@walwal-harness/cli 3.2.1 → 3.3.1

package/assets/templates/HARNESS.md

@@ -12,7 +12,7 @@
 ├── config.json                 # 하네스 설정
 ├── progress.json               # 기계 판독 상태 (세션 오케스트레이션)
 ├── progress.log                # 사람 판독 히스토리 (append-only)
-├── next-prompt.txt             # 다음 세션용 프롬프트 (claude CLI 파이프용)
+├── handoff.json                # 세션 전환 문서 (prompt, model, artifacts, regression 등)
 ├── actions/                    # 현재 활성 문서
 │   ├── pipeline.json           # Dispatcher 결정 (어떤 파이프라인인지)
 │   ├── plan.md                 # 제품 사양
@@ -143,7 +143,7 @@ pending → draft → reviewed → approved
 |------|------|
 | `.harness/progress.json` | 기계 판독 상태 (현재 에이전트, 파이프라인, 실패 정보) |
 | `.harness/progress.log` | 사람 판독 히스토리 (append-only) |
-| `.harness/next-prompt.txt` | 다음 세션용 프롬프트 (claude CLI 파이프용) |
+| `.harness/handoff.json` | 세션 전환 문서 (prompt, model, thinking_mode, artifacts, regression) |
 
 ### 실행 방법
 
@@ -164,7 +164,7 @@ Feature-level 프로그래스를 출력하고 다음 에이전트를 안내합
 /harness-generator-backend
 
 # 방법 B: claude CLI 자동 실행
-claude --prompt "$(cat .harness/next-prompt.txt)"
+claude --model $(jq -r .model .harness/handoff.json) --prompt "$(jq -r .prompt .harness/handoff.json)"
 ```
 
 #### 4. SessionStart 훅 (자동)

package/assets/templates/config.json

@@ -1,7 +1,7 @@
 {
   "harness": {
     "name": "6-Agent Production Harness",
-    "version": "3.2.0",
+    "version": "3.3.0",
     "description": "Dispatcher + NestJS MSA + React/Next.js + Playwright 기반 실무 하네스",
     "source": "https://www.anthropic.com/engineering/harness-design-long-running-apps"
   },
@@ -10,7 +10,9 @@
       "role": "사용자 요청 분석 → 파이프라인 결정 (FE-ONLY / BE-ONLY / FULLSTACK)",
       "skill": "harness-dispatcher",
       "outputs": ["actions/pipeline.json"],
-      "always_first": true
+      "always_first": true,
+      "model": "opus",
+      "thinking_mode": null
     },
     "brainstorming": {
       "role": "사용자의 러프한 요구사항을 대화형으로 구체화하여 Planner가 바로 쓸 수 있는 brainstorm-spec.md로 변환.",
@@ -33,6 +35,9 @@
       },
       "invoked_by": "dispatcher (user opt-in after suggestion)",
       "next_on_complete": "planner",
+      "model": "opus",
+      "thinking_mode": "plan",
+      "thinking_mode_description": "Claude Code plan 모드로 실행. 구조화된 계획 수립 후 실행.",
       "attribution": "Derived from obra/superpowers skills/brainstorming (MIT License)"
     },
     "planner": {
@@ -43,21 +48,30 @@
         "actions/plan.md",
         "actions/feature-list.json",
         "actions/api-contract.json"
-      ]
+      ],
+      "model": "opus",
+      "thinking_mode": "ultraplan",
+      "thinking_mode_description": "최고 수준의 계획 모드. 아키텍처 결정, API 설계, 서비스 분할에 깊은 사고 필요."
     },
     "generator-backend": {
       "role": "NestJS MSA 서비스 구현, API 엔드포인트, DB 스키마",
       "skill": "harness-generator-backend",
       "inputs": ["actions/plan.md", "actions/feature-list.json", "actions/api-contract.json", "actions/sprint-contract.md"],
       "outputs": ["code:backend", "actions/sprint-contract.md"],
-      "order": 1
+      "order": 1,
+      "model": "sonnet",
+      "thinking_mode": null,
+      "model_rationale": "코드 생성은 Sonnet이 비용 대비 효율적. 계획은 Planner가 이미 완료."
     },
     "generator-frontend": {
       "role": "React/Next.js UI 컴포넌트, 상태관리, API 연동",
       "skill": "harness-generator-frontend",
       "inputs": ["actions/plan.md", "actions/feature-list.json", "actions/api-contract.json", "actions/sprint-contract.md"],
       "outputs": ["code:frontend", "actions/sprint-contract.md"],
-      "order": 2
+      "order": 2,
+      "model": "sonnet",
+      "thinking_mode": null,
+      "model_rationale": "코드 생성은 Sonnet이 비용 대비 효율적. 계획은 Planner가 이미 완료."
     },
     "evaluator-functional": {
       "role": "Playwright로 E2E 기능 검증, API 응답/DB 상태 확인",
@@ -65,6 +79,9 @@
       "tools": ["playwright:browser_*"],
       "inputs": ["actions/sprint-contract.md", "actions/feature-list.json", "actions/api-contract.json"],
       "outputs": ["actions/evaluation-functional.md"],
+      "model": "opus",
+      "thinking_mode": "ultrathink",
+      "thinking_mode_description": "최고 수준의 사고 모드. 적대적 검증에 깊은 추론이 필요. 각 항목을 비판적으로 분석.",
       "evaluation_template": "assets/templates/evaluation-functional.md.template",
       "adversarial_rules": {
         "comment": "Evaluator 적대적 행동 규칙 — 통과 편향 제거를 위한 강제 규칙",
@@ -92,6 +109,9 @@
       "tools": ["playwright:browser_take_screenshot", "playwright:browser_resize", "playwright:browser_snapshot"],
       "inputs": ["actions/sprint-contract.md", "actions/evaluation-functional.md", "actions/feature-list.json"],
       "outputs": ["actions/evaluation-visual.md"],
+      "model": "opus",
+      "thinking_mode": "ultrathink",
+      "thinking_mode_description": "최고 수준의 사고 모드. 시각적 결함, 접근성, AI슬롭을 비판적으로 탐지.",
       "evaluation_template": "assets/templates/evaluation-visual.md.template",
       "adversarial_rules": {
         "comment": "Visual Evaluator 적대적 행동 규칙",
@@ -119,7 +139,9 @@
       "inputs": ["actions/plan.md", "actions/feature-list.json", "actions/api-contract.json", "actions/sprint-contract.md"],
       "outputs": ["code:flutter", "actions/sprint-contract.md"],
       "order": 2,
-      "fe_stack": "flutter"
+      "fe_stack": "flutter",
+      "model": "sonnet",
+      "thinking_mode": null
     },
     "evaluator-functional-flutter": {
       "role": "Flutter 앱 검증 — flutter analyze, flutter test, build_runner 일관성, 안티패턴 정적 검증",
@@ -127,7 +149,9 @@
       "tools": ["bash:flutter", "bash:dart"],
       "inputs": ["actions/sprint-contract.md"],
       "outputs": ["actions/evaluation-functional.md"],
-      "fe_stack": "flutter"
+      "fe_stack": "flutter",
+      "model": "opus",
+      "thinking_mode": "ultrathink"
     }
   },
   "flow": {
@@ -288,7 +312,6 @@
     "isolation": true,
     "state_file": ".harness/progress.json",
     "state_log": ".harness/progress.log",
-    "next_prompt_file": ".harness/next-prompt.txt",
     "handoff_file": ".harness/handoff.json",
     "context_guard": {
       "comment": "컨텍스트 분리 하드 가드레일. UserPromptSubmit 훅이 위반을 감지하면 경고를 주입한다.",

package/bin/init.js

@@ -155,10 +155,10 @@ function scaffoldHarness() {
     fs.writeFileSync(progressLog, `# Harness Progress Log\n# ${date} — Initialized\n`);
   }
 
-  // Create next-prompt.txt placeholder
-  const nextPrompt = path.join(HARNESS_DIR, 'next-prompt.txt');
-  if (!fileExists(nextPrompt) || isForce) {
-    fs.writeFileSync(nextPrompt, '');
+  // Create handoff.json placeholder
+  const handoff = path.join(HARNESS_DIR, 'handoff.json');
+  if (!fileExists(handoff) || isForce) {
+    fs.writeFileSync(handoff, '{}');
   }
 
   log('.harness/ scaffolding complete');

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@walwal-harness/cli",
-  "version": "3.2.1",
+  "version": "3.3.1",
   "description": "Production harness for AI agent engineering — Planner, Generator(BE/FE), Evaluator(Func/Visual), optional Brainstormer (requirements refinement). Supports React and Flutter FE stacks.",
   "bin": {
     "walwal-harness": "bin/init.js"
@@ -36,6 +36,6 @@
     "gotchas/"
   ],
   "dependencies": {
-    "@walwal-harness/cli": "^3.2.1"
+    "@walwal-harness/cli": "^3.3.0"
   }
 }

package/scripts/harness-next.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 # harness-next.sh — 세션 오케스트레이터
 # 현재 progress.json 상태를 읽고 다음 에이전트를 결정한다.
-# Feature-level 프로그래스를 출력하고 next-prompt.txt를 생성한다.
+# Feature-level 프로그래스를 출력하고 handoff.json을 생성한다.
 set -e
 
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
@@ -20,7 +20,7 @@ PROJECT_ROOT="$(resolve_harness_root "${1:-.}")" || {
 PROGRESS="$PROJECT_ROOT/.harness/progress.json"
 CONFIG="$PROJECT_ROOT/.harness/config.json"
 PIPELINE_JSON="$PROJECT_ROOT/.harness/actions/pipeline.json"
-NEXT_PROMPT="$PROJECT_ROOT/.harness/next-prompt.txt"
+HANDOFF="$PROJECT_ROOT/.harness/handoff.json"
 
 check_jq || exit 1
 
@@ -296,7 +296,7 @@ render_agent_bar "$PROJECT_ROOT"
 echo ""
 
 # ─────────────────────────────────────────
-# Generate next-prompt.txt
+# Generate handoff.json (unified session transition document)
 # ─────────────────────────────────────────
 if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_status" != "blocked" ]; then
   # ── Escalation check: 3회 실패 시 Planner에게 scope 축소 요청 ──
@@ -304,14 +304,17 @@ if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_st
   if [ "$retry_count" -ge "$escalate_after" ] && [ "$sprint_status" = "failed" ] && [ "$next_agent" != "planner" ]; then
     echo "  ⚠ Escalation: ${retry_count}회 실패 — Planner에게 scope 축소/접근 변경 요청"
     next_agent="planner"
-    # progress.json에 에스컬레이션 기록
     jq --arg msg "Escalated after ${retry_count} failures. Planner must review scope or approach." \
        '.next_agent = "planner" |
         .failure.message = $msg |
         .failure.retry_target = "planner"' "$PROGRESS" > "${PROGRESS}.tmp" && mv "${PROGRESS}.tmp" "$PROGRESS"
   fi
 
-  # Build prompt
+  # ── Read model & thinking mode for next agent ──
+  agent_model=$(jq -r ".agents[\"${next_agent}\"].model // \"opus\"" "$CONFIG" 2>/dev/null || echo "opus")
+  agent_thinking=$(jq -r ".agents[\"${next_agent}\"].thinking_mode // \"null\"" "$CONFIG" 2>/dev/null || echo "null")
+
+  # ── Build prompt text (embedded in handoff.json) ──
   prompt="/harness-${next_agent} 를 실행하세요."
 
   if [ "$sprint_num" -gt 0 ]; then
@@ -322,22 +325,33 @@ if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_st
     prompt+="을 진행합니다."
   fi
 
-  prompt+=$'\n'".harness/progress.json을 읽고 현재 상태를 확인하세요."
+  prompt+=$'\n'".harness/handoff.json을 읽고 컨텍스트를 확인하세요."
+
+  # Inject thinking mode instruction
+  if [ "$agent_thinking" != "null" ]; then
+    case "$agent_thinking" in
+      ultraplan)
+        prompt+=$'\n\n'"[Thinking Mode: ultraplan] /${agent_thinking} 모드를 사용하세요. 깊은 사고로 아키텍처와 설계를 수행합니다."
+        ;;
+      ultrathink)
+        prompt+=$'\n\n'"[Thinking Mode: ultrathink] /${agent_thinking} 모드를 사용하세요. 최대 추론 깊이로 비판적 검증을 수행합니다."
+        ;;
+      plan)
+        prompt+=$'\n\n'"[Thinking Mode: plan] /${agent_thinking} 모드를 사용하세요. 구조화된 계획을 수립한 후 실행합니다."
+        ;;
+    esac
+  fi
 
-  # Add failure context if retrying (include previous failure summary)
+  # Add failure context if retrying
   failure_msg=$(jq -r '.failure.message // empty' "$PROGRESS")
   if [ -n "$failure_msg" ] && [ "$failure_msg" != "null" ]; then
     prompt+=$'\n\n'"이전 실패 사유: ${failure_msg}"
     prompt+=$'\n'"같은 접근을 반복하지 말고, 실패 원인을 분석한 후 다른 전략으로 시도하세요."
   fi
 
-  echo "$prompt" > "$NEXT_PROMPT"
-
-  # ── Generate structured handoff.json ──
-  HANDOFF="$PROJECT_ROOT/.harness/handoff.json"
+  # ── Collect artifacts ──
   FEATURE_LIST="$PROJECT_ROOT/.harness/actions/feature-list.json"
 
-  # Collect available artifacts
   local -a artifacts_ready=()
   for f in plan.md feature-list.json api-contract.json sprint-contract.md evaluation-functional.md evaluation-visual.md; do
     if [ -f "$PROJECT_ROOT/.harness/actions/$f" ]; then
@@ -353,7 +367,7 @@ if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_st
     focus_features=$(jq '[.features[]? | select(.passes == null or (.passes | length) == 0 or ((.passes // []) | map(select(. == "evaluator-functional")) | length == 0)) | .id] | .[0:5]' "$FEATURE_LIST" 2>/dev/null || echo "[]")
   fi
 
-  # ── Regression data: collect previous sprint's passed AC from archive ──
+  # ── Regression data ──
   local regression_source="null"
   local prev_sprint=$((sprint_num - 1))
   local prev_archive="$PROJECT_ROOT/.harness/archive/sprint-$(printf '%03d' $prev_sprint)"
@@ -367,7 +381,7 @@ if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_st
     fi
   fi
 
-  # ── Eval-specific scoring config ──
+  # ── Eval-specific config ──
   local eval_config="null"
   local cross_validation_data="null"
   case "$next_agent" in
@@ -384,22 +398,24 @@ if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_st
       ;;
   esac
 
-  # ── Cross-Validation: evaluator-visual이면 functional 결과 파싱 ──
+  # ── Cross-Validation ──
   if [ "$next_agent" = "evaluator-visual" ]; then
     local func_eval="$PROJECT_ROOT/.harness/actions/evaluation-functional.md"
     if [ -f "$func_eval" ]; then
-      # evaluation-functional.md 내 JSON 코드블록에서 Cross-Validation Data 추출
       cross_validation_data=$(sed -n '/```json/,/```/p' "$func_eval" | tail -n +2 | head -n -1 | jq 'select(.evaluator == "functional")' 2>/dev/null || echo "null")
     fi
   fi
 
-  # Build handoff.json
+  # ── Build handoff.json (single source of truth for session transition) ──
   jq -n \
     --arg from "${current_agent:-dispatcher}" \
     --arg to "$next_agent" \
+    --arg prompt "$prompt" \
     --argjson sprint "$sprint_num" \
     --argjson retry "$retry_count" \
     --arg status "$sprint_status" \
+    --arg agent_model "$agent_model" \
+    --arg agent_thinking "$agent_thinking" \
     --arg failure_msg "${failure_msg:-}" \
     --argjson artifacts "$artifacts_json" \
     --argjson focus "$focus_features" \
@@ -410,9 +426,12 @@ if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_st
     '{
       from: $from,
       to: $to,
+      prompt: $prompt,
       sprint: $sprint,
       retry_count: $retry,
       sprint_status: $status,
+      model: $agent_model,
+      thinking_mode: (if $agent_thinking == "null" then null else $agent_thinking end),
       failure_context: (if $failure_msg != "" then $failure_msg else null end),
       artifacts_ready: $artifacts,
       focus_features: $focus,
@@ -424,12 +443,20 @@ if [ "$next_agent" != "null" ] && [ "$next_agent" != "archive" ] && [ "$agent_st
     }' > "$HANDOFF"
 
 elif [ "$next_agent" = "archive" ]; then
-  cat > "$NEXT_PROMPT" <<'PROMPT'
-Sprint 문서를 아카이브 하세요.
-.harness/actions/의 스프린트 문서를 .harness/archive/sprint-NNN/으로 이동합니다.
-.harness/progress.json을 읽고 sprint 번호를 확인하세요.
-PROMPT
+  jq -n \
+    --arg from "${current_agent:-evaluator}" \
+    --argjson sprint "$sprint_num" \
+    --arg timestamp "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
+    '{
+      from: $from,
+      to: "archive",
+      prompt: "Sprint 문서를 아카이브하세요.\n.harness/actions/의 스프린트 문서를 .harness/archive/sprint-NNN/으로 이동합니다.\n.harness/handoff.json을 읽고 sprint 번호를 확인하세요.",
+      sprint: $sprint,
+      model: "opus",
+      thinking_mode: null,
+      timestamp: $timestamp
+    }' > "$HANDOFF"
 
 else
-  echo "" > "$NEXT_PROMPT"
+  echo '{}' > "$HANDOFF"
 fi

package/scripts/harness-statusline.sh

@@ -88,6 +88,28 @@ if [ "$retry_count" -gt 0 ]; then
   retry_str=" R${retry_count}"
 fi
 
+# Model/mode for current or next agent
+CONFIG="$PROJECT_ROOT/.harness/config.json"
+active_agent="$current_agent"
+if [ "$active_agent" = "none" ] || [ "$active_agent" = "null" ]; then
+  active_agent=$(jq -r '.next_agent // "none"' "$PROGRESS" 2>/dev/null)
+fi
+model_short=""
+if [ -f "$CONFIG" ] && [ "$active_agent" != "none" ] && [ "$active_agent" != "null" ]; then
+  am=$(jq -r ".agents[\"${active_agent}\"].model // empty" "$CONFIG" 2>/dev/null)
+  at=$(jq -r ".agents[\"${active_agent}\"].thinking_mode // empty" "$CONFIG" 2>/dev/null)
+  if [ -n "$am" ]; then
+    model_short="${am}"
+    if [ -n "$at" ] && [ "$at" != "null" ]; then
+      model_short+="/${at}"
+    fi
+  fi
+fi
+
 # Build compact status line
-# Format: [S1] FULL | >backend | 2/5 feat | R0 | ctx 45% | $1.23
-echo "[S${sprint_num}] ${pl} | ${status_icon}${agent_short}${retry_str} | ${completed_features}/${total_features} feat | ctx ${context_pct}% | \$${cost}"
+# Format: [S1] FULL | >backend | sonnet | 2/5 feat | ctx 45% | $1.23
+if [ -n "$model_short" ]; then
+  echo "[S${sprint_num}] ${pl} | ${status_icon}${agent_short}${retry_str} | ${model_short} | ${completed_features}/${total_features} feat | ctx ${context_pct}% | \$${cost}"
+else
+  echo "[S${sprint_num}] ${pl} | ${status_icon}${agent_short}${retry_str} | ${completed_features}/${total_features} feat | ctx ${context_pct}% | \$${cost}"
+fi

package/scripts/lib/harness-render-progress.sh

@@ -268,9 +268,28 @@ render_progress() {
 
   # ── Next Action ──
   if [ "$next_agent" != "none" ] && [ "$next_agent" != "null" ] && [ "$agent_status" != "blocked" ]; then
+    # Read model & thinking mode from config
+    local next_model="opus"
+    local next_thinking="null"
+    if [ -f "$CONFIG" ]; then
+      next_model=$(jq -r ".agents[\"${next_agent}\"].model // \"opus\"" "$CONFIG" 2>/dev/null || echo "opus")
+      next_thinking=$(jq -r ".agents[\"${next_agent}\"].thinking_mode // \"null\"" "$CONFIG" 2>/dev/null || echo "null")
+    fi
+
+    local mode_str=""
+    if [ "$next_thinking" != "null" ]; then
+      mode_str=" [/${next_thinking}]"
+    fi
+
     echo ""
-    echo "  Next → /harness-${next_agent}"
-    echo "  Auto → claude --prompt \"\$(cat .harness/next-prompt.txt)\""
+    echo "  Next → /harness-${next_agent}  (model: ${next_model}${mode_str})"
+
+    # Build auto CLI command with model flag, reading prompt from handoff.json
+    local model_flag=""
+    if [ "$next_model" != "opus" ]; then
+      model_flag=" --model ${next_model}"
+    fi
+    echo "  Auto → claude${model_flag} --prompt \"\$(jq -r .prompt .harness/handoff.json)\""
   fi
 
   echo ""