Overview

Adds the first Report — health. The server computes a deterministic verdict and renders it as text + unicode sparklines. health separates flow (is work starting?), execution (are the runs that start succeeding?), and liveness (is the telemetry fresh?), each with a headline verdict and a suggested next action — no LLM in the verdict, so there's nothing to hallucinate.

Surfaces:

• get_report MCP tool — markdown with sparklines + 🟢/🟡/🔴 status markers (or ANSI via color).
• report MCP prompt — a slash command in hosts that support prompts.
• trigger report [key] CLI — colourised in a real terminal, plain markdown when piped.
• GET /api/v1/reports/:key — format=markdown|ansi|json.

Flow is diagnosed by a cause tree (env-limit saturation, queue throttling, key starvation, trigger spike, dequeue stall) off measured queue metrics, with a runs-snapshot fallback. One ReportViewModel, rendered to markdown / ANSI / JSON.

Also: trigger mcp now always starts the server — the interactive install wizard is gated behind trigger mcp --install. Previously a TTY dropped into the wizard, so MCP hosts that spawn the server over a PTY never got a server and timed out.

Screenshots

Degraded — env concurrency-limit saturation:

/report health        prod · last 1h · vs your 7d normal

🟡 Flow slowing  ·  🟢 Execution healthy  ·  🟢 data fresh

FLOW        🟡 at your env concurrency limit (last 40 min)
  concurrency     100/100           ▁▅▆█████   pinned 40 of last 60 min

  pending         1,910     ↑ 16×   ▁▁▂▃▅▅▇█   (normal ~120)

  start latency   p95 42s   ↑ 6×    ▁▁▂▄▆▆▇█   (normal ~7s)

  worst queue     email-sends — 82% of pending

  read: limit saturated → starts lag → backlog grows
        not workers, not platform — dequeue keeps pace at ~820/min
        nothing dead-lettered

EXECUTION   🟢 the runs that DO start are fine
  failures 1.3% (normal ~1.1%) · durations normal
  read: NOT a code problem

LIVENESS    🟢 fresh — last completion 4s ago

→ Raise the env concurrency limit
  or do nothing — backlog drains in ~2.3 min once triggers ease

Healthy:

/report health        prod · last 1h · vs your 7d normal

🟢 Flow healthy  ·  🟢 Execution healthy  ·  🟢 data fresh

FLOW        🟢 starting normally — pending 84 (normal ~120) · starts p95 6s

EXECUTION   🟢 completing normally — failures 0.9% (normal ~1.1%) · durations normal

LIVENESS    🟢 fresh — last completion 2s ago

→ nothing to do

Testing

Automated (apps/webapp/test/*)

reportHealth.test.ts — 14 tests. interpret() is pure (fixed fixtures, no IO / clock / LLM):

• Golden A — degraded (env-limit saturation): cause + evidence, anomaly window (last 40 min), queue attribution, exclusions (incl. nothing dead-lettered), read chain, driving-metric annotation, footer (raise limit + "do nothing, drains in ~2.3 min"), full markdown snapshot.
• Golden B — healthy: per-finding severities, nothing_to_do footer, full markdown snapshot.
• Cause tree — every branch: dequeue_stall, env_limit_saturation, queue_limit_throttling, key_starvation, trigger_spike, and the v1-symptom fallback each select the correct cause.
• Liveness trust guard: stale telemetry forces execution unknown + a crit summary (never a false-green from a silent feed).
• ANSI render: the terminal surface uses glyphs (✓/⚠/✕) + ANSI colour, never the markdown status emoji.
• isPendingIncreasing trend helper.

reportHealthData.test.ts — 5 tests. Drives loadHealthInput through its query seam (HealthDeps) — source selection, snapshot fallback on empty and on throw (the query-concurrency 500 guard), dlq parsing (0 vs unmeasured null), and windowMinutes from the resolved (clipped) time range. The SQL/TRQL translation is owned + tested by the query service; the ClickHouse MV aggregation by the @internal/clickhouse package tests.