@tokenoftrust/cli 1.4.0-rc.17 → 1.4.0-rc.19

package/src/commands/dev.mjs

@@ -615,9 +615,19 @@ export async function resolvePublicRendererSource(args, env = process.env, { dec
         `(or pin with --renderer-version to silence).`,
     );
   }
-  const tarball = meta?.versions?.[version]?.dist?.tarball;
-  if (!tarball) throw new Error(`no published ${pkg}@${version} on npm`);
-  return { kind: "public", version, url: tarball, strip: 1, cacheKey: `public-${version}` };
+  const dist = meta?.versions?.[version]?.dist;
+  if (!dist?.tarball) throw new Error(`no published ${pkg}@${version} on npm`);
+  // `integrity` is the artifact's CONTENT identity (npm dist.integrity, else the
+  // legacy shasum) — installRunnerTarball verifies the downloaded bytes against
+  // it and records it so a same-version corrected republish busts the cache.
+  return {
+    kind: "public",
+    version,
+    url: dist.tarball,
+    strip: 1,
+    cacheKey: `public-${version}`,
+    integrity: dist.integrity || dist.shasum || null,
+  };
 }
 
 /**
@@ -637,7 +647,17 @@ export async function resolveEntitledRendererSource(args, { client: providedClie
   if (!res?.url || !res?.version) {
     throw new Error(res?.error || "no renderer-artifact URL returned");
   }
-  return { kind: "entitled", version: res.version, url: res.url, strip: 0, cacheKey: res.version };
+  // Opportunistic content identity: recorded/verified when the MCP declares one
+  // (integrity/sha256); a server that doesn't is simply unverified (null), never
+  // an error — the cache then busts on version changes only, as before.
+  return {
+    kind: "entitled",
+    version: res.version,
+    url: res.url,
+    strip: 0,
+    cacheKey: res.version,
+    integrity: res.integrity || res.sha256 || null,
+  };
 }
 
 /**
@@ -676,7 +696,12 @@ export async function ensureRendererArtifact(args, { client: providedClient } =
 
   try {
     const runnerDir = await installRunnerTarball(
-      { source: credential.url, version: credential.version, isUrl: true },
+      {
+        source: credential.url,
+        version: credential.version,
+        isUrl: true,
+        integrity: credential.integrity || credential.sha256 || null,
+      },
       { log: (m) => console.error(m) },
     );
     setRunnerVersion(credential.version); // telemetry: stamp the entitled runner version, like ensureSampleRenderer
@@ -851,7 +876,7 @@ export async function ensureSampleRenderer(args, ctx, { env = process.env, cache
     console.error(`~ renderer: ${src.why}`);
     return installRunnerTarball(
       { source: src.source, version: sourceVersionKey(src.source), isUrl: src.isUrl },
-      { log: (m) => console.error(m) },
+      { log: (m) => console.error(m), cacheRoot },
     );
   }
 
@@ -862,12 +887,16 @@ export async function ensureSampleRenderer(args, ctx, { env = process.env, cache
   const declared = declaredRunnerVersion(args.workspace || ctx?.workspacePath);
   const wantVersion = declared || CLI_VERSION;
 
-  // Fully-offline fast path: when the WANTED version (declared, else lockstep) is
-  // already cached, reuse it without touching npm — honouring "don't hit npm when
-  // the RIGHT version is already cached" without ever reusing a version the
-  // resolution wouldn't choose. Safe because `public-<version>` only exists if a
-  // prior run fetched exactly that version. Skipped when an explicit pin is set
-  // (that must go through resolution).
+  // Offline-safe fast path: when the WANTED version (declared, else lockstep) is
+  // already cached, reuse it — honouring "don't hit npm when the RIGHT version is
+  // already cached" without ever reusing a version the resolution wouldn't choose.
+  // Safe because `public-<version>` only exists if a prior run fetched exactly
+  // that version. Skipped when an explicit pin is set (that must go through
+  // resolution). One refinement over fully-offline: a QUICK, soft-fail registry
+  // probe (publishedRunnerIntegrity) revalidates the cached CONTENT identity when
+  // npm is reachable, so a corrected republish under the same version string is
+  // picked up automatically; offline/slow/unanswerable → trust the cache exactly
+  // as before (the probe can never block or fail the run).
   if (!explicitPin) {
     const exact = pinnedPublicCacheDir(cacheRoot, wantVersion);
     if (exact) {
@@ -878,21 +907,31 @@ export async function ensureSampleRenderer(args, ctx, { env = process.env, cache
       // host's arch (rendererCacheHealthy) — a cache poisoned with the wrong-arch
       // bindings (npm/cli#4828) would otherwise be reused forever and crash astro
       // at boot with the swallowed "dev server didn't come up". When both hold,
-      // reuse it (fully offline-safe).
+      // reuse it (offline-safe).
       if (probeRunnerVersion(exact) && rendererCacheHealthy(exact)) {
+        const published = await publishedRunnerIntegrity(env, wantVersion);
+        const recorded = readCacheMarker(exact)?.integrity || null;
+        if (!published || !recorded || published === recorded) {
+          console.error(
+            `~ renderer: cached public runner ${wantVersion} (${declared ? "declared by this store" : "matches this CLI"})`,
+          );
+          setRunnerVersion(wantVersion);
+          prunePublicRunnerCache(cacheRoot, wantVersion);
+          return exact;
+        }
+        // Same version string, different published contents — a corrected
+        // republish. Fall through to resolution + a fresh install (which also
+        // verifies the new bytes against the new integrity).
+        console.error(`~ renderer: ${wantVersion} was republished with different contents — refetching the corrected artifact`);
+        prunePublicRunnerCache(cacheRoot, null); // drop ALL public-* — none matches what npm now publishes
+      } else {
         console.error(
-          `~ renderer: cached public runner ${wantVersion} (${declared ? "declared by this store" : "matches this CLI"})`,
+          rendererCacheHealthy(exact)
+            ? `~ renderer: cached runner at ${exact} can't report a version — refetching (forced upgrade)`
+            : `~ renderer: cached runner at ${exact} is missing native bindings for ${process.platform}-${process.arch} — refetching`,
         );
-        setRunnerVersion(wantVersion);
-        prunePublicRunnerCache(cacheRoot, wantVersion);
-        return exact;
+        prunePublicRunnerCache(cacheRoot, null); // drop ALL public-* — none is trustworthy
       }
-      console.error(
-        rendererCacheHealthy(exact)
-          ? `~ renderer: cached runner at ${exact} can't report a version — refetching (forced upgrade)`
-          : `~ renderer: cached runner at ${exact} is missing native bindings for ${process.platform}-${process.arch} — refetching`,
-      );
-      prunePublicRunnerCache(cacheRoot, null); // drop ALL public-* — none is trustworthy
     }
   }
 
@@ -923,8 +962,8 @@ export async function ensureSampleRenderer(args, ctx, { env = process.env, cache
   // The store preview engine (public npm ${PUBLIC_RUNNER_PACKAGE}@${pub.version}) — kept
   // out of the user's way; the setup spinner below is the visible progress.
   const dir = await installRunnerTarball(
-    { source: pub.url, version: pub.cacheKey, isUrl: true, strip: pub.strip },
-    { log: (m) => console.error(m) },
+    { source: pub.url, version: pub.cacheKey, isUrl: true, strip: pub.strip, integrity: pub.integrity },
+    { log: (m) => console.error(m), cacheRoot },
   );
   setRunnerVersion(pub.version); // telemetry: the runner version running this session
   // The pinned version is now installed under public-<version> — drop any other
@@ -949,92 +988,269 @@ function sourceVersionKey(source) {
  * authenticated (ensureRendererArtifact) and zero-login (ensureSampleRenderer)
  * paths so they cache identically.
  *
- * @param {{ source: string, version: string, isUrl?: boolean }} spec
- * @param {{ log?: (m: string) => void }} [opts]
+ * Cache-poisoning invariants (the 2026-08-18 first-run hardening):
+ *   • promote-on-success only — the install runs in a per-attempt staging dir and
+ *     is renamed into the canonical slot ONLY after it fully succeeds, so a failed
+ *     install can never become the cached artifact a later run resumes from.
+ *   • the completion marker is a manifest carrying the source's CONTENT identity
+ *     (`integrity`), so a corrected republish under the SAME version string is a
+ *     cache miss (rebuild), not a stale hit — no manual `rm -rf` ever required.
+ *   • a failed attempt auto-cleans and retries ONCE from a clean slate before
+ *     surfacing the error (transient blips heal themselves); deterministic
+ *     failures (`e.permanent`) skip the retry and fail loud immediately.
+ *
+ * @param {{ source: string, version: string, isUrl?: boolean, strip?: number, integrity?: string|null }} spec
+ *   `integrity` is the source artifact's content identity when the resolver knows
+ *   it (npm `dist.integrity`/`dist.shasum`); used to verify the downloaded bytes
+ *   and to bust a cached entry whose recorded identity no longer matches.
+ * @param {{ log?: (m: string) => void, cacheRoot?: string }} [opts]
  * @returns {Promise<string>} the cached, installed runner tree's root directory.
  */
-export async function installRunnerTarball({ source, version, isUrl = true, strip = 0 }, { log = (m) => console.error(m) } = {}) {
-  const runnerDir = join(RENDERER_CACHE_ROOT, version);
-  const marker = join(runnerDir, ".tot-cache-complete");
-  if (existsSync(marker)) {
-    // Reuse ONLY if the cached tree carries native bindings for this host's arch.
-    // A cache poisoned with the wrong-arch binaries (npm/cli#4828 — e.g. darwin-x64
-    // on an arm64 Mac) is otherwise trusted forever, and astro crashes at boot with
-    // `Cannot find native binding`, swallowed as "the dev server didn't come up".
-    // Unhealthy → fall through and rebuild from the source below.
-    if (rendererCacheHealthy(runnerDir)) return runnerDir; // already downloaded + installed
-    log(`~ store preview engine cache is missing native bindings for ${process.platform}-${process.arch} — rebuilding it…`);
+export async function installRunnerTarball(
+  { source, version, isUrl = true, strip = 0, integrity = null },
+  { log = (m) => console.error(m), cacheRoot = RENDERER_CACHE_ROOT } = {},
+) {
+  const runnerDir = join(cacheRoot, version);
+  // Staging dirs from DEAD runs (crashed/killed installs) must not leak disk
+  // forever — reap them here, the one funnel every install path goes through.
+  sweepStaleStagingDirs(cacheRoot);
+
+  const localSource = isUrl ? null : resolveLocalTarball(source);
+  // The EXPECTED content identity of the source artifact. A local tarball with no
+  // caller-provided integrity is cheap to hash on every run, so a same-path
+  // republish (new contents, same file name) busts the cache too.
+  const expected = integrity || (isUrl ? null : fileIntegrity(localSource));
+
+  const cached = readCacheMarker(runnerDir);
+  if (cached) {
+    if (!rendererCacheHealthy(runnerDir)) {
+      // A cache poisoned with the wrong-arch binaries (npm/cli#4828 — e.g.
+      // darwin-x64 on an arm64 Mac) is otherwise trusted forever, and astro
+      // crashes at boot with `Cannot find native binding`, swallowed as "the dev
+      // server didn't come up". Fall through and rebuild from the source below.
+      log(`~ store preview engine cache is missing native bindings for ${process.platform}-${process.arch} — rebuilding it…`);
+    } else if (expected && cached.integrity && expected !== cached.integrity) {
+      // CONTENT-HASH BUST: same version string, different artifact contents — a
+      // corrected republish. The cached entry is stale by identity, not by label;
+      // rebuild from the corrected source instead of serving the stale cache.
+      log(`~ the preview engine's ${version} artifact changed upstream (same version, new contents) — rebuilding…`);
+    } else {
+      return runnerDir; // already downloaded + installed (and contents still match)
+    }
   }
 
   // First run only — set the expectation so the one-time cost doesn't read as a
   // hang: this downloads + installs the renderer once, then every later run of
   // this version is a no-network cache hit.
   log(`~ first run: setting up your store preview (~a minute, one-time — cached after this)…`);
-  const localSource = isUrl ? null : resolveLocalTarball(source);
-  const archivePath = isUrl ? join(tmpdir(), `tot-renderer-${process.pid}-${Date.now()}.tar.gz`) : localSource;
-  try {
-    if (isUrl) {
-      // The fetch itself is otherwise silent (no per-byte output) and can run
-      // tens of seconds on a cold cache — tick a spinner so it never looks hung.
-      const spin = startProgress("downloading the store preview engine…");
+  // AUTO-CLEAN-AND-RETRY-ONCE: a transient failure (network blip mid-download, a
+  // registry hiccup mid-install) heals itself with one clean re-attempt instead
+  // of stopping a first run at an error only `rm -rf` folklore could clear.
+  // Bounded to one retry so a genuinely-broken source still fails loudly.
+  for (let attempt = 1; ; attempt++) {
+    const archivePath = isUrl ? join(tmpdir(), `tot-renderer-${process.pid}-${attempt}.tar.gz`) : localSource;
+    const stagingDir = `${runnerDir}.staging-${process.pid}`;
+    try {
+      if (isUrl) {
+        // The fetch itself is otherwise silent (no per-byte output) and can run
+        // tens of seconds on a cold cache — tick a spinner so it never looks hung.
+        const spin = startProgress("downloading the store preview engine…");
+        try {
+          await downloadFile(source, archivePath);
+        } finally {
+          spin.stop();
+        }
+      }
+      if (!existsSync(archivePath)) {
+        throw new Error(`renderer tarball not found: ${archivePath}`);
+      }
+      // Refuse to install bytes that don't match the source's declared identity —
+      // a truncated/corrupted download would otherwise be cached as if complete.
+      // (Transient by nature, so the retry above gets a fresh download.)
+      if (expected && !tarballMatchesIntegrity(archivePath, expected)) {
+        throw new Error(`the downloaded preview-engine tarball failed its integrity check (expected ${expected})`);
+      }
+      // The identity recorded in the completion manifest below — what future runs
+      // compare against to detect a same-version republish. Hash the actual bytes
+      // when the resolver couldn't tell us (e.g. the entitled signed-URL path).
+      const contentId = expected || fileIntegrity(archivePath);
+      rmSync(stagingDir, { recursive: true, force: true });
+      mkdirSync(stagingDir, { recursive: true });
+      extractTarball(archivePath, stagingDir, { strip });
+      // Pin the runner install to PUBLIC npm. The moat-free runner has only public
+      // deps, but the HOST's global ~/.npmrc may point `registry` at a private
+      // mirror (an internal proxy that 502s, or one an invited developer can't
+      // reach) — an invited dev's machine config must never decide where the
+      // runner's public deps come from. A project-level .npmrc wins over the user's.
+      writeFileSync(join(stagingDir, ".npmrc"), "registry=https://registry.npmjs.org/\n");
+      // The install is the long, noisy step — tick a spinner while its output goes
+      // to a log, so the terminal shows one clean line instead of the pnpm firehose.
+      // corepack setup logs to the SAME file so its failures aren't invisible (they
+      // were the silent cause of "couldn't set up the store preview engine").
+      const installLog = join(cacheRoot, `${version}.install.log`);
+      ensureCorepackPnpm(stagingDir, { logPath: installLog });
+      const installSpin = startProgress("installing the store preview engine…", {
+        stages: [{ afterMs: 20000, text: "still setting up the preview engine (first run only)…" }],
+      });
       try {
-        await downloadFile(source, archivePath);
+        await runPnpmInstall(stagingDir, { logPath: installLog });
       } finally {
-        spin.stop();
+        installSpin.stop();
       }
+      // Atomic-ish: only rename into the final, discoverable path once install
+      // succeeded, so a crashed/interrupted run never leaves a half-built cache
+      // entry that a later `tot dev` would treat as ready.
+      rmSync(runnerDir, { recursive: true, force: true });
+      renameSync(stagingDir, runnerDir);
+      // Fence a fresh install against npm/cli#4828: if the installer left the wrong
+      // arch's native bindings (or none) for this host, DON'T stamp the completion
+      // marker — an unmarked tree is never reused, so the next run reinstalls cleanly
+      // instead of caching the poison and crashing astro at boot. Fail loud + actionable
+      // rather than swallow it as "the dev server didn't come up". Permanent: the
+      // same installer on the same host would just produce the same result, so the
+      // auto-retry is skipped.
+      if (!rendererCacheHealthy(runnerDir)) {
+        await emitObstacle("renderer-native-bindings-missing");
+        const err = new CliError(
+          `the store preview engine installed but is missing its native components for ${process.platform}-${process.arch}`,
+          {
+            next: "install pnpm (`npm i -g pnpm`, or `corepack enable`) and re-run `tot start` — pnpm installs the platform-native bits npm can skip (npm/cli#4828)",
+            exitCode: 2,
+          },
+        );
+        err.permanent = true;
+        throw err;
+      }
+      writeCacheMarker(runnerDir, { version, integrity: contentId });
+      return runnerDir;
+    } catch (e) {
+      // A failed attempt must never survive on disk — not as staging debris, and
+      // (by promote-on-success) it never reached the canonical slot at all.
+      rmSync(stagingDir, { recursive: true, force: true });
+      if (e?.permanent === true || attempt >= 2) throw e;
+      log(`~ that didn't work (${String(e?.message || e).split("\n")[0]}) — retrying once from a clean slate…`);
+    } finally {
+      if (isUrl) rmSync(archivePath, { force: true });
     }
-    if (!existsSync(archivePath)) {
-      throw new Error(`renderer tarball not found: ${archivePath}`);
-    }
-    const stagingDir = `${runnerDir}.staging-${process.pid}`;
-    rmSync(stagingDir, { recursive: true, force: true });
-    mkdirSync(stagingDir, { recursive: true });
-    extractTarball(archivePath, stagingDir, { strip });
-    // Pin the runner install to PUBLIC npm. The moat-free runner has only public
-    // deps, but the HOST's global ~/.npmrc may point `registry` at a private
-    // mirror (an internal proxy that 502s, or one an invited developer can't
-    // reach) — an invited dev's machine config must never decide where the
-    // runner's public deps come from. A project-level .npmrc wins over the user's.
-    writeFileSync(join(stagingDir, ".npmrc"), "registry=https://registry.npmjs.org/\n");
-    // The install is the long, noisy step — tick a spinner while its output goes
-    // to a log, so the terminal shows one clean line instead of the pnpm firehose.
-    // corepack setup logs to the SAME file so its failures aren't invisible (they
-    // were the silent cause of "couldn't set up the store preview engine").
-    const installLog = join(RENDERER_CACHE_ROOT, `${version}.install.log`);
-    ensureCorepackPnpm(stagingDir, { logPath: installLog });
-    const installSpin = startProgress("installing the store preview engine…", {
-      stages: [{ afterMs: 20000, text: "still setting up the preview engine (first run only)…" }],
-    });
+  }
+}
+
+/**
+ * Read a cache entry's completion marker (`.tot-cache-complete`). Returns the
+ * manifest object (at least `{ integrity: string|null }`), or null when the
+ * marker is absent — i.e. the entry is incomplete/partial and must be treated
+ * as if it didn't exist. A legacy pre-manifest marker (a bare timestamp string)
+ * reads as complete-with-unknown-identity, so existing healthy caches survive
+ * the upgrade without a forced rebuild.
+ */
+export function readCacheMarker(dir) {
+  try {
+    const raw = readFileSync(join(dir, ".tot-cache-complete"), "utf8");
     try {
-      await runPnpmInstall(stagingDir, { logPath: installLog });
-    } finally {
-      installSpin.stop();
+      const m = JSON.parse(raw);
+      if (m && typeof m === "object") return { integrity: null, ...m };
+    } catch {
+      /* legacy timestamp-string marker */
     }
-    // Atomic-ish: only rename into the final, discoverable path once install
-    // succeeded, so a crashed/interrupted run never leaves a half-built cache
-    // entry that a later `tot dev` would treat as ready.
-    rmSync(runnerDir, { recursive: true, force: true });
-    renameSync(stagingDir, runnerDir);
-    // Fence a fresh install against npm/cli#4828: if the installer left the wrong
-    // arch's native bindings (or none) for this host, DON'T stamp the completion
-    // marker — an unmarked tree is never reused, so the next run reinstalls cleanly
-    // instead of caching the poison and crashing astro at boot. Fail loud + actionable
-    // rather than swallow it as "the dev server didn't come up".
-    if (!rendererCacheHealthy(runnerDir)) {
-      await emitObstacle("renderer-native-bindings-missing");
-      throw new CliError(
-        `the store preview engine installed but is missing its native components for ${process.platform}-${process.arch}`,
-        {
-          next: "install pnpm (`npm i -g pnpm`, or `corepack enable`) and re-run `tot start` — pnpm installs the platform-native bits npm can skip (npm/cli#4828)",
-          exitCode: 2,
-        },
-      );
+    return { integrity: null };
+  } catch {
+    return null; // no marker → never treat the entry as installed
+  }
+}
+
+/** Stamp a cache entry complete: version + source content identity + when. */
+function writeCacheMarker(dir, { version, integrity }) {
+  writeFileSync(
+    join(dir, ".tot-cache-complete"),
+    JSON.stringify({ version, integrity: integrity || null, completedAt: new Date().toISOString() }) + "\n",
+  );
+}
+
+/** sha512 SRI (`sha512-<base64>`, npm's `dist.integrity` format) of a file; null when unreadable. */
+function fileIntegrity(path) {
+  try {
+    return `sha512-${createHash("sha512").update(readFileSync(path)).digest("base64")}`;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Do the tarball's bytes match `expected` — an SRI string (`sha512-<b64>`, npm's
+ * `dist.integrity`) or npm's legacy `dist.shasum` (bare 40-hex sha1)? Unknown
+ * formats and probe errors return true: this check exists to catch corrupted
+ * bytes, never to block an install on a format we can't verify.
+ */
+export function tarballMatchesIntegrity(archivePath, expected) {
+  try {
+    const want = String(expected).trim();
+    const sri = /^(sha512|sha384|sha256|sha1)-([A-Za-z0-9+/=]+)$/.exec(want);
+    if (sri) {
+      return createHash(sri[1]).update(readFileSync(archivePath)).digest("base64") === sri[2];
     }
-    writeFileSync(marker, new Date().toISOString());
-  } finally {
-    if (isUrl) rmSync(archivePath, { force: true });
+    if (/^[0-9a-f]{40}$/i.test(want)) {
+      return createHash("sha1").update(readFileSync(archivePath)).digest("hex") === want.toLowerCase();
+    }
+    return true;
+  } catch {
+    return true;
+  }
+}
+
+/**
+ * Reap `<entry>.staging-<pid>` dirs left by DEAD processes — failed/killed
+ * installs used to accumulate one orphaned staging tree per attempt, leaking
+ * disk forever. A staging dir whose pid is still alive belongs to a concurrent
+ * `tot dev` mid-install and is left alone. Best-effort: never throws, and never
+ * touches this process's own staging dir (created fresh after this sweep).
+ */
+export function sweepStaleStagingDirs(cacheRoot, { pidAlive = processAlive } = {}) {
+  try {
+    if (!cacheRoot || !existsSync(cacheRoot)) return;
+    for (const name of readdirSync(cacheRoot)) {
+      const m = /\.staging-(\d+)$/.exec(name);
+      if (!m) continue;
+      const pid = Number(m[1]);
+      if (pid === process.pid || pidAlive(pid)) continue;
+      rmSync(join(cacheRoot, name), { recursive: true, force: true });
+    }
+  } catch {
+    /* best-effort cache hygiene */
+  }
+}
+
+/** Is a pid a live process? (signal 0 probe; EPERM = alive but not ours.) */
+function processAlive(pid) {
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch (e) {
+    return e?.code === "EPERM";
+  }
+}
+
+/**
+ * The registry-declared content identity (`dist.integrity`, else `dist.shasum`)
+ * of the public runner at `version` — or null when npm can't answer QUICKLY
+ * (offline, slow, 4xx/5xx, malformed). Used by ensureSampleRenderer's cached
+ * fast path to detect a same-version republish without ever making the network
+ * a hard dependency: null means "can't verify right now — trust the cache",
+ * preserving the offline-reuse behavior exactly.
+ */
+export async function publishedRunnerIntegrity(env, version, { timeoutMs = 2000, fetchFn = fetch } = {}) {
+  try {
+    const pkg = env.TOT_RUNNER_PACKAGE || PUBLIC_RUNNER_PACKAGE;
+    const registry = (env.TOT_NPM_REGISTRY || DEFAULT_NPM_REGISTRY).replace(/\/$/, "");
+    const res = await fetchFn(`${registry}/${pkg.replace("/", "%2f")}`, {
+      headers: { accept: "application/json" },
+      signal: AbortSignal.timeout(timeoutMs),
+    });
+    if (!res.ok) return null;
+    const dist = (await res.json())?.versions?.[version]?.dist;
+    return dist?.integrity || dist?.shasum || null;
+  } catch {
+    return null;
   }
-  return runnerDir;
 }
 
 /** Strip an optional file:// prefix from a local tarball path and resolve it absolute. */
@@ -1190,22 +1406,27 @@ export async function runPnpmInstall(runnerDir, { logPath, spawnFn = spawnAsyncR
         continue;
       }
       // The launcher ran; the install itself failed. That's the actionable error.
+      // (installRunnerTarball auto-retries this ONCE from a clean slate before it
+      // reaches the user — so by the time this surfaces, it failed twice.)
       await emitObstacle("install-failed");
       throw new CliError(
         `couldn't set up the store preview engine${pnpmFailureHint(logPath)}` +
           (logPath ? `\n     details: ${logPath}` : ""),
-        { next: "check the details log above, then re-run `tot start` (it resumes from the cache)" },
+        { next: "check the details log above, then re-run `tot start` (it retries from a clean slate — no cache to clear)" },
       );
     }
     // Every launcher ENOENT'd → there's no pnpm on this machine and corepack
     // couldn't provide one (corepack isn't bundled on Node 25+). npm ships with
     // every Node, so `npm i -g pnpm` is the escape hatch that always exists.
+    // Permanent: retrying can't conjure a launcher — skip the clean-slate retry.
     await emitObstacle("pnpm-missing");
-    throw new CliError(
+    const err = new CliError(
       "couldn't set up the store preview engine — pnpm isn't available on this machine" +
         (logPath ? `\n     details: ${logPath}` : ""),
       { next: "install pnpm with `npm i -g pnpm` (or `corepack enable`), then re-run `tot start`" },
     );
+    err.permanent = true;
+    throw err;
   } finally {
     if (fd !== null) closeSync(fd);
   }
@@ -1220,6 +1441,20 @@ function pnpmFailureHint(logPath) {
   if (!logPath) return " — is pnpm/corepack available on this host?";
   try {
     const tail = readFileSync(logPath, "utf8").slice(-8000);
+    // A 404 means the registry answered — a specific package/version doesn't
+    // exist there. Since installs now run from a clean slate every attempt
+    // (promote-on-success + auto-retry), this is a BROKEN RUNNER RELEASE (it
+    // references an unpublished package), not the user's cache — no `rm -rf`
+    // will help. Check this BEFORE the generic ERR_PNPM_FETCH match, since
+    // pnpm's 404 error text also contains "ERR_PNPM_FETCH".
+    const missing404 = tail.match(/ERR_PNPM_FETCH_404[^\n]*GET\s+(\S+)/i);
+    if (missing404) {
+      return (
+        ` — the preview engine references a package that isn't published (${missing404[1]});` +
+        " that's a broken preview-engine release, not your machine — try again later or pin a" +
+        " known-good version with --renderer-version"
+      );
+    }
     if (/npm\.tokenoftrust\.com|ERR_PNPM_FETCH|502|ECONNREFUSED|ETIMEDOUT|ENOTFOUND/i.test(tail)) {
       return " — the Token of Trust package registry looks unreachable right now; check your connection and retry";
     }

package/src/commands/pr.mjs

@@ -42,6 +42,24 @@ const DEFAULT_MCP_URL = "https://mcp.tokenoftrust.com";
 const DEFAULT_STOREFRONT_URL = "https://storefront.tokenoftrust.store";
 const SUBCOMMANDS = ["list", "view", "close"];
 
+/**
+ * The storefront-owned, shareable `/preview/<tenant>/pr/<N>` link — NEVER the
+ * forge/Gitea `url` (2026-08-18 incident: a raw forge PR URL reached an
+ * owner). `candidate_status` (the local-checkout MCP tool) has no
+ * `previewUrl` field at all, unlike the operator `GET /api/changes` path — so
+ * this constructs it the same way `runPrListOperator`'s caller resolves
+ * `storefrontUrl`, from the same env/--url override chain. Pure.
+ * @param {string} storefrontUrl
+ * @param {string} tenant
+ * @param {number|null|undefined} prNumber
+ * @returns {string|null}
+ */
+export function buildPreviewUrl(storefrontUrl, tenant, prNumber) {
+  if (typeof prNumber !== "number" || !tenant) return null;
+  const base = (storefrontUrl || DEFAULT_STOREFRONT_URL).trim().replace(/\/+$/, "");
+  return `${base}/preview/${tenant}/pr/${prNumber}`;
+}
+
 const USAGE = `tot pr — see and manage candidate PRs
 
   tot pr [list]              list your open candidate PRs for this store
@@ -118,9 +136,11 @@ export function matchCandidate(candidates, target) {
 /**
  * One-line candidate summary for `tot pr list` — surfaces branch ↔ PR# ↔ preview
  * URL so a dev sees, at a glance, which git branch each candidate belongs to (u4 —
- * branch-bound candidates) and where its preview lives. Prefers the candidate's
- * `previewUrl`, falling back to the PR `url`. `active` marks the one THIS checkout's
- * branch resolves to. Pure — unit-tested.
+ * branch-bound candidates) and where its preview lives. ONLY `previewUrl` (the
+ * storefront-owned `/preview/<tenant>/pr/<N>` link) — NEVER `url` (the forge/
+ * Gitea `html_url`), which must never reach a terminal (2026-08-18 incident:
+ * a raw forge PR URL reached an owner). `active` marks the one
+ * THIS checkout's branch resolves to. Pure — unit-tested.
  * @param {{prNumber?:number|null, branch?:string|null, changeId:string, state?:string|null,
  *   previewUrl?:string|null, url?:string|null}} c
  * @param {{ active?: boolean }} [opts]
@@ -128,8 +148,7 @@ export function matchCandidate(candidates, target) {
 export function formatCandidateLine(c, { active = false } = {}) {
   const pr = typeof c.prNumber === "number" ? `#${c.prNumber}` : "#—";
   const branch = c.branch ? c.branch : "(no branch)";
-  const previewUrl = c.previewUrl || c.url || null;
-  const urlPart = previewUrl ? `  ${previewUrl}` : "";
+  const urlPart = c.previewUrl ? `  ${c.previewUrl}` : "";
   const activePart = active ? "  ← active" : "";
   return `  PR ${pr}  ${branch}  ${c.changeId}  [${c.state ?? "?"}]${urlPart}${activePart}`;
 }
@@ -326,6 +345,7 @@ export async function run(argv, ctx) {
   }
 
   const baseUrl = args.mcp || env.MCP_BASE_URL || env.TOT_MCP_URL || DEFAULT_MCP_URL;
+  const storefrontUrl = args.url || env.TOT_STOREFRONT_URL || env.STOREFRONT_BASE_URL || DEFAULT_STOREFRONT_URL;
   const statePath = defaultCandidateStatePath(env);
   // Branch-bound (u4): the active-pointer namespace is scoped to the current git
   // branch, so the "← active" marker reflects THIS branch's candidate.
@@ -347,7 +367,8 @@ export async function run(argv, ctx) {
       const active = readActiveChangeId(statePath, scope);
       console.log(`Open candidate PRs for ${repo}:`);
       for (const c of candidates) {
-        console.log(formatCandidateLine(c, { active: !!active && c.changeId === active }));
+        const previewUrl = c.previewUrl ?? buildPreviewUrl(storefrontUrl, tenant, c.prNumber);
+        console.log(formatCandidateLine({ ...c, previewUrl }, { active: !!active && c.changeId === active }));
       }
       return 0;
     }
@@ -366,7 +387,9 @@ export async function run(argv, ctx) {
       if (match.headSha) console.log(`  head:             ${match.headSha}`);
       if (match.baseSha) console.log(`  base:             ${match.baseSha}`);
       console.log(`  mergeable (forge): ${match.mergeable ?? "?"}`);
-      if (match.url) console.log(`  ${match.url}`);
+      // ONLY the storefront-owned preview link -- never the raw forge/Gitea `url`.
+      const previewUrl = match.previewUrl ?? buildPreviewUrl(storefrontUrl, tenant, match.prNumber);
+      if (previewUrl) console.log(`  ${previewUrl}`);
       return 0;
     }