qa-loop.yaml

# LoopEngine 研发闭环提效流水线配置
# 平台级通用配置 - 可按项目覆盖(project_overrides)
#
# 复用 skillforge: loop/gates/escalation/agents 段保持与原 qa-loop.yaml 兼容;
# 新增: scope / confidence / checklist / agent_models 段

# ========== 循环控制 ==========
loop:
  max_iterations: 5              # 最大循环次数 (防失控)
  converge_threshold: 0          # 收敛阈值: 每轮 issue 减少多少才算收敛 (0=只要有减少)
  escalate_on_new_critical: 2    # 连续N轮出现新Critical → 升级人工 (经 AskUserQuestion)

# ========== 6 阶段调度 ==========
phases:
  - scope                 # [0] 修改范围分析 + 方案草案 (仅首轮)
  - review                # [1] 多视角审查 → Finding[]
  - fix                   # [2] 修复 open findings, 更新 solution
  - test-data             # [3] 准备测试数据 (fixtures/seeds/cassette/anonymize)
  - test                  # [4] 跑测试 + 补缺失, 输出 TestStatus
  - confidence            # [5] 每条修复打置信度 + checklist 核对
  - gate                  # [6] 门禁评估 → pass / continue / escalate

# ========== TestData 阶段配置 (测试数据来源管理) ==========
# 把"测试数据从哪来"显式化, 而非硬编码在测试里。
# 策略: synthetic(合成) / fixture(固定样本) / sample(采样真实) /
#       anonymize(真实脱敏) / replay(录制回放)
test_data:
  default_strategy: synthetic       # 默认合成 (隐私安全、可复现)
  data_sources:
    db:
      strategy: anonymize           # DB 用脱敏真实数据
      dsn: "${TEST_DB_DSN}"
      anonymize: true
    api:
      strategy: replay              # 外部 API 录制回放 (确定性, 无需在线)
      cassette_dir: "tests/cassettes"
    fixtures:
      strategy: synthetic           # 单元 fixture 合成
      cover_boundary: true          # 覆盖边界值 (对接 edge-case-hunting)
    mq:                             # 消息队列 (Kafka/RabbitMQ/RocketMQ)
      broker: kafka                 # kafka / rabbitmq / rocketmq
      strategy: real                # real(testcontainers真broker) / replay(录制回放) / embedded(内嵌)
      # 铁律: 优先 real; mock/embedded 会漏序列化+partition+消费者组重平衡问题
      bootstrap: "${TEST_KAFKA}"    # real=testcontainers地址; 留空=testcontainers自动起
      topics:                       # 测试用 topic (生产→消费)
        - { name: "order-events", role: "produce" }
        - { name: "payment-results", role: "consume" }
        - { name: "order-events.test-sink", role: "test_sink" }  # 断言专用sink
      schema:
        format: avro               # avro / protobuf / json-schema — 契约测试用
        registry: "${SCHEMA_REGISTRY}"
      consumers:
        group_isolation: per_test  # 每测试独立 group-id (防 offset 串扰)
        offset_reset: earliest      # 测试从头消费确定状态
      assertions:
        async: true                 # 异步断言 (轮询副作用, 不 sleep)
        timeout_ms: 10000           # 等待消息副作用的最长时间
  anonymization:                    # PII 脱敏规则 (DB + MQ 消息体)
    - { field: "email", rule: "hash" }
    - { field: "phone", rule: "mask" }
    - { field: "id_card", rule: "mask" }
    - { field: "payload.user.mobile", rule: "mask", scope: "mq" }   # MQ 消息体嵌套字段
  lifecycle:                         # 数据生命周期
    per_test_isolation: true         # 每个测试独立数据, 用后清理 (无共享可变状态)
    mq_cleanup: purge_after_test    # MQ: 测试后 purge topic (防跨测试污染)

# ========== Scope 阶段配置 ==========
scope:
  diff_strategy: auto          # auto=git diff / explicit=手动指定文件
  include_context_files: true  # 范围清单是否列出受影响上下文文件

# ========== 质量门禁 (gate) ==========
gates:
  no_critical_issues: true      # 无 Critical 级别 issue
  no_major_issues: false       # 无 Major 级别 issue (可放宽)
  tests_pass: true             # 所有测试通过
  coverage:
    enabled: true
    line_pct_min: 80           # 行覆盖率最低 80%
    branch_pct_min: 70         # 分支覆盖率最低 70%

# ========== Confidence 阶段配置 ==========
confidence:
  enabled: false               # 是否启用置信度门禁 (启用后所有 fixed 修复须达阈值)
  min_threshold: 0.8           # 最低置信度阈值 (0-1)
  checklist:                    # confidence-evaluator 核对项
    - "根因已定位 (非仅修症状)"
    - "修复最小化 (无 scope creep)"
    - "覆盖该路径的测试已通过"
    - "无回归风险 (相关调用方已检查)"

# ========== Agent 调度策略 ==========
agents:
  review_first: true           # 先 review 找问题再 fix
  parallel:
    review_and_hunt: true      # review 和 bug-hunt 并行 (独立分析)
  worktree_isolation: true      # 多 agent 写操作走 git worktree 隔离 (防并行冲突)
  order:
    - scope-analyzer
    - code-reviewer
    - bug-fixer
    - test-runner
    - confidence-evaluator

# ========== Review 阶段配置 (精度优化 ②③④, 默认关闭, 向后兼容) ==========
# 一键预设 (review.precision_preset): 套用 ②③④ 杠杆组合, 显式配置优先。
#   low      : 全关 (recall-first, ~71% precision / 100% recall, 默认行为)
#   balanced : ② loose + ③ verify (砍既有问题+投射性假阳性, 保 recall)
#   strict   : ② strict + ③ + ④ consensus (precision-max, 让 recall; 适合自审非门禁)
# 也可不设 preset, 单独开下面任意字段。详见 docs/PRECISION-TUNING.md
# review:
#   precision_preset: balanced
#
#   perspectives: [correctness, security, quality, tests]   # 审查视角, 可自定义
#   # ④ 杀手视角 + 共识: 抬 precision (让 recall, 当前100%有余量)
#   skeptic: true                     # 加一个 generalist 杀手视角 (全视角+反驳优先)
#   consensus_min: 2                  # finding 需被≥N视角报告才存活 (1=并集默认; 2=共识)
#   # ③ finding 二次验证器 (review 后 fix 前): 独立 agent 反驳, 丢弃假阳性
#   verify_findings: true             # 启用后需在 agent_models 配 finding-verifier 模型
#                                    # (可复用 confidence-evaluator 的模型配置)
#   # ② diff 锚定 (真实 PR 精度头号杠杆): 丢弃锚不到变更行的 finding,
#   # 砍"既有问题被当新 bug"类假阳性。需在 git 仓库内跑 (git diff 取变更行)。
#   diff_anchor: true                 # 启用后 review 后自动 git diff + 锚定过滤
#   anchor_mode: loose                # loose (默认, 变更行±context 内即留) / strict (必须落在变更行)
#   anchor_context: 3                 # loose 模式的相邻行容忍度

# ========== Agent → 模型映射 (核心: 中转代理或直连) ==========
# 每个 agent 声明 ModelConfig: provider/base_url/api_key/model
#   - base_url 指向中转代理 = 走中转 (如内部网关 / one-api / litellm)
#   - 不填 base_url            = 按 provider 直连 (anthropic/openai/...)
# 严禁硬编码模型; 改这里即可切换, 无需改代码。
agent_models:
  scope-analyzer:
    provider: openai            # 中转代理通常走 openai 兼容协议
    base_url: "https://relay.example.com/v1"   # ← 改成你的中转代理地址
    api_key: "${RELAY_API_KEY}"               # ← 从环境变量读, 勿明文
    model: "gpt-4o-mini"                      # 范围分析用快模型省成本
    fallbacks:
      - { provider: anthropic, model: "claude-haiku-4-5" }

  code-reviewer:
    provider: anthropic
    base_url: "https://relay.example.com"      # 中转代理 (无 /v1 时 anthropic 格式)
    api_key: "${RELAY_API_KEY}"
    model: "claude-sonnet-4-6"                 # 审查用强模型
    fallbacks:
      - { provider: openai, base_url: "https://relay.example.com/v1", model: "gpt-4o" }

  bug-fixer:
    provider: anthropic
    base_url: "https://relay.example.com"
    api_key: "${RELAY_API_KEY}"
    model: "claude-sonnet-4-6"
    fallbacks:
      - { provider: openai, base_url: "https://relay.example.com/v1", model: "gpt-4o" }

  test-data-preparer:
    provider: openai
    base_url: "https://relay.example.com/v1"
    api_key: "${RELAY_API_KEY}"
    model: "gpt-4o-mini"                       # 数据准备用快模型 (需写文件, 非深度推理)
    fallbacks:
      - { provider: anthropic, model: "claude-haiku-4-5" }

  test-runner:
    provider: openai
    base_url: "https://relay.example.com/v1"
    api_key: "${RELAY_API_KEY}"
    model: "gpt-4o-mini"                       # 测试规划用快模型
    fallbacks:
      - { provider: anthropic, model: "claude-haiku-4-5" }

  confidence-evaluator:
    provider: anthropic
    base_url: "https://relay.example.com"
    api_key: "${RELAY_API_KEY}"
    model: "claude-sonnet-4-6"                 # 置信度评估需深度推理
    fallbacks:
      - { provider: openai, base_url: "https://relay.example.com/v1", model: "gpt-4o" }

# ========== 上下文管理 ==========
context:
  compaction_threshold: 0.8     # 上下文使用率 > 80% 触发压缩 (复用 skillforge CompactionEngine)
  checkpoint: true              # 每轮落盘 summary 便于断点续跑

# ========== 升级策略 (loop 无法收敛时) ==========
escalation:
  on_repeated_critical: true    # 连续 escalate_on_new_critical 轮新 Critical → 升级
  on_max_iterations: true       # 达 max_iterations 未收敛 → 升级
  ask_user: true                # 升级时用 AskUserQuestion 征求决策 (不静默失败)

# ========== 项目特定覆盖示例 (取消注释按项目调整) ==========
# project_overrides:
#   coverage:
#     line_pct_min: 90   # 关键系统要求 90%
#   confidence:
#     enabled: true      # 关键系统强制置信度门禁