fix(ingest): fall back to gen_ai.conversation.id for every vendor

apps/ingest/src/ai_session.rs

@@ -35,10 +35,11 @@
 //!
 //! Detection is ordered first-match over the vendor predicates below; the
 //! session ID is the first non-empty session-granularity attribute for the
-//! matched vendor. A vendor with no session-level key of its own (its
-//! instrumentation only emits run/user-scoped IDs, or nothing) and every
-//! unknown-tier bucket read the OTel GenAI key, `gen_ai.conversation.id`,
-//! which the public docs give any emitter as the way to group its traces.
+//! matched vendor. Every vendor and unknown-tier bucket falls back to the OTel
+//! GenAI key, `gen_ai.conversation.id`, after its own keys: the public docs
+//! give any emitter that key as the way to group its traces, and for a vendor
+//! with no session-level key of its own (its instrumentation only emits
+//! run/user-scoped IDs, or nothing) it is the only one.
 //!
 //! One vendor is not a framework: `maple` matches any span carrying a
 //! `maple_ai.session.id` attribute. That is the one key an emitter both writes
@@ -842,12 +843,15 @@ struct Vendor {
     id: &'static str,
     detect: DetectFn,
     /// Session-granularity span attribute keys; the first non-empty value wins.
+    /// [`CONVERSATION_ID_KEY`] is tried last unless the list ranks it itself.
     session_keys: &'static [&'static str],
 }
 
-/// The session key of a dialect with none of its own: the OTel GenAI
-/// conversation id, which the docs tell every emitter to set.
-const CONVERSATION_ID_ONLY: &[&str] = &["gen_ai.conversation.id"];
+/// The OTel GenAI conversation id, which the docs tell every emitter to set.
+const CONVERSATION_ID_KEY: &str = "gen_ai.conversation.id";
+
+/// The session keys of a dialect with none of its own.
+const CONVERSATION_ID_ONLY: &[&str] = &[CONVERSATION_ID_KEY];
 
 /// Ordered: first match wins. `maple` leads because its key is an explicit
 /// opt-in rather than a framework fingerprint (see the module doc). Then
@@ -999,7 +1003,10 @@ static UNKNOWN_TIER: &[Vendor] = &[
     Vendor {
         id: "unknown:openinference",
         detect: detect_unknown_openinference,
-        session_keys: CONVERSATION_ID_ONLY,
+        // `session.id` only after the conversation id: Maple's browser SDK
+        // stamps its replay session under that key on every span, so it can
+        // span several conversations.
+        session_keys: &[CONVERSATION_ID_KEY, "session.id"],
     },
     Vendor {
         id: "unknown:other",
@@ -1039,9 +1046,13 @@ fn run_predicates(
         .iter()
         .chain(UNKNOWN_TIER)
         .find(|vendor| (vendor.detect)(&ctx))?;
+    let fallback =
+        (!vendor.session_keys.contains(&CONVERSATION_ID_KEY)).then_some(CONVERSATION_ID_KEY);
     let session_id = vendor
         .session_keys
         .iter()
+        .copied()
+        .chain(fallback)
         .find_map(|key| session_value(span_attrs, key));
     Some(AiClassification {
         vendor: vendor.id,
@@ -1526,6 +1537,36 @@ mod tests {
         );
     }
 
+    #[test]
+    fn vendors_with_their_own_key_fall_back_to_the_conversation_id() {
+        for (scope, span_name, vendor) in [
+            ("crewai.telemetry", "Crew.kickoff", "crewai"),
+            ("strands.telemetry.tracer", "invoke_agent", "strands"),
+            ("openinference.instrumentation.agno", "agent.run", "agno"),
+        ] {
+            classified(
+                scope,
+                span_name,
+                &[("gen_ai.conversation.id", "conv-8")],
+                &[],
+                vendor,
+                Some("conv-8"),
+            );
+            // The vendor's own key still wins when both are set.
+            classified(
+                scope,
+                span_name,
+                &[
+                    ("gen_ai.conversation.id", "conv-8"),
+                    ("session.id", "own-1"),
+                ],
+                &[],
+                vendor,
+                Some("own-1"),
+            );
+        }
+    }
+
     #[test]
     fn vendor_matched_but_session_key_absent_or_empty() {
         classified(
@@ -2331,6 +2372,28 @@ mod tests {
             "unknown:openinference",
             Some("conv-9"),
         );
+        // OpenInference's `session.id` fills in only without a conversation id:
+        // a browser replay session under that key must not merge conversations.
+        classified(
+            "",
+            "llm",
+            &[
+                ("openinference.span.kind", "LLM"),
+                ("gen_ai.conversation.id", "conv-9"),
+                ("session.id", "browser-1"),
+            ],
+            &[],
+            "unknown:openinference",
+            Some("conv-9"),
+        );
+        classified(
+            "",
+            "llm",
+            &[("openinference.span.kind", "LLM"), ("session.id", "oi-1")],
+            &[],
+            "unknown:openinference",
+            Some("oi-1"),
+        );
         classified(
             "",
             "task",

packages/agent-sessions/src/session-turns.ts

@@ -402,10 +402,10 @@ function findAnchors(ordered: readonly AiSessionSpan[]): readonly TurnAnchor[] {
 	const byConversation = new Map<string, AiSessionSpan>()
 	for (const span of ordered) {
 		const conversationId = span.genAi.conversationId
-		// Six vendors (flue, google_adk, mastra, microsoft_agent_framework,
-		// openai_agents_sdk, pydantic_ai) derive `maple_ai.session.id` FROM
-		// `gen_ai.conversation.id`, so for them the id names the session and
-		// repeats on every span — a partition of one, not a turn key.
+		// The gateway derives `maple_ai.session.id` FROM `gen_ai.conversation.id`
+		// whenever a vendor's own session key is absent (and for some vendors
+		// ahead of it), so there the id names the session and repeats on every
+		// span — a partition of one, not a turn key.
 		if (conversationId === undefined || sessionIds.has(conversationId)) continue
 		if (!byConversation.has(conversationId)) byConversation.set(conversationId, span)
 	}