Repository navigation
Releases
0.7.0-rc.24
This pre-release is the open-source build of Conductor, made from main, and covers everything merged since v0.7.0-rc.23.
Highlights
- Needs you: one list and one number. Everything that waits on a person (a question, a plan, a task out of its attempts, a run's decision) is one list in the sidebar's Needs you, and one amber number on the nav, the rail, the phone's first tab, the palette's row and the tab's title. (#956)
- One status language. A session is needs you, working, finished, running or exited, with one mark and one word wherever it shows. A run and its tasks speak the same status language. (#956)
- Compare approaches. A ticket's approved plan can launch two to four candidates of one base. Conductor measures them side by side, Brakeman can draft a comparison, and you select one candidate as a final decision. Only that candidate is then reviewed and delivered. (#980, #864, #875, #976)
- GitHub linking. Save your own GitHub token, import an issue as a ticket or link one, and see GitHub's state and checks on the ticket. You can publish an outcome back as a comment you have previewed byte for byte, and turn a failed CI job into a repair request. (#982, #848, #884, #909, #979)
- Accounts and capacity on screen. Settings lists your agents' accounts. The launch dialogs choose an account and a model, and a capacity panel shows the usage windows the agent's own tool reports. A launch on an account now runs on that account's login. (#845, #857, #899)
- Browser evidence. Open in Conductor opens an attempt's preview in a pane. A saved capture can create a ticket, be attached to one, be sent to Brakeman, or become a review's finding. (#983)
- Launch with a first message. For Claude Code and Codex, the Launch dialog asks "What should it do?". The message is typed at the agent's own prompt once the session is ready. (#956)
- Codex's tool approval form is a question. On a session of the server, Codex's form before it runs an MCP tool shows in Needs you and the question bar with its own choices. (#958)
Needs you and the app's screens
- Needs you. The page is called "Needs you" everywhere. Its nav row leaves when nothing waits and nothing was decided in the last day. Each row says its kind, and on a wide window the oldest item opens by itself. With nothing waiting the page says "Nothing needs you now." and lists what was decided in the last day. "Open on your phone" is in your menu on every page. (#956)
- The sidebar. Its sections are Needs you, Finished, Working, Shared with you and Exited. Finished holds the agents that ended their turn ("finished 12m ago"), and a run sits once, where its busiest member is. The sidebar opens at 200 px; its header is the mark, "Conductor" and your menu, which gains "Alerts…" and "Collapse to the rail". Its foot is six tiles: Needs you, Yard, Crews, Dispatch, Brakeman and More. (#956)
- A question in a narrow sidebar. Under 280 px a row that asks shows two lines of its question and an Answer button. A panel beside the sidebar holds the whole question and its choices. (#956)
- The Yard. The queue beside the grid is gone, so the grid takes the whole panel at every width. A tile's header is the session's name, chips filter by All, Needs you, Working, Finished and Running, and the Roundhouse is a view of the Yard (a Grid | Roundhouse switch). J, K and Enter are no longer Yard keys, and Alt+J and Alt+K reach the terminal again on every page. (#956)
- The session page. The inspector starts closed, the line under the name is agent, folder and branch, and Share is a quiet button. A link to a session that is gone shows its record's name. On a phone, Files, Share and Fullscreen move into More. (#956)
- A run's page. The header fits in two rows, with a "More for this run" menu. One bar and one sentence say how far the run is. The Tasks view is a table that says each task's state once, and the Graph view draws the run as lines, with a station per task. The Crews table and a live run's card draw the same small line drawing. (#956, #960)
- A run's decision. It shows its question once and one set of buttons. The recommended choice is marked on its button, and the fine print (the id, the type, each step's receipt) folds under Details, open at first only for a plan's approval and a comparison's selection. (#962)
- Dispatch's screens. A project opens with its description as one line and one toolbar: List, Board and Roadmap, search, a Filter popover with a chip for each value in use, the count, saved views and the sort. The ticket list's columns line up, on a phone too. A ticket has a strip under its title that names its sections with their states (GitHub first on a linked ticket), and an approved plan folds to one line. (#956, #983)
- Events is retired. History has a Search and Feed switch, and the Feed is the live feed the Events page held. "Alerts…" in your menu opens the alert switches and "Where each event goes". An agent's hooks and the Conductor skill are set up on the Agents page under "Reporting to Conductor". Old Events links land where their tab lives now. (#956)
- Empty and loading pages. A page with nothing to show says what the page is, why it is empty and the one action that fills it. While it reads it says so, and a failed read shows the server's words and Read again. The welcome is one sentence and three steps. (#956)
- Connect to Conductor. With no workbench token, or one the server refuses, a page to connect replaces the app. While it shows, the app behind it sends nothing. The Workbench token dialog, with Cancel, opens only from the menu, the palette and Set token. (#956)
- Colours and buttons. Every screen uses the brand's own greys. Amber is only where someone must act, and each area has one filled button. Every icon-only button has a tooltip that is also its accessible name; it opens on a resting pointer and on keyboard focus, never on touch. (#956)
- The Agents page. The list is drawn at once, and each agent fills in as its check answers ("Checking…"). Each card says whether the agent reports to Conductor. (#911, #956)
- Paste on a phone. The terminal on a touch device has a paste button. A startup question with no choices, such as Claude Code's sign-in screen, points to it. The paste sends no Enter: you press Enter in the terminal. (#907)
Dispatch and GitHub linking
- Your token. Save a GitHub token with
conductor trackers(read from standard input) or in Settings' GitHub card, which has its own Check and save. The token is checked with GitHub before it is kept, is never served, printed or logged, and is left out of every export. (#982) - Import and link. A project offers Import from GitHub with a preview. A linked ticket shows a GitHub panel: GitHub's state, Refresh, Take over and Give back, the imported images and the steps to reproduce. GitHub owns the title and the description until a person takes them over, and the ticket's status never moves because GitHub's did. A pasted GitHub address in the palette finds its ticket. (#982, #848)
- Publish to GitHub. From the GitHub panel, the ticket's Publications or a link's row. Your own text, the offered parts and "Close the issue as completed" all start off. Preview shows the server's exact bytes with their SHA-256, and Publish posts once and says who published where. Nothing is published to GitHub but what a person publishes. (#982, #876, #884)
- GitHub's text and agents. A plan's proposal can hand over the issue's steps to reproduce and evidence links. Every text a GitHub read wrote reaches an agent inside a quote mark, in the launch package and in every read of a ticket an agent makes. (#909, #979, #982)
- CI jobs as repairs. The Review view lists the GitHub CI jobs of each delivery pull request, failed first, with Read again. A failed job can be prepared as a repair request, and the repair attempt's task file quotes the job's name and the end of its log. (#982)
Yardmaster and crews (comparing approaches among them)
- New comparison. Crews has New comparison: the objective, the criteria, the checks and two to four candidates. A comparison is launched only from a ticket's approved plan revision, and the button reads "Launch N candidates". Each candidate works in a lane of its own and does not reach another lane's work. (#980, #864, #875)
- Measured apart from said. The ticket's comparison has a column per candidate, the checks with the server's verdict (comparable, unknown, not comparable), and what Conductor measured: size, dependency changes, test and build files, what the candidates share. The agents' summaries, their reviews and Brakeman's draft are shown apart, as "Supporting information, not evidence". (#980)
- The selection. When the last lane ends, one decision opens for the operator: the candidates with a result, and None of these. Nothing is preselected, a rationale is required, and the button says the answer is final ("Select approach-b (final)"). Only the operator answers it. After it, only the selected candidate can be repaired, reviewed and delivered, and Review the selected candidate opens the Review view on it. (#980)
- Crew launches. Every launch of a saved crew from the Crews page or the palette opens the crew's launch sheet, and a ticket's plan launches from the ticket's own launch dialog. The launch names the crew's revision the sheet read, and the server starts nothing for another revision. (#899, #963)
- Ledger version 12. The managed runs' ledger moves to version 12 for comparisons; no table changes. A build before this one opens such a ledger read-only. (#976)
Browser evidence: previews and captures
- Preview and capture. Open in Conductor (a task row's menu, the palette, a ticket's candidates) opens an attempt's preview in a pane, with device modes, Refresh, New tab and Capture. Capture opens an editor: a screenshot with numbered boxes and a note...
0.7.0-rc.23
This pre-release is the open-source build of Conductor, made from main, and covers everything merged since v0.7.0-rc.21.
Dispatch
- Review a candidate. A ticket's Review view shows one candidate of its run (a task's result commit) as a diff against a base you name. You can put notes on lines of either side, on whole files, on the candidate as a whole and on failed checks. The notes stay in a draft in your browser until you send them as one review: Request changes, Comment or Approve. On the run page and on a ticket's receipts, Review opens the view. (#665, #711, #722, #734, #748)
- One traceable repair round. Request changes is recorded in Dispatch, and Yardmaster applies it as the task's repair. A run the server no longer runs gets the repair through a continuation, and only when you say so. The repair attempt reads your findings, with their ids and lines, in its task file and says which ones it addressed. Nothing merges, pushes, delivers or closes by itself. (#710, #717, #722)
- Findings across candidates. A finding keeps its identity from one candidate to the next. Against a later candidate the view says where its lines went (same, moved, changed, gone, renamed), and you mark it Resolved, Still open or Withdrawn. The view also shows the candidates' lineage, the criteria as launched and as they are now, the plan revision, the expected checks and the delivery. The palette offers Review the latest candidate, Send review and Open findings, and the view works at phone width. (#722, #748)
- Failed checks. The failed-check panel prepares a repair from the check's receipt. It works out the failure's class (code, infrastructure, flaky, cancelled, unavailable) and, when Yardmaster would take it, offers Check again, which verifies the same commit again as an attempt of its own. (#697, #711, #748)
- Decisions and people on a ticket. The ticket page lists the ticket's decisions and links to their items. A plan revision's decision can name whom it is for and can be an Assignment (work a person does by hand, answered Done or Not done). (#719, #741, #761)
Decisions
- Durable decisions. An attempt's agent can put a question to you as a record in the ledger, with its alternatives, a recommendation, the evidence and what a yes authorizes. The attempt ends its turn and is given back its try: waiting costs no retry. A ticket's approved plan revision can name gates, and the tasks behind a gate wait for your answer while the run's other tasks go on. Records survive a restart, and an answer recorded while a run is stopped is applied once when the run is continued. (#664, #672, #678, #683, #730)
- One card everywhere. You answer in the same card on the run page, in the session's question bar, under Needs you, from the palette and on the ticket page, or with
conductor decision respond. The card tells apart open, recorded (with its receipt), validated and applied, and says when evidence changed after the question was asked. Sending again gets the same receipt, and a second answer is told who answered first. A share-link guest sees the question but cannot answer it. (#688, #689, #719) - Needs your decision.
/decisionslists in one place everything waiting on you: a member's question, a failed or held task, a plan revision, a Brakeman card and durable decisions. Each item shows its brief, the question, the choices, the evidence one tap away and the four steps of an action (available, recorded, validated, resumed). A failed task's item offers Check again beside Retry, Accept anyway and Cancel when the task's failed candidate can still be checked again. (#679, #684, #735, #750, #761) - From a phone. On a phone, the bottom bar's first tab is Decide, with the count of open items. When something is waiting, the app opens on Needs your decision. On the desktop, the sidebar has a Decide row and
G Ngoes there. A decision link answers a member's question and a failed task's Retry, Accept anyway and Cancel from the link's join page, served by this server or through a switchyard; a durable decision is shown there but is answered on the LAN or the desktop. In Share, "Answer decisions from this link" turns that on, and Open on your phone gives the link with a QR code. A run's decision link works through the run's desk even when no member is alive. (#684, #698, #724, #737) - People in a crew. A crew can list people beside its members. People are named labels: nothing is ever started for them. Each person can hold one responsibility (architecture owner, product owner, security reviewer or acceptance owner), and each task names its accountable owner. The run page and the attempt's task file say who answers for the work. A person's page lists their crews, tasks, tickets and the decisions for them, and
@in the palette finds people. (#671, #674, #682, #741) - Answering for someone. A decision can be for
role:…orperson:…. The operator answers it on that person's behalf, and the card and the receipt keep who answered beside whom it was for. Brakeman, when granted, can route an open decision to the operator or to a person of the run's crew. (#741, #744) - Long answers are kept. An answer within its own length bound is always kept, whatever characters it holds. (#753)
Brakeman
- Briefs an agent can write first time.
brief_submit's schema and the guide's newbriefssection describe a brief's shape, and refusals name what is accepted. A brief must hold at least one item, it records the agent's version, and a made draft's Blocks line names the other ticket by its key and title. (#603) - More evidence. Plan revisions, launches and candidates can be cited in a brief, and Brakeman can draft a plan revision for you to propose ("Propose as plan revision"). The guide lists the server's registered providers. (#605, #656)
- Granted actions. In Settings ("What it may run") you can grant Brakeman any of four operations: launch an approved plan, prepare a follow-up revision, request a repair, and route a decision. Each act runs through the operation's own route under the actor
brakeman, can be retried without being repeated, and shows its receipt in the brief, under "What Brakeman did" when no proposal of the brief names it. Theacttool, andconductor brakeman actin the guide, are offered only in a turn that holds a grant. (#612, #656, #691, #721, #744, #762) - In the background. Brakeman can wake on events (a task that failed with proof, a question left open, a run that ended or paused, a ticket made ready, a briefing time). Wakes inside a batch window are gathered into one turn, with a cooldown and a daily budget. To control it, use Settings' "In the background", the Brakeman page's role card (state, Pause, Resume, notes, the log of wakes) or
conductor brakeman role. (#655) - When the agent cannot start. A signed-out agent, Claude Code's first-run screens or an update dialog make the role unavailable, in the screen's own words, and the role tries again later. A turn only types its line at the agent's own prompt, and a screen Conductor does not recognise gets no keys. (#691, #720, #721)
- Reads the history. Brakeman's
observecan search the history, limited to what its grant covers. A session that ended before a restart can be cited from its record. (#764) - Smaller things. Settings has one Save, which also saves Brakeman's section. A Claude Code turn configured to run in the home folder runs in Brakeman's own folder instead, so Claude Code asks its trust question only once. The guide and
conductor guidesay what the caller may call and the limits it must keep, and the skill's limits match the server's. (#691, #718, #721, #727, #743, #747)
History
- Session records. Every session the server runs leaves a record: its agent and version, the agent's own session id, host, folder, repository, branch, commits before and after, and what passed through Conductor. What you type into a terminal or a reply box is never kept. The
history.sessionssetting turns records off, and each record can be forgotten. (#673) - Search. One index holds session records, runs, attempts, check receipts, tickets, comments, plan revisions, decisions, reviews and Brakeman's briefs. Search it with
~in the palette, under the plain search's History group, on the History page (/history,G H) or withconductor history. Each row is Conversation, Decision, Checked ✓/✗ or a record, gives its source, host, project, revision, time and availability, and opens the evidence itself. (#669, #681, #694, #746, #755, #763) - Resume from a record. A session's record page offers Resume, which picks the agent's conversation back up after a restart where the agent supports it, and Forget. The run page can open at a step's log or a command, and a ticket page at a comment. (#723, #731)
- Hosted sessions. A session run by
conductor hostgets a record too, from what its host reports (where it ran, the agent's version, repository and commits). It reads "host away" while the host is gone and is picked up again when the host comes back. (#822)
Yardmaster and crews
- Check again and repairs. A failed check's attempt can be checked again at the same commit, and a person's requested changes become the task's repair attempt, also through a continuation of a stopped run. (#697, #710, #717)
- Decisions in runs. Gates from a plan revision hold only the tasks they name. An attempt waiting on a decision is parked so the member can take another ready task, and its try is given back. (#672, #678)
- Launch checks. A crew launch is refused when a member's agent cannot be shown to be installed, instead of starting the lead and failing that member later. (#826)
Accounts and capacity
conductor accountsandconductor capacity. Accounts are named references to an agent's own login folder (Claude Code or Codex),...
0.7.0-rc.21
Merge pull request #601 from rocksolidlabs/issue-432 web: a button collapses the terminal under an open file, and Preview plan opens the plan rendered (#432)
0.7.0-rc.18
Merge pull request #379 from phenixrizen/issue-375 release: built from a dispatch, the tag made on publish; an example crew run by tasks
0.7.0-rc.16
The agent harness, rows 5–10 (#349, part of #334): a crew with tasks launches a managed run.
- Dispatch and results. Each task runs as an attempt in its own worktree at the commit it starts from, with the task in
.conductor/task.mdandtask.jsonand one typed line saying where it is and how to submit. A result needs a commit; the task's checks run against it one at a time, and a failing check sends the next attempt back with the log tail. Every change goes to the SQLite ledger before its side effect. - The agent's side.
GET /api/sessions/{id}/task,POST …/task/result, the MCP toolstask_contextandtask_result, andconductor task context|result. A review's needs_changes reopens the task it reviewed with the findings; what rested on the result is withdrawn. Pause, Retry, Accept anyway and Cancel carry receipts. - The run page. A Tasks view with each task's state, words, result and actions; the header counts; the graph's Tasks | Members switch; task rows in the palette.
- Continuation. After a restart the server concludes what was left open; Continue keeps the results that stood, verifies a candidate again, and numbers the next attempt after the old one.
- Messages, blockers and checkpoints. Members send durable messages (
message_send, the inbox and its acknowledgement); an attempt reports a blocker (task_blocked) and saves a checkpoint (task_checkpoint) the next attempt starts from.
Also: the hostagent's first-viewer tests give the join's ping a bound their steps cannot exceed (#345).
Known: the harness has not been run by hand with real agents yet; that is #350, in progress on this build.
v0.7.0-rc.15
v0.7.0-rc.15
rc.14 with the third batch of UI fixes from the rc.14 screens (PR #333), five issues as one stream:
- The top bar's Files and Chat buttons are gone at desktop widths, where the inspector rail has both (#329).
- An exited session's buttons line up: the row's
…sits beside its Resume arrow, both centred on the row (#331). - The Keyboard shortcuts modal, redrawn: wider on a desktop, the column names once in a header that stays put as the list scrolls, no tints, no dots for a missing key (the cell is empty and reads "none" to a screen reader), Enter as a word (#328).
- Tabs close as VS Code's do: right-click a tab, or press Shift+F10 or the Menu key on it, for Close, Close others, Close to the right, Close to the left and Close all; an unsaved tab among them asks first. The same five are actions on the tab's row in the command palette (#330).
- Every command palette row does what it says: Open Chat opened the chat and, run again, closed it; a row named for showing something shows it and never hides it. A sweep now drives every row on every page in the tests; it found the browser's Settings page offering a Save row that could never act, which is the desktop app's alone now (#332).
Everything else is rc.14's: the session page without its sideways scrollbar, the command palette and its reach, the editor beside the terminal and its rows in the palette, the inspector rail, Touched across every session, the Roundhouse strip, the shortcuts per system, Markdown previewed, a typed line acknowledged, runs surviving restarts, and the server, host and switchyard fixes since rc.12. See the rc.13 notes.
v0.7.0-rc.14
v0.7.0-rc.14
rc.13 with one fix: the horizontal scrollbar along the bottom of the session page in the desktop app (#325). The inspector rail's count pill stuck out past the rail by two pixels, and the page's body grew a scrollbar for them at every width. The pill sits inside the rail now, and the inspector clips sideways. A phone's quick-reply choices still run past a narrow window, as they did before (#327).
Everything else is rc.13's: the command palette and its reach, the editor beside the terminal and its rows in the palette, the inspector rail, Touched across every session, the Roundhouse strip, the shortcuts per system, Markdown previewed, a typed line acknowledged, runs surviving restarts, and the server, host and switchyard fixes since rc.12. See the rc.13 notes.
v0.7.0-rc.13
v0.7.0-rc.13
The workbench's UI round and the fixes since rc.12.
The workbench
- The command palette (#302, #303): Ctrl+K (⌘K on a Mac), / on any page, Alt+/ while typing into an agent. What needs you first, with its choices answered in the row; the page's own actions; Running, Exited, Runs, Shared with you, Artifacts, Go to and every other action;
>#:@!?for a kind; verbs typed as one row (stop docs,share docs view,answer review 2,say docs looks good…); a pill for the next session that needs you with]and[; Answer all; a line typed into a session says typed once its owner entered it, and a session that asks refuses it (#322). - The editor beside the terminal (#309), and its own rows in the palette: Keys → Neovim, Save, Close file, Ask the agent about the selection (#321).
- One inspector rail instead of two tab rows, Touched an Activity filter, the git polling lifted (#301); the breadcrumb strip scrolls to its end (#315).
- Touched across every session: the palette's
:lists the files every session's agent touched, from the server's index, and opens them in the session's editor (#304). - The Roundhouse strip rearranged by drag or keyboard, shown as thumbnails (#310); the collapsed sidebar marks the open session on its square (#314); the keyboard shortcuts per system, Alt+Q to leave a terminal (#312).
- Markdown previewed in the editor, Mermaid diagrams drawn in a sandboxed frame (#166); a plan Claude Code saves outside the folder opens in the editor (#289).
Server, host and switchyard
- Runs survive restarts: records written and resumed, one continuation per run,
conductor up --resume(#313). - A malformed save part closes the connection; a forgotten run revokes its switchyard links (#311).
- Hooks report over http or https, whatever the server's scheme (#300); the switchyard is this server by the ports it listens on (#305); with a public address the peers gather on its family alone (#306); a viewer moving to the relay keeps the size it held (#307); no SIGNAL frame over its limit leaves the server (#308).
- A host's
fileentries reach the server and the admin feed (they were dropped by an event-type check that predated file events), and the protocol says seven event types.
Protocol
GET /api/touched;submitwithreqIdandunlessWaiting, answered bysubmit_okor an error naming it;welcome.submitAck. Older clients and owners are unchanged. Seedocs/protocol.md.
v0.7.0-rc.12
Release candidate 12 of 0.7.0. The full entries are in docs/features.md (#299).
- The desktop app says an update once, in the workbench: a toast with Restart to update and Later, a button in the sidebar's header and an item in the tray's menu, never a dialog of its own at every check. A dialog first says what the restart ends (#294, issue #290). This is the first release the app can update itself from; the rc.11 app updating to this one still shows rc.11's dialog.
- Claude Code's permission dialogs, plan approval and AskUserQuestion offer their own choices, read off the screen, with submissions held while one shows; a choice that changes the permission mode gets no number key (#292, issue #288).
- A server is named by its address alone:
rendezvous.serverandconductor host --serverrefuse a URL with a user name, password, query or fragment; logs, errors and replies show a server URL as its scheme, host and path; a failed control dial says why in fixed words. A switchyard URL withuser:pass@behind a Basic-auth proxy now gets the refusal at start (#298). - A question answered by a key in its terminal is offered nowhere else at once; the share dialog draws a QR code of the link; the Files breadcrumb, the Explorer's path filter, the Launch dialog's folder field and an exited row's Resume button, among other small things (#284).
- Codex: its update question reads whole, and its program status (OSC 7501) weighs in place of its title from 0.163 (#268, #280).
- A host back from away clears a question answered meanwhile; the Launch dialog's host command carries
--insecurewhen the host needs it (#259, #260). - Copilot, Cursor, Antigravity and Goose report working mid-turn: install their hooks again to get the new events (#262).
- Startup questions hold a submission's Enter and a stranded line's hold; two screen-model fixes for a grown window and a synchronized frame (#265, #266, #278, #279).
- The Settings toast says when the server restarted; the Files pane works on a very short window; the Events page says what reports working (#277, #261, #264).
- The work flow on GitHub's own primitives, a local preflight on Linux, the Mac and Windows, and the weekly runs fixed (#252–#287).
Published as a prerelease so the desktop app's updater can find it (the macOS build is unsigned and cannot update itself yet).
v0.7.0-rc.11
Release candidate 11 of 0.7.0. The full entry is in docs/features.md ("v0.7.0-rc.11").
- An agent's own question stays in Needs you after its turn ends, until someone answers; the first response wins (#207, issue #193).
- A narrowed wall tile draws its rows on their rows: tiles keep a scrollback and fit without the scrollbar's gutter (#223, issue #221).
- The editor's Neovim takes a
:%ssubstitute, and a Neovim lost keeps its tab's text (#188). - A quick reply names the question it was shown for; a question reads whole in the sidebar, the Yard and the reply bar (#192, #191).
- Codex: no bell for a question it asks while it keeps working, and its state read from its window title (#194, #205).
- Hosted sessions: what a link's viewer sends passes the link's check; the host's own messages go ahead of the frames it relays and the host vouches for its server only by its own answered pings; a plain
http://orws://server on another machine needs--insecure(#198, #204, #212). - The terminal fills its pane wherever it is shown (#200, #209).
- The review process in AGENTS.md: plans reviewed before code, bounded review rounds, the backlog on GitHub issues and the project board (#184, #244).
- Test fixes for CI's flakes (#199, #210, #216, #222, #245, #246, #248, #249).
Published as a prerelease so the desktop app's updater can find it (the macOS build is unsigned and cannot update itself yet).