pm-coder 子Agent 的执行载体定义。 v3.1 变革：继承 harnesses/base/ 公共层，只保留 coder 特化逻辑。
公共逻辑（权限、安全、审计、检查点、交接、上下文、可观测）全部在 base/ 层定义， 本文件只声明 coder 独有的配置和行为。
设计参考：Claude Code 的规划管制 + 上下文工程 + 风险分级权限 + 交接棒 + 事件驱动钩子； Guides/Sensors 双闭环 + 两阶段 Code Review + 系统性调试 SOP。
harness_version: "2.0"
compatible_skill_version: ">=3.0"
compatible_framework: "Ironforge v1.0"
# ═══════════════════════════════════════
# 继承公共层
# ═══════════════════════════════════════
inherits:
- base/permission-framework.md # 三层权限执行 + spawn配置
- base/security-hooks.md # 敏感信息扫描 + 受保护路径
- base/audit-logging.md # 审计日志格式和事件
- base/checkpoint-protocol.md # 检查点自动保存和回滚
- base/handoff-protocol.md # 交接棒触发和流程
- base/context-engineering.md # 上下文分层和预算
- base/observability-config.md # HEARTBEAT和通信事件name: pm-coder
harness_type: subagent
description: |
编程执行子Agent。负责代码编写、调试、重构、测试。
通过 task 工具以 Team Mode spawn，加入项目团队协作。
内置"先规划后编码"管制、风险分级权限、自动质量钩子。
v2 新增：开发必须对齐原型设计。
spawn_config:
subagent_name: "code-explorer" # 内置 subagent 类型（代码探索+编写）
explore_phase:
mode: "plan"
max_turns: 10
execute_phase:
mode: "acceptEdits"
max_turns: 40
# name 和 team_name 由 pm-runner 动态填入specialization:
# ═══════════════════════════════════════
# 权限覆盖
# ═══════════════════════════════════════
permission_override:
execute_phase:
blocked_tools: [delete_file] # coder 不应删除文件
yellow_tools: [replace_in_file]
command_whitelist:
- npm install
- npm run build
- npm run test
- npm run lint
- tsc --noEmit
- python -m pytest
- python -m black --check
- eslint
- prettier --check
- pylint
- mypy
# ═══════════════════════════════════════
# 安全钩子覆盖
# ═══════════════════════════════════════
security_override:
protected_paths:
read_only:
add: ["src/**/*.test.*"] # 测试文件只读（防误改）
# ═══════════════════════════════════════
# 审计覆盖
# ═══════════════════════════════════════
audit_override:
additional_events:
- event_type: code_review
detail: {stage, result, findings}
# ═══════════════════════════════════════
# 检查点覆盖（v3.1 P1 稳定性）
# ═══════════════════════════════════════
checkpoint_override:
frequency: "every_step"
retention: "all"
rollback_strategy: "git_first"
additional_trigger: "每次 replace_in_file 后"
# ═══════════════════════════════════════
# 熔断覆盖（v3.1 P1 稳定性）
# ═══════════════════════════════════════
circuit_breaker_override:
agent_level:
failure_threshold: 3
window: "30min"
cooldown_base: "30min"
task_level:
retry_budget: 3
backoff: "exponential"
non_retryable_errors:
- "syntax_error" # 语法错误不重试
- "type_error" # 类型错误不重试
- "import_error" # 导入错误不重试
# ═══════════════════════════════════════
# 幂等性覆盖（v3.1 P1 稳定性）
# ═══════════════════════════════════════
idempotency_override:
replace_in_file:
old_str_min_context_lines: 3 # old_str 至少包含3行上下文
write_to_file:
always_check_exists: true # 始终检查文件是否已存在
command_pre_check:
npm_install: "node_modules/.package-lock.json 存在 → 跳过"
npm_run_build: "无前置检查（重复无害）"
npm_run_test: "无前置检查（重复无害）"
# ═══════════════════════════════════════
# 上下文预算覆盖
# ═══════════════════════════════════════
context_budget_override:
layer_1_hot: "≤ 3000 tokens"
layer_2_working: "≤ 15000 tokens"
handoff_trigger: "70% 轮次使用"
# ═══════════════════════════════════════
# 通信事件覆盖
# ═══════════════════════════════════════
communication_override:
additional_notify_events:
- event: debug_blocked
template: |
【debug_blocked】T{XXX} | pm-coder | 同一bug 2轮未修复guides_sensors:
# ─── Guides（前馈控制）───
guides:
- "角色身份 → SKILL.md Layer 1"
- "目标约束 → Goal success_criteria"
- "工作流程 → SKILL.md Steps"
- "编码标准 → references/code-standards.md"
- "审查协议 → references/code-review-protocol.md"
- "调试 SOP → references/debugging-protocol.md"
# ─── Sensors（反馈控制）───
sensors:
computational: # Coder 自执行（确定性，快，无幻觉）
- "H1 plan_compliance — plan 范围检查"
- "H2 file_ownership — 文件所有权检查"
- "H3 syntax_check — 编译/语法检查"
- "H4 heartbeat_sync — 产出物清单同步"
- "H7 full_test_suite — 测试套件"
- "H8 lint_check — 代码风格"
- "H9 deliverable_integrity — 产出物完整性"
trigger: "每次文件操作后 + 任务完成时"
lightweight_self_check: # 关键词匹配级
- "关键词匹配检查"
- "API 端点名称核对"
- "文件路径验证"
trigger: "任务完成时"
reasoning: # orchestrator 验收时执行
- "语义评估 — 对比 Goal success_criteria"
- "Spec 深度对照"
- "代码质量语义审查"
executor: "orchestrator（三角验证）"planning_first:
phases:
- phase: "explore" # Phase A: 探索（只读）
mode: "plan"
max_turns: 10
steps:
- "读取项目 HEARTBEAT"
- "读取上游任务 HEARTBEAT"
- "读取 context_pool/tech_stack.md"
- "读取 context_pool/architecture.md"
- "扫描现有代码结构"
- "识别影响范围和风险点"
output: "{context_root}/context_pool/progress/T{XXX}-plan.md"
notification: "send_message(event_type='plan_ready')"
- phase: "approval_gate" # Phase B: 审批等待
trigger: "orchestrator 审阅 plan.md"
outcomes:
approved: "进入 Phase C"
rejected: "附带修改意见 → 回到 Phase A"
needs_research: "委托 researcher → 重新规划"
- phase: "execute" # Phase C: 编码
mode: "acceptEdits"
max_turns: 40
guardrails:
- "严格按 plan.md 执行"
- "超出 plan 范围 → 请求 orchestrator 批准"
- "每完成一个里程碑 → 自动触发 post-step 钩子"hooks:
# ─── 编码前钩子 ───
pre_edit:
- name: "plan_compliance_check"
trigger: "每次 write_to_file 或 replace_in_file 前"
check: "操作是否在 plan.md 范围内"
- name: "file_ownership_check"
trigger: "每次 replace_in_file 前"
check: "文件是否被其他 Agent 占用"
# ─── 编码后钩子 ───
post_edit:
- name: "syntax_check"
trigger: "每次修改 .ts/.tsx/.py/.js/.vue 文件后"
commands: ["tsc --noEmit {file}", "python -m py_compile {file}"]
on_failure: "分析错误 → 修复 → 重试（最多3次）"
- name: "heartbeat_sync"
trigger: "每次 write_to_file / replace_in_file 成功后"
action: "更新 HEARTBEAT 产出物清单"
# ─── 步骤完成钩子 ───
post_step:
- name: "milestone_quality_gate"
trigger: "完成 plan.md 中的每个里程碑步骤后"
check: "执行对应的验证命令"
- name: "progress_report"
trigger: "每完成一个 plan step 后"
action: "send_message(event_type='task_progress')"
# ─── 任务完成钩子 ───
on_complete:
- name: "full_test_suite"
trigger: "准备发送 task_complete 前"
commands: ["npm run test", "python -m pytest"]
on_failure: "修复代码 → 重试（最多2次） → 仍失败则 task_partial_success"
- name: "lint_check"
trigger: "准备发送 task_complete 前"
commands: ["npm run lint", "pylint {src}"]
on_failure: "自动修复可修复的 → 无法修复标 warning"
- name: "deliverable_integrity"
trigger: "准备发送 task_complete 前"
check: "plan.md 文件是否都创建 + HEARTBEAT 产出物一致"
- name: "code_review_stage1"
type: "lightweight_self_check"
trigger: "Stage 2 之前"
check: "关键词匹配 + 文件完整性"
- name: "code_review_stage2"
type: "computational_sensor"
trigger: "计算型传感器全部通过后"
check: "H7/H8/H9 Hooks"
# ─── 调试模式钩子 ───
on_debug:
- name: "debug_phase_gate"
trigger: "调试类任务启动时"
check: "强制使用四阶段调试 SOP"
- name: "debug_stop_loss"
trigger: "同一 bug 2轮未修复"
action: "冻结 → send_message(task_blocked) → 建议 researcher 或人工介入"skill_loading:
tier_1_always:
- pm-coder/SKILL.md
tier_2_on_task:
- Domain Skills（vue3, electron, fastapi 等）
tier_3_on_demand:
- pm-coder/references/heartbeat-ops.md
- pm-coder/references/code-standards.md
- pm-coder/references/acceptance-criteria.md
- pm-coder/references/handoff-protocol.md
- pm-coder/references/hooks-specification.md
- pm-coder/references/code-review-protocol.md
- pm-coder/references/debugging-protocol.md
- pm-core/references/recovery-recipes.mdprompt_layers:
layer_0_intent:
- "当前执行阶段: {explore|execute}"
- "审批状态: {pending|approved|rejected}"
- "如为 execute → 附 plan.md 路径"
layer_1_identity:
- "角色身份 + 核心职责"
- "Goal success_criteria + constraints"
- "风险分级权限表"
- "可用 Skills Catalog"
layer_2_narrative:
- "项目 HEARTBEAT 摘要"
- "上游产出"
- "已完成阶段"
layer_3_focus:
- "当前任务描述 + ID + 类型"
- "输出要求（格式、位置、验收标准）"
- "上下文预算提醒"
- "记忆要求（HEARTBEAT操作节奏）"
- "三角验证流程"
- "可加载参考资料路径"| 任务类型 | 典型场景 | 额外 Skills | 预估轮次 |
|---|---|---|---|
| frontend | Vue3/React页面开发 | vue3, react, electron | 50 |
| backend | API/服务端开发 | fastapi, express, prisma | 50 |
| database | 数据库设计与实现 | sql, mongodb | 30 |
| testing | 单元/集成测试 | jest, pytest | 30 |
| debugging | Bug修复 | 视具体项目 | 30 |
| refactoring | 代码重构 | 视具体项目 | 40 |
Coder Harness v2.1 (Ironforge P1) — 2026-04-24