Function read_reasoning_budget_tokens 

fn read_reasoning_budget_tokens(
    obj: &Map<String, Value>,
    issues: &mut Vec<FieldIssue>,
) -> Option<i32>

Read the reasoning_budget_tokens field, under either name upstream accepts for it.

Exactly upstream’s range and nothing narrower: -1 <= v <= i32::MAX, with -1 meaning “defer to the launch --reasoning-budget” and 0 meaning “stop thinking immediately”. llama-server answers -2 with an HTTP 400 naming that range, so rejecting below -1 here reproduces upstream’s own verdict rather than adding a gglib opinion — the difference from read_reasoning_effort, which has no upstream verdict to reproduce.

The alias is read, and the canonical key wins

llama-server accepts THINKING_BUDGET_TOKENS_KEY as a second spelling of the same parameter (ADR 0007 finding 7c). A reader that knew only the canonical name left the alias ungoverned: it entered no layer, appeared in no discard record, was overwritten by no force-insert, and so an untrusted client’s thinking_budget_tokens outranked the operator’s resolved budget silently — the #779 shape this arc exists to close.

Whichever name arrives, the value becomes InferenceConfig::reasoning_budget_tokens and is governed like any other client-authoritative budget. With both present the canonical key wins, because that is the name gglib itself emits and the one every other surface (provenance, the audit, gglib model explain) reports. An explicit null counts as absent under either name, as it does for every other reader here.

Issues are reported against the key the client actually sent, so a refusal names the text that was in the request rather than a name the client never used.