@skyhook-io/k8s-ui 1.10.3 → 1.10.5

package/src/components/trace/reachInspector.ts

@@ -0,0 +1,667 @@
+import type { Trace, RouteResult, ResourceRef } from './types'
+import type { Mark, SevTone } from './reachMarks'
+import { routeMark, routeChip, routeTone, routeAsSeenFrom, originRouteEvidence, routeForOrigin, traceInClusterRunnable } from './reachMarks'
+import type { Origin, OriginId } from './reachOrigins'
+import { strongestGap, actionableGap, originSkipReason } from './reachOrigins'
+import { hopEvidenceFor, originProducedEvidence, type GraphNode } from './reachGraphModel'
+
+// 'run-probes' re-runs the reachability probes. It is deliberately NOT
+// 'refresh': the panel that offers it promises fresh evidence, and a static
+// refetch collects none.
+export type InspectorAction = 'run-in-cluster' | 'run-probes' | 'open-resource' | 'copy-command'
+
+export interface InspectorCTA {
+  text: string
+  primary?: boolean
+  action?: InspectorAction
+  ref?: ResourceRef
+  command?: string
+  /** Set when the CTA describes something Radar cannot do - rendered inert. */
+  disabledReason?: string
+}
+
+/** Only nodes are selectable now. Edges carry no segment-local evidence, so
+ *  clicking one could only ever repeat the path result - see `Sidebar`. */
+export type Selection = string | undefined
+
+/**
+ * What the sidebar shows.
+ *
+ * `path` is ALWAYS present: whether traffic got through, from where, with what
+ * caveats, and what to do next. That question must never require a click - it
+ * is the reason the tab exists. `resource` is ADDITIVE, appearing when a node
+ * is selected, and never replaces the diagnosis.
+ */
+export interface Sidebar {
+  path: {
+    chipTone: SevTone
+    chipText: string
+    title: string
+    /** The concrete path under test. Always visible: it was inside the collapsed
+     *  details, so the panel described a result without ever naming what was
+     *  requested. */
+    request?: string
+    body: string
+    scope: { k: string; v: string }[]
+    evidence: { mark: Mark; text: string }[]
+    notProve: string[]
+    next: { header: string; body: string; blocked?: string; ctas: InspectorCTA[]; quiet?: boolean }
+  }
+  /** Every hop on the selected path, in order - the whole story for this path
+   *  seen from this vantage, rather than a summary plus whichever node was last
+   *  clicked. Reading beat clicking: understanding a path used to take three
+   *  clicks whose panel meant something different after each one. */
+  hops: HopReport[]
+  /** Configured-but-bypassed resources: on screen for orientation, outside the
+   *  journey. The label says WHY they are not part of this vantage's path. */
+  context?: { label: string; hops: HopReport[] }
+}
+
+export interface HopDetail {
+  kind: string
+  name: string
+  chipTone: SevTone
+  chipText: string
+  body: string
+  facts: { k: string; v: string }[]
+  rows?: { mark: Mark; name: string; detail: string }[]
+  moreRows?: number
+  anomalies?: { mark: Mark; text: string }[]
+  notProve: string[]
+  openRef?: ResourceRef
+}
+
+/** Configured-but-bypassed resources, grouped under one label naming WHY they
+ *  are outside this vantage's journey. Their config sections stay readable;
+ *  their journey-state words do not apply. */
+function contextGroup(ctx: Ctx, byId: Map<string, GraphNode>): Sidebar['context'] {
+  const nodes = (ctx.contextNodeIds ?? []).map((id) => byId.get(id)).filter((n): n is GraphNode => !!n)
+  if (nodes.length === 0) return undefined
+  const label =
+    ctx.origin.id === 'apiserver'
+      ? 'CONFIGURED ENTRY — KUBERNETES RELAYED THE REQUEST PAST THIS'
+      : ctx.origin.id === 'incluster' || ctx.origin.id === 'radar-incluster'
+        ? 'CONFIGURED ENTRY — THE PROBE DIALS THE SERVICE DIRECTLY'
+        : 'PARALLEL ENTRY — NOT ON THIS ROUTE\u2019S PATH'
+  return {
+    label,
+    hops: nodes.map((n) => ({
+      ...resourceSection(n),
+      id: n.id,
+      state: 'plain' as const,
+      chipText: '',
+      expanded: false,
+    })),
+  }
+}
+
+/**
+ * A hop's place in the story.
+ *
+ *  - `break`   where the request is known to have stopped. Expanded by default;
+ *              it is the answer.
+ *  - `before`  reached before that. Collapsed - it worked, so it is context.
+ *  - `after`   never tried, because something earlier stopped. Collapsed, and
+ *              labelled so its silence is not read as health.
+ *  - `plain`   no break to order around.
+ */
+export type HopState = 'before' | 'break' | 'after' | 'plain'
+
+export interface HopReport extends HopDetail {
+  id: string
+  state: HopState
+  expanded: boolean
+  /** Set on entry hops when they are parallel members of one entry stage - the
+   *  note reads "one of N parallel entry points" instead of implying sequence. */
+  parallelCount?: number
+}
+
+const NOT_DATAPLANE = 'Nothing about the normal network path. Kubernetes relayed this for us, so routing, network policy and the mesh were all skipped.'
+const SYNTHETIC_IDENTITY =
+  'That your app can reach it. This test ran from a throwaway Pod under the namespace’s default account with no token mounted — not as your application — so anything that checks who is calling may answer differently.'
+
+function originScope(o: Origin, trace: Trace): { k: string; v: string }[] {
+  const runsIn: Record<OriginId, string> = {
+    incluster: `a throwaway Pod in ${trace.subject.namespace || 'the cluster'}`,
+    'radar-incluster': 'Radar\u2019s own Pod',
+    apiserver: 'the kube-apiserver process',
+    local: 'your workstation (outside the cluster)',
+    caller: 'the application workload’s own Pod',
+    external: 'a client on the public internet',
+  }
+  return [
+    { k: 'TESTED FROM', v: o.name },
+    { k: 'RUNS IN', v: runsIn[o.id] },
+    // The laptop row's content is where the dial came FROM ("you dialled a
+    // public address…"), not who dialled - identity and network position are
+    // different facts and must not share a label.
+    { k: o.id === 'local' ? 'NETWORK POSITION' : 'IDENTITY', v: o.identity },
+    { k: 'MECHANISM', v: o.mech },
+  ]
+}
+
+/**
+ * The next-step prompt, driven by the strongest gap Radar can CLOSE. The
+ * unreachable ceiling is stated underneath as a caveat instead of being offered
+ * as an action - a button that can never be pressed is not a next step.
+ */
+function gapNext(
+  origins: Origin[],
+  current: Origin,
+  namespace?: string,
+  multiPath?: boolean,
+  inClusterRunnable = true,
+): Sidebar['path']['next'] {
+  // The server can only test a route that carries a concrete in-cluster request.
+  // When none does, the run fails with "not supported for this subject" - so the
+  // panel must say that HERE, instead of recommending it as the strongest
+  // evidence available and letting the operator find out by clicking.
+  const notRunnable = 'No path on this resource has a request Radar can send from inside the cluster, so this test would have nothing to run.'
+  // No button here. The in-cluster run is offered ON the vantage capsule in the
+  // graph, which is the thing that would produce the missing evidence - and a
+  // third copy of the same control (header, panel, capsule) was one too many.
+  // The panel keeps the REASONING, which is what it is good at.
+  const inClusterCTA = (): InspectorCTA[] => []
+  const actionable = actionableGap(origins)
+  const ceiling = strongestGap(origins)
+  const ceilingNote = ceiling?.unsupported ? `Even then, ${ceiling.name.toLowerCase()} stays untested — ${ceiling.unavailable}` : undefined
+  const denied = origins.find((o) => o.mark === 'denied')
+  // A run is never scoped to the selected path - the server tests every declared
+  // path in one pass. Saying so here stops the picker above from reading as a
+  // filter on what the button will do.
+  const allPaths = multiPath ? ' The test covers every path on this resource, not only this one.' : ''
+
+  // Mid-run, a running origin is not a "gap" any more, so the no-gap branch
+  // below would claim Radar "already has the strongest evidence" BEFORE the
+  // result exists. Being in flight is its own state, and it is the whole truth
+  // about this panel right now.
+  const runningOrigin = origins.find((o) => o.mark === 'running')
+  if (runningOrigin) {
+    return {
+      header: 'TEST RUNNING',
+      body: `${runningOrigin.name} is testing now — the result lands here when it finishes.`,
+      ctas: [],
+    }
+  }
+
+  if (!actionable && denied) {
+    const ns = namespace || '<namespace>'
+    return {
+      header: 'ASK FOR THIS PERMISSION',
+      body: `Running an in-cluster test needs \`create\` on \`jobs\` in ${ns}. Grant it, or run the check from a workload you already control.`,
+      blocked: denied.unavailable,
+      ctas: [{ text: 'Copy the permission check', action: 'copy-command', command: `kubectl auth can-i create jobs -n ${ns}` }],
+    }
+  }
+  if (actionable && actionable.id === current.id) {
+    // You are looking at the vantage that is itself the gap.
+    return {
+      header: 'RUN THIS NEXT',
+      body: inClusterRunnable
+        ? `Nothing has been tested from ${actionable.name} yet, and it is the strongest evidence Radar can still collect.${allPaths}`
+        : `${notRunnable} Every declared path here was skipped before a request could be formed.`,
+      blocked: ceilingNote,
+      ctas: actionable.id === 'incluster' ? inClusterCTA() : [{ text: '⟳ Re-run', action: 'run-probes' }],
+    }
+  }
+  if (!actionable) {
+    // Terminal state: there is nothing to do, and a call-to-action box whose
+    // message is "no action" shouts its own irrelevance. One quiet line of
+    // closure; no Re-run (the header already owns that control - the same
+    // one-copy rule as inClusterCTA), and no ceiling caveat (WHAT THIS
+    // DOESN'T PROVE states it on this same pane).
+    return {
+      header: 'NO STRONGER TEST AVAILABLE',
+      body: `This is the strongest evidence Radar can collect for this path.${allPaths}`,
+      ctas: [],
+      quiet: true,
+    }
+  }
+  return {
+    header: 'RUN THIS NEXT',
+    body:
+      actionable.id === 'incluster' && !inClusterRunnable
+        ? `${notRunnable} Every declared path here was skipped before a request could be formed.`
+        : `${actionable.name} has not been used for this path, and is the strongest evidence Radar can still collect.${allPaths}`,
+    blocked: ceilingNote,
+    ctas: actionable.id === 'incluster' ? inClusterCTA() : [{ text: '⟳ Re-run', action: 'run-probes' }],
+  }
+}
+
+interface Ctx {
+  trace: Trace
+  route?: RouteResult
+  origin: Origin
+  origins: Origin[]
+  nodes: GraphNode[]
+  /** Where the path first broke, from the graph model. */
+  breakNodeId?: string
+  /** A Service-routing boundary: the node whose EXIT the request never got
+   *  past. The break renders on that hop with exit-phrased copy - no node is
+   *  blamed. */
+  breakAtExitOf?: string
+  /** Inline nodes that are NOT network hops (the workload). Never given
+   *  journey states - a workload is neither "reached" nor a stopping place. */
+  nonNetworkNodeIds?: string[]
+  /** Configured-but-bypassed entries - context, never the journey. */
+  contextNodeIds?: string[]
+  /** Non-network nodes spliced into the display after a journey hop. */
+  interleave?: { id: string; afterId: string }[]
+  /** >1 when the journey's entries are parallel members of one stage. */
+  entryParallelCount?: number
+  /** The journey's entry-hop node ids, from the graph model. */
+  journeyEntryNodeIds?: string[]
+  /** The selected route's chain in traversal order, from the graph model. */
+  pathNodeIds?: string[]
+  stale?: boolean
+  running?: boolean
+  /** More than one scenario is on screen, so scope has to be stated explicitly. */
+  multiPath?: boolean
+  /** The HTTP path the run requests, as chosen in "what to test". */
+  httpPath?: string
+}
+
+/**
+ * Whether an in-cluster run has anything to send.
+ *
+ * The server tests a route only when it carries a concrete InClusterRequest; a
+ * route whose probes were all skipped never becomes one, so a subject can offer
+ * paths and still have nothing runnable. Knowing this from the trace is what
+ * lets the panel say so BEFORE the operator spends a Job on it.
+ */
+
+
+/** Whether the diagnosis says nothing the headline has not already said. Both
+ *  are generated, so they collide whenever the producer falls back to the same
+ *  generic sentence - and a banner repeating the line above it reads as a second
+ *  problem rather than the same one. */
+function restatesTitle(summary: string | undefined, title: string): boolean {
+  const norm = (x: string) => x.trim().toLowerCase().replace(/[.!]$/, '')
+  return !!summary && !!title && norm(summary) === norm(title)
+}
+
+/** The persistent diagnosis: did traffic get through, from where, and what next. */
+function pathSection(ctx: Ctx): Sidebar['path'] {
+  const { trace, route, origin, origins } = ctx
+  // The route outcome is merged across origins. Without this gate the panel
+  // rendered another vantage's success under the selected vantage's name - a
+  // permanently unavailable origin could read "a real request went through"
+  // while the graph beside it said "not routable". Same lie the graph already
+  // guards against, in the surface users actually read.
+  // This origin's OWN result when the producer sent one; the coarse
+  // "did this origin produce anything" gate only remains as the fallback.
+  const ev = originRouteEvidence(route, origin.id)
+  const asSeen = ev.kind === 'none' ? undefined : ev.result
+  // A config-derived break is true of every vantage and observed by none, so it
+  // reads as a configuration fact - never as this origin's failed dial.
+  const fromConfig = ev.kind === 'config'
+  // Which KIND of derived break. Declared-config is broken whatever the cluster
+  // is doing; cluster-state is true right now and changes when the workload
+  // does. Calling the second one a configuration failure sends the reader to
+  // edit YAML when the fix is to get Pods ready.
+  const basis = ev.kind === 'config' ? ev.result.basis : undefined
+  const hasEvidence = ev.kind === 'own' || (ev.kind === 'rollup' && originProducedEvidence(origin))
+  const mark: Mark = fromConfig
+    ? 'config'
+    : hasEvidence
+      ? asSeen
+        ? routeMark(asSeen, { stale: ctx.stale, running: ctx.running })
+        : 'untested'
+      : origin.mark
+
+  const notProve: string[] = []
+  if (origin.kind === 'synthetic') notProve.push(SYNTHETIC_IDENTITY)
+  if (origin.kind === 'relayed') notProve.push(NOT_DATAPLANE)
+  const hasFrontDoor = (trace.upstreams ?? []).length > 0
+  const external = origins.find((o) => o.id === 'external')
+  if (hasFrontDoor && external?.unsupported && origin.id !== 'external') {
+    // "No request has come in from outside" is only true until Radar's own
+    // machine dials the public entry and gets an answer - that request DID come
+    // from outside. What stays unproven then is narrower: a real user's request.
+    const outsideDialAnswered = (trace.upstreams ?? []).some((h) => {
+      const m = hopEvidenceFor(h, { id: 'local' }, trace.runVantage)?.mark
+      return m === 'proved' || m === 'answered'
+    })
+    notProve.push(
+      outsideDialAnswered
+        ? 'That real users can reach it — Radar dialled the public entry from outside and got an answer, but no real user request has been observed.'
+        : 'That people on the internet can reach it — no request has come in from outside.',
+    )
+  }
+
+  const evidence: { mark: Mark; text: string }[] = []
+  const seen = new Set<string>()
+  const add = (m: Mark, text: string) => {
+    const key = text.trim().toLowerCase()
+    if (!key || seen.has(key)) return
+    seen.add(key)
+    evidence.push({ mark: m, text })
+  }
+  if ((hasEvidence || fromConfig) && asSeen?.evidence) add(mark, asSeen.evidence)
+  // What this vantage saw at each HOP. The route is built from the backend's
+  // probes, so a laptop that dialled the front door and got an answer had all
+  // of it discarded and read as "no test has been run from here" - beside a
+  // graph drawing that very dial.
+  const hopSeen = ctx.nodes
+    .filter((n) => !n.isOrigin && n.hop)
+    .map((n) => ({ node: n, ev: hopEvidenceFor(n.hop, origin, trace.runVantage, { stale: ctx.stale, running: ctx.running }) }))
+    .filter((x): x is { node: GraphNode; ev: NonNullable<ReturnType<typeof hopEvidenceFor>> } => !!x.ev)
+  if (!hasEvidence && !fromConfig) {
+    for (const { node, ev: e } of hopSeen) add(e.mark, `${node.kind.toLowerCase()} ${node.name} — ${e.title || e.label}`)
+  }
+  // The one boundary two observations can establish. Stated as the reasoning
+  // that produced it, so it reads as evidence rather than as a verdict.
+  if (ev.kind === 'own' && routeForOrigin(route, origin.id)?.failedBoundary === 'service-routing') {
+    add('failed', 'the Pods answered directly, but the Service did not — so the Service’s own routing is what breaks')
+  }
+  // Localization facts are behind-the-gate (apiserver / direct-pod) evidence.
+  // They belong to the relayed origin, not to whichever origin is selected -
+  // listing them under the in-cluster probe credited it with observations it
+  // never made.
+  for (const f of origin.id === 'apiserver' && hasEvidence ? route?.localization ?? [] : []) {
+    const layer = f.layer.toUpperCase()
+    const detail = f.detail?.trim()
+    const body = !detail ? layer : detail.toUpperCase().startsWith(layer) ? detail : `${layer} · ${detail}`
+    add(f.ok ? 'proxied' : 'failed', `${body} — checked directly, past the entry point`)
+  }
+  if (evidence.length === 0) {
+    // A route the run SKIPPED still carries why. The reason is looked up from
+    // the skip rows by the exact identity that produced it, so it appears only
+    // under the vantage whose dial was skipped - never charging the in-cluster
+    // probe with the proxy's limits, or the laptop with the proxy's timeouts.
+    const skipReason = ev.kind === 'rollup' && asSeen?.outcome === 'not-tested' ? originSkipReason(trace, origin.id, route) : undefined
+    // An in-cluster run that was ATTEMPTED and never started is not "no test
+    // has been run" - the capsule preserves the attempt and this row must not
+    // erase it. The error itself is the observation.
+    const attemptError = origin.id === 'incluster' && origin.mark === 'blocked' ? origin.unavailable : undefined
+    add(
+      mark,
+      // The attempt error outranks skip rows: a prior Job's leftover skips
+      // would otherwise paint a fresh failed run (image pull, quota) with last
+      // run's reason, one pane from a capsule saying "test couldn't run".
+      attemptError ||
+        skipReason ||
+        (origin.unsupported
+          ? 'Radar cannot test from here, so nothing has been learned this way'
+          : origin.mark === 'denied'
+            ? 'not permitted to run this test'
+            : // A vantage that CANNOT be used states why (nothing dialable from
+              // the laptop, a skipped mechanism) - "no test has been run" reads
+              // as a test someone forgot, which is a different claim.
+              origin.unavailable || 'no test has been run from here'),
+    )
+  }
+
+  const failed = mark === 'failed'
+  // Only when it is about THIS path. A trace carries one diagnosis but many
+  // routes, so an unattributed or sibling-owned cause must not be rendered as
+  // the selected path's - the route's own evidence is what remains true.
+  const rawDiagnosis = trace.diagnosis
+  const diagnosis = rawDiagnosis && (!rawDiagnosis.route || rawDiagnosis.route === route?.route) ? rawDiagnosis : undefined
+  const reachedSomething = !hasEvidence && !fromConfig && hopSeen.some((x) => x.ev.mark !== 'failed')
+  const body = basis === 'cluster-state'
+    ? 'Nothing is ready to serve this path right now, so it cannot work from any vantage. No request was sent to establish that — it is read off the current state of the cluster, and it changes when the workload does.'
+    : fromConfig
+    ? 'The configuration itself is broken, so this path cannot work from any vantage. No request was sent to establish that — it is read off what is declared.'
+    : reachedSomething
+    ? 'This vantage did reach part of the path — see what it saw below. It has no result for this route as a whole, so the end-to-end journey from here is still unproven.'
+    : !hasEvidence
+    ? (origin.unavailable || 'Nothing has been tested from here, so this says nothing about whether traffic gets through.') +
+      // When NOTHING was tested anywhere, the health dots are the only colour
+      // on the board and read as a passed test - the one state where the
+      // dot/line split needs saying out loud, not just in the caption.
+      ((trace.coverage?.tested ?? 0) === 0 ? ' The dots on the graph show each resource’s own reported health — cluster state, not a test result.' : '')
+    : failed
+    ? 'This is the first confirmed failure. Everything after it was never tried, so there is nothing to report past this point.'
+    : mark === 'proved'
+      ? 'A real request went through and the target answered.'
+      : mark === 'proxied'
+        ? 'Kubernetes relayed a request and the target answered — which shows something is serving, not that the normal path works.'
+        : mark === 'untested'
+          ? 'Nothing has been tried from here yet. Configuration may look right, but that is intent, not proof.'
+          : mark === 'stale'
+            ? 'This result predates a change to the cluster, so it is set aside rather than trusted.'
+            : mark === 'running'
+              ? 'A test is running. Earlier results stay until new ones replace them.'
+              : // A proxy-only failure wears the same amber mark as a real answer,
+                // but nothing answered - saying "the target answered" here sent
+                // the reader to debug an application response that never existed.
+                asSeen?.outcome === 'unreachable' && asSeen?.confidence === 'indirect'
+                ? 'The relayed dial failed. The proxy bypasses the real path, so this does not condemn it — but nothing answered, and the real path is still untested.'
+                : 'The target answered, but not with what was asked for.'
+
+  return {
+    chipTone: asSeen ? routeTone(asSeen, { stale: ctx.stale, running: ctx.running }) : 'unknown',
+    chipText: asSeen ? routeChip(asSeen, { stale: ctx.stale, running: ctx.running }) : 'not tested',
+    title: `${origin.name} → ${route?.target || trace.subject.name}`,
+    request: route ? `${route.route}${ctx.httpPath && ctx.httpPath !== '/' ? ` · HTTP path ${ctx.httpPath}` : ''}` : undefined,
+    body,
+    scope: [...originScope(origin, trace), ...(route ? [{ k: 'PATH', v: route.route }] : [])],
+    evidence,
+    notProve,
+    next:
+      failed && diagnosis
+        ? {
+            header: 'LIKELY CAUSE',
+            body: diagnosis.summary + (diagnosis.nextAction ? ` ${diagnosis.nextAction}` : ''),
+            ctas: [
+              ...(diagnosis.culpritResource ? [{ text: 'Open the culprit', primary: true, action: 'open-resource' as InspectorAction, ref: diagnosis.culpritResource }] : []),
+              ...(diagnosis.command ? [{ text: 'Copy the command', action: 'copy-command' as InspectorAction, command: diagnosis.command }] : []),
+            ],
+          }
+        : gapNext(origins, origin, trace.subject.namespace, ctx.multiPath, traceInClusterRunnable(trace)),
+  }
+}
+
+/** The additive detail for a selected node. Never replaces the diagnosis. */
+function resourceSection(node: GraphNode): HopDetail {
+  const hop = node.hop
+  const findings = hop?.findings ?? []
+
+  if (node.podRows) {
+    const roster = hop?.config?.pods ?? []
+    const total = hop?.config?.podTotal ?? roster.length
+    const ready = typeof hop?.meta?.ready === 'number' ? (hop.meta.ready as number) : roster.filter((p) => p.ready).length
+    const selected = typeof hop?.meta?.selected === 'number' ? (hop.meta.selected as number) : total
+    const publishNotReady = !!hop?.meta?.publishNotReadyAddresses
+    const notReady = publishNotReady ? [] : roster.filter((p) => !p.ready)
+    const omitted = total - roster.length
+    const notProve: string[] = []
+    if (omitted > 0) notProve.push(`The ${omitted} Pods that were not tested. Untested is not proven.`)
+    if (notReady.length > 0) notProve.push(`The ${notReady.length} not-ready Pod${notReady.length > 1 ? 's' : ''} — nothing was sent to them, so nothing was learned.`)
+    return {
+      kind: 'PODS',
+      name: `${ready} of ${selected} eligible`,
+      chipTone: node.tone,
+      chipText: 'backends',
+      body: publishNotReady
+        ? 'The Pods behind this Service. This Service is set to send traffic to Pods even before they report ready.'
+        : 'The Pods behind this Service. Kubernetes only sends traffic to the ones that report ready.',
+      facts: [
+        { k: 'MATCHING PODS', v: `${selected}` },
+        // Derived from readiness, NOT from observed delivery - "taking traffic"
+        // claimed evidence we do not have.
+        { k: 'ELIGIBLE', v: `${ready}` },
+        {
+          k: 'SITTING OUT',
+          v: publishNotReady ? 'none — not-ready Pods get traffic too' : notReady.length > 0 ? `${notReady.length} not ready` : 'none',
+        },
+      ],
+      rows: node.podRows,
+      moreRows: node.moreRows,
+      anomalies: node.anomalies,
+      notProve,
+    }
+  }
+
+  const c = hop?.config
+  const facts: { k: string; v: string }[] = []
+  if (c?.clusterIP) facts.push({ k: 'CLUSTER IP', v: c.clusterIP })
+  if (c?.serviceType) facts.push({ k: 'TYPE', v: c.serviceType })
+  if (c?.ports?.length) facts.push({ k: 'PORTS', v: c.ports.map((x) => `${x.port}→${x.targetPort ?? x.port}`).join(', ') })
+  if (c?.addresses?.length) facts.push({ k: 'ADDRESS', v: c.addresses.join(', ') })
+  if (c?.hostnames?.length) facts.push({ k: 'HOSTS', v: c.hostnames.join(', ') })
+  if (c?.selector) facts.push({ k: 'SELECTOR', v: Object.entries(c.selector).map(([k, v]) => `${k}=${v}`).join(', ') })
+  if (node.ref?.namespace) facts.push({ k: 'NAMESPACE', v: node.ref.namespace })
+
+  return {
+    kind: node.kind,
+    name: node.name,
+    chipTone: node.tone,
+    chipText: node.dim ? 'not on this path' : '',
+    body: node.dim
+      ? 'This entry point is attached to the resource but does not serve the host being tested.'
+      : findings.length > 0
+        ? findings[0].cause || findings[0].message
+        : 'Configuration and health of this resource.',
+    facts,
+    notProve: [],
+    openRef: node.ref,
+  }
+}
+
+/**
+ * Builds the sidebar. The diagnosis is always computed; a node selection only
+ * appends to it.
+ */
+export function buildSidebar(sel: Selection, ctx: Ctx): Sidebar {
+  const path = pathSection(ctx)
+  // The selected route's own chain, from the graph's traversal - NOT every node
+  // sorted by position. Sorting served siblings up as if they were sequential
+  // hops and included branches this route never touches.
+  const byId = new Map(ctx.nodes.map((n) => [n.id, n]))
+  const journey = (ctx.pathNodeIds ?? [])
+    .map((id) => byId.get(id))
+    .filter((n): n is GraphNode => !!n && !n.isOrigin)
+  // Display order interleaves non-network nodes (the workload) after their
+  // parent - present to read, absent from the journey's state machine.
+  const chain: GraphNode[] = []
+  for (const n of journey) {
+    chain.push(n)
+    for (const iv of ctx.interleave ?? []) {
+      if (iv.afterId === n.id) {
+        const wl = byId.get(iv.id)
+        if (wl) chain.push(wl)
+      }
+    }
+  }
+  const entryIds = new Set(ctx.journeyEntryNodeIds ?? [])
+  const nonNetwork = new Set(ctx.nonNetworkNodeIds ?? [])
+  // A boundary break anchors to the EXIT of its source hop (the Service);
+  // breakNodeId anchors to a landing node. Either way one index carries 'break'.
+  const anchorId = ctx.breakNodeId ?? ctx.breakAtExitOf
+  const breakIdx = anchorId ? chain.findIndex((n) => n.id === anchorId) : -1
+  // A break at ONE of N parallel entries orders nothing: a sibling entry is
+  // not "before" it, and the hops behind the stage were not necessarily
+  // stopped - the request may have gone through a sibling. Serial
+  // before/after semantics apply only when the break is outside the stage.
+  const breakInParallelStage =
+    breakIdx >= 0 && (ctx.entryParallelCount ?? 0) > 1 && entryIds.has(chain[breakIdx]?.id ?? '')
+  const stateOf = (i: number, id: string): HopState => {
+    // The workload is not a network hop: it is never "reached", never a place
+    // a request stops - whatever happens around it.
+    if (nonNetwork.has(id)) return 'plain'
+    if (breakIdx < 0) return 'plain'
+    if (breakInParallelStage) return i === breakIdx ? 'break' : 'plain'
+    if (i < breakIdx) return 'before'
+    return i === breakIdx ? 'break' : 'after'
+  }
+  // What to open when the reader has not chosen: the break if there is one,
+  // else the destination - the point of the path. Never everything at once,
+  // which is how a report becomes a wall.
+  const defaultOpen = breakIdx >= 0 ? breakIdx : chain.length - 1
+  return {
+    path,
+    hops: chain.map((n, i) => {
+      const section = resourceSection(n)
+      return {
+        ...section,
+        id: n.id,
+        state: stateOf(i, n.id),
+        // Exit-phrased for a boundary break: the SERVICE hop carries it, and
+        // "stopped here" would blame the hop itself for its exit's failure.
+        chipText:
+          i === breakIdx && ctx.breakAtExitOf === n.id
+            ? 'routing to its Pods breaks just past here'
+            : section.chipText,
+        parallelCount: entryIds.has(n.id) && (ctx.entryParallelCount ?? 0) > 1 ? ctx.entryParallelCount : undefined,
+        expanded: sel ? n.id === sel : i === defaultOpen,
+      }
+    }),
+    context: contextGroup(ctx, byId),
+  }
+}
+
+/** The headline verdict band. Derived from the selected scenario, never from an
+ *  aggregate that could hide a failing route behind passing siblings. */
+const VERDICT_TONE: Record<string, SevTone> = { healthy: 'healthy', degraded: 'degraded', broken: 'unhealthy', unknown: 'unknown' }
+
+export function buildVerdict(
+  trace: Trace,
+  route: RouteResult | undefined,
+  opts: { stale?: boolean; running?: boolean; pathLabel?: string; originId?: string; originName?: string } = {},
+): {
+  tone: SevTone
+  chipText: string
+  /** Set when the resource has more than one path, so the reader can tell the
+   *  route-scoped badge apart from the resource-wide headline beside it. */
+  chipScope?: string
+  scopeLabel?: string
+  title: string
+  problem?: string
+  body: string
+} {
+  // With no route there is nothing to derive a tone from, but the backend has
+  // still reached a verdict (e.g. a config fault found without probing). Falling
+  // through to 'unknown' showed a grey dot on a resource the tracer called
+  // degraded.
+  // The band sits directly above the inspector, which reads the SELECTED
+  // origin's own result. Leaving the band on the merged rollup let the two
+  // contradict each other in adjacent panes - "could not get through" over
+  // "got through" - on exactly the disagreeing traces per-vantage evidence
+  // exists to represent.
+  const seen = opts.originId ? routeAsSeenFrom(route, opts.originId) : route
+  // A route exists but this origin has no row: the badge says "not tested",
+  // and it must wear the NEUTRAL tone - painting it with the resource-wide
+  // verdict colour (amber on a degraded trace) dressed a vantage-scoped
+  // "not tested" as a warning about that vantage. The verdict fallback stays
+  // for traces with no routes at all (a config fault found without probing).
+  const tone: SevTone = opts.running ? 'info' : seen ? routeTone(seen, opts) : route ? 'unknown' : VERDICT_TONE[trace.verdict] ?? 'unknown'
+  return {
+    tone,
+    chipText: opts.running ? 'testing' : seen ? routeChip(seen, opts) : 'not tested',
+    // Title, problem and coverage describe the WHOLE resource; tone and chip
+    // follow the selected path. With several paths on screen that difference is
+    // invisible unless each side says which scope it speaks for.
+    // The badge now follows the selected path AND vantage while the headline,
+    // problem and coverage stay resource-wide. Both scopes are legitimate; what
+    // is not legitimate is leaving the reader to guess which is which, so the
+    // badge names its own scope in full.
+    // Prefixed here rather than at the render site: with no pathLabel the old
+    // template produced the visible "for from Radar on your machine".
+    chipScope: [opts.pathLabel ? `for ${opts.pathLabel}` : '', opts.originName ? `from ${opts.originName}` : '']
+      .filter(Boolean)
+      .join(' · ') || undefined,
+    scopeLabel: opts.pathLabel || opts.originName ? 'THIS RESOURCE' : undefined,
+    // A stale screen previously led with the old headline ("Reachable...") and
+    // then said underneath that the result was excluded. That is a contradiction,
+    // not an exclusion.
+    title: opts.stale
+      ? 'This result is out of date — re-test'
+      : trace.headline || route?.route || `Reachability · ${trace.subject.name}`,
+    // The diagnosis is a named fault with a culprit and a next action - it
+    // answers "why not", where the headline only says how much was tested. It
+    // is called out rather than rendered as body prose, which made the more
+    // important fact read as an explanation of the less important one.
+    // A diagnosis that only restates the headline is not a second fact. The
+    // banner rendered "couldn't actively test any route from here" directly
+    // under a title saying exactly that.
+    problem: restatesTitle(
+      trace.diagnosis?.summary,
+      opts.stale ? '' : trace.headline || route?.route || `Reachability · ${trace.subject.name}`,
+    )
+      ? undefined
+      : trace.diagnosis?.summary,
+    body: trace.diagnosis ? '' : trace.reason || '',
+  }
+}