@retasc/cli 1.3.1 → 1.3.2

package/dist/config.js

@@ -1,6 +1,6 @@
 import { homedir } from "node:os";
 import { join } from "node:path";
-import { mkdirSync, readFileSync, writeFileSync, renameSync, unlinkSync, existsSync, chmodSync, } from "node:fs";
+import { mkdirSync, readFileSync, writeFileSync, renameSync, unlinkSync, existsSync, chmodSync, openSync, closeSync, writeSync, statSync, readdirSync, } from "node:fs";
 import { randomUUID } from "node:crypto";
 // Production defaults. Overridable via env for dev/testing.
 //   RETASC_DEPLOYMENT_URL — Convex deployment (.cloud) for management calls
@@ -101,11 +101,156 @@ export function saveConfig(cfg) {
         }
         throw err;
     }
+    // Best-effort GC of temp siblings orphaned by a hard kill (SIGKILL / power loss)
+    // between writeFileSync(tmp) and renameSync above — otherwise they accumulate in
+    // ~/.retasc forever across crashes. Runs after our own rename so our temp is
+    // already gone. Age-gated (see sweepStaleTemps) so it never touches a concurrent
+    // writer's in-flight temp.
+    sweepStaleTemps(dir);
+}
+/** A temp sibling this much older than now was orphaned by a crash, not left by a
+ *  live writer — any real writeFileSync→renameSync window is milliseconds. */
+const TEMP_STALE_MS = 60_000;
+/** Delete `.config.json.<uuid>.tmp` leftovers only once they're demonstrably stale,
+ *  so an unconditional sweep can't race-delete a concurrent writer's fresh temp
+ *  (which would reintroduce the very lost-update class this file guards against). */
+function sweepStaleTemps(dir) {
+    let entries;
+    try {
+        entries = readdirSync(dir);
+    }
+    catch {
+        return; // dir vanished / unreadable — nothing to sweep
+    }
+    const now = Date.now();
+    for (const name of entries) {
+        if (!name.startsWith(".config.json.") || !name.endsWith(".tmp"))
+            continue;
+        const p = join(dir, name);
+        try {
+            if (now - statSync(p).mtimeMs > TEMP_STALE_MS)
+                unlinkSync(p);
+        }
+        catch {
+            /* raced another sweeper or a writer's own rename — fine, leave it */
+        }
+    }
+}
+// --- Cross-process config lock -------------------------------------------------
+// saveConfig is byte-atomic (temp + rename) but that is NOT isolation: two
+// concurrent load→modify→save sequences both read the old file and the last
+// rename wins, silently dropping the other's fields (RTSC-250). The dangerous
+// case: a `defaultProjectPrefix` write that began before a token refresh writes
+// back the OLD single-use refreshToken, so the next refresh fails and a
+// non-interactive MCP context can't re-run the device flow. An advisory lock file
+// serializes the whole read-modify-write across processes, so every patcher
+// re-reads the freshest config (including a just-rotated token) before writing.
+//
+// Assumes a coherent LOCAL clock: the staleness heuristic compares a lock file's
+// mtime (set by the host that created it) against this host's Date.now(). That
+// holds for the intended case — several processes on ONE machine sharing one
+// ~/.retasc. On a network home with a skewed server clock it degrades to
+// best-effort (over-eager or over-lazy stealing); we don't target that here.
+// Any whole-file mutating writer MUST go through patchConfig so it takes the
+// lock — saveConfig alone is not self-locking (logout's delete is the one benign
+// exception: racing a refresh there is a user-intent race, not a durability bug).
+/** A lock whose mtime is older than this belongs to a holder that died without
+ *  releasing it (a crash leaves the O_EXCL file behind); break it so a crash
+ *  can't wedge every future patch forever. Far larger than any real hold, which
+ *  is a synchronous read+write of a tiny file (single-digit ms). */
+const LOCK_STALE_MS = 30_000;
+/** Absolute backstop: if staleness somehow never frees the lock, force it after
+ *  this. Kept ABOVE the stale threshold so normal breaking is governed by
+ *  staleness, not by a timer that could guillotine a merely-slow live holder. */
+const LOCK_TIMEOUT_MS = 60_000;
+const LOCK_BACKOFF_MS = 25;
+/** Block this (single) thread for `ms` without busy-spinning. Node is
+ *  single-threaded; the lock we wait on is released by ANOTHER process, so
+ *  parking the thread is correct — nothing here could release it. */
+function sleepSync(ms) {
+    Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms);
+}
+/** The current holder's identity + age, or undefined if the lock just vanished. */
+function readLockHolder(lockPath) {
+    try {
+        const nonce = readFileSync(lockPath, "utf8");
+        return { nonce, ageMs: Date.now() - statSync(lockPath).mtimeMs };
+    }
+    catch {
+        return undefined; // gone between our failed create and this read → just retry
+    }
+}
+/** Run `fn` holding an exclusive on-disk lock, so its read-modify-write of the
+ *  config can't interleave with another process's. Each acquirer stamps a unique
+ *  nonce into the lock file; steal and release only ever remove a lock whose nonce
+ *  still matches the one we observed, so a holder that was judged stale and had its
+ *  lock stolen can't later delete the DIFFERENT holder's lock (which would collapse
+ *  mutual exclusion back into the lost-update this whole mechanism prevents). */
+function withConfigLock(fn) {
+    const dir = configDir();
+    mkdirSync(dir, { recursive: true, mode: 0o700 });
+    const lockPath = join(dir, ".config.json.lock");
+    const nonce = randomUUID();
+    const deadline = Date.now() + LOCK_TIMEOUT_MS;
+    let fd;
+    for (;;) {
+        try {
+            // "wx" = O_CREAT | O_EXCL | O_WRONLY — atomically fails if the lock is held.
+            fd = openSync(lockPath, "wx", 0o600);
+            writeSync(fd, nonce); // stamp our identity so steal/release can verify it
+            break;
+        }
+        catch (err) {
+            if (err.code !== "EEXIST")
+                throw err;
+            const holder = readLockHolder(lockPath);
+            if (!holder)
+                continue; // vanished → retry the create immediately
+            if (holder.ageMs > LOCK_STALE_MS || Date.now() > deadline) {
+                // Dead holder (stale) or backstop expired: break THIS holder's lock only.
+                // If a fresh acquirer replaced it since our read, the nonce differs and we
+                // leave it — never steal a live lock out from under a new owner.
+                breakLockIf(lockPath, holder.nonce);
+                continue;
+            }
+            sleepSync(LOCK_BACKOFF_MS);
+        }
+    }
+    try {
+        return fn();
+    }
+    finally {
+        try {
+            closeSync(fd);
+        }
+        catch {
+            /* already closed */
+        }
+        // Only remove the lock if it's still OURS. If we were stolen from while stalled,
+        // the file now holds another holder's nonce — leave it for them.
+        breakLockIf(lockPath, nonce);
+    }
+}
+/** Unlink the lock only if it still carries `nonce` — a nonce-checked delete, so
+ *  we never remove a lock a different process now owns. */
+function breakLockIf(lockPath, nonce) {
+    try {
+        if (readFileSync(lockPath, "utf8") === nonce)
+            unlinkSync(lockPath);
+    }
+    catch {
+        /* already gone, replaced, or unreadable — nothing safe to do */
+    }
 }
 export function patchConfig(patch) {
-    const next = { ...loadConfig(), ...patch };
-    saveConfig(next);
-    return next;
+    // Load AND save under the lock: re-reading inside the critical section is what
+    // makes this safe — a patch that only sets `defaultProjectPrefix` still picks up
+    // a refreshToken another process rotated a moment ago, instead of clobbering it.
+    return withConfigLock(() => {
+        const next = { ...loadConfig(), ...patch };
+        saveConfig(next);
+        return next;
+    });
 }
 export function isLoggedIn(cfg = loadConfig()) {
     return Boolean(cfg.token);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retasc/cli",
-  "version": "1.3.1",
+  "version": "1.3.2",
   "description": "Retasc CLI — sign in with GitHub, create projects, mint agent API keys, and wire your agent to the Retasc MCP server in one command.",
   "type": "module",
   "bin": {