ADR-0038: A configurable AI temperature
- Status: Accepted
- Date: 2026-10-31
- Amends: ADR-0035, whose temperature was fixed rather than configurable
Context
ADR-0035 asks every classification request for temperature 0, "fixed rather than
configurable", and accepts that a model taking only its default temperature can no longer
classify. The release that carries it, 1.1.0, is a minor one. Before it, no request named a
temperature, so a deployment using such a model worked; after the upgrade every one of its
changes would become a CLASSIFIER_FALLBACK. A minor release must not do that.
Such models exist for two of the three providers:
- OpenAI's reasoning models accept only their default temperature.
- Anthropic's API reference marks
temperaturedeprecated: "Models released after Claude Opus 4.6 do not support setting temperature. A value of 1.0 will be accepted for backwards compatibility, all other values will be rejected with a 400 error."
The ranges each API accepts are, from its reference: OpenAI 0 to 2, DeepSeek 0 to 2,
and Anthropic 0.0 to 1.0.
Decision
RELEASEFLOW_AI_TEMPERATUREsets the temperature of every classification request. It stays one value for every request of a deployment, as ADR-0035 intended; it is never chosen per change, per Project, or per request.- Unset or blank means
0. A deployment that does not set it sends exactly the bytes it sent before this decision, so recorded requests still match. The Compose files forward${RELEASEFLOW_AI_TEMPERATURE:-0}, because an empty value is what Compose passes for an unset variable written${…:-}. defaultleaves the temperature out of the request, so the model uses its own, as every request did before ADR-0035.- Any other value must parse as a number within the selected provider's range, inclusive. Otherwise startup fails, like the other AI settings, with a message naming the variable and the accepted range. Without a provider, nothing is sent and the value is not read.
- ADR-0035's reasons for
0stand: it remains the default, and the reason to change it is a model that cannot take it, not variety of wording.
Consequences
- A deployment using an OpenAI reasoning model, or a Claude model released after Opus 4.6,
keeps classifying after the upgrade once it sets
RELEASEFLOW_AI_TEMPERATURE=default. The CHANGELOG and the upgrade note say so before the upgrade, since nothing can detect such a model at startup. - A deployment that sets nothing behaves as ADR-0035 describes, request for request.
- A value other than
0widens the variance ADR-0035 narrowed. That is the operator's choice and applies to every change alike.