release: v2.0.6 - Intelligence & UX Optimization (2026-05-14)

src/features/company/dispatcher.ts

@@ -284,7 +284,8 @@ async function _dispatchOne(
         // hit disk / shell. The report (e.g. "✅ Created: foo.py") is
         // appended to the response so the user sees what really happened.
         let finalResponse = rawResponse || '_(empty response)_';
-        if (rawResponse && deps.executeActionTags && _hasActionTag(rawResponse)) {
+        const hasTag = !!rawResponse && _hasActionTag(rawResponse);
+        if (rawResponse && deps.executeActionTags && hasTag) {
             try {
                 const report = await deps.executeActionTags(rawResponse);
                 if (report.length > 0) {
@@ -297,12 +298,30 @@ async function _dispatchOne(
                 logError('company.dispatcher: action-tag execution failed.', { agentId, err });
                 finalResponse = `${rawResponse}\n\n---\n⚠️ Action 실행 실패: ${err}`;
             }
+        } else if (rawResponse && !hasTag && _claimsFileCreation(rawResponse)) {
+            // Hallucination guard: small models love to *narrate* file
+            // creation ("foo.py를 생성했습니다 …") without emitting the
+            // <create_file> tag — so the user sees ✅ in chat but nothing
+            // on disk. Catch the mismatch here and flag it loudly so the
+            // CEO synthesis (which reads this response) and the user both
+            // know nothing was actually written.
+            const warning = '⚠️ **실제 파일이 생성되지 않았습니다.** Agent가 파일 생성을 텍스트로 설명했지만 ConnectAI 액션 태그(`<create_file>` 등)를 사용하지 않아 디스크에 아무것도 만들어지지 않았어요. 같은 요청을 다시 시도하거나, 사용자가 직접 만드세요.';
+            finalResponse = `${rawResponse}\n\n---\n${warning}`;
+            logInfo('company.dispatcher: agent claimed creation without action tag.', { agentId });
         }
+        // `error: 'no-action-tag-but-claimed'` is *advisory* — we still let
+        // the turn complete because some agents (Writer, Researcher) are
+        // legitimately answer-only. But by flagging the agent output we
+        // mark it as not-fully-successful so the CEO synthesis can read
+        // the warning verbatim.
+        const claimedButDidnt = rawResponse && !hasTag && _claimsFileCreation(rawResponse);
         return {
             agentId, task,
             response: finalResponse,
             durationMs: Date.now() - startedAt,
-            error: rawResponse ? undefined : 'empty-response',
+            error: rawResponse
+                ? (claimedButDidnt ? 'claimed-creation-no-tag' : undefined)
+                : 'empty-response',
         };
     } catch (e: any) {
         const err = e?.message ?? String(e);
@@ -325,3 +344,27 @@ async function _dispatchOne(
 function _hasActionTag(text: string): boolean {
     return /<\s*(?:create_file|edit_file|delete_file|read_file|list_files|list_brain|run_command|read_brain|reveal_in_explorer|open_file|glob|grep)\b/i.test(text);
 }
+
+/**
+ * Heuristic: does the response *narrate* having created files/folders?
+ *
+ * We look for the combination of (a) a Korean / English creation verb and
+ * (b) a filename-like or "folder" mention. The intent is to catch the
+ * hallucination pattern where an agent writes "foo.py 파일을 생성했습니다"
+ * or "Created `bar/` directory" without emitting the corresponding
+ * `<create_file>` tag, so the dispatcher can flag it back to the CEO and
+ * the user instead of silently reporting success.
+ *
+ * Kept narrow on purpose — a *plan* like "다음에는 X를 만들어야 합니다"
+ * shouldn't trigger this. We require past-tense / completion phrasing.
+ */
+function _claimsFileCreation(text: string): boolean {
+    // Past-tense creation verbs (Korean + English).
+    const claimRe = /(?:생성했|만들었|작성했|저장했|구현했|created|wrote|saved|built|generated)/i;
+    if (!claimRe.test(text)) return false;
+    // Combined with either an explicit filename (something.ext) or the word
+    // "폴더" / "directory" / "folder" near the verb.
+    const fileLike = /\b[\w\-./]+\.(?:py|js|ts|tsx|jsx|md|json|html|css|sh|yaml|yml|sql|java|go|rs|c|cpp|rb|php)\b/i.test(text);
+    const folderLike = /(?:폴더|디렉토리|directory|folder)/i.test(text);
+    return fileLike || folderLike;
+}