# Swarm UI The swarm's own web surface and the operator's day-to-day view: served by the gateway on the **swarm apex** (`services.hyperhive.swarm.domain`), readable only by operators. The per-hive dashboard, on each hive's own domain, covers host-level detail for one hive. ## What it shows | route | what | | ---------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `/` | the start page: the swarm's name over a digital-rain backdrop, plus a paperclip offering to set up an agent while no `ruth` exists (it opens the create form with `ruth` filled in); the swarm name in the header links here | | `/hives` | the hive directory, each hive with its last reported status | | `/agents` | every agent: status, config PR, wanted state, inference settings, resource caps; create agents (with their permissions, optionally from an [archetype](archetypes.md)), link forge, matrix and GitHub accounts | | `/agents//terminal` | one agent's live terminal | | `/agents//subagents//terminal` | one of that agent's subagents' live terminal | | `/alerts` | the alerts the swarm's Grafana holds now: the firing ones, with a button for the pending, no-data and error ones and a link to Grafana's own alert list; read through the controller with a Viewer service-account token | | `/jobs` | the controller's job graph — where agent creation and credential mints show progress | | `/schedules` | the scheduled prompts: add, edit, pause, resume and remove | | `/broadcast` | send one message to several agents at once, picked from a grid of their icons | | `/messages` | the swarm-wide message feed: every message between agents, and to them from the operator, reminders and compaction, with a history window and a filter on sender or recipient | | `/issues` | a cross-repo issue report | | `/review` | the review classifier: its recent runs, a try-out on a real PR, and its settings | | `/builds` | the swarm builder's recent builds of agent config PRs, with what nix printed when one failed | Everything it shows comes from [`swarm-controller`](../../swarm-controller/README.md). Pages that refresh on a timer share one polling rule: the page sends no new request while the last one is still running, so an older answer never overtakes a slow one, and a failed refresh keeps the rows already on screen under an error. An agent created here or with `swarmctl agent create` starts `paused`; set it `up` from its card. A card shows **login needed** when the agent's own turn-state header says its claude has no session, and a **log in** button that runs the claude login from a dialog: open the URL it shows, then paste the code back. The credentials go to the agent through the secret store ([`claude-login.md`](claude-login.md)). The agent's own web UI login still works too. ### Agent and subagent terminals An agent's detail panel on `/agents` shows a small live preview of its terminal, and below it the agent's **subagents**: every subagent the swarm queue holds terminal rows for from the last 24 hours. Picking one shows its terminal in the same preview, and **expand** opens it in its own tab, as it does for the agent. The list refreshes every 10 seconds, so a subagent the agent spawns shows up once it writes output. A subagent that drops out of the list closes its terminal, and a later subagent with the same name starts unpicked. On a terminal's own page (`/agents//terminal`, or a subagent's) the header has an **N running subagents** badge. It opens a list of the agent's subagents, each a link to its terminal; on a subagent's page the current one gets a marker and the list starts with a link back to the agent's own terminal. A subagent counts as running when it published a row in the last 15 minutes. The swarm doesn't report more than when the newest row came, so a subagent that thinks quietly for longer drops out of the count. The list shows the others below, dimmed, as **not running**. An agent's terminal opens with its last 100 rows from the past 24 hours, then follows new ones; a reconnect picks up after the last row it showed. A subagent terminal does the same with that subagent's rows, and has no turn-state badges. The settings menu in the header has a **hide debug output** toggle. It drops the muted debug rows (thinking, coalesced ticks, harness chatter) from every terminal, the preview and the tabs, and rows already on screen hide or show as you flip it. The browser keeps the choice for the swarm UI, apart from the per-agent pages'. A terminal's own tab takes the agent's icon as its browser-tab icon, so tabs of different agents tell apart. A subagent's tab shows its agent's icon. An agent with no icon keeps the hyperhive mark. The tab also has an agent badge, with the agent's icon and name, ahead of the connection badge. It opens the agent's details in a popover under the badge: the same status, wanted state, inference settings, resource caps and linked accounts as the panel on `/agents`, without the terminal preview. The popover doesn't dim the page. Selecting outside it, or Escape, closes it. From the keyboard, opening a popover moves focus into it, and Escape puts focus back on the badge. The filter popovers on table headers work the same way. You can stop or pause the agent, or link an account, without leaving the terminal. A subagent's tab opens its agent's details. ### Review `/review` is the one page for the [review classifier](review-classifier.md), with three sections in a column. **try a classification** and **classifier settings** come first, each folded to a single bar: the bar of the settings says what they're now, and what you typed in a section stays when you fold it. **recent runs** is under them and always open. The sections share nothing and ask for their own data. #### Recent runs The **recent runs** section lists what the [review classifier](review-classifier.md) has run, newest first, one card per run. A card shows the repo and PR, the head commit, whether the run was a try-out or a requested review, how long ago it started, how long it took, how many classifications ran, and the outcome as a badge. Select a card to open it: it shows the same result table as the [try-out](#review-try-out), the overall verdict, the reason when the outcome has one, and a link to the posted review when there is one. An outcome can be more than approve or request changes: | badge | what happened | | ------------------- | -------------------------------------------------- | | `approved` | the run posted an approval | | `changes requested` | the run posted a request for changes | | `comment only` | there was no clean verdict, so it posted a comment | | `queue full` | the queue was full, so it turned the run away | | `post failed` | the run finished but posting the review failed | | `done` | a try-out finished | | `failed` | the run failed; the card says why | The page refreshes every 30 seconds, and a card you opened stays open when it does. An empty list says there are no runs yet, and a failed load shows the reason the controller gave. The page shows no token counts, cache counts or cost, and stores and posts nothing. #### Try a classification The **try a classification** section shows how the [review classifier](review-classifier.md) answers a classification on a real PR, before you put it in a repo's config. Give it a repo as `owner/name` and a PR number, then write a classification: a name, a description, a polarity (positive or negative) and a threshold from 0 to 100. **add a classification** adds more, and **run** starts the run. The repo field offers the repos that have an open PR, and once you've picked or typed a repo, the PR field offers its open PRs by number and title. Both fields stay free text: type a repo or a PR the lists lack, such as a merged one. If a list can't load, the page shows the reason and the field still takes what you type. **load the repo's rules** fills the classifications from the rules file on the repo's default branch, so you can run the repo's own rules or start from them and edit. It reads the file each time, so a rule merged a minute ago shows up. The page names the file and the commit it read, and asks before it replaces classifications you have written. If the repo has no rules file, the file doesn't parse, or it holds more than ten rules, the page shows the reason and leaves your classifications as they were. A run takes a minute or more, and up to half an hour for ten classifications. The page says it's running and shows how long it has been, then lists each classification with its verdict, the model's certainty, the threshold, whether the verdict counted and the reason; the posted review folds those fields into four columns ([`review-classifier.md`](review-classifier.md)). A classification that couldn't run shows its error where the reason would be. Under the table are the overall verdict, approve or request changes, and the head commit the run read. A failed run shows its error, for example when the swarm has no API key yet. What you typed stays after a run, so you can edit the classification and run it again. The page saves no classification and no result, and a run posts nothing to the forge; like every table in the UI, the result table remembers its column filters in the browser. It shows no token counts, cache counts or cost. When a classification answers the way you want, **open a PR** proposes the classifications on the page to the repo. The classifier's forge account pushes a branch that adds them to the repo's rules file, with you as the commit's author, and the page links to the forge's page for opening the PR from that branch, where you log in as usual. A classification named like one the file already holds replaces it, and the file's other rules stay. It needs no PR number. If the forge has no account of yours yet, log in to the forge once first. If the classifier's account may not push to the repo, the page says so: a repo admin adds it as a collaborator with write access ([the service account](review-classifier.md#the-service-account)). #### Classifier settings The **classifier settings** section shows the runtime, model and effort the [review classifier](review-classifier.md) runs with and how many requested reviews run at once, and lets you change them. The runtime is **API key** (claude on the Anthropic key) or **ACP** (opencode on a provider). On the API key runtime the model is free text, one word as claude's `--model` takes it, such as `haiku` or `sonnet`. Choosing ACP shows the provider's fields in place of the model: its `https://` URL, its model name, and what it charges in USD per million tokens for input and output, with optional cache read and cache write prices. The effort is one of low, medium, high, xhigh and max, or the runtime's own default. Concurrent reviews is a whole number from 1 to 8; try-outs don't count toward it. **save** is on once something changed and the fields are valid. A save changes the runtime, model, effort and limit, and the provider too while the runtime is ACP, so switching back to API key keeps the stored provider for next time. If the controller refuses it, the page shows the reason and leaves the form as you typed it. A change takes effect from the next run, with no restart. The page has no API key field: admins set each runtime's key in the bao UI. ### Scheduled prompts `/schedules` has one card per scheduled prompt. The **name** is the card's title, with the **agents** it fires for, the message, how often it repeats and when it fires next. **edit** turns the card into its own form; **pause** and **resume** stop and restart it; **remove** deletes it after a confirmation. **+ schedule** opens an empty form above the list, and the filters narrow the list by name, agent or state (`active`, `paused`, or `past` for a one-shot whose time has passed). `past` is the controller's clock against the start time, not a record that an agent fired, so you can still pause a `past` schedule. A schedule repeats **every** some number of seconds, minutes, hours or days, or fires **once**. Its start time is the moment a one-shot fires, and the origin of the fire times of a repeating one. The form previews the next three fire times. Changing the start or the repeat moves them; editing only the name, message or agents doesn't, and the form warns when an edit would. A start in the past fires the latest missed time once, right away, in every agent chosen, and the form says so. The form offers the agents of the roster that `/agents` lists. The page reloads that roster with the schedules, so an agent created after you opened the page shows up without a reload. If the roster fails to load, the form says so instead of claiming there are no agents. Only an operator can change a schedule here or with `swarmctl schedule`; an agent has no way to add one. The page needs the controller's `/api/schedules` routes. ### Sending a message to an agent Under an agent's terminal, in the preview and in its own tab, is a **message** box. **send** hands the text to the agent's harness, and the agent takes it as an inbox message from your authelia username. **sent** means the harness was listening and holds the message. If the harness isn't running, or isn't connected to the swarm queue, the box shows an error and the agent gets nothing. Subagents have no message box. The message doesn't go through the agent's hive, so the hive dashboard's message history doesn't list it. The agent's answer shows in its terminal; an agent that calls `send` to your username gets an unknown-recipient error. A harness that restarts before taking its turn loses the message.
How a message reaches the agent The gateway sets `Remote-User` from authelia's answer on every `/api/` request, replacing any the browser sent. `POST /api/agents//input` takes that header as the sender and sends `{from, body}` as a NATS request on `$SWARM.input.`. swarm-controller is the only queue client allowed to publish there. The agent's harness subscribes to its own subject and queues each message for its turn loop with its reminders. Only the agent's own queue credential may subscribe there. The harness doesn't reply, so swarm-controller answers 202 when the request times out after two seconds, and 503 when the queue reports no subscriber. `POST /api/agents/broadcast` takes `{agents, body}` and sends the same message to each agent the same way, up to 16 at once. It answers 200 with `{results: [{agent, sent, reason?}]}`, one entry per agent in request order, and `reason` holds the text the single-agent route would have answered with. One agent not listening doesn't stop the others. swarm-controller refuses the whole request and sends nothing when `body` or `agents` is empty, when `agents` names an agent twice, names more than 64, or names one that `GET /api/agents` doesn't list. `GET /api/messages/stream` is an SSE feed of the messages between agents, and from senders that are no agent (the operator, `reminder`, `compact`) to an agent. It reads the terminal rows on `$SWARM.term.` and has no hive in the path. Each event's `data` is `{ts, from, to, body, body_format?}`. An agent's `send` row gives one event from that agent to the recipient in the row's summary. A turn started by a non-agent sender gives one event to the agent whose terminal holds the row. A turn started by an agent gives none, since that agent's `send` row already did. The query `since_minutes` (default 60, at most 1440, 0 for live only) sends recent history first, oldest first, read from each agent's newest 100 rows. A message an agent drained with an in-turn `recv` starts no turn and so is missing. The **messages** page (`/messages`) reads that feed. A **history** select picks the window, from live only up to 24 hours, and a filter keeps the messages whose sender or recipient contains the text you type. Each message is one row of the same list an agent's terminal uses, a summary line with the time, the two parties and the start of the body, and the whole body under it. The list follows the newest message unless you scroll up, and holds the last 2000. The feed has no event id, so a reconnect sends the history again, and the page empties its list each time the connection opens, then fills it from the replay. Changing the window opens a new stream. A command has its own subject, so the harness never mistakes typed text for one. `POST /api/agents//cancel` sends `"cancel"` as a NATS request on `$SWARM.control.`, and only swarm-controller may publish there. The harness answers `{acted}`, which says whether a turn was running when the cancel stopped it. It runs the same code as the hive dashboard's cancel button. Unlike input, the harness answers a command, so swarm-controller returns that answer as 200, 503 for no subscriber, and 504 when the harness held the command and didn't answer within ten seconds. A 504 means unknown, not cancelled.
Where the subagent list and rows come from The agent's subagent daemon publishes each subagent's terminal rows on `$SWARM.term..sub.` with the agent's own queue credential, into a stream named `term-sub-` that swarm-controller creates for every declared agent. The stream keeps rows for 24 hours. swarm-controller serves the list as `GET /api/agents//subagents`, the subagents named by the subjects in that stream, each with `last_row_unix`, the stream's publish time of its newest row, or `null` when that row isn't read within two seconds. The lookups run concurrently, sixteen at a time, so the list waits at most two seconds for each sixteen subagents. swarm-controller also relays one subagent's rows as SSE on `GET /api/agents//subagents//term/stream`. Like the agent's terminal route below, it creates a short-lived consumer on that stream, here filtered to the one subagent, which sends its newest 100 rows and then each new one, with the same `Last-Event-ID` resume. The stream holds every subagent of the agent, so for a fresh reader the consumer starts at the oldest row and swarm-controller drops all but the newest 100 before sending. An agent with no store identity, or no queue address, publishes no subagent terminals.
Where an agent's terminal history comes from The harness publishes the agent's terminal rows on `$SWARM.term.`, into a stream named `term-main-` that swarm-controller creates for every declared agent, with the same 24 hours or 64 MiB as the subagent stream. That stream holds this one subject, and the subagent stream holds only subjects below `$SWARM.term..sub`, so they never overlap. `GET /api/agents//term/stream` creates a short-lived consumer on it that starts 100 rows back, so the SSE stream sends those rows and then each new one. Each event's `id` is the row's stream sequence, and a request with `Last-Event-ID` resumes after it. An id past the newest row of the stream, as after swarm-controller recreates the stream, starts over as a fresh reader. When the consumer stops sending its heartbeat, as after a queue restart, the route ends the response and the browser reconnects. Until the stream exists, the route sends live rows only.
### Issue report The **issues** page lists the open issues of every repo that has one, from `GET /api/issue-report`, or of one repo from `GET /api/repos/{org}/{repo}/issue-report`. Each row shows whether an open dependency blocks the issue and how many other issues depend on it. The list sorts by that count by default, and **hide blocked** drops the blocked rows. Picking a repo fetches its report again. The page remembers the pick, so a repo that no longer has an open issue stays in the **repo** select with the label **no open issues**, and the table stays filtered to it. The repo list and the report fail on their own: a repo list that fails to load shows its error beside the report, and a report that fails to load clears the table instead of leaving the previous repo's issues under the error. ### Builds The **builds** page lists the agent config PR heads the swarm builder built, newest first, from `GET /api/builder/runs`. Each card shows the repo and PR, the first seven characters of the commit, how the build ended (**built**, **failed** or **error**), when it started and how long it took. Select a card to read it: for a failed build, what `nix` printed; for a built one, the system closure; for an error, why no build came back. The **Details** link on a `swarm-builder` commit status opens this page as `/builds?run=`, with that build's card open and in view. The builder keeps its newest 100 builds, so a link to an older one shows a note instead. ### Broadcasting a message The **broadcast** page sends one message to several agents at once. Pick the agents from a grid of their icons. **running** shows the agents that report themselves running and **all** shows every agent; the browser keeps the choice. **select all** ticks every agent the filter shows, and **clear** unticks them. The page skips an agent you ticked and then hid with the filter. Type the message and press **send to N agents**. The text goes exactly as typed, as an inbox message from your authelia username, the same as the box under an agent's terminal. One broadcast may name at most 64 agents, and the button stays off until you untick the extra ones. A broadcast to many agents can take a few seconds to answer, and the button reads **sending…** meanwhile. The page then lists what happened to each agent. An agent that wasn't listening shows as not sent, with the reason, and the others still got the message. **select the N that failed** ticks just those, switching to **all** if the filter hid any, so you can send the same message again. When every agent got the message the box clears; otherwise the text stays. ### Inference settings The agent's detail panel shows its model and effort under **inference**, each with where it comes from: the swarm, its agent.nix, its own pick, or the default. Below them are the subagent model list the swarm sets, or "its own" when it sets none, and the subagent cap, which reads "no cap" when unset. **edit** opens a dialog with the four fields, filled with what the swarm stores; saving does a `PUT /api/agents/{agent}/inference` of the whole value. A field you leave empty stays out of it, so the agent keeps its own value for that field; an empty max subagents is no cap. **clear the swarm's setting** sends the `DELETE`. The panel reads `GET /api/agents/{agent}/inference`; before the agent's first report reaches the controller, it shows the swarm's stored values only. [inference](inference.md) describes the settings themselves. ### Editing an agent's instructions The **instructions** row of an agent's detail panel has an **edit** button. It opens `AGENTS.md`, the file the agent reads its instructions from, in a plain text box. The text is the copy the agent last published, so the box still opens for a stopped agent, but **save** needs the agent running. Save sends the swarm a diff against the version you opened, and the agent applies it to its own file with git. When the agent changed the file in the meantime, the swarm refuses the diff. The box then keeps your text, shows the agent's version below it, and offers two ways on: **keep my text, replace the agent's changes**, which makes your text the new base for the next save, or **use the agent's version**, which drops yours. A diff that git refuses for any other reason shows git's message and leaves the file as it was. Closing the dialog keeps an unsaved edit for the next time you open it. ### Resource caps The agent's detail panel shows its CPU quota and memory max under **resource caps**. A cap is the swarm's, so it follows the agent to whichever hive runs it, and it replaces what the hive says for that field. A field the swarm leaves out shows as the hive deciding it. **edit** opens a dialog with the two fields, filled with what the swarm stores. A cpu quota is a percentage of one core (`400%` is four cores). A memory max is a size such as `512M` or `8G`, a percentage of the machine's memory, or `infinity`. Saving does a `PUT /api/agents/{agent}/limits` of the filled fields. A `PUT` replaces the stored value whole, so an empty box clears that cap and the hive decides that field again. **clear the swarm's caps** sends the `DELETE`. The controller checks each value against systemd's own syntax, because systemd ignores a value it can't parse and the agent would run uncapped, and it refuses, for example, a percentage with a leading zero. The dialog shows the refusal's reason under the fields and leaves them as you typed them. A saved change isn't instant. The hive reads the caps about every 30 seconds, and a running agent takes a changed cap at its next spawn or rebuild, so the row shows the new value before the agent runs under it. The panel reads `GET /api/agents/{agent}/limits`. [hivectl](../tools/hivectl.md#per-agent-resource-limits) sets the hive's own limits. ### Settings at creation The **+ agent** dialog has two collapsed sections, **resource caps** and **inference**, with the same fields as the detail panel's editors. Leave the fields empty and the request is the plain name and hive; collapsing a section doesn't withdraw what you typed in it. Fill them and `POST /api/agents` carries them as optional `limits` and `inference` objects, the same shapes the two `PUT` routes take. The controller checks them with the `PUT` rules before it queues anything, so a bad value gets a refusal with its reason and creates no agent. It stores them ahead of the job graph, before it declares the agent, so they are in place before the agent first runs. If the creation then fails, the controller puts back what it replaced. Each agent name keys its settings and destroying an agent doesn't clear them, so a name set up before keeps its stored settings when you leave the fields empty. ### Initial AGENTS.md at creation A third collapsed section, **instructions**, holds a plain text box for the text the new agent's `AGENTS.md` starts with. Left empty, or holding only whitespace, it sends nothing and the agent starts without the file. Otherwise the text goes as typed in the optional `agents_md` field of `POST /api/agents`. The agent writes it on its first start, and only when it has no `AGENTS.md` yet, so creating an agent again under a name that has one never replaces it. The controller refuses a text over 64 KiB. Its reason shows under the section, open or not, and the swarm creates no agent. After a successful create the box empties. The [instructions editor](#editing-an-agents-instructions) changes the file later. The [swarm README](README.md#an-agents-agentsmd) has the wire details. ### Linking external accounts Each agent on `/agents` opens three dialogs that write a credential for it into the swarm secret store through swarm-controller. All three are blind set actions: no route hands a token back. swarm-controller refuses to link a name that already holds an account; delete that account first to replace it. The agent's detail panel lists its linked accounts under **accounts**, one row per account: kind, name and host. Opening an agent makes one request, `GET /api/hives/{hive}/agents/{agent}/linked-accounts`, which returns every account of that agent as names and hosts and never a credential. When the panel fails to read the accounts, the inference settings or the resource caps, it shows the reason as text. A list the panel already loaded stays on screen beneath the error. | kind | one row per | name | host | | ------- | ------------------------------------------- | ----------- | ------------------------ | | matrix | `swarm/agents//matrix/` | the account | its `homeserver`, if set | | forgejo | `swarm/agents//forge/