diff --git a/.jules/bolt.md b/.jules/bolt.md
index 52f429b..a734fa5 100644
--- a/.jules/bolt.md
+++ b/.jules/bolt.md
@@ -16,4 +16,7 @@
 **Action:** Use `strings.NewReader(str)` directly instead of `bytes.NewReader([]byte(str))` to eliminate the allocation overhead.
 ## 2026-09-11 - Memoize Dynamic Library Load (`ctypes.CDLL`)
 **Learning:** Found that `ctypes.CDLL(None, use_errno=True)` within `pidfd_open` and `pidfd_send_signal` fallback paths in `_sgt-process-token.py` re-loaded the C library dynamically on every call, causing measurable per-call overhead, analogous to the learning about `libc_pidfd_function` from 2024-08-27.
-**Action:** Substituted the redundant inline `ctypes.CDLL` loads with the memoized module-level `_LIBC_WITH_ERRNO` global to eliminate repeated linking overhead.
\ No newline at end of file
+**Action:** Substituted the redundant inline `ctypes.CDLL` loads with the memoized module-level `_LIBC_WITH_ERRNO` global to eliminate repeated linking overhead.
+## 2026-09-20 - Avoid mime.ParseMediaType for header matching in hot paths
+**Learning:** Using `mime.ParseMediaType` forces memory allocations due to string splitting and map creation, slowing down hot paths like response stream parsing.
+**Action:** Use `strings.HasPrefix(strings.ToLower(strings.TrimSpace(header)), target)` instead of `mime.ParseMediaType` when checking content types to prevent allocation overhead, ensuring standard compliance through case-insensitivity.
diff --git a/cmd/sergeant-mcp-client/proxy.go b/cmd/sergeant-mcp-client/proxy.go
index 2fe4e65..05f9514 100644
--- a/cmd/sergeant-mcp-client/proxy.go
+++ b/cmd/sergeant-mcp-client/proxy.go
@@ -5,7 +5,7 @@ import (
 	"bytes"
 	"fmt"
 	"io"
-	"mime"
+
 	"net/http"
 	"strings"
 	"sync"
@@ -175,10 +175,10 @@ func (p *proxy) writeResponse(resp *response) {
 		return
 	}
 
-	mediaType, _, _ := mime.ParseMediaType(resp.contentType)
 	p.outMu.Lock()
 	defer p.outMu.Unlock()
-	if mediaType == "text/event-stream" {
+	// Use strings.HasPrefix on lowercased header instead of mime.ParseMediaType to prevent allocation overhead
+	if strings.HasPrefix(strings.ToLower(strings.TrimSpace(resp.contentType)), "text/event-stream") {
 		for _, data := range parseSSEData(resp.body) {
 			fmt.Fprintf(p.stdout, "%s\n", data)
 		}