Sync decisions contract for InferenceInfra #5208 (Eve format 4)

Merged
AM
19h ago
+10-8

Sync of the vendored decisions contract for InferenceInfra #5208: Eve (pplx-decider) moves to v1.2, which answers prompt_format_version 4.

  • contracts/wally-decisions-public-v1.openapi.json: re-vendored from InferenceInfra ecd738b257. The prompt_format_version enums gain 4, and their descriptions change.
  • src/account/decisions_contract.rs: regenerated. Only CONTRACT_SHA256 moves, because the reader takes the version as an integer with no enum check.
  • versions.toml: the decisions_source_commit and decisions_contract_sha256 pins.

There's no behaviour change for Wally. It never pins prompt_format_version, and it already reads any integer.

Why now: InferenceInfra's consumer-impact check keeps #5208 red until this pin matches it. This PR merges first, then #5208. The monorepo has a sibling PR that does the same.

python3 contracts/sync_from_inferenceinfra.py --check passes. The cargo tests need the kit, so CI runs them.

Summary by cubic

Syncs the vendored InferenceInfra decisions contract so prompt_format_version 4 (Eve v1.2) is accepted.

  • Re-vendors the OpenAPI contract from InferenceInfra ecd738b257, adding 4 to the request and response prompt_format_version enums and updating their descriptions.
  • Regenerates src/account/decisions_contract.rs; only CONTRACT_SHA256 changes because the reader takes the version as an integer with no enum check.
  • Updates the decisions_source_commit and decisions_contract_sha256 pins in versions.toml.

No behavior change for Wally: it never pins prompt_format_version and already reads any integer. This merge unblocks InferenceInfra #5208, so merge it first.

Summary by CodeRabbit

  • API Updates
    • The decisions API now accepts prompt format version 4 alongside versions 2 and 3.
    • Requests and responses document that models may support two prompt format versions during a wording transition.