@azumag/opencode-rate-limit-fallback 1.0.15 → 1.0.16

package/dist/index.js

@@ -316,28 +316,36 @@ export const RateLimitFallback = async ({ client, directory }) => {
                 parts: parts,
                 model: { providerID: nextModel.providerID, modelID: nextModel.modelID },
             };
-            toast("Retrying", `Using ${nextModel.providerID}/${nextModel.modelID}`, "info").catch(() => { });
-            // Abort to cancel the retry loop, then promptAsync immediately.
-            // CRITICAL: no delay between abort and promptAsync.
-            // - In headless (opencode run), server disposes ~8ms after session idle.
-            //   Any delay (even 50ms) means promptAsync arrives after server is dead.
-            // - promptAsync returns immediately (just queues), so it fires before
-            //   the server can shut down.
-            // - Do NOT use prompt (sync) here — it causes abort flag race condition
-            //   in TUI mode (prompt immediately interrupted).
+            // CRITICAL PATH: abort → promptAsync with NO delay between them.
+            //
+            // In headless mode (opencode run), the server disposes within milliseconds
+            // after session goes idle (observed ~8ms in production logs — this is not
+            // a guaranteed bound, just empirically observed). Any delay (setTimeout,
+            // awaited toast, etc.) means promptAsync arrives after the server is dead.
+            //
+            // The await on promptAsync waits for the HTTP round-trip (server acknowledgment),
+            // NOT for prompt completion — generation runs asynchronously on the server.
+            //
+            // Do NOT use prompt() (sync) here — it triggers the abort flag race condition
+            // in TUI mode, causing the new prompt to be immediately interrupted.
             try {
                 await client.session.abort({ path: { id: sessionID } });
                 logToFile(`abort succeeded for session ${sessionID}`);
             }
             catch (abortErr) {
-                logToFile(`abort failed (non-critical): ${abortErr}`);
+                // If abort fails, the session may still be in its retry loop.
+                // We still send promptAsync as best-effort: when the retry loop eventually
+                // completes (timeout or success), the queued prompt should be processed.
+                logToFile(`abort failed: ${abortErr} — sending promptAsync as best-effort`);
             }
             await client.session.promptAsync({
                 path: { id: sessionID },
                 body: promptBody,
             });
-            logToFile(`promptAsync sent successfully for session ${sessionID} with model ${nextModel.providerID}/${nextModel.modelID}`);
-            toast("Fallback Successful", `Now using ${nextModel.modelID}`, "success").catch(() => { });
+            logToFile(`promptAsync sent for session ${sessionID} with model ${nextModel.providerID}/${nextModel.modelID}`);
+            // Toasts are fire-and-forget: placed AFTER the critical path so they cannot
+            // interfere with the abort→promptAsync timing, even if they fail in headless.
+            toast("Fallback Active", `Now using ${nextModel.modelID}`, "success").catch(() => { });
             retryState.delete(stateKey);
             // Clear fallback flag to allow next fallback if needed
             fallbackInProgress.delete(sessionID);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@azumag/opencode-rate-limit-fallback",
-  "version": "1.0.15",
+  "version": "1.0.16",
   "description": "OpenCode plugin that automatically switches to fallback models when rate limited",
   "type": "module",
   "main": "./dist/index.js",