diff --git a/Sources/NucleicCore/Claude/ClaudeTokenProxy.swift b/Sources/NucleicCore/Claude/ClaudeTokenProxy.swift
index b55f6fde..507396b2 100644
--- a/Sources/NucleicCore/Claude/ClaudeTokenProxy.swift
+++ b/Sources/NucleicCore/Claude/ClaudeTokenProxy.swift
@@ -154,14 +154,16 @@ public actor ClaudeTokenProxy {
         }
     }
 
-    /// Serialize an upstream response back to the client: status line, headers (hop-by-hop stripped,
-    /// `Connection: close` added, `Content-Length`/`Transfer-Encoding` dropped since we stream and
-    /// close), then the body chunks as they arrive.
+    /// Serialize an upstream response back to the client: status line, headers (hop-by-hop and stale
+    /// representation framing stripped), `Connection: close`, then the body chunks as they arrive.
+    /// URLSession transparently decodes compressed upstream bodies, so forwarding the original
+    /// `Content-Encoding` would make Claude's fetch decode those bytes a second time (`ZlibError`).
     static func writeResponse(_ response: UpstreamResponse, to writer: ByteWriter) async {
         var head = "HTTP/1.1 \(response.status) \(reasonPhrase(response.status))\r\n"
         for (name, value) in response.headers {
             let lower = name.lowercased()
-            if lower == "content-length" || lower == "transfer-encoding" || lower == "connection" { continue }
+            if lower == "content-length" || lower == "transfer-encoding"
+                || lower == "content-encoding" || lower == "connection" { continue }
             head += "\(name): \(value)\r\n"
         }
         head += "Connection: close\r\n\r\n"
@@ -238,6 +240,11 @@ public actor ClaudeTokenProxy {
         for (name, value) in request.headers {
             urlRequest.setValue(value, forHTTPHeaderField: name)
         }
+        // Claude's fetch advertises gzip/br. URLSession decodes compressed responses before exposing
+        // their bytes but can retain the upstream Content-Encoding header, which would make the
+        // downstream fetch decompress the decoded stream again. Request identity as the primary
+        // defense; writeResponse also drops stale encoding metadata defensively.
+        urlRequest.setValue("identity", forHTTPHeaderField: "Accept-Encoding")
         if !request.body.isEmpty { urlRequest.httpBody = request.body }
 
         #if canImport(FoundationNetworking)
diff --git a/Tests/NucleicCoreTests/ClaudeTokenProxyTests.swift b/Tests/NucleicCoreTests/ClaudeTokenProxyTests.swift
index e44ffb59..e98cf3a3 100644
--- a/Tests/NucleicCoreTests/ClaudeTokenProxyTests.swift
+++ b/Tests/NucleicCoreTests/ClaudeTokenProxyTests.swift
@@ -152,7 +152,11 @@ struct ClaudeTokenProxyTests {
         let stub = UpstreamStub { _ in
             Self.response(
                 status: 200,
-                headers: [("content-type", "text/event-stream"), ("content-length", "999")],
+                headers: [
+                    ("content-type", "text/event-stream"),
+                    ("content-encoding", "gzip"),
+                    ("content-length", "999"),
+                ],
                 chunks: ["event: a\ndata: 1\n\n", "event: b\ndata: 2\n\n"])
         }
         let proxy = ClaudeTokenProxy(tokenProvider: { "t" }, upstream: stub.upstream())
@@ -167,6 +171,8 @@ struct ClaudeTokenProxyTests {
         #expect(out.contains("Connection: close"))
         // We stream + close, so the upstream framing headers are dropped (they'd mislead the client).
         #expect(!out.lowercased().contains("content-length: 999"))
+        // URLSession has already decoded the body; advertising gzip makes fetch double-decode it.
+        #expect(!out.lowercased().contains("content-encoding"))
         #expect(out.contains("event: a\ndata: 1\n\n"))
         #expect(out.contains("event: b\ndata: 2\n\n"))
     }
@@ -208,7 +214,12 @@ struct ClaudeTokenProxyTests {
     #if canImport(Network) || os(Linux)
     @Test func loopbackListenerInjectsTokenEndToEnd() async throws {
         let stub = UpstreamStub { _ in
-            Self.response(status: 200, headers: [("content-type", "text/plain")], chunks: ["ok"])
+            // Model URLSession's decoded bytes plus the stale upstream gzip header that caused
+            // Claude/undici to throw ZlibError before the proxy normalized response metadata.
+            Self.response(
+                status: 200,
+                headers: [("content-type", "text/plain"), ("content-encoding", "gzip")],
+                chunks: ["ok"])
         }
         let proxy = ClaudeTokenProxy(tokenProvider: { "fresh" }, upstream: stub.upstream())
         let port = try await proxy.start()