deer-flow/frontend/tests/unit/core/citations/sources.test.ts

import { describe, expect, it } from "@rstest/core";

import {
  extractCitationSources,
  formatCitationMarkdownReference,
} from "@/core/citations/sources";

describe("extractCitationSources", () => {
  it("extracts citation markdown links in first-seen order", () => {
    const markdown = [
      "Deep research needs evidence [citation:Paper A](https://example.com/a).",
      "A second claim cites [citation:Report B](https://news.example.org/report?x=1).",
    ].join("\n");
    const firstIndex = markdown.indexOf("[citation:Paper A]");
    const secondIndex = markdown.indexOf("[citation:Report B]");

    expect(extractCitationSources(markdown)).toEqual([
      {
        id: "https://example.com/a",
        title: "Paper A",
        url: "https://example.com/a",
        domain: "example.com",
        count: 1,
        occurrences: [{ index: firstIndex, title: "Paper A" }],
      },
      {
        id: "https://news.example.org/report?x=1",
        title: "Report B",
        url: "https://news.example.org/report?x=1",
        domain: "news.example.org",
        count: 1,
        occurrences: [{ index: secondIndex, title: "Report B" }],
      },
    ]);
  });

  it("deduplicates repeated citation URLs and preserves occurrence titles", () => {
    const markdown = [
      "First [citation:Original Title](https://example.com/research).",
      "Later [citation:Updated Title](https://example.com/research).",
    ].join("\n");
    const firstIndex = markdown.indexOf("[citation:Original Title]");
    const secondIndex = markdown.indexOf("[citation:Updated Title]");

    expect(extractCitationSources(markdown)).toEqual([
      {
        id: "https://example.com/research",
        title: "Original Title",
        url: "https://example.com/research",
        domain: "example.com",
        count: 2,
        occurrences: [
          { index: firstIndex, title: "Original Title" },
          { index: secondIndex, title: "Updated Title" },
        ],
      },
    ]);
  });

  it("ignores normal links, image links, and citations inside fenced code", () => {
    const markdown = [
      "[Normal](https://example.com/normal)",
      "![citation:Image](https://example.com/image.png)",
      "```md",
      "[citation:Example](https://example.com/example)",
      "```",
      "Real source [citation:Real](https://example.com/real).",
    ].join("\n");
    const realIndex = markdown.indexOf("[citation:Real]");

    expect(extractCitationSources(markdown)).toEqual([
      {
        id: "https://example.com/real",
        title: "Real",
        url: "https://example.com/real",
        domain: "example.com",
        count: 1,
        occurrences: [{ index: realIndex, title: "Real" }],
      },
    ]);
  });

  it("keeps every source when citations are directly adjacent", () => {
    const markdown =
      "[citation:A](https://example.com/a)[citation:B](https://example.com/b)[citation:C](https://example.com/c)";

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/a",
      "https://example.com/b",
      "https://example.com/c",
    ]);
  });

  it("keeps URLs that contain multiple balanced parenthetical groups", () => {
    const markdown = "[citation:W](https://en.wikipedia.org/wiki/Foo_(a)_(b))";

    expect(extractCitationSources(markdown)[0]).toMatchObject({
      url: "https://en.wikipedia.org/wiki/Foo_(a)_(b)",
      domain: "en.wikipedia.org",
    });
  });

  it("ignores citations inside inline code spans", () => {
    const markdown =
      "Example: `[citation:X](https://x.com/inline)` then real [citation:Real](https://example.com/real).";

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps a citation whose paragraph sits between two unclosed backtick runs", () => {
    // An inline span cannot cross a blank line, so neither stray backtick opens
    // a span here and the citation is a rendered link, not sample code.
    const markdown = [
      "Install it with `npm i",
      "",
      "The upstream guide is [citation:Docs](https://example.com/docs).",
      "",
      "Then run `npm start` to serve it.",
    ].join("\n");

    const sources = extractCitationSources(markdown);

    expect(sources.map((source) => source.url)).toEqual([
      "https://example.com/docs",
    ]);
    expect(sources[0]?.occurrences[0]?.index).toBe(
      markdown.indexOf("[citation:Docs]"),
    );
  });

  it("still masks a code span that closes inside its own paragraph", () => {
    const markdown = [
      "Install it with `npm i",
      "",
      "Example: `[citation:Fake](https://example.com/fake)` then real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("ends a paragraph on a CRLF blank line, as the inline scanner does", () => {
    const markdown =
      "Install it with `npm i\r\n\r\nThe upstream guide is [citation:Docs](https://example.com/docs).\r\n\r\nThen run `npm start` to serve it.";

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/docs",
    ]);
  });

  it("ignores citations inside an unclosed fenced code block", () => {
    const markdown = [
      "Streaming output:",
      "```md",
      "[citation:Streaming](https://example.com/streaming)",
    ].join("\n");

    expect(extractCitationSources(markdown)).toEqual([]);
  });

  it("ends an inline span at a heading, thematic break or list marker", () => {
    // CommonMark ends the paragraph at these block boundaries too, so "Use
    // `npm" cannot steal the opener of `npm start` two lines later and the
    // citation between them is a rendered link, not sample code.
    for (const boundary of ["## Title", "---", "- item"]) {
      const markdown = [
        "Use `npm",
        boundary,
        "See [citation:X](https://example.com/x) and `npm start`.",
      ].join("\n");

      expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
        "https://example.com/x",
      ]);
    }
  });

  it("ends an inline span at a blank line inside a blockquote", () => {
    const markdown = [
      "> Install it with `npm i",
      ">",
      "> The guide is [citation:Docs](https://example.com/docs) and `npm start`.",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/docs",
    ]);
  });

  it("masks citations inside a fenced block nested in a list item", () => {
    // The fence is indented past the list marker, so recognising it needs the
    // container prefix; the blank line inside keeps it open across paragraphs.
    const markdown = [
      "- run:",
      "    ```md",
      "    [citation:Fake1](https://example.com/fake1)",
      "",
      "    [citation:Fake2](https://example.com/fake2)",
      "    ```",
      "",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("masks a blockquoted fence, and stops at the line that ends the quote", () => {
    // Fake1 is inside the fence. The blank line has no `>` marker, so it closes
    // the blockquote — and the fence with it — which makes Fake2 a rendered link
    // again. Verified against remark-parse rather than by eye.
    const markdown = [
      "> ```md",
      "> [citation:Fake1](https://example.com/fake1)",
      "",
      "> [citation:Fake2](https://example.com/fake2)",
      "> ```",
      "",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake2",
      "https://example.com/real",
    ]);
  });

  it("masks a dedented citation that lands in an indented code block", () => {
    // Losing the `>` closes the block quote and the fence inside it, but the
    // escaping line is not free: four columns of indentation at the top level
    // is an indented code block, so Fake still renders as code. Verified
    // against remark-parse, whose tree is quote>code, code, quote>code, paragraph.
    const markdown = [
      "> ```md",
      "    [citation:Fake](https://example.com/fake)",
      "> ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("counts a tab-indented escaping line as the same four-column run", () => {
    // A tab reaches the next tab stop, so `\t` is four columns and `\t\t` eight;
    // both are indented code, and the marker padding on the opener is irrelevant.
    const markdown = [
      ">\t```md",
      "\t[citation:Fake](https://example.com/fake)",
      ">\t```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps an indented run across a blank line inside it", () => {
    // The CRLF variant pins the blankness classification: after a split on
    // "\n" the blank line still carries its "\r", and reading it as content
    // cleared the run and let Fake2 through.
    for (const eol of ["\n", "\r\n"]) {
      const markdown = [
        "> ```md",
        "    [citation:Fake1](https://example.com/fake1)",
        "",
        "    [citation:Fake2](https://example.com/fake2)",
        "> ```",
        "Real [citation:Real](https://example.com/real).",
      ].join(eol);

      expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
        "https://example.com/real",
      ]);
    }
  });

  it("masks the indented code block that follows the blank line ending a quoted fence", () => {
    // The blank line ends the blockquote, and the four-space line after it is a
    // top-level indented code block, so Fake stays code instead of surfacing as
    // a phantom source. Verified against remark-parse.
    const markdown = [
      "> ```md",
      "",
      "    [citation:Fake](https://example.com/fake)",
      "> ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("renders the paragraph that follows the blank line ending a quoted fence", () => {
    // Two columns is paragraph territory, so the pending quote exit must not
    // mask the line that classifies it: this pins the boundary against the
    // four-column case above.
    const markdown = [
      "> ```md",
      "",
      "  [citation:Para](https://example.com/para)",
      "> ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/para",
      "https://example.com/real",
    ]);
  });

  it("measures a quote escape against the list item the quote sat in", () => {
    // The quote is a child of the list item and the item outlives it, so four
    // columns of indentation are only two past the item's content column: the
    // citation is paragraph content, not an indented code block. Verified
    // against remark-parse.
    const markdown = [
      "- item",
      "  > ```md",
      "    [citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("still masks a quote escape indented past the surviving list item", () => {
    // Six columns are four past the item's content column, so the escaped line
    // does start an indented code block — and the unindented line after it ends
    // the item and the run, rendering again.
    const markdown = [
      "- item",
      "  > ```md",
      "      [citation:Fake](https://example.com/fake)",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("ends the indented run at a dedent so that citation renders again", () => {
    // Two columns is inside the item-less paragraph the run gives way to, so the
    // second citation is a rendered link while the four-column one above stays
    // code.
    const markdown = [
      "> ```md",
      "    [citation:Code](https://example.com/code)",
      "  [citation:Para](https://example.com/para)",
      "> ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/para",
      "https://example.com/real",
    ]);
  });

  it("resumes normal scanning once an indented run meets a quoted line", () => {
    // The run ends at `> [citation:Quoted]`, which is quote content again, and
    // the `> ``` ` below it opens the fence that the final line escapes.
    const markdown = [
      "> ```md",
      "    [citation:Fake](https://example.com/fake)",
      "> [citation:Quoted](https://example.com/quoted)",
      "> ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/quoted",
      "https://example.com/real",
    ]);
  });

  it("does not mask an escaping citation three columns in", () => {
    // Three columns is not enough to start an indented code block, so the line
    // is a top-level paragraph and its citation is a real rendered link. This
    // pins the threshold against the four-column case above.
    const markdown = [
      "> ```md",
      "   [citation:Fake](https://example.com/fake)",
      "> ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake",
      "https://example.com/real",
    ]);
  });

  it("does not open a fence from a backtick run in the middle of a line", () => {
    const markdown = [
      "Run ```md please",
      "[citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps a fence open across a shorter backtick run inside it", () => {
    const markdown = [
      "````md",
      "```",
      "[citation:Inside](https://example.com/inside)",
      "```",
      "````",
      "",
      "Outside [citation:Outside](https://example.com/outside).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/outside",
    ]);
  });

  it("keeps an indented top-level fence open for unindented content", () => {
    // The opener's own indentation is not container indentation: content lines
    // may sit further left than the marker, and the closer only has to come back
    // within three columns of the opener.
    const markdown = [
      "Intro text.",
      "  ```md",
      "[citation:Fake](https://example.com/fake)",
      "  ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("treats a four-space indented marker inside a column-zero fence as content", () => {
    const markdown = [
      "```md",
      "    ```",
      "[citation:Fake](https://example.com/fake)",
      "```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("does not open a fence from a marker indented four columns at the top level", () => {
    // Four columns past the container's content column is an indented code
    // block, not a fence opener, so the citation below it renders as a link.
    const markdown = [
      "    ```md",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("still opens a fence from a marker indented three columns at the top level", () => {
    // The limit is three columns, so the next case up pins the boundary rather
    // than overshooting it.
    const markdown = [
      "   ```md",
      "[citation:Fake](https://example.com/fake)",
      "   ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("does not open a fence from a list marker with no space after it", () => {
    // `-` has to be followed by indentation to start a list item, so `-```md`
    // is ordinary paragraph text and nothing behind it is code.
    const markdown = [
      "-```md",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("counts a tab after a list marker as columns when it sizes the item", () => {
    // `-<TAB>` advances to the next four-column stop, so this item's content
    // starts at column four and the citation two spaces in has left the fence
    // rather than sitting inside it.
    const markdown = [
      "-\t```md",
      "  [citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("ends a fence when a line drops out of the inner blockquote", () => {
    const markdown = [
      ">> ```md",
      ">> [citation:Fake](https://example.com/fake)",
      "> Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("lets a citation after a list-nested quote escape its fence", () => {
    // `- > ```md` opens the quote behind a list marker, so the line has no
    // leading `>` for a line-start pattern to find even though the fence really
    // is quoted. The dedented citation below sits two columns past the item's
    // content column, which is a paragraph rather than an indented code block.
    const markdown = [
      "- > ```md",
      "    [citation:Fake](https://example.com/fake)",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake",
      "https://example.com/real",
    ]);
  });

  it("ends a list-nested quote fence at its first unquoted line", () => {
    const markdown = [
      "- > ```md",
      "    [citation:One](https://example.com/one)",
      "     [citation:Two](https://example.com/two)",
      "\t[citation:Three](https://example.com/three)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/one",
      "https://example.com/two",
      "https://example.com/three",
    ]);
  });

  it("escapes a quote fence nested behind a star marker", () => {
    const markdown = [
      "* > ```md",
      "  [citation:Fake](https://example.com/fake)",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake",
      "https://example.com/real",
    ]);
  });

  it("escapes a quote fence nested behind an ordered marker", () => {
    // `1. ` gives the item a content column of three, so four columns of
    // indentation is one column of content rather than an indented code block.
    const markdown = [
      "1. > ```md",
      "    [citation:Fake](https://example.com/fake)",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake",
      "https://example.com/real",
    ]);
  });

  it("keeps a list-nested quote open for a line that carries its marker", () => {
    // The mirror of the cases above: `>` on the second line means the quote did
    // not end, so nothing about the escape path applies here.
    const markdown = [
      "- >> ```md",
      ">> [citation:Fake](https://example.com/fake)",
      "> [citation:Mid](https://example.com/mid)",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake",
      "https://example.com/mid",
      "https://example.com/real",
    ]);
  });

  it("ends a list-nested fence at a closer that is itself inside the quote", () => {
    const markdown = [
      "- > ```md",
      "> [citation:Fake](https://example.com/fake)",
      "> ```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake",
      "https://example.com/real",
    ]);
  });

  it("gives a quoted closer its full three-column budget", () => {
    // The one optional space after `>` belongs to the block quote marker, not to
    // the fence, so `>` plus four spaces indents the closer by three columns and
    // it closes the block.
    const markdown = [
      "> ```md",
      "> [citation:Fake](https://example.com/fake)",
      ">    ```",
      "> [citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("does not let a marker with an info string close the fence", () => {
    // A closing fence is bare, so ` ```text ` here is literal content and the
    // citation after it stays inside the block instead of becoming a source.
    const markdown = [
      "```md",
      "```text",
      "[citation:Fake](https://example.com/fake)",
      "```",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("does not close a fence from list-looking content inside it", () => {
    // Inside a fenced block `- ``` ` is literal content: a closing fence may
    // carry indentation and a marker, nothing else. Reading it as a closer leaks
    // Fake, and the real closer then masks away Real.
    const markdown = [
      "```md",
      "- ```",
      "[citation:Fake](https://example.com/fake)",
      "```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("closes a list-nested fence at the item's content column", () => {
    // The same rule one container deeper: `- ```md` opens the fence at the list
    // item's content column, `  - ``` ` stays literal content, and the bare
    // `  ``` ` at that column closes it. Verified against remark-parse, not by
    // eye.
    const markdown = [
      "- ```md",
      "  - ```",
      "  [citation:Fake](https://example.com/fake)",
      "  ```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("does not open a backtick fence whose info string holds a backtick", () => {
    // ` ```md `x` ` is paragraph text, so the citation below it is a link the
    // reader can reach and must not be masked away.
    const markdown = [
      "```md `x`",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("opens a tilde fence whatever its info string says", () => {
    // Tilde fences take backticks in the info string, so the boundary above is
    // about the marker character rather than a blanket rule.
    const markdown = [
      "~~~md `x`",
      "[citation:Fake](https://example.com/fake)",
      "~~~",
      "Real [citation:Real](https://example.com/real).",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps a literal quote marker inside an open list fence", () => {
    // Once a fence is open its lines are literal, so the `>` here is code rather
    // than a new block quote. Reading it as a container made the line look
    // dedented to column zero, which escaped the column-two list fence and
    // surfaced the sample's citation while reinterpreting the real closer as a
    // top-level opener.
    const markdown = [
      "- ```md",
      "  > example",
      "  [citation:Fake](https://example.com/fake)",
      "  ```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps a citation on a literal quote line inside a list fence", () => {
    // Same shape with the citation on the quoted line itself: it stays code.
    const markdown = [
      "- ```md",
      "  > [citation:Fake](https://example.com/fake)",
      "  ```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps a citation inside a fence opened behind several list markers", () => {
    // `- - - ```md` is a fence in the third list item, whose content column is
    // six. Recording only the first marker put the fence at column two, which
    // rejected the valid opener as four columns past its container.
    const markdown = [
      "- - - ```md",
      "      [citation:Fake1](https://example.com/fake1)",
      "",
      "      [citation:Fake2](https://example.com/fake2)",
      "      ```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps a literal quote marker inside a quoted list fence", () => {
    // The fence already consumed the quote it was opened in, so the second `>`
    // here is literal code. `linePosition()` stripped both markers, which made
    // the sample line look dedented past the fence's column and ended the fence
    // one line early, surfacing the citation the renderer shows as code.
    const markdown = [
      "> - ```md",
      ">   > example",
      ">   [citation:Fake](https://example.com/fake)",
      ">   ```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("keeps the list item open across a lazy paragraph continuation", () => {
    // An unindented line of paragraph text continues the item's paragraph rather
    // than leaving it, so the item's content column still measures the fence two
    // columns in. Popping the item there recorded the fence at column zero, which
    // swallowed the unindented citation to end of input.
    const markdown = [
      "- note",
      "lazy continuation",
      "",
      "  ~~~md",
      "  sample",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("reads five columns after a list marker as indented code, not padding", () => {
    // Only up to four columns of whitespace pad a list marker; the fifth belongs
    // to an indented code block inside an item whose content column is two.
    // Consuming the whole run recorded column six, so the two-space fence looked
    // like a dedent that left the item and the citation after it hid to EOF.
    const markdown = [
      "-     code",
      "",
      "  ~~~md",
      "  sample",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("restores the enclosing list when an ordinary quote ends", () => {
    // The blank line ends `> note`, but the list item outlives the quote, so the
    // two-space fence still belongs to an item whose content column is two. The
    // citation at column zero leaves that item and ends the unclosed fence with
    // it, which makes it a rendered link.
    const markdown = [
      "- item",
      "  > note",
      "",
      "  ~~~md",
      "  [citation:Fake](https://example.com/fake)",
      "  sample",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("restores the enclosing list when a quote ends at the fence line", () => {
    // The same exit without a blank line: a fence cannot be lazily continued by
    // a block quote, so `~~~md` already left the quote and the item's column is
    // what the following lines measure against.
    const markdown = [
      "- item",
      "  > note",
      "  ~~~md",
      "  [citation:Fake](https://example.com/fake)",
      "  sample",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("restores the enclosing list after a two-deep quote ends", () => {
    const markdown = [
      "- item",
      "  >> note",
      "",
      "  ~~~md",
      "  [citation:Fake](https://example.com/fake)",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("does not open a fence from an ordered marker that cannot interrupt a paragraph", () => {
    // An ordered list that does not start at 1 cannot interrupt a paragraph, so
    // all three lines are one paragraph and its citation renders.
    const markdown = [
      "Intro",
      "2. ~~~md",
      "   [citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("does not open a fence from a non-interrupting marker inside an item paragraph", () => {
    const markdown = [
      "- item",
      "  2. ~~~md",
      "     [citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("still opens a fence from a `1.` marker that does interrupt the paragraph", () => {
    const markdown = [
      "Intro",
      "1. ~~~md",
      "   [citation:Fake](https://example.com/fake)",
    ].join("\n");

    expect(extractCitationSources(markdown)).toEqual([]);
  });

  it("still opens a fence from a `2.` marker after a block boundary", () => {
    // The blank line closes the paragraph, so the ordered item starts a real
    // container and its fence hides the citation again.
    const markdown = [
      "Intro",
      "",
      "2. ~~~md",
      "   [citation:Fake](https://example.com/fake)",
    ].join("\n");

    expect(extractCitationSources(markdown)).toEqual([]);
  });

  it("keeps the citation a newly entered quote renders as a paragraph", () => {
    // Pinned against the parser rather than against a rule: remark's tree for
    // these lines is blockquote > [paragraph "2. ```md", paragraph with the
    // Fake link, code], so `2. ```md` never becomes a list item inside the
    // fresh quote and `>    ``` ` is the opener of an unclosed fence. Both
    // citations render, and Sources has to surface both.
    const markdown = [
      "Intro",
      "> 2. ```md",
      ">",
      ">    [citation:Fake](https://example.com/fake)",
      ">    ```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/fake",
      "https://example.com/real",
    ]);
  });

  it("ends the paragraph for a setext underline over LF", () => {
    const markdown = [
      "Intro",
      "---",
      "2. ```md",
      "   [citation:Fake1](https://example.com/fake1)",
      "",
      "   [citation:Fake2](https://example.com/fake2)",
      "   ```",
      "[citation:Real](https://example.com/real)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("ends the paragraph for a setext underline over CRLF", () => {
    // The trailing "\r" belongs to the line ending, so `---` is still a setext
    // underline and the paragraph is over: `2.` opens a real list, both of its
    // citations are fenced code, and the `   ``` ` line closes that fence rather
    // than starting a new one that would hide the rendered citation to EOF.
    const markdown = [
      "Intro",
      "---",
      "2. ```md",
      "   [citation:Fake1](https://example.com/fake1)",
      "",
      "   [citation:Fake2](https://example.com/fake2)",
      "   ```",
      "[citation:Real](https://example.com/real)",
    ].join("\r\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/real",
    ]);
  });

  it("uses the source domain when the citation label is generic", () => {
    const markdown = "See [citation:Source](https://www.example.com/path).";

    expect(extractCitationSources(markdown)[0]).toMatchObject({
      title: "example.com",
      domain: "example.com",
      url: "https://www.example.com/path",
    });
  });
});

describe("scans a deeply repeated blockquote chain in one pass", () => {
  // tests/e2e/user-message-plain-text.spec.ts renders an AI message whose whole
  // body is one line of thousands of `>` markers, and the citation pass runs on
  // that text before the renderer does. Reading the containers a fence marker
  // may sit behind has to reach each position once: an alternation that can take
  // a space either as a step of its own or as the padding of the next marker
  // describes the same prefix many ways, so a line carrying no fence leaves the
  // engine to enumerate them all. Measured against the pattern this file guards,
  // one check costs about a second at 26 markers and roughly doubles for each
  // marker added, so a three-thousand-marker line never returns. The budget sits
  // two orders of magnitude above what a single pass takes, which pins the growth
  // rather than the machine.
  const shapes: Array<[string, string]> = [
    ["> ".repeat(26) + "deep", "26-level quote"],
    ["- " + "> ".repeat(26) + "deep-list", "26-level quote behind a list"],
    ["> ".repeat(3000) + "deep", "3000-level quote"],
    ["- " + "> ".repeat(3000) + "deep-list", "3000-level quote behind a list"],
  ];

  for (const [markdown, label] of shapes) {
    it(`masks nothing in the ${label} shape within one pass of the line`, () => {
      const started = performance.now();
      const sources = extractCitationSources(markdown);

      expect(sources).toEqual([]);
      expect(performance.now() - started).toBeLessThan(500);
    });
  }

  it("still reaches a citation that follows a deep quote chain", () => {
    // Skipping the scan would keep every case above green, so the same shape is
    // checked with real work to find after it.
    const markdown = [
      "> ".repeat(26) + "deep",
      "[citation:Deep](https://example.com/deep)",
    ].join("\n");

    expect(extractCitationSources(markdown).map((s) => s.url)).toEqual([
      "https://example.com/deep",
    ]);
  });
});

describe("formatCitationMarkdownReference", () => {
  it("formats a source as a reusable markdown reference", () => {
    const [source] = extractCitationSources(
      "Evidence [citation:Paper A](https://example.com/a).",
    );

    expect(formatCitationMarkdownReference(source!)).toBe(
      "[Paper A](https://example.com/a)",
    );
  });
});