"""Agent loop — 角色agent循环认领执行任务（v3.4.0）。 v3.4.0 新增： - 所有agent具备文件读写 + git pull/push 能力 - agent 产出实际代码文件到项目仓库，而非仅文档 - PM agent 负责 clone 项目关联仓库、初始化工作区 - develop agent 产出可直接运行的源码文件 - 每个agent完成后自动 git commit + push 任务链：requirement → design → develop → test → deploy """ import asyncio import json import os import subprocess import logging logger = logging.getLogger("pipeline.agent_loop") ROLE_ALIASES = { 'developer': 'develop', 'dev': 'develop', 'development': 'develop', 'coding': 'develop', 'requirement': 'requirement', 'requirements': 'requirement', 'requirement_analysis': 'requirement', 'designer': 'design', 'ui': 'design', 'ux': 'design', 'testing': 'test', 'qa': 'test', 'tester': 'test', 'deployment': 'deploy', 'release': 'deploy', 'operation': 'ops', 'operations': 'ops', 'maintenance': 'ops', '运维': 'ops', 'pm': 'pm', 'project_manager': 'pm', '项目经理': 'pm', 'manager': 'pm', } ROLE_CHAIN = { 'requirement': 'design', 'design': 'develop', 'develop': 'test', 'test': 'deploy', 'deploy': None, } TASK_REVIEW = 'review' TASK_APPROVED = 'approved' # 允许的 shell 工作目录前缀（安全限制） _ALLOWED_WORKDIRS = [ os.path.expanduser('~/pipeline_ws'), '/d/pipeline/workspaces', '/tmp/pipeline_workspaces', '/tmp/pipeline_ws', ] def _normalize_role(role): r = (role or '').strip().lower() return ROLE_ALIASES.get(r, r) def _get_db(): from sqlor.dbpools import DBPools db = DBPools() if not db.databases: from appPublic.jsonConfig import getConfig config = getConfig() if config.databases: db.databases = config.databases return db def _resolve_workspace(workspace_dir): if not workspace_dir: workspace_dir = os.path.expanduser('~/pipeline_ws/default') try: parent = os.path.dirname(workspace_dir) if parent and not os.access(parent, os.W_OK): proj_name = os.path.basename(workspace_dir) workspace_dir = os.path.expanduser(f'~/pipeline_ws/{proj_name}') except Exception: workspace_dir = os.path.expanduser(f'~/pipeline_ws/{os.path.basename(workspace_dir) or "default"}') os.makedirs(workspace_dir, exist_ok=True) return workspace_dir # ── Agent 工具函数 ── def _is_safe_workdir(workdir): """检查目录是否在允许范围内。""" wd = os.path.abspath(workdir) for allowed in _ALLOWED_WORKDIRS: awd = os.path.abspath(os.path.expanduser(allowed)) if wd.startswith(awd): return True return False async def _run_shell(command, workdir, timeout=120): """安全执行 shell 命令。返回 {"rc": int, "stdout": str, "stderr": str}""" cwd = os.path.abspath(workdir) if workdir else os.path.expanduser('~/pipeline_ws') if not _is_safe_workdir(cwd): return {"rc": -1, "stdout": "", "stderr": f"安全限制：目录 {cwd} 不在允许范围"} if not os.path.isdir(cwd): return {"rc": -1, "stdout": "", "stderr": f"目录不存在: {cwd}"} try: proc = await asyncio.create_subprocess_shell( command, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE, cwd=cwd, executable='/bin/bash') try: stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=timeout) except asyncio.TimeoutError: proc.kill() await proc.wait() return {"rc": -1, "stdout": "", "stderr": f"命令超时（{timeout}s）"} return {"rc": proc.returncode or 0, "stdout": stdout.decode('utf-8', 'replace')[-8000:], "stderr": stderr.decode('utf-8', 'replace')[-4000:]} except Exception as e: return {"rc": -1, "stdout": "", "stderr": str(e)[:500]} async def _write_code_file(filepath, content): """写代码文件，自动创建父目录。""" try: d = os.path.dirname(filepath) if d: os.makedirs(d, exist_ok=True) with open(filepath, 'w', encoding='utf-8') as f: f.write(content or '') return True, filepath except Exception as e: return False, str(e) async def _git_setup(workdir): """确保 git 用户已配置。""" r = await _run_shell('git config user.email', workdir, 5) if 'opencomputing' not in r.get('stdout', '') and '@' not in r.get('stdout', ''): await _run_shell('git config user.email "pipeline@opencomputing.cn"', workdir, 5) await _run_shell('git config user.name "Pipeline Agent"', workdir, 5) async def _git_commit_push(workdir, commit_message, branch='main'): """git add → commit → push。""" await _git_setup(workdir) r1 = await _run_shell('git add -A', workdir, 10) r2 = await _run_shell(f'git diff --cached --stat', workdir, 10) if not r2.get('stdout', '').strip(): return {"rc": 0, "message": "没有变更需要提交"} r3 = await _run_shell(f'git commit -m "{commit_message}"', workdir, 15) if r3['rc'] != 0: return {"rc": r3['rc'], "message": f"commit 失败: {r3['stderr'][:200]}"} r4 = await _run_shell(f'git push origin {branch}', workdir, 30) return {"rc": r4['rc'], "message": f"push {'成功' if r4['rc']==0 else '失败'}: {r4['stderr'][:200]}"} async def _git_clone(repo_url, target_dir, branch='main'): """克隆仓库到目标目录。已存在则 pull。""" if os.path.isdir(os.path.join(target_dir, '.git')): r = await _run_shell(f'git checkout {branch} && git pull origin {branch}', target_dir, 30) return {"rc": r['rc'], "message": f"已存在，pull: {r['stdout'][:200]}"} parent = os.path.dirname(target_dir) os.makedirs(parent, exist_ok=True) r = await _run_shell(f'git clone -b {branch} {repo_url} {target_dir}', parent, 120) return {"rc": r['rc'], "message": f"clone: {r['stdout'][:200] if r['rc']==0 else r['stderr'][:200]}"} # ── Prompts ── ROLE_PROMPT = """你是软件开发产线项目中的「__ROLE__」角色agent。 当前任务：__TITLE__ 参数：__PARAMS__ __QNA__ ## 工作环境 - 项目根目录：__WORKSPACE__ - 代码仓库位于：__WORKSPACE__/repos/ 下各子目录 - 你有能力读写文件、执行git操作 ## 产出要求 __ROLE_SPECIFIC__ ## 输出格式（严格JSON，不要markdown包裹） { "status": "done", "deliverable_type": "类型(requirement_doc/design_doc/code_files/test_report/deploy_doc)", "summary": "一段话概述产出", "result": "当产出是文档时，这里是完整文档内容", "files": [ {"path": "repos/仓库名/src/相对路径/文件.java", "content": "文件完整内容"}, {"path": "repos/仓库名/pom.xml", "content": "..."} ], "git_commit_message": "feat: 简短描述本次变更", "need_more_info": false, "question": "" } - 如果是文档类产出(requirement/design)，用 result 字段输出文档正文 - 如果是代码产出(develop)，必须用 files 数组输出每个源码文件，每个元素含 path 和 content - 测试产出用 result 输出测试报告，如有测试脚本则用 files - 部署产出用 files 输出部署配置/脚本，用 result 输出部署说明 - 缺少关键信息时：{"status":"need_info","question":"问题"} ## Git 操作说明 - 产出代码文件将自动写入 files 中指定的路径并 git commit + push - 你只需要在 files 中指定正确的仓库路径即可""" ROLE_SPECIFICS = { 'requirement': """你是需求分析师。输出完整的需求规格文档（Markdown），用 result 字段。 文档应包含：项目概述、功能需求、非功能需求、用户角色、业务流程、验收标准。""", 'design': """你是系统设计师。输出完整的设计文档（Markdown），用 result 字段。 文档应包含：系统架构、技术选型、数据库设计(ER图描述)、API接口设计、前端页面规划、安全设计。 同时可输出 proto/接口定义文件到 files。""", 'develop': """你是开发工程师。你必须输出实际可运行的代码文件！ 用 files 数组输出每个源码文件，准确指定在仓库中的路径。 files 中的 path 格式：repos/{仓库名}/src/... 或 repos/{仓库名}/pom.xml 等。 每个文件包含完整可运行的代码。 同时用 result 字段输出代码说明。""", 'test': """你是测试工程师。输出测试报告（Markdown）到 result 字段。 如有测试脚本/用例代码，输出到 files。 报告应包含：测试策略、测试用例、执行结果、Bug清单、覆盖率。""", 'deploy': """你是部署运维工程师。输出部署文档（Markdown）到 result 字段。 必须输出部署配置/脚本到 files：Dockerfile/docker-compose.yml/nginx配置等。 files 中的 path 格式：repos/{仓库名}/deploy/...""", } PM_PROMPT = """你是项目经理（PM）。审核交付件并推动项目前进。 ## 项目 工作目录：__WORKSPACE__ 关联仓库：__REPOS__ ## 待审核任务 标题：__TITLE__ 角色：__ROLE__ 交付件摘要：__SUMMARY__ 交付件内容(前6000字)：__DELIVERABLE__ ## 仓库状态 __REPO_STATE__ ## 决策 输出纯JSON： 1. 通过：{"status":"approved","comment":"审核意见","next_task_title":"下阶段标题","next_task_description":"详细描述"} 2. 驳回：{"status":"rejected","comment":"驳回原因","questions":"修改要求"} 3. 完成：{"status":"completed","comment":"项目总结"} 审核标准： - requirement：需求是否清晰完整可量化 - design：方案合理、覆盖需求、技术可行 - develop：代码文件是否实际产出、目录结构是否合理、是否可编译运行 - test：测试覆盖充分、发现问题记录完整 - deploy：部署配置完整、可一键部署""" # ── Agent 核心逻辑 ── async def _claim_task(sor, tenant_id, role, state='submitted', match_role=True): role = _normalize_role(role) where_role = "AND (role=${role}$ OR role='')" if match_role else "" recs = await sor.sqlExe( "SELECT id, title, params, pipeline_id, tenant_id, role FROM pipeline_tasks " "WHERE tenant_id=${tid}$ AND state=${state}$ " + where_role + " " "AND NOT EXISTS (SELECT 1 FROM pipeline_task_steps s WHERE s.task_id=pipeline_tasks.id) " "ORDER BY created_at ASC LIMIT 1", {"tid": tenant_id, "state": state, "role": role}) if not recs: return None task = recs[0] task_id = task.id from appPublic.uniqueID import getID claim_token = getID() await sor.sqlExe( "UPDATE pipeline_tasks SET state='running', claimed_by=${cb}$ " "WHERE id=${tid}$ AND state=${state}$", {"cb": claim_token, "tid": task_id, "state": state}) check = await sor.sqlExe( "SELECT id FROM pipeline_tasks WHERE id=${tid}$ AND state='running' AND claimed_by=${cb}$", {"tid": task_id, "cb": claim_token}) if not check: logger.info(f"claim lost race: task={task_id}") return None return task async def _get_workspace_dir(sor, project_id): recs = await sor.sqlExe("SELECT workspace_dir FROM sd_projects WHERE id=${pid}$", {"pid": project_id}) ws = getattr(recs[0], 'workspace_dir', '') if recs else '' return _resolve_workspace(ws) async def _build_qna_section(sor, task_id): from .questions import get_task_qna qna = await get_task_qna(task_id) if not qna: return "" lines = ["历史问答："] for item in qna: src = "客户" if item.get('answer_source') == "customer" else "主agent" lines.append(f"问：{item.get('question','')}") lines.append(f"答({src})：{item.get('answer','')}") return "\n".join(lines) def _parse_result(raw): raw = (raw or "").strip() if raw.startswith("```"): raw = raw.split("\n", 1)[1].rsplit("```", 1)[0] try: d = json.loads(raw) if isinstance(d, dict): return d except (json.JSONDecodeError, ValueError): pass return {"status": "done", "result": raw} async def _get_deliverable_content(sor, task_id): recs = await sor.sqlExe( "SELECT content, deliverable_type FROM pipeline_deliverables " "WHERE task_id=${tid}$ ORDER BY created_at DESC LIMIT 1", {"tid": task_id}) if recs: return getattr(recs[0], 'content', '') or '', getattr(recs[0], 'deliverable_type', '') or '' return '', '' async def _get_project_repos(sor, project_id): """获取项目关联的所有仓库。""" recs = await sor.sqlExe( "SELECT id, repo_name, repo_url, default_branch, local_path FROM sd_project_repos " "WHERE project_id=${pid}$", {"pid": project_id}) return [{'id': getattr(r, 'id', ''), 'name': getattr(r, 'repo_name', ''), 'url': getattr(r, 'repo_url', ''), 'branch': getattr(r, 'default_branch', 'main'), 'local_path': getattr(r, 'local_path', '') or ''} for r in (recs or [])] async def _setup_repos(sor, workspace_dir, project_id): """PM：clone 所有项目关联仓库到 workspace/repos/。""" repos = await _get_project_repos(sor, project_id) results = [] repos_dir = os.path.join(workspace_dir, 'repos') os.makedirs(repos_dir, exist_ok=True) for repo in repos: target = os.path.join(repos_dir, repo['name']) r = await _git_clone(repo['url'], target, repo['branch']) results.append({'repo': repo['name'], **r}) return results async def _get_repo_state(workspace_dir): """获取仓库当前状态（供PM审核时查看）。""" repos_dir = os.path.join(workspace_dir, 'repos') if not os.path.isdir(repos_dir): return "暂无仓库" lines = [] for name in sorted(os.listdir(repos_dir)): rp = os.path.join(repos_dir, name) if os.path.isdir(rp) and os.path.isdir(os.path.join(rp, '.git')): r = await _run_shell('git log --oneline -3', rp, 5) lines.append(f"\n[{name}]") lines.append(r.get('stdout', '') or '(空仓库)') return "\n".join(lines) if lines else "仓库无提交记录" async def _get_next_role(current_role): return ROLE_CHAIN.get(_normalize_role(current_role)) async def _create_next_task(sor, project_id, task, next_role, pm_comment=''): from appPublic.uniqueID import getID title = getattr(task, 'title', '') or '' params_str = getattr(task, 'params', '{}') or '{}' try: params = json.loads(params_str) if isinstance(params_str, str) else params_str except (json.JSONDecodeError, TypeError): params = {} new_title = f"{title}（{next_role}阶段）" new_params = {**params, 'previous_role': _normalize_role(getattr(task, 'role', '')), 'previous_task_id': getattr(task, 'id', ''), 'pm_comment': pm_comment} new_task_id = getID() await sor.C('pipeline_tasks', { 'id': new_task_id, 'tenant_id': project_id, 'pipeline_id': 'role_task', 'owner_id': 'pm', 'title': new_title, 'params': json.dumps(new_params, ensure_ascii=False), 'role': next_role, 'state': 'submitted', 'claimed_by': None, }) logger.info(f"_create_next_task: {next_role} task={new_task_id}") return new_task_id, new_title # ── 角色 Agent ── async def role_agent_run(project_id, role, agent_id=None, model_name=None): role = _normalize_role(role) if role == 'pm': return {"status": "idle", "message": "PM请使用 pm_review_run"} db = _get_db() async with db.sqlorContext("pipeline") as sor: task = await _claim_task(sor, project_id, role) if not task: return {"status": "idle", "message": "没有待办任务"} task_id = task.id title = getattr(task, "title", "") or "" params_str = getattr(task, "params", "{}") or "{}" workspace_dir = await _get_workspace_dir(sor, project_id) # 确保 repos 目录存在 repos_dir = os.path.join(workspace_dir, 'repos') if not os.path.isdir(repos_dir): os.makedirs(repos_dir, exist_ok=True) role_specific = ROLE_SPECIFICS.get(role, ROLE_SPECIFICS.get('develop', '')) qna_section = await _build_qna_section(sor, task_id) prompt = (ROLE_PROMPT .replace('__ROLE__', role) .replace('__TITLE__', title) .replace('__PARAMS__', params_str) .replace('__QNA__', qna_section) .replace('__WORKSPACE__', workspace_dir) .replace('__ROLE_SPECIFIC__', role_specific)) from .llm_bridge import llm_call try: raw = await llm_call(prompt, model=model_name, temperature=0.4) except Exception as e: await sor.sqlExe("UPDATE pipeline_tasks SET state='failed' WHERE id=${tid}$", {"tid": task_id}) logger.error(f"role_agent_run llm failed: task={task_id} err={e}") return {"status": "failed", "task_id": task_id, "error": str(e)[:200]} parsed = _parse_result(raw) if parsed.get("status") == "need_info" and parsed.get("question"): from .questions import agent_ask qid = await agent_ask(project_id, task_id, role, parsed["question"], context={"partial": parsed.get("partial", ""), "title": title}) return {"status": "need_info", "task_id": task_id, "question_id": qid, "question": parsed["question"]} # ── 处理产出 ── from appPublic.uniqueID import getID did = getID() result_text = parsed.get("result") or raw deliverable_type = parsed.get("deliverable_type") or role summary = parsed.get("summary", "") # 写入实际代码文件（如果有 files 数组） files_written = [] code_files = parsed.get("files") or [] if isinstance(code_files, list): for f in code_files: if isinstance(f, dict) and f.get("path") and f.get("content"): abs_path = os.path.join(workspace_dir, f["path"]) ok, msg = await _write_code_file(abs_path, f["content"]) if ok: files_written.append(abs_path) logger.info(f"code file written: {abs_path}") else: logger.error(f"code file failed: {abs_path} err={msg}") # 写交付件文档 file_path = '' if result_text: deliverable_dir = os.path.join(workspace_dir, 'deliverables', role) os.makedirs(deliverable_dir, exist_ok=True) safe_type = (deliverable_type or 'deliverable').replace('/', '_') file_path = os.path.join(deliverable_dir, f"{task_id}_{safe_type}.md") try: with open(file_path, 'w', encoding='utf-8') as f: f.write(result_text or '') logger.info(f"deliverable written: {file_path}") except Exception as e: logger.error(f"deliverable write failed: {file_path} err={e}") file_path = '' # 写DB await sor.C("pipeline_deliverables", { "id": did, "project_id": project_id, "task_id": task_id, "deliverable_type": deliverable_type, "title": title, "content": result_text, "file_path": file_path, "quality_score": 80, "review_status": "pending", "created_by": agent_id or role, }) # Git commit + push（如果有产出文件） commit_msg = parsed.get("git_commit_message") or f"{role}: {title[:80]}" git_result = {"rc": 0, "message": "无git操作"} if files_written: # 在 repos 目录下的每个 git 仓库各自 commit for repo_name in os.listdir(repos_dir): repo_path = os.path.join(repos_dir, repo_name) if os.path.isdir(os.path.join(repo_path, '.git')): git_result = await _git_commit_push(repo_path, commit_msg) # 置为 review 状态 await sor.sqlExe( "UPDATE pipeline_tasks SET state=${st}$, claimed_by=NULL WHERE id=${tid}$", {"st": TASK_REVIEW, "tid": task_id}) logger.info(f"role_agent_run completed: task={task_id} role={role} " f"deliverable={did} files={len(files_written)} git={git_result.get('rc',-1)}") return {"status": "completed", "task_id": task_id, "deliverable_id": did, "files_written": len(files_written), "git_result": git_result} # ── PM 审核 ── async def pm_review_run(project_id, agent_id=None, model_name=None): db = _get_db() async with db.sqlorContext("pipeline") as sor: task = await _claim_task(sor, project_id, '', state=TASK_REVIEW, match_role=False) if not task: return {"status": "idle", "message": "没有待审核任务"} task_id = task.id title = getattr(task, "title", "") or "" task_role = _normalize_role(getattr(task, "role", "") or "") workspace_dir = await _get_workspace_dir(sor, project_id) # 首次审核时，确保仓库已 clone repos = await _get_project_repos(sor, project_id) if repos: await _setup_repos(sor, workspace_dir, project_id) deliverable_content, deliverable_type = await _get_deliverable_content(sor, task_id) if not deliverable_content: await sor.sqlExe("UPDATE pipeline_tasks SET state='submitted' WHERE id=${tid}$", {"tid": task_id}) return {"status": "rejected", "task_id": task_id, "reason": "没有交付件"} summary = deliverable_content[:200].replace('\n', ' ') content_preview = deliverable_content[:6000] if len(deliverable_content) > 6000: content_preview += "\n...（截断）" repo_state = await _get_repo_state(workspace_dir) repos_str = ", ".join([r['name'] for r in repos]) if repos else "无" prompt = (PM_PROMPT .replace('__TITLE__', title) .replace('__ROLE__', task_role) .replace('__SUMMARY__', summary) .replace('__DELIVERABLE__', content_preview) .replace('__WORKSPACE__', workspace_dir) .replace('__REPOS__', repos_str) .replace('__REPO_STATE__', repo_state)) from .llm_bridge import llm_call try: raw = await llm_call(prompt, model=model_name, temperature=0.3) except Exception as e: await sor.sqlExe("UPDATE pipeline_tasks SET state='failed' WHERE id=${tid}$", {"tid": task_id}) return {"status": "failed", "task_id": task_id, "error": str(e)[:200]} parsed = _parse_result(raw) decision = parsed.get("status", "rejected") comment = parsed.get("comment", "") if decision == "approved": from appPublic.uniqueID import getID pm_did = getID() await sor.C("pipeline_deliverables", { "id": pm_did, "project_id": project_id, "task_id": task_id, "deliverable_type": "pm_review", "title": f"PM审核：{title}", "content": json.dumps(parsed, ensure_ascii=False), "file_path": os.path.join(workspace_dir, 'deliverables', 'pm', f"{task_id}_review.md"), "quality_score": 100, "review_status": "approved", "created_by": agent_id or "pm", }) await sor.sqlExe( "UPDATE pipeline_deliverables SET review_status='approved', review_comment=${cm}$ " "WHERE task_id=${tid}$", {"cm": comment, "tid": task_id}) await sor.sqlExe( "UPDATE pipeline_tasks SET state=${st}$ WHERE id=${tid}$", {"st": TASK_APPROVED, "tid": task_id}) next_role = await _get_next_role(task_role) if next_role: next_tid, next_title = await _create_next_task(sor, project_id, task, next_role, comment) return {"status": "approved", "task_id": task_id, "next_task_id": next_tid, "next_role": next_role, "comment": comment} else: return {"status": "completed", "task_id": task_id, "comment": comment or "项目完成"} elif decision == "rejected": from .questions import agent_ask rejection_q = parsed.get("questions") or comment or "交付件不满足要求" await agent_ask(project_id, task_id, "pm", rejection_q, context={"pm_comment": comment, "deliverable_type": deliverable_type}) await sor.sqlExe("UPDATE pipeline_tasks SET state='submitted' WHERE id=${tid}$", {"tid": task_id}) return {"status": "rejected", "task_id": task_id, "comment": comment, "question": rejection_q} else: await sor.sqlExe("UPDATE pipeline_tasks SET state='completed' WHERE id=${tid}$", {"tid": task_id}) return {"status": "completed", "task_id": task_id, "comment": comment or "项目完成"} async def role_agent_loop(project_id, role, agent_id=None, model_name=None, max_iterations=10): results = [] for _ in range(max_iterations): r = await role_agent_run(project_id, role, agent_id, model_name) results.append(r) if r["status"] in ("idle", "need_info", "failed"): break await asyncio.sleep(1) return results async def pm_review_loop(project_id, agent_id=None, model_name=None, max_iterations=20): results = [] for _ in range(max_iterations): r = await pm_review_run(project_id, agent_id, model_name) results.append(r) if r["status"] in ("idle", "failed"): break await asyncio.sleep(1) return results # 向后兼容 async def run_agent_loop(project_id, role_name, model_name=None): role = _normalize_role(role_name or "develop") if role == 'pm': return await pm_review_run(project_id, model_name=model_name) return await role_agent_run(project_id, role_name or "develop", model_name=model_name) async def agent_loop(project_id, role_name, model_name=None, max_iterations=10): role = _normalize_role(role_name or "develop") if role == 'pm': return await pm_review_loop(project_id, model_name=model_name, max_iterations=max_iterations) return await role_agent_loop(project_id, role_name or "develop", model_name=model_name, max_iterations=max_iterations)