k0-board 0.4.0 → 0.5.0

package/server/mirror.js

@@ -0,0 +1,1048 @@
+import fs from 'node:fs'
+import path from 'node:path'
+import * as db from './db.js'
+import * as settings from './settings.js'
+import { matter } from '../web/md.js'
+
+// ── The readable copy ────────────────────────────────────────────────────────
+// The board's database is the source of truth. `.k0/` is the same backlog written down where a
+// person — and git — can read it: one file per epic, one per story, rewritten whole whenever the
+// item changes. No diffing, no merging of what changed: a section is printed from the database
+// every time, so a file can never drift into saying something the board does not.
+//
+// The files are not a cache. They are the only copy of this work that survives a lost database,
+// which is what `importRepo` is for, and it is why the format has to be parsable as well as
+// readable: everything `writeStory` prints, `importRepo` puts back.
+//
+// Two consequences of "readable" run through the whole file and are worth knowing before reading
+// any single function. The first is that people edit these files: whatever this module does not
+// print, it does not touch, header keys included. The second is that what it does print is
+// markdown somebody wrote, headings and all — see `shiftHeadings`.
+//
+// Nothing here throws at its caller. A story is renamed on the board, the row is written, and the
+// copy on disk is a copy: a read-only mount or a full disk gives back a sentence, never an
+// exception that marks the edit as failed after the database has already taken it.
+//
+// This module formats and it parses. What state a story is in, when it may go to `Done`, which
+// story to pick up next — none of that is decided here.
+
+const K0 = '.k0'
+
+/** The sections this module knows how to print. Anything else in the file is somebody's, and stays. */
+const STORY_SECTIONS = ['Why', 'Prompt', 'Discussion', 'Decisions', 'Plan', 'Verification', 'Log']
+const EPIC_SECTIONS = ['Why', 'Discussion', 'Decisions', 'Stories']
+
+// The header keys this module owns, lowercased the way `matter` hands them back. A key somebody
+// added by hand — `owner:`, `ticket:` — is theirs and is printed back untouched, for the same
+// reason a section they wrote is: a header that quietly loses a line is a header nobody writes in
+// twice.
+const STORY_KEYS = ['key', 'order', 'epic', 'parent', 'state', 'color', 'starred', 'depends_on', 'sessions', 'lang',
+  'created', 'updated', 'completed']
+const EPIC_KEYS = ['key', 'order', 'state', 'lang', 'created', 'updated']
+
+// ── Where it may write ───────────────────────────────────────────────────────
+
+const real = (p) => {
+  try {
+    return fs.realpathSync(p)
+  } catch {
+    return null
+  }
+}
+
+/**
+ * This repository's `.k0/`, or null when there is nowhere to write. Three separate refusals.
+ *
+ * A repository that has been moved or deleted is not an error: there is simply nowhere to write.
+ *
+ * A path that is not a git repository is one k0 was handed rather than one it found. A story's
+ * `project_path` arrives on a request, and `POST /api/backlog/story { project_path: '/Users/me' }`
+ * is a well-formed request: without this, a mistyped path is a `.k0/` folder appearing in a
+ * directory nobody asked for. The promise this module makes is not "inside `<repo>/.k0/`" — that
+ * is trivially true whatever `<repo>` says — it is "nowhere the user did not put a repository".
+ *
+ * And a `.k0` that is a symbolic link somewhere else is a repository that has been rigged. `.k0/`
+ * is committed, so it travels with a clone; a `.k0` pointing at `..` would turn every write in
+ * this file into a write into the working tree, and every delete into a delete out of it.
+ */
+function where(repoPath) {
+  const base = real(repoPath)
+  if (!base) return null
+  try {
+    if (!fs.statSync(base).isDirectory()) return null
+    // A worktree's `.git` is a file holding a path rather than a directory, so this asks only
+    // whether it is there.
+    if (!fs.existsSync(path.join(base, '.git'))) return null
+  } catch {
+    return null
+  }
+  const dir = path.join(base, K0)
+  const found = real(dir)
+  return found && found !== dir ? null : dir
+}
+
+const there = (repoPath) => !!where(repoPath)
+
+/**
+ * Every path this module touches is resolved through here, and anything that lands outside
+ * `<repo>/.k0/` throws instead of being written.
+ *
+ * A comment saying "only writes inside .k0/" would be worth nothing: the paths are built from
+ * titles, and a title is whatever the user typed. So it is checked twice, the way `files.js`
+ * checks the paths the viewer is asked for. First the path as written — that is where a `..` in a
+ * slug lives. Then where it really points, because a lexical check reads a symbolic link as a
+ * name: `.k0/stories/K42-old.md` linked at `~/.ssh/config` passes any amount of `startsWith` and
+ * is then renamed by `place` and overwritten by `put`. A path that is not there yet has nowhere
+ * else it could be pointing, and is allowed: that one is about to be created.
+ */
+function inside(repoPath, ...parts) {
+  const dir = where(repoPath)
+  if (!dir) throw new Error(`mirror.js has nowhere to write: ${repoPath} is not a repository it may put a ${K0}/ in`)
+  const target = path.resolve(dir, ...parts)
+  if (target !== dir && !target.startsWith(dir + path.sep)) {
+    throw new Error(`mirror.js will not touch ${target}: it is outside ${dir}`)
+  }
+  const found = real(target)
+  if (found && found !== target && !found.startsWith(dir + path.sep)) {
+    throw new Error(`mirror.js will not touch ${target}: it leads to ${found}, outside ${dir}`)
+  }
+  return target
+}
+
+/**
+ * With the backlog switched off k0 is the board it has always been, and a repository must come
+ * out of that untouched — no folder appearing in a working tree the user did not ask for. The
+ * switch is read here rather than trusted to every caller, because it only takes one that forgot.
+ */
+const enabled = () => settings.read().backlog
+
+/** What went wrong, without the error class in front of it: this ends up on somebody's screen. */
+const sentence = (err) => String(err?.message ?? err ?? '').replace(/^[A-Z]+:\s*/, '').trim()
+
+/**
+ * The one place a filesystem failure becomes a sentence.
+ *
+ * A repository on a read-only mount, a checkout whose permissions somebody tightened, a full
+ * disk: the board has already taken the edit, and the request that made it must not come back
+ * carrying `EACCES` as though the edit had not happened. `worktree.js` does the same thing with
+ * git's words, for the same reason.
+ */
+function attempt(fallback, fn) {
+  try {
+    return { ...fallback, ...fn(), error: null }
+  } catch (err) {
+    return { ...fallback, error: `The ${K0}/ copy could not be written: ${sentence(err)}` }
+  }
+}
+
+function put(repoPath, rel, text) {
+  const file = inside(repoPath, rel)
+  fs.mkdirSync(path.dirname(file), { recursive: true })
+  fs.writeFileSync(file, text)
+  return file
+}
+
+// ── Names ────────────────────────────────────────────────────────────────────
+
+/**
+ * The title, lowercased, everything that is not a letter or a digit collapsed to one dash,
+ * trimmed, and cut at forty characters — a file name a person can read at a glance in a diff.
+ *
+ * It is deliberately not unique: two stories called "fix the API" slug the same. The key in
+ * front of it is what keeps them apart, and the key is the only part of the name anything ever
+ * matches on. `worktree.js` imports this rather than keeping a copy: a worktree directory and a
+ * `.k0/` file are named for the same story, and two rules that were once the same rule drift.
+ */
+export function slug(title) {
+  return String(title ?? '')
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, '-')
+    .replace(/^-+|-+$/g, '')
+    .slice(0, 40)
+    .replace(/-+$/, '')
+}
+
+const fileName = (key, title) => {
+  const s = slug(title)
+  return s ? `${key}-${s}.md` : `${key}.md`
+}
+
+/**
+ * Where this item's file goes, moving the one that is already there if the title has changed.
+ *
+ * Renaming rather than writing a second file is the whole point: two files claiming to be K42,
+ * one of them stale, is exactly the disagreement `.k0/` exists to make impossible — and the
+ * stale one is the one somebody would read.
+ */
+function place(repoPath, folder, key, title) {
+  const want = fileName(key, title)
+  const dir = inside(repoPath, folder)
+  const mine = new RegExp(`^${key}(-.*)?\\.md$`)
+  if (fs.existsSync(dir)) {
+    for (const name of fs.readdirSync(dir)) {
+      if (name === want || !mine.test(name)) continue
+      fs.renameSync(inside(repoPath, folder, name), inside(repoPath, folder, want))
+    }
+  }
+  return `${folder}/${want}`
+}
+
+// ── The order ────────────────────────────────────────────────────────────────
+// `order` is the row's own `sort_hint`, written down because it is the one thing about a file
+// that the folder itself cannot say: a directory listing is alphabetical, and the board's order
+// is whatever it was dragged into. It is copied and not computed, so there is no second rule here
+// to drift from the first — and on the way back in it is what puts the files in the order the
+// board had them.
+
+/** An epic's own stories, without the tasks they were split into: those hang off their story. */
+const topLevel = (stories) => stories.filter((s) => !s.parent_story_id)
+
+// ── Headings inside a section ────────────────────────────────────────────────
+// The Why, a prompt, an approved plan and the answer to a round are all markdown somebody wrote,
+// and markdown people write has headings in it — this repository's own plan format opens with
+// `## TLDR`. Printed as it stands, that line is indistinguishable from `## Plan`: the parser cuts
+// the file there, the plan comes back as the paragraph above its first heading, and the rest of
+// it re-appears at the bottom as though a person had written it, in pieces, out of order.
+//
+// So a heading inside a section is printed one level deeper than it was written, which is also
+// where it belongs: `## TLDR` under `## Plan` is a third-level heading of the document it is
+// actually part of. Reading lifts it back. A first-level `#` is left where it is — it is not what
+// cuts the file, and pushing it down is what would turn it into something that does.
+//
+// One thing does not survive: a sixth-level heading has no seventh level to go to and comes back
+// as a fifth. A plan with six levels of heading has troubles this file cannot fix.
+//
+// Fenced code is never touched, here or in `split`. A `## ` inside a fence is a line of somebody's
+// example, and cutting the file there would take their code block apart.
+
+const FENCE = /^\s*(```+|~~~+)/
+const HEADING = /^##\s+(.+?)\s*$/
+const DEEPER = /^(#{2,6})(\s.*)$/
+const H1 = /^#\s+\S/
+
+/** Each line, with whether it is inside a fenced code block. Fence lines count as fenced. */
+function* walk(text) {
+  let fence = null
+  for (const line of String(text ?? '').split('\n')) {
+    const f = FENCE.exec(line)
+    if (!f) {
+      yield { line, fenced: fence !== null }
+      continue
+    }
+    if (fence === null) fence = f[1][0]
+    else if (f[1][0] === fence) fence = null
+    yield { line, fenced: true }
+  }
+}
+
+function shiftHeadings(text, by) {
+  const out = []
+  for (const { line, fenced } of walk(text)) {
+    const h = fenced ? null : DEEPER.exec(line)
+    out.push(h ? '#'.repeat(Math.min(6, Math.max(2, h[1].length + by))) + h[2] : line)
+  }
+  return out.join('\n')
+}
+
+const deepen = (text) => shiftHeadings(text, 1)
+const lift = (text) => shiftHeadings(text, -1)
+
+// ── Printing ─────────────────────────────────────────────────────────────────
+
+/** Blocks separated by one blank line, one newline at the end, nothing empty in between. */
+const page = (blocks) =>
+  blocks
+    .map((b) => (b == null ? '' : String(b).replace(/\s+$/, '')))
+    .filter((b) => b !== '')
+    .join('\n\n') + '\n'
+
+const frontMatter = (pairs) =>
+  ['---', ...pairs.filter(([, v]) => v !== null && v !== undefined).map(([k, v]) => `${k}: ${v}`), '---'].join('\n')
+
+const section = (name, text) => {
+  const body = String(text ?? '').trim()
+  return body ? `## ${name}\n\n${body}` : null
+}
+
+/** A section whose content is markdown somebody wrote, so its own headings go one level down. */
+const prose = (name, text) => section(name, deepen(String(text ?? '').trim()))
+
+const at = (ms) => new Date(Number(ms) || 0).toISOString()
+const day = (ms) => at(ms).slice(0, 10)
+const stamp = (ms) => (Number(ms) ? at(ms) : null)
+
+const roundsBlock = (rounds) =>
+  rounds
+    .map((r) => {
+      const head = r.estimated_total ? `### Round ${r.n} of ${r.estimated_total}` : `### Round ${r.n}`
+      const q = `**Q** ${deepen(r.question || '')}`.trimEnd()
+      const a = `**A** ${deepen(r.answer || '')}`.trimEnd()
+      return [head, q, a].join('\n')
+    })
+    .join('\n\n')
+
+/**
+ * The decisions this file is about, each under the name the counter-check calls it by.
+ *
+ * A story's list holds its epic's decisions as well as its own, because they are what the story
+ * is actually held to and because the `## Verification` section below already prints verdicts
+ * about them — `K7·D3 violated`. Without them here, the only plain-text record of the work names
+ * a broken rule and never says anywhere what that rule was, and the reader has to know to go and
+ * open the epic's file to find out.
+ *
+ * They are named and not copied. `n` numbers per owner, so a bare `D3` would be the story's own
+ * third decision, and a file claiming the epic's rule as the story's is the fork that superseding
+ * one of them would then have to be done twice to undo. `parseDecisions` reads `- **D<n>**` and
+ * nothing else, so on the way back in these lines are skipped and the story is rebuilt with
+ * exactly the decisions it owns.
+ */
+const decisionsBlock = (decisions, labelOf) =>
+  decisions
+    .map((d) => {
+      const by = d.superseded_by ? labelOf(d.superseded_by) : null
+      const from = d.from ? ` *(from ${d.from})*` : ''
+      return `- **${d.label}** ${d.text}${from}${by ? ` *(superseded by ${by})*` : ''}`
+    })
+    .join('\n')
+
+const MARK = { pass: 'x', fail: '!', skip: '-', todo: ' ' }
+
+// The evidence goes on a line of its own, indented under the item, and not after a dash on the
+// same line. A check is a sentence, the house writes sentences with em dashes in them, and there
+// is no rule — cut at the first dash, cut at the last — that can tell a sentence's own dash from
+// the one in front of the evidence. Underneath it there is nothing to tell apart.
+const checksBlock = (items) =>
+  items
+    .map((c) => {
+      const line = `- [${MARK[c.state] ?? ' '}] ${c.n}. ${String(c.text ?? '').replace(/\s*\n\s*/g, ' ').trim()}`
+      const why = String(c.evidence ?? '').trim()
+      return why ? `${line}\n  ${why.split('\n').join('\n  ')}` : line
+    })
+    .join('\n')
+
+/**
+ * The verdicts, run by run, then the checklist.
+ *
+ * The label comes from the effective set and never from the check row's number alone. A story
+ * inherits its epic's decisions and is counter-checked against them one by one, so a verdict may
+ * be about `K7·D3` — the epic's third decision — while the story has a D3 of its own saying
+ * something else entirely. `D3` in the file would put that verdict on the wrong sentence on the
+ * way back in, and the file would still look right.
+ */
+function verificationBlock(storyId, checks) {
+  const label = new Map(db.effectiveDecisions(storyId).map((d) => [d.id, d.label]))
+  const runs = new Map()
+  for (const c of db.listDecisionChecks(storyId)) {
+    // A decision deleted since the run was made has no sentence left to be a verdict about.
+    if (!label.has(c.decision_id)) continue
+    if (!runs.has(c.run)) runs.set(c.run, [])
+    runs.get(c.run).push(c)
+  }
+  const blocks = [...runs.keys()]
+    .sort((a, b) => a - b)
+    .map((run) => {
+      const lines = runs.get(run).map((c) => {
+        // One verdict, one line. `parseVerification` reads these line by line and throws away
+        // anything that is not indented under a checklist item, so a `violated` whose evidence
+        // runs to a second sentence on a second line — which is what "the file and line, and one
+        // sentence saying how" regularly arrives as — would come back cut off at the newline, and
+        // the person rebuilding from `.k0/` would read half the reason his story was blocked.
+        const why = String(c.evidence ?? '').replace(/\s*\n\s*/g, ' ').trim()
+        return `- **${label.get(c.decision_id)}** ${c.verdict}${why ? ` — ${why}` : ''}`
+      })
+      return [`### Run ${run}`, ...lines].join('\n')
+    })
+  if (checks.length) blocks.push(checksBlock(checks))
+  return blocks.join('\n\n')
+}
+
+// A log entry can be a paragraph. The continuation lines are indented so the entry stays one
+// bullet to the eye and one entry to the parser, which would otherwise read every line after the
+// first as an entry of its own with no date. A blank line inside an entry is printed blank rather
+// than as two spaces: every editor strips trailing whitespace, so a paragraph break written that
+// way is one the next save deletes.
+const logBlock = (entries) =>
+  entries
+    .map((e) => {
+      const who = e.session_id ? ` · session ${e.session_id}` : ''
+      const [first, ...rest] = String(e.text ?? '').split('\n')
+      const text = [first, ...rest.map((l) => (l.trim() ? `  ${l}` : ''))].join('\n')
+      return `- ${day(e.at)}${who} — ${text}`
+    })
+    .join('\n')
+
+/** What is already in the file, or an empty one when there is nothing there yet to keep. */
+function existing(file) {
+  try {
+    return matter(fs.readFileSync(file, 'utf8'))
+  } catch {
+    return { data: {}, body: '' }
+  }
+}
+
+/**
+ * The parts of the file that are nobody's business but the person who wrote them.
+ *
+ * `.k0/` is committed, so it gets edited by hand: notes, links, a paragraph somebody wanted next
+ * to the story, a `ticket:` line in the header. Rewriting the file from the database would take
+ * those with it, and after the first time nobody would ever write anything there again. What this
+ * module prints, it owns; everything else comes back, unread.
+ */
+const keptSections = (prev, known) =>
+  split(prev.body)
+    .sections.filter((s) => !known.includes(s.heading))
+    .map((s) => [`## ${s.heading}`, ...s.lines].join('\n').replace(/\s+$/, ''))
+
+const keptKeys = (prev, known) => Object.entries(prev.data).filter(([k]) => !known.includes(k))
+
+/** Paragraphs, so the standfirst can be one sentence and the notes under it can be somebody's. */
+const paragraphs = (linesIn) => linesIn.join('\n').trim().split(/\n\s*\n/).filter((p) => p.trim())
+
+const afterHeading = (lead) => lead.slice(lead.findIndex((l) => H1.test(l)) + 1)
+
+/**
+ * Everything above the first heading except the standfirst, which is printed from the database.
+ *
+ * `printed` says whether there is a standfirst going out this time, and it is not always: an epic
+ * has none at all, and a story whose description is empty prints none either. Dropping the first
+ * paragraph regardless takes one paragraph of somebody's notes with it on every write, and the
+ * write after that takes the next — a file left alone for a week comes back with the notes gone
+ * and nothing anywhere saying they were ever there.
+ */
+const keptLead = (prev, printed) =>
+  paragraphs(afterHeading(split(prev.body).lead))
+    .slice(printed ? 1 : 0)
+    .join('\n\n')
+
+// ── Writing ──────────────────────────────────────────────────────────────────
+
+/** One story or task, printed whole. `file` is null when there was nowhere to write. */
+export function writeStory(storyId) {
+  const story = db.getStory(storyId)
+  if (!story || !enabled() || !there(story.project_path)) return { file: null, error: null }
+  return attempt({ file: null }, () => {
+    const file = storyFile(story)
+    // The epic's file carries this story's line and its progress count, so it goes stale the moment
+    // a story under it changes state or title. It is one small file: rewrite it rather than invent
+    // a way of knowing whether it needed it.
+    const epic = story.epic_id ? db.getEpic(story.epic_id) : null
+    if (epic) epicFile(epic)
+    return { file }
+  })
+}
+
+function storyFile(story) {
+  const repo = story.project_path
+  ensureReadme(repo)
+  const rel = place(repo, 'stories', story.key, story.title)
+  const prev = existing(inside(repo, rel))
+  const epic = story.epic_id ? db.getEpic(story.epic_id) : null
+  const parent = story.parent_story_id ? db.getStory(story.parent_story_id) : null
+  // The epic's first, because they were taken first and they shaped the ones under them — the
+  // same order `db.effectiveDecisions` puts them in for the plan and for the counter-check.
+  const inherited = epic ? db.epicDecisions(epic.id).map((d) => ({ ...d, from: epic.key })) : []
+  const decisions = [...inherited, ...db.listDecisions(story.id).map((d) => ({ ...d, label: `D${d.n}` }))]
+  const labelOf = (id) => decisions.find((d) => d.id === id)?.label ?? null
+  const checks = db.listCheckItems(story.id)
+  const sessions = db.listSessions(story.id).map((s) => s.session_id).filter(Boolean)
+
+  const text = page([
+    frontMatter([
+      ['key', story.key],
+      ['order', story.sort_hint ?? 0],
+      ['epic', epic ? epic.key : null],
+      // Only a task has one, so a story's header stays exactly what the format says it is. It has
+      // to be written: without it a restored task comes back as a top-level story and the work it
+      // was split out of loses its pieces without a word.
+      ['parent', parent ? parent.key : null],
+      ['state', story.state],
+      // The colour is the user's own grouping across a hundred post-its, and it is a column like
+      // any other. Left out of the header it is not a smaller board after a rebuild, it is a board
+      // where every post-it is yellow and nothing says the grouping ever existed.
+      ['color', story.color || 'yellow'],
+      ['starred', story.starred ? 'true' : 'false'],
+      ['depends_on', `[${db.dependenciesOf(story.id).map((d) => d.key).join(', ')}]`],
+      ['sessions', `[${sessions.join(', ')}]`],
+      ['lang', story.lang || ''],
+      // The dates are in the header because the board reads them: how old a story is decides the
+      // warm columns and the `Old` fold. Restored without them, every story in a repository claims
+      // to have been touched at the same instant and the ordering the user reads without thinking
+      // is flat.
+      ['created', stamp(story.created_at)],
+      ['updated', stamp(story.updated_at)],
+      ['completed', stamp(story.completed_at)],
+      ...keptKeys(prev, STORY_KEYS),
+    ]),
+    `# ${story.key} · ${story.title}`,
+    // The line the post-it shows, printed where a document puts its standfirst: no heading of its
+    // own, because it is not a section — it is the story said in one sentence. Anything a person
+    // wrote under it stays under it, and is not read back as part of that sentence.
+    deepen(String(story.description ?? '').trim()),
+    keptLead(prev),
+    `## Why\n\n${deepen(String(story.body ?? '').trim())}`.trimEnd(),
+    prose('Prompt', story.prompt),
+    section('Discussion', roundsBlock(db.listRounds({ storyId: story.id }))),
+    section('Decisions', decisionsBlock(decisions, labelOf)),
+    prose('Plan', story.plan),
+    section('Verification', verificationBlock(story.id, checks)),
+    section('Log', logBlock(db.listLog(story.id))),
+    ...keptSections(prev, STORY_SECTIONS),
+  ])
+  return put(repo, rel, text)
+}
+
+/** One epic: why it exists, the discussion it came from, its decisions, and what is under it. */
+export function writeEpic(epicId) {
+  const epic = db.getEpic(epicId)
+  if (!epic || !enabled() || !there(epic.project_path)) return { file: null, error: null }
+  return attempt({ file: null }, () => ({ file: epicFile(epic) }))
+}
+
+function epicFile(epic) {
+  ensureReadme(epic.project_path)
+  const rel = place(epic.project_path, 'epics', epic.key, epic.title)
+  const prev = existing(inside(epic.project_path, rel))
+  const stories = topLevel(db.storiesOfEpic(epic.id))
+  const done = stories.filter((s) => s.state === 'Done').length
+  const decisions = db.listEpicDecisions(epic.id).map((d) => ({ ...d, label: `D${d.n}` }))
+  const labelOf = (id) => decisions.find((d) => d.id === id)?.label ?? null
+
+  const list = [
+    `${done}/${stories.length} done`,
+    '',
+    ...stories.map((s) => `- **${s.key}** · ${s.title} — ${s.state}`),
+  ].join('\n')
+
+  const text = page([
+    frontMatter([
+      ['key', epic.key],
+      ['order', epic.sort_hint ?? 0],
+      ['state', epic.state],
+      ['lang', epic.lang || ''],
+      ['created', stamp(epic.created_at)],
+      ['updated', stamp(epic.updated_at)],
+      ...keptKeys(prev, EPIC_KEYS),
+    ]),
+    `# ${epic.key} · ${epic.title}`,
+    keptLead(prev),
+    `## Why\n\n${deepen(String(epic.body ?? '').trim())}`.trimEnd(),
+    // An epic is discussed and decided before a single story of it exists, and those decisions
+    // apply to every story under it. They live here for the same reason: there is nowhere else
+    // they could live without being copied, and a copied decision forks the first time it changes.
+    section('Discussion', roundsBlock(db.listRounds({ epicId: epic.id }))),
+    section('Decisions', decisionsBlock(decisions, labelOf)),
+    section('Stories', list),
+    ...keptSections(prev, EPIC_SECTIONS),
+  ])
+  return put(epic.project_path, rel, text)
+}
+
+const README = `# .k0
+
+This folder is k0's readable copy of the backlog for this repository. The board's database is
+the source of truth: every file here is printed from it whenever the epic or the story changes,
+whole, so a file can never end up claiming something the board does not.
+
+- \`epics/\` — one file per epic: why it exists, the rounds of questions it came out of, the
+  decisions taken, and the stories under it.
+- \`stories/\` — one file per story or task: why, the discussion, the decisions, the plan, the
+  counter-check and the log of what happened.
+
+Write in them if you like: anything under a heading k0 does not print is left exactly where you
+put it, and so is anything you add to the header at the top or to the notes under the title.
+Everything else is rewritten on the next change.
+
+Commit the folder. It is the only copy of this work that outlives the database, and k0 reads it
+back when it opens a repository it has no stories for.
+
+That is also why the board's git mark went from a tick to a dot the day this folder appeared:
+it is honest — there is something here that is not committed yet — and committing it puts the
+mark back. If you would rather k0 never wrote any of this, switch the backlog off in its
+settings: no folder is created in any repository after that.
+`
+
+/**
+ * The folder explains itself the moment it appears, whoever made it appear. A `.k0/` turning up
+ * in somebody's working tree with nothing in it but a file called `K42-fix-api.md` is a folder
+ * they will delete. Written once and never again: it is the one file here that is nobody's copy
+ * of anything, so an edit to it is an edit, not drift.
+ */
+function ensureReadme(repoPath) {
+  if (!fs.existsSync(inside(repoPath, 'README.md'))) put(repoPath, 'README.md', README)
+}
+
+/**
+ * Every epic and every story of a repository, and the README once.
+ *
+ * It never deletes a file whose story it cannot find. That sounds like tidiness and is in fact
+ * the one thing that would destroy the backup: a database that has just been lost, or a k0
+ * pointed at an empty one, would sync a repository and take the whole folder with it. Deleting
+ * a story deletes its file, through `removeStory`, where there is something that actually says
+ * the story is gone.
+ */
+export function syncRepo(repoPath) {
+  const out = { epics: 0, stories: 0, error: null }
+  if (!enabled() || !there(repoPath)) return out
+  return attempt(out, () => {
+    const epics = db.listEpics(repoPath)
+    const stories = db.storiesOfProject(repoPath)
+    // Nothing to write is not the same as nothing to do wrong. A sweep runs over every repository
+    // k0 knows about, and writing the README first would put a `.k0/` into all eighteen of somebody's
+    // checkouts the day they updated — folders explaining a backlog that none of those repositories
+    // has. The folder appears when the first story does, and not before.
+    if (!epics.length && !stories.length) return { epics: 0, stories: 0 }
+    ensureReadme(repoPath)
+    for (const epic of epics) epicFile(epic)
+    for (const story of stories) storyFile(story)
+    return { epics: epics.length, stories: stories.length }
+  })
+}
+
+/**
+ * The file of something that is gone. The key is what identifies it and not the name: the title
+ * may well have changed between the last write and the delete, and a file left behind here comes
+ * back as a story of its own the next time a lost database is rebuilt from the folder.
+ */
+function drop(folder, key, repoPath) {
+  const out = { gone: false, error: null }
+  if (!there(repoPath)) return out
+  const k = /^K?(\d+)$/.exec(String(key ?? '').trim())
+  if (!k) return out
+  return attempt(out, () => {
+    const dir = inside(repoPath, folder)
+    if (!fs.existsSync(dir)) return { gone: false }
+    const mine = new RegExp(`^K${k[1]}(-.*)?\\.md$`)
+    let gone = false
+    for (const name of fs.readdirSync(dir)) {
+      if (!mine.test(name)) continue
+      fs.rmSync(inside(repoPath, folder, name))
+      gone = true
+    }
+    return { gone }
+  })
+}
+
+export const removeStory = (key, repoPath) => drop('stories', key, repoPath)
+
+export const removeEpic = (key, repoPath) => drop('epics', key, repoPath)
+
+// ── Reading ──────────────────────────────────────────────────────────────────
+// The way back. A file that has been edited into something this cannot understand is skipped and
+// said so, never thrown over: an import runs on a repository the user has not looked at in a
+// year, and one bad file must not take the other two hundred with it.
+
+/** The body cut into its `##` sections, with whatever came before the first one kept apart. */
+function split(body) {
+  const lead = []
+  const sections = []
+  let cur = null
+  for (const { line, fenced } of walk(body)) {
+    const h = fenced ? null : HEADING.exec(line)
+    if (h) {
+      cur = { heading: h[1], lines: [] }
+      sections.push(cur)
+      continue
+    }
+    ;(cur ? cur.lines : lead).push(line)
+  }
+  return { lead, sections }
+}
+
+const textOf = (linesIn) => linesIn.join('\n').trim()
+
+const sectionText = (sections, name) => {
+  const s = sections.find((x) => x.heading === name)
+  return s ? textOf(s.lines) : ''
+}
+
+/** A section this module printed: its headings were pushed down on the way out and come back up. */
+const proseText = (sections, name) => lift(sectionText(sections, name))
+
+const sectionLines = (sections, name) => sections.find((x) => x.heading === name)?.lines ?? []
+
+const listValue = (v) =>
+  String(v ?? '')
+    .replace(/^\[|\]$/g, '')
+    .split(',')
+    .map((s) => s.trim())
+    .filter(Boolean)
+
+const keyNumber = (v) => {
+  const m = /^K?(\d+)$/.exec(String(v ?? '').trim())
+  return m ? Number(m[1]) : null
+}
+
+const moment = (v) => {
+  const ms = Date.parse(String(v ?? '').trim())
+  return Number.isFinite(ms) ? ms : null
+}
+
+function parseRounds(linesIn) {
+  const rounds = []
+  let cur = null
+  let field = null
+  for (const line of linesIn) {
+    const head = /^###\s+Round\s+(\d+)(?:\s+of\s+(?:about\s+)?(\d+))?/.exec(line)
+    if (head) {
+      cur = { n: Number(head[1]), estimated_total: Number(head[2] || 0), question: [], answer: [] }
+      rounds.push(cur)
+      field = null
+      continue
+    }
+    if (!cur) continue
+    const q = /^\*\*Q\*\*\s?(.*)$/.exec(line)
+    if (q) {
+      field = 'question'
+      cur.question.push(q[1])
+      continue
+    }
+    const a = /^\*\*A\*\*\s?(.*)$/.exec(line)
+    if (a) {
+      field = 'answer'
+      cur.answer.push(a[1])
+      continue
+    }
+    if (field) cur[field].push(line)
+  }
+  return rounds.map((r) => ({ ...r, question: lift(textOf(r.question)), answer: lift(textOf(r.answer)) }))
+}
+
+function parseDecisions(linesIn) {
+  const out = []
+  for (const line of linesIn) {
+    const m = /^-\s+\*\*D(\d+)\*\*\s+(.*?)(?:\s*\*\(superseded by D(\d+)\)\*)?\s*$/.exec(line)
+    if (!m) continue
+    out.push({ n: Number(m[1]), text: m[2].trim(), superseded_by_n: m[3] ? Number(m[3]) : null })
+  }
+  return out
+}
+
+const STATE_OF = { x: 'pass', X: 'pass', '!': 'fail', '-': 'skip', ' ': 'todo', '': 'todo' }
+
+/**
+ * The runs of verdicts and the checklist.
+ *
+ * A verdict carries the epic's key when it is about an inherited decision — `**K7·D3**` — and the
+ * key is kept, not thrown away: the story has a D3 of its own, and the two are different
+ * sentences. `importRepo` resolves each one against the decision it names before recording it.
+ */
+function parseVerification(linesIn) {
+  const runs = new Map()
+  const checks = []
+  let run = null
+  let item = null
+  for (const line of linesIn) {
+    const head = /^###\s+Run\s+(\d+)/.exec(line)
+    if (head) {
+      run = Number(head[1])
+      item = null
+      if (!runs.has(run)) runs.set(run, [])
+      continue
+    }
+    const verdict = /^-\s+\*\*(?:(K\d+)·)?D(\d+)\*\*\s+(kept|violated|na)(?:\s+—\s+(.*))?\s*$/.exec(line)
+    if (verdict && run) {
+      item = null
+      runs.get(run).push({
+        epic_key: verdict[1] || null,
+        decision_n: Number(verdict[2]),
+        verdict: verdict[3],
+        evidence: (verdict[4] || '').trim(),
+      })
+      continue
+    }
+    const box = /^-\s+\[(.?)\]\s+(?:(\d+)\.\s+)?(.*)$/.exec(line)
+    if (box) {
+      item = { text: box[3].trim(), evidence: '', state: STATE_OF[box[1]] ?? 'todo' }
+      checks.push(item)
+      continue
+    }
+    // Whatever is indented under an item is that item's evidence, however many dashes it has in it.
+    if (item && /^\s{2,}\S/.test(line)) item.evidence = [item.evidence, line.trim()].filter(Boolean).join('\n')
+    else if (line.trim()) item = null
+  }
+  return { runs: [...runs.entries()].sort((a, b) => a[0] - b[0]), checks }
+}
+
+/**
+ * The log, dates and all.
+ *
+ * The date is in the file and is put back: a story whose log runs from January to September is the
+ * account of what happened and when, and restored with every line stamped the day of the rebuild
+ * it stops being one. A paragraph break inside an entry is a line with nothing on it — an editor
+ * that strips trailing whitespace takes the indent off it — so blank lines are held rather than
+ * appended, kept when another continuation line follows and dropped when the entry has ended.
+ */
+function parseLog(linesIn) {
+  const out = []
+  let blanks = 0
+  for (const line of linesIn) {
+    const m = /^-\s+(\d{4}-\d{2}-\d{2})(?:\s+·\s+session\s+(\S+))?\s+—\s+(.*)$/.exec(line)
+    if (m) {
+      out.push({ at: moment(`${m[1]}T00:00:00.000Z`), session_id: m[2] || null, text: m[3] })
+      blanks = 0
+      continue
+    }
+    if (!out.length) continue
+    if (!line.trim()) {
+      blanks++
+      continue
+    }
+    if (!/^\s{2,}\S/.test(line)) {
+      blanks = 0
+      continue
+    }
+    out[out.length - 1].text += '\n'.repeat(blanks + 1) + line.trim()
+    blanks = 0
+  }
+  return out
+}
+
+/** The title as written in the heading; the key in front of it is not part of it. */
+const titleFrom = (lead, fallback) => {
+  const line = lead.find((l) => H1.test(l))
+  if (!line) return fallback
+  return line.replace(/^#\s+/, '').replace(/^K\d+\s+·\s+/, '').trim() || fallback
+}
+
+/**
+ * The standfirst: the first paragraph under the heading and only that one.
+ *
+ * The rest of what somebody wrote up there stays where they put it, and `keptLead` prints it back.
+ * Taking the lot would put two paragraphs of notes on a post-it that shows one line, and `/api/backlog`
+ * would return them as the story's summary.
+ */
+const descriptionFrom = (lead) => lift(paragraphs(afterHeading(lead))[0] ?? '')
+
+/** The title a file name can still give back when somebody has deleted the heading. */
+const titleFromName = (name) => name.replace(/^K\d+-?/, '').replace(/\.md$/, '').replace(/-/g, ' ').trim()
+
+function parseFile(name, text) {
+  const { data, body } = matter(text)
+  const key = keyNumber(data.key)
+  if (!key) return { why: 'the front matter has no key' }
+  const { lead, sections } = split(body)
+  const title = titleFrom(lead, titleFromName(name))
+  if (!title) return { why: 'there is no title' }
+  return { key, data, title, lead, sections }
+}
+
+const readFolder = (repoPath, folder) => {
+  const dir = inside(repoPath, folder)
+  if (!fs.existsSync(dir)) return []
+  return fs
+    .readdirSync(dir)
+    .filter((n) => n.endsWith('.md'))
+    .sort()
+    .map((name) => ({ name, text: fs.readFileSync(inside(repoPath, folder, name), 'utf8') }))
+}
+
+/**
+ * The files of one folder, parsed, with the ones that cannot be used said out loud.
+ *
+ * Two files claiming the same key is the case worth the code. It happens for an ordinary reason —
+ * somebody renames `K42-fix-api.md` by hand, so the next write does not recognise it and puts a
+ * fresh `K42-fix-api.md` beside it, or a git merge leaves a copy — and the module's own promise is
+ * that two stories carrying K42 is a board that can no longer say which one anything refers to.
+ * Importing both makes exactly that board: every `parent` and every `depends_on` pointing at K42
+ * would attach to whichever file happened to sort later. The first one wins, in the order the
+ * folder reads, and the other is named in the report so the user can see what was left out.
+ */
+function usable(repoPath, folder, report) {
+  const seen = new Map()
+  const out = []
+  for (const { name, text } of readFolder(repoPath, folder)) {
+    const parsed = parseFile(name, text)
+    if (parsed.why) {
+      report.skipped.push({ file: `${folder}/${name}`, why: parsed.why })
+      continue
+    }
+    const clash = seen.get(parsed.key)
+    if (clash) {
+      report.skipped.push({ file: `${folder}/${name}`, why: `K${parsed.key} is already claimed by ${clash}` })
+      continue
+    }
+    seen.set(parsed.key, `${folder}/${name}`)
+    out.push(parsed)
+  }
+  return out
+}
+
+/**
+ * Files in the order the board had them in, which is what `order` was written down for. A file
+ * with no `order` — hand-written, or older than this format — sorts by its key, so it lands after
+ * nothing in particular rather than at the top of everything.
+ */
+const byOrder = (a, b) => (Number(a.data.order ?? a.key) - Number(b.data.order ?? b.key)) || a.key - b.key
+
+/**
+ * The files back into the database.
+ *
+ * This is the backup path, and it runs exactly when the database is empty for this repository —
+ * a new machine, a lost `~/.k0`, a board somebody deleted. It refuses to run on a repository
+ * that already has work in it: importing twice does not merge, it duplicates, and two stories
+ * carrying the key K42 is a board that can no longer say which one anything refers to.
+ *
+ * Everything `writeStory` and `writeEpic` print comes back. Three things are not printed and so
+ * cannot: the minute of the day a decision was taken or a log entry written, whether a decision
+ * came out of a discussion or a plan, and whether a check was ticked by the user or by Claude.
+ * None of the three is worth a line of a file somebody has to read.
+ */
+export function importRepo(repoPath) {
+  const report = { epics: 0, stories: 0, skipped: [], already: false, error: null }
+  if (!there(repoPath)) return report
+  return attempt(report, () => {
+    if (!fs.existsSync(inside(repoPath))) return {}
+    if (db.storiesOfProject(repoPath).length || db.listEpics(repoPath).length) return { already: true }
+    return restore(repoPath, report)
+  })
+}
+
+function restore(repoPath, report) {
+  const epicIds = new Map()
+  // Per epic, the file's D-numbers against the rows they became: a story's verdict may be about
+  // its epic's D3, and only the epic knows which row that is.
+  const epicDecisions = new Map()
+  const epicFiles = usable(repoPath, 'epics', report).sort(byOrder)
+  epicFiles.forEach((f, i) => {
+    const epic = db.createEpic({
+      project_path: repoPath,
+      title: f.title,
+      body: proseText(f.sections, 'Why'),
+      lang: f.data.lang || '',
+      sort_hint: i,
+      key_num: f.key,
+    })
+    if (f.data.state === 'Done') db.patchEpic(epic.id, { state: 'Done' })
+    for (const r of parseRounds(sectionLines(f.sections, 'Discussion'))) db.addEpicRound(epic.id, r)
+    const rows = restoreDecisions(parseDecisions(sectionLines(f.sections, 'Decisions')), (d) =>
+      db.addEpicDecision(epic.id, d)
+    )
+    epicIds.set(f.key, epic.id)
+    epicDecisions.set(f.key, rows)
+    report.epics++
+  })
+
+  const storyIds = new Map()
+  const storyFiles = usable(repoPath, 'stories', report).sort(byOrder)
+  storyFiles.forEach((f, i) => {
+    const story = db.createStory({
+      project_path: repoPath,
+      title: f.title,
+      description: descriptionFrom(f.lead),
+      body: proseText(f.sections, 'Why'),
+      prompt: proseText(f.sections, 'Prompt'),
+      lang: f.data.lang || '',
+      color: f.data.color || 'yellow',
+      epic_id: epicIds.get(keyNumber(f.data.epic)) ?? null,
+      sort_hint: i,
+      key_num: f.key,
+    })
+    storyIds.set(f.key, story.id)
+    report.stories++
+
+    db.patchStory(story.id, { plan: proseText(f.sections, 'Plan'), starred: f.data.starred === 'true' })
+    // Through `setState` and not as a column of the insert: `Done` is also `completed_at`, and a
+    // story restored straight into the table would come back saying Done while the board, which
+    // reads the date, went on drawing it as unfinished.
+    if (db.STATES.includes(f.data.state) && f.data.state !== story.state) db.setState(story.id, f.data.state)
+
+    for (const r of parseRounds(sectionLines(f.sections, 'Discussion'))) db.addRound(story.id, r)
+    const own = restoreDecisions(parseDecisions(sectionLines(f.sections, 'Decisions')), (d) =>
+      db.addDecision(story.id, d)
+    )
+
+    const { runs, checks } = parseVerification(sectionLines(f.sections, 'Verification'))
+    db.setCheckItems(story.id, checks)
+    // A verdict names the decision it is about — `D3`, or `K7·D3` when the story inherited it —
+    // and is recorded against that row's id. Matching on the number alone would put a verdict
+    // about the epic's third decision onto the story's third decision, which is another sentence.
+    const inherited = (key) => epicDecisions.get(keyNumber(key)) ?? new Map()
+    for (const [run, results] of runs) {
+      const rows = []
+      for (const r of results) {
+        const row = (r.epic_key ? inherited(r.epic_key) : own).get(r.decision_n)
+        if (row) rows.push({ decision_id: row.id, verdict: r.verdict, evidence: r.evidence })
+      }
+      db.recordRunChecks(story.id, run, rows)
+    }
+    for (const entry of parseLog(sectionLines(f.sections, 'Log'))) db.addLogEntry(story.id, entry)
+    // Oldest first, each one closing the one before it, which is how the story kept its history
+    // of attempts. They come back looking alive, because that is what attaching a session means;
+    // the watcher is a second behind and marks dead whichever of them is not really running.
+    for (const sid of listValue(f.data.sessions)) db.attachSession(story.id, sid)
+  })
+
+  // Second pass, and it has to be one: a story can depend on a story further down the folder, and
+  // a task can be read before the story it was split out of exists.
+  for (const f of storyFiles) {
+    const id = storyIds.get(f.key)
+    if (!id) continue
+    const parent = storyIds.get(keyNumber(f.data.parent))
+    // A hand-edited pair of files each naming the other as parent is a loop, and a loop poisons
+    // every later write of either story rather than showing up here. It is refused where it can
+    // still be said out loud.
+    if (parent && parent !== id && !descends(parent, id)) db.patchStory(id, { parent_story_id: parent })
+    else if (parent) report.skipped.push({ file: `stories/K${f.key}`, why: 'its parent is one of its own tasks' })
+    for (const dep of listValue(f.data.depends_on)) {
+      const on = storyIds.get(keyNumber(dep))
+      if (on) db.addDependency(id, on)
+    }
+  }
+
+  // Last of all, because everything above moves `updated_at` as it writes, and the point of the
+  // dates is that they are the ones the board had.
+  for (const f of epicFiles) restoreTimes('epic', epicIds.get(f.key), f.data)
+  for (const f of storyFiles) restoreTimes('story', storyIds.get(f.key), f.data)
+
+  // Handed back rather than left in `report` for the caller to read. `attempt` builds its answer
+  // as `{ ...fallback, ...fn() }`, and the spread of `fallback` is evaluated BEFORE `fn()` runs —
+  // so anything counted into `report` while restoring is copied at nought and thrown away. The
+  // restore itself worked; it just reported that it had done nothing, which is worse than failing.
+  return { epics: report.epics, stories: report.stories, skipped: report.skipped }
+}
+
+/** Whether `id` is somewhere above `candidate` in the tree, which would make a parent a loop. */
+function descends(candidate, id) {
+  const seen = new Set()
+  let at = candidate
+  while (at && !seen.has(at)) {
+    if (at === id) return true
+    seen.add(at)
+    at = db.getStory(at)?.parent_story_id ?? null
+  }
+  return false
+}
+
+function restoreTimes(kind, id, data) {
+  if (!id) return
+  db.restoreTimes(kind, id, {
+    created_at: moment(data.created),
+    updated_at: moment(data.updated),
+    completed_at: moment(data.completed),
+  })
+}
+
+/**
+ * Decisions in the order they were taken, then the supersedings.
+ *
+ * Two passes, because a decision is regularly replaced by one written after it: D2 superseded by
+ * D7 cannot be recorded until D7 has a row and an id of its own. `add` hands back the row, and
+ * the file's numbers are matched to those ids rather than trusted as ids themselves — which is
+ * also what makes a gap survive. Somebody deletes the D3 line by hand; the file still says D4 and
+ * D5, the verdicts written about them still say D4 and D5, and the map keeps those numbers
+ * pointing at the right sentences whatever the database chose to call them. The number goes down
+ * with the decision as well, because `n` is the decision's name and a renumbered one is a
+ * counter-check about somebody else's rule.
+ */
+function restoreDecisions(parsed, add) {
+  const rows = new Map()
+  for (const d of [...parsed].sort((a, b) => a.n - b.n)) {
+    rows.set(d.n, add({ n: d.n, text: d.text }))
+  }
+  for (const d of parsed) {
+    const row = rows.get(d.n)
+    const by = d.superseded_by_n == null ? null : rows.get(d.superseded_by_n)
+    if (row && by) db.supersedeDecision(row.id, by.id)
+  }
+  return rows
+}