fix(google): cap max_tokens at 65536 on the Gemini 3.8 Live models - #591

gemini-3.8-live and gemini-3.8-live-extended-thinking set no max_tokens constraint, so celeste sends any value. Google documents a fixed output limit for both. The model pages (Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking) each list "Input token limit 131,072 / Output token limit 65,536". GET /v1beta/models/{id} returns outputTokenLimit: 65536 for both. The GenerateContent reference defines maxOutputTokens as "The maximum number of tokens to include in a response candidate. Note: The default value varies by model, see the Model.output_token_limit attribute". The Live API reference takes a GenerationConfig in BidiGenerateContentSetup.generationConfig and does not list maxOutputTokens among the unsupported fields. On Google Cloud, the Gemini 3.8 Live page lists "Maximum output tokens 64K", the same 65,536. This is a documented output figure, which is what CLAUDE.md's MAX_TOKENS rule requires.

The constraint has an effect because max_tokens already reaches the setup frame. MaxTokensMapper (GOOGLE_PARAMETER_MAPPERS, wire MaxOutputTokensMapper) writes generationConfig.maxOutputTokens, and GoogleLiveClient._build_request puts the result under setup. The existing wire-contract row covers this mapping.

Fix:

• src/celeste/modalities/live/providers/google/models.py: LiveParameter.MAX_TOKENS: Range(min=1, max=65536) on both entries. This is the same shape as the Google text models.
• No test change: the mapper list is unchanged, and test_parameter_wire_contracts.py already has the Live maxOutputTokens row.

Validation:

• make ci passes (762 unit tests).

• Live, Gemini API key. One turn per case through celeste.live.connect(...) on wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent, prompt "Say hello in one short sentence.":

  • Both models, max_tokens=65536: the setup frame carries generationConfig.maxOutputTokens: 65536, setupComplete arrives, and the turn ends with turnComplete (audio plus transcript).
  • Both models, max_tokens=65537: rejected locally with ConstraintViolationError: Must be between 1 and 65536, got 65537.

• The field is honored on gemini-3.8-live, with the prompt "Tell me a long, detailed story about a lighthouse keeper":

  max_tokens	response tokens	thinking tokens
  150	71	57
  400	275	48
  not set	1,486	78

API reference vs. live serving (recorded per CLAUDE.md): the Live setup only checks that the value is positive.

• maxOutputTokens: 0, sent through extra_body to skip celeste's check, closes the socket with 1007 "max_output_tokens must be positive". That confirms min=1.
• 65537 and 1000000 sent the same way are accepted, and the turn completes normally.
• So nothing live enforces max=65536. The cap comes from the documented output-token limit and outputTokenLimit, as it does for the Google text models.
• Side observation, not changed here: with max_tokens=20 on gemini-3.8-live, the server sends setupComplete and then no modelTurn or turnComplete for 45 s. The small budget seems to go entirely to thinking tokens.
• Google Cloud (GoogleADC) was not called live: the local ADC project lacks aiplatform.endpoints.predict.