🦋 New version release

[brentrager]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@smooai/smooth-operator-core@1.10.0

Minor Changes

• 418d996: Carry cost/usage provenance and the gateway response id out of the engine, so nothing downstream can
  publish a fabricated number as a measurement.

  collect_stream invents a usage struct whenever a response carries no usage chunk — which LiteLLM
  does for every smooth-* alias — hardcoding prompt_tokens = 0 and estimating completion_tokens
  as content.len() / 4. The estimate is kept, because budget enforcement needs something to multiply,
  but it is now flagged: LlmResponse.usage_estimated, aggregated onto AgentEvent::Completed as
  usage_estimated. This is what let production record input_tokens = 0 on every streamed turn
  beside an output count that only ever tracked the reply's length. The old comment there claimed the
  estimate would "produce a real cost number against ModelPricing"; it does not, and it now says so.

  AgentEvent::Completed also gains cost_estimated — true once any call in the run was priced from
  the local ModelPricing table instead of the gateway's own figure. That table cannot price aliased
  routes and returns the free tier for anything it does not recognise, so a tainted total may be a wild
  under-count while looking exact. Cost and usage provenance are tracked separately on purpose: the
  gateway reports cost on an HTTP header and usage on an SSE chunk, so either can be authoritative
  while the other is a guess.

  Finally, the gateway's response id (chatcmpl-…, or msg_… on the Anthropic-native path) is now
  captured off both the streaming and non-streaming paths and surfaced as LlmResponse.response_id and
  AgentEvent::Completed.response_id. It was previously discarded at deserialization. It is the join
  key to LiteLLM_SpendLogs.request_id, whose row carries the gateway's authoritative dollars and
  its real prompt/completion counts — the only trustworthy source for either while the flags above can
  be set.

  All three fields cross the Temporal activity boundary via ModelCallOutput, so a durable replay
  cannot silently launder an estimate back into a measurement. smooth-operator-temporal's core
  requirement moves 1.8 -> 1.9 accordingly.

  Note for consumers matching exhaustively on StreamEvent: this adds a ResponseId variant.

@smooai/smooth-operator-temporal@1.9.3

Patch Changes

• Updated dependencies [418d996]
  • @smooai/smooth-operator-core@1.10.0