Distributed builders
For God Mode holders: the Admin Console section that shows the machines that build STRIDE, the controls for working with them, and how to add another machine. Back to Administering the system.
Distributed builders: the machines that build STRIDE
STRIDE distributed builders shows the build fleet at a glance: the machines that build STRIDE, one orchestrator and any number of builders (the one you may know as the buddy is a builder), what each is working on, and any question one of them is waiting on you to answer. A switch at the top chooses the view: Work (everything on this page up to Adding a machine), Machines, Fleet settings and Setup guide, described in the sections below. Only God Mode holders can see it in full. A System Admin that a God Mode holder gave Builders access sees it as a requester (see Builders access); any other System Admin, who administers one organization, does not see this section. It reads again every 10 seconds while the section is open and stops when you leave it or switch to another browser tab; the line beside Refresh says when it last read, and turns to Stale if a read failed or the reading is old.
- All work, and one entry for each machine in the fleet (the orchestrator, then each builder by name), choose whose machines, work, questions and activity are shown.
- Machines: one card each, the orchestrator first. Each says whether it is Up, Suspect (nothing heard for over 5 minutes), Hung (the machine is alive but its work has not reported for 30 minutes) or Down, how long ago its heartbeat was, whether it is Running or Draining, its lanes, its Agents (the sessions and subagents it is running), how many tasks it holds, its Browsers (headless browsers running now, against their budget, such as 3 / 14), its load and database connections, and why it is below its ceiling when it is. A figure the machine’s last heartbeat did not include reads not reported; hover it for why. It is not a fault in this page. Drain asks first, then has that machine finish the lanes it is running, release the tasks it has not started and stop claiming new ones; Resume lets it claim again. A machine paused on purpose reads Paused, never Suspect, with Paused on purpose, how long ago and the reason it was given. Pause sits beside Drain: the machine finishes the lanes it is running and claims nothing new until you press Resume; it asks first. A machine that is up but reports a setup problem it cannot fix itself reads Needs attention, with the problem and its fix.
- Work: every task by state — Building, Blocked or failed, Queued, Delivered, waiting to integrate and, folded at the end, Shipped (delivered work already merged, with the release or commit it went out in) — with the machine, release, model, its age and its latest note. Delivered and shipped work is gray: it is done, not a problem. A building task shows an estimated progress bar: it never passes 95% before the task is delivered, turns amber past its estimate, and says No estimate yet rather than inventing one.
- Recent activity: what the machines did, newest first.
Answering a question. Waiting on you shows only the questions still open; the ones already answered fold into Answered underneath, with what was chosen. Each open question lists the options the machine offered, with its Recommended one marked. Choose one, add a note if you like, and press Send answer. You can only choose one of the offered options; the note goes with it and is never read as an instruction. A question that offered no options cannot be answered here. The machine acts on your answer at its next check-in.
Decisions that are not about a task. A builder can ask you something that belongs to no task, often in answer to a prompt you sent from Talk to the builders. It waits in Waiting on you like any other, with a line saying which of your prompts it answers (quoted), or Not about a task. It is also drawn under that conversation: beneath the builder’s reply, or beneath your prompt while nothing has replied yet. Answer it in either place; it is one decision.
Releasing a claim. On a task a machine is building, Release this claim puts it back on the queue for any builder to take, for when that machine is stuck or gone. Add a reason if you like, then confirm; the machine stops working on it and keeps nothing it had not saved. Only a God Mode holder sees the button; the database refuses everyone else.
Distributed builders: blocked or failed work
When a builder reports work blocked (it needs something it cannot get) or failed (it tried and the work does not hold), the task waits under Blocked or failed. The Blocked or failed figure at the top is a button: press it to open that group and jump to it. Each task shows the start of its report, and Show log opens the whole report and the task’s brief as plain text, with Copy log. Every God Mode holder can read the log.
- Retry with a note puts the task back on the queue with your one-line note added to its brief. It keeps its branch, so the next builder carries on from where it stopped.
- Take it local moves the task to the orchestrator; the buddy can no longer claim it.
- Dismiss marks it handled, with a one-line reason. It leaves the group and the count, and stays in Recent activity; nothing is deleted.
- Ask the machine why opens Talk to the builders with the machine that held the task chosen and a question about what blocked it written in. Nothing is sent until you press Send.
Only a God Mode holder can use them; the database refuses everyone else. Each one asks you to confirm first, and a note or reason is one line of at most 500 characters that must not look like a password or key. If the task has moved on in the meantime, the reason is shown under it and nothing changes.
Distributed builders: the five views and Waiting on you
The section has a control for choosing what it shows: Waiting on you, Board, Talk, Requests and Specs. Board is the machines, work and activity described above, and it is what opens first. The figures and the All work, Buddy, Orchestrator choice at the top show on Board only. The choice is for the visit only; it is not remembered. Request a spec sits beside the control and appears only for a God Mode holder who can send.
Waiting on you is everything that needs you, in one list, with the oldest and the blocking ones first. Its button carries the count, for example Waiting on you · 3, wherever you are in the section. Each item says which machine it is from and what it is about, and how long it has waited. It holds three kinds of item: Decisions (the questions a builder asked, answered with option cards as before), Specs to review (a spec is ready for you; Open spec opens it), and Paths to choose (a request you sent with Let the machine recommend; it replied with two options and its recommendation, and the choice is yours). Answering an item removes it from the list at once. When nothing needs you it reads Nothing is waiting on you.
Distributed builders: asking for a spec
Press Request a spec to describe a feature or a change and have a build machine write it up for you, instead of typing it into a command line. The machine reads your request and writes a spec, a wireframe and a build and test plan for you to review. Nothing is built until you approve it. Only a God Mode holder can send a request, and it acts with your authority, inside the hard stops, like a prompt from Talk to the builders.
- Title (3 to 120 characters, one line) and Describe it (20 to 8,000 characters; a counter shows how many you have used).
- Areas (optional): chips for the part of STRIDE it touches, such as Tasks, Notes, Calendar, Team boards, OKRs, Admin, Phone, Website, Email, AI, Database, Builders or Other. Pick up to eight.
- Screenshots (optional, up to 4): Paste or choose an image (PNG, JPEG or WebP, resized to about 1 MB). Nothing in a screenshot is hidden for you, so check it before you send.
- How to proceed. Write a spec first is the default: the machine writes the spec, wireframe and plans for you to review. Build it directly is for a small, obvious change; the machine still asks you if it finds an effect on data, authorization or AI. Let the machine recommend makes it reply with both options and its recommendation, and you choose (it appears in Waiting on you).
- Send to. Let STRIDE pick is the default. It chooses the less busy of the two machines among those heard from in the last 5 minutes: the one with fewer requests in progress, then fewer tasks, and the Orchestrator on a tie. It says which one and why before you send. Orchestrator or Buddy sends it to that machine, each with a dot showing whether it is up. Either machine can write a spec.
Press Send request (or Ctrl+Enter, ⌘+Enter on a Mac). The text must not look like a password or key; if it does, the kind is named and nothing is sent. If a screenshot fails to attach, the request is already created and the button reads Try again, with Withdraw it beside it.
What you see after you send. Each request in Requests (Open or All) has a card that follows it:
- Queued for Buddy (or the Orchestrator), with how long it has waited. With Let STRIDE pick it also says Picked by STRIDE and why. A request waits safely for a machine that is offline: the card says the machine hasn’t been heard from and for how long.
- Moved to the other machine: a request sent with Let STRIDE pick that nobody started in 10 minutes moves once, if the other machine is up. A request you sent to a named machine never moves by itself.
- Waiting for the machine’s working session, then is reading your request with three bouncing dots, once a working session has claimed it. Being picked up is not the same as being started.
- Waiting for your choice on how to proceed, for a request you sent with Let the machine recommend.
- Writing the spec, with when it started; then Spec ready for review with Open spec. After that the card follows the spec: Revision requested, Approved, Building and Shipped.
- Declined (the machine says why) or Withdrawn end a request. Couldn’t be delivered means the machine could not take it; it stays queued for when it is back.
Withdraw. Withdraw on a card asks Withdraw this request? first. Before a machine has taken it, it is simply taken off the queue; after, the machine that has it is told to stop and not to publish a spec for it. Until a machine has claimed a request, Send to the Orchestrator (or Send to Buddy) on its card sends it to the other machine instead.
Distributed builders: reading and reviewing a spec
A finished spec appears under Specs with its status, its version and how many decisions and comments are open. The filter chooses Needs review, Approved, Building, Shipped, Withdrawn or All. The email STRIDE sends you when a spec is ready names the spec, and the spec is waiting in Distributed builders → Specs. Opening it shows who wrote it (a build machine, never a person), and tabs you can move between with the left and right arrow keys:
- Overview: what it is, what changed since the previous version, the decisions still open, and your request.
- Spec, Build plan and Test plan: the documents, with a table of Contents for the spec. Raw HTML in a document is shown as plain text, and only secure web links are links.
- Wireframe: the machine’s interactive mock-up, with the Static views (screenshots) beneath it. The wireframe is a simulation. It was written by a build machine, runs in a locked frame, and cannot reach your account, your data or the network; its caption says so. Stop removes it, and Open interactive starts it again. A wireframe that cannot load leaves the static views.
- Decisions: the questions the spec asks, with the machine’s recommended option marked. A blocking decision must be answered before you can approve; one that is not blocking can be left, and Your call (hover for Go with the machine’s recommendation) takes the recommendation in one press. A spec that asks nothing says No decisions needed.
- Feedback: your comments and what each version did about them. History: every version, with when it was published and what happened to it.
Commenting. Press Comment on any heading in the spec, or add one on the whole spec, and press Add comment. A comment is a draft until you request changes or approve; the draft count shows beside Request changes, and a draft can be deleted. Nothing reaches the machine before then.
Requesting changes. Request changes sends your drafts and an optional note as one message to the machine that wrote the spec (to the other machine, and the dialog says so, if that one has been silent for 30 minutes). It asks you to confirm it is you first: Email me a code, enter the 6-digit code, and press Confirm and send. The spec reads Revision requested, with dots while the machine revises, then the next version arrives with What changed since the previous version. The old version stays readable, with its comments.
Distributed builders: approving, building and shipped
When you are happy, press Approve…. The dialog is Approve and build:
- Phases to build: the phases the build plan declares, all ticked; untick any you do not want yet. A plan with no phases approves the whole spec. You must choose at least one.
- Blocking decisions must all be answered; if any is open, the dialog lists it and offers Go to decisions. A decision that is not blocking and still open is built to the machine’s own recommendation.
- Note to the orchestrator (optional) goes with the approval.
- Confirm it’s you. Approving needs a fresh second-factor check: Email me a code, enter the 6-digit code (Send a new code if it did not arrive), then press Confirm and approve. If you confirmed recently, it simply asks to Approve and build.
An approval covers exactly that version, and exactly the phases you chose. Approving version 3 never approves version 4: if the machine publishes a newer version, the approval is cleared, the spec is Ready for review again, and you approve the new version separately. An older version is shown read only.
Building and shipped. Approving tells the Orchestrator, which builds it using either machine and alone publishes to the live site. The spec reads Approved, then Building (with the release it is in), then Shipped once the last approved phase is live; the request card follows it. Phases you did not approve stay listed as not approved. The Orchestrator may still ask you questions on the way; they appear in Waiting on you. A spec you no longer want is dropped by withdrawing its request.
Distributed builders: talking to the builders
Under the Machines cards, Talk to the builders lets you prompt the build machines from the page instead of their command line. Choose the Orchestrator, a builder by name (for example the Buddy) or All, write what you would have typed in the CLI, and press Send (or Ctrl+Enter, ⌘+Enter on a Mac). Only a God Mode holder can send an instruction, and every God Mode holder can. A prompt acts with your authority, inside the hard stops, and is logged and kept 30 days. A System Admin with Builders access can send too, but theirs is a request, never an instruction (below).
Builders access: a System Admin's view
A God Mode holder can give one System Admin Builders access from their Profile in People. That person sees this section read-only: the machines and what they are working on, with no Drain, Pause, Resume, recover, release, log or answer controls, and only the prompts they sent themselves. Their Talk box is labeled as a request, up to 20 an hour. The builders may answer a request, investigate and prepare a branch, spec or report, but nothing is built to ship, applied, deployed or pushed on a request: anything that would change STRIDE goes to a God Mode holder for approval. Which kind a prompt is comes from the database, from the sender's role, never from what the prompt says. Builders access ends when the person stops being a System Admin or the account is deactivated.
Canned prompts sit under the box, the same kind of chips as the canned openers under a comment: Need anything from me?, Run autonomously, Stuck?, Running max?, Distribute evenly and Release status. Pressing one puts its whole prompt in the box, replacing what is there; nothing is sent until you press Send, so you can add to it first. Hover a chip to read its prompt.
- Your prompt shows on the right with its receipt. ✓ Sent comes first, with how long it has waited; then ✓✓ Picked up, with the time a session or a responder took it; then ✓✓ Answered. The words carry the state and the ticks repeat it. All is one prompt with a receipt for each machine.
- Three bouncing dots appear on the left, where the reply will land, with the role of the machine that has your prompt and a running working time. They mean one thing: a session or a responder has taken your prompt and is working on it. The reply replaces the dots in place, and the thread keeps your place.
- The status line says what the machine is doing now, with its step out of the total when it reports one, and how long ago it wrote that line, for example running the fleet selftest · step 2 of 4 · 28s ago. The words are written by the machine itself, not by STRIDE, and the time counts from its last line. It sits in the dots’ bubble when there are dots, and under the prompt when there are none. A machine that has not written a line yet shows the dots and their working time alone.
- Waking a fresh session. When the buddy has to start a fresh session for your prompt, the status line reads Waking a fresh session until a session picks the prompt up; then the machine’s own line replaces it.
- Withdraw. While your prompt is still waiting to be picked up (Sent or Not delivered), Withdraw takes it back before a machine picks it up. No machine can pick it up afterwards, and the receipt reads Withdrawn. Once a session or a responder has picked it up, Withdraw is no longer offered: STRIDE cannot stop work that is already running. (Drain on the machine card stops it claiming new work.) If the database refuses, the reason is shown in the notice above the message box, in the database’s own words, and clears at the next good reading or the next send.
- Ask for an update. When a picked-up prompt has gone quiet for a minute, the status line says so and Ask for an update appears under it. It asks the machine to post a progress reply; it does not stop the work. You can ask at most once a minute. After you ask, the line reads Asked for an update · 40s ago until the machine’s next word, and the button returns if it is still quiet a minute later.
- No word for a minute. Also, after a minute with no word from the machine, while the dots run, the faint line beside them says no word for 1m, and the figure grows as the silence does. That is elapsed time only: the machine may still be working, and the line does not say whether it is alive. The dots keep going until the reply arrives.
- Replies show on the left with the machine that wrote them and who answered: the working session, or a responder.
- Orchestrator → Buddy marks a message the orchestrator sent the buddy itself, on the left with its state. It carries the orchestrator’s authority, never yours: it can redirect or pause the buddy’s work, but never answer one of your decisions.
- Decisions. When a reply needs your decisions, each one appears under it as a question card with its Recommended option marked. Answer it there exactly as under Waiting on you.
- Not delivered means the machine could not take the prompt: the reason is shown, and the prompt stays queued where the machine can still pick it up when it is back.
- A prompt that is refused is not sent, and the reason is shown under the box: too many prompts in an hour, something that looks like a password or key (the kind is named, the text is never repeated back), or an account that does not hold God Mode. Your text stays in the box.
When there are no dots. The dots never show while nothing is working on your prompt, so they never claim work that is not happening. The receipt says what is true instead:
- Sent, and nothing has taken it yet. On the Buddy the dots usually arrive 5 to 15 seconds after Send, and later when a responder has to start. Until then there are no dots.
- The machine has not been heard from. When a machine is Suspect or Down, the receipt says so with how long, for example Buddy hasn’t been heard from for 31m. It picks this up when it’s back. There are no dots.
- The orchestrator’s lead session is offline. Your prompt waits for it, and a status-only reply says The orchestrator’s lead session is offline. Your message waits for it. There are no dots, because no session is working on it.
- Not delivered. The receipt turns red and gives the reason. There are no dots.
- This page cannot check. When its last good reading is more than 45 seconds old, or the device is offline, the dots stop and the line says can’t check right now with when it last checked, beside Try again. Stopped dots mean this page cannot see the machine. They do not mean the machine has stopped.
How soon a responder answers, per machine. The Buddy has a backup responder. When its working session has not taken a prompt up, a responder on the Buddy’s Mac starts about 2 seconds after the prompt has waited, and its reply says it came from a responder. If the responder cannot start, or runs and does not take the prompt up, the prompt is marked Not delivered with the reason; a model usage limit is named as one. The Orchestrator works differently. When its lead session has been away for 2 minutes and your prompt has waited 2 minutes, a status-only reply says the lead session is offline. Your prompt then waits for the lead session to answer it; nothing answers it in its place.
While a prompt is waiting for a machine, the conversation reads again every 3 seconds; otherwise it follows the section’s 10 seconds. A prompt waiting for an offline lead session does not hold the 3-second read, and the 10-second read shows the reply when the lead session returns. Both stop when you leave the section or switch to another browser tab.
Distributed builders: the Machines view and Fleet settings
The Machines view lists every machine in the fleet, whatever stage of its life it is in, and is where you add, set up, change and retire them. Only God Mode holders can see it and use it; the database refuses everyone else, including a System Admin. A fleet with no machine yet shows No machines in this fleet yet, with Add the orchestrator and Read the setup guide.
- One card per machine, with its name, its role (Orchestrator or Builder) and a chip for where it is in its life: Pending (a code is waiting for the machine to use it), Setting up (Setting up: 3 of 7 steps done, and the step it is on), Active (with the same health chip as the Work view: Up, Suspect, Hung or Down), Paused, Setup failed (Setup stopped at “a step”, with the reason), a code that expired unused, a credential that was revoked, or Retired. A card also shows its settings in one line, and Settings v4 not yet applied while the machine has not picked up your latest save. Show retired brings back the machines you retired; they are hidden by default.
- Chromes across the fleet sums the headless browsers every machine may run. A line such as Chromes across the fleet: 16 of 16 in business hours, 24 of 24 off-hours sits above the cards and warns when the machines together would go over the fleet’s ceiling (see Fleet settings).
- A machine with no code waiting reads No code is waiting. Issue one to set this machine up. with Issue a code. A machine with a waiting code offers Show code (only in the browser that issued it: STRIDE never keeps the code where it could show it again), Issue a new code and Cancel the code. Canceling asks first (Cancel the code for the machine): the code stops working at once and nothing joins the fleet.
- Actions on each card opens a menu: Edit settings; Pause (the machine finishes its lanes and claims nothing new, and the card shows Paused; no machine can lift a pause you set) or Resume; Rotate credential; Make this the orchestrator (on a builder); Download setup file; Re-enroll; and Retire this machine, in red. Each asks first, and the ones that change who may act as a machine ask for your second factor too.
Edit settings opens that machine’s settings: how many lanes and headless browsers it may run in business hours and off-hours, the port block its lanes use on that machine only (blocks on different machines may overlap), the load above which it stops claiming, which model tiers its lanes may use (a restriction only: the orchestrator still picks each task’s tier), when it may claim work (Always, Off-hours only, Business hours only or Never (rests)), how quickly it answers a prompt, and the labels for what it may take (the tags journeys and db-admin mean browser and database work; other labels are matched against what a task says it needs). More settings holds the responder, heartbeat and watch timings, Keep the Mac awake while it works and the local status page. A save is a new version; the machine applies it within a minute and the card shows Settings vN not yet applied until it has. If someone else saved first you are told, and Load the latest shows theirs. Saving is refused when the machines together would exceed the fleet’s ceiling: Machines’ business-hours Chromes add up to the figure, Saving is refused until they fit.
Rotate credential offers Rotate (recommended), which swaps the machine to a new credential on
its next call and keeps the old one working for 10 minutes, so nothing is needed on the machine. A machine also renews its own credential
by itself 60 days after it was issued; every renewal lasts 90 days and there is no overall end, and STRIDE emails the platform notification
address at each one. A machine whose credential is not due cannot renew it early. It offers
Revoke now, for a lost or compromised machine: the credential stops at once, the tasks the machine had claimed go
back to the queue, and the machine can only return by Re-enroll with a new code. Make this the
orchestrator starts a hand-over: the current orchestrator is asked to finish its integration and drain, and when it reports nothing running
the roles swap in one step (a fleet has exactly one orchestrator, because there is a single pipeline to main). If it has been silent for 30 minutes you
may tick Force the move now. The new orchestrator also needs its GitHub account allowed to push main; until then
it can integrate but not ship. Cancel the hand-over on the notice undoes it before it completes.
Retire this machine asks you to type its name: its credential is revoked at once, its claimed tasks go back to the
queue, and it leaves the list (its history stays for 30 days). Retiring the orchestrator warns that nothing can ship until another machine is made
the orchestrator.
Fleet settings holds what every machine shares: the fleet’s name, its time zone (for example
America/New_York), Business days and Business hours start and end, the two ceilings (Headless Chromes,
all machines, business hours and off-hours, the most browsers the whole fleet may run at once, because every machine tests
against one database), DDL proofs at once, whole fleet (database migration proofs run one at a time by default), the database
connection ceiling, how long before a machine reads Suspect, Down or Hung, when a task counts as stale, and the Branch prefix builders’ branches
start with (it ends with a slash, for example partner/). Decision emails go to the platform notification address, which must
belong to a God Mode holder; it is set outside this page. Save fleet settings is refused with the reason when a value is out of range or the machines
no longer fit.
Distributed builders: adding a machine
You can add a Mac to the fleet from STRIDE, as the orchestrator or as a builder. The machine’s role, name and settings are kept in STRIDE, not on the machine, and the machine reads them when it sets itself up. You prepare one thing here (a one-time code and a setup file) and then run the file on the new Mac. Nothing secret is in the file.
- Add a machine. On Machines, press Add a machine. Step 1, Role and name: choose
Orchestrator (it integrates every finished branch, runs the release sweep and is the only machine allowed to push
main; the card is disabled with its reason when the fleet already has one) or Builder (it claims work the orchestrator posts, builds each task on its own branch and reports back; it can never pushmain), and give it a Name you will call it. Its machine id is made from the name. Press Next: settings. - Settings. Step 2 starts from Fleet defaults for the role, or you can use Copy from a machine you already have. Set what you need now; every setting can be changed later. The fleet’s headless-browser ceilings are checked before you save. Press Save and issue a code.
- Confirm it’s you. A code lets a computer join your build fleet, so STRIDE asks for your second factor first, every time: it emails a 6-digit code, which you enter here. (You can issue at most 10 codes an hour.)
- The code and the setup file. Step 3 shows the enrollment code once, as
XXXX-XXXX-XXXX, with a countdown: Works once · expires in 15 minutes. Copy code copies it. Only you see it, only its first four characters are kept readable afterwards, and a new code for the same machine cancels the previous one. Then choose how to set up the Mac: Download setup prompt (Claude Code) (a prompt you paste into Claude Code on the new Mac), Download setup script (the same steps as a script for a person), or Show the command (the one line that runs the script, with a reminder that it will ask for the code). Both files are made in your browser from the machine’s own settings and the fleet’s, so they are already configured for that machine; the file name and its first lines say which version of the settings they carry (a file made before you edited the settings still applies the current ones when it runs, and says so). Neither file contains the code or any secret, so sending one by email or AirDrop is fine. - On the new Mac. Move the file over. In Claude Code, open an empty folder and paste the prompt; or run the script in Terminal. It checks what the Mac needs, offers to install what is missing, asks for the folder (it proposes one), clones the repository if needed, and then asks for the code. Type the code into the terminal yourself: it is not echoed, never passed on a command line and never enters Claude’s conversation.
- Watch it come up. The dialog and the machine’s card tick through the seven steps as the Mac reports them, under The machine’s progress, and the machine turns Active when its first heartbeat arrives. If a step fails, the card reads Setup failed and says which step and why; fix it on the Mac and run the file again (it resumes from the first step not done), or press Issue a new code.
If the code is not used. It expires after 15 minutes, or when you Cancel the code; the card then reads The code expired unused and nothing has joined the fleet. Press Issue a new code when the Mac is ready. A code is refused after 5 wrong tries in 10 minutes from one place, and it only works while you still hold God Mode. STRIDE emails the platform notification address when a machine enrolls and when a machine’s credential is revoked or renewed.
Running the file again on a Mac that is already set up only re-applies its current settings: it asks for no code and changes nothing on the checklist. On a Mac set up as a different machine it stops and says so, Retire it first, or use that machine’s setup file. To bring a machine back after a revoke, choose Re-enroll from its Actions and use the new code.
What STRIDE does not hold. A new machine needs only the fleet’s own channel, which is a per-machine credential kept in the Mac’s Keychain, shown here by its first characters only, and revocable at once. A machine tagged journeys or db-admin, or the orchestrator, also needs the project’s workload secrets file for the work it does. STRIDE never stores or sends those: the orchestrator sends them to the Mac the way it does today, and the setup stops at its self-test with that instruction until they are there. A builder with neither tag needs no such file.
Distributed builders: the setup guide
The Setup guide view is the short version of the section above, Add a machine in six lines, for when you are standing at the Mac. It lists what a machine needs installed and the seven steps the page ticks off. The same list is under Add a machine in the dialog.
What the Mac needs, by role. Every machine needs:
- macOS 14 or later, with a logged-in desktop session: the supervisor runs as launchd agents in your login session.
- Xcode Command Line Tools, which bring Git and Python 3.9 or later; the setup offers
xcode-select --installand waits. - Homebrew, to install the rest. The setup prints Homebrew’s official install command and stops; it never pipes a remote script itself.
- GitHub CLI, signed in to github.com with access to the repository, and Claude Code, signed in. The sign-ins are yours to do.
The orchestrator, and any machine tagged journeys, also needs Google Chrome for headless browser journeys; the orchestrator, and any machine tagged db-admin, also needs Deno. Those two tags and the orchestrator also need the checkout’s workload secrets file, sent by the orchestrator as described under Add a machine. The orchestrator also reports whether the owner tools file is present: it is reported, never required, and the setup never installs it.
The seven steps, in order. A step cannot be reported until the one before it is:
- Code accepted: STRIDE took the code.
- Credential stored: the Mac kept its credential in the Keychain and read it back.
- Repository ready: The checkout is a clone of the fleet’s repository, up to date, with the push guard in place and its role recorded.
- Claude Code settings: the standard settings and the deny list are merged in.
- Supervisor installed: the background agents for its role are generated from its settings and loaded.
- Self-test passed: it can reach the fleet with its credential, its settings are readable, a push to
mainis refused for a builder, and no owner password is anywhere on it. - First heartbeat: the machine reports in and turns Active.
The setup never puts a credential in the repository or in the workload secrets file, never signs in as you, and never changes Claude Code permissions beyond the fleet’s standard deny list.
On the Mac, afterwards. The setup leaves a small command line you can use: python3 _docs/tools/dispatch.py whoami says what STRIDE
calls this Mac, its role, and its settings and credential (only the credential’s first characters), dispatch.py config get <setting> reads one of its
settings, dispatch.py budget says how many headless browsers it may run now, and dispatch.py rotate swaps its credential by hand. A credential lasts 90 days, and whoami says when a rotation is due
(within 30 days of expiry); Rotate credential on the Machines view asks the machine to rotate on its next call.