feat(runs): pause/resume/cancel + unified status primitives + notification center

[santoshkumarradha]

Summary

Adds pause / resume / cancel lifecycle controls for workflow runs across the runs table, run-detail page, and a new bulk action bar — together with a notification center in the top bar and a unified set of status primitives so every status pixel in the app reads from one source of truth (getStatusTheme() in utils/status.ts).

The branch also honestly represents the asymmetry of the cancel pathway: in-flight HTTP calls to agents can't be killed, so root_execution_status is exposed separately from the children-aggregated status and the UI uses it for the kebab, dot, and live-elapsed decisions.

Lifecycle controls

• Per-row kebab on the runs table — Pause / Resume / Cancel via RunLifecycleMenu with AlertDialog confirmation, gated on root status.
• Bulk action bar at the bottom of the runs table — multi-select Pause / Resume / Cancel using Promise.allSettled so partial failures don't block the rest.
• Run detail page header cluster — same lifecycle actions inline; cancel remains available for pending / queued / waiting roots, not just running.
• Cancellation strip on the detail page — when the root is cancelled but children are still draining, an honest muted strip shows "N nodes still finishing".

Notification center

• Bell + popover lives in the sticky top bar (next to the ⌘K hint) with (N) unread prefix on document.title.
• Compact tree grouped by run, collapsed by default, one-line latest-event summary, hover tooltip for full message.
• Semantic icons per event kind — pause / resume / cancel / error / complete / start / info, no more "everything is a green check".
• Sonner toasts for transient notifications with neutral cards + thin type-tinted left borders (no richColors flooding).
• useRunNotification(opts) with an eventKind parameter; persistent log capped at 50 entries.

Unified status primitives

• StatusTheme interface in utils/status.ts extended with icon: LucideIcon and motion: "none" | "live".
• StatusDot, StatusIcon, StatusPill primitives in components/ui/status-pill.tsx — all read from getStatusTheme(), no hardcoded colors anywhere.
• StatusDot has a motion-safe:animate-ping halo for live motion statuses, frozen otherwise.
• Badge swapped Phosphor → Lucide and now derives shouldSpin from getStatusTheme(canonical).motion === "live" instead of hardcoding variant === "running".
• WorkflowNode deletes its duplicate STATUS_TONE_TOKEN_MAP and switch-based getStatusIcon and routes through the same theme.

Backend

• WorkflowRunSummary now carries RootExecutionStatus populated from agg.RootStatus.
• RunSummaryAggregation adds RootStatus *string; QueryRunSummaries and getRunAggregation select and normalize paused_count + root_status.
• deriveStatusFromCounts adds an explicit paused branch before the succeeded fallback so all-paused runs no longer collapse. Terminal check excludes paused, so completed_at / duration_ms stay nil.
• Run trace and DAG graph now desaturate child bars/nodes when the root is terminal but children are still running, so the visual matches the data model.

Live elapsed scaling

• RunsPage cells use an adaptive liveTickIntervalMs(ageMs) (1s → 5s → 30s → frozen past 1h) so visible-row timers don't pin a fan.
• NewDashboardPage duration cell uses isTerminalStatus(effectiveStatus) instead of aggregate run.terminal, so cancelled/paused roots freeze the timer even while children drain.

Multi-pass review

A 6-agent parallel review (code-quality, security, go-backend, frontend-a11y, performance, architecture) was run via Codex CLI + Gemini CLI in subprocesses. Surfaced 21 findings: 0 critical, 3 high, 13 medium, 5 low, with 0 medium-or-higher security issues.

Fixed in commit 47d5a90:

• H1 workflow_runs.go — paused branch in deriveStatusFromCounts
• H2 RunsPage — StatusMenuDot uses StatusDot; filteredRuns uses root_execution_status ?? r.status
• H3 WorkflowNode — duplicate status map deleted, routes through getStatusTheme()
• M1+M2 badge.tsx — Lucide icons + theme-driven motion
• M3+M4 RunDetailPage — statusVariant removed, header uses StatusPill
• M5 RunsPage filteredRuns — root-effective status fallback
• M6 NewDashboardPage — isTerminalStatus(effective) instead of run.terminal
• M7 RunDetailPage lifecycle cluster — Cancel available for non-terminal non-running roots
• M8 NotificationBell — removed nested role="button" over a real <button>
• M9 RunLifecycleMenu — kebab opacity /40 → /70 for WCAG contrast
• M10 status-pill StatusDot — role="img" + aria-label when label is hidden
• L2 NotificationBell — timestamp text-[10px]/80 → text-[11px] text-muted-foreground
• L3 RunLifecycleMenu — aria-busy={isPending} on the trigger

Consciously deferred:

• M11 execution_records.go ranked-root CTE — medium-confidence edge case; current MAX(CASE ...) projection is correct under the one-root-per-run invariant. Will revisit if multi-root data shows up.
• L1 / L4 / L5 — selectedRuns Map memo, WorkflowDAGLightweightResponse typing, nested-button a11y note. Non-blocking polish.

Verification

• npx tsc --noEmit — clean (exit 0)
• go build ./... — clean (exit 0)
• go vet ./... — clean (exit 0)
• No new lint errors in any touched file (3 remaining warnings are pre-existing on unchanged lines)
• Manual smoke-tested via demo workflow harness (/tmp/af_demo_runs.py) firing 7 mixed multi-node runs (sequential, parallel, nested pipelines)

Test plan

• Open /runs, fire several long runs, confirm per-row kebab → Pause / Resume / Cancel each show confirm dialogs and dispatch correctly
• Multi-select 3+ rows, exercise the bulk action bar for each lifecycle action, confirm partial failures surface in toasts
• Cancel a multi-node run while children are mid-flight, confirm:
  • the row's status dot flips to cancelled tone
  • the live duration freezes
  • the run-detail page shows the muted "N nodes still finishing" strip
  • per-node graph still spins for nodes that haven't yielded
• Pause a run, confirm the runs table, dashboard, and run-detail badge all show paused (not succeeded), completed_at stays empty
• Open the bell, fire a mix of start / pause / resume / cancel / error / complete events, confirm semantic icons render and (N) appears in the tab title
• Tab through the kebab + bell with keyboard only, confirm focus rings + aria-busy announce correctly
• Resize to mobile, confirm the bulk action bar and lifecycle cluster don't overflow

🤖 Generated with Claude Code