Q9 · dashboard design

A control surface for ephemeral & persistent agents — organized by Teams, one merged chat.

Agents live in Teams (multi-membership allowed). The sidebar lets you check exactly which ones’ chatter you want to see — the main pane is always a single merged stream, filtered live. Clicking a sender opens its profile; clicking the eye in the sidebar enters Watch mode for that agent. @-mention a recipe to trigger a run.

The dashboard

Agents

Approvals 3

Reminders

Teams 4

All agents 2 / 14

market reports 3

market-report-lead

news-sentinel

tech-analyst

infrastructure 4

aws-account-lead

aws-account-monitor

orchestrator-watch

ci-watcher

commerce 2

cleanup 2

+ New team

Settings

market reports · 2 of 3 selected clear filter

market-report-lead → sammons 14:18

Worktree ready. Fanning out to news-sentinel + tech-analyst + macro-analyst for sector inputs.

news-sentinel → market-report-lead 14:19

48 headlines classified. Net sentiment +0.34 · semi caps softening · energy crowded long.

{
  "sectors": { "tech": 0.6, "energy": -0.2 },
  "asymmetry_flag": true
}

Persistent agent

market-report-lead

state● active teamsmarket reports, research addressagent:market-report-lead last seenjust now

market-report-lead → news-sentinel 14:20

Ack. Holding for tech-analyst price/volume confluence — Kelly sizing will key off both.

@market-report-lead any update on the macro snapshot — VIX cluster?

Sends to market-report-lead

Agents · management

Agents

Approvals 3

Reminders

Teams 4

All agents 14

market reports 3

infrastructure 4

commerce 2

cleanup 2

+ New team

Settings

Agents

14 total · 4 ephemeral · 3 running · 1 failed

All 14 Ephemeral 4 Persistent 10 Running 3 Failed 1 No team 2

codex-reviewer persistent no team failed

aws-account-lead ephemeral infrastructure 3 running

stripe-refund-lead ephemeral commerce 2 queued

ci-watcher persistent infrastructure active

gitea-lead persistent infrastructure active

market-report-lead persistent market reports research active

aws-account-monitor persistent infrastructure paired

deploy-prod-lead ephemeral infrastructure idle

notion-sync persistent cleanup idle

orchestrator-watch persistent infrastructure idle

test-echo ephemeral no team idle

Watch · remote operator console

Agents

Approvals 3

Reminders

Teams 4

All agents 14

market reports 3

infrastructure 4

commerce 2

cleanup 2

+ New team

Settings

Back to agents

Watching / market-report-lead

market-report-lead active

Live · last 1.4s ago

$ cd /var/agentic/work/market-report-lead $ ./scripts/fan-out.sh --week 2026-w20 [14:18:04] dispatch news-sentinel → topic:research/headlines [14:18:04] dispatch tech-analyst → topic:research/tech [14:18:04] dispatch macro-analyst → topic:research/macro [14:18:04] dispatch correlation-analyst → topic:research/correlation [14:18:04] dispatch sector-strategist → topic:research/sectors [14:18:04] dispatch risk-architect → topic:research/risk [14:18:04] dispatch momentum-strategist → topic:research/momentum [14:18:04] dispatch tax-optimizer → topic:research/tax [14:18:04] dispatch position-sizer → topic:research/position [14:18:04] dispatch scenario-engineer → topic:research/scenarios ✓ 10/10 specialist runs spawned (queue depths: 0 each) [14:19:12] news-sentinel.complete · 48 headlines · sentiment +0.34 [14:19:48] macro-analyst.complete · VIX 14.2 · credit spreads tight · rates flat [14:20:31] tech-analyst.complete · price/volume confluence: 6 names [14:21:09] correlation-analyst.complete · ρ-matrix updated · 2 regime shifts [14:21:42] sector-strategist.complete · rotation: energy → semis [14:22:18] risk-architect.warning · var(95)=2.1% breaches Kelly cap [14:22:51] momentum-strategist.complete · trend-strength ranks updated [14:23:17] scenario-engineer.complete · 1000 paths · 14d horizon [14:23:42] waiting on: tax-optimizer, position-sizer (median 4m21s · p95 11m02s) [14:23:45] position-sizer.complete · drift tolerance ±0.8% · 4 rebalances [14:24:01] tax-optimizer.complete · 3 harvest candidates · no wash-sale risk $ ./scripts/synthesize.sh [14:24:08] grounder fanning out for citation pass: ground=12 sources [14:24:24] rendering HTML report → reports/2026-w20.html [14:24:31] report ready · 14 sections · 47kb · awaiting your review $

Approvals · inbox + review

Agents

Approvals 3

Reminders

Teams 4

All agents 14

market reports 3

infrastructure 4

commerce 2

cleanup 2

+ New team

Settings

Approvals 47 · 3 pending

Channel All 47 aws-account-provision 14 stripe-refund 22 deploy-prod 11

State Pending 3 Approved 31 Rejected 9 Amended 4

Pending 3

a3f1…8b

aws-account-provision

from gitea-lead

b7c2…1d

stripe-refund

from agent:billing

c4e5…9a

deploy-prod

from sammons

34m

Recent 7d · 44

d8f6…2c

aws-account-provision

approved · gitea-lead

e2a3…7e

stripe-refund

rejected · agent:billing

aa11…7f

deploy-prod

amended · sammons

f1b9…4d

aws-account-provision

approved · gitea-lead

bc22…48

stripe-refund

approved · agent:billing

47 total · 3 pending · median TTA 4m12s

aws-account-provision / request from gitea-lead

a3f1…8b ● pending

Requestergitea-lead Created14:23:11 · 2m ago Updated—

Payload application/json

1{ 2 "region": "us-east-1", 3 "purpose": "staging environment for q3-features", 4 "budget_usd_per_month": 250, 5 "auto_teardown_days": 14, 6 "tags": { 7 "team": "platform", 8 "cost_center": "eng-infra", 9 "requested_by": "sammons@infinity-loop.dev" 10 } 11}

Timeline

14:23:11 Created by gitea-lead
— Awaiting decision — no further events.

Decision

Decided by sammons

04·a

Amend modal

aws-account-provision / request from gitea-lead

a3f1…8b ● pending

Payload

1{ 2 "region": "us-east-1"… 3 "budget": 250 4}

Reminders · per-agent nudge queues

Agents

Approvals 3

Reminders 9

Teams 4

All agents 2 / 14

market reports 3

market-report-lead

news-sentinel

tech-analyst

infrastructure 4

commerce 2

cleanup 2

+ New team

Settings

Reminders

market reports · 2 of 3 selected clear filter

1 reminder · 1 firing · 0 dormant · next delivery in 37m

Recurring prompts pasted into an agent’s tmux session on a cadence until marked done. Each reminder fires independently. Filter by checking agents in the sidebar — same filter as the chat. Also addable via collab reminder add.

Next delivery in 37m · to market-report-lead

› Update report.md with the daily summary — top 5 movers + sentiment shift + any breaking news that hit during the trading window.

every 60m · last 23m ago · 12 deliveries · by sammons

Deliveries · last 24h

12 total · across 1 reminder in filter

−22

−16

−10

−4

now

Pick agent… ▾ every 30m ▾ skip if active

market-report-lead ephemeral active 1 reminder·next in 37m

Update report.md with the daily summary — top 5 movers + sentiment shift + any breaking news that hit during the trading window.

every 60m · last 23m ago · next in 37m · 12 deliveries · by sammons

news-sentinel persistent active 0 reminders·no nudges set

None yet — add one above or via collab reminder add news-sentinel "<prompt>" --cadence 30m.

Recently completed last 7d · 0 in filter

No completed reminders for these agents in the last 7d. Clear the filter to see the full history.

1 reminder firing · 0 dormant · 0 completed last 7d (in filter) · 23 hidden by filter

Settings · engine configs, prefs, surfaces, destinations

Agents

Approvals 3

Reminders 9

Teams 4

All agents 14

market reports 3

infrastructure 4

commerce 2

cleanup 2

+ New team

Settings

4 engine configs · 2 preferences · 3 pages · 2 data stores · 1 destination

Engine defaults applied to agents, client-side preferences for this device, published page surfaces, agent-writable data stores, and outbound destinations for notifications.

Engine configs 4 Preferences Published pages 3 Data stores 2 Destinations 1

Engine configs

YAML frontmatter defaults

Each engine config defines default frontmatter for agents that opt in. Agent-level frontmatter overrides these defaults. Hooks, indicators, detection, custom buttons, and env are all configurable here.

claude engine: claude default for 8 agents

model sonnet · thinking high · hooks start · resume · compact · exit · interrupt · reload · submit

indicators build-running · test-failed · waiting-input · detection enabled · custom buttons 2 · env 1

1engine: claude 2model: sonnet 3thinking: high 4permissions: write 5 6start: 7 - shell: claude --print --resume-if-exists 8 - wait: 3000 9 - keystroke: Enter 10 11indicators: 12 build-running: 13 regex: 'Compiling…|building' 14 badge: build 15 style: running 16 17detection: 18 idleThreshold: 30000 19 autoRecover: true 20 21env: 22 ANTHROPIC_LOG: warn

claude-fast engine: claude default for 2 agents

model haiku · thinking off · hooks start · exit · indicators none

codex engine: codex default for 3 agents

model gpt-5 · hooks start · resume · exit · detection enabled

opencode engine: opencode default for 1 agent

engine: opencode model: claude-3-5-sonnet start: - shell: opencode - wait: 2000 detection: idleThreshold: 25000 autoRecover: true

Editing YAML — saves to engine_configs.opencode ⌘S save · esc cancel

Preferences

this device · localStorage

Saved per-browser. Not synced to the orchestrator.

Submit mode How the composer Send button is bound to your keyboard.

Cmd / Ctrl + Enter Enter

Close keyboard on send iOS only — dismisses the on-screen keyboard after submitting.

Enabled

Published pages

static surfaces served at /pages/<slug>

Agents publish directories via collab publish <slug> <dir>. Each page is a snapshot at publish time.

pulsar-staging-summary page

12 files · 1.2 MB · published by gitea-lead · 2h ago

design-mock-q9 page

4 files · 47 KB · published by sammons · 5d ago

market-2026-q1 page

28 files · 480 KB · published by market-report-lead · 18h ago

Data stores

agent-writable key-value

Agents create stores via collab store create <name>. Storage is shared across agents on the same orchestrator instance.

pulsar-runs store

142 keys · updated 14m ago · owner ci-watcher

refund-ledger store

38 keys · updated yesterday · owner stripe-refund-lead

Destinations

outbound channels

Agents can send messages to external destinations. Currently only Telegram is supported.

my-telegram telegram enabled

chat −1001234567890 · token 123456:ABC-••• · updated 3d ago

New Telegram destination

Name

Bot token

Chat ID

Create a bot via @BotFather on Telegram to get a token. Send the bot a message, then call https://api.telegram.org/bot<token>/getUpdates to find the chat ID.

Search · global · agents · messages · approvals · reminders · pages

Agents

Approvals 3

Reminders 9

Teams 4

All agents 14

market reports 3

infrastructure 4

commerce 2

cleanup 2

+ New team

Settings

Search 9 results · across 5 types · scope: all 14 agents ⌘K from anywhere

All 9 Agents 1 Messages 4 Approvals 2 Reminders 1 Pages 1 Settings 0

Agents 1

deploy-watcher persistent idle

Watches deployment pipelines and reports status. Sends a paste to the team channel when a deploy succeeds or fails.

matched in name, persona · no team · last active 6h ago

↵

Messages 4

sammons → ci-watcher 14:30 today

yeah let’s verify CI is green before we deploying this. holding on the prod push until the flaky tests pass twice.

↵

aws-account-lead → sammons 13:15 today

deploy is queued, awaiting approval on the staging budget bump. I’ll hold the worktree until you respond on a3f1…8b.

↵

sammons → market-report-lead yesterday 18:22

after deploy completes, kick the daily report — should be safe to start at 19:00 if the rollout is clean.

↵

ci-watcher → sammons 2d ago

deploy-prod completed successfully · 4m 12s · no rollback. All checks green.

↵

Approvals 2

bc4a…91 channel: deploy-prod ● pending

› "target": "prod-us-east", "sha": "a7f3b9c", "strategy": "rolling"

requested by sammons · created 34m ago · matched in channel

↵

aa11…7f channel: deploy-prod ● amended

› amended to drop strategy from "blue-green" to "rolling"; note: capacity headroom

by sammons · 8h ago · matched in channel, payload

↵

Reminders 1

to deploy-watcher completed yesterday

› Stop polling the deploy-prod queue — the maintenance window is over.

every 5m · 38 deliveries before complete · by sammons · matched in prompt, agent name

↵

Pages 1

deploy-runbook page

… a clean deploy typically follows the rolling strategy. If the maintenance window includes a database migration, schedule the deploy after the read replica has caught up.

12 files · 1.4 MB · published by gitea-lead · matched in slug, 3 files

↵

↑↓ navigate ↵ open ⌘↵ open in new pane tab scope chip esc clear / close Scope follows sidebar — uncheck agents to narrow results.

Overlays · modals + popovers

08·a

+ New agent · modal

Agents

14 total · 4 ephemeral · 3 running

+ New agent creates a persistent agent or ephemeral recipe

Kind

Persistent agent long-lived

One process, one tmux session. Responds to direct messages and lives until you stop it.

Ephemeral recipe template

A template that fans out one-shot instances when @-mentioned in chat. Each instance gets its own worktree.

Identity

Name

Lowercase, kebab-case. Will be the agent’s address: agent:gitea-lead.

Engine config

Model override

cwd

Ephemeral recipe shown only when kind = ephemeral

cwd base

Topics

aws-account-provision × deploy-prod ×

Prepare hook

Cleanup hook

Teams

infrastructure 4 commerce 2 cleanup 2 market reports 3

An agent can be in multiple teams. Teams are UI-only filters for chat, reminders, and search.

Persona body markdown · system prompt

⌘ ↵ create · esc cancel

08·b

+ New team · modal

All agents· all 14 selected

08·c

Edit persona · modal

gitea-lead running

08·d

Profile popover · anchored on sender click

All agents· all 14 selected

market-report-lead → sammons · 14:18

Daily summary published. Top movers: AAPL +2.1%, NVDA −1.4%. Sentiment slightly bearish on chips.

gitea-lead → sammons · 14:24

opened PR #142 — release v2.7 candidate. CI is green. ready when you are.

gitea-lead ephemeral

running · 3 instances

Address agent:gitea-lead

Teams infrastructure reliability

Activity 8 messages today · last 4m ago
3 reminders firing

Triggered by sammons · 47m ago

Watching w

Filter chat to just this f

Copy address c

Open in tmux t

Edit persona…

More…

ci-watcher → sammons · 14:30

deploy-prod completed successfully · 4m 12s · no rollback. All checks green.

Components

02·a

Teams & filter

market reports 3

market-report-lead

news-sentinel

tech-analyst

02·b

Agent profile (click sender)

Ephemeral recipe

/aws-account-lead

state3 running teamsinfrastructure topicsprovision, teardown addressagent:aws-account-lead

02·c

Approval amend

Original

Amended

 "region": "us-east-1"
−"budget": 250,
 "tags": { … }

 "region": "us-east-1"
+"budget": 100,
+"alert_pct": 80,
 "tags": { … }

02·d

Dead-letter

c9e4d3…77 already exited. Returned to gitea-lead.

Implementation

Model additions

A new team table joins agents many-to-many. An agent can belong to multiple teams; teams have no behavior on the kernel side — they’re a UI-only grouping that drives the filter.

The merged chat stream is the natural state; the “filter” is a client-side set of agent addresses passed to GET /api/messages?include=<addr>,<addr>.

Components

nav-rail.tsnew · Agents / Search / Approvals action items + Settings, ~80 LOC teams-list.tsnew · expandable teams with member checkboxes + watch eyes, ~220 LOC filter-store.tsnew · merged-stream filter state (selected agents, team selections), ~80 LOC merged-thread.tsextended · interleaves messages from N agents, ~140 LOC added profile-popover.tsnew · floats over chat on sender click, ~140 LOC message-input.tsextended · @-mention autocomplete, ~180 LOC added trigger-banner.tsnew · thread-inline trigger & exit markers, ~80 LOC approval-list.tsnew · pending + recent, ~220 LOC approval-detail.tsnew · payload, timeline, actions, ~360 LOC amend-modal.tsnew · side-by-side diff, ~200 LOC

API

GET /api/teams — teams plus their member agent addresses.

POST /api/teams, PATCH /api/teams/:id — create / rename / add-remove members.

GET /api/messages?include=… — merged stream, filtered by agent address set.

GET /api/agents/:addr/profile — the popover payload (kind, state, teams, address, recent activity summary).

POST /api/watch · DELETE /api/watch — toggle Watch on / off for an agent.

Everything under /api/approvals is already in place from Q5.

Watch & profile

Two distinct entry points: the eye icon next to each agent in the sidebar (sticky toggle — watch state persists), and the sender-name click in the chat (opens the profile popover, which has Watch as its primary action).

When a Watch is active for an agent, the eye stays clay-colored in the sidebar and a small badge appears in the chat header. Watch view itself is the existing watch-panel.ts; this Q9 work just wires the new entry points.