Increase limits for LLM output tokens (#32)

Author: mckeea

.github/scripts/ai/update_versions_and_changelogs.py

@@ -816,9 +816,8 @@ def _do_call():
         print("=" * 70)
         print(result_text)
         print("=" * 70)
-        raise Exception(
-            f"Batch {batch_num}/{total_batches} failed: Invalid JSON response"
-        )
+        # Empty -> batch_with_retry splits and retries the halves instead of aborting.
+        return {}
     except Exception as e:
         print(f"\n❌ ERROR: Batch {batch_num}/{total_batches} processing failed")
         print(f"    Error: {e}")

.github/scripts/helpers/gemini_client.py

@@ -272,9 +272,8 @@ def _strip_code_fences(text: str) -> str:
 # Keeps the inline request body small and avoids the ~128k inline token limit.
 FILE_API_THRESHOLD_TOKENS = 30_000
 
-# Cap output tokens to reduce TPM consumption (input + output must fit in quota).
-# 4096 gives changelog generation enough room for 7 files × ~200 tokens each.
-MAX_OUTPUT_TOKENS = 4_096
+# 4_096 truncated the 15-file changelog batches mid-JSON; 2.5-flash allows 65k.
+MAX_OUTPUT_TOKENS = 32_768
 
 
 def _upload_prompt_as_file(prompt: str) -> object: