Local ONNX inference (embeddings + first-class classify) and AI-gateway OpenTelemetry improvements

.github/workflows/fixture-freshness.yml

@@ -54,7 +54,7 @@ jobs:
 
       # Cache only the standalone tool's target dir; the parent
       # workspace's cache is owned by other workflows.
-      - uses: Swatinem/rust-cache@v2.7.3
+      - uses: Swatinem/rust-cache@v2
         with:
           shared-key: fixture-freshness
           workspaces: test/fixtures/refresh-tool

.github/workflows/licensing-conformance.yml

@@ -176,7 +176,7 @@ jobs:
 
       - name: open PR with refreshed vendored copies
         if: steps.diff.outputs.changed == '1'
-        uses: peter-evans/create-pull-request@v6
+        uses: peter-evans/create-pull-request@v7
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
           branch: chore/licensing-schemas-refresh

.github/workflows/perf-regression.yml

@@ -93,13 +93,13 @@ jobs:
       # would let the second build trample the first's lockfile state
       # and skew the bench.
       - name: cache (candidate)
-        uses: Swatinem/rust-cache@v2.7.3
+        uses: Swatinem/rust-cache@v2
         with:
           shared-key: perf-regression-candidate
           workspaces: candidate-src
 
       - name: cache (baseline)
-        uses: Swatinem/rust-cache@v2.7.3
+        uses: Swatinem/rust-cache@v2
         with:
           shared-key: perf-regression-baseline
           workspaces: baseline-src

.github/workflows/synthetic.yml

@@ -54,7 +54,7 @@ jobs:
       # `bench-synthetic` is its own workspace, so it gets a dedicated
       # cache key. That keeps the proxy workspace cache from being
       # invalidated by churn on the synthetic harness.
-      - uses: Swatinem/rust-cache@v2.7.3
+      - uses: Swatinem/rust-cache@v2
         with:
           shared-key: synthetic
           workspaces: bench-synthetic
@@ -189,7 +189,7 @@ jobs:
 
       - uses: dtolnay/rust-toolchain@stable
 
-      - uses: Swatinem/rust-cache@v2.7.3
+      - uses: Swatinem/rust-cache@v2
         with:
           shared-key: chaos-hot-reload
 

crates/sbproxy-ai/Cargo.toml

@@ -11,6 +11,10 @@ sbproxy-config.workspace = true
 sbproxy-cache.workspace = true
 sbproxy-security.workspace = true
 sbproxy-platform.workspace = true
+# WOR-1223: local embedding source for the semantic cache. Only the
+# sidecar client is used here; the in-process embedder lives in sbproxy-core
+# because sbproxy-classifiers depends on sbproxy-ai (would be a cycle).
+sbproxy-classifier-client.workspace = true
 
 serde.workspace = true
 serde_json.workspace = true

crates/sbproxy-ai/src/handler.rs

@@ -71,6 +71,15 @@ pub struct AiHandlerConfig {
     /// See `sbproxy_security::pii::PiiConfig` for the rule schema.
     #[serde(default)]
     pub pii: Option<sbproxy_security::pii::PiiConfig>,
+    /// WOR-1228: when `true`, emit the prompt text as the OpenInference
+    /// `input.value` span attribute so trace backends (Phoenix, Langfuse)
+    /// show the actual conversation, not just token counts. Off by default
+    /// because prompt content is sensitive: when on, the text is routed
+    /// through the configured `pii` redactor (if any) and the always-on
+    /// secret redactor before it lands on the span. Enable only with `pii`
+    /// configured and a trace backend inside your trust boundary.
+    #[serde(default)]
+    pub trace_content: bool,
     /// Opaque semantic-cache configuration block. The OSS proxy
     /// stores this verbatim and surfaces it through the stream cache
     /// recorder hook so the enterprise implementation can read its
@@ -857,6 +866,7 @@ mod tests {
             resilience: None,
             shadow: None,
             pii: None,
+            trace_content: false,
             semantic_cache: None,
             prompts: None,
             usage_parser: "auto".to_string(),
@@ -885,6 +895,7 @@ mod tests {
             resilience: None,
             shadow: None,
             pii: None,
+            trace_content: false,
             semantic_cache: None,
             prompts: None,
             usage_parser: "auto".to_string(),
@@ -913,6 +924,7 @@ mod tests {
             resilience: None,
             shadow: None,
             pii: None,
+            trace_content: false,
             semantic_cache: None,
             prompts: None,
             usage_parser: "auto".to_string(),
@@ -942,6 +954,7 @@ mod tests {
             resilience: None,
             shadow: None,
             pii: None,
+            trace_content: false,
             semantic_cache: None,
             prompts: None,
             usage_parser: "auto".to_string(),