failproofai 1.0.1-beta.1 → 1.0.2-beta.0

package/scripts/release-announcement.mjs

@@ -0,0 +1,686 @@
+/**
+ * Build the Discord webhook payload for a STABLE release.
+ *
+ * Called by the `announce` job in `.github/workflows/publish.yml`, after the
+ * registry check and the four `verify-install` legs have proved the release is
+ * real.
+ *
+ * **The GitHub Release body is the source of record; CHANGELOG.md is the
+ * fallback.** Stable releases are cut from the GitHub Releases page, and the
+ * notes written there are what the maintainer decided this release says —
+ * chosen for an audience, ordered on purpose, sometimes rewritten from the
+ * changelog entirely. Announcing from CHANGELOG.md instead would publish a
+ * DIFFERENT summary than the one on the release page, in the channel where more
+ * people read it. So the body wins whenever there is one, and the changelog
+ * covers the two cases where there is not: an empty release body, and a
+ * `workflow_dispatch` build, which has no release event at all.
+ *
+ * Three other things are load-bearing and easy to get wrong:
+ *
+ * **From the changelog, a stable release carries its whole beta line.** Somebody
+ * on `latest` moving 1.0.0 -> 1.0.1 receives everything that shipped across
+ * `1.0.1-beta.*` as well as whatever the release commit itself carried, and the
+ * `## 1.0.1` section deliberately does NOT restate it ("Everything below this
+ * heading shipped across the `1.0.0-beta.*` line" — the 1.0.0 section says so
+ * in as many words). Announcing only the stable section would describe a
+ * release as a handful of entries when it is forty, so `collectRelease` gathers
+ * `## <version>` and every `## <version>-*` section under it and merges them.
+ *
+ * **Only the first sentence of an entry survives.** This changelog's entries are
+ * paragraphs — several are over 2000 characters, which is the entire Discord
+ * message limit on its own. The first sentence is written as a headline in
+ * every entry in the file, so it is the one summary already there rather than
+ * one this script invents. GitHub's own generated notes are one PR title per
+ * bullet, so the same pass leaves them untouched.
+ *
+ * **The mention goes in `content`, never in the embed.** Discord does not
+ * resolve mentions inside embeds — a `<@&id>` there renders as the raw string
+ * and pings nobody. Paired with `allowed_mentions: {parse: [], roles: [id]}`
+ * so that the release role is the ONLY thing that can be pinged: release notes
+ * reach the embed unescaped, and `parse: []` is what stops an `@everyone` in
+ * them from becoming one.
+ */
+import { existsSync, readFileSync, writeFileSync } from "node:fs";
+import { resolve, dirname } from "node:path";
+import { fileURLToPath, pathToFileURL } from "node:url";
+
+const REPO_ROOT = resolve(dirname(fileURLToPath(import.meta.url)), "..");
+
+/** The brand pink — `HUES.pink` in `src/hooks/tui.ts`, as Discord's integer. */
+export const BRAND_COLOR = 0xff2e88;
+
+/**
+ * Entries under these headings are left out of the embed. A notification is
+ * not the changelog; it links to it. Dependency bumps are the one category
+ * nobody reads a release announcement for, and they routinely outnumber
+ * everything else. `collectRelease` still counts them, so the "+N more" line
+ * stays honest, and a release that is ONLY dependency bumps falls back to
+ * showing them rather than announcing nothing.
+ */
+const DEPRIORITIZED_GROUPS = new Set(["dependencies"]);
+
+/** Highlights shown per heading before the rest collapse into a "+N more". */
+const MAX_ENTRIES_PER_GROUP = 6;
+
+/** A headline longer than this is cut at a word boundary. */
+const MAX_HEADLINE = 180;
+
+/** How much of a stable section's prose preamble reaches the embed. */
+const MAX_LEAD = 600;
+
+// Discord's documented maxima. Exceeding any one of them is a 400 from the
+// webhook, i.e. a release that publishes fine and announces nothing.
+const LIMIT_CONTENT = 2000;
+const LIMIT_DESCRIPTION = 4096;
+const LIMIT_FIELD_VALUE = 1024;
+const LIMIT_TITLE = 256;
+const LIMIT_EMBED_TOTAL = 6000;
+
+// ── Changelog parsing ───────────────────────────────────────────────────────
+
+/**
+ * Split `CHANGELOG.md` into `## <version> — <date>` sections.
+ *
+ * The separator is matched as em dash, en dash or hyphen because the file is
+ * hand-written and a heading is exactly the kind of line where that slips. The
+ * date is optional for the same reason: a heading with no date is still a
+ * release, and refusing to parse it would take the announcement down over
+ * punctuation.
+ *
+ * Sections are returned in FILE order (newest first) and duplicate versions are
+ * kept as separate sections — the file genuinely carries two `## 1.0.0-beta.13`
+ * headings and two `## 1.0.0-beta.15` headings, and dropping either would drop
+ * real entries. `collectRelease` merges them.
+ */
+export function parseChangelog(text) {
+  const lines = String(text).split(/\r?\n/);
+  const sections = [];
+  let current = null;
+
+  for (const line of lines) {
+    const heading = /^##\s+(\S+)\s*(?:[—–-]\s*(.*))?$/.exec(line);
+    if (heading) {
+      current = { version: heading[1], date: (heading[2] ?? "").trim(), body: [] };
+      sections.push(current);
+      continue;
+    }
+    if (current) current.body.push(line);
+  }
+
+  return sections.map((s) => ({ ...s, body: s.body.join("\n") }));
+}
+
+/**
+ * Split a section body into `### <name>` groups plus any prose that precedes
+ * the first one.
+ *
+ * Bullets are joined across continuation lines: an entry runs until the next
+ * line that starts a new bullet or a new heading. Every entry in this file is
+ * currently one long line, which is exactly why a wrapped one would otherwise
+ * be silently truncated at the wrap.
+ */
+export function parseSectionBody(body) {
+  const lines = String(body).split(/\r?\n/);
+  const lead = [];
+  const groups = [];
+  let group = null;
+  let entry = null;
+
+  const flush = () => {
+    if (entry !== null && group) group.entries.push(entry.join(" ").trim());
+    entry = null;
+  };
+
+  for (const line of lines) {
+    const heading = /^###\s+(.+?)\s*$/.exec(line);
+    if (heading) {
+      flush();
+      group = { name: heading[1], entries: [] };
+      groups.push(group);
+      continue;
+    }
+
+    const bullet = /^[-*]\s+(.*)$/.exec(line);
+    if (bullet && group) {
+      flush();
+      entry = [bullet[1]];
+      continue;
+    }
+
+    if (entry !== null) {
+      // A blank line ends an entry; anything else continues it.
+      if (line.trim() === "") flush();
+      else entry.push(line.trim());
+      continue;
+    }
+
+    if (!group) lead.push(line);
+  }
+  flush();
+
+  return { lead: lead.join("\n").trim(), groups };
+}
+
+/**
+ * Gather the stable section for `version` plus every prerelease section under
+ * it, merging groups of the same name.
+ *
+ * Merging by name is not tidiness: `1.0.1-beta.2` carries TWO `### Fixes`
+ * headings, so keying on the heading text is what keeps the second one's
+ * entries from replacing the first one's — and what keeps `Fixes` from
+ * appearing twice in the embed.
+ *
+ * Only the stable section's prose preamble is kept. A beta's preamble is
+ * written for people tracking betas, and stacking four of them would bury the
+ * entries.
+ */
+export function collectRelease(changelogText, version) {
+  const sections = parseChangelog(changelogText);
+  const mine = sections.filter((s) => s.version === version || s.version.startsWith(`${version}-`));
+  if (mine.length === 0) return null;
+
+  const byName = new Map();
+  let lead = "";
+  let date = "";
+
+  for (const section of mine) {
+    const parsed = parseSectionBody(section.body);
+    if (section.version === version) {
+      if (!lead) lead = parsed.lead;
+      if (!date) date = section.date;
+    }
+    for (const group of parsed.groups) {
+      const key = group.name.trim().toLowerCase();
+      if (!byName.has(key)) byName.set(key, { name: group.name.trim(), entries: [] });
+      byName.get(key).entries.push(...group.entries);
+    }
+  }
+
+  const groups = [...byName.values()].map((g) => ({
+    name: g.name,
+    entries: g.entries.map(summarizeEntry),
+  }));
+
+  return {
+    version,
+    date,
+    lead,
+    groups,
+    total: groups.reduce((n, g) => n + g.entries.length, 0),
+    sections: mine.map((s) => s.version),
+  };
+}
+
+/**
+ * A GitHub Release body -> the same `{lead, groups, total}` shape a changelog
+ * section produces, so everything downstream is source-agnostic.
+ *
+ * Two shapes arrive here and both have to work:
+ *
+ *   * **GitHub's generated notes** — `## What's Changed`, then one
+ *     `* <PR title> by @someone in <pull url>` per merged PR, then a
+ *     `**Full Changelog**: <compare url>` line. The trailing attribution is
+ *     stripped and the pull URL becomes the `#NNN` link, because "by @x in
+ *     https://github.com/…/pull/123" is 60 characters of noise per line in a
+ *     message with a 4096-character budget.
+ *
+ *   * **Hand-written notes** — arbitrary markdown, usually `###` sections with
+ *     `-` bullets, which is the changelog's own shape.
+ *
+ * Notes with no headings at all are still notes: their bullets land in one
+ * unnamed group rather than being dropped, and prose with no bullets at all
+ * becomes the lead. The empty case returns null so the caller can fall back.
+ */
+export function parseReleaseBody(text) {
+  const body = String(text ?? "").trim();
+  if (!body) return null;
+
+  const lead = [];
+  const groups = [];
+  let group = null;
+  let entry = null;
+  let compareUrl = null;
+
+  const flush = () => {
+    if (entry !== null) {
+      const joined = entry.join(" ").trim();
+      if (joined) {
+        if (!group) {
+          group = { name: "Highlights", entries: [] };
+          groups.push(group);
+        }
+        group.entries.push(joined);
+      }
+    }
+    entry = null;
+  };
+
+  for (const line of body.split(/\r?\n/)) {
+    const heading = /^#{2,4}\s+(.+?)\s*$/.exec(line);
+    if (heading) {
+      flush();
+      group = { name: heading[1].replace(/^what's changed$/i, "Changes"), entries: [] };
+      groups.push(group);
+      continue;
+    }
+
+    // GitHub appends this to every generated body. We render our own link, and
+    // a compare URL is a better one than a blob URL when it is offered.
+    const full = /^\*{0,2}Full Changelog\*{0,2}:\s*(\S+)\s*$/i.exec(line.trim());
+    if (full) {
+      flush();
+      compareUrl = full[1];
+      continue;
+    }
+
+    const bullet = /^\s*[-*+]\s+(.*)$/.exec(line);
+    if (bullet) {
+      flush();
+      entry = [bullet[1]];
+      continue;
+    }
+
+    if (entry !== null) {
+      if (line.trim() === "") flush();
+      else entry.push(line.trim());
+      continue;
+    }
+
+    if (!group) lead.push(line);
+  }
+  flush();
+
+  const parsed = groups
+    .map((g) => ({ name: g.name, entries: g.entries.map(summarizeEntry) }))
+    .filter((g) => g.entries.length > 0);
+
+  const leadText = lead.join("\n").trim();
+  if (parsed.length === 0 && !leadText) return null;
+
+  return {
+    lead: leadText,
+    groups: parsed,
+    total: parsed.reduce((n, g) => n + g.entries.length, 0),
+    compareUrl,
+    sections: ["the GitHub Release body"],
+  };
+}
+
+/**
+ * Pick the source and return the notes, or null when neither has anything.
+ * The precedence — release body, then changelog — is the whole point; see the
+ * module header.
+ */
+export function chooseNotes({ releaseBody, changelog, version }) {
+  return (
+    parseReleaseBody(releaseBody) ??
+    (changelog ? collectRelease(changelog, version) : null)
+  );
+}
+
+// ── Entry summarizing ───────────────────────────────────────────────────────
+
+/**
+ * Abbreviations whose trailing period does not end a sentence. Without these,
+ * "e.g. the daemon" splits after "e.g." and the headline is two words long.
+ */
+const ABBREVIATIONS = new Set(["e.g", "i.e", "cf", "vs", "etc", "approx", "incl", "no", "fig"]);
+
+/**
+ * The first sentence of `text`, with inline code spans protected.
+ *
+ * Periods are everywhere in this changelog that are not sentence ends —
+ * `policy-evaluator.ts`, `1.0.15`, `~/.failproofai/run/` — and almost all of
+ * them sit inside backticks, so the spans are masked before the search and
+ * restored after. What is left is a period (or `!`/`?`) that may be followed by
+ * closing emphasis (`**`, `_`, `` ` ``, `)`, `"`), then whitespace, then the
+ * start of a new sentence. Entries in this file routinely open with a whole
+ * bolded sentence — `**Stop the digest shipping secrets verbatim.**` — which is
+ * why the closing markers have to be consumed before the boundary, not after.
+ *
+ * A candidate boundary that would leave less than `MIN_SENTENCE` characters is
+ * skipped: it is nearly always a false positive on an initial or a stray
+ * abbreviation, and a six-character headline is worse than a long one. The
+ * length is measured on the RESTORED text — an entry opening with a long path
+ * in backticks masks down to three characters, and measuring the masked form
+ * rejected every real boundary after it.
+ */
+export function firstSentence(text) {
+  const MIN_SENTENCE = 12;
+  const spans = [];
+  const masked = String(text).replace(/`[^`]*`/g, (m) => {
+    spans.push(m);
+    return `${spans.length - 1}`;
+  });
+  const restore = (s) => s.replace(/(\d+)/g, (_, i) => spans[Number(i)]);
+
+  const boundary = /([.!?])([*_`)"'\]]*)(\s+)/g;
+  let match;
+  while ((match = boundary.exec(masked)) !== null) {
+    const end = match.index + match[1].length + match[2].length;
+    const candidate = restore(masked.slice(0, end)).trim();
+    if (candidate.length < MIN_SENTENCE) continue;
+
+    const before = masked.slice(0, match.index);
+    const word = /([A-Za-z.]+)$/.exec(before)?.[1] ?? "";
+    if (ABBREVIATIONS.has(word.toLowerCase().replace(/\.$/, ""))) continue;
+
+    // A lone capital before the period is an initial ("J. Smith"), not an end.
+    if (/(^|\s)[A-Z]$/.test(before)) continue;
+
+    return candidate;
+  }
+
+  return restore(masked).trim();
+}
+
+/** Cut at a word boundary, so a headline never ends mid-token. */
+export function truncate(text, max) {
+  const s = String(text);
+  if (s.length <= max) return s;
+  const cut = s.slice(0, max - 1);
+  const space = cut.lastIndexOf(" ");
+  return `${(space > max * 0.6 ? cut.slice(0, space) : cut).replace(/[\s,;:—–-]+$/, "")}…`;
+}
+
+/**
+ * One entry — from either source — into `{headline, pr}`.
+ *
+ * The PR reference is read off the END of the WHOLE entry before the first
+ * sentence is taken, because that is where it lives in both formats and the
+ * first sentence never contains it. Three trailing forms are recognised: the
+ * changelog's `(#123)`, GitHub's `by @someone in <pull url>`, and a bare pull
+ * URL. `(#PR)` and other non-numeric placeholders yield no link rather than a
+ * broken one.
+ */
+export function summarizeEntry(entry) {
+  let text = String(entry).trim();
+  let pr = null;
+
+  const attribution = /\s+by\s+@[\w-]+\s+in\s+https?:\/\/\S*?\/pull\/(\d+)\/?\s*$/i.exec(text);
+  const bareUrl = /\s+https?:\/\/\S*?\/pull\/(\d+)\/?\s*$/i.exec(text);
+  const parenRef = /\(#(\d+)\)\s*$/.exec(text);
+
+  if (attribution) {
+    pr = Number(attribution[1]);
+    text = text.slice(0, attribution.index);
+  } else if (bareUrl) {
+    pr = Number(bareUrl[1]);
+    text = text.slice(0, bareUrl.index);
+  } else if (parenRef) {
+    pr = Number(parenRef[1]);
+  }
+
+  return {
+    headline: truncate(firstSentence(text.replace(/\s*\(#[^)]*\)\s*$/, "")), MAX_HEADLINE),
+    pr,
+  };
+}
+
+// ── Payload ─────────────────────────────────────────────────────────────────
+
+/**
+ * Order the groups for display: the deprioritized ones last, everything else in
+ * the order the changelog introduced it. Returns `[shown, hiddenCount]`.
+ */
+function selectGroups(groups) {
+  const kept = groups.filter((g) => g.entries.length > 0);
+  const primary = kept.filter((g) => !DEPRIORITIZED_GROUPS.has(g.name.toLowerCase()));
+  // A release that is nothing but dependency bumps still gets a description.
+  const shown = primary.length > 0 ? primary : kept;
+  const hidden = kept
+    .filter((g) => !shown.includes(g))
+    .reduce((n, g) => n + g.entries.length, 0);
+  return [shown, hidden];
+}
+
+/**
+ * Assemble the description within `budget`, dropping WHOLE groups rather than
+ * cutting mid-sentence, and always ending on the changelog link.
+ *
+ * The link is the part that must survive: everything above it is a sample, and
+ * the link is what makes the sample honest. The first version of this appended
+ * it and then truncated the finished string, which on a release the size of
+ * 1.0.0 cut the link off and ended the message on "Stop sending anything about
+ * a…" — a notification that showed a third of a release and pointed nowhere.
+ *
+ * The tail line states what was left out, counted rather than implied, so a
+ * reader can tell "that was everything" from "that was the first six".
+ *
+ * @param {{
+ *   lead: string,
+ *   blocks: Array<{text: string, entries: number}>,
+ *   deprioritized: number,
+ *   changelogUrl: string,
+ *   budget: number,
+ * }} options
+ */
+export function fitDescription({ lead, blocks, deprioritized, changelogUrl, budget }) {
+  const tailFor = (dropped) => {
+    const notes = [];
+    if (dropped > 0) notes.push(`${dropped} more entr${dropped === 1 ? "y" : "ies"}`);
+    if (deprioritized > 0) notes.push(`${deprioritized} dependency update${deprioritized === 1 ? "" : "s"}`);
+    const suffix = notes.length > 0 ? ` — ${notes.join(" and ")} not shown` : "";
+    return `[Full changelog](${changelogUrl})${suffix}`;
+  };
+
+  const total = blocks.reduce((n, b) => n + b.entries, 0);
+  const parts = lead ? [lead] : [];
+  let used = 0;
+
+  for (const block of blocks) {
+    const candidate = [...parts, block.text, tailFor(total - used - block.entries)].join("\n\n");
+    if (candidate.length > budget) break;
+    parts.push(block.text);
+    used += block.entries;
+  }
+
+  const out = [...parts, tailFor(total - used)].join("\n\n");
+  // A lead long enough to crowd out every group is the only way to still be
+  // over budget here, and cutting it is preferable to dropping the link.
+  return out.length <= budget ? out : `${truncate(lead, Math.max(0, budget - tailFor(total).length - 2))}\n\n${tailFor(total)}`;
+}
+
+function renderGroup(group, repo) {
+  const lines = [`**${group.name}**`];
+  for (const entry of group.entries.slice(0, MAX_ENTRIES_PER_GROUP)) {
+    const link = entry.pr ? ` ([#${entry.pr}](https://github.com/${repo}/pull/${entry.pr}))` : "";
+    lines.push(`• ${entry.headline}${link}`);
+  }
+  const rest = group.entries.length - MAX_ENTRIES_PER_GROUP;
+  if (rest > 0) lines.push(`• …and ${rest} more`);
+  return lines.join("\n");
+}
+
+/**
+ * Assemble the webhook body.
+ *
+ * `notes` may be null — a stable release with notes in neither source is a
+ * mistake `publish.yml` checks for at preflight, but the announcement still
+ * goes out without highlights rather than not at all. The release has already
+ * shipped by the time this runs; silence is the worse failure.
+ *
+ * @param {{
+ *   version: string,
+ *   repo: string,
+ *   notes: ReturnType<typeof collectRelease> | ReturnType<typeof parseReleaseBody>,
+ *   roleId?: string | null,
+ *   releaseUrl?: string | null,
+ *   timestamp?: string | null,
+ * }} options
+ */
+export function buildDiscordPayload({
+  version,
+  repo,
+  notes,
+  roleId = null,
+  releaseUrl = null,
+  timestamp = null,
+}) {
+  const tag = `v${version}`;
+  const release = releaseUrl || `https://github.com/${repo}/releases/tag/${tag}`;
+  const mention = roleId ? `<@&${roleId}> ` : "";
+
+  // Only the FIRST paragraph of a stable section's preamble. 1.0.0's runs to
+  // four, and flattening them all into the description produced a wall of prose
+  // above the entries anybody opened the message to read.
+  const lead = notes?.lead ? truncate(notes.lead.split(/\n\s*\n/)[0].replace(/\s*\n\s*/g, " "), MAX_LEAD) : "";
+  const [shown, deprioritized] = notes ? selectGroups(notes.groups) : [[], 0];
+
+  // A generated release body ends on a compare link, which says more than a
+  // blob link to the file does. Without one, the changelog at this tag.
+  const changelogUrl = notes?.compareUrl || `https://github.com/${repo}/blob/${tag}/CHANGELOG.md`;
+  const blocks = shown.map((g) => ({ text: renderGroup(g, repo), entries: g.entries.length }));
+
+  const description = fitDescription({
+    lead,
+    blocks,
+    deprioritized,
+    changelogUrl,
+    budget: LIMIT_DESCRIPTION,
+  });
+
+  const embed = {
+    title: truncate(`failproofai ${tag}`, LIMIT_TITLE),
+    url: release,
+    color: BRAND_COLOR,
+    description,
+    fields: [
+      {
+        name: "Install",
+        value: truncate("```sh\nnpm install -g failproofai\n```", LIMIT_FIELD_VALUE),
+      },
+      {
+        name: "Links",
+        value: truncate(
+          [
+            `[Release notes](${release})`,
+            `[npm](https://www.npmjs.com/package/failproofai/v/${version})`,
+            "[Docs](https://docs.befailproof.ai/)",
+          ].join(" · "),
+          LIMIT_FIELD_VALUE,
+        ),
+      },
+    ],
+    footer: { text: `published from ${repo}` },
+  };
+  if (timestamp) embed.timestamp = timestamp;
+
+  // Discord's 6000-character ceiling is over the SUM of an embed's text fields,
+  // which no single limit above covers. Everything but the description is fixed
+  // and small, so the description is re-fitted against what is left rather than
+  // chopped — same rule as above, the changelog link outlives the cut.
+  const fixed = embed.title.length + embed.footer.text.length + embed.fields.reduce((n, f) => n + f.name.length + f.value.length, 0);
+  if (fixed + embed.description.length > LIMIT_EMBED_TOTAL) {
+    embed.description = fitDescription({
+      lead,
+      blocks,
+      deprioritized,
+      changelogUrl,
+      budget: Math.max(0, LIMIT_EMBED_TOTAL - fixed),
+    });
+  }
+
+  return {
+    content: truncate(`${mention}**failproofai ${tag}** is out.`, LIMIT_CONTENT),
+    // `parse: []` is the guard, not the roles list: it stops @everyone/@here
+    // and any user mention that changelog prose happens to contain. The role
+    // is then re-allowed by id, so exactly one thing in this message pings.
+    allowed_mentions: { parse: [], roles: roleId ? [String(roleId)] : [] },
+    embeds: [embed],
+  };
+}
+
+// ── CLI ─────────────────────────────────────────────────────────────────────
+
+/** `plural(1, "entry", "entries")` -> `"1 entry"`. */
+function plural(n, one, many) {
+  return `${n} ${n === 1 ? one : many}`;
+}
+
+function parseArgs(argv) {
+  const out = {};
+  for (let i = 0; i < argv.length; i++) {
+    const arg = argv[i];
+    if (!arg.startsWith("--")) continue;
+    const eq = arg.indexOf("=");
+    if (eq !== -1) out[arg.slice(2, eq)] = arg.slice(eq + 1);
+    else out[arg.slice(2)] = argv[i + 1] && !argv[i + 1].startsWith("--") ? argv[++i] : "true";
+  }
+  return out;
+}
+
+export function main(argv = process.argv.slice(2), io = console) {
+  const args = parseArgs(argv);
+  const version = args.version;
+  const repo = args.repo || process.env.GITHUB_REPOSITORY || "failproof-ai/failproofai";
+  if (!version) {
+    io.error(
+      "usage: release-announcement.mjs --version <x.y.z> [--repo owner/name] [--notes-file FILE] " +
+        "[--role-id ID] [--release-url URL] [--out FILE] [--check]",
+    );
+    process.exitCode = 1;
+    return null;
+  }
+
+  // The release body arrives as a FILE, never as an argument. It is arbitrary
+  // markdown a human typed into a web form, and putting it on a command line
+  // (or into a shell string) is how a backtick in someone's release notes
+  // becomes command substitution in the release pipeline.
+  const notesPath = args["notes-file"] && args["notes-file"] !== "true" ? args["notes-file"] : null;
+  const releaseBody = notesPath && existsSync(notesPath) ? readFileSync(notesPath, "utf8") : "";
+
+  const changelogPath = args.changelog || resolve(REPO_ROOT, "CHANGELOG.md");
+  const changelog = existsSync(changelogPath) ? readFileSync(changelogPath, "utf8") : "";
+  const notes = chooseNotes({ releaseBody, changelog, version });
+  const source = parseReleaseBody(releaseBody) ? "the GitHub Release body" : `${changelogPath}`;
+
+  // `--check` is publish.yml's preflight gate: it answers "will this release
+  // have anything to announce" before anything is built, which is the one point
+  // in the pipeline where the answer "no" is nearly free to act on.
+  if (args.check === "true") {
+    if (!notes) {
+      io.error(
+        `::error file=CHANGELOG.md::Nothing to announce for ${version}: the GitHub Release body is ` +
+          `empty and CHANGELOG.md has no section for ${version} (or any ${version}-* prerelease). ` +
+          `A stable release announces itself in Discord, so it would post with no release notes in ` +
+          `it. Write the release notes, or add the changelog section, and re-run.`,
+      );
+      process.exitCode = 1;
+      return null;
+    }
+    io.error(`${version} has release notes (${plural(notes.total, "entry", "entries")}, from ${source}).`);
+    return notes;
+  }
+
+  if (!notes) {
+    // A warning rather than an error: the release is already on npm by the
+    // time this runs, and an announcement with no highlights still tells
+    // people the version exists and how to install it.
+    io.error(`::warning::No release notes for ${version} in the release body or ${changelogPath} — announcing without highlights.`);
+  } else {
+    io.error(`Collected ${plural(notes.total, "entry", "entries")} from ${source}.`);
+  }
+
+  const payload = buildDiscordPayload({
+    version,
+    repo,
+    notes,
+    roleId: args["role-id"] && args["role-id"] !== "true" ? args["role-id"] : null,
+    releaseUrl: args["release-url"] && args["release-url"] !== "true" ? args["release-url"] : null,
+    timestamp: args.timestamp && args.timestamp !== "true" ? args.timestamp : new Date().toISOString(),
+  });
+
+  const json = JSON.stringify(payload, null, 2);
+  if (args.out && args.out !== "true") {
+    writeFileSync(args.out, json);
+    io.error(`Wrote ${args.out} (${json.length} bytes).`);
+  } else {
+    io.log(json);
+  }
+  return payload;
+}
+
+if (process.argv[1] && import.meta.url === pathToFileURL(process.argv[1]).href) {
+  main();
+}