feat(dashboard): Mission Control — unified fleet command center for builds, chats & sub-agents

[umi-appcoder]

Status (2026-07-24): Phases 1–3 and the mobile screen have shipped (#426, #432); the board is now three columns — #435 folded "Needs review" into Done, since the outcome badge on each finished card already carried that signal. Phase 4 (fleet scope) is dropped by design: workspaces are per-person, so a cross-pod board would aggregate other people's agents — the per-workspace board already shows everything you own (this also sidesteps the #360 cross-agent visibility question). Remaining work is tracked in the reworked Phases section below.

Problem

Running more than one agent at a time means juggling three tabs: builds live in Build, hypervisor chats in Hypervisor, and orchestrator sub-agents are buried inside a build's Subagents sub-tab. There is no single answer to the only question that matters when you run a fleet: "what needs me right now?" A build blocked 14 minutes on a permission prompt is invisible unless you happen to open that tab. This also compounds #267 (sidebar overload) — each new work surface today becomes another tab.

Industry context: every major platform converged on exactly this surface in 2026 (Devin Desktop's Agent Command Center, GitHub Agent HQ, Claude Code's claude agents dashboard, Cursor's cloud-agent dashboard). With Ona acquired by OpenAI, there is no open-source self-hosted equivalent — Coder's closest offering is a paid add-on. kube-coder already has every data source; it's missing only the unifying view.

Proposal: Mission Control

A new top-level tab: a board that unifies builds + hypervisor chats + sub-agents (and later, other pods via the controller) into one priority queue, organized by what needs a human:

Running → Waiting on you → Needs review → Done today

Card anatomy

• Kind (BUILD / CHAT / SUB-AGENT), assistant + model, elapsed/wait time
• AI-written one-line headline of what the agent is doing right now ("Wiring per-trigger history into server.py — 8 files touched")
• Repo + worktree branch, lineage (parent build ↔ spawned sub-agents, cross-linked)
• Footer actions: Reply · Watch · Transcript · Kill (Retry on failed, Resume on parked)

The killer column: "Waiting on you"

When an agent blocks on a permission prompt or a question, the actual prompt renders on the card with quick-reply buttons — answer from the board without opening the session. Amber edge + wait-age ("⏸ 14m") makes stalls impossible to miss. Reuses the quick-reply plumbing from #276 and the waiting-state detection behind WaitingBadge.

Needs review

Finished work shows evidence chips (tsc ✓ · vitest 214 ✓ · screenshots ×4 · e2e ✕ 1 flaky) and a PR deep-link, so review starts from evidence, not from a raw diff.

Detail drawer

Selecting a card opens a drawer: activity timeline (started from issue → read files → spawned sub-agents → ran tests → currently editing), live token/tool-call rate, and a follow-up composer — steer any agent without leaving the board.

Pulse strip

One-glance fleet state: 3 running · 2 waiting on you · 1 needs review · oldest wait 14m · events 26/min.

Scope switcher

"This workspace / Fleet (N pods)" — phase 4 aggregates every workspace through the controller using the existing Bearer remote-task API.

Light mode

Mobile

Columns collapse to stacked swimlanes, Running/Waiting first — the "run your fleet from your phone" surface. Pairs naturally with Walkie: "what's my fleet doing?" → spoken pulse-strip summary.

Why this is cheap relative to how big it feels

Almost every data source already exists:

Board element	Existing source
Build cards, states, quick replies	/api/claude/tasks (+ waiting detection, #276)
Chat cards	/api/hypervisor/threads + activity classification (#312)
Sub-agent cards + lineage	agent-orchestrator list_subagents / task-context lineage
Timeline drawer	structured transcript (#294) + activity timeline (#298)
Live updates	/api/events SSE
Fleet scope	controller + Bearer remote-task API

Genuinely new: a unified queue endpoint normalizing the three sources into one card schema, a cheap headline generator (classifier pass over recent transcript events), and the board UI.

GET /api/missioncontrol/queue
→ [{ id, kind: build|chat|subagent, assistant, model, state: running|waiting|review|done,
     headline, title, repo, branch, started_at, waiting_since?, waiting_prompt?,
     lineage: { parent?, children[] }, evidence?, links: { transcript, pr?, app? } }]

Phases

• 1. Queue endpoint + read-only board — shipped in feat(dashboard): Mission Control — unified agent queue across builds, chats & sub-agents (#425) #426
• 2. Actions on cards — quick replies, follow-ups, kill — shipped in feat(dashboard): Mission Control — unified agent queue across builds, chats & sub-agents (#425) #426
• 3. Detail drawer — timeline, output tail, follow-up composer — shipped in feat(mission): detail drawer — per-card timeline, output tail & follow-up composer (#425) #432 (headlines are log-tail heuristics for now; classifier pass below)
• 4. Fleet scope — dropped by design (2026-07-24): workspaces are per-person; no cross-pod aggregation needed
• 5. Mobile — stacked-swimlane Mission Control screen shipped in feat(dashboard): Mission Control — unified agent queue across builds, chats & sub-agents (#425) #426 (voice summary below)

Remaining:

• Classifier-written headlines — replace the log-tail heuristic (server.py marks the deferral) with a cheap classifier pass over recent transcript events
• Evidence chips on Done cards — tsc/vitest/screenshot/PR chips, re-homed onto Done after refactor(mission): fold Needs review into Done — three-column board #435 removed the "Needs review" column
• Walkie voice summary — "what are my agents doing?" → spoken pulse-strip readout for this workspace

Mockups

Interactive HTML mock (uses the real design tokens from tokens.css): docs/mockups/mission-control/index.html — assets on branch mission-control-mockup-assets.

Related

#267 (sidebar overload — this absorbs the "too many places to look" problem structurally) · #312 (activity classification — reused for card kinds) · #276 (quick replies) · #250 (mobile parity) · #363 (product metrics could feed the pulse strip) · #360 (cross-agent output reading — fleet scope must respect whatever posture is decided there)

[umi-appcoder]

Phases 1–2 (queue endpoint + interactive board on web and mobile) are implemented in #426. Remaining phases — classifier headlines, detail drawer, fleet scope, Walkie voice summary — stay tracked here.

[umi-appcoder]

Phase 3's detail drawer (per-card timeline, output tail, follow-up composer) is implemented in #432. Still tracked here: classifier-written headlines, fleet scope, Walkie voice summary.

[umi-appcoder]

Board simplification + scope decision:

• refactor(mission): fold Needs review into Done — three-column board #435 (merged) folded "Needs review" into Done — the board is now three columns (Waiting on you → Running → Done). The ✓/✗ outcome badge on each finished card carries the review signal; a separate parking column added little and cramped the desktop strip (feat(desktop): centered composer, Mission Control strip & bottom dock (#433) #434) and mobile swimlanes.
• Phase 4 (fleet scope) is dropped by design. Workspaces are per-person, so a cross-pod board would mostly surface other people's agents — the per-workspace board already answers "what needs me right now?" for everything you own. This also sidesteps the security/feat: cross-agent output reading is unscoped today — decide posture, then design "link chat" properly #360 cross-agent visibility question. If one person ever runs several workspaces, the lighter answer is per-pod notifications, not a merged board.

Issue body updated accordingly. Still open here: classifier-written headlines, evidence chips (now on Done cards), and the Walkie voice summary.

[imran31415]

I thinnk the remaining items can be scoped as sep tasks.