fix(kiro): stabilize conversationId across prompt compression (#2630)

Integrated into release/v3.8.3

open-sse/handlers/chatCore.ts

@@ -2106,6 +2106,7 @@ export async function handleChatCore({
   const allMessages = compressionBody?.messages || body?.contents || body?.request?.contents || [];
   let cavemanOutputModeApplied = false;
   let cavemanOutputModeIntensity: string | null = null;
+  let preCompressionBody: typeof body | null = null;
   if (body && Array.isArray(allMessages) && allMessages.length > 0) {
     let estimatedTokens = estimateTokens(JSON.stringify(allMessages));
     let promptCompressionEnabled = false;
@@ -2557,6 +2558,10 @@ export async function handleChatCore({
       `Checking compression: ${estimatedTokens} tokens vs ${threshold} threshold (${contextLimit} limit, ${reservedTokens} reserved)`
     );
 
+    // Capture pre-compression body so translators can access original message
+    // content even after compression alters it (e.g. stable Kiro conversationId).
+    preCompressionBody = body;
+
     if (promptCompressionEnabled && estimatedTokens > threshold) {
       log?.info?.(
         "CONTEXT",
@@ -2933,6 +2938,7 @@ export async function handleChatCore({
           preserveDeveloperRole,
           preserveCacheControl,
           signatureNamespace: connectionId,
+          ...(preCompressionBody ? { preCompressionBody } : {}),
         }
       );
     }

open-sse/translator/index.ts

@@ -132,6 +132,7 @@ export function translateRequest(
     preserveDeveloperRole?: boolean;
     preserveCacheControl?: boolean;
     signatureNamespace?: string | null;
+    preCompressionBody?: Record<string, unknown> | null;
   }
 ) {
   let result = body;
@@ -185,12 +186,16 @@ export function translateRequest(
       if (targetFormat !== FORMATS.OPENAI) {
         const fromOpenAI = getRequestTranslator(FORMATS.OPENAI, targetFormat);
         if (fromOpenAI) {
-          const translationCredentials = options?.signatureNamespace
-            ? {
-                ...(credentials && typeof credentials === "object" ? credentials : {}),
-                _signatureNamespace: options.signatureNamespace,
-              }
-            : credentials;
+          const hasNs = options?.signatureNamespace != null;
+          const hasPreCompression = options?.preCompressionBody != null;
+          const translationCredentials =
+            hasNs || hasPreCompression
+              ? {
+                  ...(credentials && typeof credentials === "object" ? credentials : {}),
+                  ...(hasNs ? { _signatureNamespace: options.signatureNamespace } : {}),
+                  ...(hasPreCompression ? { _preCompressionBody: options.preCompressionBody } : {}),
+                }
+              : credentials;
           result = fromOpenAI(model, result, stream, translationCredentials);
         }
       }

open-sse/translator/request/openai-to-kiro.ts

@@ -746,8 +746,33 @@ export function buildKiroPayload(model, body, stream, credentials) {
   // upstream Kiro/AWS conversation context, leaking prior state across
   // sessions. See conversionMessages() above for the `__synthetic` marker.
   const NAMESPACE_KIRO = "34f7193f-561d-4050-bc84-9547d953d6bf";
+
+  // Priority 1: Extract first user message from pre-compression body (passed by chatCore before
+  // compressContext runs). This keeps conversationId stable even when compression alters content.
+  // Priority 2: Deterministic hash from first user message in translated history (fallback).
+  const preCompressionBody = credentials?._preCompressionBody as
+    | Record<string, unknown>
+    | null
+    | undefined;
+  const preCompressionMessages = Array.isArray(preCompressionBody?.messages)
+    ? preCompressionBody.messages
+    : null;
+  const preCompressionFirstUser = preCompressionMessages?.find(
+    (m: Record<string, unknown>) => m.role === "user"
+  );
+  const seedFromPreCompression = preCompressionFirstUser
+    ? typeof preCompressionFirstUser.content === "string"
+      ? preCompressionFirstUser.content
+      : Array.isArray(preCompressionFirstUser.content)
+        ? (preCompressionFirstUser.content as Array<{ type: string; text?: string }>)
+            .filter((b) => b.type === "text")
+            .map((b) => b.text || "")
+            .join(" ")
+        : ""
+    : "";
   const firstRealUserTurn = history.find((h) => h?.userInputMessage?.content && !h.__synthetic);
-  const firstContent = firstRealUserTurn?.userInputMessage?.content || finalContent;
+  const firstContent =
+    seedFromPreCompression || firstRealUserTurn?.userInputMessage?.content || finalContent;
 
   // Use uuidv5 with the hash of the system prompt / first message to maintain AWS Builder ID context cache
   payload.conversationState.conversationId = uuidv5(