@azumag/opencode-rate-limit-fallback 1.0.17 → 1.0.18

package/dist/index.js

@@ -316,32 +316,35 @@ export const RateLimitFallback = async ({ client, directory }) => {
                 parts: parts,
                 model: { providerID: nextModel.providerID, modelID: nextModel.modelID },
             };
-            // CRITICAL PATH: abort → promptAsync with NO delay between them.
+            // CRITICAL PATH: promptAsync BEFORE abort.
             //
-            // In headless mode (opencode run), the server disposes shortly after
-            // session goes idle. Any delay (setTimeout, awaited toast, etc.) risks
-            // promptAsync arriving after the server is dead.
+            // In headless mode (opencode run), abort → promptAsync fails because:
+            //   1. abort triggers server dispose sequence
+            //   2. promptAsync is accepted and message created
+            //   3. Server starts processing (busy) but dispose interrupts it (idle)
+            //   4. server.instance.disposed — all within ~6ms
+            //
+            // By sending promptAsync FIRST, the server knows there is pending work
+            // before abort triggers the dispose check. When abort cancels the retry
+            // loop and the session goes idle, the server should process the queued
+            // prompt instead of disposing.
             //
             // promptAsync: HTTP POST /session/{id}/prompt_async → 204 (SDK sdk.gen.js).
-            // prompt (sync): blocks until generation completes → abort flag race in TUI,
-            //   server dispose in headless. Do NOT use.
+            // prompt (sync): blocks until generation completes — do NOT use.
             const t0 = Date.now();
+            await client.session.promptAsync({
+                path: { id: sessionID },
+                body: promptBody,
+            });
+            logToFile(`promptAsync completed for session ${sessionID} (${Date.now() - t0}ms) with model ${nextModel.providerID}/${nextModel.modelID}`);
+            const t1 = Date.now();
             try {
                 await client.session.abort({ path: { id: sessionID } });
-                logToFile(`abort succeeded for session ${sessionID} (${Date.now() - t0}ms)`);
+                logToFile(`abort succeeded for session ${sessionID} (${Date.now() - t1}ms, total ${Date.now() - t0}ms)`);
             }
             catch (abortErr) {
-                // If abort fails, the session may still be in its retry loop.
-                // We still send promptAsync as best-effort: when the retry loop eventually
-                // completes (timeout or success), the queued prompt should be processed.
-                logToFile(`abort failed (${Date.now() - t0}ms): ${abortErr} — sending promptAsync as best-effort`);
+                logToFile(`abort failed (${Date.now() - t1}ms): ${abortErr}`);
             }
-            const t1 = Date.now();
-            await client.session.promptAsync({
-                path: { id: sessionID },
-                body: promptBody,
-            });
-            logToFile(`promptAsync completed for session ${sessionID} (${Date.now() - t1}ms, total ${Date.now() - t0}ms) with model ${nextModel.providerID}/${nextModel.modelID}`);
             // Toast is best-effort notification. The toast() function (line ~185) has
             // built-in fallback: showToast failure → app.log. After promptAsync the
             // server may already be disposing, so both showToast and app.log could fail.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@azumag/opencode-rate-limit-fallback",
-  "version": "1.0.17",
+  "version": "1.0.18",
   "description": "OpenCode plugin that automatically switches to fallback models when rate limited",
   "type": "module",
   "main": "./dist/index.js",