@oneuptime/common 13.0.4 → 13.0.5

package/Utils/AI/InvestigationReport.ts

@@ -0,0 +1,1789 @@
+/*
+ * Pure helpers for reading an AI investigation report.
+ *
+ * The investigation engine posts one markdown blob: a branded heading, the
+ * model's own "Summary / Most likely root cause / Evidence / Suggested next
+ * steps" note, a server-authored "Evidence checked" list and a footer
+ * (AIInvestigationEngine.buildBrandedMarkdown). The dashboard wants to lay
+ * those parts out separately, and the server wants the evidence labels and
+ * the incident/alert numbers the note mentions.
+ *
+ * Everything here is text-only and dependency-free so it loads the same in
+ * the dashboard bundle, the API server and Common jest. None of it turns
+ * model-authored text into links or HTML: callers get plain strings and
+ * numbers, and resolve anything clickable from their own database rows.
+ */
+
+export const AI_ROOT_CAUSE_HEADING_TEXT: string =
+  "AI — Automated Root Cause Analysis";
+
+export enum InvestigationReportSectionKind {
+  Summary = "Summary",
+  RootCause = "RootCause",
+  Evidence = "Evidence",
+  NextSteps = "NextSteps",
+  Other = "Other",
+}
+
+export interface InvestigationReportSection {
+  kind: InvestigationReportSectionKind;
+  // The label as written, e.g. "Most likely root cause".
+  title: string;
+  // Section body, label and its trailing separator removed, trimmed.
+  markdown: string;
+}
+
+export interface InvestigationEvidenceCheckedEntry {
+  // "C1"
+  citationId: string;
+  label: string;
+  rowCount: number;
+}
+
+export interface InvestigationReportFooter {
+  // Footer text without the surrounding * emphasis markers.
+  text: string;
+  modelName?: string | undefined;
+  queryCount?: number | undefined;
+}
+
+export interface ParsedInvestigationReport {
+  /*
+   * True when a Summary or a Most likely root cause section with a body was
+   * recognised — i.e. the report is worth laying out section by section.
+   */
+  isStructured: boolean;
+  /*
+   * Markdown before the first recognised section, brand heading removed,
+   * trimmed ("" when there is none). With no recognised section at all this
+   * is the whole body.
+   */
+  preamble: string;
+  /*
+   * Recognised sections with a non-empty body, in document order. The
+   * Evidence checked block and the footer are NOT included.
+   */
+  sections: Array<InvestigationReportSection>;
+  // Convenience: the first section body of each kind, or undefined.
+  summary?: string | undefined;
+  rootCause?: string | undefined;
+  evidence?: string | undefined;
+  nextSteps?: string | undefined;
+  /*
+   * The server-authored "Evidence checked" list: the LAST such block, unless
+   * the footer's query count shows the server could not have written it.
+   */
+  evidenceChecked: Array<InvestigationEvidenceCheckedEntry>;
+  footer?: InvestigationReportFooter | undefined;
+  /*
+   * The whole report with the brand heading, the Evidence checked block and
+   * the footer removed.
+   */
+  bodyMarkdown: string;
+}
+
+export interface InvestigationEventReferenceToken {
+  // null = no qualifier word; the caller applies a default.
+  kind: "incident" | "alert" | null;
+  number: number;
+  // Index of "#" in the input.
+  start: number;
+  // Index after the last digit.
+  end: number;
+}
+
+export interface ExtractedInvestigationEventReference {
+  kind: "incident" | "alert";
+  number: number;
+}
+
+export const DEFAULT_EVENT_REFERENCE_LIMIT: number = 25;
+
+/*
+ * Citation marker "[C12]". A fresh RegExp per call because the global flag
+ * makes a shared instance stateful (lastIndex) across callers.
+ */
+export function getCitationMarkerRegex(): RegExp {
+  return /\[(C\d{1,3})\]/g;
+}
+
+// "EvidenceChecked" is internal: it is parsed into entries, never a section.
+type LabelKind = InvestigationReportSectionKind | "EvidenceChecked";
+
+/*
+ * Known labels, keyed by their normalised (lower-cased, whitespace-collapsed)
+ * text. A Map rather than an object literal so a heading such as
+ * "constructor" can never resolve to an Object.prototype member.
+ */
+const KNOWN_LABELS: Map<string, LabelKind> = new Map<string, LabelKind>([
+  ["summary", InvestigationReportSectionKind.Summary],
+  ["tl;dr", InvestigationReportSectionKind.Summary],
+  ["tl; dr", InvestigationReportSectionKind.Summary],
+  ["tldr", InvestigationReportSectionKind.Summary],
+  ["most likely root cause", InvestigationReportSectionKind.RootCause],
+  ["likely root cause", InvestigationReportSectionKind.RootCause],
+  ["probable root cause", InvestigationReportSectionKind.RootCause],
+  ["root cause", InvestigationReportSectionKind.RootCause],
+  ["root cause hypothesis", InvestigationReportSectionKind.RootCause],
+  ["evidence", InvestigationReportSectionKind.Evidence],
+  ["key evidence", InvestigationReportSectionKind.Evidence],
+  ["supporting evidence", InvestigationReportSectionKind.Evidence],
+  ["key findings", InvestigationReportSectionKind.Evidence],
+  ["findings", InvestigationReportSectionKind.Evidence],
+  ["suggested next steps", InvestigationReportSectionKind.NextSteps],
+  ["next steps", InvestigationReportSectionKind.NextSteps],
+  ["recommended next steps", InvestigationReportSectionKind.NextSteps],
+  ["recommended actions", InvestigationReportSectionKind.NextSteps],
+  ["recommendations", InvestigationReportSectionKind.NextSteps],
+  ["evidence checked", "EvidenceChecked"],
+]);
+
+const BRAND_HEADING_MARKER: string = "automated root cause analysis";
+const FOOTER_PREFIX: string = "Investigated automatically by OneUptime AI";
+
+const LINE_ENDING_REGEX: RegExp = /\r\n?/g;
+const WHITESPACE_RUN_REGEX: RegExp = /\s+/g;
+const BLANK_LINE_SPLIT_REGEX: RegExp = /\n[ \t]*\n/;
+
+// Fences may sit inside list items, so any indentation opens one.
+const FENCE_OPEN_REGEX: RegExp = /^[ \t]*(`{3,}|~{3,})(.*)$/;
+const FENCE_CLOSE_REGEX: RegExp = /^[ \t]*(`{3,}|~{3,})[ \t]*$/;
+
+// Four spaces or a tab is an indented code block, never a section label.
+const INDENTED_CODE_REGEX: RegExp = /^(?: {4}|\t)/;
+const ATX_HEADING_REGEX: RegExp = /^ {0,3}#{1,6}(?:[ \t]+(.*))?$/;
+const THEMATIC_BREAK_REGEX: RegExp =
+  /^ {0,3}(?:(?:-[ \t]*){3,}|(?:\*[ \t]*){3,}|(?:_[ \t]*){3,})$/;
+
+const LEADING_PARENTHETICAL_REGEX: RegExp = /^[ \t]*(\([^()]{1,80}\))/;
+/*
+ * Longest text worth looking up as a label: the longest known label plus a
+ * decorative prefix, a colon and a "(medium confidence)" style qualifier.
+ */
+const MAX_KNOWN_LABEL_INPUT_LENGTH: number = 160;
+const LEADING_NUMBERING_REGEX: RegExp = /^\d{1,2}[.)][ \t]+/;
+const ENDS_WITH_SENTENCE_PUNCTUATION_REGEX: RegExp = /[.!?]$/;
+// A label needs some text beyond emphasis markers and punctuation ("**_**").
+const HAS_LABEL_TEXT_REGEX: RegExp = /[^\s*_~`#:.—–-]/;
+const CITATION_MARKER_TEST_REGEX: RegExp = /\[C\d{1,3}\]/;
+
+/*
+ * "**Label** rest" / "__Label__ rest". The optional second group lets
+ * "***Label***" (bold italic) unwrap cleanly instead of leaving a stray "*".
+ */
+const STRONG_LEAD_REGEX: RegExp = /^(\*\*|__)(\*|_)?(.+?)\2?\1(.*)$/;
+const SEPARATOR_PREFIX_REGEX: RegExp = /^(?:[:—–]|-{1,2})[ \t]*/;
+// The text before the first ":" of a plain "Summary: body" line.
+const PLAIN_LABEL_REGEX: RegExp = /^[A-Za-z][A-Za-z \t;,()]{0,60}$/;
+
+// Bold-only lines longer than this read as emphasised prose, not a label.
+const MAX_BOLD_LABEL_LENGTH: number = 80;
+
+// Bold and plain labels rank below every ATX heading level (1-6).
+const BOLD_LABEL_RANK: number = 7;
+
+const LIST_ITEM_REGEX: RegExp = /^[ \t]*(?:[-*+]|\d{1,3}[.)])(?:[ \t]+|$)/;
+const LIST_CONTINUATION_REGEX: RegExp = /^[ \t]{2,}\S/;
+const LEADING_WHITESPACE_REGEX: RegExp = /^[ \t]/;
+/*
+ * "- **[C1]** label — 7 row(s)". The label is greedy so a label that itself
+ * contains " — " keeps everything up to the LAST " — N row(s)".
+ */
+const EVIDENCE_CHECKED_ENTRY_REGEX: RegExp =
+  /^[ \t]*(?:[-*+]|\d{1,3}[.)])[ \t]+(?:\*\*|__)?\[(C\d{1,3})\](?:\*\*|__)?(.+)[ \t][—–-][ \t]+(\d+)[ \t]+rows?(?:\(s\))?[ \t]*$/i;
+// The entry regex backtracks over the label; never feed it a runaway line.
+const MAX_EVIDENCE_CHECKED_LINE_LENGTH: number = 2000;
+
+const FOOTER_EMPHASIS_REGEX: RegExp = /^(\*\*|__|\*|_)([\s\S]+)\1$/;
+const FOOTER_USING_REGEX: RegExp = /\busing[ \t]+/;
+const FOOTER_QUERIES_RUN_REGEX: RegExp = /\bquer(?:y|ies)[ \t]+run\b/i;
+const DIGIT_REGEX: RegExp = /\d/;
+
+const EVENT_REFERENCE_PATTERN: string = "#(\\d{1,9})(?!\\w)";
+const WORD_CHARACTER_REGEX: RegExp = /\w/;
+const WHITESPACE_CHARACTER_REGEX: RegExp = /\s/;
+const FORBIDDEN_BEFORE_REFERENCE: string = "&/#";
+/*
+ * Punctuation that may sit between a qualifier word and its "#": "incidents:
+ * #1", "alerts (#2", "alert [#3]" and, where the raw markdown is scanned,
+ * "**Alerts** #4". Bounded so a stray word far back is never read.
+ */
+const QUALIFIER_GAP_PUNCTUATION: string = ":([*_";
+const MAX_QUALIFIER_GAP_PUNCTUATION: number = 4;
+/*
+ * Words that put a "#N" in some other numbering: a scheduled maintenance
+ * event, an episode, a pull request, a build, a list position. The number is
+ * not an incident or an alert, so it must not fall back to the subject's
+ * kind and link to whichever record happens to share it.
+ */
+const NON_EVENT_QUALIFIER_WORDS: Set<string> = new Set<string>([
+  "maintenance",
+  "maintenances",
+  "episode",
+  "episodes",
+  "pr",
+  "prs",
+  "mr",
+  "mrs",
+  "pull",
+  "request",
+  "requests",
+  "issue",
+  "issues",
+  "ticket",
+  "tickets",
+  "bug",
+  "bugs",
+  "commit",
+  "commits",
+  "build",
+  "builds",
+  "deploy",
+  "deploys",
+  "deployment",
+  "deployments",
+  "release",
+  "releases",
+  "version",
+  "versions",
+  "run",
+  "runs",
+  "job",
+  "jobs",
+  "pipeline",
+  "pipelines",
+  "step",
+  "steps",
+  "attempt",
+  "attempts",
+  "retry",
+  "retries",
+  "item",
+  "items",
+  "line",
+  "lines",
+  "row",
+  "rows",
+  "port",
+  "ports",
+  "pod",
+  "pods",
+  "replica",
+  "replicas",
+  "node",
+  "nodes",
+  "shard",
+  "shards",
+  "worker",
+  "workers",
+  "instance",
+  "instances",
+  "page",
+  "pages",
+  "phase",
+  "phases",
+  "stage",
+  "stages",
+  "option",
+  "options",
+  "rank",
+  "priority",
+  "no",
+  "number",
+  "monitor",
+  "monitors",
+]);
+// Longer than any qualifier or non-event word: never worth lower-casing.
+const MAX_QUALIFIER_WORD_LENGTH: number = 16;
+const SIBLING_SEPARATOR_REGEX: RegExp =
+  /^(?:[ \t\n]*,[ \t\n]*(?:(?:and|or)[ \t\n]+)?|[ \t\n]+(?:and|or)[ \t\n]+)$/i;
+// ", and " with generous whitespace; a longer gap is never a list separator.
+const MAX_SIBLING_SEPARATOR_LENGTH: number = 16;
+
+/*
+ * Stands in for an inline code span while scanning for references: not
+ * whitespace (so a qualifier cannot reach across it), not a word character
+ * and not one of the characters that block a reference.
+ */
+const CODE_SPAN_PLACEHOLDER: string = String.fromCharCode(0xfffc);
+
+interface LabelLineMatch {
+  // null: not a known label (it may still open an "Other" section).
+  kind: LabelKind | null;
+  title: string;
+  inlineBody: string;
+  /*
+   * A heading, or a bold-only line that reads like a label. Only these may
+   * open an "Other" section or an Evidence checked block.
+   */
+  isBare: boolean;
+  // The ATX heading level (1-6); 0 or undefined for a bold or plain label.
+  headingLevel?: number | undefined;
+}
+
+interface SectionStart {
+  index: number;
+  kind: InvestigationReportSectionKind;
+  title: string;
+  inlineBody: string;
+  headingLevel: number;
+  /*
+   * Set when the section's label turned out to be a group heading: it had no
+   * body of its own and was directly followed by another label. The depth
+   * rank of that first sub-label (see getLabelRank); later labels at that
+   * rank or deeper stay inside the section too.
+   */
+  subLabelRank?: number | undefined;
+}
+
+interface SectionCandidate {
+  index: number;
+  match: LabelLineMatch;
+}
+
+interface FooterMatch {
+  breakIndex: number;
+  footer: InvestigationReportFooter;
+}
+
+function normaliseLineEndings(markdown: string): string {
+  if (typeof markdown !== "string") {
+    return "";
+  }
+
+  return markdown.replace(LINE_ENDING_REGEX, "\n");
+}
+
+/*
+ * Marks every line that belongs to a fenced code block, fence lines
+ * included. Nothing inside a fence may start a section, a footer or an
+ * evidence block. An unclosed fence runs to the end of the document, as it
+ * does when the markdown is rendered.
+ */
+function markFencedLines(lines: Array<string>): Array<boolean> {
+  const fenced: Array<boolean> = [];
+  let openFence: string | null = null;
+
+  for (const line of lines) {
+    if (openFence === null) {
+      const open: RegExpExecArray | null = FENCE_OPEN_REGEX.exec(line);
+      const marker: string = open?.[1] || "";
+      const info: string = open?.[2] || "";
+
+      // A backtick fence's info string cannot contain a backtick ("```a```" is inline code).
+      if (marker && !(marker.charAt(0) === "`" && info.includes("`"))) {
+        openFence = marker;
+        fenced.push(true);
+        continue;
+      }
+
+      fenced.push(false);
+      continue;
+    }
+
+    fenced.push(true);
+
+    const close: RegExpExecArray | null = FENCE_CLOSE_REGEX.exec(line);
+    const closeMarker: string = close?.[1] || "";
+
+    if (
+      closeMarker &&
+      closeMarker.charAt(0) === openFence.charAt(0) &&
+      closeMarker.length >= openFence.length
+    ) {
+      openFence = null;
+    }
+  }
+
+  return fenced;
+}
+
+// Emoji and pictographic symbols a model likes to put in front of a heading.
+function isDecorativeCodePoint(codePoint: number): boolean {
+  return (
+    (codePoint >= 0x2000 && codePoint <= 0x2bff) ||
+    codePoint >= 0x1f000 ||
+    codePoint === 0xfe0f ||
+    codePoint === 0x200d
+  );
+}
+
+function stripLeadingDecoration(label: string): string {
+  let cursor: number = 0;
+
+  while (cursor < label.length) {
+    const codePoint: number = label.codePointAt(cursor) || 0;
+    const character: string = String.fromCodePoint(codePoint);
+
+    if (
+      !WHITESPACE_CHARACTER_REGEX.test(character) &&
+      !isDecorativeCodePoint(codePoint)
+    ) {
+      break;
+    }
+
+    cursor += character.length;
+  }
+
+  return label.slice(cursor).replace(LEADING_NUMBERING_REGEX, "");
+}
+
+/*
+ * The trailing-edge helpers below are plain string scans on purpose. The
+ * obvious regexes (/[ \t]*:[ \t]*$/, /(?:^|[ \t]+)#+[ \t]*$/ ...) backtrack
+ * quadratically over a long run of spaces, and this text is model-authored:
+ * one 20 KB line froze the parser for a minute when they were regexes.
+ */
+function endsWithColon(value: string): boolean {
+  return value.trimEnd().endsWith(":");
+}
+
+function stripTrailingColon(value: string): string {
+  const trimmed: string = value.trimEnd();
+
+  return trimmed.endsWith(":") ? trimmed.slice(0, -1).trimEnd() : trimmed;
+}
+
+// "Most likely root cause (medium confidence)" -> "Most likely root cause".
+function stripTrailingParenthetical(value: string): string {
+  const trimmed: string = value.trimEnd();
+
+  if (!trimmed.endsWith(")")) {
+    return trimmed;
+  }
+
+  const open: number = trimmed.lastIndexOf("(");
+
+  if (open === -1 || trimmed.slice(open + 1, -1).includes(")")) {
+    return trimmed;
+  }
+
+  return trimmed.slice(0, open).trimEnd();
+}
+
+// "## Summary ##" -> "Summary"; "## C#" keeps its hash.
+function stripAtxClosingSequence(value: string): string {
+  const trimmed: string = value.trimEnd();
+  let end: number = trimmed.length;
+
+  while (end > 0 && trimmed.charAt(end - 1) === "#") {
+    end--;
+  }
+
+  if (end === trimmed.length) {
+    return trimmed;
+  }
+
+  if (end === 0) {
+    return "";
+  }
+
+  const before: string = trimmed.charAt(end - 1);
+
+  return before === " " || before === "\t"
+    ? trimmed.slice(0, end).trimEnd()
+    : trimmed;
+}
+
+/*
+ * Splits "Summary — body" / "Summary – body" / "Summary - body" at the first
+ * dash separator. A hyphen only separates with whitespace on both sides, so
+ * "on-call" never splits.
+ */
+function splitAtFirstDash(
+  value: string,
+): { label: string; rest: string } | null {
+  for (let index: number = 0; index < value.length; index++) {
+    const character: string = value.charAt(index);
+
+    if (character === "—" || character === "–") {
+      return { label: value.slice(0, index), rest: value.slice(index + 1) };
+    }
+
+    if (character !== "-" || index === 0) {
+      continue;
+    }
+
+    const previous: string = value.charAt(index - 1);
+    let after: number = index + 1;
+
+    if (value.charAt(after) === "-") {
+      after++;
+    }
+
+    const next: string = value.charAt(after);
+
+    if (
+      (previous === " " || previous === "\t") &&
+      (next === " " || next === "\t")
+    ) {
+      return { label: value.slice(0, index), rest: value.slice(after) };
+    }
+  }
+
+  return null;
+}
+
+/*
+ * Removes wrapping emphasis and a trailing colon, inside or outside the
+ * emphasis: "**Summary:**", "**Summary**:", "__Summary__", "***Summary***".
+ */
+function cleanLabel(raw: string): string {
+  let label: string = stripTrailingColon(raw.trim());
+
+  for (const delimiter of ["**", "__", "*", "_"]) {
+    const isWrapped: boolean =
+      label.length > delimiter.length * 2 &&
+      label.startsWith(delimiter) &&
+      label.endsWith(delimiter) &&
+      !label
+        .slice(delimiter.length, label.length - delimiter.length)
+        .includes(delimiter);
+
+    if (isWrapped) {
+      label = stripTrailingColon(
+        label.slice(delimiter.length, label.length - delimiter.length).trim(),
+      );
+    }
+  }
+
+  return label;
+}
+
+function tidyTitle(label: string): string {
+  const collapsed: string = label.replace(WHITESPACE_RUN_REGEX, " ").trim();
+  const stripped: string = stripLeadingDecoration(collapsed).trim();
+
+  return stripped || collapsed;
+}
+
+function lookupLabel(label: string): LabelKind | null {
+  // Every known label is short; anything longer is prose.
+  if (label.length > MAX_KNOWN_LABEL_INPUT_LENGTH) {
+    return null;
+  }
+
+  const key: string = stripTrailingColon(
+    stripTrailingParenthetical(stripLeadingDecoration(label)),
+  )
+    .replace(WHITESPACE_RUN_REGEX, " ")
+    .trim()
+    .toLowerCase();
+
+  return KNOWN_LABELS.get(key) || null;
+}
+
+function looksLikeBoldLabel(title: string): boolean {
+  return (
+    HAS_LABEL_TEXT_REGEX.test(title) &&
+    title.length <= MAX_BOLD_LABEL_LENGTH &&
+    !ENDS_WITH_SENTENCE_PUNCTUATION_REGEX.test(title) &&
+    !CITATION_MARKER_TEST_REGEX.test(title)
+  );
+}
+
+/*
+ * "**Label**", "**Label:**", "**Label** — body", "**Label:** body",
+ * "**Label** (medium confidence) — body". A bold phrase followed by prose
+ * with no separator ("**Root cause** is unclear") is not a label line.
+ */
+function matchStrongLead(content: string): LabelLineMatch | null {
+  const lead: RegExpExecArray | null = STRONG_LEAD_REGEX.exec(content);
+
+  if (!lead) {
+    return null;
+  }
+
+  const inner: string = lead[3] || "";
+  let rest: string = lead[4] || "";
+  const label: string = cleanLabel(inner);
+
+  if (!label) {
+    return null;
+  }
+
+  const innerEndsWithColon: boolean = endsWithColon(inner);
+  let qualifier: string = "";
+  const parenthetical: RegExpExecArray | null =
+    LEADING_PARENTHETICAL_REGEX.exec(rest);
+
+  if (parenthetical) {
+    qualifier = parenthetical[1] || "";
+    rest = rest.slice(parenthetical[0].length);
+  }
+
+  const restTrimmed: string = rest.trim();
+  let inlineBody: string = "";
+  let isBare: boolean = false;
+
+  if (restTrimmed === "" || restTrimmed === ":") {
+    isBare = qualifier === "";
+  } else {
+    const separator: RegExpExecArray | null =
+      SEPARATOR_PREFIX_REGEX.exec(restTrimmed);
+
+    if (separator) {
+      inlineBody = restTrimmed.slice(separator[0].length).trim();
+    } else if (innerEndsWithColon) {
+      // "**Root cause:** body" and "**Root cause:** (medium confidence) body".
+      inlineBody = restTrimmed;
+    } else {
+      return null;
+    }
+  }
+
+  const title: string = tidyTitle(qualifier ? `${label} ${qualifier}` : label);
+  const kind: LabelKind | null = lookupLabel(label);
+
+  isBare = isBare && looksLikeBoldLabel(title);
+
+  if (kind === null && !isBare) {
+    return null;
+  }
+
+  return { kind, title, inlineBody, isBare };
+}
+
+// "Summary: body" — only for known labels.
+function matchPlainLabel(content: string): LabelLineMatch | null {
+  const colon: number = content.indexOf(":");
+
+  if (colon <= 0) {
+    return null;
+  }
+
+  const label: string = content.slice(0, colon).trim();
+
+  if (!PLAIN_LABEL_REGEX.test(label)) {
+    return null;
+  }
+
+  const kind: LabelKind | null = lookupLabel(label);
+
+  if (kind === null) {
+    return null;
+  }
+
+  const inlineBody: string = content.slice(colon + 1).trim();
+
+  return {
+    kind,
+    title: tidyTitle(label),
+    inlineBody,
+    isBare: inlineBody === "",
+  };
+}
+
+function matchHeading(headingText: string): LabelLineMatch | null {
+  const text: string = stripAtxClosingSequence(headingText).trim();
+
+  if (!text) {
+    return null;
+  }
+
+  const full: string = cleanLabel(text);
+  const fullKind: LabelKind | null = lookupLabel(full);
+
+  if (fullKind !== null) {
+    return {
+      kind: fullKind,
+      title: tidyTitle(full),
+      inlineBody: "",
+      isBare: true,
+    };
+  }
+
+  // "## Summary: body", "## **Summary** — body", "## Summary — body".
+  const lead: LabelLineMatch | null =
+    text.startsWith("**") || text.startsWith("__")
+      ? matchStrongLead(text)
+      : matchPlainLabel(text);
+
+  if (lead && lead.kind !== null) {
+    return { ...lead, isBare: lead.inlineBody === "" };
+  }
+
+  const dash: { label: string; rest: string } | null = splitAtFirstDash(text);
+
+  if (dash) {
+    const dashLabel: string = cleanLabel(dash.label);
+    const dashKind: LabelKind | null = lookupLabel(dashLabel);
+    const dashBody: string = dash.rest.trim();
+
+    if (dashKind !== null) {
+      return {
+        kind: dashKind,
+        title: tidyTitle(dashLabel),
+        inlineBody: dashBody,
+        isBare: dashBody === "",
+      };
+    }
+  }
+
+  // Any other heading may open an "Other" section, titled as written.
+  return {
+    kind: null,
+    title: tidyTitle(full || text),
+    inlineBody: "",
+    isBare: true,
+  };
+}
+
+// "### Evidence" -> 3. Only called on a line ATX_HEADING_REGEX accepted.
+function countLeadingHashes(line: string): number {
+  const content: string = line.trimStart();
+  let level: number = 0;
+
+  while (level < content.length && content.charAt(level) === "#") {
+    level++;
+  }
+
+  return level;
+}
+
+/*
+ * How deep a label sits in the outline: its heading level, with bold and
+ * plain labels below every heading.
+ */
+function getLabelRank(match: LabelLineMatch): number {
+  return match.headingLevel || BOLD_LABEL_RANK;
+}
+
+function matchSectionStart(line: string): LabelLineMatch | null {
+  // "*****" is a thematic break, not an empty bold label.
+  if (INDENTED_CODE_REGEX.test(line) || THEMATIC_BREAK_REGEX.test(line)) {
+    return null;
+  }
+
+  const heading: RegExpExecArray | null = ATX_HEADING_REGEX.exec(line);
+
+  if (heading) {
+    const headingMatch: LabelLineMatch | null = matchHeading(heading[1] || "");
+
+    return headingMatch
+      ? { ...headingMatch, headingLevel: countLeadingHashes(line) }
+      : null;
+  }
+
+  const content: string = line.trim();
+
+  if (!content) {
+    return null;
+  }
+
+  if (content.startsWith("**") || content.startsWith("__")) {
+    return matchStrongLead(content);
+  }
+
+  return matchPlainLabel(content);
+}
+
+function isBrandHeading(line: string): boolean {
+  const heading: RegExpExecArray | null = ATX_HEADING_REGEX.exec(line);
+
+  return Boolean(
+    heading && (heading[1] || "").toLowerCase().includes(BRAND_HEADING_MARKER),
+  );
+}
+
+/*
+ * The footer is the LAST thematic break (outside code) followed by nothing
+ * but one emphasis paragraph starting with the OneUptime AI footer text.
+ */
+function findFooter(
+  lines: Array<string>,
+  fenced: Array<boolean>,
+): FooterMatch | null {
+  let breakIndex: number = -1;
+
+  for (let index: number = lines.length - 1; index >= 0; index--) {
+    if (!fenced[index] && THEMATIC_BREAK_REGEX.test(lines[index] || "")) {
+      breakIndex = index;
+      break;
+    }
+  }
+
+  if (breakIndex < 0) {
+    return null;
+  }
+
+  const trailing: string = lines
+    .slice(breakIndex + 1)
+    .join("\n")
+    .trim();
+
+  if (!trailing || BLANK_LINE_SPLIT_REGEX.test(trailing)) {
+    return null;
+  }
+
+  const paragraph: string = trailing
+    .split("\n")
+    .map((line: string): string => {
+      return line.trim();
+    })
+    .join(" ");
+
+  const emphasis: RegExpExecArray | null =
+    FOOTER_EMPHASIS_REGEX.exec(paragraph);
+  const text: string = (emphasis?.[2] || "").trim();
+
+  if (!text.startsWith(FOOTER_PREFIX)) {
+    return null;
+  }
+
+  const footer: InvestigationReportFooter = { text };
+
+  const modelName: string | undefined = readFooterModelName(text);
+
+  if (modelName) {
+    footer.modelName = modelName;
+  }
+
+  const queryCount: number | undefined = readFooterQueryCount(text);
+
+  if (queryCount !== undefined) {
+    footer.queryCount = queryCount;
+  }
+
+  return { breakIndex, footer };
+}
+
+/*
+ * "... telemetry using gpt-4.1-mini. This is ..." -> "gpt-4.1-mini". The
+ * name ends at the first "." followed by whitespace or the end, so dots
+ * inside a model name survive. A scan rather than a lazy regex, which
+ * rescans the rest of the text for every "using".
+ */
+function readFooterModelName(text: string): string | undefined {
+  const using: RegExpExecArray | null = FOOTER_USING_REGEX.exec(text);
+
+  if (!using) {
+    return undefined;
+  }
+
+  const start: number = using.index + using[0].length;
+
+  for (let index: number = start; index < text.length; index++) {
+    if (
+      text.charAt(index) === "." &&
+      (index + 1 === text.length ||
+        WHITESPACE_CHARACTER_REGEX.test(text.charAt(index + 1)))
+    ) {
+      return text.slice(start, index).trim() || undefined;
+    }
+  }
+
+  return undefined;
+}
+
+// "... 7 queries run ..." / "... 1 query run ..." -> 7 / 1.
+function readFooterQueryCount(text: string): number | undefined {
+  const queries: RegExpExecArray | null = FOOTER_QUERIES_RUN_REGEX.exec(text);
+
+  if (!queries) {
+    return undefined;
+  }
+
+  let cursor: number = queries.index - 1;
+  const spaceEnd: number = cursor;
+
+  while (
+    cursor >= 0 &&
+    (text.charAt(cursor) === " " || text.charAt(cursor) === "\t")
+  ) {
+    cursor--;
+  }
+
+  if (cursor === spaceEnd) {
+    return undefined;
+  }
+
+  const digitsEnd: number = cursor + 1;
+
+  while (cursor >= 0 && DIGIT_REGEX.test(text.charAt(cursor))) {
+    cursor--;
+  }
+
+  const digits: string = text.slice(cursor + 1, digitsEnd);
+
+  // The engine's tool-call budget is tiny; a huge number is not a count.
+  if (!digits || digits.length > 9) {
+    return undefined;
+  }
+
+  return parseInt(digits, 10);
+}
+
+/*
+ * The lines an "Evidence checked" block owns: its label line, then list
+ * items (with their indented continuations) and the blank lines between
+ * them. The first other line ends the block, so prose that follows it is
+ * never swallowed.
+ */
+function findEvidenceCheckedBlockEnd(
+  lines: Array<string>,
+  fenced: Array<boolean>,
+  labelIndex: number,
+  contentEnd: number,
+): number {
+  let end: number = labelIndex + 1;
+  let cursor: number = labelIndex + 1;
+  let insideList: boolean = false;
+
+  while (cursor < contentEnd) {
+    const line: string = lines[cursor] || "";
+
+    if (fenced[cursor]) {
+      break;
+    }
+
+    if (!line.trim()) {
+      cursor++;
+      continue;
+    }
+
+    if (LIST_ITEM_REGEX.test(line)) {
+      insideList = true;
+    } else if (!(insideList && LIST_CONTINUATION_REGEX.test(line))) {
+      break;
+    }
+
+    cursor++;
+    end = cursor;
+  }
+
+  return end;
+}
+
+function parseEvidenceCheckedEntries(
+  lines: Array<string>,
+): Array<InvestigationEvidenceCheckedEntry> {
+  const entries: Array<InvestigationEvidenceCheckedEntry> = [];
+  const seen: Set<string> = new Set<string>();
+
+  for (const line of lines) {
+    if (line.length > MAX_EVIDENCE_CHECKED_LINE_LENGTH) {
+      continue;
+    }
+
+    const entry: RegExpExecArray | null =
+      EVIDENCE_CHECKED_ENTRY_REGEX.exec(line);
+
+    if (!entry) {
+      continue;
+    }
+
+    const citationId: string = (entry[1] || "").toUpperCase();
+    const label: string = (entry[2] || "").trim();
+
+    if (!citationId || !label || seen.has(citationId)) {
+      continue;
+    }
+
+    seen.add(citationId);
+    entries.push({
+      citationId,
+      label,
+      rowCount: parseInt(entry[3] || "0", 10),
+    });
+  }
+
+  return entries;
+}
+
+function createEmptyReport(): ParsedInvestigationReport {
+  return {
+    isStructured: false,
+    preamble: "",
+    sections: [],
+    summary: undefined,
+    rootCause: undefined,
+    evidence: undefined,
+    nextSteps: undefined,
+    evidenceChecked: [],
+    footer: undefined,
+    bodyMarkdown: "",
+  };
+}
+
+/*
+ * Joins the kept lines of [from, to). Untrimmed on purpose: a section body
+ * is appended to its inline first line, and the blank line between them is
+ * what keeps them separate paragraphs.
+ */
+function joinKeptLines(
+  lines: Array<string>,
+  kept: Array<boolean>,
+  from: number,
+  to: number,
+): string {
+  const output: Array<string> = [];
+  let lastWasRemoved: boolean = false;
+
+  for (let index: number = from; index < to; index++) {
+    if (!kept[index]) {
+      lastWasRemoved = true;
+      continue;
+    }
+
+    const line: string = lines[index] || "";
+    const previous: string | undefined = output[output.length - 1];
+
+    /*
+     * Keep a blank line where a block was cut out, so the text on either
+     * side cannot merge into one paragraph.
+     */
+    if (
+      lastWasRemoved &&
+      line.trim() !== "" &&
+      (previous === undefined || previous.trim() !== "")
+    ) {
+      output.push("");
+    }
+
+    lastWasRemoved = false;
+    output.push(line);
+  }
+
+  return output.join("\n");
+}
+
+/*
+ * Every kept line outside code that reads like a section label, in order.
+ * An indented line inside a list continues a list item ("1. Roll back", then
+ * "   Evidence: pool at 100%"), so it never starts a section: that would cut
+ * the list in two and could hand the rest of it to another section kind.
+ */
+function findSectionCandidates(
+  lines: Array<string>,
+  fenced: Array<boolean>,
+  kept: Array<boolean>,
+  contentEnd: number,
+): Array<SectionCandidate> {
+  const candidates: Array<SectionCandidate> = [];
+  let insideList: boolean = false;
+
+  for (let index: number = 0; index < contentEnd; index++) {
+    const line: string = lines[index] || "";
+
+    // Blank lines keep the list context: a list item may span paragraphs.
+    if (!kept[index] || fenced[index] || !line.trim()) {
+      continue;
+    }
+
+    const isIndented: boolean = LEADING_WHITESPACE_REGEX.test(line);
+
+    if (THEMATIC_BREAK_REGEX.test(line)) {
+      insideList = false;
+    } else if (LIST_ITEM_REGEX.test(line)) {
+      insideList = true;
+    } else if (!isIndented) {
+      insideList = false;
+    } else if (insideList) {
+      continue;
+    }
+
+    const match: LabelLineMatch | null = matchSectionStart(line);
+
+    if (match) {
+      candidates.push({ index, match });
+    }
+  }
+
+  return candidates;
+}
+
+/*
+ * Turns label candidates into the section starts the report is cut at.
+ *
+ * Recognised labels always start a section (one with no body is dropped
+ * later, as before). Unknown headings and bold-only labels, including a
+ * model-authored "Evidence checked", start an "Other" section only once the
+ * structured part of the note has begun, and only when they are not part of
+ * the section already open:
+ *
+ *   - a heading deeper than the recognised heading that opened the section
+ *     ("## Evidence" then "### Metrics") belongs to it;
+ *   - a label directly under a label with no body of its own makes that
+ *     label a group heading ("## Most likely root cause" then "### Pool
+ *     exhaustion"), and the group's later labels at the same depth or deeper
+ *     stay in it too;
+ *   - a label with nothing under it ("**Confidence: medium**" at the end) is
+ *     a line of the open section, not an empty section of its own.
+ *
+ * So a section title is never lost because its body opens with a sub-label,
+ * and no content is ever dropped.
+ */
+function resolveSectionStarts(
+  lines: Array<string>,
+  kept: Array<boolean>,
+  allCandidates: Array<SectionCandidate>,
+  contentEnd: number,
+): Array<SectionStart> {
+  /*
+   * Kept lines with text before each index, so "is there any content in
+   * [from, to)" is O(1) however many labels a section swallows.
+   */
+  const contentBefore: Array<number> = [0];
+
+  for (let index: number = 0; index < lines.length; index++) {
+    const hasText: boolean =
+      kept[index] === true && (lines[index] || "").trim() !== "";
+    contentBefore.push((contentBefore[index] || 0) + (hasText ? 1 : 0));
+  }
+
+  const hasKeptContent: (from: number, to: number) => boolean = (
+    from: number,
+    to: number,
+  ): boolean => {
+    return (
+      to > from && (contentBefore[to] || 0) - (contentBefore[from] || 0) > 0
+    );
+  };
+
+  const isRecognised: (match: LabelLineMatch) => boolean = (
+    match: LabelLineMatch,
+  ): boolean => {
+    return match.kind !== null && match.kind !== "EvidenceChecked";
+  };
+
+  /*
+   * A heading at the same or a shallower level than the heading before it is
+   * its peer, never its sub-label ("## Impact" then "## Timeline").
+   */
+  const isPeerHeading: (openLevel: number, match: LabelLineMatch) => boolean = (
+    openLevel: number,
+    match: LabelLineMatch,
+  ): boolean => {
+    const level: number = match.headingLevel || 0;
+
+    return openLevel > 0 && level > 0 && level <= openLevel;
+  };
+
+  // A label with inline text ("Evidence checked: logs") is just a line of prose.
+  const candidates: Array<SectionCandidate> = allCandidates.filter(
+    (candidate: SectionCandidate): boolean => {
+      return (
+        kept[candidate.index] === true &&
+        (isRecognised(candidate.match) || candidate.match.isBare)
+      );
+    },
+  );
+
+  /*
+   * Whether an "Other" label at this position has a body: text before the
+   * next label, or a next label that would become its sub-label and itself
+   * has a body. Walked backwards so a run of labels costs linear time.
+   */
+  const hasBody: Array<boolean> = candidates.map((): boolean => {
+    return false;
+  });
+
+  for (
+    let position: number = candidates.length - 1;
+    position >= 0;
+    position--
+  ) {
+    const candidate: SectionCandidate | undefined = candidates[position];
+    const next: SectionCandidate | undefined = candidates[position + 1];
+
+    if (!candidate) {
+      continue;
+    }
+
+    hasBody[position] =
+      hasKeptContent(candidate.index + 1, next ? next.index : contentEnd) ||
+      Boolean(
+        next &&
+          !isRecognised(next.match) &&
+          !isPeerHeading(candidate.match.headingLevel || 0, next.match) &&
+          hasBody[position + 1],
+      );
+  }
+
+  const starts: Array<SectionStart> = [];
+
+  candidates.forEach((candidate: SectionCandidate, position: number): void => {
+    const match: LabelLineMatch = candidate.match;
+    const headingLevel: number = match.headingLevel || 0;
+
+    if (match.kind !== null && match.kind !== "EvidenceChecked") {
+      starts.push({
+        index: candidate.index,
+        kind: match.kind,
+        title: match.title,
+        inlineBody: match.inlineBody,
+        headingLevel,
+      });
+      return;
+    }
+
+    const open: SectionStart | undefined = starts[starts.length - 1];
+
+    // Before the first recognised section, labels are preamble content.
+    if (!open) {
+      return;
+    }
+
+    const rank: number = getLabelRank(match);
+
+    // A sub-heading of the recognised heading that opened the section.
+    if (
+      open.kind !== InvestigationReportSectionKind.Other &&
+      open.headingLevel > 0 &&
+      headingLevel > open.headingLevel
+    ) {
+      return;
+    }
+
+    // The first label under a label with no body: that label is a group heading.
+    if (
+      open.inlineBody === "" &&
+      !hasKeptContent(open.index + 1, candidate.index) &&
+      !isPeerHeading(open.headingLevel, match)
+    ) {
+      open.subLabelRank = rank;
+      return;
+    }
+
+    // The group's later labels, at its sub-labels' depth or deeper.
+    if (open.subLabelRank !== undefined && rank >= open.subLabelRank) {
+      return;
+    }
+
+    // A label with nothing under it is a line of the open section.
+    if (!hasBody[position]) {
+      return;
+    }
+
+    starts.push({
+      index: candidate.index,
+      kind: InvestigationReportSectionKind.Other,
+      title: match.title,
+      inlineBody: "",
+      headingLevel,
+    });
+  });
+
+  return starts;
+}
+
+export function parseInvestigationReport(
+  markdown: string,
+): ParsedInvestigationReport {
+  const normalised: string = normaliseLineEndings(markdown);
+
+  if (!normalised.trim()) {
+    return createEmptyReport();
+  }
+
+  const lines: Array<string> = normalised.split("\n");
+  const fenced: Array<boolean> = markFencedLines(lines);
+  const kept: Array<boolean> = lines.map((): boolean => {
+    return true;
+  });
+
+  // 1. The branded heading, when it is the first thing in the document.
+  const firstContentIndex: number = lines.findIndex((line: string): boolean => {
+    return line.trim() !== "";
+  });
+
+  if (
+    firstContentIndex >= 0 &&
+    !fenced[firstContentIndex] &&
+    isBrandHeading(lines[firstContentIndex] || "")
+  ) {
+    kept[firstContentIndex] = false;
+  }
+
+  // 2. The footer.
+  const footerMatch: FooterMatch | null = findFooter(lines, fenced);
+  const contentEnd: number = footerMatch
+    ? footerMatch.breakIndex
+    : lines.length;
+
+  for (let index: number = contentEnd; index < lines.length; index++) {
+    kept[index] = false;
+  }
+
+  // 3. Every line that could start a section.
+  const candidates: Array<SectionCandidate> = findSectionCandidates(
+    lines,
+    fenced,
+    kept,
+    contentEnd,
+  );
+
+  /*
+   * 4. Only the LAST bare "Evidence checked" label can be the server's list.
+   * Earlier ones were written by the model and stay ordinary content.
+   */
+  let evidenceChecked: Array<InvestigationEvidenceCheckedEntry> = [];
+  let evidenceCheckedIndex: number = -1;
+
+  for (const candidate of candidates) {
+    if (candidate.match.kind === "EvidenceChecked" && candidate.match.isBare) {
+      evidenceCheckedIndex = candidate.index;
+    }
+  }
+
+  if (evidenceCheckedIndex >= 0) {
+    const blockEnd: number = findEvidenceCheckedBlockEnd(
+      lines,
+      fenced,
+      evidenceCheckedIndex,
+      contentEnd,
+    );
+    const entries: Array<InvestigationEvidenceCheckedEntry> =
+      parseEvidenceCheckedEntries(
+        lines.slice(evidenceCheckedIndex + 1, blockEnd),
+      );
+    const queryCount: number | undefined = footerMatch?.footer.queryCount;
+
+    /*
+     * The server writes one entry per citation, at most one citation per
+     * query, and no block at all when there are no citations. A footer
+     * saying "0 queries run", or fewer queries than entries, therefore means
+     * the model wrote this block: it stays in the body as model content.
+     * Without a readable count (no footer, an older format) it is trusted.
+     */
+    const isServerBlock: boolean =
+      queryCount === undefined ||
+      (queryCount > 0 && entries.length <= queryCount);
+
+    if (isServerBlock) {
+      evidenceChecked = entries;
+
+      for (
+        let index: number = evidenceCheckedIndex;
+        index < blockEnd;
+        index++
+      ) {
+        kept[index] = false;
+      }
+    }
+  }
+
+  // 5. Resolve candidates into section starts.
+  const starts: Array<SectionStart> = resolveSectionStarts(
+    lines,
+    kept,
+    candidates,
+    contentEnd,
+  );
+
+  const firstStart: SectionStart | undefined = starts[0];
+  const preamble: string = joinKeptLines(
+    lines,
+    kept,
+    0,
+    firstStart ? firstStart.index : contentEnd,
+  ).trim();
+
+  const sections: Array<InvestigationReportSection> = [];
+
+  starts.forEach((start: SectionStart, position: number): void => {
+    const next: SectionStart | undefined = starts[position + 1];
+    const body: string = joinKeptLines(
+      lines,
+      kept,
+      start.index + 1,
+      next ? next.index : contentEnd,
+    );
+    const sectionMarkdown: string = `${start.inlineBody}\n${body}`.trim();
+
+    if (!sectionMarkdown) {
+      return;
+    }
+
+    sections.push({
+      kind: start.kind,
+      title: start.title,
+      markdown: sectionMarkdown,
+    });
+  });
+
+  const firstOfKind: (
+    kind: InvestigationReportSectionKind,
+  ) => string | undefined = (
+    kind: InvestigationReportSectionKind,
+  ): string | undefined => {
+    return sections.find((section: InvestigationReportSection): boolean => {
+      return section.kind === kind;
+    })?.markdown;
+  };
+
+  const summary: string | undefined = firstOfKind(
+    InvestigationReportSectionKind.Summary,
+  );
+  const rootCause: string | undefined = firstOfKind(
+    InvestigationReportSectionKind.RootCause,
+  );
+
+  return {
+    isStructured: summary !== undefined || rootCause !== undefined,
+    preamble,
+    sections,
+    summary,
+    rootCause,
+    evidence: firstOfKind(InvestigationReportSectionKind.Evidence),
+    nextSteps: firstOfKind(InvestigationReportSectionKind.NextSteps),
+    evidenceChecked,
+    footer: footerMatch?.footer,
+    bodyMarkdown: joinKeptLines(lines, kept, 0, lines.length).trim(),
+  };
+}
+
+type EventReferenceQualifier = "incident" | "alert" | "blocked";
+
+/*
+ * The qualifier word directly before a reference: "incident #12", "Alerts
+ * #3", "prior incidents: #6954", "alerts (#100". "blocked" means a word from
+ * another numbering ("PR #45"). Scans backwards over a bounded gap and word,
+ * so the cost never grows with the text.
+ */
+function readQualifierBefore(
+  text: string,
+  hashIndex: number,
+): EventReferenceQualifier | null {
+  let cursor: number = hashIndex - 1;
+  let punctuationCount: number = 0;
+
+  while (cursor >= 0) {
+    const character: string = text.charAt(cursor);
+
+    if (WHITESPACE_CHARACTER_REGEX.test(character)) {
+      cursor--;
+      continue;
+    }
+
+    if (
+      QUALIFIER_GAP_PUNCTUATION.includes(character) &&
+      punctuationCount < MAX_QUALIFIER_GAP_PUNCTUATION
+    ) {
+      punctuationCount++;
+      cursor--;
+      continue;
+    }
+
+    break;
+  }
+
+  // Nothing between the word and "#" ("incident-#12" has a hyphen there).
+  if (cursor === hashIndex - 1) {
+    return null;
+  }
+
+  const wordEnd: number = cursor + 1;
+  const wordStartLimit: number = wordEnd - MAX_QUALIFIER_WORD_LENGTH - 1;
+
+  while (
+    cursor >= 0 &&
+    cursor >= wordStartLimit &&
+    WORD_CHARACTER_REGEX.test(text.charAt(cursor))
+  ) {
+    cursor--;
+  }
+
+  // The word runs on past the longest qualifier, so it is none of them.
+  if (cursor >= 0 && WORD_CHARACTER_REGEX.test(text.charAt(cursor))) {
+    return null;
+  }
+
+  let wordStart: number = cursor + 1;
+
+  // "__Alerts__ #3": "_" is a word character, but here it is emphasis.
+  while (wordStart < wordEnd && text.charAt(wordStart) === "_") {
+    wordStart++;
+  }
+
+  const word: string = text.slice(wordStart, wordEnd).toLowerCase();
+
+  if (word === "incident" || word === "incidents") {
+    return "incident";
+  }
+
+  if (word === "alert" || word === "alerts") {
+    return "alert";
+  }
+
+  if (NON_EVENT_QUALIFIER_WORDS.has(word)) {
+    return "blocked";
+  }
+
+  return null;
+}
+
+interface ScannedEventReference {
+  qualifier: EventReferenceQualifier | null;
+  number: number;
+  start: number;
+  end: number;
+}
+
+/*
+ * Every "#123" in one text run with the qualifier it reads, "blocked" ones
+ * included. tokenizeEventReferences documents the rules.
+ */
+function scanEventReferences(text: string): Array<ScannedEventReference> {
+  const scanned: Array<ScannedEventReference> = [];
+
+  if (typeof text !== "string" || !text.includes("#")) {
+    return scanned;
+  }
+
+  const referenceRegex: RegExp = new RegExp(EVENT_REFERENCE_PATTERN, "g");
+  let match: RegExpExecArray | null = referenceRegex.exec(text);
+
+  while (match) {
+    const start: number = match.index;
+    const digits: string = match[1] || "";
+    const before: string = start > 0 ? text.charAt(start - 1) : "";
+    const isBlocked: boolean =
+      before !== "" &&
+      (WORD_CHARACTER_REGEX.test(before) ||
+        FORBIDDEN_BEFORE_REFERENCE.includes(before));
+
+    if (!isBlocked && digits) {
+      const end: number = start + 1 + digits.length;
+      const previous: ScannedEventReference | undefined =
+        scanned[scanned.length - 1];
+      let qualifier: EventReferenceQualifier | null = readQualifierBefore(
+        text,
+        start,
+      );
+
+      if (
+        qualifier === null &&
+        previous &&
+        previous.qualifier !== null &&
+        start - previous.end <= MAX_SIBLING_SEPARATOR_LENGTH &&
+        SIBLING_SEPARATOR_REGEX.test(text.slice(previous.end, start))
+      ) {
+        qualifier = previous.qualifier;
+      }
+
+      scanned.push({
+        qualifier,
+        number: parseInt(digits, 10),
+        start,
+        end,
+      });
+    }
+
+    match = referenceRegex.exec(text);
+  }
+
+  return scanned;
+}
+
+/*
+ * Finds "#123" style references in plain text (one text run). A reference is
+ * "#" + 1-9 digits, not preceded by a word character, "&", "/" or "#", and not
+ * followed by a word character.
+ *
+ * An immediately preceding qualifier word ("incident", "incidents", "alert",
+ * "alerts", case-insensitive) sets `kind`. It may be separated from the "#"
+ * by whitespace and a little punctuation (":", "(", "[", "*", "_"). A
+ * qualifier also carries to directly following comma/"and"/"or"-separated
+ * siblings ("incidents #1, #2 and #3").
+ *
+ * A number after a word from another numbering ("scheduled maintenance #42",
+ * "PR #45", "step #1") is not returned at all, and neither are its siblings
+ * ("PRs #45, #46"): it is not an incident or alert of any kind.
+ */
+export function tokenizeEventReferences(
+  text: string,
+): Array<InvestigationEventReferenceToken> {
+  const tokens: Array<InvestigationEventReferenceToken> = [];
+
+  for (const reference of scanEventReferences(text)) {
+    if (reference.qualifier === "blocked") {
+      continue;
+    }
+
+    tokens.push({
+      kind: reference.qualifier,
+      number: reference.number,
+      start: reference.start,
+      end: reference.end,
+    });
+  }
+
+  return tokens;
+}
+
+/*
+ * Replaces inline code spans with a placeholder so references inside them
+ * are ignored. A span opens with a run of N backticks and closes at the next
+ * run of exactly N; an unmatched run is literal text.
+ */
+function maskInlineCode(text: string): string {
+  if (!text.includes("`")) {
+    return text;
+  }
+
+  let output: string = "";
+  let cursor: number = 0;
+
+  while (cursor < text.length) {
+    const openStart: number = text.indexOf("`", cursor);
+
+    if (openStart === -1) {
+      output += text.slice(cursor);
+      break;
+    }
+
+    let openEnd: number = openStart;
+
+    while (openEnd < text.length && text.charAt(openEnd) === "`") {
+      openEnd++;
+    }
+
+    const runLength: number = openEnd - openStart;
+    let search: number = openEnd;
+    let closeEnd: number = -1;
+
+    while (search < text.length) {
+      const nextStart: number = text.indexOf("`", search);
+
+      if (nextStart === -1) {
+        break;
+      }
+
+      let nextEnd: number = nextStart;
+
+      while (nextEnd < text.length && text.charAt(nextEnd) === "`") {
+        nextEnd++;
+      }
+
+      if (nextEnd - nextStart === runLength) {
+        closeEnd = nextEnd;
+        break;
+      }
+
+      search = nextEnd;
+    }
+
+    if (closeEnd === -1) {
+      output += text.slice(cursor, openEnd);
+      cursor = openEnd;
+      continue;
+    }
+
+    output += text.slice(cursor, openStart) + CODE_SPAN_PLACEHOLDER;
+    cursor = closeEnd;
+  }
+
+  return output;
+}
+
+/*
+ * References across a whole markdown document, skipping fenced code blocks
+ * and inline code spans. Returns unique {kind, number} pairs in first-seen
+ * order, with null kinds replaced by `defaultKind`, capped at `limit`. An
+ * unqualified number the document also qualifies as the other kind is left
+ * out (see below). Callers resolving a posted report should pass its
+ * bodyMarkdown, so the server's own evidence labels are not read as prose.
+ */
+export function extractEventReferences(
+  markdown: string,
+  defaultKind: "incident" | "alert",
+  limit: number = DEFAULT_EVENT_REFERENCE_LIMIT,
+): Array<ExtractedInvestigationEventReference> {
+  const references: Array<ExtractedInvestigationEventReference> = [];
+  const maxReferences: number = Math.floor(limit);
+
+  if (!(maxReferences > 0)) {
+    return references;
+  }
+
+  const normalised: string = normaliseLineEndings(markdown);
+
+  if (!normalised.includes("#")) {
+    return references;
+  }
+
+  const lines: Array<string> = normalised.split("\n");
+  const fenced: Array<boolean> = markFencedLines(lines);
+
+  /*
+   * Group the non-code lines into blocks separated by fences and blank
+   * lines: a code span never crosses either, so an unmatched backtick in one
+   * paragraph cannot hide references in the next.
+   */
+  const blocks: Array<string> = [];
+  let current: Array<string> = [];
+
+  lines.forEach((line: string, index: number): void => {
+    if (fenced[index] || !line.trim()) {
+      if (current.length > 0) {
+        blocks.push(current.join("\n"));
+        current = [];
+      }
+      return;
+    }
+
+    current.push(line);
+  });
+
+  if (current.length > 0) {
+    blocks.push(current.join("\n"));
+  }
+
+  const scanned: Array<ScannedEventReference> = [];
+
+  for (const block of blocks) {
+    for (const reference of scanEventReferences(maskInlineCode(block))) {
+      scanned.push(reference);
+    }
+  }
+
+  /*
+   * An unqualified "#12" is taken as the subject's kind, except when the
+   * same document also names 12 as something else: explicitly the OTHER
+   * kind ("alert #12 fired ... #12 kept firing" on an incident) or in another
+   * numbering ("scheduled maintenance #12 ... during #12"). The bare number
+   * most likely means that, so it is not looked up as the subject's kind at
+   * all: it stays plain text rather than linking to an unrelated record.
+   */
+  const otherKind: "incident" | "alert" =
+    defaultKind === "incident" ? "alert" : "incident";
+  const numbersNamedOtherwise: Set<number> = new Set<number>();
+
+  for (const reference of scanned) {
+    if (
+      reference.qualifier === otherKind ||
+      reference.qualifier === "blocked"
+    ) {
+      numbersNamedOtherwise.add(reference.number);
+    }
+  }
+
+  const seen: Set<string> = new Set<string>();
+
+  for (const token of scanned) {
+    if (
+      token.qualifier === "blocked" ||
+      (token.qualifier === null && numbersNamedOtherwise.has(token.number))
+    ) {
+      continue;
+    }
+
+    const kind: "incident" | "alert" = token.qualifier || defaultKind;
+    const key: string = `${kind}:${token.number}`;
+
+    if (seen.has(key)) {
+      continue;
+    }
+
+    seen.add(key);
+    references.push({ kind, number: token.number });
+
+    if (references.length >= maxReferences) {
+      return references;
+    }
+  }
+
+  return references;
+}