🦋 New version release

dotnet/core/src/SmooAI.SmoothOperator.Core.csproj

@@ -15,7 +15,7 @@
   <!-- NuGet package metadata -->
   <PropertyGroup>
     <PackageId>SmooAI.SmoothOperator.Core</PackageId>
-    <Version>1.9.0</Version>
+    <Version>1.10.0</Version>
     <Authors>SmooAI</Authors>
     <Company>SmooAI</Company>
     <Description>Native C# implementation of the smooth-operator agent engine — the in-process, Microsoft.Extensions.AI-based sibling of the Rust smooai-smooth-operator-core. Drive an agentic tool-calling loop over any IChatClient; author tools from ordinary C# methods with AIFunctionFactory. Learns from Microsoft Agent Framework idioms.</Description>

go/version.go

@@ -5,4 +5,4 @@ package e2e
 // language artifacts. The real Go "publish" is a git tag (go/v<Version>); this
 // constant is the anchor that scripts/sync-versions.mjs keeps in sync with the
 // canonical npm version on every changeset release.
-const Version = "1.9.0"
+const Version = "1.10.0"

python/core/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "smooai-smooth-operator-core"
-version = "1.9.0"
+version = "1.10.0"
 description = "Native Python implementation of the smooth-operator agent engine — an in-process, OpenAI-compatible agentic tool-calling loop with knowledge grounding. The Python sibling of the Rust reference engine and the C# core."
 readme = "README.md"
 license = { text = "MIT" }

rust/smooth-operator-core/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "smooai-smooth-operator-core"
-version = "1.9.0"
+version = "1.10.0"
 edition = "2021"
 license = "MIT"
 repository = "https://github.com/SmooAI/smooth-operator-core"

rust/smooth-operator-temporal/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "smooai-smooth-operator-temporal"
-version = "1.9.0"
+version = "1.10.0"
 edition = "2021"
 description = "Optional Temporal-backed durable execution backend for the smooth-operator agent engine (ADR-030)."
 license = "MIT"

typescript/core/CHANGELOG.md

@@ -1,5 +1,40 @@
 # @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.
+
 ## 1.9.0
 
 ### Minor Changes

typescript/core/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@smooai/smooth-operator-core",
-	"version": "1.9.0",
+	"version": "1.10.0",
 	"description": "Native TypeScript implementation of the smooth-operator agent engine — an in-process, OpenAI-compatible agentic tool-calling loop with knowledge grounding. The TypeScript sibling of the Rust reference engine, the C# core, and the Python core.",
 	"type": "module",
 	"license": "MIT",

typescript/temporal/CHANGELOG.md

@@ -1,5 +1,12 @@
 # @smooai/smooth-operator-temporal
 
+## 1.9.3
+
+### Patch Changes
+
+- Updated dependencies [418d996]
+  - @smooai/smooth-operator-core@1.10.0
+
 ## 1.9.2
 
 ### Patch Changes

typescript/temporal/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@smooai/smooth-operator-temporal",
-	"version": "1.9.2",
+	"version": "1.9.3",
 	"description": "Optional Temporal-backed durable execution backend for the smooth-operator agent engine (ADR-030). The TypeScript sibling of the Rust `smooth-operator-temporal` crate — an agent turn runs as a Temporal workflow whose model call and tool invocations are activities, driving the same `driveTurn` loop as the in-process executor.",
 	"type": "module",
 	"license": "MIT",