Unexpected truncation of thought summary with Gemini 2.5 Pro

[giacomoni]

Sometimes, the thought summary returned when setting include_thoughts=True and thinking_budget=32768 is clearly truncated, although the total number of tokens used reported in the response is below the max budget:

GenerateContentResponseUsageMetadata(
  candidates_token_count=2220,
  prompt_token_count=13216,
  prompt_tokens_details=[
    ModalityTokenCount(
      modality=<MediaModality.TEXT: 'TEXT'>,
      token_count=9088
    ),
    ModalityTokenCount(
      modality=<MediaModality.DOCUMENT: 'DOCUMENT'>,
      token_count=4128
    ),
  ],
  thoughts_token_count=5041,
  total_token_count=20477
)

Has anyone encountered this issue before? I am particularly interested to understand if the truncation of the summary is a reflection of a truncation occurring in the original chain of thoughts, or whether this truncation is a bug occurring somewhere in the backend during the generation of the summary itself.

The response contains the full structured output, so the workflow does not break.