Thought signature in agent response (no code changes introduced this behavior)

[Noladev]

Hi,
My agent responses went from: content":{"parts":[{"text":"Ah, hello there, friend! It's good to hear from you. How can I help you on your creative journey today?"}]
to:
{"content":{"parts":[{"thought_signature":"CpUCAcu98PDiQfIyH8Qj9nB2UX_OMAAoFNMn6C4jFrQH1hkbGeYhaD8W7vIkxIeI9vTeD-wBnJvNpekIX2icWWoBMrYmvojY9ujEdSJUN0QTaXPWzisz_ZP_uEChHwbloQnlLTTsBDmKaKthD9jouRui7E2iaJWVEGR1V9jQETD7jIhu1uF_dn6Cv-FeBypV38_kILv4Ciocl77DR4ZTx

I'm using ADK agents deployed on Vertex AI Agent Engine. The chat feature in my app worked until today...When I checked the logs I noticed the above.

[akshaygaur-atrium]

yes i too got the similar issue but with warning first
WARNING: Warning: there are non-text parts in the response: ['thought_signature'], returning concatenated text result from text parts. Check the full candidates.content.parts accessor to get the full model response.
ERROR: FastAPI encoder could not JSON-encode chunk of type <class 'dict'> during encoding: 'utf-8' codec can't decode byte 0xd6 in position 1: invalid continuation byte. Chunk data: {'content': {'parts': [{'thought_signature': b'\n\xd6\x01\x01\xcb\xbd\xf0\xf0baT\x00\x80~\x9e\xd6\x8e\x19N\xaf\xcb\xdfD\x05\x83\xc1\xfd>:\x01i\xaa5\x84\x12\xc1\xb3\xa0A\x8e}\xeb\x85\x9d\xa5\x81P\x9e\xb1\xeb\x9b\xaf\xdb\xcb\x92Uy\xcf\xf6\x1f\x92\x86'\x9dDvZ\x93\x8b\xfe\x05\xe4\xc01\n\xf2B\xad\xf6D{\x18r\x8cbN\x89\x89\xb9\x85R}\xf9\x80\xccb!\xca\xbd6\x00mS\xb7g\x00\x8bbo2\xa2\x97\x0c\x80Q '\x14\x10\xb6\\x87,\xbe\x01t\xee\x97K\xf6\xed\xeaoQ~\xdb\xe6\x9e\xce.c%\x8e\xec\x875\xc5\xcb\x87\xa4\xb8\xdbX\xb0X1g{:\xc7\xe2}\n\x0f\x18\x0c\x9f@\x8eM\xbd\xb64)\x04>\xcfL\x9c@h\xee9{\x84\xa6F\x92\xc0\xdd\x91\x93\x87\x8cv\x00\xea:|\xab\xd88\xe9c"\xb9D'\xb5j\x1b$\xa2\x0c\x1d\xb9\x8b', 'text': 'Hello! How can I help you today?'}], 'role': 'model'}, 'usage_metadata': {'candidates_token_count': 9, 'candidates_tokens_details': [{'modality': <MediaModality.TEXT: 'TEXT'>, 'token_count': 9}], 'prompt_token_count': 1864, 'prompt_tokens_details': [{'modality': <MediaModality.TEXT: 'TEXT'>, 'token_count': 1864}], 'thoughts_token_count': 46, 'total_token_count': 1919, 'traffic_type': <TrafficType.ON_DEMAND: 'ON_DEMAND'>}, 'invocation_id': 'e-8c72928f-xxxxxx-0f2d40d84e9b', 'author': 'root_agent', 'actions': {'state_delta': {}, 'artifact_delta': {}, 'requested_auth_configs': {}}, 'id': '', 'timestamp': 1752038537.067571}

[Noladev]

I was able to resolve this for now by changing the model fyi. This was occurring on 2.5 flash

[akshaygaur-atrium]

we were using gemini-2.5-pro
which model are you currently using.

[fufufukakaka]

Got the same issue for gemini-2.5-pro

[sachints-sk]

Got Same Issue for gemini-2.5-flash

[Oximus2025]

Same issue with 2.5 flash, downgrading to 2.0 makes the whole agent worthless in my workflow

[evildeedz]

Same our entire prod has stopped, this needs to get fixed asap. Gemini 2.0 is terrible with our agent

[Oximus2025]

It appears that the issue is coming with the Tools usage, for us, Gemini 2.0 cannot manage to use Tools, so we get the response from the AI, but once we use 2.5, it fails to respond correctly with the tool and get the above error message. The tools either 2.0 or 2.5 is not working.

[KSonPham]

same

[ieccAI]

any solution ?
problem in model or in ADK version ?

[suvigyajain0101]

Potential mitigation options ATM -

1. Turn off thinking
2. Switch to gemini-2.0-flash

[DamiMartinez]

same

[chalosalvador]

yes i too got the similar issue but with warning first WARNING: Warning: there are non-text parts in the response: ['thought_signature'], returning concatenated text result from text parts. Check the full candidates.content.parts accessor to get the full model response. ERROR: FastAPI encoder could not JSON-encode chunk of type <class 'dict'> during encoding: 'utf-8' codec can't decode byte 0xd6 in position 1: invalid continuation byte. Chunk data: {'content': {'parts': [{'thought_signature': b'\n\xd6\x01\x01\xcb\xbd\xf0\xf0baT\x00\x80~\x9e\xd6\x8e\x19N\xaf\xcb\xdfD\x05\x83\xc1\xfd>:\x01i\xaa5\x84\x12\xc1\xb3\xa0A\x8e}\xeb\x85\x9d\xa5\x81P\x9e\xb1\xeb\x9b\xaf\xdb\xcb\x92Uy\xcf\xf6\x1f\x92\x86'\x9dDvZ\x93\x8b\xfe\x05\xe4\xc01\n\xf2B\xad\xf6D{\x18r\x8cbN\x89\x89\xb9\x85R}\xf9\x80\xccb!\xca\xbd6\x00mS\xb7g\x00\x8bbo2\xa2\x97\x0c\x80Q '\x14\x10\xb6\x87,\xbe\x01t\xee\x97K\xf6\xed\xeaoQ~\xdb\xe6\x9e\xce.c%\x8e\xec\x875\xc5\xcb\x87\xa4\xb8\xdbX\xb0X1g{:\xc7\xe2}\n\x0f\x18\x0c\x9f@\x8eM\xbd\xb64)\x04>\xcfL\x9c@h\xee9{\x84\xa6F\x92\xc0\xdd\x91\x93\x87\x8cv\x00\xea:|\xab\xd88\xe9c"\xb9D'\xb5j\x1b$\xa2\x0c\x1d\xb9\x8b', 'text': 'Hello! How can I help you today?'}], 'role': 'model'}, 'usage_metadata': {'candidates_token_count': 9, 'candidates_tokens_details': [{'modality': <MediaModality.TEXT: 'TEXT'>, 'token_count': 9}], 'prompt_token_count': 1864, 'prompt_tokens_details': [{'modality': <MediaModality.TEXT: 'TEXT'>, 'token_count': 1864}], 'thoughts_token_count': 46, 'total_token_count': 1919, 'traffic_type': <TrafficType.ON_DEMAND: 'ON_DEMAND'>}, 'invocation_id': 'e-8c72928f-xxxxxx-0f2d40d84e9b', 'author': 'root_agent', 'actions': {'state_delta': {}, 'artifact_delta': {}, 'requested_auth_configs': {}}, 'id': '', 'timestamp': 1752038537.067571}

I'm getting this same error in the logs. This seems to be on the Agent Engine side because I'm able to run the agent locally and with adk api_server I only see the error when calling the agent deployed to Agent Engine.