archilyzer

commit 9f315744f6b5f6658fad5cf6256338ea9f3aeb53
parent 558efa98ab0e8bf9e986662dd84839d89b155896
Author: I Mean I'm Just Saying <imeanimjustsaying@kiwifarms.st>
Date:   Fri, 17 Jul 2026 20:16:59 -0400

parakeet-stitch: validate cached window JSON before reusing it

Resume treated any existing win-NNNN.json as a finished window via existsSync.
A 0-byte or truncated file left by a run interrupted mid-write was then trusted
and blew up the final stitch on JSON.parse. Reuse a cached window only when the
file is non-empty and parses back to a { start, doc } payload; otherwise treat
it as absent so the window is re-transcribed.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Diffstat:
M
scripts/parakeet-stitch.mjs
 | 
17
++++++++++++++++-

1 file changed, 16 insertions(+), 1 deletion(-)
diff --git a/scripts/parakeet-stitch.mjs b/scripts/parakeet-stitch.mjs
@@ -298,6 +298,21 @@ async function main() {
   }
   await writeFile(metaPath, JSON.stringify(meta));
   const winPath = (i) => path.join(workDir, `win-${String(i).padStart(4, "0")}.json`);
+  // A cached window counts only if it's non-empty and parses back to a
+  // { start, doc } payload. A 0-byte or truncated file (e.g. from a prior run
+  // interrupted mid-write) is treated as absent so the window is re-transcribed,
+  // rather than being trusted here and crashing the final stitch on JSON.parse.
+  const cachedWindow = async (i) => {
+    if (!existsSync(winPath(i))) return false;
+    try {
+      const raw = await readFile(winPath(i), "utf8");
+      if (!raw.trim()) return false;
+      const parsed = JSON.parse(raw);
+      return parsed != null && typeof parsed === "object" && "doc" in parsed;
+    } catch {
+      return false;
+    }
+  };
 
   // Graceful pause: on SIGTERM/SIGINT, finish the in-flight window then stop.
   // Completed window JSONs stay in the work dir; we do NOT stitch (the final
@@ -317,7 +332,7 @@ async function main() {
     for (let i = 0; i < N; i += 1) {
       const start = offsets[i];
       // Resume: reuse a window already transcribed in a prior run.
-      if (existsSync(winPath(i))) {
+      if (await cachedWindow(i)) {
         progress(`segment ${i + 1}/${N} @${round3(start)}s — cached`);
         done += 1;
         continue;