fix(truncation): priorStages slice 한도가 너무 짧아 review 가 잘린 코드 받던 버그

자기야 발견: 다랑이 review 결과에 "코드가 끊긴다" 가 반복 등장. 다랑이의
실제 review 메시지를 확인해 보니 narang 이 만든 todo-debug-mode.html 의
<script> 가 `const form = document.getElementById(` 에서 잘려서 반복 round
마다 같은 잘림 지점을 지적했음.

원인 추적:
- narang junior 의 LLM 출력 원본 (junior-01-*.md) 은 깔끔하게 </html> 로
  종료. LLM 자체는 멀쩡.
- spawn.ts buildSuccessResult 의 summary slice(0, 2000) 가 1차로 자름
- 그 위 aggregated slice(0, 6000) 가 0차로 자름
- runner extractStageText 의 review issues slice(0, 2000) 도 추가 한도
- LLM-emit reason 도 spawn.ts parseReviewVerdict 에서 1000/800 자로 잘림

todo HTML 한 파일이 5KB 정도 되니 6000 자 한도에서 3분의 1 잘려나감 →
darang 은 결과적으로 절반짜리 코드를 받음 → 영원히 REQUEST_CHANGES.

수정 (모두 LLM context 200k+ 안에서 안전한 한도):
- spawn.ts aggregated: 6_000 → 64_000
- spawn.ts buildSuccessResult.summary: 2_000 → 64_000
- spawn.ts parseReviewVerdict reason: 1_000 → 16_000
- spawn.ts parseDeployVerdict reason: 800/1000 → 16_000
- runner.ts extractStageText review issues: 2_000 → 32_000

검증: 다음 실행에서 darang 이 동일한 잘림 지점을 지적하지 않으면 OK.

sister-agent/src/spawn.ts

@@ -223,10 +223,16 @@ export async function executeInvocation(
       });
     }
 
+    // Aggregate manager + all child outputs into a single text blob that
+    // gets passed to the next stage as priorStages. The downstream agent
+    // (especially the reviewer) needs to see ACTUAL CODE — not a snippet
+    // — to make a meaningful judgement, so the cap is generous. Cap is
+    // sized for full HTML/JS/CSS files; LLM context windows are 200k+ so
+    // 64KB stays well inside budget even after 4 stages of accumulation.
     const aggregated = [managerWork.text, ...childTexts]
       .filter(Boolean)
       .join("\n\n---\n\n")
-      .slice(0, 6000);
+      .slice(0, 64_000);
 
     return buildSuccessResult(
       req.stage,
@@ -540,10 +546,10 @@ function parseReviewVerdict(text: string): {
   // 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) };
+    return { verdict: "ABORT", reason: text.slice(0, 16_000) };
   }
   if (rcIdx >= 0) {
-    return { verdict: "REQUEST_CHANGES", reason: text.slice(0, 1000) };
+    return { verdict: "REQUEST_CHANGES", reason: text.slice(0, 16_000) };
   }
   if (approveIdx >= 0) {
     return { verdict: "APPROVE", reason: "" };
@@ -551,7 +557,7 @@ function parseReviewVerdict(text: string): {
   // 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),
+    reason: "Reviewer did not emit an APPROVE / REQUEST_CHANGES marker. Raw text:\n" + text.slice(0, 16_000),
   };
 }
 
@@ -574,7 +580,7 @@ function parseDeployVerdict(text: string): {
   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) };
+    return { verdict: "DEPLOY_FAILED", reason: text.slice(0, 16_000) };
   }
   if (doneIdx >= 0) {
     return { verdict: "DEPLOY_DONE", reason: "" };
@@ -584,7 +590,7 @@ function parseDeployVerdict(text: string): {
     verdict: "DEPLOY_FAILED",
     reason:
       "Deployer did not emit a DEPLOY_DONE / DEPLOY_FAILED marker. Raw text:\n" +
-      text.slice(0, 800),
+      text.slice(0, 16_000),
   };
 }
 
@@ -601,7 +607,10 @@ function buildSuccessResult(
   } | null,
   producedFiles: string[] = [],
 ): HandoffMessage {
-  const summary = outputText?.slice(0, 2000) ?? "";
+  // The summary is the payload the next stage will see as priorStages
+  // text. Reviewer needs to see actual code, not a snippet, so the cap
+  // matches the upstream aggregation (64KB).
+  const summary = outputText?.slice(0, 64_000) ?? "";
   switch (stage) {
     case "plan":
       return {

src/orchestrator/runner.ts

@@ -348,7 +348,11 @@ function extractStageText(h: HandoffMessage): string {
     }
     case "review": {
       if (h.payload?.issues && h.payload.issues.length > 0) {
-        return `Review ${h.verdict}: ${JSON.stringify(h.payload.issues).slice(0, 2000)}`;
+        // Generous cap so the next implement loop sees the full reviewer
+        // critique (not just the first 2KB). Reviewer reason text can be
+        // multiple paragraphs and the implement junior needs all of it
+        // to fix the right things.
+        return `Review ${h.verdict}: ${JSON.stringify(h.payload.issues).slice(0, 32_000)}`;
       }
       return `Review verdict: ${h.verdict}`;
     }