now

Composite top-level status surface. The current runtime grounds state at dispatch time, then composes an OpenUI surface through compose_inline so the right rail reflects what is happening across skills, agents, and dispatches.

Use this when the user asks broad intent questions like "show me what's happening right now", "give me a system overview", or "show me the status". Do not route through a per-skill .openui file.

Rubric

criteria:
  - name: resource_file_exists
    kind: deterministic
    check: "The runtime can compose a Now surface from live context without a per-skill .openui file."
  - name: queries_emit
    kind: deterministic
    check: "OpenUI Lang includes Query() calls to /agents, /recent, /full-state; all queries complete within 30s timeout."
  - name: data_renders
    kind: deterministic
    check: "Agent activity, recent skill runs, and system KPIs render as Cards with Tables or Stacks; no null/undefined values in display."
  - name: dashboard_complete
    kind: judge
    check: "Dashboard shows all 4 sections (agent activity, recent runs, KPIs, skills summary); metrics are accurate vs actual endpoint values."

The dashboard IS the verb. The chat renders the OpenUI Lang composition generated from live context and can persist it as artifact lang_body if it becomes a reusable Live App.

Dashboard shape

The Now dashboard shows:

1. Agent activity - running agents, their status, how long they have been running
2. Recent skill runs - the last 10 skill evaluations with scores and timestamps
3. System KPIs - total agents, active dispatches, recent activity count
4. Skills summary - total skills available, recent skill count

Each panel is a Card with a Table or Stack of key metrics. The dashboard queries multiple endpoints:

• /agents - list of agents and their current status
• /recent - recent skill evaluations (evals.ndjson tail)
• /full-state - high-level summary of dispatch activity

All data is live; the dashboard refreshes by re-rendering when the user asks again.

No prose steps

The dashboard is purely a UI resource. It has no executable steps - the Steps section is deliberately omitted. The render happens client-side when the snappy-chat app mounts the shape.

now - loader

Per-turn rules for the live status dashboard. Full skill: state/skills/now/SKILL.md. Live-context-injection skill, eval=auto-shape. No backend execution; the dispatch handler injects fresh JSON and the model composes from it.

Critical Rules

1. Live data flows via prompt injection, not at render time. The retired state/skills/now/resources/ui.openui (deleted 2026-05-11) used OpenUI Query() bindings that fetched data when the renderer mounted. The new mechanism is the opposite direction: state/lib/live-context-builder.ts does the reads at dispatch time and splices the JSON into the system prompt. The model sees fresh listAgents() + readRecentEvals(30) + readAgentsSummary() data each turn, composes Lang via compose_inline, and the renderer mounts the result. Static Lang, dynamic content.

1. Each call produces a different surface. The injected payloads change between calls (last_tick_at advances, recent eval rows shift). Static dashboards that look identical between calls indicate the live-context splice didn't fire. Diagnose: check state/log/head-screen.log for cache iter=1 prompt_tokens=... - the now-intent path adds ~2400 tokens vs a plain "hello" baseline.

1. Regex coverage is narrow on purpose. isNowDashboardIntent() matches the welcome starter (Show what is running right now, ...), plus shorter variants ("what is running now", "system overview", "what are the agents doing"). It deliberately does NOT match "now I want to..." or "show me the now skill" or "now what?" - those should not trigger live-status data injection. If you add a new intent phrasing, update the regex AND the SKILL.md examples.

1. Splice fires in BOTH chat and operator modes. Operator mode includes liveContextBlock unconditionally (full-assembly path). Chat mode used to skip it for token-bloat reasons; the dispatch handler now opts the now-dashboard turn into the splice via isNowDashboardIntent(intent) so the welcome starter (which lands in chat mode by default) gets the data. Other chat turns stay lean - no regression to general conversation latency.

1. Use the JSON verbatim. The injected envelope says explicitly "use this JSON verbatim; do not paraphrase or invent fields." Composed Lang must reference real agent IDs from the payload, not invented names like "Agent A" or "Worker 1".

Commands

Send any intent matching isNowDashboardIntent() to render. Examples:

• "show what is running now"
• "what is happening right now"
• "what are the agents doing"
• "system overview"
• "give me a status update"

The welcome screen starter card "Show what is running now" fires the canonical phrase.

Self-Test

• [ ] Understand live-context injection happens at dispatch time, not at render time
• [ ] Know isNowDashboardIntent() is the routing gate in state/lib/live-context-builder.ts
• [ ] Know the splice fires in both chat and operator modes
• [ ] Know each turn must produce different Lang because the injected JSON differs
• [ ] Know the model must use real agent IDs from the payload, not invented names

Found a gap? Edit this file. <!-- footer-injection-point -->