@walwal-harness/cli 3.7.4 → 4.0.0-alpha.10

package/scripts/harness-team-worker.sh

@@ -0,0 +1,415 @@
+#!/bin/bash
+# harness-team-worker.sh — Team Worker: Feature-level Gen→Eval loop (v4.0)
+#
+# 1 Team = 1 프로세스. Feature Queue에서 feature를 꺼내
+# Gen→Gate→Eval 루프를 claude -p 헤드리스로 자율 실행한다.
+#
+# Usage:
+#   bash scripts/harness-team-worker.sh <team_id> [project-root]
+#
+# Environment:
+#   MAX_ATTEMPTS=3  Feature당 최대 Gen→Eval 시도 횟수
+#   GEN_MODEL=sonnet  Generator 모델
+#   EVAL_MODEL=opus   Evaluator 모델
+
+set -uo pipefail
+
+TEAM_ID="${1:?Usage: harness-team-worker.sh <team_id> [project-root]}"
+shift || true
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+
+# ── Resolve project root ──
+PROJECT_ROOT="${1:-}"
+if [ -z "$PROJECT_ROOT" ]; then
+  dir="$(pwd)"
+  while [ "$dir" != "/" ]; do
+    if [ -d "$dir/.harness" ]; then PROJECT_ROOT="$dir"; break; fi
+    dir="$(dirname "$dir")"
+  done
+fi
+
+if [ -z "$PROJECT_ROOT" ] || [ ! -d "$PROJECT_ROOT/.harness" ]; then
+  echo "[T${TEAM_ID}] .harness/ not found."
+  exit 1
+fi
+
+QUEUE="$PROJECT_ROOT/.harness/actions/feature-queue.json"
+FEATURES="$PROJECT_ROOT/.harness/actions/feature-list.json"
+CONFIG="$PROJECT_ROOT/.harness/config.json"
+PROGRESS_LOG="$PROJECT_ROOT/.harness/progress.log"
+QUEUE_MGR="$SCRIPT_DIR/harness-queue-manager.sh"
+
+# ── Lock file for git operations (prevent race conditions between teams) ──
+GIT_LOCK="$PROJECT_ROOT/.harness/.git-lock"
+
+MAX_ATTEMPTS="${MAX_ATTEMPTS:-3}"
+GEN_MODEL="${GEN_MODEL:-sonnet}"
+EVAL_MODEL="${EVAL_MODEL:-opus}"
+
+if [ -f "$CONFIG" ]; then
+  _gm=$(jq -r '.agents["generator-frontend"].model // empty' "$CONFIG" 2>/dev/null)
+  _em=$(jq -r '.agents["evaluator-functional"].model // empty' "$CONFIG" 2>/dev/null)
+  if [ -n "$_gm" ]; then GEN_MODEL="$_gm"; fi
+  if [ -n "$_em" ]; then EVAL_MODEL="$_em"; fi
+fi
+
+# ── ANSI helpers ──
+BOLD="\033[1m"
+DIM="\033[2m"
+GREEN="\033[32m"
+YELLOW="\033[33m"
+RED="\033[31m"
+CYAN="\033[36m"
+RESET="\033[0m"
+
+ts() { date +"%H:%M:%S"; }
+
+log() {
+  echo -e "[$(ts)] ${BOLD}T${TEAM_ID}${RESET} $*"
+}
+
+log_progress() {
+  echo "$(date +"%Y-%m-%d") | team-${TEAM_ID} | ${1} | ${2}" >> "$PROGRESS_LOG"
+}
+
+# ── Git lock — serialize git checkout/merge across teams ──
+acquire_git_lock() {
+  local max_wait=60 waited=0
+  while [ -f "$GIT_LOCK" ]; do
+    sleep 1
+    waited=$((waited + 1))
+    if [ "$waited" -ge "$max_wait" ]; then
+      log "${RED}Git lock timeout (${max_wait}s). Removing stale lock.${RESET}"
+      rm -f "$GIT_LOCK"
+      break
+    fi
+  done
+  echo "T${TEAM_ID}" > "$GIT_LOCK"
+}
+
+release_git_lock() {
+  rm -f "$GIT_LOCK"
+}
+
+# ── Pre-eval gate ──
+run_pre_eval_gate() {
+  local cwd="$PROJECT_ROOT"
+
+  if [ -f "$CONFIG" ]; then
+    _cwd=$(jq -r '.flow.pre_eval_gate.frontend_cwd // empty' "$CONFIG" 2>/dev/null)
+    if [ -n "$_cwd" ] && [ "$_cwd" != "null" ]; then
+      cwd="$PROJECT_ROOT/$_cwd"
+    fi
+  fi
+
+  local checks=()
+  if [ -f "$CONFIG" ]; then
+    mapfile -t checks < <(jq -r '.flow.pre_eval_gate.frontend_checks[]' "$CONFIG" 2>/dev/null)
+  fi
+
+  if [ ${#checks[@]} -eq 0 ]; then
+    checks=("npx tsc --noEmit" "npx eslint src/")
+  fi
+
+  local all_pass=true fail_cmds=""
+  for cmd in "${checks[@]}"; do
+    if (cd "$cwd" && timeout 120s bash -c "$cmd" >/dev/null 2>&1); then
+      log "  ${GREEN}✓${RESET} $cmd"
+    else
+      log "  ${RED}✗${RESET} $cmd"
+      all_pass=false
+      fail_cmds+="$cmd; "
+    fi
+  done
+
+  [ "$all_pass" = true ]
+}
+
+# ── Build generator prompt ──
+build_gen_prompt() {
+  local fid="$1" attempt="$2" feedback="${3:-}"
+
+  local fobj
+  fobj=$(jq --arg fid "$fid" '.features[] | select(.id == $fid)' "$FEATURES" 2>/dev/null)
+  local fname fdesc ac_json deps_json
+  fname=$(echo "$fobj" | jq -r '.name // .description // ""')
+  fdesc=$(echo "$fobj" | jq -r '.description // ""')
+  ac_json=$(echo "$fobj" | jq -c '.ac // []')
+  deps_json=$(echo "$fobj" | jq -c '.depends_on // []')
+
+  local project_name
+  project_name=$(jq -r '.project_name // ""' "$PROJECT_ROOT/.harness/progress.json" 2>/dev/null)
+
+  cat <<PROMPT
+You are Generator-Frontend for a harness engineering project.
+
+PROJECT: ${project_name}
+CONVENTIONS: Read CONVENTIONS.md if it exists.
+
+YOUR TASK: Implement ONLY feature ${fid}: ${fname}
+Description: ${fdesc}
+Dependencies (already implemented): ${deps_json}
+Acceptance Criteria: ${ac_json}
+
+Read these files for context:
+- .harness/actions/feature-list.json (filter to ${fid})
+- .harness/actions/api-contract.json (relevant endpoints)
+- .harness/actions/plan.md (overall design)
+
+RULES:
+- Implement ONLY this single feature
+- Do NOT modify code belonging to other features
+- Follow existing code patterns and CONVENTIONS.md
+- When done, stage and commit with: git add -A && git commit -m 'feat(${fid}): ${fname}'
+PROMPT
+
+  if [ "$attempt" -gt 1 ] && [ -n "$feedback" ]; then
+    cat <<RETRY
+
+PREVIOUS EVAL FEEDBACK (attempt ${attempt}):
+${feedback}
+
+Fix the issues above. Focus specifically on the failed criteria.
+RETRY
+  fi
+}
+
+# ── Build evaluator prompt ──
+build_eval_prompt() {
+  local fid="$1"
+
+  local fobj
+  fobj=$(jq --arg fid "$fid" '.features[] | select(.id == $fid)' "$FEATURES" 2>/dev/null)
+  local fname ac_json
+  fname=$(echo "$fobj" | jq -r '.name // .description // ""')
+  ac_json=$(echo "$fobj" | jq -c '.ac // []')
+
+  local passed_list
+  passed_list=$(jq -r '.queue.passed // [] | join(", ")' "$QUEUE" 2>/dev/null)
+
+  cat <<PROMPT
+You are Evaluator-Functional for a harness engineering project.
+
+TASK: Evaluate feature ${fid}: ${fname}
+
+Acceptance Criteria to verify:
+${ac_json}
+
+Previously passed features (regression check): [${passed_list}]
+
+SCORING RUBRIC (R1-R5):
+R1: API Contract compliance (25%)
+R2: Acceptance Criteria full pass (25%)
+R3: Negative tests (20%)
+R4: E2E scenario (15%)
+R5: Error handling & edge cases (15%)
+
+PASS threshold: 2.80 / 3.00
+FAIL: any AC not met, any regression failure
+
+You MUST output this exact block (parseable by automation):
+---EVAL-RESULT---
+FEATURE: ${fid}
+VERDICT: PASS or FAIL
+SCORE: X.XX
+FEEDBACK: one paragraph summary
+---END-EVAL-RESULT---
+PROMPT
+}
+
+# ── Parse eval result (macOS-compatible, no grep -P) ──
+parse_eval_result() {
+  local output="$1"
+
+  local verdict score feedback
+  verdict=$(echo "$output" | grep -E '^VERDICT:' | sed 's/VERDICT:[[:space:]]*//' | head -1)
+  score=$(echo "$output" | grep -E '^SCORE:' | sed 's/SCORE:[[:space:]]*//' | head -1)
+  feedback=$(echo "$output" | grep -E '^FEEDBACK:' | sed 's/FEEDBACK:[[:space:]]*//' | head -1)
+
+  echo "${verdict:-UNKNOWN}|${score:-0.00}|${feedback:-no feedback}"
+}
+
+# ══════════════════════════════════════════
+# Main Worker Loop
+# ══════════════════════════════════════════
+log "${CYAN}Team ${TEAM_ID} started${RESET} (gen=${GEN_MODEL}, eval=${EVAL_MODEL}, max=${MAX_ATTEMPTS})"
+log_progress "start" "Team ${TEAM_ID} worker started"
+
+while true; do
+  # ── Dequeue next feature ──
+  feature_id=$(bash "$QUEUE_MGR" dequeue "$TEAM_ID" "$PROJECT_ROOT" 2>/dev/null)
+
+  if [ -z "$feature_id" ] || [[ "$feature_id" == "["* ]]; then
+    log "${DIM}No features ready. Waiting 10s...${RESET}"
+    sleep 10
+
+    # Check if completely done
+    remaining=$(jq '(.queue.ready | length) + (.queue.blocked | length) + (.queue.in_progress | length)' "$QUEUE" 2>/dev/null || echo "1")
+    if [ "${remaining}" -eq 0 ] 2>/dev/null; then
+      log "${GREEN}${BOLD}ALL FEATURES COMPLETE. Team ${TEAM_ID} exiting.${RESET}"
+      log_progress "complete" "All features done"
+      exit 0
+    fi
+    continue
+  fi
+
+  log "${CYAN}▶ Dequeued ${feature_id}${RESET}"
+  log_progress "dequeue" "${feature_id}"
+
+  # ── Create feature branch (with lock) ──
+  branch="feature/${feature_id}"
+  acquire_git_lock
+  (cd "$PROJECT_ROOT" && git checkout main 2>/dev/null && git checkout -b "$branch" 2>/dev/null) || \
+  (cd "$PROJECT_ROOT" && git checkout "$branch" 2>/dev/null) || true
+  release_git_lock
+  log "Branch: ${branch}"
+
+  # ── Gen→Eval Loop ──
+  attempt=1
+  eval_feedback=""
+  passed=false
+
+  while [ "$attempt" -le "$MAX_ATTEMPTS" ]; do
+    log "${BOLD}── Attempt ${attempt}/${MAX_ATTEMPTS} ──${RESET}"
+
+    # ── Generate ──
+    log "Gen ${feature_id} (${GEN_MODEL})..."
+    bash "$QUEUE_MGR" update_phase "$feature_id" "gen" "$attempt" "$PROJECT_ROOT" 2>/dev/null
+
+    gen_prompt=$(build_gen_prompt "$feature_id" "$attempt" "$eval_feedback")
+
+    gen_start=$(date +%s)
+    log "${DIM}  claude -p --dangerously-skip-permissions --model ${GEN_MODEL}${RESET}"
+    gen_output=$(cd "$PROJECT_ROOT" && claude -p "$gen_prompt" \
+      --dangerously-skip-permissions \
+      --model "$GEN_MODEL" \
+      --output-format text 2>&1 | tee /dev/stderr) 2>&1 || true
+    gen_elapsed=$(( $(date +%s) - gen_start ))
+
+    files_changed=$(cd "$PROJECT_ROOT" && git diff --name-only 2>/dev/null | wc -l | tr -d ' ')
+    log "Gen done (${gen_elapsed}s) — ${files_changed} files"
+    log_progress "gen" "${feature_id} attempt ${attempt}: ${files_changed} files, ${gen_elapsed}s"
+
+    # Auto-commit
+    (cd "$PROJECT_ROOT" && git add -A && git commit -m "feat(${feature_id}): gen attempt ${attempt}" --no-verify 2>/dev/null) || true
+
+    # ── Pre-eval gate ──
+    log "Pre-eval gate..."
+    bash "$QUEUE_MGR" update_phase "$feature_id" "gate" "$attempt" "$PROJECT_ROOT" 2>/dev/null
+
+    if ! run_pre_eval_gate "$feature_id"; then
+      log "${RED}Gate FAIL — retrying gen${RESET}"
+      eval_feedback="Pre-eval gate failed: type check or lint errors. Fix compilation and lint issues."
+      attempt=$((attempt + 1))
+      continue
+    fi
+
+    # ── Evaluate ──
+    log "Eval ${feature_id} (${EVAL_MODEL})..."
+    bash "$QUEUE_MGR" update_phase "$feature_id" "eval" "$attempt" "$PROJECT_ROOT" 2>/dev/null
+
+    eval_prompt=$(build_eval_prompt "$feature_id")
+
+    eval_start=$(date +%s)
+    log "${DIM}  claude -p --dangerously-skip-permissions --model ${EVAL_MODEL}${RESET}"
+    eval_output=$(cd "$PROJECT_ROOT" && claude -p "$eval_prompt" \
+      --dangerously-skip-permissions \
+      --model "$EVAL_MODEL" \
+      --output-format text 2>&1 | tee /dev/stderr) 2>&1 || true
+    eval_elapsed=$(( $(date +%s) - eval_start ))
+
+    # Parse result
+    result_line=$(parse_eval_result "$eval_output")
+    verdict=$(echo "$result_line" | cut -d'|' -f1)
+    score=$(echo "$result_line" | cut -d'|' -f2)
+    feedback=$(echo "$result_line" | cut -d'|' -f3-)
+
+    log_progress "eval" "${feature_id} attempt ${attempt}: ${verdict} (${score}) ${eval_elapsed}s"
+
+    if [ "$verdict" = "PASS" ]; then
+      log "${GREEN}${BOLD}✓ PASS${RESET} ${feature_id} — ${score}/3.00 (${eval_elapsed}s)"
+      passed=true
+      break
+    else
+      log "${RED}✗ FAIL${RESET} ${feature_id} — ${score}/3.00 (${eval_elapsed}s)"
+      log "${DIM}  ${feedback}${RESET}"
+      eval_feedback="$feedback"
+      attempt=$((attempt + 1))
+    fi
+  done
+
+  # ══════════════════════════════════════════
+  # Phase 3: Branch merge with conflict handling
+  # ══════════════════════════════════════════
+  if [ "$passed" = true ]; then
+    log "Merging ${branch} → main..."
+    acquire_git_lock
+
+    merge_ok=false
+
+    # Attempt 1: straight merge
+    if (cd "$PROJECT_ROOT" && git checkout main 2>/dev/null && git merge --no-ff "$branch" -m "merge: ${feature_id} PASS" 2>/dev/null); then
+      merge_ok=true
+    else
+      # Attempt 2: abort failed merge, rebase, re-eval gate, then merge
+      log "${YELLOW}Conflict detected — rebasing ${branch} onto main...${RESET}"
+      (cd "$PROJECT_ROOT" && git merge --abort 2>/dev/null) || true
+      (cd "$PROJECT_ROOT" && git checkout "$branch" 2>/dev/null) || true
+
+      if (cd "$PROJECT_ROOT" && git rebase main 2>/dev/null); then
+        log "Rebase OK. Re-running gate..."
+
+        if run_pre_eval_gate "$feature_id"; then
+          log "Gate still PASS after rebase."
+          if (cd "$PROJECT_ROOT" && git checkout main 2>/dev/null && git merge --no-ff "$branch" -m "merge: ${feature_id} PASS (rebased)" 2>/dev/null); then
+            merge_ok=true
+          fi
+        else
+          log "${RED}Gate FAIL after rebase — needs re-gen${RESET}"
+        fi
+      else
+        log "${RED}Rebase failed — conflicts too complex${RESET}"
+        (cd "$PROJECT_ROOT" && git rebase --abort 2>/dev/null) || true
+      fi
+    fi
+
+    release_git_lock
+
+    if [ "$merge_ok" = true ]; then
+      # Clean up feature branch
+      (cd "$PROJECT_ROOT" && git branch -d "$branch" 2>/dev/null) || true
+
+      bash "$QUEUE_MGR" pass "$feature_id" "$PROJECT_ROOT" 2>/dev/null
+      log_progress "pass" "${feature_id} merged to main"
+
+      # Update feature-list.json passes
+      if [ -f "$FEATURES" ]; then
+        jq --arg fid "$feature_id" '
+          .features |= map(
+            if .id == $fid then
+              .passes = ((.passes // []) + ["generator-frontend", "evaluator-functional"] | unique)
+            else . end
+          )
+        ' "$FEATURES" > "${FEATURES}.tmp" && mv "${FEATURES}.tmp" "$FEATURES"
+      fi
+
+      log "${GREEN}${BOLD}✓ ${feature_id} DONE${RESET}"
+    else
+      log "${RED}${BOLD}Merge failed — ${feature_id} marked as failed${RESET}"
+      (cd "$PROJECT_ROOT" && git checkout main 2>/dev/null) || true
+      bash "$QUEUE_MGR" fail "$feature_id" "$PROJECT_ROOT" 2>/dev/null
+      log_progress "merge-fail" "${feature_id}"
+    fi
+
+  else
+    log "${RED}${BOLD}✗ ${feature_id} FAILED after ${MAX_ATTEMPTS} attempts${RESET}"
+    acquire_git_lock
+    (cd "$PROJECT_ROOT" && git checkout main 2>/dev/null) || true
+    release_git_lock
+    bash "$QUEUE_MGR" fail "$feature_id" "$PROJECT_ROOT" 2>/dev/null
+    log_progress "fail" "${feature_id} after ${MAX_ATTEMPTS} attempts"
+  fi
+
+  sleep 2
+done

package/scripts/harness-user-prompt-submit.sh

@@ -51,7 +51,37 @@ current_agent=${CURRENT_AGENT} (running) 인데 /harness-${REQUESTED_SKILL} 호
   fi
 fi
 
-# ── Compact context 주입 ──
+# ── v4 Parallel Mode 감지 ──
+FEATURE_QUEUE="$CWD/.harness/actions/feature-queue.json"
+if [ -f "$FEATURE_QUEUE" ]; then
+  V4_PASSED=$(jq '.queue.passed | length' "$FEATURE_QUEUE" 2>/dev/null || echo 0)
+  V4_TOTAL=$(jq '[.queue.ready, (.queue.blocked | keys), (.queue.in_progress | keys), .queue.passed, .queue.failed] | flatten | length' "$FEATURE_QUEUE" 2>/dev/null || echo 0)
+  V4_FAILED=$(jq '.queue.failed | length' "$FEATURE_QUEUE" 2>/dev/null || echo 0)
+
+  # Log user prompt to progress.log (truncated to 80 chars)
+  PROGRESS_LOG="$CWD/.harness/progress.log"
+  if [ -n "$PROMPT" ] && [ -f "$PROGRESS_LOG" ]; then
+    PROMPT_SHORT=$(echo "$PROMPT" | tr '\n' ' ' | sed 's/  */ /g' | cut -c1-80)
+    # Skip logging for empty or very short prompts
+    if [ ${#PROMPT_SHORT} -gt 2 ]; then
+      echo "$(date +"%Y-%m-%d %H:%M") | user-prompt | input | ${PROMPT_SHORT}" >> "$PROGRESS_LOG"
+    fi
+  fi
+
+  cat <<EOF
+[harness-v4] ${V4_PASSED}/${V4_TOTAL} features passed | ${V4_FAILED} failed
+
+## v4 Parallel Mode Active
+- 3 Agent Teams이 feature-queue에서 자율적으로 Gen→Eval 루프 실행 중
+- 당신은 **오케스트레이터** 역할: 대시보드 모니터링, 실패 대응, 수동 개입
+- /harness-generator-*, /harness-evaluator-* 스킬 호출 금지 (Teams가 처리)
+- 할 수 있는 것: 코드 리뷰, 아키텍처 결정, failed feature 분석, requeue 판단
+- skip: "harness skip" 시 일반 대화
+EOF
+  exit 0
+fi
+
+# ── Compact context 주입 (v3 mode) ──
 cat <<EOF
 [harness] S${SPRINT_NUM} | ${PIPELINE} | agent=${CURRENT_AGENT} (${AGENT_STATUS}) | next=${NEXT_AGENT}
 ${CONTEXT_WARNING}

package/skills/evaluator-functional/SKILL.md

@@ -60,6 +60,32 @@ disable-model-invocation: true
 5. `actions/api-contract.json` — 기대 API 동작
 6. `.harness/progress.json`
 
+## v4 Feature-Level Mode (Parallel Agent Teams)
+
+v4에서 Team Worker가 `claude -p`로 호출할 때, 프롬프트에 `FEATURE_ID`가 지정된다.
+
+### Feature-Level Rules
+- `feature-list.json`에서 **지정된 FEATURE_ID의 AC만** 검증
+- Regression: `feature-queue.json`의 `passed` 목록에 있는 Feature들의 AC 재검증
+- Cross-Validation: Feature 단위에서는 skip (Sprint-End에서 수행)
+- Visual Evaluation: Feature 단위에서는 skip (Sprint-End에서 수행)
+- 출력 형식: `---EVAL-RESULT---` 블록 (Worker가 파싱 가능)
+
+### Feature-Level Scoring
+- 동일한 R1-R5 루브릭 적용
+- PASS 기준: 2.80/3.00 (변경 없음)
+- 1건이라도 Regression 실패 시 FAIL (변경 없음)
+
+### Output Format (Machine-Parseable)
+```
+---EVAL-RESULT---
+FEATURE: F-XXX
+VERDICT: PASS or FAIL
+SCORE: X.XX
+FEEDBACK: one paragraph summary
+---END-EVAL-RESULT---
+```
+
 ## Evaluation Steps
 
 ### Step 0: IA Structure Compliance (GATE)

package/skills/generator-frontend/SKILL.md

@@ -43,6 +43,23 @@ disable-model-invocation: true
 
 **Backend 통합 러너가 동작 중이어야 함.** Gateway 미응답 시 → STOP.
 
+## v4 Feature-Level Mode (Parallel Agent Teams)
+
+v4에서 Team Worker가 `claude -p`로 호출할 때, 프롬프트에 `FEATURE_ID`가 지정된다.
+
+### Feature-Level Rules
+- `feature-list.json`에서 **지정된 FEATURE_ID만** 필터하여 구현
+- 다른 Feature의 코드를 수정하지 않음
+- `depends_on`에 명시된 Feature는 이미 구현/머지 완료된 상태
+- Feature branch (`feature/F-XXX`)에서 작업, 완료 시 commit
+- Sprint Contract는 작성하지 않음 (v4에서는 Feature 단위로 관리)
+
+### Feature-Level Prompt Template
+Worker가 전달하는 프롬프트에는 다음이 포함됨:
+- `FEATURE_ID`, `feature_name`, `description`, `ac` (Acceptance Criteria)
+- `depends_on` (이미 완료된 의존 Feature 목록)
+- Eval 재시도 시: 이전 Eval의 피드백
+
 ## Sprint Workflow
 
 1. **Sprint Contract FE 섹션 추가** — 컴포넌트, API 연동, 성공 기준