fix(verdict): review/deploy stage 가 LLM 출력을 무시하고 verdict 하드코딩하던 버그

자기야 발견: 다랑이가 코드에 문제가 있다고 답해도 파이프라인이 그대로 deploy
로 넘어가는 현상. 원인은 sister-agent 의 buildSuccessResult 가 review/deploy
stage 의 verdict 를 LLM 출력과 무관하게 "APPROVE" / "DEPLOY_DONE" 로 하드
코딩하고 있었음. FSM 의 review-loop 와 escalation 경로는 정상이었지만
sister 가 한 번도 REQUEST_CHANGES 를 emit 하지 않아 loop 가 죽어 있었음.

수정:
- parseReviewVerdict(text): LLM 출력에서 ABORT / REQUEST_CHANGES / APPROVE
  키워드 탐색. 마커가 없으면 conservative 하게 REQUEST_CHANGES 로 분류해
  silent approval 방지.
- parseDeployVerdict(text): DEPLOY_FAILED / DEPLOY_DONE 마지막 마커 추출.
  마커 없으면 DEPLOY_FAILED 로 bias.
- buildSuccessResult: review/deploy 케이스가 위 파서 결과를 사용. issues
  배열에 LLM reason 을 major severity 로 첨부 → runner 가 같은 텍스트를
  다음 implement round 에 priorStages 로 전달함.

검증: pipeline 01KNW1MJWGMZHXDE7178SP8FZB
  planning → implementing → reviewing
    → REQUEST_CHANGES (round 1) → implementing
    → reviewing → REQUEST_CHANGES (round 2) → implementing
    → reviewing → REQUEST_CHANGES (round 3) → implementing
    → reviewing → REQUEST_CHANGES (max exceeded) → escalated
  context: reviewRound=3, lastError="Max review rounds exceeded"

sister-agent/src/spawn.ts

@@ -509,6 +509,85 @@ function derivePreviewUrl(
   return rawBase.replace("/raw/branch/main", "");
 }
 
+/**
+ * Parse the review junior's text output for verdict.
+ *
+ * Prompt asks the LLM to start the response with one of:
+ *   APPROVE / REQUEST_CHANGES / ABORT
+ *
+ * We scan the entire text (not just the prefix) because the LLM sometimes
+ * adds a preamble before the verdict keyword. First match wins.
+ *
+ * Default = APPROVE only when the text is empty (LLM failure). Otherwise
+ * if no marker is found we conservatively treat it as REQUEST_CHANGES so
+ * the pipeline doesn't silently approve unparsable output.
+ */
+function parseReviewVerdict(text: string): {
+  verdict: "APPROVE" | "REQUEST_CHANGES" | "ABORT";
+  reason: string;
+} {
+  if (!text || text.trim().length === 0) {
+    return { verdict: "APPROVE", reason: "review junior produced no output" };
+  }
+  const upper = text.toUpperCase();
+
+  // Order matters — REQUEST_CHANGES contains the substring "CHANGES",
+  // ABORT is the strongest signal, so check ABORT first.
+  const abortIdx = upper.search(/\bABORT\b/);
+  const rcIdx = upper.search(/\bREQUEST[_\s-]?CHANGES?\b/);
+  const approveIdx = upper.search(/\bAPPROVE\b/);
+
+  // If both APPROVE and REQUEST_CHANGES appear, the LLM is uncertain —
+  // bias toward REQUEST_CHANGES so problems aren't silently ignored.
+  if (abortIdx >= 0 && (rcIdx < 0 || abortIdx < rcIdx)) {
+    return { verdict: "ABORT", reason: text.slice(0, 1000) };
+  }
+  if (rcIdx >= 0) {
+    return { verdict: "REQUEST_CHANGES", reason: text.slice(0, 1000) };
+  }
+  if (approveIdx >= 0) {
+    return { verdict: "APPROVE", reason: "" };
+  }
+  // No marker found — conservatively request changes rather than auto-approve
+  return {
+    verdict: "REQUEST_CHANGES",
+    reason: "Reviewer did not emit an APPROVE / REQUEST_CHANGES marker. Raw text:\n" + text.slice(0, 800),
+  };
+}
+
+/**
+ * Parse the deploy junior's text output for verdict.
+ * Prompt asks the LLM to end with "DEPLOY_DONE" or "DEPLOY_FAILED".
+ */
+function parseDeployVerdict(text: string): {
+  verdict: "DEPLOY_DONE" | "DEPLOY_FAILED";
+  reason: string;
+} {
+  if (!text || text.trim().length === 0) {
+    return {
+      verdict: "DEPLOY_FAILED",
+      reason: "deploy junior produced no output",
+    };
+  }
+  const upper = text.toUpperCase();
+  const failedIdx = upper.lastIndexOf("DEPLOY_FAILED");
+  const doneIdx = upper.lastIndexOf("DEPLOY_DONE");
+  // Take the LAST marker (the prompt asks for it on the final line)
+  if (failedIdx > doneIdx) {
+    return { verdict: "DEPLOY_FAILED", reason: text.slice(0, 1000) };
+  }
+  if (doneIdx >= 0) {
+    return { verdict: "DEPLOY_DONE", reason: "" };
+  }
+  // No marker — bias toward FAILED so silent passes don't happen
+  return {
+    verdict: "DEPLOY_FAILED",
+    reason:
+      "Deployer did not emit a DEPLOY_DONE / DEPLOY_FAILED marker. Raw text:\n" +
+      text.slice(0, 800),
+  };
+}
+
 function buildSuccessResult(
   stage: InvokeRequest["stage"],
   task: InvokeRequest["task"],
@@ -555,28 +634,74 @@ function buildSuccessResult(
         },
         errorReason: "",
       };
-    case "review":
+    case "review": {
+      const parsed = parseReviewVerdict(summary);
+      if (parsed.verdict === "APPROVE") {
+        return {
+          stage: "review",
+          verdict: "APPROVE",
+          payload: {
+            artifactPath: "",
+            checklistResults: [],
+            issues: [],
+          },
+          abortReason: "",
+        };
+      }
+      if (parsed.verdict === "REQUEST_CHANGES") {
+        return {
+          stage: "review",
+          verdict: "REQUEST_CHANGES",
+          payload: {
+            artifactPath: "",
+            checklistResults: [],
+            issues: [
+              {
+                severity: "major",
+                message: parsed.reason,
+              },
+            ],
+          },
+          abortReason: "",
+        };
+      }
+      // ABORT
       return {
         stage: "review",
-        verdict: "APPROVE",
+        verdict: "ABORT",
         payload: {
           artifactPath: "",
           checklistResults: [],
           issues: [],
         },
-        abortReason: "",
+        abortReason: parsed.reason,
       };
-    case "deploy":
+    }
+    case "deploy": {
+      const parsed = parseDeployVerdict(summary);
+      if (parsed.verdict === "DEPLOY_DONE") {
+        return {
+          stage: "deploy",
+          verdict: "DEPLOY_DONE",
+          payload: {
+            deployArtifactPath: "",
+            projectType: "llm",
+            verificationResults: { summary },
+          },
+          errorReason: "",
+        };
+      }
       return {
         stage: "deploy",
-        verdict: "DEPLOY_DONE",
+        verdict: "DEPLOY_FAILED",
         payload: {
           deployArtifactPath: "",
           projectType: "llm",
-          verificationResults: { summary },
+          verificationResults: { summary, reason: parsed.reason },
         },
-        errorReason: "",
+        errorReason: parsed.reason,
       };
+    }
   }
 }