msd-core/scripts/gen-adr-index.cjs

#!/usr/bin/env node
'use strict';

/**
 * Generates the ADR index table in docs/adr/README.md from the ADR files
 * themselves, and validates the corpus' lifecycle invariants.
 *
 * The index is a DERIVED artifact: it is regenerated from every
 * `docs/adr/<id>-<slug>.md` on disk, so it cannot silently drift out of date
 * the way a hand-maintained table does. CI re-runs this with `--check` and
 * fails on any diff or invariant violation.
 *
 * Invariants enforced (see docs/adr/README.md "Lifecycle rules"):
 *   1. Every ADR declares `- **Status:** <Token>` with Token in STATUSES.
 *   2. A Superseded/Retired ADR names its successor as a markdown link to the
 *      target file — never a bare "ADR-N", which is ambiguous (ADR-0010 and
 *      ADR-0011 each resolve to more than one file).
 *   3. Supersession is symmetric: if A supersedes B, B records superseded-by A.
 *   4. An ADR whose H1 declares an id must match its filename's id.
 *   5. The committed index equals the generated index.
 *
 * Usage:
 *   node scripts/gen-adr-index.cjs              # print the index to stdout
 *   node scripts/gen-adr-index.cjs --write      # rewrite the index in README.md
 *   node scripts/gen-adr-index.cjs --check      # exit 1 if stale or invalid
 */

const fs = require('node:fs');
const path = require('node:path');

const { ExitError, runMain } = require('./lib/cli-exit.cjs');

const ROOT = path.resolve(__dirname, '..');
const ADR_DIR = path.join(ROOT, 'docs', 'adr');
const README_PATH = path.join(ADR_DIR, 'README.md');

const START_MARKER = '<!-- ADR-INDEX:START — generated by scripts/gen-adr-index.cjs; do not edit by hand -->';
const END_MARKER = '<!-- ADR-INDEX:END -->';

/**
 * The canonical status vocabulary.
 *
 * `Legacy` and `Retired` are deliberately distinct from `Superseded`:
 *   - Superseded — a specific newer ADR replaced this decision. Names it.
 *   - Retired    — the thing this ADR decided no longer exists at all, and no
 *                  single ADR replaced it (e.g. a deleted package boundary).
 *   - Legacy     — frozen historical record, kept for provenance, not a
 *                  pattern to imitate.
 *
 * NOTE: `Legacy` describes a DECISION's standing, not a filename. The
 * `0001-`..`0012-` sequential *naming* era is legacy, but many of those ADRs
 * (e.g. 0002, 0004, 0008, 0009) are Accepted and load-bearing today. Do not
 * conflate the two: grep the naming rule in README.md, not this enum.
 */
const STATUSES = ['Accepted', 'Proposed', 'Superseded', 'Legacy', 'Retired'];

/**
 * Header fields that assert a lifecycle relation.
 *
 * Two DISTINCT relations, deliberately not conflated:
 *
 *   supersedes — the target decision is REPLACED. The target's status becomes
 *     Superseded and it must name this ADR. (ADR-0174 → ADR-0005.)
 *
 *   subsumes — the target decision still HOLDS, but a broader ADR now frames
 *     it; the target keeps its Accepted status and becomes a component of the
 *     larger decision. (ADR-1239/EoS subsumes ADR-1016 "as the declarative
 *     adapter" — the descriptor is still real and still correct.)
 *
 * Both directions are symmetry-checked, but only `supersedes` implies a status
 * change on the target. Collapsing subsumption into supersession would mark
 * four live, load-bearing ADRs as dead — the opposite of the truth.
 */
const RELATION_FIELDS = new Map([
  ['supersedes', { kind: 'supersedes', dir: 'out' }],
  ['superseded by', { kind: 'supersedes', dir: 'in' }],
  ['subsumes', { kind: 'subsumes', dir: 'out' }],
  ['subsumed by', { kind: 'subsumes', dir: 'in' }],
]);

/** Relation kinds and the header field a reader should add to fix each gap. */
const RELATION_SPEC = {
  supersedes: { out: 'Supersedes', in: 'Superseded by' },
  subsumes: { out: 'Subsumes', in: 'Subsumed by' },
};

/**
 * A relation field whose value opens with "nothing"/"none"/"n/a" asserts the
 * absence of the relation, whatever prose follows it.
 */
const NEGATED_RELATION_RE = /^\s*(?:nothing|none|n\/a|[—–-])\s*(?:$|[;,.]|\s)/i;

/**
 * Header fields appear in two shapes across the corpus, both legitimate:
 *   bullet — `- **Status:** Accepted`
 *   table  — `| **Status** | Accepted |`
 * Yield [field, value] for either.
 */
function* headerFields(header) {
  const bullet = /^\s*[-*]\s*\*\*([^*:]+?)(?::)?\*\*\s*(.*)$/gm;
  let m;
  while ((m = bullet.exec(header)) !== null) yield [m[1].trim(), m[2].trim()];

  const row = /^\s*\|\s*\*\*([^*|]+?)(?::)?\*\*\s*\|\s*(.*?)\s*\|\s*$/gm;
  while ((m = row.exec(header)) !== null) yield [m[1].trim(), m[2].trim()];
}

/** Numeric identity of an ADR: "0011" and "11" are the same id. */
function canonicalId(raw) {
  return String(raw).replace(/^0+(?=\d)/, '');
}

/** The documented filename shape: `<issue#>-<kebab-slug>.md`. */
const ADR_FILENAME_RE = /^[0-9]+-[a-z0-9-]+\.md$/;

function adrFiles() {
  return fs
    .readdirSync(ADR_DIR)
    .filter((f) => f.endsWith('.md') && f !== 'README.md')
    .filter((f) => fs.statSync(path.join(ADR_DIR, f)).isFile())
    .sort();
}

/**
 * Split the directory into files this tool can parse and files it cannot.
 *
 * A file without a numeric prefix is not merely unparseable — it is invisible
 * to the index, which is the failure this gate exists to prevent. Report it as
 * a violation naming the convention, rather than crashing on `match(...)[1]`
 * or silently skipping it.
 */
function partitionAdrFiles() {
  const conforming = [];
  const nonConforming = [];
  for (const f of adrFiles()) (ADR_FILENAME_RE.test(f) ? conforming : nonConforming).push(f);
  return { conforming, nonConforming };
}

/** Extract the leading bullet-field header block (everything before the first `##`). */
function headerBlock(text) {
  const body = text.split(/\r?\n/);
  const stop = body.findIndex((l) => /^##\s/.test(l));
  return (stop === -1 ? body : body.slice(0, stop)).join('\n');
}

/**
 * A relation may also be declared as a whole SECTION rather than a header field.
 * ADR-0174 is the exemplar: a `## Supersedes` heading over a table whose first
 * column links each superseded ADR and whose remaining columns explain why.
 * That is the richest form in the corpus and must count — reading only the
 * header block would report the repo's best-documented supersession as missing.
 *
 * Returns { supersedes: [file…], subsumes: [file…] } from matching sections.
 */
const RELATION_SECTION_RE = /^##\s+(Supersedes|Subsumes)\b[^\n]*$/i;

function relationSections(text) {
  const lines = text.split(/\r?\n/);
  const out = { supersedes: [], subsumes: [] };
  for (let i = 0; i < lines.length; i++) {
    const m = lines[i].match(RELATION_SECTION_RE);
    if (!m) continue;
    const kind = m[1].toLowerCase() === 'supersedes' ? 'supersedes' : 'subsumes';
    // Collect until the next heading of any level.
    let j = i + 1;
    const body = [];
    for (; j < lines.length && !/^#{1,6}\s/.test(lines[j]); j++) body.push(lines[j]);
    const chunk = body.join('\n');
    if (NEGATED_RELATION_RE.test(chunk.trim())) continue;
    out[kind].push(...linkedAdrFiles(chunk));
    i = j - 1;
  }
  return out;
}

/** All markdown links to sibling ADR files inside a chunk of text. */
function linkedAdrFiles(text) {
  const out = [];
  const re = /\]\(\s*(?:\.\/)?([0-9]+-[a-z0-9-]+\.md)\s*\)/gi;
  let m;
  while ((m = re.exec(text)) !== null) out.push(m[1]);
  return out;
}

/** Bare `ADR-123` / `ADR 123` mentions that are NOT part of a markdown link. */
function bareAdrRefs(text) {
  const withoutLinks = text.replace(/\[[^\]]*\]\([^)]*\)/g, '');
  const out = [];
  const re = /\bADR[-\s]0*(\d+)\b/gi;
  let m;
  while ((m = re.exec(withoutLinks)) !== null) out.push(canonicalId(m[1]));
  return out;
}

function parseAdr(file) {
  const full = path.join(ADR_DIR, file);
  const text = fs.readFileSync(full, 'utf8');
  const lines = text.split(/\r?\n/);

  // `fileId` is the numeric identity used for comparison ("0011" === "11");
  // `displayId` preserves the filename's prefix exactly as written, because the
  // corpus and its cross-references say "ADR-0001" and "ADR-58", not "ADR-1".
  const rawId = file.match(/^([0-9]+)-/)[1];
  const fileId = canonicalId(rawId);
  const displayId = rawId;

  const h1 = (lines.find((l) => /^#\s/.test(l)) || '').replace(/^#\s+/, '').trim();
  // Title as displayed: drop a leading "ADR-123 — " / "ADR-123: " prefix and a
  // trailing "[Proposed]"-style status bracket, both of which the index renders
  // from structured fields instead.
  const title = h1
    .replace(/^ADR[-\s]?0*\d+\s*(?:[—:-]\s*)?/i, '')
    .replace(/\s*\[(?:Proposed|Accepted|Superseded|Legacy|Retired)\]\s*$/i, '')
    .trim();

  const declaredIdMatch = h1.match(/^ADR[-\s]?0*(\d+)\b/i);
  const declaredId = declaredIdMatch ? canonicalId(declaredIdMatch[1]) : null;

  const header = headerBlock(text);

  let statusRaw = null;
  // relations[kind][dir] = [{field, value, links, bare}]
  const relations = { supersedes: { out: [], in: [] }, subsumes: { out: [], in: [] } };

  for (const [field, value] of headerFields(header)) {
    if (field.toLowerCase() === 'status') {
      if (statusRaw === null) statusRaw = value;
      continue;
    }
    // "Supersedes (generalizes)" / "Subsumes as adapters" → "supersedes" / "subsumes"
    const key = field.toLowerCase().replace(/\s*\([^)]*\)\s*/g, ' ').replace(/\s+as\s+.*$/, '').trim();
    const spec = RELATION_FIELDS.get(key);
    if (!spec) continue;
    // "Supersedes: nothing; amends the ADR-1239 harness" asserts NO relation. Such a
    // field routinely name-drops other ADRs in its prose ("related", "amends", "builds
    // on"); reading those as supersession claims invents links that were never made.
    if (NEGATED_RELATION_RE.test(value)) continue;
    relations[spec.kind][spec.dir].push({ field, value, links: linkedAdrFiles(value), bare: bareAdrRefs(value) });
  }

  const statusToken = statusRaw ? (statusRaw.match(/^([A-Za-z]+)/) || [])[1] : null;

  // A "Superseded by X" written into the Status line itself is the relation.
  if (statusRaw && /^Superseded\b/i.test(statusRaw)) {
    relations.supersedes.in.push({ field: 'Status', value: statusRaw, links: linkedAdrFiles(statusRaw), bare: bareAdrRefs(statusRaw) });
  }

  // `## Supersedes` / `## Subsumes` sections count as OUT claims (ADR-0174's table).
  const sections = relationSections(text);
  for (const kind of ['supersedes', 'subsumes']) {
    if (sections[kind].length === 0) continue;
    relations[kind].out.push({ field: `## ${kind === 'supersedes' ? 'Supersedes' : 'Subsumes'} section`, value: '', links: sections[kind], bare: [] });
  }

  return { file, fileId, displayId, title, declaredId, statusRaw, statusToken, relations, text };
}

function buildCorpus() {
  const { conforming, nonConforming } = partitionAdrFiles();
  const adrs = conforming.map(parseAdr);
  const byFile = new Map(adrs.map((a) => [a.file, a]));
  const byId = new Map();
  for (const a of adrs) {
    if (!byId.has(a.fileId)) byId.set(a.fileId, []);
    byId.get(a.fileId).push(a);
  }
  return { adrs, byFile, byId, nonConforming };
}

function validate({ adrs, byFile, byId, nonConforming }) {
  const errors = [];
  const add = (file, msg) => errors.push(`${file}: ${msg}`);

  for (const f of nonConforming) {
    add(
      f,
      'filename does not match the `<issue#>-<kebab-slug>.md` convention, so it cannot appear in the index. ' +
        'Rename it (see docs/adr/README.md "Naming Convention"), or move it out of docs/adr/ if it is not an ADR.',
    );
  }

  for (const a of adrs) {
    if (!a.statusToken) {
      add(a.file, 'no `- **Status:** <Token>` field found in the header block.');
      continue;
    }
    if (!STATUSES.includes(a.statusToken)) {
      add(a.file, `status "${a.statusToken}" is not one of ${STATUSES.join(' | ')} (full line: "${a.statusRaw}").`);
    }
    if (a.declaredId && a.declaredId !== a.fileId) {
      add(a.file, `H1 declares ADR-${a.declaredId} but the filename says ${a.fileId}. The id must match the filename.`);
    }

    // A Superseded ADR must point at its successor by FILE LINK.
    if (a.statusToken === 'Superseded') {
      const links = a.relations.supersedes.in.flatMap((r) => r.links);
      if (links.length === 0) {
        const bare = a.relations.supersedes.in.flatMap((r) => r.bare);
        add(
          a.file,
          bare.length
            ? `status is Superseded and mentions ADR-${bare.join('/')} but not as a markdown link to the file. ` +
              'Bare ids are ambiguous (ADR-0010 and ADR-0011 each resolve to multiple files) — link the target file.'
            : 'status is Superseded but names no successor. Write `Superseded by [ADR-N](N-slug.md)`.',
        );
      }
    }

    // Every relation link must resolve; every bare id must be linked (and exist).
    for (const kind of Object.keys(RELATION_SPEC)) {
      for (const dir of ['out', 'in']) {
        for (const rel of a.relations[kind][dir]) {
          for (const l of rel.links) {
            if (!byFile.has(l)) add(a.file, `"${rel.field}" links "${l}", which does not exist in docs/adr/.`);
          }
          // The synthetic relation lifted out of the Status line is already covered by
          // the dedicated Superseded check above; reporting it again just duplicates.
          if (rel.field === 'Status') continue;
          // Ids already linked ANYWHERE in this field. A field legitimately
          // reads "…([ADR-58](58-x.md)) — see 'Relation to ADR-58' below": the
          // trailing prose repeats an id that is linked earlier, and flagging
          // that would be noise. But an id that appears ONLY bare is an
          // unchecked claim — and testing `rel.links.length` instead of the
          // specific id silently dropped every bare claim in a field that
          // happened to carry one link.
          const linkedIds = new Set(rel.links.map((l) => (byFile.get(l) || {}).fileId).filter(Boolean));
          for (const b of rel.bare) {
            if (linkedIds.has(b)) continue;
            const candidates = byId.get(b) || [];
            if (candidates.length === 0) {
              add(a.file, `"${rel.field}" names ADR-${b}, which does not exist in docs/adr/. If it is an ISSUE number, write "#${b}" — not "ADR-${b}".`);
            } else {
              add(
                a.file,
                `"${rel.field}" names ADR-${b} without a file link` +
                  (candidates.length > 1 ? ` (ambiguous — resolves to ${candidates.length} files: ${candidates.map((c) => c.file).join(', ')})` : '') +
                  '. Link the target file so the relation is checkable.',
              );
            }
          }
        }
      }
    }
  }

  // Symmetry, per relation kind: A -out-> B  <=>  B -in-> A.
  //
  // Only a RATIFIED (Accepted) claimant is owed the back-link. A Proposed ADR's
  // supersession claim is prospective — it has not taken effect, so stamping its
  // target as superseded would assert something untrue (ADR-857 is Proposed and
  // claims to generalize ADR-0011/ADR-58, both of which are Accepted and live).
  // When such an ADR is ratified to Accepted, this check starts demanding the
  // back-links at exactly the right moment.
  const OPPOSITE = { out: 'in', in: 'out' };
  for (const a of adrs) {
    for (const kind of Object.keys(RELATION_SPEC)) {
      for (const dir of ['out', 'in']) {
        // The ratification guard applies to the OUT direction only: an unratified
        // ADR's claim over someone else is prospective and must not obligate the
        // target. The IN direction is this ADR's statement about ITSELF ("I am
        // superseded by X") and is always owed a reciprocal — guarding it too
        // would skip every Superseded ADR (statusToken !== 'Accepted') and leave
        // dangling one-way claims unchecked, which is the bug this gate exists
        // to catch.
        if (dir === 'out' && a.statusToken !== 'Accepted') continue;
        for (const target of new Set(a.relations[kind][dir].flatMap((r) => r.links))) {
          const b = byFile.get(target);
          if (!b) continue;
          const back = new Set(b.relations[kind][OPPOSITE[dir]].flatMap((r) => r.links));
          if (back.has(a.file)) continue;
          const needed = RELATION_SPEC[kind][OPPOSITE[dir]];
          const claim = dir === 'out' ? `it ${kind} this ADR` : `it is ${kind === 'supersedes' ? 'superseded' : 'subsumed'} by this ADR`;
          add(
            target,
            `${a.file} declares ${claim}, but this ADR does not record it. ` +
              `Add \`- **${needed}:** [ADR-${a.displayId}](${a.file})\` so a reader of THIS file learns the decision moved on.`,
          );
        }
      }
    }
  }

  return errors;
}

const GROUPS = [
  {
    heading: 'Active decisions',
    blurb: 'These govern the system as it stands. Cite these.',
    match: (a) => a.statusToken === 'Accepted',
  },
  {
    heading: 'Proposed',
    blurb: 'Decided in principle, not yet ratified. Do not cite as settled architecture.',
    match: (a) => a.statusToken === 'Proposed',
  },
  {
    heading: 'Superseded, Retired, and Legacy',
    blurb: 'Historical record. **Do not follow these** — each names what replaced it, or why it was retired.',
    match: (a) => ['Superseded', 'Retired', 'Legacy'].includes(a.statusToken),
  },
];

/**
 * Render ADR-authored text (a title) into a markdown table cell.
 *
 * Three hazards, all from text this script does not control:
 *   - `|` would split the cell and corrupt the row.
 *   - An HTML comment would be emitted verbatim into README.md. A title
 *     containing the END marker relocates it, so the NEXT `--write` splices
 *     against the wrong boundary and silently eats the rest of the file.
 *     Escaping `<`/`>` makes a comment sequence unformable, which also blocks
 *     any other HTML injected through a title.
 *   - A backslash is markdown's own escape character, so it MUST be escaped
 *     first. Escaping `|` → `\|` without it turns the input `\|` into `\\|`,
 *     which markdown reads as a literal backslash followed by an UNESCAPED
 *     pipe — re-opening the cell break the pipe escape exists to prevent.
 *     Order is load-bearing: backslash first, then everything that emits one.
 */
function cellText(text) {
  return String(text)
    .replace(/\\/g, '\\\\')
    .replace(/\|/g, '\\|')
    .replace(/</g, '&lt;')
    .replace(/>/g, '&gt;')
    .replace(/\r?\n/g, ' ')
    .trim();
}

function linkCell(files, byFile) {
  if (files.length === 0) return '—';
  return files.map((l) => `[ADR-${(byFile.get(l) || {}).displayId || '?'}](${l})`).join(', ');
}

function renderIndex(corpus) {
  const { byFile } = corpus;
  const out = [START_MARKER, ''];

  for (const g of GROUPS) {
    const rows = corpus.adrs.filter(g.match).sort((x, y) => Number(x.fileId) - Number(y.fileId));
    if (rows.length === 0) continue;

    out.push(`### ${g.heading} (${rows.length})`, '', g.blurb, '');
    const isHistorical = g.heading.startsWith('Superseded');
    // "Read first" points at the broader ADR that now frames this one. It is how a
    // reader of a still-Accepted component decision (e.g. the runtime descriptor)
    // discovers the wider decision that reframed it (e.g. EoS) instead of assuming
    // the component IS the architecture.
    out.push(
      isHistorical ? '| ADR | Title | Status | Replaced by |' : '| ADR | Title | Status | Read first |',
      isHistorical ? '|-----|-------|--------|-------------|' : '|-----|-------|--------|------------|',
    );
    for (const a of rows) {
      const cells = [`[ADR-${a.displayId}](${a.file})`, cellText(a.title), a.statusToken];
      cells.push(
        isHistorical
          ? linkCell([...new Set(a.relations.supersedes.in.flatMap((r) => r.links))], byFile)
          : linkCell([...new Set(a.relations.subsumes.in.flatMap((r) => r.links))], byFile),
      );
      out.push(`| ${cells.join(' | ')} |`);
    }
    out.push('');
  }

  out.push(
    `_${corpus.adrs.length} ADRs. Generated by \`scripts/gen-adr-index.cjs\` — run \`--write\` after adding or restatusing an ADR._`,
    '',
    END_MARKER,
  );
  return out.join('\n');
}

function spliceIntoReadme(readme, index) {
  const start = readme.indexOf(START_MARKER);
  const end = readme.indexOf(END_MARKER);
  if (start === -1 || end === -1) {
    throw new ExitError(
      1,
      `docs/adr/README.md is missing the index markers.\nExpected:\n  ${START_MARKER}\n  ${END_MARKER}\n`,
    );
  }
  return readme.slice(0, start) + index + readme.slice(end + END_MARKER.length);
}

function main() {
  const [, , flag] = process.argv;

  const corpus = buildCorpus();
  const errors = validate(corpus);

  if (errors.length > 0 && flag !== '--write') {
    process.stderr.write(
      `docs/adr/ has ${errors.length} lifecycle violation(s).\n` +
        'See docs/adr/README.md "Lifecycle rules" for the contract.\n\n',
    );
    for (const e of errors) process.stderr.write(`  ✗ ${e}\n`);
    process.stderr.write('\n');
    throw new ExitError(1);
  }

  const index = renderIndex(corpus);

  if (flag === '--check') {
    const readme = fs.readFileSync(README_PATH, 'utf8');
    const expected = spliceIntoReadme(readme, index);
    if (expected !== readme) {
      process.stderr.write(
        'docs/adr/README.md index is stale. Run:\n  node scripts/gen-adr-index.cjs --write\n\n',
      );
      throw new ExitError(1);
    }
    process.stdout.write(`docs/adr/README.md index is up to date (${corpus.adrs.length} ADRs).\n`);
  } else if (flag === '--write') {
    const readme = fs.readFileSync(README_PATH, 'utf8');
    fs.writeFileSync(README_PATH, spliceIntoReadme(readme, index));
    process.stdout.write(`Wrote ADR index into ${README_PATH} (${corpus.adrs.length} ADRs).\n`);
    if (errors.length > 0) {
      process.stderr.write(`\n${errors.length} lifecycle violation(s) remain — --check will fail:\n\n`);
      for (const e of errors) process.stderr.write(`  ✗ ${e}\n`);
    }
  } else {
    process.stdout.write(index + '\n');
  }
}

runMain(main);