fhirsmith 0.18.2 → 1.0.0

package/library/memory-guard.js

@@ -0,0 +1,428 @@
+/**
+ * Process memory guard - the one place that decides how close to its memory limit the
+ * server is, so the request gate (503) and the in-operation check (deadCheck, 422) agree.
+ *
+ * The limit is the cgroup limit the kernel will enforce (MemoryMax on the systemd unit, or
+ * the container limit); with none, the machine's total RAM. Two thresholds, as fractions of
+ * the limit, both measured against process RSS (with a role, the cgroup's private memory: see
+ * below):
+ *
+ *  - rejectFraction: new requests get 503 + Retry-After before any work (or body parsing)
+ *    is done. Shedding load early is what gives in-flight work and the GC room to recover.
+ *  - abortFraction:  operations already running are aborted at their next deadCheck (422
+ *    too-costly). Higher than rejectFraction, so the gate acts first.
+ *
+ * Both are configurable in config.server.memory: { "rejectFraction": 0.85, "abortFraction": 0.92 }
+ *
+ * While refusing, the guard looks now and then for memory it can get back: a forced last-resort
+ * GC (the JS heap, and the freed pages V8 keeps for reuse), then malloc_trim (memory the C
+ * allocator holds free - closed SQLite connections, above all; library/native-memory.js), if
+ * "trimAllocator" is true: by default only with a role (library/role.js), so a server with no
+ * role only forces the GC, as before.
+ *
+ * A master and its tx-workers (--role, cluster.md) share one cgroup, and the kernel's limit is
+ * on all of them together, so with a role (config.server.memory.group) the measure is the
+ * group's: the cgroup's private memory (memory.stat anon), read by each process itself. Every
+ * process refuses new requests above rejectFraction; above abortFraction every process aborts
+ * the operations it has running. Not only the largest: a process's size doesn't say whether it
+ * is busy, and a large idle one (the master, say) would leave a worker's runaway $expand to the
+ * kernel's OOM killer. This needs a cgroup with a limit of its own (MemoryMax on the unit or
+ * scope, a container limit); without one each process measures its own RSS, as one does.
+ *
+ * @module library/memory-guard
+ */
+
+const fs = require('fs');
+const os = require('os');
+const v8 = require('v8');
+const vm = require('vm');
+const path = require('path');
+const { readCgroupMemoryLimit, ownCgroupDir, readCgroupAnon, readCgroupProcessRss } = require('./cgroup-memory');
+const { trimNativeMemory, canTrimNativeMemory } = require('./native-memory');
+
+const DEFAULT_REJECT_FRACTION = 0.85;
+// malloc_trim (trimAllocator): off unless asked for, so a plain server is as it was. With a
+// role, library/role.js turns it on (ROLE_TRIM_ALLOCATOR) unless the config says false
+const DEFAULT_TRIM_ALLOCATOR = false;
+const DEFAULT_ABORT_FRACTION = 0.92;
+// RSS is cheap to read, but the gate runs on every request; a sample this old is fine
+const SAMPLE_MS = 250;
+// While refusing, look for memory to get back (a forced GC, then malloc_trim) at most this
+// often. Refusing requests removes the allocation activity that would otherwise trigger a GC,
+// and nothing returns the C allocator's free memory by itself, so without this RSS might never
+// fall back below the threshold and the server would stay shut.
+const GC_INTERVAL_MS = 5000;
+// A look that gets back less than a tenth of the excess doubles the wait before the next, up
+// to this: when most of RSS is in use (the SNOMED and LOINC cache buffers), they free little,
+// and every full GC stops the process. One that gets back more keeps the wait: closing
+// thousands of SQLite connections takes seconds, and the next look gets back the rest
+const GC_MAX_INTERVAL_MS = 60000;
+// Between those looks, malloc_trim alone, before a request is refused, at most this often and
+// with no back-off: memory used in a burst is freed when the burst ends, and the C allocator
+// keeps it resident until asked, so a quiet server could sit just over the threshold and
+// refuse everything until the next look, up to GC_MAX_INTERVAL_MS away. With nothing to give
+// back, a trim takes well under a millisecond (about 20 us measured, on a 3.3 GB heap)
+const TRIM_INTERVAL_MS = 1000;
+// The forced GC's form. A plain gc() collects the JS heap but keeps the pages it frees for
+// reuse, resident and outside heapTotal, so they still count against the limit: after bursts of
+// value set searches, a server held 2.7 GB of anon memory with a 1.0 GB heap, and plain GCs gave
+// back nothing. A last-resort GC gives those pages back (that server: 1.4 GB, 1.1 GB with
+// malloc_trim). "flavor" is in V8 12 (Node 22) on; Node 18 took the object for a lesser GC than
+// a plain gc(), so before V8 12, plain gc()
+const LAST_RESORT_GC = Number(process.versions.v8.split('.')[0]) >= 12;
+const LAST_RESORT_GC_ARG = { type: 'major', execution: 'sync', flavor: 'last-resort' };
+// the group measure reads a file: the in-operation check (about ten a request) reuses one
+// at most this old
+const GROUP_SAMPLE_MS = 100;
+// RssFile (/proc/self/status) is read at most this often. The in-operation check runs many
+// times a request, and a read of /proc/self/status costs twice what the RSS itself does;
+// between reads, a fresh RSS less the last RssFile is a close enough estimate
+const FILE_SAMPLE_MS = 200;
+// the in-operation check reuses a measure at most this old. It runs about ten times a request,
+// and RSS moves little in 10 ms
+const ABORT_SAMPLE_MS = 10;
+
+const state = {
+  limit: 0,
+  source: null,
+  rejectFraction: DEFAULT_REJECT_FRACTION,
+  abortFraction: DEFAULT_ABORT_FRACTION,
+  rejectAt: 0,
+  abortAt: 0,
+  trimAllocator: DEFAULT_TRIM_ALLOCATOR,
+  lastResortGc: LAST_RESORT_GC, // the forced GC's form (see LAST_RESORT_GC)
+  sampleTime: 0,
+  sampleRss: 0,
+  // a role (master, tx-worker): the cgroup measured, from ownCgroupDir; else null
+  group: null,
+  groupWanted: false,
+  // set by measureExcludingFileBacked()
+  excludeFileBacked: false,
+  fileTime: 0,
+  fileRss: 0 // the last RssFile read (see fileBackedRss)
+};
+
+function fraction(v, dflt) {
+  return (typeof v === 'number' && v > 0 && v <= 1) ? v : dflt;
+}
+
+/**
+ * The limit and the two thresholds for config.server.memory, and with opts.group the cgroup to
+ * measure (one with the limit on it, else null). configure() uses it, and so does the launcher
+ * (cluster.js), so both agree on where the threshold is.
+ * @param {object} [opts] - { rejectFraction, abortFraction, group, limit and cgroup (tests) }
+ */
+function thresholds(opts = {}) {
+  let limit = opts.limit;
+  let source = limit ? 'configured' : null;
+  let cg = null;
+  if (!limit) {
+    cg = readCgroupMemoryLimit();
+    limit = cg.limit;
+    source = cg.source;
+  }
+  let group = null;
+  if (opts.group) {
+    // only a cgroup that is the group's own, with the limit on it
+    const own = opts.cgroup !== undefined ? opts.cgroup : ownCgroupDir();
+    const limitFile = opts.cgroup !== undefined ? null : (cg || readCgroupMemoryLimit()).source;
+    if (own && (opts.cgroup !== undefined || (limitFile && path.dirname(limitFile) === own.dir)) && readCgroupAnon(own) > 0) {
+      group = own;
+    }
+  }
+  if (!limit) {
+    limit = os.totalmem();
+    source = 'os.totalmem()';
+  }
+  const rejectFraction = fraction(opts.rejectFraction, DEFAULT_REJECT_FRACTION);
+  const abortFraction = fraction(opts.abortFraction, DEFAULT_ABORT_FRACTION);
+  return { limit, source, group, rejectFraction, abortFraction, rejectAt: limit * rejectFraction, abortAt: limit * abortFraction };
+}
+
+/**
+ * (Re)initialise. Called once at startup from server.js with config.server.memory; the module
+ * also self-initialises with defaults so code that loads without server.js (tests, tools) works.
+ * @param {object} [opts] - as thresholds(), and trimAllocator
+ */
+function configure(opts = {}) {
+  Object.assign(state, thresholds(opts));
+  state.groupWanted = !!opts.group;
+  state.trimAllocator = (opts.trimAllocator ?? DEFAULT_TRIM_ALLOCATOR) === true;
+  state.sampleTime = 0;
+}
+
+/**
+ * Measure RSS less its file-backed pages (Linux only). With SNOMED CT cache files memory-mapped
+ * ("snomed": {"load": "mmap"} in config.json) RSS includes every mapped page that has been read, which the kernel drops
+ * and re-reads as it needs to - counting them would refuse requests on a server that is fine.
+ */
+function measureExcludingFileBacked() {
+  state.excludeFileBacked = process.platform === 'linux';
+  state.sampleTime = 0;
+}
+
+/**
+ * RssFile as last read, read again if that was FILE_SAMPLE_MS ago or more, or if the sample
+ * has been dropped (sampleTime 0: configure, measureExcludingFileBacked, a test).
+ */
+function fileBackedRss() {
+  const now = Date.now();
+  if (state.sampleTime === 0 || now - state.fileTime >= FILE_SAMPLE_MS) {
+    state.fileRss = rssFileBytes();
+    state.fileTime = now;
+  }
+  return state.fileRss;
+}
+
+function rssFileBytes() {
+  try {
+    const m = /^RssFile:\s+(\d+) kB/m.exec(fs.readFileSync('/proc/self/status', 'utf8'));
+    return m ? Number(m[1]) * 1024 : 0;
+  } catch {
+    return 0;
+  }
+}
+
+/** This process's RSS, less reclaimable file-backed pages when using mapped files. */
+function ownRss() {
+  return process.memoryUsage.rss() - (state.excludeFileBacked ? fileBackedRss() : 0);
+}
+
+function rss(fresh = false) {
+  const now = Date.now();
+  if (fresh || now - state.sampleTime > SAMPLE_MS) {
+    state.sampleRss = (state.group && readCgroupAnon(state.group)) || ownRss();
+    state.sampleTime = now;
+  }
+  return state.sampleRss;
+}
+
+/** True if new requests should be refused (uses a sample at most SAMPLE_MS old). */
+function shouldReject() {
+  return state.rejectAt > 0 && rss() > state.rejectAt;
+}
+
+let gcFn;
+
+/**
+ * global.gc if node was started with --expose-gc, which is how it should be started (the
+ * Dockerfile and npm start do; add it to the node command of a systemd unit). Otherwise
+ * expose it at runtime: the flag can be set late, and a fresh context then has gc - this
+ * works, but Node does not promise it will keep working. null if neither works.
+ */
+function getGc() {
+  if (gcFn !== undefined) {
+    return gcFn;
+  }
+  gcFn = null;
+  if (typeof global.gc === 'function') {
+    gcFn = global.gc;
+  } else {
+    try {
+      v8.setFlagsFromString('--expose-gc');
+      const fn = vm.runInNewContext('gc');
+      if (typeof fn === 'function') {
+        gcFn = fn;
+      }
+    } catch {
+      // no gc available - the guard still works, it just recovers more slowly
+    }
+  }
+  return gcFn;
+}
+
+let aborts = 0;
+let lastAbortLog = 0;
+let abortLog = null;
+
+/**
+ * Current RSS in bytes if running operations should be aborted, else 0. A measure at most
+ * ABORT_SAMPLE_MS old (GROUP_SAMPLE_MS for a group; file-backed pages are read at most
+ * FILE_SAMPLE_MS ago), or fresh
+ * if the sample was dropped (configure, measureExcludingFileBacked).
+ * A trip is written to the server log (at most once a minute, with a running count) - the
+ * caller's own message only goes to the aborted operation's diagnostics.
+ * @param {string} [place] - where the check was made, for the log
+ */
+function overAbortThreshold(place = 'unknown') {
+  if (state.abortAt <= 0) {
+    return 0;
+  }
+  const r = rss(Date.now() - state.sampleTime > (state.group ? GROUP_SAMPLE_MS : ABORT_SAMPLE_MS));
+  if (r <= state.abortAt) {
+    return 0;
+  }
+  aborts++;
+  const now = Date.now();
+  if (now - lastAbortLog > 60000) {
+    lastAbortLog = now;
+    try {
+      if (!abortLog) {
+        abortLog = require('./logger').getInstance().child({ module: 'memory' });
+      }
+      abortLog.warn(`Memory critical: ${gb(r)} GB of ${gb(state.limit)} GB - aborting running operations (at ${place}; ${aborts} aborted so far)`);
+    } catch {
+      // logging must never stop the abort
+    }
+  }
+  return r;
+}
+
+function gb(bytes) {
+  return (bytes / 1024 / 1024 / 1024).toFixed(1);
+}
+
+function describe() {
+  const group = !state.groupWanted ? '' : state.group
+    ? `, for the whole cgroup ${state.group.dir} (its private memory)`
+    : ', measured as this process\'s RSS (no cgroup limit of the service\'s own to measure as a group)';
+  return `Memory limit: ${gb(state.limit)} GB (from ${state.source})${group}; ` +
+    (state.excludeFileBacked && !state.group ? 'measured as RSS less file-backed pages; ' : '') +
+    `new requests refused above ${gb(state.rejectAt)} GB (${Math.round(state.rejectFraction * 100)}%), ` +
+    `operations aborted above ${gb(state.abortAt)} GB (${Math.round(state.abortFraction * 100)}%); ` +
+    `while refusing, a forced ${state.lastResortGc ? 'last-resort ' : ''}GC` + (!state.trimAllocator ? ' (no malloc_trim: trimAllocator is off)'
+      : canTrimNativeMemory() ? ' and malloc_trim' : ' (malloc_trim not available here)');
+}
+
+/**
+ * Express middleware: 503 + Retry-After while RSS is over the reject threshold. Install it
+ * before the body parsers so a refused request costs nothing. While refusing, it looks for
+ * memory to get back now and then and re-measures - every gcInterval ms, backing off to
+ * gcMaxInterval while the looks get back little: a forced last-resort GC (where the runtime has
+ * only a plain one, not while the JS heap is too small for that to help), then malloc_trim
+ * (unless trimAllocator is false); and between those, malloc_trim alone at most every
+ * trimInterval ms, before refusing a request. It logs when refusing starts (then once a minute,
+ * with how long and how many) and when it stops.
+ * @param {object} [opts] - { log, exempt: [path prefixes never refused], retryAfter (seconds),
+ *   gcInterval (ms), gcMaxInterval (ms), heapUsed (function returning bytes, for tests),
+ *   gc (function, for tests: as global.gc),
+ *   trim (function, for tests: as trimNativeMemory), trimInterval (ms), onReject (optional observation callback,
+ *   called only when sending a 503) }
+ */
+function middleware(opts = {}) {
+  const exempt = opts.exempt || [];
+  const retryAfter = String(opts.retryAfter || 30);
+  const log = opts.log;
+  const gcInterval = opts.gcInterval || GC_INTERVAL_MS;
+  const gcMaxInterval = Math.max(opts.gcMaxInterval || GC_MAX_INTERVAL_MS, gcInterval);
+  const heapUsed = opts.heapUsed || (() => process.memoryUsage().heapUsed);
+  const trim = opts.trim || trimNativeMemory;
+  const trimInterval = opts.trimInterval || TRIM_INTERVAL_MS;
+  let gcWait = gcInterval; // the current wait between forced GCs (backs off, see below)
+  let refusingSince = 0; // when the current spell of refusing began (0 = not refusing)
+  let refused = 0;       // refused in the current spell
+  let lastLog = 0;
+  let lastGc = 0;
+  let lastTrim = 0;
+
+  const secs = (ms) => Math.round(ms / 1000);
+
+  return (req, res, next) => {
+    if (exempt.some(p => req.path === p || req.path.startsWith(p + '/'))) {
+      return next();
+    }
+    let over = shouldReject();
+    const now = Date.now();
+
+    // Refusing starves the GC of the allocation that would normally trigger it, so while
+    // refusing, force a full collection now and then. A last-resort GC also gives back the
+    // pages V8 keeps for reuse, which no heap figure counts (0.9 GB of them under a 0.13 GB
+    // heapTotal, in one measure), so the heap's size says nothing of what it can free: it
+    // always runs, and the back-off bounds its cost. A plain gc() frees heap and nothing else,
+    // so with one, only if the JS heap is at least the amount over the threshold. Then
+    // malloc_trim: the C allocator keeps what is freed outside the JS heap (closed SQLite
+    // connections) resident until asked.
+    // Re-measure; a look that got back little doubles the wait before the next
+    if (over && now - lastGc >= gcWait) {
+      lastGc = now;
+      const before = state.sampleRss;
+      const excess = before - state.rejectAt;
+      const heap = heapUsed();
+      const gc = opts.gc || getGc();
+      // a group: every process refusing does this, so each need only free its share
+      const share = state.group ? Math.max(1, readCgroupProcessRss(state.group).size) : 1;
+      const done = [];
+      if (!state.lastResortGc && heap < excess / share) {
+        if (log) {
+          log.info(`Memory: no forced GC - the JS heap (${gb(heap)} GB) is less than the ${gb(excess)} GB over the threshold`);
+        }
+      } else if (gc) {
+        try {
+          if (state.lastResortGc) {
+            gc(LAST_RESORT_GC_ARG);
+          } else {
+            gc();
+          }
+          done.push(state.lastResortGc ? 'forced last-resort GC' : 'forced GC');
+        } catch {
+          // ignore
+        }
+      }
+      if (state.trimAllocator && trim() !== null) {
+        done.push('malloc_trim');
+        lastTrim = now;
+      }
+      if (done.length) {
+        const after = rss(true);
+        over = after > state.rejectAt;
+        gcWait = !over ? gcInterval : before - after >= excess / 10 ? gcWait : Math.min(gcWait * 2, gcMaxInterval);
+        if (log) {
+          log.info(`Memory: ${done.join(' and ')} while refusing - ${gb(before)} GB -> ${gb(after)} GB` +
+            (over ? `; next in ${secs(gcWait)}s` : ''));
+        }
+      } else {
+        gcWait = Math.min(gcWait * 2, gcMaxInterval);
+      }
+    } else if (over && state.trimAllocator && now - lastTrim >= trimInterval) {
+      lastTrim = now;
+      const before = state.sampleRss;
+      if (trim() !== null) {
+        const after = rss(true);
+        over = after > state.rejectAt;
+        if (log && (!over || before - after >= 0.1 * 2 ** 30)) {
+          log.info(`Memory: malloc_trim before refusing - ${gb(before)} GB -> ${gb(after)} GB`);
+        }
+      }
+    }
+
+    if (!over) {
+      gcWait = gcInterval;
+      if (refusingSince) {
+        if (log) {
+          log.warn(`Memory recovered: ${gb(state.sampleRss)} GB of ${gb(state.limit)} GB - accepting requests again after ${secs(now - refusingSince)}s (${refused} refused)`);
+        }
+        refusingSince = 0;
+        refused = 0;
+      }
+      return next();
+    }
+
+    if (!refusingSince) {
+      refusingSince = now;
+      lastLog = 0;
+    }
+    refused++;
+    try {
+      opts.onReject?.();
+    } catch {
+      // Observability must never change whether this request is refused.
+    }
+    if (log && now - lastLog > 60000) {
+      log.warn(`Memory high: ${gb(state.sampleRss)} GB of ${gb(state.limit)} GB - refusing new requests with 503 for ${secs(now - refusingSince)}s (${refused} refused)`);
+      lastLog = now;
+    }
+    const text = `Server is under memory pressure (${gb(state.sampleRss)} GB of ${gb(state.limit)} GB); try again shortly`;
+    res.status(503);
+    res.setHeader('Retry-After', retryAfter);
+    res.setHeader('Connection', 'close');
+    res.setHeader('Content-Type', 'application/fhir+json');
+    res.end(JSON.stringify({
+      resourceType: 'OperationOutcome',
+      issue: [{ severity: 'error', code: 'transient', details: { text }, diagnostics: text }]
+    }));
+  };
+}
+
+configure();
+
+module.exports = { thresholds, configure, measureExcludingFileBacked, shouldReject, overAbortThreshold, describe, middleware, gb, getGc, _state: state };

package/library/native-memory.js

@@ -0,0 +1,52 @@
+/**
+ * Return the C allocator's free memory to the OS: glibc malloc_trim(0), through the optional
+ * koffi FFI dependency that the SNOMED mapping uses (tx/sct/mmap.js).
+ *
+ * Why: SQLite connections (one per code system provider, and an operation can open thousands)
+ * are allocated by the C library's malloc in libuv's threads, in its per-thread arenas. When
+ * they close, glibc keeps the freed memory in those arenas, resident, and gives back only a
+ * free stretch at the very top of each. After a 20,000-entry LOINC $batch-validate-code was
+ * aborted, a worker held 4.7 GB of such free memory: no GC can touch it, so the memory guard
+ * stayed shut. malloc_trim(0) releases the free pages inside the arenas as well (it took that
+ * worker from 8.0 GB to 3.9 GB in half a second). It frees nothing that is in use.
+ *
+ * Linux with glibc only. Elsewhere (macOS, Windows, musl without malloc_trim), or without
+ * koffi, it does nothing and returns null.
+ *
+ * @module library/native-memory
+ */
+let trim; // undefined: not tried yet; null: not available
+
+function load() {
+  if (trim === undefined) {
+    trim = null;
+    if (process.platform === 'linux') {
+      try {
+        trim = require('koffi').load(null).func('int malloc_trim(size_t pad)');
+      } catch {
+        // no koffi, or no malloc_trim in this C library
+      }
+    }
+  }
+  return trim;
+}
+
+/** @returns {boolean} whether trimNativeMemory() can do anything here */
+function canTrimNativeMemory() {
+  return !!load();
+}
+
+/** @returns {boolean|null} whether memory was released, or null if this can't be done here */
+function trimNativeMemory() {
+  if (!load()) {
+    return null;
+  }
+  try {
+    return trim(0) !== 0;
+  } catch {
+    trim = null;
+    return null;
+  }
+}
+
+module.exports = { trimNativeMemory, canTrimNativeMemory };

package/library/npm-audit.js

@@ -205,10 +205,16 @@ class NpmAudit {
    * Dashboard line: this server's version, and whether there is a newer release
    */
   renderRelease() {
+    const text = this.releaseText();
+    return text ? `<p>${text}</p>` : '';
+  }
+
+  /** The release, as inline html ('' if the installed version isn't known) */
+  releaseText() {
     if (!this.installedVersion) {
       return '';
     }
-    let html = `<p><strong>FHIRsmith release:</strong> this server runs ${escape(this.installedVersion)}`;
+    let html = `<strong>FHIRsmith release:</strong> this server runs ${escape(this.installedVersion)}`;
     const cmp = compareVersions(this.installedVersion, this.latestVersion);
     if (cmp === null) {
       html += ' (the latest release could not be checked)';
@@ -218,7 +224,7 @@ class NpmAudit {
     } else {
       html += ', the latest release';
     }
-    return html + '</p>';
+    return html;
   }
 
   /**
@@ -414,35 +420,55 @@ class NpmAudit {
   /**
    * Dashboard: always shown, so a check that is failing or switched off is visible too
    */
-  renderDashboard(now = Date.now()) {
-    let html = '<div id="npm-audit">';
+  /**
+   * The release and the audit, as one line of inline html (the top of /dashboard): the advice
+   * and the findings are in renderFindings
+   */
+  renderLine(now = Date.now()) {
     if (!this.enabled) {
-      return html + '<p><strong>npm audit:</strong> disabled (npmAudit.enabled = false)</p></div>';
+      return '<strong>npm audit:</strong> disabled (npmAudit.enabled = false)';
+    }
+    const parts = [];
+    const release = this.releaseText();
+    if (release) {
+      parts.push(release);
     }
-    html += this.renderRelease();
     if (!this.checkedAt) {
-      html += '<p><strong>npm audit:</strong> ';
-      html += this.lastError ? `<span style="color:#b00">could not be run: ${escape(this.lastError)}</span>` : 'not run yet';
-      return html + '</p></div>';
+      parts.push('<strong>npm audit:</strong> ' +
+        (this.lastError ? `<span style="color:#b00">could not be run: ${escape(this.lastError)}</span>` : 'not run yet'));
+    } else {
+      const colour = this.findings.length === 0 ? '#070' : (this.isSerious() ? '#b00' : '#b60');
+      let audit = `<strong>npm audit:</strong> <span style="color:${colour}">${escape(this.summary())}</span>` +
+        ` in ${this.packageCount} packages, checked ${escape(describeAgo(this.checkedAt.getTime(), now))}`;
+      if (this.lastError) {
+        audit += ` <span style="color:#b00">(the latest check failed: ${escape(this.lastError)})</span>`;
+      }
+      parts.push(audit);
     }
-    const colour = this.findings.length === 0 ? '#070' : (this.isSerious() ? '#b00' : '#b60');
-    html += `<p><strong>npm audit:</strong> <span style="color:${colour}">${escape(this.summary())}</span>` +
-      ` in ${this.packageCount} packages, checked ${escape(describeAgo(this.checkedAt.getTime(), now))}`;
-    if (this.lastError) {
-      html += ` <span style="color:#b00">(the latest check failed: ${escape(this.lastError)})</span>`;
+    return parts.join(' · ');
+  }
+
+  /** The advice and the table of findings, when there are any ('' otherwise) */
+  renderFindings() {
+    if (!this.enabled || !this.checkedAt || this.findings.length === 0) {
+      return '';
     }
-    html += '</p>';
-    if (this.findings.length > 0) {
-      html += `<p>${escape(this.advice())}</p>`;
-      html += '<table class="grid"><tr><th>Severity</th><th>Package</th><th>Installed</th><th>Vulnerable</th><th>Advisory</th></tr>';
-      for (const f of this.findings) {
-        const link = /^https:\/\//.test(f.url) ? `<a href="${escape(f.url)}">${escape(f.title || f.url)}</a>` : escape(f.title);
-        html += `<tr><td>${escape(f.severity)}</td><td>${escape(f.name)}</td><td>${escape(f.installed.join(', '))}</td>` +
-          `<td>${escape(f.range)}</td><td>${link}</td></tr>`;
-      }
-      html += '</table>';
+    let html = `<p>${escape(this.advice())}</p>`;
+    html += '<table class="grid"><tr><th>Severity</th><th>Package</th><th>Installed</th><th>Vulnerable</th><th>Advisory</th></tr>';
+    for (const f of this.findings) {
+      const link = /^https:\/\//.test(f.url) ? `<a href="${escape(f.url)}">${escape(f.title || f.url)}</a>` : escape(f.title);
+      html += `<tr><td>${escape(f.severity)}</td><td>${escape(f.name)}</td><td>${escape(f.installed.join(', '))}</td>` +
+        `<td>${escape(f.range)}</td><td>${link}</td></tr>`;
     }
-    return html + '</div>';
+    return html + '</table>';
+  }
+
+  /**
+   * The dashboard's npm audit: one line (renderLine), and only when there are findings, the advice
+   * and the table. The home page's banner links here (#npm-audit)
+   */
+  renderDashboard(now = Date.now()) {
+    return `<div id="npm-audit"><p>${this.renderLine(now)}</p>${this.renderFindings()}</div>`;
   }
 }
 

package/library/openapi-doc.js

@@ -429,10 +429,12 @@ const PAGE_STYLE = `<style>
 .openapi-ref .try-body { max-height: 30em; overflow: auto; }
 </style>`;
 
+// Headings: the page this goes into is titled by the module's template with an <h2>, so the
+// reference starts at <h3> (Endpoints, Schemas) - <h4> per operation and per schema, <h5> for
+// the parts of an operation - and the API's title is not a heading of its own
 function buildHtml(spec, BASE_PATH) {
   let html = PAGE_STYLE + '<div class="openapi-ref">';
-  html += `<h1>${escape(spec.info.title)} API</h1>`;
-  html += `<p class="lead">${escape(spec.info.summary || '')}</p>`;
+  html += `<p class="lead"><strong>${escape(spec.info.title)} API</strong>${spec.info.summary ? ': ' + escape(spec.info.summary) : ''}</p>`;
   html += `<p>Machine-readable description (OpenAPI ${escape(spec.openapi)}): ` +
     `<a href="${BASE_PATH}/openapi.json">openapi.json</a> &middot; ` +
     `<a href="${BASE_PATH}/openapi.yaml">openapi.yaml</a></p>`;
@@ -441,7 +443,7 @@ function buildHtml(spec, BASE_PATH) {
     html += `<p>See also: <a href="${escape(spec.externalDocs.url)}">${escape(spec.externalDocs.description || spec.externalDocs.url)}</a></p>`;
   }
 
-  html += '<h2>Endpoints</h2>';
+  html += '<h3>Endpoints</h3>';
   // an index, when there are enough of them to need one
   const all = Object.entries(spec.paths).flatMap(([p, item]) => METHODS.filter(m => item[m]).map(m => [p, m, item[m]]));
   if (all.length > 12) {
@@ -464,7 +466,7 @@ function buildHtml(spec, BASE_PATH) {
       }
       first = false;
       html += `<div class="panel panel-default" id="${escape(op.operationId || '')}"><div class="panel-heading">`;
-      html += `<span class="label label-primary">${method.toUpperCase()}</span> <code>${escape(BASE_PATH + p)}</code> &mdash; ${escape(op.summary || '')}`;
+      html += `<h4 class="panel-title"><span class="label label-primary">${method.toUpperCase()}</span> <code>${escape(BASE_PATH + p)}</code> &mdash; ${escape(op.summary || '')}</h4>`;
       html += '</div><div class="panel-body">';
       html += markdown(op.description);
       html += renderParameters(spec, [...(item.parameters || []), ...(op.parameters || [])]);
@@ -474,7 +476,7 @@ function buildHtml(spec, BASE_PATH) {
     }
   }
 
-  html += '<hr/><h2>Schemas</h2>';
+  html += '<hr/><h3>Schemas</h3>';
   for (const [name, schema] of Object.entries((spec.components && spec.components.schemas) || {})) {
     html += renderSchema(name, schema);
   }

package/library/package-manager.js

@@ -351,6 +351,11 @@ class CIBuildClient {
 
 class PackageManager {
     totalDownloaded = 0;
+    /**
+     * {Set<string>} the package folders (id#version, in the cache folder) that fetch and
+     * fetchUrl have returned - what has been used, for pruning the rest (Library.pruneUnusedCaches)
+     */
+    usedPackages = new Set();
 
     /**
      * @param {string[]} packageServers - Ordered list of package server URLs
@@ -377,6 +382,7 @@ class PackageManager {
         // Check cache first
         const cachedPath = await this.checkCache(packageId, resolvedVersion);
         if (cachedPath) {
+            this.usedPackages.add(cachedPath);
             return cachedPath;
         }
 
@@ -386,6 +392,7 @@ class PackageManager {
         this.totalDownloaded = this.totalDownloaded + packageData.length;
         // Extract to cache
         const extractedPath = await this.extractToCache(packageId, resolvedVersion, packageData);
+        this.usedPackages.add(extractedPath);
 
         return extractedPath;
     }
@@ -629,6 +636,7 @@ class PackageManager {
             }
 
             this.totalDownloaded = this.totalDownloaded + packageData.length;
+            this.usedPackages.add(finalName);
             return finalName;
         } catch (error) {
             console.error(`Error fetching package from URL ${url}: ${error.message}`);