@azumag/opencode-rate-limit-fallback 1.0.16 → 1.0.17

package/dist/index.js

@@ -318,33 +318,34 @@ export const RateLimitFallback = async ({ client, directory }) => {
             };
             // 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.
+            // 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.
             //
-            // 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.
+            // 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.
+            const t0 = Date.now();
             try {
                 await client.session.abort({ path: { id: sessionID } });
-                logToFile(`abort succeeded for session ${sessionID}`);
+                logToFile(`abort succeeded for session ${sessionID} (${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: ${abortErr} — sending promptAsync as best-effort`);
+                logToFile(`abort failed (${Date.now() - t0}ms): ${abortErr} — sending promptAsync as best-effort`);
             }
+            const t1 = Date.now();
             await client.session.promptAsync({
                 path: { id: sessionID },
                 body: promptBody,
             });
-            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.
+            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.
+            // The outer .catch() ensures even total toast() failure never blocks or throws.
             toast("Fallback Active", `Now using ${nextModel.modelID}`, "success").catch(() => { });
             retryState.delete(stateKey);
             // Clear fallback flag to allow next fallback if needed

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@azumag/opencode-rate-limit-fallback",
-  "version": "1.0.16",
+  "version": "1.0.17",
   "description": "OpenCode plugin that automatically switches to fallback models when rate limited",
   "type": "module",
   "main": "./dist/index.js",