Merge pull request '修正 Codex Action 重複輸出最後回答' (#4) from develop into master

Reviewed-on: actions/codex#4

.gitea/ai-review/exclusions.json

@@ -148,5 +148,11 @@
     "role": "Maya",
     "original_finding": "對於 codex 執行失敗的各種細節（權限不足、找不到 binary 等）都統一處理為 status: 1，測試未驗證具體錯誤來源。",
     "reason": "此 GitHub Action 對外輸出以 `status=completed|failed` 與 `output` 訊息表達成功或失敗；process exit code 只需反映 Codex 子程序結果或通用啟動失敗。細分 ENOENT、EACCES 等 Action exit code 會改變既有介面且對 workflow 判斷價值有限，本次已用簡潔 output 訊息與測試覆蓋具體錯誤來源。"
+  },
+  {
+    "location": "app/main.js:142",
+    "role": "Rogue",
+    "original_finding": "移除這些無謂的重新編碼比較，直接嘗試解碼。",
+    "reason": "AI 對話收斂判定為誤報（問題在最新程式碼中不成立或不適用）"
   }
 ]

.gitea/ai-review/findings.json

@@ -1 +1,16 @@
-[]
+[
+  {
+    "level": "warning",
+    "role": "Bard",
+    "location": "app/main.js:241",
+    "problem": "`runCodex` 中 `resultOutput` 的計算邏輯過於緊湊且不易維護，且直接進行檔案同步讀取，若檔案過大可能阻塞 Node.js 事件循環。",
+    "suggestion": "建議拆解複雜的邏輯為獨立的輔助函數，並改用非同步的檔案讀取方式。"
+  },
+  {
+    "level": "info",
+    "role": "Bard",
+    "location": "app/main.js:321",
+    "problem": "確保字串結尾換行的樣板程式碼邏輯重複出現。",
+    "suggestion": "建立 `ensureNewline` 共用工具函式。"
+  }
+]

app/main.js

@@ -183,11 +183,13 @@ function readExecutionConfig() {
   };
 }
 
-function codexExecArgs(model, prompt) {
+function codexExecArgs(model, prompt, lastMessageFile) {
   return [
     "exec",
     "--dangerously-bypass-approvals-and-sandbox",
     "--skip-git-repo-check",
+    "--output-last-message",
+    lastMessageFile,
     "--model",
     model,
     prompt,
@@ -197,7 +199,10 @@ function codexExecArgs(model, prompt) {
 function runCodex(model, prompt) {
   return new Promise((resolve) => {
     const { timeoutMs, outputLimitBytes, workspace } = readExecutionConfig();
-    const args = codexExecArgs(model, prompt);
+    const tempDir = makeTempDir(process.env.CODEX_HOME || "/root/.codex");
+    const lastMessageFile = path.join(tempDir, "last-message.txt");
+    tempFiles.trackFile(lastMessageFile);
+    const args = codexExecArgs(model, prompt, lastMessageFile);
 
     // This Docker Action runs inside an ephemeral CI container where Codex must be
     // able to edit the checked-out workspace without interactive approvals.
@@ -210,9 +215,6 @@ function runCodex(model, prompt) {
       output.append(Buffer.from(`Codex execution timed out after ${timeoutMs} ms.\n`));
     }, timeoutMs);
 
-    child.stdout.pipe(process.stdout);
-    child.stderr.pipe(process.stdout);
-
     child.stdout.on("data", (chunk) => {
       output.append(chunk);
     });
@@ -235,7 +237,11 @@ function runCodex(model, prompt) {
         output.append(Buffer.from(`Codex process terminated by signal ${signal}.\n`));
       }
 
-      resolve({ status: code ?? 1, output: output.toString() });
+      const lastMessage = fs.existsSync(lastMessageFile) ? fs.readFileSync(lastMessageFile, "utf8") : null;
+      const diagnosticOutput = output.toString();
+      const resultOutput = code === 0 && lastMessage !== null ? lastMessage : diagnosticOutput || lastMessage || "";
+
+      resolve({ status: code ?? 1, output: resultOutput });
     });
   });
 }
@@ -312,6 +318,9 @@ function setupAuth(oauth, codexHome) {
 async function runCodexAction({ oauth, model, codexHome, prompt }) {
   const lockHandle = setupAuth(oauth, codexHome);
   const result = await runCodex(model, prompt);
+  if (result.output) {
+    process.stdout.write(result.output.endsWith("\n") ? result.output : `${result.output}\n`);
+  }
   appendGithubOutput(result.status === 0 ? "completed" : "failed", result.output);
 
   if (lockHandle !== undefined) {

tests/entrypoint_test.sh

@@ -56,6 +56,26 @@ SH
   chmod +x "$dir/codex"
 }
 
+make_fake_transcript_codex() {
+  local dir="$1"
+
+  cat > "$dir/codex" <<'SH'
+#!/bin/sh
+last_message_file=""
+while [ "$#" -gt 0 ]; do
+  if [ "$1" = "--output-last-message" ]; then
+    shift
+    last_message_file="$1"
+  fi
+  shift
+done
+
+printf '%s\n' 'codex' 'final answer' 'tokens used' '123' 'final answer' >&2
+printf '%s\n' 'final answer' > "$last_message_file"
+SH
+  chmod +x "$dir/codex"
+}
+
 encoded_json() {
   printf '%s' "$1" | base64 | tr -d '\n'
 }
@@ -269,6 +289,20 @@ test_codex_output_truncation() {
   rm -rf "$tmpdir"
 }
 
+test_last_message_output_without_transcript_duplicate() {
+  local tmpdir
+  tmpdir="$(mktemp -d)"
+  make_fake_transcript_codex "$tmpdir"
+
+  run_entrypoint "$tmpdir" OAUTH="$(encoded_json '{}')" MODEL="gpt-test" PROMPT="hello"
+
+  [[ "$(grep -c "final answer" "$tmpdir/stdout")" -eq 1 ]] || die "stdout printed final answer more than once"
+  grep -q "status=completed" "$tmpdir/github-output" || die "last message completed status"
+  [[ "$(grep -c "final answer" "$tmpdir/github-output")" -eq 1 ]] || die "github output included transcript duplicate"
+  ! grep -q "tokens used" "$tmpdir/github-output" || die "github output included codex transcript"
+  rm -rf "$tmpdir"
+}
+
 test_invalid_execution_config_uses_defaults() {
   local tmpdir
   tmpdir="$(mktemp -d)"
@@ -299,6 +333,7 @@ test_missing_codex_command_output
 test_codex_timeout_output
 test_codex_signal_output
 test_codex_output_truncation
+test_last_message_output_without_transcript_duplicate
 test_invalid_execution_config_uses_defaults
 
 echo "entrypoint tests passed"