The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.
Harness Pro: cable contract with the daemon
The daemon (cli/src/cable/*) talks to every device over the same framed JSON cable. The Harness Pro (Pod UI)
and the round dial are two clients of that one daemon. They differ only in what their hello says.
The hello capabilities are the contract. A capability key in the hello turns on extra daemon output for that session. A device that names none of them (the round dial, or an older Pro) gets exactly the traffic it gets today: no new field, no new message, byte-identical frames. A daemon that predates a capability simply ignores the key, and the Pro must still work with empty bodies (plan, Review Focus 1).
Sources: branch dev/firmware-pro at 0c73edf72 (read-only), merge base with main 59ce00535; main at
10c9e7420. Firmware side: SRC/main/cable_client.c, ui/habitat/ui_habitat.c,
ui/habitat/pro_player_library.h (SRC = devices/harness-device/firmware in the read-only worktree).
1. Pro hello (owner decision 7)
Full key list, in the order the Pro sends it:
| Key | Value | Meaning |
|---|---|---|
fw |
firmware version string | existing |
product |
CABLE_PRODUCT |
existing; the daemon releases the port if it does not match |
proto |
number | existing |
hw |
board name | existing, informational |
board |
"pro" |
new on main: names the client kind, for logs |
swarms |
"members-v1" |
capability |
recap |
"long-v1" |
capability |
activity |
"steps-v1" |
capability |
stats |
"turn-v1" |
capability |
player |
"library-v1" |
capability |
superTab |
"wait-v1" |
capability: Super Tab waits on question (below) |
swarmCard |
"v1" |
capability: the Swarm cards and the Lead view (frames below) |
swarmChat |
"v1" |
capability: the chat the desktop has open on each Lead's face, and its notices (the round dial's, spec 2026-10-30-harness-pro-swarm-parity) |
swarmBuddy |
"v1" |
capability: the PM and the Buddy in each Swarm project (the round dial's, same spec) |
ota |
"v1" |
capability: the Pro takes firmware offers |
mac |
aa:bb:... |
existing |
settings |
object | existing: brightness, character, face, voiceLang, muted, quiet, straightTitle, focusFace, scrollReversed, round |
Dropped from the branch's hello: speech (pcm16-v1). The Pro does not take speech output.
A capability is a positive string match; any other value (absent, wrong string, wrong type) means "off".
Capabilities are per session: they reset on a new port and on close, like the branch's playerCapable.
2. Capability table
| Capability | Turns on | JSON shape and limits |
|---|---|---|
swarms: "members-v1" |
agentIds and members on each swarms item |
{t:"swarms", selected, items:[{id, name, agents, panes, agentIds?:string[], members?:[{id, name, engine, machine, online:bool}]}], tiles:[{x1,y1,x2,y2,a}]}. One frame, hard cap CABLE_MAX_PAYLOAD = 8192 B; the daemon budgets 8192 - 256 B. Shrink order when too big: ids + members for all tabs, then ids for all tabs, then ids for the selected tab only, then none. Firmware caps: 24 tabs (SWARMS_MAX), 24 ids per tab (SWARM_MEMBERS_MAX), id < 48 B (ID_MAX). members only names ids that are in agentIds; it covers ids the roster and library cannot name (an offline machine's agent). online is false only when the daemon knows that machine is offline; absent or unknown reads as online. |
recap: "long-v1" |
full on summary (live and restore:true) |
{t:"summary", agentId, recap, text, full?:string, restore?:true, ...}. full is the whole answer's prose: markdown stripped, whitespace collapsed, at most 1200 chars, at most 3600 UTF-8 bytes, cut on a word boundary with …. Omitted (not empty) when the answer has no text. Firmware recap buffer is 2048 B. recap and text are unchanged. |
activity: "steps-v1" |
step, action, elapsedSeconds on turn.activity |
{t:"turn.activity", agentId, text, action?:string, elapsedSeconds?:int, step?:string}. step: one tool card as "Tool · arg" (a path argument keeps its last segment), at most 120 chars, sent as {text:"", step} the moment the tool starts; the firmware cuts at 400. action: latest assistant-written progress paragraph, whitespace collapsed, at most 188 bytes (firmware keeps 192 B). elapsedSeconds: integer 0..2592000, parsed from the engine footer (1h 2m 3s); anything else is omitted. action also rides on the first turn.activity right after turn.started. |
player: "library-v1" |
feature player.library in welcome.features; the player.library.get request and the player.library reply |
Request (device to host): {t:"player.library.get", offset:int 0..1000000, request:int 0..4294967295}; anything else is ignored. Reply: {t:"player.library", request, offset, total, machines, rows:[...]}. rows has at most 6 (page size); offset is clamped to max(0, total-6); the page is never re-sorted (desktop order). Row: {id, machineId, name, engine, status, ageSeconds}; id, machineId non-empty and < 48 B; name <= 95 B, engine <= 15 B (control chars to spaces, UTF-8 safe); status one of idle working question finished failed paused offline; ageSeconds is -1 (unknown) or 0..315360000. The firmware rejects the whole page if count != min(6, total - offset). request echoes the latest request; a reply for an older request is dropped. Resent at most every 10 s unless the page changed. |
superTab: "wait-v1" |
question frames for Super Tab waits |
{t:"question", agentId, id:"supertab:…", questions:[{q}], superTab:string ≤48 B, attention:"decision"|"stuck"}; see "Super Tab waits on question". Without it the daemon sends none. |
swarmCard: "v1" |
supertabs, supertab, supertab.recap, supertab.focus to the Pro; supertab.recap.get, supertab.open, supertab.select from it |
See "Swarm cards". No such frame reaches a device that does not send it (the round dial, older Pro firmware). |
Order of arrival the firmware tolerates: any. Every field above is optional on the device side; a missing one leaves the Pod showing an empty body, never an error.
3. Port inventory
Branch commits touching cli/ and desktop/ since the merge base: 92283ad62 (dock companion, unrelated to Pod),
4051f87f4 (Harness Player 1), 07ea2fd55 (every session in the Player library), 0c73edf72 (Pod). Firmware
reads below are SRC/main/....
| Feature (branch files) | (a) Pod reads it | (b) On main | (c) Port |
|---|---|---|---|
Swarm members: agentIds on swarms items; swarmMembers flag; members metadata (memberMeta, SwarmMemberMeta); 8 KB shrink in swarmsMessage; agentIds added to the change key; setSwarms also notes overview agents. From 0c73edf72 (+ metadata in the same branch). |
Yes: cable_client.c:763-790 (agentIds, members), ui_habitat.c:7052 -> pod_ui_swarms. |
No. CableSwarm has no agentIds; listSwarms drops them (cableHost.ts:437) though AppSwarms.swarms[].agentIds already exists, and desktop already sends it. The change key counts only agents, so a reorder or move inside a tab is not re-sent. |
Yes, gated on swarms=members-v1. cable/cableSession.ts (CableSwarm.agentIds, swarmMembers, memberMeta, swarmsMessage, key), cable/cableHost.ts (listSwarms keeps agentIds). Without the flag the frame must not change (the key may include ids only when the flag is on, so a round dial's push cadence is unchanged too). |
Long recap: deriveLongRecap; full on summary; fullText in recentSummaries/RecentTurn for restore; DeviceFleet passes fullText. 0c73edf72. |
Yes: cable_client.c:1197-1202 -> ui_project_recap_full -> pod_ui_recap. |
Partly. The mirror already stores full text (recapReads.ts RecentRecap.fullText, mirror.lastFullText); the cable layer drops it. |
Yes, gated on recap=long-v1. lib/deviceRecap.ts (deriveLongRecap, 1200/3600 limits), cable/cableSession.ts (summary(..., fullText), restore loop, CableHost.recentSummaries type), cable/cableHost.ts (RecentTurn.fullText, recentSummaries), device/deviceFleet.ts (fullText in the recap cache), cli.ts (pass mirror.lastFullText(...) to cable.summary). |
Tool step: cableStepFor, turnStep, step tee in cli.ts. 0c73edf72. |
Yes: cable_client.c:1179 -> ui_project_step. |
No. | Yes, gated on activity=steps-v1. cable/cableHost.ts (cableStepFor), cable/cableSession.ts (turnStep, drops the frame without the flag), cable/cableFleet.ts (turnStep fan-out), cli.ts (tee before cableEventFor). The branch sends step to every device ("older firmware ignores it"); on main it must not reach a round dial. |
Player activity: action + elapsedSeconds on turn.activity; terminalActivitySnapshot; activitySnapshot host hook; CommanderMirror.latestUpdate and the update field on processing; update passthrough in cableEventFor, FleetEvent, DeviceFleet. 4051f87f4. |
Yes: cable_client.c:1173-1177 -> ui_project_player_activity -> pod_ui_turn("activity") (ui_habitat.c:5919-5940). Inventory R3. |
No. terminalActivity returns the footer text only. |
Yes, gated on activity=steps-v1. cable/terminalActivity.ts, cable/cableHost.ts (activitySnapshot, cableEventFor.update), cable/cableSession.ts (sendActivity, readActivity, turnStarted(..., update)), cable/machineFleet.ts + device/deviceFleet.ts (carry update), lib/commander.ts (playerUpdate, latestUpdate, update on processing; latestAction is unused by the daemon wiring and can stay behind), cli.ts (activitySnapshot wiring). Check that update on the commander_event payload does not change what other consumers (WiFi device, desktop) receive; gate or strip it if it does. |
Player library: cable/playerLibrary.ts (playerPage, PLAYER_PAGE_SIZE=6), player.library.get, syncPlayerLibrary, welcome feature player.library, host.playerLibrary (overview sessions, else the flat agent list). 4051f87f4, 07ea2fd55. |
Yes: cable_client.c:1281-1283 (pro_player_library_parse), request at cable_client.c:204; feature bit in cable_features.c. |
No. listAgentsFlat exists. |
Yes, gated on player=library-v1. New cable/playerLibrary.ts; cable/cableSession.ts (player.library.get case, syncPlayerLibrary in the tick and after restore, player.library in features, per-session resets); cable/cableHost.ts (playerLibrary() fallback only). |
Player overview: player.overview, player.context, syncPlayerOverview, AppSwarms.overview, appSwarmsFrom overview parse in localWsServer.ts. 4051f87f4. |
No. cable_client.c:1285-1316 parses them but ui_player_context has no Pod code and pro_app_context is a no-op (owner decision 4); player.overview is round-only (Pro ignores it). |
No. | No. Not on the wire for the Pro, so nothing to port. The plan's Task 8 interface lists player.context; this document drops it because Pod does not read it. Add it later if a Pod view needs it. |
Library from the desktop inventory: overview.sessions in app_swarms, playerSessions parse in localWsServer.ts, status/lastActivityAt per session. 07ea2fd55. |
Partly: Pod reads status and ageSeconds from each row (pro_player_library_parse), but any valid value works. |
No. | Open decision (section 6). Not ported by default: the daemon fallback lists every agent the router knows with status:"idle", ageSeconds:-1. |
Creature voice: creatureVoice.ts, creatureVoiceConfig.ts, CableHost.creatureVoice. |
No. It feeds cableSpeech, and the Pro drops speech from its hello. |
No. | No. Speech synthesis for the round dial only. |
Speech frames: cableSpeech.ts, cableSession.speech, CableType.Speech = 0x05, speech.state. |
No (CABLE_TYPE_SPEECH and speech.* are in the inventory's ignored list). |
No. | No. Also avoids the frame-type 0x05 clash noted in memory (Pro Player Studio). |
Models menu: models_menu_controller.dart validUntil, models_menu_test.dart. |
No. It only feeds player.context allowance. |
No. | No. |
Desktop Player overview: AppNotifier.playerOverview, _schedulePlayerAnnouncement, overview in app_swarms, player_overview_test.dart. |
No (see overview and library rows). | No. | No, unless the open decision below is taken. |
Desktop swarms announce test, dial_swarms_announce_test.dart and the dialAnnouncementSenderForTest hook. |
No, test seam only. | No test. | No (see section 4). |
CableFleet.sessionForPort, Speech wiring in cli.ts dock companion. 92283ad62. |
No. | No. | No. Unrelated prototype session hook. |
welcome.features stays as on main plus player.library (only when player=library-v1 and the host can list). |
Yes: cable_features.c reads it; more than 16 entries, or a non-string entry, makes the firmware read 0. |
No. | Part of the library row. Keep the list <= 16. |
Where it landed on main
Main moved the devices into their own process (services/devices.ts), so the branch's cli.ts wiring became:
the step tee and update pass-through in services/devices.ts (card, toDial); the timer through a new optional
core.agents.activitySnapshot (core/api.ts, core/main.ts, core/devicesLink.ts, services/devicesProcess.ts;
the timer is parsed by elapsedSecondsOf in lib/terminalActivity.ts); action is the latest update seen on that
agent's processing card (cleared at the turn's end), not read from the core. The live summary gets full from
recentSummaries(agentId)[0].fullText when that turn's text equals the card's. Not ported: the branch's
agents key change (total|) and recording the swarms key only after a send, since neither is a Pro field and both
change a round dial's cadence.
Port order (all in cli/src/)
cable/cableSession.ts: hello parse ofswarms,recap,activity,player,board(new, main parses none today); reset on new port and on close.- The four features above, each behind its flag, with the specs in plan Task 8 step 2: a round hello (no capabilities) gets identical frames; Pro hello gets steps, full recap, members, a 6-row library.
cli.tswiring.
4. Desktop
No desktop change is required for Pod on main.
- Main's
AppNotifier._announceOpenPanesToDialalready buildsswarmRowswith'agentIds'per swarm (desktop/lib/state/app_state.dart, in theswarmRowsblock, thensendTerminalFrame('app_swarms', {...})), andAppSwarms.swarms[].agentIdsis already parsed byappSwarmsFrominlocalWsServer.ts. The daemon only drops it inlistSwarms. So tabs and members come from a daemon-only port. - Reorder and move re-send: the announce runs from the layout-change path (
_announceOpenPanesToDial()right after_layoutRevision++), and the branch'sdial_swarms_announce_test.dartis a test of that existing path, not of a new change. The branch diff inapp_state.darttouches the announce only to addoverviewand the test seam. What was missing was the daemon's change key (it hid a same-count reorder), which the daemon port fixes. - The remaining desktop diff (
playerOverview,_schedulePlayerAnnouncement,models_menu_controller.dart) feedsplayer.overview,player.contextand the library'ssessions; Pod reads none of the first two. - Plan Step 4 says to add the desktop test only if desktop is touched; it is not.
5. Messages the Pro ignores (round-only)
The Pro counts these as unknown and does nothing (debug log only): wifi.configure (a Pro feature, decide
separately), every speech.* and the speech PCM frame (0x05), models, notif.replace, notif.seen,
question.state, answer.receipt, draft.state, voice.draft, voice.question, voice.search, voice.form,
selection.state, carry.state, form.state, visit.state, metrics.state (and metrics.get),
player.overview, player.context (parsed in the shared firmware but unused by Pod), machine.selected,
machine.error, toast, and the notification fields of summary (silent, quiet, name, machine), though it reads tier (below).
The Pro does use: welcome, ping, agents.*, agent.updated, machines.*, swarms, focus,
turn.started/activity/done/error, summary, question, question.close, player.library, voice.transcript,
voice.error; and sends hello, machines.list, agents.list, player.library.get, agent.open, scroll,
pong, the voice stream and voice.cancel.
6. Open points
- Library status and age. Without the desktop
overview.sessionsport, the host's rows arestatus:"idle"andageSeconds:-1for every session. The cable session overridesstatuswith what it has seen itself:workingfromturn.started(or a step, or a live footer) untilturn.done,turn.erroror a non-silentsummary, andquestionfromquestionuntilquestion.close(thenworkingagain). Without that, the page the daemon resends every 10 s put a working agent toidleon the Pro until its next event (2026-10-11). The age is still unknown. Porting the rest meansdesktop/lib/state/app_state.dart(playerOverview),localWsServer.ts(playerSessions) andAppSwarms.overview. Owner call. player.contextin the plan. Plan Task 8 lists it as produced; Pod does not read it. Not ported here.updateonprocessing. It rides the sharedcommander_event; confirm no other consumer of the card changes behavior before portinglatestUpdate(see the Player activity row).
Firmware offers: a hello with board:"pro" gets no fw.offer (no manifest lookup, no transfer) until the Pro has its own OTA channel; a hello without board is offered as before.
7. Losing bytes on the USB path: seq and resync
The Pro's USB-Serial/JTAG path loses bytes host to device, each loss under one 64-byte packet. A small frame
(ping, turn.done, focus, question.close...) can vanish whole and nothing notices. Three layers catch it:
- Count below the driver. The firmware carries a copy of ESP-IDF 5.5.4's
esp_driver_usb_serial_jtag(firmware/components/) with two atomic counters; thealiveheartbeat printsisr_rx(bytes the ISR pulled from the hardware FIFO) andring_drop(ring-buffer pushes that failed). Ifisr_rxis short of what the daemon wrote, the loss is on the wire; ifring_dropmoves, the reader is too slow. - Partial frames. A frame that stops arriving mid-way (its tail lost) is dropped after 300 ms of silence, counted as corrupt and resynced at once, so the next frame is not swallowed by its pending length.
seq. The daemon adds an integerseqto every JSON frame it sends to a device whose hello saysboard:"pro". Eachwelcomerestarts the count at 0 (the welcome itself carriesseq:0), so a keepalive hello restarts it too. Round dials get noseq: their frames are byte-identical to before. The Pro takes the firstseqof a session as the base and expects each next one to be exactly +1; anything else is a gap (a frame was lost) and counts inseq_gapson the heartbeat. A frame withoutseq(an older daemon) is not checked.
A gap, or a frame that failed its CRC, makes the Pro send {t:"resync"}, at most once per 2 s (and not during a
firmware transfer, which has its own resend). The daemon answers with a full snapshot, also rate-limited to once per
2 s: machines.*, swarms, agents.*, notif.replace (the notices, including question rows), and a question
frame for each question still open. It does not replay recent summary history (the attach push does); a lost
summary is not repaired by a resync. The welcome of a Pro lists resync in features; a Pro whose welcome does not (an older daemon) sends the old
swarms.list and agents.list instead. A welcome without seq turns the gap check off for that session.
summary.tier
summary carries tier: "alert" (beep, wake, toast, Notices entry), "list" (Notices entry only: no beep,
wake or toast) or "silent" (the card updates, nothing else). The host still sends the older flags beside it
for firmware that predates tier: alert sets neither, list sets quiet: true, silent sets silent: true.
question and turn.error carry no tier; treat both as alert.
Super Tab waits on question
When a Super Tab member running on this machine needs a person (its assignment's work.status is needs_human or
blocked), the daemon sends it as a question with two optional fields: superTab (the tab's name, at most 48 bytes)
and attention ("decision" or "stuck"). Only a device whose hello says superTab: "wait-v1" (the Pro) is sent
one; the round dial would open a question screen that cannot load. questions[0].q is the decision asked for, or
Stuck: <reason>; the question has no key or options and is not answered on a device (question.read says so; an
answer or answer.reviewed for its id is dropped), and it does not hold Talk back: talking to the agent is how a
person unblocks it. Its id is supertab:<assignment>:<status>, under 64 bytes. It is closed with question.close
when the status moves on, the attempt or agent changes, or the tab is paused.
The Pro keeps such a question through the agent's turns (turn.started, turn.done, turn.error, a summary, a
library sweep), so it alerts once: a finished turn meanwhile neither beeps nor replaces its Notices entry or toast, a
failed one still alerts. Closed, the agent takes the state its turns left it in (idle or done, working only mid-turn).
A link drop ends it as it ends any question (it becomes a plain one the next turn ends); the daemon asks it again when
the Pro's session comes back (its agents.list) and on a resync. The library page reports a waiting agent as
question. An engine's own question for the same agent replaces it (the daemon closes the wait first) and the daemon
asks the wait again once that question closes.
Swarm cards (swarmCard: "v1")
The account's Swarms (the shared multi-agent projects of the Super Tab) as Home cards, each with its project Leads as pets; a Lead opens on its last message in the Swarm
chat (the Lead view). The daemon reads the Swarms from the backend; the Pro only draws them and reports what the person picks. Limits: at most 16 Swarms
(SWARMS_CARD_MAX), 12 Leads per Swarm (SWARM_LEADS_MAX); ids [A-Za-z0-9_-], < 48 B; names cut to 48 bytes; message text to 1,500 bytes (SWARM_TEXT_MAX - 1) on a
UTF-8 boundary. A frame over 512 JSON tokens is refused by the Pro's guard, so the list goes as supertabs plus one supertab per Swarm (12 Leads is about 200 tokens).
Daemon to Pro:
{"t":"supertabs","ids":["<superTabId>",...]} the set and order of the Swarms (at most 16); a Swarm not listed is dropped
{"t":"supertab","id":"<superTabId>","tab":"<desk tab id or empty>","name":"openharness","count":4,
"leads":[{"id":"<leadAgentId>","projectId":"<projectId>","project":"mobile","engine":"claude","owner":"","local":"<harnessId or empty>",
"state":"ask|stuck|work|done","order":1,"since":134}]} one per Swarm, replacing it; `leads` in card order
{"t":"supertab.recap","tab":"<superTabId>","lead":"<leadAgentId>","text":"...","at":1792560000,"none":false} a Lead's last message (answers recap.get, and is pushed while that Lead is open)
{"t":"supertab.focus","tab":"<superTabId>","lead":"<leadAgentId or empty>"} the desktop opened this Swarm project (empty: the Swarm, no project)
tablinks a Swarm to the desk tab of the same Swarm (empty: the desktop has none open); a linked desk tab shows the Swarm's card, an unlinked Swarm is one more Home page after the tabs.leadscome most urgent first (ask,stuck,work,done), ties in project order, and only the 12 most urgent are sent.order(optional) is the project's 0-based place in the Swarm as the app lists it (an old Swarm without projects is 0), so the round dial can show the Leads in the app's order; values may skip where a project has no Lead or was cut. The Pro puts the Leads inorder(ties, and a Lead without one, keep their place as sent): the card, its pages and the Lead view's swipe all follow it, whatever each Lead's state. The capsule still counts stuck, needs you and working.state:ask(needs you),stuck,work,done.sinceis the seconds into the project's earliest turn on the daemon's machine (the Lead's or a worker's; at least 1, and the Pro counts on from it). A Lead at work without one (remote, or working only from open assignments) shows "working" with no clock; asupertabis re-sent when that turn's start moves.localis the harness id of a Lead that runs on this account's machine, else empty. A Lead withlocaloffers the usual talk control in its view; a remote one (ownernames the person whose machine) has none.supertab.recapwithnone:true(or no text) is "No message yet". A recap for a Swarm or Lead the Pro does not hold is dropped.supertab.focusis acted on likefocusfor an agent: not while recording or updating, not within the touch quiet window (0.8 s on Home, 4 s elsewhere), and a Lead the Swarm list does not hold yet waits 5 s for it. It opens that Lead's view and sends nosupertab.select(the desktop is already there).- A link drop clears the Swarm list and any pending focus; the daemon sends
supertabsand thesupertabframes again on attach.
Pro to daemon:
{"t":"supertab.recap.get","tab":"<superTabId>","lead":"<leadAgentId>"} the Lead's last message (sent when its view opens and its text is not LOADED)
{"t":"supertab.open","tab":"<superTabId>","lead":"<leadAgentId or empty>"} the recap on screen; the lead is empty (tab names the Swarm it was in) when it closes
{"t":"supertab.select","tab":"<superTabId>","lead":"<leadAgentId>"} the person picked this Lead on the Pro: the desktop opens that Swarm tab and project
- Tapping a Lead's pet on the card, or swiping the Lead view to the previous / next Lead (card order, wrapping), sends
supertab.select, thensupertab.open(andsupertab.recap.getunless the text is already there; its state goes LOADING, drawn as "..."). Asupertab.focussendssupertab.openandsupertab.recap.getonly. - What the daemon believes is open follows what is on screen: leaving the view (back, Home, Notices, the Lead removed with none left) sends
supertab.openwith an emptylead; Notices over the view closes it and Back from Notices opens it again. A closed Lead's cached text is dropped (only the open Lead gets pushes, so it would go stale). - After a link drop nothing is sent until the Swarm list is back; then an open view sends
supertab.openandsupertab.recap.getagain. - A Lead that leaves the list while its view is open moves the view to the Lead that took its place (the next, else the last); with none left, or the Swarm gone, the view goes Home.
- Frames are sent without waiting: a busy link loses one.
The face chat and its notices (swarmChat: "v1")
With swarmChat, the daemon sends the round dial's frames (docs/device-voice-swarm-chat.md, spec 2026-10-28-round-dial-swarm-chat-design.md):
"leads":[{..., "chat":"lead|pm|group", "chatId":"<conversationId>", "chatName":"PM"}] the chat the desktop has open in that project, else the Lead's
{"t":"supertab.recap", ..., "chatId":"<conversationId>"} the last message of that chat
{"t":"supertab.notice","tab":"<superTabId>","lead":"<leadAgentId>","chat":"lead|pm","chatId":"<conversationId>","id":"<messageId>","text":"...","at":1792560000}
{"t":"supertab.notice.read","tab":"<superTabId>","lead":"<leadAgentId>","chatId":"<conversationId>"}
- A
pmorgroupchat is named under the project on the Lead view;lead(or none) draws the view as before. - A recap whose
chatIdis not the Lead's current chat is dropped. When a Lead's chat changes, an open Lead shows "..." until the new chat's message arrives. supertab.noticefiles one Notices row per Lead (its project, tagged "· ") and toasts, unless that Lead's view is on screen. A tap on either opens the Lead and sendssupertab.select.supertab.notice.readremoves the row whosechatIdmatches. A Swarm that leavessupertabs, or a Lead that leaves its Swarm'ssupertabframe, takes its rows and their toast with it: access lost sends no read.stateisworkonly for a working group and otherwisedone;askandstuckare not sent to aswarmChatdevice.supertab.focuswith an emptylead(the desktop opened the Swarm with no project, e.g. its Chat with PM) opens the Lead last shown in that Swarm, else its first.- Voice follows the face only from the Lead view. A
voice.beginthe Pro sends from a Lead view carries"face":true: the daemon posts the words in that Lead's face chat, the side it holds (the Buddy chat on the Buddy side), and a failed Buddy post is avoice.error, never a terminal turn. Withoutface(the plain agent room, which shows no chat) the daemon treats aboard:"pro"hello'svoice.beginas it did beforeswarmChat: the words post in the Lead conversation, never in the group, PM or Buddy chat the desktop has open. An older daemon ignoresface.
PM | Buddy (swarmBuddy: "v1")
{"t":"supertab", ..., "side":"pm|buddy", "leads":[{..., "chat":"buddy", "buddies":true,
"buddy":{"chatId":"<conversationId>","name":"diego's Buddy","engine":"claude","cells":25513,"color":"#d6a79a","live":true}}]}
{"t":"supertab.side","tab":"<superTabId>","side":"pm|buddy"} Pro to daemon: a tap on the pill
- With
buddies, the Lead view has a PM | Buddy pill at the top. A tap sendssupertab.sideand shows the segment pressed until the nextsupertabfor that Swarm, which carries the side the daemon keeps (the newest of the Pro and the desktop wins). - The face chat
buddydraws the Buddy's avatar (cells,color: the desktop pattern), its name in lilac and "offline" whileliveis false. With the Buddy side and nobuddy: "No Buddy here yet". - A
supertab.noticewithchat:"buddy"is quiet: no toast, no Notices row, no beep; a lilac dot on the pill's Buddy segment until the Buddy side opens. - A link drop clears the dot.