diff --git a/dist/providers/google-gemini-cli.js b/dist/providers/google-gemini-cli.js
index 0000000..1111111 100644
--- a/dist/providers/google-gemini-cli.js
+++ b/dist/providers/google-gemini-cli.js
@@ -248,6 +248,12 @@ async function* streamGeminiCli(model, context, credentials, options) {
                         break; // Success, exit retry loop
                     }
                     const errorText = await response.text();
+                    // PATCH: Fail immediately on 429 for Antigravity to let caller rotate accounts.
+                    // Antigravity rate limits can have very long retry delays (10+ minutes).
+                    if (isAntigravity && response.status === 429) {
+                        console.log(`[pi-ai] 429 rate limit - failing fast to rotate account`);
+                        throw new Error(`Cloud Code Assist API error (${response.status}): ${errorText}`);
+                    }
                    // Check if retryable
                    if (attempt < MAX_RETRIES && isRetryableError(response.status, errorText)) {
                        // Use server-provided delay or exponential backoff

diff --git a/dist/providers/openai-responses.js b/dist/providers/openai-responses.js
index 0000000..1111111 100644
--- a/dist/providers/openai-responses.js
+++ b/dist/providers/openai-responses.js
@@ -397,9 +397,17 @@ function convertMessages(model, context) {
         }
         else if (msg.role === "assistant") {
             const output = [];
+            // OpenAI Responses rejects `reasoning` items that are not followed by a `message`.
+            // Tool-call-only turns (thinking + function_call) are valid assistant turns, but
+            // their stored reasoning items must not be replayed as standalone `reasoning` input.
+            const hasTextBlock = msg.content.some((b) => b.type === "text");
             for (const block of msg.content) {
                 // Do not submit thinking blocks if the completion had an error (i.e. abort)
                 if (block.type === "thinking" && msg.stopReason !== "error") {
                     if (block.thinkingSignature) {
+                        if (!hasTextBlock)
+                            continue;
                         const reasoningItem = JSON.parse(block.thinkingSignature);
                         output.push(reasoningItem);
                     }
                 }
                 else if (block.type === "text") {

diff --git a/dist/providers/openai-codex-responses.js b/dist/providers/openai-codex-responses.js
index 0000000..1111111 100644
--- a/dist/providers/openai-codex-responses.js
+++ b/dist/providers/openai-codex-responses.js
@@ -434,9 +434,17 @@ function convertMessages(model, context) {
         }
         else if (msg.role === "assistant") {
             const output = [];
+            // OpenAI Responses rejects `reasoning` items that are not followed by a `message`.
+            // Tool-call-only turns (thinking + function_call) are valid assistant turns, but
+            // their stored reasoning items must not be replayed as standalone `reasoning` input.
+            const hasTextBlock = msg.content.some((b) => b.type === "text");
             for (const block of msg.content) {
                 if (block.type === "thinking" && msg.stopReason !== "error") {
                     if (block.thinkingSignature) {
+                        if (!hasTextBlock)
+                            continue;
                         const reasoningItem = JSON.parse(block.thinkingSignature);
                         output.push(reasoningItem);
                     }
                 }
                 else if (block.type === "text") {