Make gpt-5.6-luna usable: per-model temperature, json_schema graders

[adamjohnwright]

Two things in this repo stopped gpt-5.6-luna working at all. Both fixed; the default model is unchanged — this makes the switch possible and tested, it does not make it.

1. Temperature

get_llm hardcoded temperature=0.0. The gpt-5.5/5.6 families and gpt-6-astra reject it:

Unsupported value: temperature does not support 0.0 with this model. Only the default (1) value is supported.

as a 400 on the first request, not at construction — so nothing notices until a user asks a question. resolve_temperature() sends 1.0 for those families, 0.0 for everything else, LLM_TEMPERATURE overrides.

The trap worth recording

Sending no temperature is not possible on langchain-openai 0.2.14: omitting the argument makes ChatOpenAI send its own pydantic default of 0.7, which these models refuse just as firmly as 0.0.

The first version of this branch did exactly that. model_dump(exclude_unset=True) agreed the field was unset — and the request still carried 0.7. Only running it against the live API showed it. That is constitution Article I in one line, so the test file says so.

What temperature 1 costs

Determinism. Measured, 10 runs each on four inputs:

input	gpt-4o-mini @ 0.0	gpt-5.6-luna @ 1.0
science question → intent	reactome ×10	reactome ×10
how-to question → intent	userguide ×10	userguide ×10
benign text → safety	true ×10	true ×10
prompt injection → safety	false ×10	false ×10

Stable on what was tested. Not a guarantee.

2. Structured output

The three graders used the default function_calling method, which the gpt-5.6 family refuses on /v1/chat/completions:

Function tools with reasoning_effort are not supported for gpt-5.6-luna ... use /v1/responses or set reasoning_effort to none.

and langchain-openai 0.2.14 has no Responses API support. method="json_schema" uses response_format instead — verified for all three graders on both models, so this is not a luna-only path that would rot untested.

(reasoning_effort="none" also works, but it turns the reasoning off, which is the point of the model. With json_schema, every effort level works.)

End-to-end

Through AgentGraph.ainvoke on the React-to-Me profile — the same entry point bin/chat-chainlit.py uses — with the Release95 bundle and three real questions. Both models answer; 6 LLM calls each.

	s/question	input tok	output tok
gpt-4o-mini	22.5	~2685	see note
gpt-5.6-luna	41.2	~3238	see note

Note: the OpenAI callback undercounts streamed completions inconsistently between the two, so treat output tokens as indicative only. Latency and call count are solid.

Pricing I cannot verify — no API reports it. That is the one input to the "just as cheap" question I could not check.

Qualitatively, on "Which complexes contain EGFR?" luna named four specific complexes with Reactome links; gpt-4o-mini gave a general description naming none. n=3, no rubric — a signal, not a measurement.