@@ -876,15 +876,15 @@ const verdictContract = "VERDICT BLOCK: " +
 	"inconclusive=true when you could not determine the cause: root_cause empty, the blocking question first in unresolved, verify or investigate steps offered. healthy and inconclusive are mutually exclusive. " +
 	"root_cause: the paragraph under the headline, two or three plain sentences for a reader who will not open the analysis: what failed, the mechanism, and the one result that shows it, with since-when if a result gives it. No [[radar:...]] markers here or in summary. GitHub-flavored markdown; backticks only for an identifier the reader needs exactly, not for every name. Claim nothing the ledger does not cover; empty when healthy or inconclusive. " +
 	"root_cause_evidence_refs: at most 3 refs of the checks that establish WHY, copied exactly from their [[radar:evidence-ref=ev_...]] markers; empty when root_cause is empty. " +
-	"evidence: at most 8 results from this investigation that returned data (a call that errored or was denied cannot be cited; say in the story what it would have shown), each with its ref copied exactly, a role (cause: establishes why; symptom: what it looks like; context: checked and worth seeing; benign: looks adverse but is not a live problem; demoted: related but less relevant; rules_out: this result contradicts a hypothesis), an optional one-sentence claim (required for benign, demoted and rules_out, 200 characters max), and a gap: one clause naming what this result does not cover that a reader might assume it does (the other restarts, the other pods, earlier history), written to stand on its own under the card, a fragment rather than a sentence: 'live spec only, not the Git manifest', never 'Shows the live spec, not the manifest'; omitted when nothing is assumed. A gap is a limit of the evidence, never a licence to claim past it. Refs may repeat with different subjects. When a result covers several resources, pods, containers or log streams, subject names the one you mean (kind and name required; group is the API group and is empty for core kinds such as Pod, ConfigMap, Secret, Service, Event and Node; for a listing, name the entry you mean, or give kind and namespace with no name when the point is what the listing does not contain; namespace, container, stream and observation as needed, observation being the evidence kind, exactly one of: resource, logs, events, changes, metrics, metrics:cpu / metrics:memory / metrics:restarts for a diagnose bundle's charts, alerts with name as the rule's name, or issue; any other word leaves the item unplaced). The diagnose bundle is never about one thing: every citation of it needs a subject. " +
+	"evidence: at most 8 results from this investigation that returned data (a call that errored or was denied cannot be cited; say in the story what it would have shown), each with its ref copied exactly, a role (cause: establishes why; symptom: what it looks like; context: checked and worth seeing; benign: looks adverse but is not a live problem; demoted: related but less relevant; rules_out: this result contradicts a hypothesis), an optional one-sentence claim (required for benign, demoted and rules_out, 200 characters max), and a gap: one clause naming what this result does not cover that a reader might assume it does (the other restarts, the other pods, earlier history), written to stand on its own under the card, a fragment rather than a sentence: 'live spec only, not the Git manifest', never 'Shows the live spec, not the manifest'; omitted when nothing is assumed. A gap is a limit of the evidence, never a licence to claim past it. Refs may repeat with different subjects. When a result covers several resources, pods, containers or log streams, subject names the one you mean (kind and name required; group is the API group and is empty for core kinds such as Pod, ConfigMap, Secret, Service, Event and Node; for a listing, name the entry you mean, or give kind and namespace with no name when the point is what the listing does not contain; namespace, container, stream and observation as needed, observation being the evidence kind, exactly one of: resource (anything a call returned about a resource: a single resource, a listing, a ranking, a posture card, a Helm release, a permissions check, a neighborhood), logs, events, changes, metrics, metrics:cpu / metrics:memory / metrics:restarts for a diagnose bundle's charts, alerts with name as the rule's name, or issue; any other word leaves the item unplaced). The diagnose bundle is never about one thing: every citation of it needs a subject. " +
 	"ruled_out: at most 5 hypotheses, each pointing at the 0-based evidence item whose result contradicts it, meaning the hypothesis predicts something that result shows to be false (a container that ran does not rule out the wrong image; an empty window does not rule out an earlier change); a hypothesis nothing contradicts is untested and belongs in unresolved, not here. " +
 	"steps: up to 6 actions someone can take, empty on a healthy verdict, each self-contained and copy-pasteable in GitHub-flavored markdown with inline code for commands (a fenced bash block only for real multi-line scripts, fence lines on their own). kind: mitigate changes the cluster to restore service, verify settles an unresolved item, investigate gathers more; these are purposes, not an order. precondition: the condition under which the step is right, empty when there is none. A warning or a thing not to do belongs in the story, not in steps. " +
 	"recommended_index: the one mitigate step Apply should perform, 1-based, safe, targeted, reversible, with no precondition and depending on nothing you inferred or left untested (a cluster type read off a node name is an inference; an intent you could not read is untested); when a step depends on such a thing it carries it as its precondition, the discriminating verify step comes first, and recommended_index is 0. recommended_reason: one clause on why it is the safe pick and, when the fix restores a behaviour the evidence cannot confirm is the wanted one (a default config, an earlier template), a few words saying so; empty when 0. " +
 	"revises_assessment: false on an initial or verification turn. confidence: 0..1. "
 
 const storyGuidance = "STORY: prose an on-call engineer reads top to bottom, after the verdict block. Its job is the why: how the evidence leads to the headline, what else was checked, and what the evidence does not cover. It need not repeat the summary; open with whatever the reader must see first, knowing they will look for the result behind the headline. Length and shape are yours: as short as the case allows, never a bullet list restating results. An on-call reader gives it about a minute, so most cases fit in 150-250 words and only a case with several moving parts needs more; a healthy story stops once present health, the limit of any adverse signal, and what would reopen the question are clear. A healthy story says how long the target has been in this state only when a cited result shows it (a pod's start time, a rollout's completion, an event's age); restart or rollout age alone does not establish continuous health, so otherwise say it is healthy at the time of this check and that how long is unknown. Radar's audit findings are posture, not part of the story unless one bears on the cause. " +
 	"PLACING A CARD: put [[radar:evidence=N]] alone on a line, where N is the 0-based index into evidence (the index, never the ev_ reference: [[radar:evidence-ref=...]] is the ledger's marker on a tool result and must not appear in the story, the summary or root_cause), and Radar renders that result's card there; [[radar:evidence=N|compact]] renders its header and your role label only, not your note; the same marker inside a sentence places the card under that paragraph if nothing else places it, and otherwise is a reference back to the card; prefer the own-line form, which lets you choose where the card lands. A marker inside a code span, code block or blockquote is literal text; the story holds no ```json fence of its own. Two or three cards is typical; 6 is the ceiling, not a target. " +
-	"Only results Radar renders as cards can be placed: diagnose, get_resource, list_resources or list_namespaces with at least one item, get_events, get_pod_logs, get_workload_logs, get_changes, get_neighborhood, get_topology, get_helm_release, get_subject_permissions, get_prometheus_rules, query_prometheus and issues; a search, a top_resources ranking, an empty listing, list_helm_releases, list_packages, get_dashboard, discover_metrics, get_cluster_audit and get_cluster_upgrade_readiness can be cited in evidence but have no card, so never place them. WHEN TO PLACE: your judgement. Place a card where seeing the result changes how the sentence lands: the log line that names the error, the spec field that is wrong, the chart that shows when. Not every claim needs a card. Every result about this resource that Radar captured is listed under Captured results regardless, and every read is in Activity, so cite a result in evidence without placing it when a card would interrupt the reading, or place it compact when the reader needs the fact but not the detail. Close with what would change the answer, or one clause saying nothing open would, unless that is already plain. "
+	"Only results Radar renders as cards can be placed: diagnose, get_resource, list_resources, list_namespaces, list_helm_releases, list_packages or search with at least one row, get_events, get_pod_logs, get_workload_logs, get_changes, get_neighborhood, get_topology, get_helm_release (with its hooks, history and values), get_subject_permissions (with its effective rules), get_prometheus_rules, query_prometheus, discover_metrics, issues, top_resources (a ranking; the workload's own row is marked), and get_cluster_audit or get_cluster_upgrade_readiness (only the findings about this resource or its relations become a card; a scan with none becomes a receipt); an empty listing or get_dashboard can be cited in evidence but has no card, so never place it. WHEN TO PLACE: your judgement. Place a card where seeing the result changes how the sentence lands: the log line that names the error, the spec field that is wrong, the chart that shows when. Not every claim needs a card. Every result about this resource that Radar captured is listed under Captured results regardless, and every read is in Activity, so cite a result in evidence without placing it when a card would interrupt the reading, or place it compact when the reader needs the fact but not the detail. Close with what would change the answer, or one clause saying nothing open would, unless that is already plain. "
 
 // followUpInstruction keeps Findings stable across ordinary questions.
 // Saved runs show agents restating the root cause on most answers; without an
 
@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"log"
+	"sort"
 	"strings"
 
 	"github.com/modelcontextprotocol/go-sdk/mcp"
@@ -131,6 +132,9 @@ func handleGetHelmRelease(ctx context.Context, req *mcp.CallToolRequest, input g
 		result["managedByFluxHelmRelease"] = detail.ManagedByFluxHelmRelease
 	}
 
+	if len(detail.History) > 0 {
+		result["history"] = newestHelmRevisions(detail.History, defaultHelmHistoryLimit)
+	}
 	if len(detail.Hooks) > 0 {
 		result["hooks"] = detail.Hooks
 	}
@@ -333,6 +337,21 @@ func cloneHelmValue(value any) any {
 	}
 }
 
+// A release's stored history can run to hundreds of revisions; the detail
+// read carries the newest few so the Findings card can show the trail, and
+// include=history carries all of it.
+const defaultHelmHistoryLimit = 10
+
+func newestHelmRevisions(history []helm.HelmRevision, limit int) []helm.HelmRevision {
+	if len(history) <= limit {
+		return history
+	}
+	sorted := make([]helm.HelmRevision, len(history))
+	copy(sorted, history)
+	sort.SliceStable(sorted, func(i, j int) bool { return sorted[i].Revision > sorted[j].Revision })
+	return sorted[:limit]
+}
+
 // parseIncludes parses a comma-separated include string into a set.
 func parseIncludes(s string) map[string]bool {
 	result := make(map[string]bool)
 
@@ -84,3 +84,18 @@ func TestRedactedHelmValuesRedactsSecretsAndKeepsReferences(t *testing.T) {
 		t.Fatalf("input mutated: %#v", input["password"])
 	}
 }
+
+func TestNewestHelmRevisionsKeepsTheNewestAndTheOrderOfShortHistories(t *testing.T) {
+	short := []helm.HelmRevision{{Revision: 1}, {Revision: 2}}
+	if got := newestHelmRevisions(short, 10); len(got) != 2 || got[0].Revision != 1 {
+		t.Fatalf("short history changed: %+v", got)
+	}
+	long := make([]helm.HelmRevision, 0, 12)
+	for i := 1; i <= 12; i++ {
+		long = append(long, helm.HelmRevision{Revision: i})
+	}
+	got := newestHelmRevisions(long, 10)
+	if len(got) != 10 || got[0].Revision != 12 || got[9].Revision != 3 {
+		t.Fatalf("long history = %d entries, first %d, last %d; want 10, 12, 3", len(got), got[0].Revision, got[len(got)-1].Revision)
+	}
+}
@@ -1,3 +1,5 @@
+import { StoryExcerpt } from "./StoryExcerpt";
+import { storyExcerpt } from "./investigationEvidence/excerpt";
 import {
   type InvestigationCaseItem,
   investigationCaseObservationKey,
@@ -10,13 +12,7 @@ import {
   useId,
 } from "react";
 import { clsx } from "clsx";
-import {
-  Collapse,
-  CollapseChevron,
-  TerminalBlock,
-  TerminalBlockLabel,
-  stripAnsi,
-} from "@skyhook-io/k8s-ui";
+import { Collapse, CollapseChevron } from "@skyhook-io/k8s-ui";
 import { parseLogLine } from "../../utils/log-format";
 import { evidenceDisplaySnapshot } from "./investigationEvidencePresentation";
 import {
@@ -41,13 +37,10 @@ import {
   uniquePrimarySources,
   phaseLabel,
 } from "./investigationEvidence/cardParts";
-import { VISIBLE_LOG_EVIDENCE_LINES } from "./investigationEvidence/bodies/streams";
 import {
-  InventoryRow,
   listingScopeNamespace,
   namedInventoryRows,
 } from "./investigationEvidence/bodies/inventory";
-import { MetricsBody } from "./investigationEvidence/bodies/metrics";
 import {
   EvidenceBody,
   evidenceHasDetails,
@@ -187,41 +180,28 @@ export function EvidenceCard({
     observation.data,
     observation.summary,
   );
-  const storyLogLines =
-    storyCard && observation.data.type === "logs"
-      ? (observation.data.logs?.lines ?? [])
-          .slice(-VISIBLE_LOG_EVIDENCE_LINES)
-          .map((line) => stripAnsi(line))
-      : [];
-  // A story log card shows the newest two selected lines always — the ones a
-  // story cites from a tail. Expanding appends the earlier lines below them,
-  // under a divider, so nothing above them moves.
-  const storyLogHead = storyLogLines.slice(-2);
-  const storyLogRest = storyLogLines.slice(0, -2);
-  // A story log card carries its cited lines inline, so the only thing left to
-  // expand is the earlier lines; the record keeps the full logs body.
-  const inlineLogs = storyCard && observation.data.type === "logs";
-  // A chart placed as the cause or the symptom is the card: it shows inline,
-  // and the fold has nothing left to add.
-  const inlineMetrics =
-    storyCard &&
-    !compact &&
-    observation.data.type === "metrics" &&
-    cardItems.some((item) => item.role === "cause" || item.role === "symptom");
-  // Each row a citation names is the part of the listing the sentence rests
-  // on; an entry the listing does not hold is often the point, so the card
-  // says so rather than showing another row.
-  const storyInventoryRows =
-    storyCard && !compact && observation.data.type === "inventory"
+  const excerpt = storyCard
+    ? storyExcerpt(observation, cardItems, {
+        compact,
+        scopeNamespace: listingScopeNamespace(observation.source),
+      })
+    : undefined;
+  const citedRows =
+    observation.data.type === "inventory"
       ? namedInventoryRows(
           observation.data.resources,
           cardItems,
           listingScopeNamespace(observation.source),
-        )
-      : [];
+        ).flatMap((named) => (named.row ? [named.row] : []))
+      : undefined;
+  // A card whose content shows inline has nothing left behind the fold but
+  // the earlier lines or the history.
+  const inlineLogs = excerpt?.kind === "lines";
+  const inlineMetrics = excerpt?.kind === "chart";
+  const storyLogRest = excerpt?.kind === "lines" ? excerpt.rest : [];
   const hasExpandableBody =
     (hasEvidenceDetails && !inlineLogs && !inlineMetrics) || meaningfulHistory;
-  const canExpand = !compact && (hasExpandableBody || storyLogLines.length > 2);
+  const canExpand = !compact && (hasExpandableBody || storyLogRest.length > 0);
   const revealHistory = investigationEvidenceShouldRevealHistory(
     group,
     revealSourceId,
@@ -434,80 +414,17 @@ export function EvidenceCard({
               ))}
             </div>
           ) : null}
-          {storyLogHead.length > 0 ? (
-            <div
-              data-story-log-lines
-              className={
-                prominence === "primary" ? "px-3 pb-2.5" : "px-2.5 pb-2"
-              }
-            >
-              <TerminalBlock
-                footer={
-                  storyLogRest.length > 0 ? (
-                    <Collapse open={open && canExpand}>
-                      <TerminalBlockLabel divider>
-                        Earlier lines
-                      </TerminalBlockLabel>
-                      <pre className="overflow-x-auto whitespace-pre-wrap break-words px-3 pb-2.5 font-mono text-xs leading-relaxed text-[var(--terminal-text)]">
-                        {storyLogRest.join("\n")}
-                      </pre>
-                    </Collapse>
-                  ) : null
-                }
-              >
-                {storyLogHead.join("\n")}
-              </TerminalBlock>
-            </div>
-          ) : null}
-          {storyInventoryRows.length > 0 ? (
-            <div
-              data-story-inventory-rows
-              className={
-                prominence === "primary" ? "px-3 pb-2.5" : "px-2.5 pb-2"
-              }
-            >
-              <div className="rounded-md border border-theme-border">
-                {storyInventoryRows.map((entry, index) =>
-                  entry.row ? (
-                    <InventoryRow
-                      key={entry.key}
-                      resource={entry.row}
-                      divider={index > 0}
-                    />
-                  ) : (
-                    <p
-                      key={entry.key}
-                      data-story-inventory-absent={
-                        entry.matches === 0 ? "" : undefined
-                      }
-                      className={clsx(
-                        "px-2.5 py-1.5 font-mono text-xs text-theme-text-secondary",
-                        index > 0 && "border-t border-theme-border/60",
-                      )}
-                    >
-                      {entry.matches === 0
-                        ? `Not in this listing: ${entry.label}`
-                        : `${entry.label}: ${entry.matches} entries match in this listing`}
-                    </p>
-                  ),
-                )}
-              </div>
-            </div>
-          ) : null}
-          {inlineMetrics && observation.data.type === "metrics" ? (
-            <div
-              data-story-metrics
-              className={
-                prominence === "primary" ? "px-3 pb-2.5" : "px-2.5 pb-2"
-              }
-            >
-              <MetricsBody
-                data={observation.data}
-                changeCoverage={metricsMarkersBySource?.get(
-                  observation.source.id,
-                )}
-              />
-            </div>
+          {excerpt ? (
+            <StoryExcerpt
+              excerpt={excerpt}
+              data={observation.data}
+              prominence={prominence}
+              open={open}
+              canExpand={canExpand}
+              changeCoverage={metricsMarkersBySource?.get(
+                observation.source.id,
+              )}
+            />
           ) : null}
         </>
       ) : cardItems.some((item) => item.claim || item.role) ? (
@@ -550,6 +467,7 @@ export function EvidenceCard({
                   changeCoverage={metricsMarkersBySource?.get(
                     observation.source.id,
                   )}
+                  citedRows={citedRows}
                 />
               ) : null}
               {meaningfulHistory ? (