#!/usr/bin/env python3 """ 宿主编排层（无 GUI）：对齐 ChatGPT / Codex 桌面版宿主该干的事， 而不是再写一个 agent 循环。 ChatGPT / Codex 桌面版作为宿主，核心就这几件（见 README 的对齐表）： 1. 线程生命周期：建、续跑、fork、归档、删除、列表、持久化 2. 并行派发：多 thread 同时跑，每任务一个 git worktree 隔离 3. 子代理委派：主 thread 派子 thread，子 thread 结果汇总回主 thread （模型内子代理是另一层，见 examples/subagent-prompt.txt； 这里是宿主级委派，用 parentTurnId / rootTurnId 把委派链串起来） 4. 长任务 / 目标循环：同一 thread 多回合，直到验收命令通过或到上限 5. 评审队列：diff + 状态 + 合并 / 丢弃，人工在合并前把关 6. 自动化：按间隔到期的任务自动派发，结果进评审队列（无推送， 这是云端 ChatGPT 才有的部分，见 README 未对齐表） 7. 边界：宿主定 cwd、沙箱、审批、模型、MCP/skills，agent 只在边界内跑 实现方式：一个 app-server 进程托管多个 thread（官方 thread manager 就是这么设计的），本文件用后台读线程按 id / threadId 分发响应和 turn/completed 通知，所以多个 thread 的 turn 可以并发跑。 线程状态本身持久化在 CODEX_HOME，host-state.json 只存宿主自己的 任务登记（task -> thread/worktree/状态/diff），不是 agent 状态的 第二份拷贝。 用法示例： python3 host.py dispatch --repo work/host-repo --task "给 sample.py 加 add2 函数" --name add2 python3 host.py list python3 host.py status add2 python3 host.py review add2 python3 host.py merge add2 # 评审通过后才合并，冲突会 abort python3 host.py discard add2 # 评审不通过，删 worktree 和分支 python3 host.py run --task add2 --prompt "再补一个测试" python3 host.py fork --task add2 --name add2-v2 python3 host.py delegate --parent-task add2 --task "只读检查 add2 有没有边界问题" --name add2-review python3 host.py goal --repo work/host-repo --task "创建 result.txt，内容只有一行 goal-ok" \ --verify "test -f result.txt" --name goal1 python3 host.py parallel --jobs examples/host-jobs.json python3 host.py automations --add --name daily-check --repo work/host-repo \ --task "只读检查 sample.py 有没有明显问题" --interval 86400 python3 host.py run-due 默认 CODEX_HOME 是 ~/.codex；测试时用 --codex-home 指向隔离目录。 默认用 PATH 里的 codex；本机 VM 需用配对二进制时设 CODEX_BIN=.../bin/codex（start.sh 已自动优先用 bin/codex）。 """ import argparse import json import os import queue import subprocess import sys import threading import time import uuid from concurrent.futures import ThreadPoolExecutor HERE = os.path.dirname(os.path.abspath(__file__)) DEFAULT_STATE = os.path.join(HERE, "host-state.json") def default_codex_bin(): if os.environ.get("CODEX_BIN"): return os.environ["CODEX_BIN"] paired = os.path.join(HERE, "bin", "codex") if os.path.exists(paired): return paired return "codex" class Server: """单个 app-server 进程，多 thread 并发。""" def __init__(self, codex_bin=None, codex_home=None, approval="deny"): self.codex_bin = codex_bin or default_codex_bin() env = dict(os.environ) if codex_home: os.makedirs(codex_home, exist_ok=True) env["CODEX_HOME"] = codex_home self.proc = subprocess.Popen( [self.codex_bin, "app-server"], stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, env=env, ) self.send_lock = threading.Lock() self.next_id = 1 self.pending = {} self.pending_lock = threading.Lock() self.turn_queues = {} self.events = [] self.events_lock = threading.Lock() self.approval = approval self.reader = threading.Thread(target=self._read_loop, daemon=True) self.reader.start() def _answer_server_request(self, msg): # MCP 工具审批不是通用 approval：服务端发的是 # mcpServer/elicitation/request（mode=form，requestedSchema）， # 按 MCP elicitation 协议回 action=accept/decline + content， # 回通用 {"decision": ...} 会被当成拒绝（2026-10-25 实测）。 if msg.get("method") == "mcpServer/elicitation/request": if self.approval == "auto": result = {"action": "accept", "content": {}} else: result = {"action": "decline"} self._send({"jsonrpc": "2.0", "id": msg["id"], "result": result}) return decision = "approve" if self.approval == "auto" else "deny" self._send({"jsonrpc": "2.0", "id": msg["id"], "result": {"decision": decision}}) def _read_loop(self): try: for raw in self.proc.stdout: if not raw.strip(): continue try: msg = json.loads(raw) except Exception: continue if "method" in msg and "id" in msg: self._answer_server_request(msg) continue if "id" in msg: with self.pending_lock: q = self.pending.get(msg["id"]) if q is not None: q.put(msg) continue # 通知 params = msg.get("params") or {} thread_id = params.get("threadId") if not thread_id and isinstance(params.get("thread"), dict): thread_id = params["thread"].get("id") with self.events_lock: self.events.append(msg) if msg.get("method") == "turn/completed" and thread_id: q = self.turn_queues.get(thread_id) if q is not None: q.put(msg) except Exception: pass def _send(self, obj): with self.send_lock: self.proc.stdin.write((json.dumps(obj) + "\n").encode()) self.proc.stdin.flush() def request(self, method, params=None, timeout=120): rid = self.next_id self.next_id += 1 q = queue.Queue() with self.pending_lock: self.pending[rid] = q try: self._send({"jsonrpc": "2.0", "id": rid, "method": method, "params": params or {}}) return q.get(timeout=timeout) finally: with self.pending_lock: self.pending.pop(rid, None) def notify(self, method, params=None): self._send({"jsonrpc": "2.0", "method": method, "params": params or {}}) def initialize(self): resp = self.request("initialize", { "clientInfo": {"name": "codex-host", "version": "0.1"}, }, timeout=60) if not resp or "error" in resp: raise RuntimeError(f"initialize 失败: {resp}") self.notify("initialized", {}) return resp["result"] def create_thread(self, cwd, model=None, sandbox="workspace-write", approval_policy="never", instructions=None): params = {"cwd": os.path.abspath(cwd), "approvalPolicy": approval_policy, "sandbox": sandbox} if model: params["model"] = model if instructions: params["developerInstructions"] = instructions resp = self.request("thread/start", params, timeout=120) if not resp or "error" in resp: raise RuntimeError(f"thread/start 失败: {resp}") return resp["result"]["thread"]["id"] def list_threads(self): resp = self.request("thread/list", {}, timeout=60) if not resp or "error" in resp: raise RuntimeError(f"thread/list 失败: {resp}") return resp["result"] def resume_thread(self, thread_id): resp = self.request("thread/resume", {"threadId": thread_id}, timeout=120) if not resp or "error" in resp: raise RuntimeError(f"thread/resume 失败: {resp}") return resp["result"] def fork_thread(self, thread_id, cwd=None, sandbox=None): params = {"threadId": thread_id} if cwd: params["cwd"] = os.path.abspath(cwd) if sandbox: params["sandbox"] = sandbox resp = self.request("thread/fork", params, timeout=120) if not resp or "error" in resp: raise RuntimeError(f"thread/fork 失败: {resp}") # fork 响应里 thread 字段形态随版本略有差异，尽量都取 r = resp["result"] if isinstance(r, dict): t = r.get("thread") or {} return t.get("id") or r.get("threadId") or r.get("id") return None def archive_thread(self, thread_id): resp = self.request("thread/archive", {"threadId": thread_id}, timeout=60) if resp and "error" in resp: raise RuntimeError(f"thread/archive 失败: {resp}") return resp def run_turn(self, thread_id, prompt, model=None, effort=None, parent_turn_id=None, root_turn_id=None, timeout=600, sandbox=None): q = queue.Queue() self.turn_queues[thread_id] = q start_events = len(self.events) try: params = {"threadId": thread_id, "input": [{"type": "text", "text": prompt}]} if model: params["model"] = model if effort: params["effort"] = effort if sandbox: # turn 级 sandboxPolicy 会覆盖 thread 级设置，且对后续 # turn 也生效。这是修复 fork 换 cwd 后被重置为 read-only # 的关键：fork 时传 sandbox 参数服务端不认，turn 里传才认。 tmap = {"workspace-write": "workspaceWrite", "read-only": "readOnly", "danger-full-access": "dangerFullAccess"} params["sandboxPolicy"] = {"type": tmap.get(sandbox, sandbox)} if parent_turn_id: params["parentTurnId"] = parent_turn_id if root_turn_id: params["rootTurnId"] = root_turn_id resp = self.request("turn/start", params, timeout=120) if not resp or "error" in resp: return {"status": "failed", "text": "", "error": resp} turn_id = None if isinstance(resp.get("result"), dict): t = resp["result"].get("turn") or {} turn_id = t.get("id") try: done = q.get(timeout=timeout) except queue.Empty: return {"status": "timeout", "text": "", "turn_id": turn_id} turn = (done.get("params") or {}).get("turn") or {} text = "" for item in turn.get("items") or []: if item.get("type") == "agentMessage" and \ item.get("phase") == "final_answer": text = item.get("text") or text if not text: # 兜底：从本回合事件里找 final_answer with self.events_lock: evs = list(self.events[start_events:]) for ev in evs: p = ev.get("params") or {} item = p.get("item") or {} if item.get("type") == "agentMessage" and \ item.get("phase") == "final_answer" and \ item.get("text"): text = item["text"] return {"status": turn.get("status", "completed"), "text": text, "turn_id": turn.get("id") or turn_id, "root_turn_id": turn.get("rootTurnId"), "error": turn.get("error")} finally: self.turn_queues.pop(thread_id, None) def close(self): try: self.proc.stdin.close() except Exception: pass try: self.proc.wait(timeout=5) except Exception: self.proc.kill() # ---------------- 宿主状态 / worktree / diff ---------------- # 生产加固：state 锁。锁只包住“读/写”这两个短临界区，不包住 # 整个模型 turn（turn 可能跑几分钟，包住它会把别的 host 进程 # 活活饿死）。保存时先在锁里重读磁盘，再把内存里的任务按名 # 覆盖、自动化按名合并，这样两个进程各改各的任务时，后保存 # 的不会把先保存的覆盖掉。tmp 文件带 pid，避免两个进程用 # 同一个 .tmp 互相踩。锁文件在 state 文件旁边，名如 # host-state.json.lock，用 fcntl.flock（Linux/macOS 都有）。 import contextlib try: import fcntl except ImportError: # Windows 等无 fcntl 的平台：退化为无锁 fcntl = None @contextlib.contextmanager def _state_lock(path, exclusive=True): lock_path = path + ".lock" os.makedirs(os.path.dirname(os.path.abspath(lock_path)), exist_ok=True) f = open(lock_path, "a+") try: if fcntl is not None: fcntl.flock(f.fileno(), fcntl.LOCK_EX if exclusive else fcntl.LOCK_SH) yield finally: if fcntl is not None: try: fcntl.flock(f.fileno(), fcntl.LOCK_UN) except Exception: pass f.close() def _read_state_file(path): if os.path.exists(path): try: return json.load(open(path)) except Exception: pass return {"tasks": {}, "automations": []} def load_state(path): with _state_lock(path, exclusive=False): return _read_state_file(path) def save_state(path, state): with _state_lock(path, exclusive=True): # 锁内重读 + 合并：内存任务覆盖同名磁盘任务，磁盘上 # 别的进程新增的任务保留；自动化按 name 合并，内存优先。 disk = _read_state_file(path) merged_tasks = dict(disk.get("tasks") or {}) merged_tasks.update(state.get("tasks") or {}) disk_autos = {a.get("name"): a for a in (disk.get("automations") or []) if isinstance(a, dict)} for a in (state.get("automations") or []): if isinstance(a, dict) and a.get("name"): disk_autos[a["name"]] = a merged = {"tasks": merged_tasks, "automations": list(disk_autos.values())} # 保留内存 state 里 tasks/automations 之外的顶层字段 for k, v in state.items(): if k not in ("tasks", "automations"): merged[k] = v tmp = f"{path}.tmp.{os.getpid()}" json.dump(merged, open(tmp, "w"), indent=2, ensure_ascii=False) os.replace(tmp, path) # 让调用方内存里的 state 也跟上合并结果，避免下次保存 # 又把别的进程的任务挤掉（调用方继续用这份 state 时）。 state.clear() state.update(merged) def log_event(state_path, message): """run-due / merge / discard 等运维事件写日志，不依赖标准输出。 日志在 state 文件同目录的 host-run.log，定时任务跑完后 看这个文件就知道派发了什么、跳过了什么。""" try: log_path = os.path.join(os.path.dirname(os.path.abspath(state_path)), "host-run.log") ts = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) with open(log_path, "a", encoding="utf-8") as f: f.write(f"{ts} {message}\n") except Exception: pass def git(args, cwd): return subprocess.run(["git"] + args, cwd=cwd, capture_output=True, text=True) def make_worktree(repo, task_name): """一任务一 worktree；不是 git 仓库时退化为独立目录拷贝。""" repo = os.path.abspath(repo) base = os.path.join(HERE, "work", "host-worktrees") os.makedirs(base, exist_ok=True) path = os.path.join(base, task_name) if os.path.exists(path): return path, None if os.path.isdir(os.path.join(repo, ".git")): branch = f"host/{task_name}" r = git(["worktree", "add", path, "-b", branch, "HEAD"], repo) if r.returncode != 0: # 分支已存在时直接用它 r = git(["worktree", "add", path, branch], repo) if r.returncode != 0: raise RuntimeError(f"git worktree 失败: {r.stderr}") return path, branch # 非 git：复制一份 import shutil shutil.copytree(repo, path) return path, None def task_diff(worktree): if os.path.isdir(os.path.join(worktree, ".git")) or \ subprocess.run(["git", "-C", worktree, "rev-parse"], capture_output=True).returncode == 0: st = git(["status", "--porcelain"], worktree).stdout diff = git(["diff", "HEAD"], worktree).stdout untracked = git(["ls-files", "--others", "--exclude-standard"], worktree).stdout # 生产评审要点：新文件默认是 untracked，git diff 里看不到内容， # 只列文件名等于没评审。把未追踪文本文件的内容也带上（截断）。 contents = {} for rel in untracked.split(): p = os.path.join(worktree, rel) if rel.endswith(".pyc") or "__pycache__" in rel: continue try: with open(p, encoding="utf-8") as f: contents[rel] = f.read(8000) except Exception: contents[rel] = "" return {"status": st, "diff": diff, "untracked": untracked, "untracked_contents": contents} return {"status": "", "diff": "", "untracked": "\n".join(sorted(os.listdir(worktree)))} def dispatch(server, state, name, prompt, repo, model, effort, sandbox, approval_policy, use_worktree=True, parent_task=None): if name in state["tasks"]: raise RuntimeError(f"task 已存在: {name}，先换名或 discard 它") if repo and not os.path.isabs(repo) and not os.path.exists(repo): cand = os.path.join(HERE, repo) if os.path.exists(cand): repo = cand if not sandbox: sandbox = "workspace-write" cwd = repo branch = None if use_worktree and repo: cwd, branch = make_worktree(repo, name) thread_id = server.create_thread(cwd, model=model, sandbox=sandbox, approval_policy=approval_policy) task = {"name": name, "thread_id": thread_id, "cwd": cwd, "repo": os.path.abspath(repo) if repo else None, "branch": branch, "sandbox": sandbox, "status": "running", "prompt": prompt, "parent_task": parent_task, "created_at": time.time(), "turns": []} state["tasks"][name] = task return task def run_task_turn(server, state, task, prompt, model, effort, sandbox=None): # task 自带 sandbox 时优先用它（只读评审任务不能被全局默认覆盖）， # 否则用调用方显式传的 sandbox；两者都没有时不传 sandboxPolicy， # 保留 thread 原有设置（旧任务没有 sandbox 记录时尤其重要）。 use_sandbox = task.get("sandbox") or sandbox result = server.run_turn(task["thread_id"], prompt, model=model, effort=effort, sandbox=use_sandbox) task["turns"].append({"prompt": prompt, "result": result}) task["status"] = "completed" if result["status"] == "completed" \ else "failed" task["last_text"] = result.get("text", "") return result def _load_config_file(path): if path and os.path.exists(path): try: return json.load(open(path)) except Exception as e: raise RuntimeError(f"配置文件读失败 {path}: {e}") return {} def _resolve_path(base_dir, p): if not p: return p if os.path.isabs(p): return p return os.path.join(base_dir, p) def main(): ap = argparse.ArgumentParser() ap.add_argument("--codex-home", default=None) ap.add_argument("--codex-bin", default=None) ap.add_argument("--config", default=None, help="项目配置文件，默认 host.py 同目录的 host-config.json（若存在）") ap.add_argument("--state", default=None) ap.add_argument("--model", default=None) ap.add_argument("--effort", default=None) ap.add_argument("--sandbox", default=None) ap.add_argument("--approval-policy", default=None) ap.add_argument("--approval", default=None, choices=["auto", "deny"]) sub = ap.add_subparsers(dest="cmd", required=True) p = sub.add_parser("init") p.add_argument("--repo", default=None) p.add_argument("--force", action="store_true") sub.add_parser("config") p = sub.add_parser("dispatch") p.add_argument("--name", required=True) p.add_argument("--task", required=True) p.add_argument("--repo", default=None) p.add_argument("--no-worktree", action="store_true") sub.add_parser("list") p = sub.add_parser("status"); p.add_argument("--task", required=True) p = sub.add_parser("review"); p.add_argument("--task", required=True) p = sub.add_parser("merge"); p.add_argument("--task", required=True) p = sub.add_parser("discard"); p.add_argument("--task", required=True) p = sub.add_parser("archive"); p.add_argument("--task", required=True) p = sub.add_parser("fork"); p.add_argument("--task", required=True); p.add_argument("--name", required=True); p.add_argument("--sandbox", default=None) p = sub.add_parser("compact"); p.add_argument("--task", required=True) p = sub.add_parser("run"); p.add_argument("--task", required=True); p.add_argument("--prompt", required=True) p = sub.add_parser("delegate") p.add_argument("--parent-task", required=True); p.add_argument("--task", required=True) p.add_argument("--name", required=True); p.add_argument("--report-to-parent", action="store_true") p = sub.add_parser("goal") p.add_argument("--name", required=True); p.add_argument("--task", required=True) p.add_argument("--repo", default=None); p.add_argument("--verify", required=True) p.add_argument("--max-rounds", type=int, default=3) p = sub.add_parser("parallel"); p.add_argument("--jobs", required=True) p = sub.add_parser("automations"); p.add_argument("--add", action="store_true") p.add_argument("--name"); p.add_argument("--task"); p.add_argument("--repo") p.add_argument("--interval", type=int, default=86400) p.add_argument("--cron-line", action="store_true", help="只打印 run-due 的 crontab 行，不装机") p.add_argument("--install-cron", action="store_true", help="把 run-due 写进当前用户 crontab（幂等，已有则不重复）") p.add_argument("--cron-every-minutes", type=int, default=30) sub.add_parser("run-due") args = ap.parse_args() # 配置解析：CLI > 配置文件 > 内置默认。配置文件里绝不放密钥， # API key 只走环境变量（中转站用 XCM_API_KEY 那种），这是硬规则。 config_path = args.config or os.path.join(HERE, "host-config.json") cfg = _load_config_file(config_path) if os.path.exists(config_path) else {} cfg_dir = os.path.dirname(os.path.abspath(config_path)) if args.cmd == "init": if os.path.exists(config_path) and not args.force: print(f"配置文件已存在：{config_path}，加 --force 才会覆盖") return init_cfg = { "repo": args.repo or cfg.get("repo") or "", "model": "gpt-6.1-sol", "effort": "low", "sandbox": "workspace-write", "approval_policy": "never", "approval": "deny", "codex_home": "", "state": "host-state.json", "_note": "API key 不放这里，只走环境变量", } json.dump(init_cfg, open(config_path, "w"), indent=2, ensure_ascii=False) print(f"配置文件已写：{config_path}，请按需改 repo/model 后再用") return # 有效配置（init/config 之外的命令都用它） args.model = args.model or cfg.get("model") or "gpt-6.1-sol" args.effort = args.effort or cfg.get("effort") or "low" args.sandbox = args.sandbox or cfg.get("sandbox") args.approval_policy = args.approval_policy or cfg.get("approval_policy") or "never" args.approval = args.approval or cfg.get("approval") or "deny" args.codex_home = args.codex_home or _resolve_path(cfg_dir, cfg.get("codex_home")) args.codex_bin = args.codex_bin or _resolve_path(cfg_dir, cfg.get("codex_bin")) args.state = args.state or _resolve_path(cfg_dir, cfg.get("state")) or DEFAULT_STATE if args.cmd == "config": print(json.dumps({ "config_file": config_path if os.path.exists(config_path) else None, "model": args.model, "effort": args.effort, "sandbox": args.sandbox, "approval_policy": args.approval_policy, "approval": args.approval, "codex_home": args.codex_home, "codex_bin": args.codex_bin, "state": args.state, "repo": cfg.get("repo"), }, indent=2, ensure_ascii=False)) return # dispatch/goal/automations 的 repo 缺省用配置里的 repo if args.cmd in ("dispatch", "goal"): if not args.repo: args.repo = cfg.get("repo") if not args.repo: raise RuntimeError("缺少 --repo，也没有在 host-config.json 里配 repo") if args.cmd == "automations" and (args.cron_line or args.install_cron): # cron 行生成：只打印 / 装机，不跑模型，不花钱。 bin_path = args.codex_bin or default_codex_bin() log_path = os.path.join(os.path.dirname(os.path.abspath(args.state)), "host-run.log") line = (f"*/{args.cron_every_minutes} * * * * cd {HERE} && " f"CODEX_BIN={bin_path} python3 host.py --config {config_path} " f"run-due >> {log_path} 2>&1 # codex-host-run-due") if args.cron_line: print(line) print("注意：这行本身不生效，要装进 crontab 才生效；" "VM 会被平台回收，crontab 不一定能活过回收，" "常驻机器用 crontab，VM 上更稳的是 Muse cron 定时调 run-due。") return if args.install_cron: cur = subprocess.run(["crontab", "-l"], capture_output=True, text=True).stdout or "" if "codex-host-run-due" in cur: print("crontab 里已有 codex-host-run-due，未重复安装") return new = (cur.rstrip("\n") + "\n" + line + "\n") if cur.strip() else line + "\n" subprocess.run(["crontab", "-"], input=new, text=True, check=True) print(f"已安装 crontab：{line}") return state = load_state(args.state) srv = Server(codex_bin=args.codex_bin, codex_home=args.codex_home, approval=args.approval) try: srv.initialize() if args.cmd == "dispatch": task = dispatch(srv, state, args.name, args.task, args.repo, args.model, args.effort, args.sandbox, args.approval_policy, use_worktree=not args.no_worktree) # 生产教训：先落盘任务登记再跑 turn，中途被杀/超时后 # 任务还在 host-state 里，可用 run 续跑；否则 turn 中断 # 就等于任务丢了（2026-10-25 委派评审 300 秒超时实测）。 save_state(args.state, state) result = run_task_turn(srv, state, task, args.task, args.model, args.effort, args.sandbox) save_state(args.state, state) print(json.dumps({"task": args.name, "thread_id": task["thread_id"], "cwd": task["cwd"], "status": task["status"], "text": result.get("text", "")}, indent=2, ensure_ascii=False)) elif args.cmd == "list": for name, t in state["tasks"].items(): print(f"{name}: thread={t['thread_id'][:8]} status={t['status']} " f"cwd={t['cwd']} branch={t.get('branch')}") print(f"app-server threads: " f"{json.dumps(srv.list_threads(), ensure_ascii=False)[:500]}") elif args.cmd == "status": t = state["tasks"][args.task] print(json.dumps(t, indent=2, ensure_ascii=False)) elif args.cmd == "review": t = state["tasks"][args.task] d = task_diff(t["cwd"]) t["diff"] = d save_state(args.state, state) print(f"status:\n{d['status']}\nuntracked:\n{d['untracked']}\n" f"diff:\n{d['diff'][:8000]}") for rel, content in (d.get("untracked_contents") or {}).items(): print(f"\n--- untracked file: {rel} ---\n{content}") elif args.cmd == "merge": t = state["tasks"][args.task] if not t.get("repo") or not t.get("branch"): print("非 git worktree 任务，无法 merge，直接看 diff 后手动处理") return # 关键一步：worktree 里的改动默认是未提交的，直接 merge 分支 # 合不到任何东西。先在 worktree 里提交，再把分支合并进主仓。 git(["add", "-A"], t["cwd"]) c = git(["commit", "-m", f"host: {args.task}"], t["cwd"]) if c.returncode != 0 and "nothing to commit" not in c.stdout + c.stderr: print(f"worktree 提交失败：{c.stderr}") return r = git(["merge", "--no-ff", t["branch"], "-m", f"host: {args.task}"], t["repo"]) if r.returncode != 0: git(["merge", "--abort"], t["repo"]) print(f"合并冲突，已 abort：{r.stderr}") return git(["worktree", "remove", "--force", t["cwd"]], t["repo"]) git(["branch", "-d", t["branch"]], t["repo"]) t["status"] = "merged" save_state(args.state, state) print(f"merged {t['branch']} into {t['repo']}") elif args.cmd == "discard": t = state["tasks"][args.task] if t.get("repo") and t.get("branch"): git(["worktree", "remove", "--force", t["cwd"]], t["repo"]) git(["branch", "-D", t["branch"]], t["repo"]) t["status"] = "discarded" save_state(args.state, state) print(f"discarded {args.task}") elif args.cmd == "compact": t = state["tasks"][args.task] # 注意：compact 要求 thread 已在当前 server 里加载， # 直接对未加载的旧 thread 调会报 thread not found， # 所以先 resume 再 compact，这是协议层面的先后顺序。 srv.resume_thread(t["thread_id"]) resp = srv.request("thread/compact/start", {"threadId": t["thread_id"]}, timeout=120) print(f"compact {args.task}: {resp}") elif args.cmd == "archive": t = state["tasks"][args.task] srv.archive_thread(t["thread_id"]) t["status"] = "archived" save_state(args.state, state) print(f"archived {args.task}") elif args.cmd == "fork": t = state["tasks"][args.task] # 生产教训：fork 只分叉 thread 历史，不自动隔离工作区。 # 旧实现复用原 worktree，合并后 worktree 已删时 fork 出来的 # thread 连 cwd 都没了，只能只读。正确做法：有 repo 时给 # fork 单独建一个新 worktree，并把 cwd 传给 thread/fork。 new_cwd = t["cwd"] new_branch = t.get("branch") if t.get("repo"): old_cwd = t["cwd"] new_cwd, new_branch = make_worktree(t["repo"], args.name) if os.path.isdir(old_cwd): # 未合并任务：新 worktree 只有主仓内容，没有原任务的 # 改动。把原 worktree 的文件（不含 .git）拷过去， # fork 出来的 thread 历史和工作区才对得上。 import shutil for root, dirs, files in os.walk(old_cwd): dirs[:] = [d for d in dirs if d not in (".git", "__pycache__")] for fn in files: if fn.endswith(".pyc") or fn == ".git": # .git 在 worktree 里是一个指向 gitdir 的文件， # 拷过去会把新 worktree 的指针指回旧 worktree， # 导致 git worktree 校验失败（2026-10-25 实测）。 continue src = os.path.join(root, fn) rel = os.path.relpath(src, old_cwd) dst = os.path.join(new_cwd, rel) os.makedirs(os.path.dirname(dst), exist_ok=True) shutil.copy2(src, dst) # fork 不传 sandbox 时服务端默认 read-only（实测）， # 所以必须显式传原任务的沙箱，旧任务没存沙箱时默认 workspace-write。 fork_sandbox = args.sandbox or t.get("sandbox") or "workspace-write" new_id = srv.fork_thread(t["thread_id"], cwd=new_cwd, sandbox=fork_sandbox) nt = dict(t) nt.update({"name": args.name, "thread_id": new_id, "cwd": new_cwd, "branch": new_branch, "status": "forked", "parent_task": args.task}) state["tasks"][args.name] = nt save_state(args.state, state) print(f"forked {args.task} -> {args.name} thread={new_id} " f"cwd={new_cwd}") elif args.cmd == "run": t = state["tasks"][args.task] srv.resume_thread(t["thread_id"]) result = run_task_turn(srv, state, t, args.prompt, args.model, args.effort, args.sandbox) save_state(args.state, state) print(json.dumps(result, indent=2, ensure_ascii=False)) elif args.cmd == "delegate": parent = state["tasks"][args.parent_task] # 子 thread：cwd 与父任务相同 worktree 时只读更安全， # 这里默认沿用父 cwd + read-only，避免子任务改坏父任务的改动 sub_thread = srv.create_thread(parent["cwd"], model=args.model, sandbox="read-only", approval_policy="never") sub = {"name": args.name, "thread_id": sub_thread, "cwd": parent["cwd"], "repo": parent.get("repo"), "branch": None, "sandbox": "read-only", "status": "running", "prompt": args.task, "parent_task": args.parent_task, "created_at": time.time(), "turns": []} state["tasks"][args.name] = sub save_state(args.state, state) result = run_task_turn(srv, state, sub, args.task, args.model, args.effort, args.sandbox) save_state(args.state, state) print(json.dumps({"sub_task": args.name, "status": sub["status"], "text": result.get("text", "")}, indent=2, ensure_ascii=False)) if args.report_to_parent: srv.resume_thread(parent["thread_id"]) rep = run_task_turn(srv, state, parent, f"子任务 {args.name} 的结论如下，请汇总到主任务结论里，" f"不要重复做子任务的工作：\n{result.get('text', '')}", args.model, args.effort, args.sandbox) save_state(args.state, state) print(f"reported to parent: {rep.get('text', '')[:1000]}") elif args.cmd == "goal": task = dispatch(srv, state, args.name, args.task, args.repo, args.model, args.effort, args.sandbox, args.approval_policy) save_state(args.state, state) ok = False last_verify = "" for round_no in range(1, args.max_rounds + 1): if round_no == 1: prompt = args.task else: # 生产教训：只说“验收未通过” agent 不知道改什么， # 2026-10-25 实测 prod-retry 3 轮都原地不动。 # 必须把验收命令和失败输出喂回去，才是可用的目标循环。 prompt = (f"上一轮验收未通过，请针对性修复，不要重复已完成的部分。\n" f"目标：{args.task}\n" f"验收命令：{args.verify}\n" f"上一轮验收输出：\n{last_verify}") result = run_task_turn(srv, state, task, prompt, args.model, args.effort, args.sandbox) v = subprocess.run(args.verify, shell=True, cwd=task["cwd"], capture_output=True, text=True) last_verify = (v.stdout + v.stderr).strip()[:1000] or \ f"exit code {v.returncode}" print(f"round {round_no}: turn={result['status']} " f"verify_rc={v.returncode} verify_out={last_verify[:300]}") if v.returncode == 0: ok = True break task["status"] = "goal-met" if ok else "goal-unmet" save_state(args.state, state) print(f"goal {'met' if ok else 'unmet'} after rounds") elif args.cmd == "parallel": jobs = json.load(open(args.jobs)) # 先全部派发（建 thread + worktree），再并发跑 turn tasks = [] for job in jobs: t = dispatch(srv, state, job["name"], job["task"], job.get("repo", ""), args.model, args.effort, job.get("sandbox", args.sandbox), args.approval_policy, use_worktree=bool(job.get("repo"))) tasks.append((job, t)) save_state(args.state, state) def _run(job_task): job, t = job_task return job["name"], run_task_turn(srv, state, t, job["task"], args.model, args.effort, args.sandbox) with ThreadPoolExecutor(max_workers=len(tasks)) as ex: for name, result in ex.map(_run, tasks): print(f"{name}: {result['status']} " f"{result.get('text', '')[:300]}") save_state(args.state, state) elif args.cmd == "automations": if args.add: state["automations"].append({ "name": args.name, "task": args.task, "repo": args.repo or cfg.get("repo"), "interval": args.interval, "last_run": 0}) save_state(args.state, state) print(f"automation added: {args.name} every {args.interval}s") else: print(json.dumps(state["automations"], indent=2, ensure_ascii=False)) elif args.cmd == "run-due": now = time.time() for auto in state["automations"]: if now - auto.get("last_run", 0) < auto["interval"]: log_event(args.state, f"run-due skip {auto['name']} " f"(interval {auto['interval']}s, not due)") continue name = f"{auto['name']}-{int(now)}" task = dispatch(srv, state, name, auto["task"], auto["repo"], args.model, args.effort, args.sandbox, args.approval_policy) save_state(args.state, state) result = run_task_turn(srv, state, task, auto["task"], args.model, args.effort, args.sandbox) auto["last_run"] = now log_event(args.state, f"run-due {auto['name']} -> {name}: " f"{result['status']}") print(f"automation {auto['name']} -> {name}: " f"{result['status']} (进评审队列，见 host.py status/review)") save_state(args.state, state) finally: srv.close() if __name__ == "__main__": main()