtime's repo list.
* fix(plugins): harden workspace.rename against twin folders and hidden characters
Folder renames require a unique folder id, so twins on two hosts are not both renamed,
and fail when no folder has the id. Names refuse C1, bidi and zero-width characters,
which could make a label read as another workspace's name.
* fix(plugins): workspace.rename matches a sidebar rename for folders and reports kept names
A folder rename also cancels a pending first-agent-message rename, so the automatic
title cannot replace the plugin's name. A rename Orca does not apply (during a WSL
move) is a conflict instead of a silent success. workspace.list's project is optional,
so mocks written for host API 1.4 still validate. More invisible characters are refused.