diff --git a/.changeset/preserve-file-encoding.md b/.changeset/preserve-file-encoding.md
index 034e6540f3..2bbfe07f03 100644
--- a/.changeset/preserve-file-encoding.md
+++ b/.changeset/preserve-file-encoding.md
@@ -2,7 +2,7 @@
 "@kilocode/cli": minor
 ---
 
-The agent now detects and preserves the original text encoding of files when reading and editing them, so non-UTF-8 files are displayed correctly to the model and written back in their original encoding.
+The agent now detects and preserves the original text encoding of files when reading and editing them, so non-UTF-8 files are displayed correctly to the model and written back in their original encoding. New files are still created as UTF-8 without BOM — detection only applies when overwriting or editing an existing file.
 
 Supported: UTF-8 (with or without BOM), UTF-16 with BOM, and common legacy Latin and CJK encodings (Shift_JIS, EUC-JP, GB2312, Big5, EUC-KR, Windows-1251, KOI8-R, ISO-8859, and others).
 
diff --git a/packages/opencode/src/kilocode/encoding.ts b/packages/opencode/src/kilocode/encoding.ts
index f846da9871..70fc51330e 100644
--- a/packages/opencode/src/kilocode/encoding.ts
+++ b/packages/opencode/src/kilocode/encoding.ts
@@ -109,11 +109,14 @@ export namespace Encoding {
   export function encode(text: string, encoding: string): Buffer {
     // iconv-lite's UTF codecs strip/ignore BOMs, but we support "UTF-X with BOM"
     // as a distinct variant. Prepend the BOM manually so round-tripping keeps
-    // the original byte signature intact.
-    if (encoding === UTF8_BOM) return Buffer.concat([UTF8_BOM_BYTES, iconv.encode(text, "utf-8")])
+    // the original byte signature intact. Strip a leading U+FEFF from `text`
+    // first so we never emit a double BOM when the decoded text already
+    // contains one (e.g. if a tool round-trips content verbatim).
+    const body = text.charCodeAt(0) === 0xfeff ? text.slice(1) : text
+    if (encoding === UTF8_BOM) return Buffer.concat([UTF8_BOM_BYTES, iconv.encode(body, "utf-8")])
     const lower = encoding.toLowerCase()
-    if (lower === "utf-16le") return Buffer.concat([Buffer.from([0xff, 0xfe]), iconv.encode(text, encoding)])
-    if (lower === "utf-16be") return Buffer.concat([Buffer.from([0xfe, 0xff]), iconv.encode(text, encoding)])
+    if (lower === "utf-16le") return Buffer.concat([Buffer.from([0xff, 0xfe]), iconv.encode(body, encoding)])
+    if (lower === "utf-16be") return Buffer.concat([Buffer.from([0xfe, 0xff]), iconv.encode(body, encoding)])
     return iconv.encode(text, encoding)
   }
 
diff --git a/packages/opencode/test/kilocode/tool-encoding.test.ts b/packages/opencode/test/kilocode/tool-encoding.test.ts
index 2ac56a8d00..1e03493fd6 100644
--- a/packages/opencode/test/kilocode/tool-encoding.test.ts
+++ b/packages/opencode/test/kilocode/tool-encoding.test.ts
@@ -223,6 +223,35 @@ describe("tool encoding preservation", () => {
         }),
       ),
     )
+
+    // Guard against double-BOM regressions: if the model ever hands back content
+    // that already starts with U+FEFF (e.g. by round-tripping literal bytes),
+    // writing it to a BOM-encoded file must still produce exactly one BOM.
+    const bomCases: Array<[string, string, Buffer]> = [
+      ["UTF-8 with BOM", UTF8_BOM, Buffer.from([0xef, 0xbb, 0xbf])],
+      ["UTF-16 LE", "utf-16le", Buffer.from([0xff, 0xfe])],
+      ["UTF-16 BE", "utf-16be", Buffer.from([0xfe, 0xff])],
+    ]
+    for (const [label, encoding, bom] of bomCases) {
+      it.live(`does not emit a double BOM for ${label} when content starts with U+FEFF`, () =>
+        provideTmpdirInstance((dir) =>
+          Effect.gen(function* () {
+            const filepath = path.join(dir, "file.txt")
+            yield* putEncoded(filepath, "hello", encoding)
+            yield* markRead(filepath)
+
+            yield* runWrite({ filePath: filepath, content: "\uFEFFgoodbye" })
+
+            const bytes = yield* loadBytes(filepath)
+            // Exactly one BOM prefix, immediately followed by encoded payload.
+            expect(bytes.subarray(0, bom.length).equals(bom)).toBe(true)
+            expect(bytes.subarray(bom.length, bom.length * 2).equals(bom)).toBe(false)
+            const decoded = yield* loadDecoded(filepath, encoding)
+            expect(decoded).toBe("goodbye")
+          }),
+        ),
+      )
+    }
   })
 
   describe("EditTool preserves existing file encoding across edits", () => {