"""FastAPI server backing the local web GUI. Wraps a single global :class:`LocalCodingAgent`, exposes JSON endpoints for chat, slash commands, and saved sessions, and serves the static SPA. """ from __future__ import annotations import asyncio from contextlib import asynccontextmanager import csv import hashlib import shlex import json import os import queue import re import shutil import signal import subprocess import threading import time import venv from dataclasses import dataclass, field, replace from pathlib import Path from threading import Lock, RLock from typing import Any, Callable from urllib import error, request from urllib.parse import quote, unquote, urlparse from uuid import uuid4 from fastapi import FastAPI, HTTPException, Request from fastapi.responses import FileResponse, JSONResponse, Response, StreamingResponse from fastapi.staticfiles import StaticFiles from pydantic import BaseModel, Field from src.agent_session import AgentMessage, AgentSessionState from src.agent_runtime import LocalCodingAgent from src.agent_slash_commands import get_slash_command_specs from src.agent_types import ( AgentPermissions, AgentRuntimeConfig, ModelConfig, ) from src.bundled_skills import ALWAYS_ENABLED_HIDDEN_SKILL_NAMES, get_bundled_skills from src.data_agent_inputs import DataAgentInputError, load_input_sources from src.jupyter_runtime import ( DEFAULT_JUPYTER_WORKSPACE_ROOT, JupyterRuntimeError, JupyterRuntimeManager, JupyterRuntimeSession, ) from src.mcp_runtime import MCPRuntime, MCPServerProfile from src.openai_compat import OpenAICompatClient, OpenAICompatError from src.personal_memory import ( SKILL_MEMORY_DIRNAME, USER_MEMORY_FILENAME, PersonalMemoryManager, ) from src.bash_bg_store import ( ACTIVE_BG_STATUSES, TERMINAL_BG_STATUSES, BashBgStore, BgTaskSpec, BgTaskStatus, ) from src.run_state_store import ACTIVE_RUN_STATUSES, RunStateStore from src.session_store import ( DEFAULT_AGENT_SESSION_DIR, StoredAgentSession, deserialize_runtime_config, load_agent_session, save_agent_session, serialize_model_config, serialize_runtime_config, ) from src.token_budget import calculate_token_budget STATIC_DIR = Path(__file__).resolve().parents[2] / 'frontend' / 'legacy-static' API_TOOL_CONTENT_MAX_CHARS = 20000 FEISHU_NPM_REGISTRY = 'https://pkgs.d.xiaomi.net/artifactory/api/npm/mi-npm/' FEISHU_MCP_PACKAGE = '@mi/feishu-mcp-pro@latest' FEISHU_MCP_SERVER_NAME = 'feishu-mcp-pro' FEISHU_DOCUMENT_SUFFIXES = {'.docx', '.md', '.txt'} FEISHU_SPREADSHEET_SUFFIXES = {'.csv', '.xlsx'} FEISHU_SUPPORTED_DOCUMENT_SUFFIXES = FEISHU_DOCUMENT_SUFFIXES | FEISHU_SPREADSHEET_SUFFIXES FEISHU_UNSUPPORTED_DOCUMENT_SUFFIXES = {'.json', '.jsonl'} FEISHU_DOC_MARKDOWN_MAX_CHARS = 100_000 FEISHU_SPREADSHEET_MAX_ROWS = 5000 FEISHU_SPREADSHEET_MAX_COLS = 200 FEISHU_SPREADSHEET_MAX_CELL_CHARS = 5000 FEISHU_SPREADSHEET_WRITE_BATCH_ROWS = 500 FEISHU_ONLINE_DOCS_FILENAME = 'online-docs.json' JUPYTER_WORKSPACE_FILENAME = 'jupyter_workspace.json' JUPYTER_WORKSPACES_REGISTRY_FILENAME = 'workspaces.json' JUPYTER_STREAM_CHUNK_BYTES = 1024 * 1024 FEISHU_REMOTE_DOC_MAX_BYTES = 50 * 1024 * 1024 FEISHU_REMOTE_SHEET_MAX_BYTES = 100 * 1024 * 1024 @dataclass class FeishuLoginProcess: process: subprocess.Popen[str] started_at: float output: list[str] = field(default_factory=list) login_url: str | None = None user_code: str | None = None _FEISHU_LOGIN_PROCESSES: dict[str, FeishuLoginProcess] = {} _FEISHU_LOGIN_LOCK = threading.Lock() _FEISHU_DOC_MAP_LOCK = threading.Lock() VALIDATED_CHAT_MODEL_PROVIDERS = { # 这些 provider 已验证可以在当前 WebUI 中使用。 # 部分模型族会在客户端按模型 id 切换到专用兼容路由。 'azure_openai', 'hunyuan', 'minimax', 'moonshot', 'ppio', 'siliconflow', 'tongyi', 'vertex_ai', 'volcengine_maas', 'wenxin', 'xiaomi', 'zhipuai', } UNSUPPORTED_TOOL_CHAT_MODEL_IDS = { # 这些模型来自 /models，但带 tools/tool_choice 的最小 Agent 请求实测失败。 'azure_openai/gpt-5.5', 'azure_openai/grok-3', 'azure_openai/grok-3-mini', 'minimax/abab6.5t-chat', 'ppio/gemini-2.0-flash-20250609', 'siliconflow/deepseek-ai/deepseek-ocr', 'siliconflow/deepseek-ai/deepseek-r1-distill-qwen-14b', 'siliconflow/deepseek-ai/deepseek-r1-distill-qwen-32b', 'siliconflow/deepseek-ai/deepseek-r1-distill-qwen-7b', 'siliconflow/qwen/qwen2.5-coder-32b-instruct', 'siliconflow/qwen/qwen2.5-vl-32b-instruct', 'siliconflow/qwen/qwen2.5-vl-72b-instruct', 'siliconflow/qwen/qwen3-235b-a22b', 'siliconflow/qwen/qwen3-235b-a22b-thinking-2507', 'siliconflow/qwen/qwen3-30b-a3b-instruct-2507', 'siliconflow/qwen/qwen3-vl-8b-instruct', 'siliconflow/qwen/qwen3-vl-8b-thinking', 'siliconflow/thudm/glm-4-9b-chat', 'siliconflow/thudm/glm-4.1v-9b-thinking', 'siliconflow/zai-org/glm-4.5', 'tongyi/deepseek-r1-distill-qwen-14b', 'tongyi/deepseek-r1-distill-qwen-32b', 'tongyi/qwen-mt-plus', 'tongyi/qwen-mt-turbo', 'tongyi/qwen-plus-0919', 'tongyi/qwen2.5-0.5b-instruct', 'tongyi/qwen2.5-1.5b-instruct', 'tongyi/qwen3-0.6b', 'tongyi/qwen3-1.7b', 'tongyi/qwen3-30b-a3b', 'tongyi/qwen3-4b', 'tongyi/qwen3-8b', 'vertex_ai/gemini-2.0-flash-001', 'vertex_ai/gemini-2.0-flash-lite-001', 'vertex_ai/gemini-2.5-computer-use-preview-10-2025', 'vertex_ai/gemini-2.5-flash', 'vertex_ai/gemini-2.5-flash-lite', 'vertex_ai/gemini-2.5-flash-lite-preview-09-2025', 'vertex_ai/gemini-2.5-pro', 'vertex_ai/gemini-3-flash-preview', 'vertex_ai/gemini-3.1-flash-lite-preview', 'vertex_ai/gemini-3.1-pro-preview', 'vertex_ai/gemini-3.1-pro-preview-pt', 'wenxin/ernie-speed-128k', 'wenxin/ernie-speed-8k', 'xiaomi/mi-brag-vl', 'xiaomi/midashenglm-7b-1021-bf16', 'xiaomi/milm2.1-13b-chat', 'xiaomi/mimo-vl-7b-rl', 'xiaomi/mimo-vl-7b-rl-0808', 'xiaomi/paddleocr-vl-0.9b', 'xiaomi/qwen-235b-a22b', 'xiaomi/qwen2.5-72b-instruct-gptq-int4', 'xiaomi/qwen2.5-vl-72b-instruct-awq', 'xiaomi/qwen25-coder-7b', 'xiaomi/qwen3-235b-a22b', 'xiaomi/qwen3-235b-a22b-instruct-2507', 'xiaomi/qwen3-32b', } # WebUI 的快速入口只展示可以通过对话输入框安全执行的 slash command。 # 这些命令仍然保留在终端 `/` 列表里，但不适合作为 WebUI 快捷项。 WEBUI_HIDDEN_SLASH_COMMANDS = { 'exit', 'quit', 'feedback', 'bug', 'upgrade', 'stickers', 'chrome', 'install-github-app', 'install-slack-app', 'privacy-settings', 'mobile', 'ios', 'android', 'desktop', 'app', 'vim', 'theme', } WEBUI_SLASH_COMMAND_DESCRIPTIONS_ZH = { 'help': '查看内置 slash command 帮助。', 'context': '查看当前会话上下文用量估算。', 'context-raw': '查看原始环境、用户上下文和系统上下文快照。', 'token-budget': '查看当前 token 预算窗口、保留量和提示词长度限制。', 'mcp': '查看本地 MCP 清单和资源数量。', 'search': '查看搜索运行状态、切换搜索提供方，或执行一次网页搜索。', 'remote': '查看远程运行状态，或激活远程目标/配置。', 'worktree': '查看受管 git worktree 状态，或进入/退出当前 worktree 会话。', 'account': '查看账号运行状态或已配置账号档案。', 'ask': '查看 ask-user 运行状态或历史记录。', 'login': '激活本地账号档案或临时身份。', 'logout': '清除当前本地账号会话。', 'config': '查看配置状态、有效配置、配置来源或某个配置值。', 'lsp': '查看 LSP 状态，或执行符号、定义、引用、悬停、调用层级和诊断查询。', 'remotes': '列出本地远程配置。', 'ssh': '激活 SSH 远程目标/配置。', 'teleport': '激活 Teleport 远程目标/配置。', 'direct-connect': '激活直连远程目标/配置。', 'deep-link': '激活 deep-link 远程目标/配置。', 'disconnect': '断开当前远程运行目标。', 'resources': '列出本地 MCP 资源，可按关键词过滤。', 'resource': '按 URI 渲染一个本地 MCP 资源。', 'tasks': '查看本地运行时任务列表，可按状态过滤。', 'workflows': '列出从工作流清单发现的本地 workflows。', 'workflow': '查看或运行一个本地 workflow。', 'triggers': '列出从清单发现的本地远程触发器。', 'trigger': '查看或运行一个远程触发器。', 'teams': '列出本地协作团队配置。', 'team': '查看一个本地协作团队。', 'messages': '查看所有团队或某个团队的协作消息。', 'task-next': '查看本地任务列表中的下一个可执行任务。', 'plan': '查看当前本地运行计划。', 'task': '按 id 查看一个本地任务。', 'prompt': '渲染当前生效的系统提示词。', 'permissions': '查看当前工具权限模式。', 'hooks': '查看已发现的本地 hook 和 policy 清单。', 'trust': '查看 workspace 信任模式、受管设置和安全环境变量。', 'model': '查看或修改当前 Agent 实例使用的模型。', 'tools': '列出已注册工具，以及当前权限是否允许调用。', 'agents': '列出本地 Agent 配置，或查看某个 Agent 定义。', 'memory': '查看当前加载的 CLAUDE.md 记忆包和已发现文件。', 'status': '查看当前运行时和会话状态摘要。', 'clear': '清理当前进程中的临时运行状态。', 'compact': '总结并压缩当前对话，释放上下文空间。', 'cost': '查看当前会话的总耗时和费用估算。', 'diff': '查看当前工作区未提交改动。', 'files': '列出当前会话上下文中已加载的文件。', 'copy': '把上一条助手回复写入临时文件。', 'export': '导出当前对话为文本文件。', 'stats': '查看会话使用统计。', 'tag': '给当前会话添加或移除可搜索标签。', 'rename': '重命名当前对话。', 'branch': '基于当前对话创建一个分支/副本。', 'effort': '查看或设置模型推理强度。', 'doctor': '诊断并验证 claw-code 安装和配置。', 'commit': '创建 git commit。', 'pr-comments': '读取 GitHub PR 评论。', 'resume': '恢复之前的对话。', 'add-dir': '添加新的工作目录。', 'skills': '列出可用 Skills。', 'fast': '切换 fast mode。', 'rewind': '把对话恢复到之前的 checkpoint。', 'output-style': '已废弃：请使用 /config 修改输出风格。', 'release-notes': '查看本地 release notes 或变更日志链接。', 'extra-usage': '查看 extra-usage 配置链接。', 'passes': '查看 Claude Code guest passes 信息。', 'rate-limit-options': '查看账号触发限流时的可选处理方式。', 'reload-plugins': '重新加载本地插件清单并报告数量。', 'voice': '切换当前 workspace 的 voice mode 设置。', 'sandbox-toggle': '查看 sandbox 状态，或排除一条命令模式。', 'keybindings': '打印或创建本地快捷键配置文件。', 'btw': '向模型快速提一个旁路问题，不修改当前任务状态。', 'version': '打印当前 Agent 版本。', 'init': '初始化 CLAUDE.md 代码库说明文件。', 'ide': '查看检测到的 IDE/终端集成状态。', 'plugin': '列出已安装插件，或查看插件子命令用法。', 'remote-env': '列出远程环境，或设置默认配置。', 'bridge': '查看 remote-control bridge 状态。', 'remote-setup': '检查 Claude Code on the web 的准备状态。', } # --------------------------------------------------------------------------- # Agent state holder # --------------------------------------------------------------------------- @dataclass(frozen=True) class AgentInstanceConfig: cwd: Path model: str base_url: str api_key: str timeout_seconds: float allow_shell: bool allow_write: bool SKILL_SETTINGS_FILENAME = 'skill_settings.json' class RunProcessRegistry: def __init__(self) -> None: self._lock = Lock() self._processes: set[Any] = set() def add(self, process: Any) -> None: with self._lock: self._processes.add(process) def discard(self, process: Any) -> None: with self._lock: self._processes.discard(process) def kill_all(self) -> None: with self._lock: processes = list(self._processes) for process in processes: try: if process.poll() is None: self._signal_process_tree(process, signal.SIGTERM) except OSError: continue deadline = time.monotonic() + 1.0 for process in processes: try: remaining = max(0.0, deadline - time.monotonic()) process.wait(timeout=remaining) except Exception: self._signal_process_tree(process, signal.SIGKILL) try: process.wait(timeout=1.0) except Exception: pass @staticmethod def _signal_process_tree(process: Any, sig: signal.Signals) -> None: try: if os.name == 'posix': pgid = os.getpgid(process.pid) if pgid == process.pid: os.killpg(pgid, sig) return if sig == signal.SIGTERM: process.terminate() else: process.kill() except ProcessLookupError: return except OSError: try: if sig == signal.SIGTERM: process.terminate() else: process.kill() except OSError: pass @dataclass class RunRecord: run_id: str account_key: str session_id: str status: str started_at: float updated_at: float cancel_event: threading.Event process_registry: RunProcessRegistry pending_prompt: str = '' current_stage: str = '' error: str = '' events: list[dict[str, Any]] = field(default_factory=list) class RunManager: def __init__(self) -> None: self._lock = RLock() self._runs: dict[str, RunRecord] = {} self._latest_by_session: dict[tuple[str, str], str] = {} def start(self, account_key: str, session_id: str, pending_prompt: str = '') -> RunRecord: now = time.time() record = RunRecord( run_id=uuid4().hex, account_key=account_key, session_id=session_id, status='queued', started_at=now, updated_at=now, cancel_event=threading.Event(), process_registry=RunProcessRegistry(), pending_prompt=pending_prompt, ) with self._lock: self._runs[record.run_id] = record self._latest_by_session[(account_key, session_id)] = record.run_id return record def update(self, run_id: str, *, status: str | None = None, stage: str | None = None, error: str | None = None) -> None: with self._lock: record = self._runs.get(run_id) if record is None: return if status is not None: record.status = status if stage is not None: record.current_stage = stage if error is not None: record.error = error record.updated_at = time.time() def record_event(self, run_id: str, event: dict[str, object]) -> dict[str, Any] | None: normalized = _normalize_run_event(event) if normalized is None: return None normalized['recorded_at'] = time.time() with self._lock: record = self._runs.get(run_id) if record is None: return normalized record.events.append(normalized) if len(record.events) > 160: del record.events[: len(record.events) - 160] record.updated_at = time.time() return normalized def finish(self, run_id: str, status: str) -> None: self.update(run_id, status=status) def cancel_latest(self, account_key: str, session_id: str) -> bool: with self._lock: record = self._select_session_record(account_key, session_id) return self.cancel_record(record) def cancel_run(self, run_id: str) -> bool: with self._lock: record = self._runs.get(run_id) return self.cancel_record(record) def cancel_session(self, account_key: str, session_id: str) -> list[str]: """Cancel every non-terminal run for a session. The UI can hold a stale run_id after refresh/replay while a newer request is queued behind the same session lock. Cancelling only the stale id leaves the queued run alive and makes the composer look impossible to stop. Session-level cancel is the user-facing intent. """ with self._lock: records = [ record for record in self._runs.values() if record.account_key == account_key and record.session_id == session_id and record.status not in {'completed', 'failed', 'cancelled'} ] cancelled: list[str] = [] for record in records: if self.cancel_record(record): cancelled.append(record.run_id) return cancelled def cancel_record(self, record: RunRecord | None) -> bool: if record is None or record.status in {'completed', 'failed', 'cancelled'}: return False record.cancel_event.set() record.process_registry.kill_all() self.update(record.run_id, status='cancelled', stage='用户已取消') return True def snapshot_latest(self, account_key: str, session_id: str) -> dict[str, Any] | None: with self._lock: record = self._select_session_record(account_key, session_id) if record is None: return None return { 'run_id': record.run_id, 'session_id': record.session_id, 'status': record.status, 'current_stage': record.current_stage, 'started_at': record.started_at, 'updated_at': record.updated_at, 'elapsed_ms': max(0, int((time.time() - record.started_at) * 1000)), 'pending_prompt': record.pending_prompt, 'error': record.error, 'events': [dict(event) for event in record.events], } def _select_session_record( self, account_key: str, session_id: str, ) -> RunRecord | None: records = [ record for record in self._runs.values() if record.account_key == account_key and record.session_id == session_id ] running = [ record for record in records if record.status == 'running' ] if running: return max(running, key=lambda item: item.updated_at) queued = [ record for record in records if record.status == 'queued' ] if queued: return max(queued, key=lambda item: item.started_at) run_id = self._latest_by_session.get((account_key, session_id)) return self._runs.get(run_id or '') class AgentState: """Holds account-scoped agent instances, config, and execution locks.""" def __init__( self, *, cwd: Path, model: str, base_url: str, api_key: str, timeout_seconds: float, allow_shell: bool, allow_write: bool, session_directory: Path, ) -> None: self.session_directory = session_directory self._lock = RLock() self._agents: dict[str, LocalCodingAgent] = {} self._run_locks: dict[str, Lock] = {} self._account_configs: dict[str, AgentInstanceConfig] = {} self.run_manager = RunManager() self.run_state_store = RunStateStore(self.session_directory.parent / 'run_state.db') self.bash_bg_store = BashBgStore(self.session_directory.parent / 'bash_bg.db') self.event_loop: 'asyncio.AbstractEventLoop | None' = None self.jupyter_runtime_manager = JupyterRuntimeManager() self._default_config = AgentInstanceConfig( cwd=cwd.resolve(), model=model, base_url=base_url, api_key=api_key, timeout_seconds=timeout_seconds, allow_shell=allow_shell, allow_write=allow_write, ) self.memory_manager = PersonalMemoryManager( self.session_directory.parent / 'accounts', self.model_config_for, ) self.memory_manager.start() @property def cwd(self) -> Path: return self._default_config.cwd @property def model(self) -> str: return self._default_config.model @property def base_url(self) -> str: return self._default_config.base_url @property def api_key(self) -> str: return self._default_config.api_key @property def allow_shell(self) -> bool: return self._default_config.allow_shell @property def allow_write(self) -> bool: return self._default_config.allow_write def _account_key(self, account_id: str | None) -> str: return _safe_account_id(account_id) if account_id else '__default__' def _session_key(self, account_id: str | None, session_id: str | None = None) -> str: session_part = _safe_session_id(session_id) or '__shared__' return f'{self._account_key(account_id)}:{session_part}' def _config_for(self, account_id: str | None) -> AgentInstanceConfig: if not account_id: return self._default_config key = self._account_key(account_id) config = self._account_configs.get(key) if config is None: config = replace(self._default_config) self._account_configs[key] = config return config def config_for(self, account_id: str | None) -> AgentInstanceConfig: with self._lock: return self._config_for(account_id) def model_config_for(self, account_id: str | None) -> ModelConfig: """给后台辅助任务使用的模型配置。 AgentInstanceConfig 还包含 cwd、权限等运行态字段；调用模型时需要的是 OpenAICompatClient 支持的 ModelConfig，否则后台记忆整理会缺少 temperature 等字段。 """ config = self.config_for(account_id) return ModelConfig( model=config.model, base_url=config.base_url, api_key=config.api_key, timeout_seconds=config.timeout_seconds, ) def _set_config_for(self, account_id: str | None, config: AgentInstanceConfig) -> None: if not account_id: self._default_config = config return self._account_configs[self._account_key(account_id)] = config def _account_base(self, account_id: str | None) -> Path: if not account_id: return self.session_directory.parent safe_id = _safe_account_id(account_id) return (self.session_directory.parent / 'accounts' / safe_id).resolve() def account_paths(self, account_id: str | None) -> dict[str, Path]: if not account_id: base = self.session_directory.parent return { 'base': base, 'sessions': self.session_directory, 'scratchpad': self.session_directory, 'uploads': self.session_directory, 'outputs': self.session_directory, 'python_env': base / 'python' / '.venv', 'memory': base / 'memory', } base = self._account_base(account_id) return { 'base': base, 'sessions': base / 'sessions', 'scratchpad': base / 'sessions', 'uploads': base / 'sessions', 'outputs': base / 'sessions', 'python_env': base / 'python' / '.venv', 'memory': base / 'memory', } def _skill_settings_path(self, account_id: str | None) -> Path: return self.account_paths(account_id)['base'] / SKILL_SETTINGS_FILENAME def _load_disabled_skill_names(self, account_id: str | None) -> set[str]: path = self._skill_settings_path(account_id) try: payload = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return set() disabled = payload.get('disabled_skill_names') if not isinstance(disabled, list): return set() return { str(name).strip().lower() for name in disabled if str(name).strip() } def _save_disabled_skill_names( self, account_id: str | None, disabled_skill_names: set[str], ) -> None: path = self._skill_settings_path(account_id) path.parent.mkdir(parents=True, exist_ok=True) payload = { 'disabled_skill_names': sorted( name for name in disabled_skill_names if name not in ALWAYS_ENABLED_HIDDEN_SKILL_NAMES ) } path.write_text( json.dumps(payload, ensure_ascii=False, indent=2, sort_keys=True) + '\n', encoding='utf-8', ) def enabled_skill_names(self, account_id: str | None) -> tuple[str, ...]: config = self._config_for(account_id) disabled = self._load_disabled_skill_names(account_id) enabled: list[str] = [] for skill in get_bundled_skills(config.cwd): if not skill.user_invocable: continue lowered = skill.name.lower() if ( lowered in ALWAYS_ENABLED_HIDDEN_SKILL_NAMES or lowered not in disabled ): enabled.append(skill.name) return tuple(enabled) def set_skill_enabled( self, account_id: str | None, skill_name: str, enabled: bool, ) -> None: with self._lock: skill_name = skill_name.strip() if not skill_name: raise ValueError('skill name must not be empty') lowered = skill_name.lower() if lowered in ALWAYS_ENABLED_HIDDEN_SKILL_NAMES: raise ValueError( f'skill is always enabled and cannot be changed: {skill_name}' ) config = self._config_for(account_id) if not any( skill.name.lower() == lowered and skill.user_invocable for skill in get_bundled_skills(config.cwd) ): raise ValueError(f'unknown skill: {skill_name}') disabled = self._load_disabled_skill_names(account_id) if enabled: disabled.discard(lowered) else: disabled.add(lowered) self._save_disabled_skill_names(account_id, disabled) account_prefix = f'{self._account_key(account_id)}:' for key in list(self._agents): if key.startswith(account_prefix): self._agents.pop(key, None) def set_all_skills_enabled( self, account_id: str | None, enabled: bool, ) -> None: with self._lock: config = self._config_for(account_id) configurable_skill_names = { skill.name.lower() for skill in get_bundled_skills(config.cwd) if skill.user_invocable and skill.name.lower() not in ALWAYS_ENABLED_HIDDEN_SKILL_NAMES } disabled = self._load_disabled_skill_names(account_id) if enabled: disabled.difference_update(configurable_skill_names) else: disabled.update(configurable_skill_names) self._save_disabled_skill_names(account_id, disabled) self._clear_agents_for_account(account_id) def _clear_agents_for_account(self, account_id: str | None) -> None: account_prefix = f'{self._account_key(account_id)}:' for key in list(self._agents): if key.startswith(account_prefix): self._agents.pop(key, None) def sync_skills_from_git(self, account_id: str | None) -> dict[str, Any]: """Pull the current git branch and refresh account-scoped skill state.""" with self._lock: config = self._config_for(account_id) repo = config.cwd if not (repo / '.git').exists(): raise ValueError(f'当前工作目录不是 git 仓库：{repo}') def run_git(args: list[str], *, timeout: int = 60) -> subprocess.CompletedProcess[str]: env = { **os.environ, 'GIT_TERMINAL_PROMPT': '0', } proc = subprocess.run( ['git', *args], cwd=repo, capture_output=True, text=True, timeout=timeout, env=env, ) if proc.returncode != 0: detail = (proc.stderr or proc.stdout or '').strip() raise ValueError(f'git {" ".join(args)} 失败：{detail}') return proc dirty = run_git(['status', '--porcelain', '--untracked-files=no']).stdout.strip() if dirty: raise ValueError('工作区存在未提交的 tracked 改动，已停止同步 Skill。') branch = run_git(['branch', '--show-current']).stdout.strip() or 'main' before = run_git(['rev-parse', 'HEAD']).stdout.strip() run_git(['fetch', 'origin'], timeout=120) pull = run_git(['pull', '--ff-only', 'origin', branch], timeout=120) after = run_git(['rev-parse', 'HEAD']).stdout.strip() changed_files: list[str] = [] if before != after: changed_files = [ line.strip() for line in run_git( ['diff', '--name-only', f'{before}..{after}'], timeout=60, ).stdout.splitlines() if line.strip() ] self._clear_agents_for_account(account_id) skills = get_bundled_skills(config.cwd) return { 'branch': branch, 'before': before, 'after': after, 'updated': before != after, 'changed_files': changed_files, 'skill_count': sum(1 for skill in skills if skill.user_invocable), 'message': pull.stdout.strip() or pull.stderr.strip(), } def _build_agent(self, account_id: str | None = None) -> LocalCodingAgent: config = self._config_for(account_id) paths = self.account_paths(account_id) for directory in paths.values(): if directory.name != '.venv': directory.mkdir(parents=True, exist_ok=True) self._ensure_python_env(paths['python_env']) permissions = AgentPermissions( allow_file_write=config.allow_write, allow_shell_commands=config.allow_shell, ) runtime_config = AgentRuntimeConfig( cwd=config.cwd, permissions=permissions, stream_model_responses=True, session_directory=paths['sessions'], scratchpad_root=paths['scratchpad'], python_env_dir=paths['python_env'], enabled_skill_names=self.enabled_skill_names(account_id), auto_compact_threshold_tokens=180_000, ) model_config = ModelConfig( model=config.model, base_url=config.base_url, api_key=config.api_key, timeout_seconds=config.timeout_seconds, ) return LocalCodingAgent( model_config=model_config, runtime_config=runtime_config, ) def agent_for( self, account_id: str | None = None, session_id: str | None = None, ) -> LocalCodingAgent: with self._lock: key = self._session_key(account_id, session_id) agent = self._agents.get(key) if agent is None: agent = self._build_agent(account_id) self._agents[key] = agent return agent def run_lock_for( self, account_id: str | None = None, session_id: str | None = None, ) -> Lock: key = self._session_key(account_id, session_id) with self._lock: lock = self._run_locks.get(key) if lock is None: lock = Lock() self._run_locks[key] = lock return lock def update( self, *, model: str | None = None, base_url: str | None = None, api_key: str | None = None, cwd: str | None = None, allow_shell: bool | None = None, allow_write: bool | None = None, account_id: str | None = None, ) -> None: with self._lock: config = self._config_for(account_id) if model is not None: normalized_model = _normalize_model_id(model) if normalized_model is None: raise ValueError('model must not be empty') config = replace(config, model=normalized_model) if base_url is not None: config = replace(config, base_url=base_url) if api_key is not None: config = replace(config, api_key=api_key) if cwd is not None: resolved = Path(cwd).expanduser().resolve() if not resolved.is_dir(): raise ValueError(f'cwd does not exist: {resolved}') config = replace(config, cwd=resolved) if allow_shell is not None: config = replace(config, allow_shell=allow_shell) if allow_write is not None: config = replace(config, allow_write=allow_write) self._set_config_for(account_id, config) account_prefix = f'{self._account_key(account_id)}:' for key in list(self._agents): if key.startswith(account_prefix): self._agents.pop(key, None) def snapshot(self, account_id: str | None = None) -> dict[str, Any]: with self._lock: config = self._config_for(account_id) agent = self.agent_for(account_id) paths = self.account_paths(account_id) return { 'model': config.model, 'base_url': config.base_url, 'cwd': str(config.cwd), 'account_id': _safe_account_id(account_id) if account_id else None, 'account_directory': str(paths['base']), 'session_directory': str(paths['sessions']), 'upload_directory': str(paths['uploads']), 'output_directory': str(paths['outputs']), 'python_env_directory': str(paths['python_env']), 'allow_shell': config.allow_shell, 'allow_write': config.allow_write, 'active_session_id': agent.active_session_id, } def lock(self) -> Lock: return self._lock def _ensure_python_env(self, env_dir: Path) -> None: python_bin = env_dir / 'bin' / 'python' pip_bin = env_dir / 'bin' / 'pip' if python_bin.exists() and pip_bin.exists(): return env_dir.parent.mkdir(parents=True, exist_ok=True) venv.EnvBuilder(with_pip=True, symlinks=False).create(env_dir) # --------------------------------------------------------------------------- # Request models # --------------------------------------------------------------------------- class ChatRequest(BaseModel): prompt: str = Field(min_length=1) runtime_context: str | None = None resume_session_id: str | None = None account_id: str | None = None session_id: str | None = None class RunCancelRequest(BaseModel): session_id: str = Field(min_length=1) account_id: str | None = None run_id: str | None = None class StateUpdate(BaseModel): model: str | None = None base_url: str | None = None api_key: str | None = None cwd: str | None = None allow_shell: bool | None = None allow_write: bool | None = None account_id: str | None = None class ModelListRequest(BaseModel): base_url: str | None = None api_key: str | None = None account_id: str | None = None class JupyterWorkspaceBindRequest(BaseModel): session_id: str = Field(min_length=1) base_url: str = Field(min_length=1) password: str = Field(min_length=1) workspace_root: str = DEFAULT_JUPYTER_WORKSPACE_ROOT account_id: str | None = None class SavedJupyterWorkspaceBindRequest(BaseModel): session_id: str = Field(min_length=1) account_id: str | None = None class SkillPreferenceUpdate(BaseModel): skill: str | None = None enabled: bool apply_all: bool = False account_id: str | None = None class SkillSyncRequest(BaseModel): account_id: str | None = None class SessionUpdate(BaseModel): title: str | None = Field(default=None, max_length=80) is_training: bool | None = None class FeishuAccountRequest(BaseModel): account_id: str | None = None class FeishuOnlineDocRequest(BaseModel): path: str = Field(min_length=1) title: str | None = None folder_token: str | None = None account_id: str | None = None class AdminLoginRequest(BaseModel): username: str = Field(min_length=1) password: str = Field(min_length=1) class AdminAccountCreateRequest(BaseModel): account_id: str = Field(min_length=2) class MemoryUpdateRequest(BaseModel): account_id: str = Field(min_length=1) content: str def _append_runtime_context(current: str | None, addition: str) -> str: parts = [part.strip() for part in (current, addition) if part and part.strip()] return '\n\n'.join(parts) # --------------------------------------------------------------------------- # App factory # --------------------------------------------------------------------------- def create_app(state: AgentState) -> FastAPI: @asynccontextmanager async def lifespan(app: FastAPI): # noqa: ARG001 state.event_loop = asyncio.get_running_loop() _bash_bg_manager.set_chat_runner(_run_chat_payload) await _bash_bg_manager.recover_from_store(state) scanner_task = asyncio.create_task(_watcher_scanner_loop(state)) try: yield finally: scanner_task.cancel() _watcher_manager.cancel_all() _bash_bg_manager.cancel_all() state.event_loop = None app = FastAPI(title='Claw Code GUI', version='1.0', lifespan=lifespan) # ------------- static + index ------------------------------------------ app.mount( '/static', StaticFiles(directory=str(STATIC_DIR)), name='static', ) @app.get('/', include_in_schema=False) async def root() -> FileResponse: return FileResponse(STATIC_DIR / 'index.html') # ------------- info ------------------------------------------------------ @app.get('/api/state') async def get_state(account_id: str | None = None) -> dict[str, Any]: return state.snapshot(account_id) @app.post('/api/state') async def post_state(payload: StateUpdate) -> dict[str, Any]: try: state.update(**payload.model_dump(exclude_none=True)) except ValueError as exc: raise HTTPException(status_code=400, detail=str(exc)) return state.snapshot(payload.account_id) @app.get('/api/jupyter/session') async def get_jupyter_session( session_id: str, account_id: str | None = None, ) -> dict[str, Any]: safe_session_id = _safe_session_id(session_id) if not safe_session_id: raise HTTPException(status_code=400, detail='session_id is required') account_key = state._account_key(account_id) runtime = _jupyter_runtime_for_session( state, account_id, safe_session_id, ) if runtime is None: return { 'connected': False, 'account_id': account_key, 'session_id': safe_session_id, } return runtime.to_dict() @app.get('/api/jupyter/files') async def list_jupyter_files( session_id: str, account_id: str | None = None, ) -> dict[str, Any]: safe_session_id = _safe_session_id(session_id) if not safe_session_id: raise HTTPException(status_code=400, detail='session_id is required') runtime = _jupyter_runtime_for_session(state, account_id, safe_session_id) if runtime is None: return { 'connected': False, 'session_id': safe_session_id, 'input': [], 'output': [], } try: return { 'connected': True, 'session_id': safe_session_id, 'workspace_cwd': runtime.binding.workspace_cwd, 'input': runtime.list_files('input', kind='input'), 'output': runtime.list_files('output', kind='output'), } except JupyterRuntimeError as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.get('/api/jupyter/file') async def download_jupyter_file( session_id: str, path: str, account_id: str | None = None, ) -> Response: safe_session_id = _safe_session_id(session_id) if not safe_session_id: raise HTTPException(status_code=400, detail='session_id is required') runtime = _jupyter_runtime_for_session(state, account_id, safe_session_id) if runtime is None: raise HTTPException(status_code=404, detail='Jupyter runtime is not connected') try: response, filename = runtime.open_file_stream(path) except JupyterRuntimeError as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc def stream_chunks() -> Any: try: for chunk in response.iter_content(chunk_size=JUPYTER_STREAM_CHUNK_BYTES): if chunk: yield chunk finally: response.close() return StreamingResponse( stream_chunks(), media_type=_content_type_for_filename(filename), headers={ 'content-disposition': ( f'attachment; filename="{_ascii_download_filename(filename)}"; ' f"filename*=UTF-8''{_url_quote_filename(filename)}" ) }, ) @app.post('/api/jupyter/bind-session') async def bind_jupyter_session( payload: JupyterWorkspaceBindRequest, ) -> dict[str, Any]: safe_session_id = _safe_session_id(payload.session_id) if not safe_session_id: raise HTTPException(status_code=400, detail='session_id is required') account_key = state._account_key(payload.account_id) try: runtime = await asyncio.to_thread( state.jupyter_runtime_manager.bind_session, account_id=account_key, session_id=safe_session_id, base_url=payload.base_url, password=payload.password, workspace_root=payload.workspace_root, project_root=state.config_for(payload.account_id).cwd, ) except JupyterRuntimeError as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc except Exception as exc: raise HTTPException( status_code=502, detail=f'Unable to connect Jupyter runtime: {exc}', ) from exc _save_jupyter_runtime_binding( state.account_paths(payload.account_id)['sessions'], runtime, ) _register_jupyter_workspace( state, payload.account_id, base_url=payload.base_url, workspace_root=payload.workspace_root, password=payload.password, ) return runtime.to_dict() @app.get('/api/jupyter/workspaces') async def list_jupyter_workspaces( account_id: str | None = None, ) -> list[dict[str, Any]]: entries = _load_jupyter_workspaces_registry(state, account_id) return [_public_jupyter_workspace_entry(e) for e in entries] @app.post('/api/jupyter/workspaces/{workspace_id}/bind') async def bind_saved_jupyter_workspace( workspace_id: str, payload: SavedJupyterWorkspaceBindRequest, ) -> dict[str, Any]: safe_session_id = _safe_session_id(payload.session_id) if not safe_session_id: raise HTTPException(status_code=400, detail='session_id is required') entries = _load_jupyter_workspaces_registry(state, payload.account_id) match = next((e for e in entries if e.get('id') == workspace_id), None) if match is None: raise HTTPException(status_code=404, detail='workspace not found') password = match.get('password') base_url = match.get('base_url') workspace_root = match.get('workspace_root') or DEFAULT_JUPYTER_WORKSPACE_ROOT if not (isinstance(password, str) and isinstance(base_url, str)): raise HTTPException(status_code=400, detail='workspace entry is incomplete') account_key = state._account_key(payload.account_id) try: runtime = await asyncio.to_thread( state.jupyter_runtime_manager.bind_session, account_id=account_key, session_id=safe_session_id, base_url=base_url, password=password, workspace_root=workspace_root, project_root=state.config_for(payload.account_id).cwd, ) except JupyterRuntimeError as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc except Exception as exc: raise HTTPException( status_code=502, detail=f'Unable to connect Jupyter runtime: {exc}', ) from exc _save_jupyter_runtime_binding( state.account_paths(payload.account_id)['sessions'], runtime, ) _register_jupyter_workspace( state, payload.account_id, base_url=base_url, workspace_root=workspace_root, password=password, label=match.get('label') if isinstance(match.get('label'), str) else None, ) return runtime.to_dict() @app.delete('/api/jupyter/workspaces/{workspace_id}') async def delete_saved_jupyter_workspace( workspace_id: str, account_id: str | None = None, ) -> dict[str, Any]: entries = _load_jupyter_workspaces_registry(state, account_id) kept = [e for e in entries if e.get('id') != workspace_id] if len(kept) == len(entries): raise HTTPException(status_code=404, detail='workspace not found') _save_jupyter_workspaces_registry(state, account_id, kept) return {'ok': True} @app.get('/api/slash-commands') async def list_slash_commands() -> list[dict[str, Any]]: commands: list[dict[str, Any]] = [] for spec in get_slash_command_specs(): if any(name in WEBUI_HIDDEN_SLASH_COMMANDS for name in spec.names): continue commands.append( { 'names': list(spec.names), 'primary': spec.names[0], 'description': WEBUI_SLASH_COMMAND_DESCRIPTIONS_ZH.get( spec.names[0], spec.description, ), 'source': 'slash_command', 'webui_supported': True, } ) return commands @app.get('/api/skills') async def list_skills(account_id: str | None = None) -> list[dict[str, Any]]: config = state.config_for(account_id) disabled = state._load_disabled_skill_names(account_id) return [ { 'name': skill.name, 'description': skill.description, 'when_to_use': skill.when_to_use, 'aliases': list(skill.aliases), 'allowed_tools': list(skill.allowed_tools), 'enabled': skill.name.lower() not in disabled, 'configurable': skill.name.lower() not in ALWAYS_ENABLED_HIDDEN_SKILL_NAMES, } for skill in get_bundled_skills(config.cwd) if skill.user_invocable and skill.name.lower() not in ALWAYS_ENABLED_HIDDEN_SKILL_NAMES ] @app.patch('/api/skills') async def update_skill_preference(payload: SkillPreferenceUpdate) -> list[dict[str, Any]]: try: if payload.apply_all: state.set_all_skills_enabled(payload.account_id, payload.enabled) elif payload.skill: state.set_skill_enabled(payload.account_id, payload.skill, payload.enabled) else: raise ValueError('skill must be provided unless apply_all is true') except ValueError as exc: raise HTTPException(status_code=400, detail=str(exc)) return await list_skills(payload.account_id) @app.post('/api/skills/sync') async def sync_skills(payload: SkillSyncRequest) -> dict[str, Any]: try: result = await asyncio.to_thread( state.sync_skills_from_git, payload.account_id, ) except (ValueError, subprocess.TimeoutExpired) as exc: raise HTTPException(status_code=400, detail=str(exc)) result['skills'] = await list_skills(payload.account_id) account_key = state._account_key(payload.account_id) config = state.config_for(payload.account_id) result['remote_runtime_sync'] = await asyncio.to_thread( state.jupyter_runtime_manager.sync_account, account_key, config.cwd, ) return result # ------------- memory ---------------------------------------------------- @app.get('/api/memory/user') async def get_user_memory(account_id: str) -> dict[str, Any]: try: return state.memory_manager.list_user_memory(_safe_account_id(account_id)) except Exception as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.put('/api/memory/user') async def put_user_memory(payload: MemoryUpdateRequest) -> dict[str, Any]: try: return state.memory_manager.update_user_memory( _safe_account_id(payload.account_id), payload.content, ) except Exception as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.get('/api/memory/skills') async def get_skill_memories(account_id: str) -> list[dict[str, Any]]: try: return state.memory_manager.list_skill_memories(_safe_account_id(account_id)) except Exception as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.get('/api/memory/events') async def get_memory_events(account_id: str) -> dict[str, Any]: try: safe_account = _safe_account_id(account_id) return { 'queue': state.memory_manager.queue_snapshot(safe_account), 'events': state.memory_manager.list_recent_events(safe_account, limit=20), } except Exception as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.get('/api/memory/skills/{skill_name}') async def get_skill_memory(skill_name: str, account_id: str) -> dict[str, Any]: try: return state.memory_manager.read_skill_memory( _safe_account_id(account_id), skill_name, ) except Exception as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.put('/api/memory/skills/{skill_name}') async def put_skill_memory( skill_name: str, payload: MemoryUpdateRequest, ) -> dict[str, Any]: try: return state.memory_manager.update_skill_memory( _safe_account_id(payload.account_id), skill_name, payload.content, ) except Exception as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.delete('/api/memory/skills/{skill_name}') async def delete_skill_memory(skill_name: str, account_id: str) -> dict[str, Any]: try: return state.memory_manager.delete_skill_memory( _safe_account_id(account_id), skill_name, ) except Exception as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc @app.get('/api/memory/queue') async def get_memory_queue(account_id: str | None = None) -> dict[str, Any]: return state.memory_manager.queue_snapshot( _safe_account_id(account_id) if account_id else None ) @app.get('/api/memory/events') async def get_memory_events( account_id: str, limit: int = 50, ) -> list[dict[str, Any]]: return state.memory_manager.list_recent_events( _safe_account_id(account_id), limit=limit, ) # ------------- admin ----------------------------------------------------- @app.post('/api/admin/login') async def admin_login(payload: AdminLoginRequest) -> dict[str, Any]: if payload.username == 'admin' and payload.password == 'admin': return {'token': _admin_token(), 'username': 'admin'} raise HTTPException(status_code=401, detail='管理员账号或密码错误') @app.get('/api/admin/summary') async def admin_summary(token: str) -> dict[str, Any]: _require_admin_token(token) return _build_admin_summary(state) @app.get('/api/admin/accounts') async def admin_accounts(token: str) -> list[dict[str, Any]]: _require_admin_token(token) return _list_admin_accounts(state) @app.post('/api/admin/accounts') async def admin_create_account( token: str, payload: AdminAccountCreateRequest, ) -> dict[str, Any]: _require_admin_token(token) return _admin_create_account(state, payload.account_id) @app.delete('/api/admin/accounts/{account_id}') async def admin_delete_account(account_id: str, token: str) -> dict[str, Any]: _require_admin_token(token) return _admin_delete_account(state, account_id) @app.get('/api/admin/memory') async def admin_memory(token: str, account_id: str | None = None) -> dict[str, Any]: _require_admin_token(token) safe_account = _safe_account_id(account_id) if account_id else None return { 'queue': state.memory_manager.queue_snapshot(safe_account), 'events': ( state.memory_manager.list_recent_events(safe_account, limit=100) if safe_account else [] ), } @app.get('/api/models') async def list_models_get(account_id: str | None = None) -> dict[str, Any]: config = state.config_for(account_id) return _list_backend_models(config.base_url, config.api_key) @app.post('/api/models') async def list_models_post(payload: ModelListRequest) -> dict[str, Any]: config = state.config_for(payload.account_id) return _list_backend_models( payload.base_url or config.base_url, payload.api_key or config.api_key, ) # ------------- Feishu integration --------------------------------------- @app.get('/api/integrations/feishu/status') async def feishu_status(account_id: str | None = None) -> dict[str, Any]: return _feishu_status_payload(state, account_id) @app.post('/api/integrations/feishu/login') async def feishu_login(payload: FeishuAccountRequest) -> dict[str, Any]: return _start_feishu_login(state, payload.account_id) @app.post('/api/integrations/feishu/logout') async def feishu_logout(payload: FeishuAccountRequest) -> dict[str, Any]: _stop_feishu_login(payload.account_id) paths = _feishu_paths(state, payload.account_id) result = _run_feishu_cli(paths, ['logout'], timeout_seconds=30) status = _feishu_status_payload(state, payload.account_id) status['logout_output'] = result['output'] return status @app.post('/api/files/online-doc') async def create_feishu_online_doc(payload: FeishuOnlineDocRequest) -> dict[str, Any]: file_path, online_doc_key = _resolve_feishu_source_file( state, payload.account_id, payload.path, ) suffix = file_path.suffix.lower() if suffix in FEISHU_UNSUPPORTED_DOCUMENT_SUFFIXES: raise HTTPException( status_code=400, detail='当前先不支持 json/jsonl 转在线文档。', ) if suffix not in FEISHU_SUPPORTED_DOCUMENT_SUFFIXES: raise HTTPException( status_code=400, detail='当前仅支持 md、txt、csv、docx、xlsx 转在线文档。', ) status = _feishu_status_payload(state, payload.account_id) if not status.get('logged_in'): raise HTTPException( status_code=409, detail={ 'code': 'feishu_not_logged_in', 'message': '当前账号还没有完成飞书授权，请先授权后再转换。', 'status': status, }, ) paths = _feishu_paths(state, payload.account_id) runtime = MCPRuntime(servers=(_feishu_server_profile(paths),)) title = _clean_feishu_doc_title(payload.title or file_path.stem) kind = 'sheet' if suffix in FEISHU_SPREADSHEET_SUFFIXES else 'doc' try: if kind == 'sheet': rendered, metadata = _create_feishu_online_spreadsheet( runtime, file_path, title=title, folder_token=payload.folder_token, ) else: markdown = _convert_file_to_feishu_markdown(file_path, title=title) arguments: dict[str, Any] = {'title': title, 'markdown': markdown} if payload.folder_token and payload.folder_token.strip(): arguments['folder_token'] = payload.folder_token.strip() rendered, metadata = runtime.call_tool( 'doc_create', arguments=arguments, server_name=FEISHU_MCP_SERVER_NAME, max_chars=API_TOOL_CONTENT_MAX_CHARS, timeout_seconds=60.0, ) except DataAgentInputError as exc: raise HTTPException(status_code=400, detail=str(exc)) except OSError as exc: raise HTTPException(status_code=400, detail=f'读取文件失败: {exc}') except Exception as exc: target_name = '飞书表格' if kind == 'sheet' else '飞书文档' raise HTTPException(status_code=502, detail=f'{target_name}创建失败: {exc}') url = _extract_first_url(rendered) if url: _record_feishu_online_doc( state, payload.account_id, file_path=file_path, map_key=online_doc_key, title=title, url=url, kind=kind, ) return { 'ok': True, 'title': title, 'url': url, 'kind': kind, 'file_path': str(file_path), 'file_type': suffix.lstrip('.'), 'result': rendered, 'metadata': metadata, } # ------------- sessions -------------------------------------------------- @app.get('/api/sessions') async def list_sessions(account_id: str | None = None) -> list[dict[str, Any]]: directory = state.account_paths(account_id)['sessions'] if not directory.exists(): return [] results: list[dict[str, Any]] = [] for path in sorted(_iter_session_files(directory), key=_session_mtime, reverse=True): try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): continue session_id = str(data.get('session_id', _session_id_from_path(path))) usage = data.get('usage') if isinstance(data.get('usage'), dict) else {} model_config = ( data.get('model_config') if isinstance(data.get('model_config'), dict) else {} ) messages = data.get('messages') or [] title = data.get('title') preview = '' for msg in messages: if isinstance(msg, dict) and msg.get('role') == 'user': content = msg.get('content', '') if isinstance(content, str) and not _is_internal_message(content): preview = _strip_session_context(content)[:120] break results.append( { 'session_id': session_id, 'turns': data.get('turns', 0), 'tool_calls': data.get('tool_calls', 0), 'preview': title if isinstance(title, str) and title.strip() else preview, 'modified_at': _session_mtime(path), 'model': model_config.get('model'), 'usage': usage, 'is_training': bool(data.get('is_training', False)), } ) return results @app.get('/api/sessions/{session_id}') async def get_session(session_id: str, account_id: str | None = None) -> dict[str, Any]: directory = state.account_paths(account_id)['sessions'] safe_id = _safe_session_id(session_id) if safe_id is None: raise HTTPException(status_code=404, detail='Session not found') try: stored = load_agent_session(safe_id, directory=directory) except FileNotFoundError: raise HTTPException(status_code=404, detail='Session not found') return _serialize_stored_session(stored) @app.delete('/api/sessions/{session_id}') async def delete_session(session_id: str, account_id: str | None = None) -> dict[str, Any]: directory = state.account_paths(account_id)['sessions'] safe_id = _safe_session_id(session_id) if safe_id is None or not _delete_session_files(directory, safe_id): raise HTTPException(status_code=404, detail='Session not found') return {'deleted': True, 'session_id': safe_id} @app.patch('/api/sessions/{session_id}') async def update_session( session_id: str, payload: SessionUpdate, account_id: str | None = None, ) -> dict[str, Any]: directory = state.account_paths(account_id)['sessions'] safe_id = _safe_session_id(session_id) if safe_id is None: raise HTTPException(status_code=400, detail='Invalid session id') if payload.title is None and payload.is_training is None: raise HTTPException(status_code=400, detail='Nothing to update') result: dict[str, Any] = {'session_id': safe_id} if payload.title is not None: title = _clean_manual_session_title(payload.title) if title is None: raise HTTPException(status_code=400, detail='Invalid session title') if not _update_session_title(directory, safe_id, title): raise HTTPException(status_code=404, detail='Session not found') result['title'] = title if payload.is_training is not None: value = bool(payload.is_training) if not _update_session_training(directory, safe_id, value): raise HTTPException(status_code=404, detail='Session not found') result['is_training'] = value return result @app.get('/api/training/step-detail') async def get_training_step_detail( session_id: str, step: str, run_id: str | None = None, account_id: str | None = None, ) -> Response: sessions_dir = state.account_paths(account_id)['sessions'] state_path = _session_state_path(sessions_dir, session_id) state_entries, _ = _read_program_state(state_path) # Per-round: if frontend passed run_id, use that round's actual runDic # so historical step views (e.g., R1's augment) read the right # workflow directory. Fall back to "max in state" only when run_id # is absent or the round has no resolvable runDic — preserves prior # behaviour for unbundled callers. run_dic: int | None = None if isinstance(run_id, str) and run_id.strip(): rid = run_id.strip() per_round = _resolve_run_dic_per_round(state_entries) # augment 处理的是「上一轮评测的错例」，产物按 R{n-1} 的 runDic 命名 # （data_clean_/、augment_.jsonl ...），不是本轮 eval workflow if step == 'augment': m = re.match(r'^[Rr](\d+)$', rid) if m and int(m.group(1)) >= 1: run_dic = per_round.get(f'R{int(m.group(1)) - 1}') if run_dic is None: run_dic = per_round.get(rid) if run_dic is None: run_dic = _resolve_run_dic_from_state(state_entries) runtime = _jupyter_runtime_for_session(state, account_id, session_id) artifact_paths = _step_artifact_paths(step, run_dic, runtime) if not artifact_paths: payload = { 'step': step, 'run_dic': run_dic, 'format': 'empty', 'content': '', 'source_path': None, 'fetched_at_ms': int(time.time() * 1000), 'error': '该 step 没有定义产物路径', } return Response( content=json.dumps(payload, ensure_ascii=False), media_type='application/json', headers={'Cache-Control': 'no-store'}, ) # 多分区表格视图：augment（H1 修改 / H2 仿写）与 dist-analysis（候选 + 报告）共用 # path_indices 指向 _step_artifact_paths 返回的路径列表；多个候选按顺序尝试， # 用第一个能读到的（让 dist-analysis 的 NFS 主路径 + jupyter pod 本地兜底 # 路径合并成同一个 section，避免出现重复卡片）。 multi_section_specs: dict[str, list[dict[str, Any]]] = { 'augment': [ {'title': 'H1 确认修改（complex 翻转 → 原文件 in-place 改标）', 'max_rows': None, 'path_indices': [0]}, {'title': 'H2 仿写候选 raw（GPT 原始输出，未过 sanity）', 'max_rows': None, 'path_indices': [1]}, {'title': 'H2 仿写最终（本轮新增训练样本，合入 augment.jsonl）', 'max_rows': None, 'path_indices': [2]}, {'title': 'label-master 标签复核（H1+H2 全量 verdict）', 'max_rows': None, 'path_indices': [3]}, ], 'dist-analysis': [ {'title': '问题分析报告（workflow.md）', 'max_rows': None, 'path_indices': [0]}, {'title': 'H1 预计修改候选', 'max_rows': None, 'path_indices': [1, 2]}, ], } if step == 'dist-analysis': # 从 iteration_log.jsonl 读 H1 假设短标签，注入到候选集 section 标题 h1_label = await asyncio.to_thread( _read_iteration_log_h1_label, runtime, run_dic ) if h1_label: for spec in multi_section_specs['dist-analysis']: if spec['title'] == 'H1 预计修改候选': spec['title'] = f'H1 预计修改候选（{h1_label}）' break if step in multi_section_specs: section_specs = multi_section_specs[step] table_sections: list[dict[str, Any]] = [] tried: list[str] = [] errors: list[str] = [] for spec in section_specs: title = spec['title'] max_rows = spec.get('max_rows') candidate_paths = [ artifact_paths[i] for i in spec.get('path_indices', []) if 0 <= i < len(artifact_paths) ] if not candidate_paths: continue # 多候选按顺序尝试，用第一个能读到的；都读不到则用首个候选作为缺失占位 content: str | None = None path_str: str = candidate_paths[0] last_err: str | None = None for cand in candidate_paths: tried.append(cand) content_try, err = await asyncio.to_thread( _read_artifact, runtime, cand ) if content_try is not None: content = content_try path_str = cand break last_err = err if content is None: if last_err: errors.append(f'{title}: {last_err}') table_sections.append( { 'title': title, 'source_path': candidate_paths[0], 'kind': 'missing', 'columns': [], 'rows': [], 'total_rows': 0, 'shown_rows': 0, 'error': last_err or '产物文件不存在', } ) continue fmt = _detect_artifact_format(path_str) try: if fmt == 'markdown': # markdown 直接以原文渲染，不走表格 table_sections.append( { 'title': title, 'source_path': path_str, 'kind': 'markdown', 'columns': [], 'rows': [], 'total_rows': 0, 'shown_rows': 0, 'body': content, } ) continue if fmt == 'csv': cols, table_rows, total = _csv_to_table(content, max_rows) kind = 'csv' elif fmt in {'jsonl', 'json'}: cols, table_rows, total = _jsonl_to_table( content, max_rows ) kind = 'jsonl' else: lines = content.splitlines() total = len(lines) capped = lines if max_rows is None else lines[:max_rows] cols = ['line'] table_rows = [[_truncate_cell(line)] for line in capped] kind = 'text' except Exception as exc: # noqa: BLE001 errors.append(f'{title}: parse failed: {exc}') table_sections.append( { 'title': title, 'source_path': path_str, 'kind': 'error', 'columns': [], 'rows': [], 'total_rows': 0, 'shown_rows': 0, 'error': f'解析失败：{exc}', } ) continue table_sections.append( { 'title': title, 'source_path': path_str, 'kind': kind, 'columns': cols, 'rows': table_rows, 'total_rows': total, 'shown_rows': len(table_rows), } ) if step == 'dist-analysis': # 软警告：dist-analysis 标 complete 时若 H1 候选 CSV 缺失，把 section # 的兜底 "产物文件不存在" 换成更明确的提示。某些轮次（hparam 类假设、 # 无标签问题）确实可能没有候选，所以不 fail 这一步、不阻塞 human-check。 target_run = (run_id or '').strip() last_status: str | None = None for entry in state_entries: if not isinstance(entry, dict) or entry.get('step') != 'dist-analysis': continue eid = (entry.get('run_id') or '').strip() if target_run and eid and eid != target_run: continue if 'status' in entry: last_status = entry.get('status') if last_status == 'complete': rd_label = str(run_dic) if run_dic is not None else '' for sec in table_sections: if ( str(sec.get('title', '')).startswith('H1 预计修改候选') and sec.get('kind') == 'missing' ): sec['error'] = ( f'未检测到 output/relabel_candidates_{rd_label}.csv —— ' f'若本轮无 H1 候选可忽略；如有候选请按 SKILL.md 「阶段一」' f'落盘 file,line,query,old_label,是否改(1/0),建议新label。' ) break payload = { 'step': step, 'run_dic': run_dic, 'format': 'tables', 'sections': table_sections, 'source_path': artifact_paths[0], 'tried_paths': tried, 'fetched_at_ms': int(time.time() * 1000), } if errors: payload['error'] = '; '.join(errors) return Response( content=json.dumps(payload, ensure_ascii=False), media_type='application/json', headers={'Cache-Control': 'no-store'}, ) last_error: str | None = None for path_str in artifact_paths: content, err = await asyncio.to_thread( _read_artifact, runtime, path_str ) if content is not None: fmt = _detect_artifact_format(path_str) if fmt == 'jsonl': # 只取前 50 行，避免超大 content = '\n'.join(content.splitlines()[:50]) payload = { 'step': step, 'run_dic': run_dic, 'format': fmt, 'content': content, 'source_path': path_str, 'fetched_at_ms': int(time.time() * 1000), } return Response( content=json.dumps(payload, ensure_ascii=False), media_type='application/json', headers={'Cache-Control': 'no-store'}, ) if err: last_error = err payload = { 'step': step, 'run_dic': run_dic, 'format': 'empty', 'content': '', 'source_path': artifact_paths[0] if artifact_paths else None, 'fetched_at_ms': int(time.time() * 1000), 'error': last_error or '产物文件不存在或读取失败', 'tried_paths': artifact_paths, } return Response( content=json.dumps(payload, ensure_ascii=False), media_type='application/json', headers={'Cache-Control': 'no-store'}, ) @app.get('/api/training/pipeline') async def get_training_pipeline( session_id: str | None = None, account_id: str | None = None, ) -> Response: sessions_dir = state.account_paths(account_id)['sessions'] state_path = _session_state_path(sessions_dir, session_id) state_entries, _ = _read_program_state(state_path) iter_count = await asyncio.to_thread( _read_iteration_log_count, session_id, agent_state=state, account_id=account_id, ) metrics_history = await asyncio.to_thread( _read_iteration_log_metrics_history, session_id, agent_state=state, account_id=account_id, ) payload = await asyncio.to_thread( _hardcoded_autoresearch_pipeline, session_id, sessions_dir=sessions_dir, state_entries=state_entries, model_config=state.model_config_for(account_id), agent_state=state, account_id=account_id, iteration_log_count=iter_count, metrics_history=metrics_history, ) failure_status = await asyncio.to_thread( _read_session_failure_status, sessions_dir, session_id, ) run_active = _is_run_active(state, account_id, session_id) _apply_program_state( payload, state_entries, iteration_log_count=iter_count or 0, session_failure_status=failure_status, run_active=run_active, session_id=session_id, ) return Response( content=json.dumps(payload, ensure_ascii=False), media_type='application/json', headers={'Cache-Control': 'no-store'}, ) @app.get('/api/training/pipeline/stream') async def stream_training_pipeline( request: Request, session_id: str | None = None, account_id: str | None = None, ) -> StreamingResponse: sessions_dir = state.account_paths(account_id)['sessions'] state_path = _session_state_path(sessions_dir, session_id) async def event_gen(): last_signature: str | None = None last_heartbeat = time.monotonic() refresher = asyncio.create_task( _iter_count_refresh_loop( session_id, agent_state=state, account_id=account_id, ) ) metrics_refresher = asyncio.create_task( _metrics_history_refresh_loop( session_id, agent_state=state, account_id=account_id, ) ) try: while True: if await request.is_disconnected(): break state_entries, state_mtime = _read_program_state(state_path) iter_count = await asyncio.to_thread( _read_iteration_log_count, session_id, agent_state=state, account_id=account_id, ) metrics_history = await asyncio.to_thread( _read_iteration_log_metrics_history, session_id, agent_state=state, account_id=account_id, ) payload = await asyncio.to_thread( _hardcoded_autoresearch_pipeline, session_id, sessions_dir=sessions_dir, state_entries=state_entries, agent_state=state, account_id=account_id, iteration_log_count=iter_count, metrics_history=metrics_history, ) failure_status = await asyncio.to_thread( _read_session_failure_status, sessions_dir, session_id, ) run_active = _is_run_active(state, account_id, session_id) _apply_program_state( payload, state_entries, iteration_log_count=iter_count or 0, session_failure_status=failure_status, run_active=run_active, session_id=session_id, ) payload_str = json.dumps(payload, ensure_ascii=False) signature = f'{state_mtime}|{hash(payload_str)}' if signature != last_signature: last_signature = signature yield f'data: {payload_str}\n\n' now = time.monotonic() if now - last_heartbeat > 15: yield ': heartbeat\n\n' last_heartbeat = now await asyncio.sleep(0.5) finally: refresher.cancel() try: await refresher except (asyncio.CancelledError, Exception): # noqa: BLE001 pass metrics_refresher.cancel() try: await metrics_refresher except (asyncio.CancelledError, Exception): # noqa: BLE001 pass return StreamingResponse( event_gen(), media_type='text/event-stream', headers={ 'Cache-Control': 'no-store', 'X-Accel-Buffering': 'no', 'Connection': 'keep-alive', }, ) @app.get('/api/context-budget') async def get_context_budget( session_id: str | None = None, account_id: str | None = None, ) -> dict[str, Any]: with state.lock(): if session_id: directory = state.account_paths(account_id)['sessions'] safe_id = _safe_session_id(session_id) if safe_id is None: raise HTTPException(status_code=404, detail='Session not found') try: stored = load_agent_session(safe_id, directory=directory) except FileNotFoundError: raise HTTPException(status_code=404, detail='Session not found') session = _session_state_from_stored(stored) model = str(stored.model_config.get('model') or state.model) runtime_config = deserialize_runtime_config(stored.runtime_config) else: agent = state.agent_for(account_id) session = agent.last_session or agent.build_session() model = agent.model_config.model runtime_config = agent.runtime_config snapshot = calculate_token_budget( session=session, model=model, budget_config=runtime_config.budget_config, output_schema=runtime_config.output_schema, ) return _serialize_token_budget(snapshot) def _merge_run_snapshot( memory_snapshot: dict[str, Any] | None, stored_snapshot: dict[str, Any] | None, ) -> dict[str, Any] | None: if memory_snapshot is None: return stored_snapshot if stored_snapshot is None: merged = dict(memory_snapshot) else: merged = {**stored_snapshot, **memory_snapshot} if not memory_snapshot.get('events') and stored_snapshot.get('events'): merged['events'] = stored_snapshot.get('events') if memory_snapshot.get('current_stage'): merged['current_stage'] = memory_snapshot.get('current_stage') if ( memory_snapshot.get('status') not in ACTIVE_RUN_STATUSES and stored_snapshot.get('elapsed_ms') is not None ): merged['elapsed_ms'] = stored_snapshot.get('elapsed_ms') if stored_snapshot.get('finished_at') is not None: merged['finished_at'] = stored_snapshot.get('finished_at') merged['cancellable'] = merged.get('status') in ACTIVE_RUN_STATUSES return merged @app.get('/api/runs/latest') async def latest_run( session_id: str, account_id: str | None = None, ) -> dict[str, Any]: safe_id = _safe_session_id(session_id) if safe_id is None: raise HTTPException(status_code=400, detail='Invalid session id') account_key = state._account_key(account_id) snapshot = state.run_manager.snapshot_latest(account_key, safe_id) stored_snapshot = state.run_state_store.snapshot_latest(account_key, safe_id) if ( snapshot is not None and stored_snapshot is not None and stored_snapshot.get('status') in ACTIVE_RUN_STATUSES and snapshot.get('run_id') == stored_snapshot.get('run_id') and snapshot.get('status') not in ACTIVE_RUN_STATUSES ): state.run_state_store.finish( str(stored_snapshot.get('run_id') or ''), status=str(snapshot.get('status') or 'interrupted'), elapsed_ms=( int(snapshot['elapsed_ms']) if isinstance(snapshot.get('elapsed_ms'), (int, float)) else None ), stage=str(snapshot.get('current_stage') or ''), error=( str(snapshot.get('error')) if snapshot.get('error') is not None else None ), ) stored_snapshot = state.run_state_store.snapshot_latest(account_key, safe_id) if snapshot is not None: return _merge_run_snapshot(snapshot, stored_snapshot) or snapshot if stored_snapshot is not None: if stored_snapshot.get('status') in ACTIVE_RUN_STATUSES: interrupted = state.run_state_store.mark_interrupted_if_active( str(stored_snapshot.get('run_id') or ''), ) return interrupted or stored_snapshot return stored_snapshot directory = state.account_paths(account_id)['sessions'] try: stored = load_agent_session(safe_id, directory=directory) except FileNotFoundError: return {'session_id': safe_id, 'status': 'idle'} if _stored_session_has_incomplete_tail(stored): return { 'session_id': safe_id, 'status': 'interrupted', 'current_stage': '上次运行已中断，后台没有正在执行的进程', } return {'session_id': safe_id, 'status': 'idle'} @app.post('/api/runs/cancel') async def cancel_run(payload: RunCancelRequest) -> dict[str, Any]: safe_id = _safe_session_id(payload.session_id) if safe_id is None: raise HTTPException(status_code=400, detail='Invalid session id') account_key = state._account_key(payload.account_id) cancelled_ids: set[str] = set() if payload.run_id and state.run_manager.cancel_run(payload.run_id): cancelled_ids.add(payload.run_id) cancelled_ids.update(state.run_manager.cancel_session(account_key, safe_id)) stored_cancelled = state.run_state_store.finish_active_for_session( account_key, safe_id, status='cancelled', stage='用户已取消', ) cancelled_ids.update(stored_cancelled) cancelled = bool(cancelled_ids) if cancelled: _mark_session_interrupted( state.account_paths(payload.account_id)['sessions'], safe_id, status='cancelled', ) return { 'session_id': safe_id, 'cancelled': cancelled, 'cancelled_run_ids': sorted(cancelled_ids), } def _run_chat_payload( request: ChatRequest, event_sink: Any | None = None, ) -> dict[str, Any]: requested_session_id = _safe_session_id( request.resume_session_id or request.session_id ) or uuid4().hex account_key = state._account_key(request.account_id) prompt = request.prompt.strip() run_record = state.run_manager.start(account_key, requested_session_id, prompt) state.run_state_store.start( run_id=run_record.run_id, account_key=account_key, session_id=requested_session_id, pending_prompt=prompt, started_at=run_record.started_at, ) run_lock = state.run_lock_for(request.account_id, requested_session_id) agent = state.agent_for(request.account_id, requested_session_id) config = state.config_for(request.account_id) session_directory = state.account_paths(request.account_id)['sessions'] jupyter_runtime = _jupyter_runtime_for_session( state, request.account_id, requested_session_id, ) runtime_context = request.runtime_context memory_context = state.memory_manager.render_injection( request.account_id, state.enabled_skill_names(request.account_id), ) if memory_context: runtime_context = _append_runtime_context(runtime_context, memory_context) if jupyter_runtime is not None: runtime_context = _append_runtime_context( runtime_context, jupyter_runtime.render_context(), ) def emit_agent_event(event: dict[str, object]) -> None: stage = _runtime_event_stage(event) if stage: state.run_manager.update(run_record.run_id, stage=stage) state.run_state_store.update(run_record.run_id, stage=stage) stored_event = state.run_manager.record_event(run_record.run_id, event) if stored_event is not None: state.run_state_store.record_event(run_record.run_id, stored_event) _emit_runtime_event(event_sink, event) previous_title = _read_session_title( session_directory, requested_session_id, ) fallback_title = _save_in_progress_session( directory=session_directory, agent=agent, session_id=requested_session_id, prompt=prompt, ) if run_lock.locked(): queued_event = { 'type': 'run_queued', 'run_id': run_record.run_id, 'session_id': requested_session_id or '', 'account_id': request.account_id or '', } stored_event = state.run_manager.record_event(run_record.run_id, queued_event) if stored_event is not None: state.run_state_store.record_event(run_record.run_id, stored_event) _emit_runtime_event( event_sink, queued_event, ) with run_lock: if run_record.cancel_event.is_set(): state.run_manager.finish(run_record.run_id, 'cancelled') state.run_state_store.finish( run_record.run_id, status='cancelled', elapsed_ms=max(0, int((time.time() - run_record.started_at) * 1000)), stage='已取消排队中的请求', ) return { 'final_output': '已取消排队中的请求。', 'turns': 0, 'tool_calls': 0, 'transcript': (), 'session_id': requested_session_id, 'usage': {}, 'total_cost_usd': 0.0, 'stop_reason': 'cancelled', } state.run_manager.update(run_record.run_id, status='running', stage='后端开始执行') state.run_state_store.update( run_record.run_id, status='running', stage='后端开始执行', cancellable=True, ) started_event = { 'type': 'run_started', 'run_id': run_record.run_id, 'started_at': run_record.started_at, 'session_id': requested_session_id or '', 'account_id': request.account_id or '', } stored_event = state.run_manager.record_event(run_record.run_id, started_event) if stored_event is not None: state.run_state_store.record_event(run_record.run_id, stored_event) _emit_runtime_event(event_sink, started_event) event_loop = state.event_loop def _bg_register(spec: BgTaskSpec) -> None: if event_loop is None: raise RuntimeError( 'event loop unavailable; backend not started via lifespan' ) asyncio.run_coroutine_threadsafe( _bash_bg_manager.register(state, spec), event_loop, ) def _bg_status_query(task_id: str) -> 'BgTaskStatus | None': return _bash_bg_manager.status_query(state, task_id) def _bg_kill_request(task_id: str) -> bool: if event_loop is None: return False future = asyncio.run_coroutine_threadsafe( _bash_bg_manager.cancel(state, task_id), event_loop, ) try: return bool(future.result(timeout=15)) except Exception: # noqa: BLE001 return False agent.tool_context = replace( agent.tool_context, cancel_event=run_record.cancel_event, process_registry=run_record.process_registry, jupyter_runtime=jupyter_runtime, bg_register=_bg_register, bg_status_query=_bg_status_query, bg_kill_request=_bg_kill_request, account_id=account_key, session_id=requested_session_id, run_id=run_record.run_id, ) started_at = time.perf_counter() try: try: if request.resume_session_id is not None: try: stored = load_agent_session( requested_session_id, directory=session_directory, ) except FileNotFoundError: raise HTTPException( status_code=404, detail='Session to resume not found', ) stored = _sanitize_stored_session_for_resume(stored) result = agent.resume( prompt, stored, runtime_context=runtime_context, event_sink=emit_agent_event, ) else: result = agent.run( prompt, session_id=requested_session_id, runtime_context=runtime_context, event_sink=emit_agent_event, ) except Exception as exc: _mark_session_interrupted( session_directory, requested_session_id, status=( 'cancelled' if run_record.cancel_event.is_set() else 'failed' ), detail=str(exc), ) state.run_manager.update( run_record.run_id, status=( 'cancelled' if run_record.cancel_event.is_set() else 'failed' ), error=str(exc), ) state.run_state_store.finish( run_record.run_id, status=( 'cancelled' if run_record.cancel_event.is_set() else 'failed' ), elapsed_ms=max( 0, int((time.time() - run_record.started_at) * 1000), ), error=str(exc), ) raise finally: agent.tool_context = replace( agent.tool_context, cancel_event=None, process_registry=None, jupyter_runtime=None, bg_register=None, bg_status_query=None, bg_kill_request=None, account_id=None, session_id=None, run_id=None, ) elapsed_ms = max(0, int((time.perf_counter() - started_at) * 1000)) payload = _serialize_run_result(result) payload['elapsed_ms'] = elapsed_ms used_skills = _extract_used_skill_names(payload.get('transcript') or ()) if not used_skills: used_skills = tuple(state.enabled_skill_names(request.account_id)) if result.session_id: _annotate_last_assistant_elapsed( session_directory, result.session_id, elapsed_ms, ) _ensure_session_title( session_directory, result.session_id, model=config.model, base_url=config.base_url, api_key=config.api_key, previous_title=previous_title, fallback_title=fallback_title, ) _annotate_transcript_elapsed(payload, elapsed_ms) try: state.memory_manager.enqueue_interaction( account_id=request.account_id, session_id=result.session_id, user_prompt=prompt, assistant_output=result.final_output, skills=used_skills, model=config.model, ) except Exception: pass state.run_manager.finish( run_record.run_id, 'cancelled' if run_record.cancel_event.is_set() else 'completed', ) state.run_state_store.finish( run_record.run_id, status='cancelled' if run_record.cancel_event.is_set() else 'completed', elapsed_ms=max(0, int((time.time() - run_record.started_at) * 1000)), ) if run_record.cancel_event.is_set() and result.session_id: _mark_session_interrupted( session_directory, result.session_id, status='cancelled', ) return payload # ------------- chat ------------------------------------------------------ @app.post('/api/chat') async def chat(request: ChatRequest) -> dict[str, Any]: prompt = request.prompt.strip() if not prompt: raise HTTPException(status_code=400, detail='Prompt is empty') def _run() -> dict[str, Any]: return _run_chat_payload(request) try: payload = await asyncio.to_thread(_run) except HTTPException: raise except Exception as exc: # surface the error in the UI return JSONResponse( status_code=500, content={ 'error': str(exc), 'error_type': type(exc).__name__, }, ) return payload @app.post('/api/chat/stream') async def chat_stream(request: ChatRequest) -> StreamingResponse: prompt = request.prompt.strip() if not prompt: raise HTTPException(status_code=400, detail='Prompt is empty') event_queue: queue.Queue[dict[str, Any] | None] = queue.Queue() def emit_event(event: dict[str, object]) -> None: event_queue.put({'kind': 'event', 'event': event}) def worker() -> None: try: payload = _run_chat_payload(request, event_sink=emit_event) except HTTPException as exc: event_queue.put( { 'kind': 'error', 'status_code': exc.status_code, 'detail': exc.detail, } ) except Exception as exc: event_queue.put( { 'kind': 'error', 'status_code': 500, 'error': str(exc), 'error_type': type(exc).__name__, } ) else: event_queue.put({'kind': 'result', 'payload': payload}) finally: event_queue.put(None) threading.Thread(target=worker, daemon=True).start() stream_started = time.perf_counter() def generate() -> Any: while True: try: item = event_queue.get(timeout=15) except queue.Empty: yield json.dumps( { 'kind': 'event', 'event': { 'type': 'server_heartbeat', 'elapsed_ms': max( 0, int((time.perf_counter() - stream_started) * 1000), ), }, }, ensure_ascii=False, ) + '\n' continue if item is None: break yield json.dumps(item, ensure_ascii=False) + '\n' return StreamingResponse( generate(), media_type='application/x-ndjson', headers={ 'Cache-Control': 'no-cache, no-transform', 'X-Accel-Buffering': 'no', }, ) @app.post('/api/clear') async def clear_state(account_id: str | None = None) -> dict[str, Any]: with state.lock(): state.agent_for(account_id).clear_runtime_state() return state.snapshot(account_id) return app # --------------------------------------------------------------------------- # Serialization helpers # --------------------------------------------------------------------------- def _serialize_run_result(result: Any) -> dict[str, Any]: return { 'final_output': result.final_output, 'turns': result.turns, 'tool_calls': result.tool_calls, 'transcript': [_normalize_transcript_entry(entry) for entry in result.transcript], 'session_id': result.session_id, 'usage': result.usage.to_dict(), 'total_cost_usd': result.total_cost_usd, 'stop_reason': result.stop_reason, } def _normalize_transcript_entry(entry: dict[str, Any]) -> dict[str, Any]: content = entry.get('content', '') if entry.get('role') == 'tool' and isinstance(content, str): content = _truncate_api_text(content, API_TOOL_CONTENT_MAX_CHARS) out: dict[str, Any] = { 'role': entry.get('role', ''), 'content': content, } for key in ('name', 'tool_call_id', 'tool_calls', 'metadata', 'message_id'): if key in entry and entry[key] not in (None, '', [], {}): out[key] = entry[key] for key in ('state', 'stop_reason'): if key in entry and entry[key] not in (None, ''): out[key] = entry[key] return out def _truncate_api_text(text: str, limit: int) -> str: if len(text) <= limit: return text head = text[: limit // 2] tail = text[-(limit // 2) :] return f'{head}\n...[truncated for api response]...\n{tail}' def _serialize_stored_session(stored: StoredAgentSession) -> dict[str, Any]: return { 'session_id': stored.session_id, 'turns': stored.turns, 'tool_calls': stored.tool_calls, 'messages': [_normalize_transcript_entry(dict(m)) for m in stored.messages], 'usage': stored.usage, 'total_cost_usd': stored.total_cost_usd, 'model': stored.model_config.get('model'), 'is_training': stored.is_training, } def _stored_session_has_incomplete_tail(stored: StoredAgentSession) -> bool: _, changed = _trim_incomplete_message_tail(stored.messages) if changed: return True budget_state = stored.budget_state if isinstance(stored.budget_state, dict) else {} return budget_state.get('status') in {'queued', 'running'} def _is_run_active(state: 'AgentState', account_id: str | None, session_id: str | None) -> bool | None: """True iff `run_state_store` says the latest run for this session is in `queued`/`running`. None when the lookup is impossible (no session id, or no row yet). The pipeline endpoint uses this to detect orphan `running` step entries — agent wrote `step:running`, finished its turn cleanly, and never wrote the closing entry; from the user's POV the run is idle but the panel was still spinning. """ if not session_id: return None try: account_key = state._account_key(account_id) snapshot = state.run_state_store.snapshot_latest(account_key, session_id) except Exception: return None if snapshot is None: return None return snapshot.get('status') in ACTIVE_RUN_STATUSES def _read_session_failure_status( directory: Path, session_id: str | None, ) -> str | None: """Returns 'failed' / 'cancelled' / 'interrupted' iff the stored session was explicitly tombstoned by `_mark_session_interrupted`. Returns None for healthy sessions AND for live runs (budget_state.status='running'/'queued'). The pipeline panel uses this to paint a failed run red. We deliberately do NOT use `_stored_session_has_incomplete_tail` here — that helper is for detecting incomplete tails when LOADING a session for resume, where budget_state='running' implies a dead writer. Polled live, 'running' just means the agent is still chugging. """ if not session_id: return None try: stored = load_agent_session(session_id, directory=directory) except (FileNotFoundError, OSError, json.JSONDecodeError): return None bs = stored.budget_state if isinstance(stored.budget_state, dict) else {} status = bs.get('status') if status in ('failed', 'cancelled', 'interrupted'): return str(status) return None def _mark_session_interrupted( directory: Path, session_id: str, *, status: str, detail: str = '', ) -> None: try: stored = load_agent_session(session_id, directory=directory) except (FileNotFoundError, OSError, json.JSONDecodeError): return trimmed, changed = _trim_incomplete_message_tail(stored.messages) budget_state = ( dict(stored.budget_state) if isinstance(stored.budget_state, dict) else {} ) if not changed and budget_state.get('status') not in {'queued', 'running'}: return messages = list(trimmed) if not _last_message_is_run_status(messages, status): messages.append( { 'role': 'assistant', 'content': _interrupted_session_message(status), 'state': 'final', 'stop_reason': status, 'metadata': { 'kind': 'run_status', 'status': status, 'detail': detail[:500], 'created_at_ms': int(time.time() * 1000), }, } ) budget_state['status'] = status budget_state['interrupted_at'] = int(time.time()) save_agent_session( replace( stored, messages=tuple(messages), budget_state=budget_state, ), directory=directory, ) def _sanitize_stored_session_for_resume( stored: StoredAgentSession, ) -> StoredAgentSession: trimmed, changed = _trim_incomplete_message_tail(stored.messages) if trimmed and _is_pending_user_message(trimmed[-1]): trimmed = trimmed[:-1] changed = True messages = _without_run_status_messages(trimmed) if len(messages) != len(trimmed): changed = True if not changed: return stored budget_state = ( dict(stored.budget_state) if isinstance(stored.budget_state, dict) else {} ) budget_state['status'] = 'interrupted' return replace( stored, messages=messages, budget_state=budget_state, ) def _trim_incomplete_message_tail( messages: tuple[dict[str, Any], ...] | tuple[Any, ...], ) -> tuple[tuple[dict[str, Any], ...], bool]: trimmed = [dict(message) for message in messages if isinstance(message, dict)] changed = False while trimmed and _is_incomplete_session_message(trimmed[-1]): trimmed.pop() changed = True while trimmed and _assistant_has_trailing_tool_call(trimmed[-1]): trimmed.pop() changed = True return tuple(trimmed), changed def _is_incomplete_session_message(message: dict[str, Any]) -> bool: state = message.get('state') if isinstance(state, str) and state not in {'', 'final'}: return True metadata = message.get('metadata') if isinstance(metadata, dict): if ( message.get('role') != 'user' and metadata.get('placeholder') is True and metadata.get('status') == 'running' ): return True if metadata.get('phase') in {'starting', 'running'} and message.get('role') == 'tool': return True return False def _is_pending_user_message(message: dict[str, Any]) -> bool: metadata = message.get('metadata') return ( message.get('role') == 'user' and isinstance(metadata, dict) and metadata.get('placeholder') is True and metadata.get('status') == 'running' ) def _assistant_has_trailing_tool_call(message: dict[str, Any]) -> bool: if message.get('role') != 'assistant': return False tool_calls = message.get('tool_calls') return isinstance(tool_calls, list) and bool(tool_calls) def _without_run_status_messages( messages: tuple[dict[str, Any], ...], ) -> tuple[dict[str, Any], ...]: return tuple( message for message in messages if not _is_run_status_message(message) ) def _is_run_status_message(message: dict[str, Any]) -> bool: metadata = message.get('metadata') return isinstance(metadata, dict) and metadata.get('kind') == 'run_status' def _last_message_is_run_status(messages: list[dict[str, Any]], status: str) -> bool: if not messages: return False metadata = messages[-1].get('metadata') return ( isinstance(metadata, dict) and metadata.get('kind') == 'run_status' and metadata.get('status') == status ) def _interrupted_session_message(status: str) -> str: if status == 'cancelled': return '上一次任务已取消，后台没有正在执行的进程。你可以继续回复，或重新发起任务。' if status == 'failed': return '上一次任务异常中断，后台没有正在执行的进程。你可以继续回复，或重新发起任务。' return '上一次任务已中断，后台没有正在执行的进程。你可以继续回复，或重新发起任务。' def _runtime_event_stage(event: dict[str, object]) -> str: event_type = event.get('type') if event_type == 'content_delta': return '' if event_type == 'tool_start': stage_note = str(event.get('assistant_content') or '').strip() if stage_note.startswith(('进度：', '进度:')): return stage_note tool_name = str(event.get('tool_name') or '').strip() return f'调用工具 {tool_name}' if tool_name else '正在调用工具' if event_type == 'tool_delta': tool_name = str(event.get('tool_name') or '').strip() return f'{tool_name} 输出中' if tool_name else '工具输出中' if event_type == 'tool_result': tool_name = str(event.get('tool_name') or '').strip() return f'工具完成 {tool_name}' if tool_name else '工具调用完成' if event_type == 'final_text_start': return '正在整理回复' if event_type == 'final_text_end': return '回复整理完成' if event_type == 'user_review_required': return '等待用户 review' if event_type == 'continuation_request': return '继续补全回复' if event_type in { 'prompt_length_check', 'prompt_length_recovery', 'auto_compact_summary', 'auto_compact_circuit_breaker', }: return '整理上下文' if event_type == 'task_budget_exceeded': return '达到任务预算限制' return '' _RUN_EVENT_TYPES = { 'run_queued', 'run_started', 'content_delta', 'tool_start', 'tool_delta', 'tool_result', 'final_text_start', 'final_text_end', 'user_review_required', 'continuation_request', 'prompt_length_check', 'prompt_length_recovery', 'auto_compact_summary', 'auto_compact_circuit_breaker', 'task_budget_exceeded', } def _normalize_run_event(event: dict[str, object]) -> dict[str, Any] | None: event_type = event.get('type') if not isinstance(event_type, str) or event_type not in _RUN_EVENT_TYPES: return None allowed = { 'type', 'run_id', 'session_id', 'account_id', 'tool_name', 'tool_call_id', 'arguments', 'delta', 'assistant_content', 'message_id', 'stream', 'ok', 'metadata', 'reason', 'continuation_index', 'turn_index', 'strategy', 'projected_input_tokens', 'soft_input_limit_tokens', 'hard_input_limit_tokens', 'exceeds_hard_limit', 'exceeds_soft_limit', } normalized = { key: _json_safe_limited(value, parent_key=key) for key, value in event.items() if key in allowed and value is not None } normalized['type'] = event_type return normalized _VERBOSE_STRING_KEYS = {'code', 'command', 'content', 'delta', 'script', 'stdout', 'output'} _VERBOSE_STRING_LIMIT = 20000 def _json_safe_limited(value: Any, *, depth: int = 0, parent_key: str | None = None) -> Any: if depth > 4: return str(value)[:500] if isinstance(value, str): limit = _VERBOSE_STRING_LIMIT if parent_key in _VERBOSE_STRING_KEYS else 2000 return value[:limit] if isinstance(value, (int, float, bool)) or value is None: return value if isinstance(value, dict): return { str(key)[:200]: _json_safe_limited(item, depth=depth + 1, parent_key=str(key)) for key, item in list(value.items())[:80] } if isinstance(value, (list, tuple)): return [_json_safe_limited(item, depth=depth + 1, parent_key=parent_key) for item in value[:80]] try: json.dumps(value) return value except TypeError: return str(value)[:1000] def _emit_runtime_event( event_sink: Any | None, event: dict[str, object], ) -> None: if event_sink is None: return event_sink(event) def _save_in_progress_session( *, directory: Path, agent: LocalCodingAgent, session_id: str | None, prompt: str, ) -> str | None: safe_id = _safe_session_id(session_id) if safe_id is None or not prompt: return None initial_title = _derive_initial_session_title(prompt) pending_metadata = {'status': 'running', 'placeholder': True} session_path = _session_json_path(directory, safe_id) if session_path.exists(): # 续聊也要先落盘用户消息。否则前端流断开或刷新时，用户刚发的内容只 # 存在于浏览器内存里，会出现“消息丢了但后端状态还在”的错觉。 try: stored = load_agent_session(safe_id, directory=directory) except (FileNotFoundError, OSError, json.JSONDecodeError): return None messages = list(stored.messages) if messages and _is_pending_user_message(dict(messages[-1])): messages.pop() messages.append( { 'role': 'user', 'content': prompt, 'state': 'final', 'metadata': pending_metadata, 'message_id': f'user_pending_{int(time.time() * 1000)}', } ) budget_state = ( dict(stored.budget_state) if isinstance(stored.budget_state, dict) else {} ) budget_state['status'] = 'running' try: save_agent_session( replace( stored, messages=tuple(messages), budget_state=budget_state, ), directory=directory, ) except OSError: return None return None # 新会话的第一轮执行可能很久。先写一个运行中占位，避免刷新页面后找不到会话。 scratchpad_directory = ( agent.runtime_config.scratchpad_root / safe_id / 'scratchpad' ).resolve() try: scratchpad_directory.mkdir(parents=True, exist_ok=True) session = agent.build_session(None, scratchpad_directory=scratchpad_directory) session.append_user( prompt, metadata=pending_metadata, message_id='user_pending_0', ) stored = StoredAgentSession( session_id=safe_id, model_config=serialize_model_config(agent.model_config), runtime_config=serialize_runtime_config(agent.runtime_config), system_prompt_parts=session.system_prompt_parts, user_context=dict(session.user_context), system_context=dict(session.system_context), messages=session.transcript(), turns=0, tool_calls=0, usage={}, total_cost_usd=0.0, file_history=(), budget_state={'status': 'running'}, plugin_state={}, scratchpad_directory=str(scratchpad_directory), ) saved_path = save_agent_session(stored, directory=directory) if initial_title: try: data = json.loads(saved_path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return initial_title data['title'] = initial_title data['title_source'] = 'first_message' data['title_generated_at'] = int(time.time()) _write_session_metadata(saved_path, data) return initial_title except OSError: return None def _derive_initial_session_title(prompt: str) -> str | None: # 第一条消息刚发出时先给侧边栏一个可读标题，后续再由模型摘要精修。 stripped = _strip_session_context(prompt) stripped = re.sub(r'\s+', ' ', stripped).strip() stripped = stripped.strip('"\'“”‘’`') if not stripped: return None title = _clean_session_title(stripped) return title or None def _session_state_from_stored(stored: StoredAgentSession) -> AgentSessionState: return AgentSessionState( system_prompt_parts=stored.system_prompt_parts, user_context=stored.user_context, system_context=stored.system_context, messages=[ AgentMessage.from_openai_message(message) for message in stored.messages if isinstance(message, dict) ], ) def _serialize_token_budget(snapshot: Any) -> dict[str, Any]: return { 'model': snapshot.model, 'context_window_tokens': snapshot.context_window_tokens, 'projected_input_tokens': snapshot.projected_input_tokens, 'message_tokens': snapshot.message_tokens, 'chat_overhead_tokens': snapshot.chat_overhead_tokens, 'reserved_output_tokens': snapshot.reserved_output_tokens, 'reserved_compaction_buffer_tokens': snapshot.reserved_compaction_buffer_tokens, 'reserved_schema_tokens': snapshot.reserved_schema_tokens, 'hard_input_limit_tokens': snapshot.hard_input_limit_tokens, 'soft_input_limit_tokens': snapshot.soft_input_limit_tokens, 'overflow_tokens': snapshot.overflow_tokens, 'soft_overflow_tokens': snapshot.soft_overflow_tokens, 'exceeds_hard_limit': snapshot.exceeds_hard_limit, 'exceeds_soft_limit': snapshot.exceeds_soft_limit, 'token_counter_backend': snapshot.token_counter_backend, 'token_counter_source': snapshot.token_counter_source, 'token_counter_accurate': snapshot.token_counter_accurate, } def _list_backend_models(base_url: str, api_key: str) -> dict[str, Any]: req = request.Request( _join_url(base_url, '/models'), headers={ 'Authorization': f'Bearer {api_key}', 'api-key': api_key, 'Content-Type': 'application/json', }, method='GET', ) try: with request.urlopen(req, timeout=10) as response: payload = json.loads(response.read().decode('utf-8')) except error.HTTPError as exc: detail = exc.read().decode('utf-8', errors='replace') raise HTTPException( status_code=502, detail=f'HTTP {exc.code} from model backend: {detail}', ) from exc except (error.URLError, OSError, json.JSONDecodeError) as exc: raise HTTPException( status_code=502, detail=f'Unable to list models from {base_url}: {exc}', ) from exc models = _normalize_model_list(payload) return { 'models': models, 'raw_count': _raw_model_count(payload), 'filtered_count': len(models), } def _normalize_model_list(payload: Any) -> list[dict[str, str]]: data = payload.get('data') if isinstance(payload, dict) else payload if not isinstance(data, list): return [] models_by_key: dict[str, dict[str, str]] = {} for item in data: if isinstance(item, str): normalized = _normalize_model_id(item) if normalized is not None: models_by_key.setdefault(normalized.lower(), {'id': normalized}) continue if not isinstance(item, dict): continue model_id = item.get('id') or item.get('model') or item.get('name') if not isinstance(model_id, str) or not model_id: continue model_type = _optional_model_string(item.get('model_type') or item.get('type')) if not _is_llm_model(model_id, model_type): continue provider = _optional_model_string( item.get('owned_by') or item.get('provider') or item.get('owner') ) if provider and provider not in VALIDATED_CHAT_MODEL_PROVIDERS: continue normalized = _normalize_model_id(model_id, provider=provider) if normalized is not None and normalized.lower() in UNSUPPORTED_TOOL_CHAT_MODEL_IDS: continue if normalized is not None: entry = {'id': normalized} if provider: entry['provider'] = provider if model_type: entry['model_type'] = model_type models_by_key.setdefault(normalized.lower(), entry) return sorted(models_by_key.values(), key=lambda item: item['id'].lower()) def _raw_model_count(payload: Any) -> int: data = payload.get('data') if isinstance(payload, dict) else payload return len(data) if isinstance(data, list) else 0 def _optional_model_string(value: Any) -> str | None: if not isinstance(value, str): return None stripped = value.strip() return stripped or None def _normalize_model_id(model: str, *, provider: str | None = None) -> str | None: value = model.strip() if not value: return None if provider and not value.lower().startswith(f'{provider.lower()}/'): return f'{provider}/{value}' return value def _is_llm_model(model_id: str, model_type: str | None) -> bool: lower = model_id.strip().lower() blocked_fragments = ( 'asr', 'audio', 'embedding', 'image_generation', 'rerank', 'speech', 'text2image', 'transcribe', 'translation', 'tts', 'voiceclone', 'voicedesign', ) if any(fragment in lower for fragment in blocked_fragments): return False if model_type: return model_type.strip().lower() == 'llm' return True def _join_url(base_url: str, suffix: str) -> str: return f'{base_url.rstrip("/")}/{suffix.lstrip("/")}' def _content_type_for_filename(filename: str) -> str: suffix = Path(filename).suffix.lower() if suffix == '.json': return 'application/json; charset=utf-8' if suffix in {'.jsonl', '.txt', '.md', '.csv'}: return 'text/plain; charset=utf-8' if suffix == '.xlsx': return 'application/vnd.openxmlformats-officedocument.spreadsheetml.sheet' if suffix == '.docx': return 'application/vnd.openxmlformats-officedocument.wordprocessingml.document' return 'application/octet-stream' def _ascii_download_filename(filename: str) -> str: safe = re.sub(r'[^A-Za-z0-9._-]+', '_', filename).strip('._') return safe[:120] or 'download' def _url_quote_filename(filename: str) -> str: return quote(filename, safe='') def _annotate_transcript_elapsed(payload: dict[str, Any], elapsed_ms: int) -> None: transcript = payload.get('transcript') if not isinstance(transcript, list): return for entry in reversed(transcript): if not isinstance(entry, dict) or entry.get('role') != 'assistant': continue metadata = entry.get('metadata') if not isinstance(metadata, dict): metadata = {} entry['metadata'] = metadata metadata['elapsed_ms'] = elapsed_ms return def _annotate_last_assistant_elapsed( directory: Path, session_id: str, elapsed_ms: int, ) -> None: path = _session_json_path(directory, session_id) try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return messages = data.get('messages') if not isinstance(messages, list): return for message in reversed(messages): if not isinstance(message, dict) or message.get('role') != 'assistant': continue metadata = message.get('metadata') if not isinstance(metadata, dict): metadata = {} message['metadata'] = metadata metadata['elapsed_ms'] = elapsed_ms try: path.write_text(json.dumps(data, ensure_ascii=False, indent=2), encoding='utf-8') except OSError: return return def _read_session_title(directory: Path, session_id: str | None) -> str | None: if not session_id: return None path = _session_json_path(directory, session_id) try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return None title = data.get('title') if isinstance(title, str) and title.strip(): return title.strip() return None def _ensure_session_title( directory: Path, session_id: str, *, model: str, base_url: str, api_key: str, previous_title: str | None = None, fallback_title: str | None = None, ) -> None: path = _session_json_path(directory, session_id) try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return if previous_title: data['title'] = previous_title _write_session_metadata(path, data) return title = data.get('title') title_source = data.get('title_source') if ( isinstance(title, str) and title.strip() and title_source != 'first_message' ): return messages = data.get('messages') if not isinstance(messages, list): if fallback_title and not (isinstance(title, str) and title.strip()): data['title'] = fallback_title data['title_source'] = 'first_message' data['title_generated_at'] = int(time.time()) _write_session_metadata(path, data) return user_messages = _session_user_messages(messages) if not user_messages: if fallback_title and not (isinstance(title, str) and title.strip()): data['title'] = fallback_title data['title_source'] = 'first_message' data['title_generated_at'] = int(time.time()) _write_session_metadata(path, data) return generated = _generate_session_title( user_messages, model=model, base_url=base_url, api_key=api_key, ) if not generated: if fallback_title and not (isinstance(title, str) and title.strip()): data['title'] = fallback_title data['title_source'] = 'first_message' data['title_generated_at'] = int(time.time()) _write_session_metadata(path, data) return data['title'] = generated data['title_source'] = 'llm' data['title_generated_at'] = int(time.time()) _write_session_metadata(path, data) def _session_user_messages(messages: list[Any]) -> list[str]: user_messages: list[str] = [] for message in messages: if not isinstance(message, dict) or message.get('role') != 'user': continue content = message.get('content') if not isinstance(content, str) or _is_internal_message(content): continue stripped = _strip_session_context(content) if stripped: user_messages.append(stripped) return user_messages def _generate_session_title( user_messages: list[str], *, model: str, base_url: str, api_key: str, ) -> str | None: conversation = '\n'.join( f'用户第{index}轮：{message[:240]}' for index, message in enumerate(user_messages[-6:], start=1) ) payload = { 'model': model, 'temperature': 0.2, 'max_tokens': 32, 'messages': [ { 'role': 'system', 'content': ( '你是会话标题生成器。请根据用户对话生成一个中文短标题，' '要求 4 到 12 个字，只输出标题，不要解释，不要引号。' ), }, {'role': 'user', 'content': conversation}, ], } client = OpenAICompatClient( ModelConfig( model=model, base_url=base_url, api_key=api_key, temperature=0.2, ) ) try: turn = client.complete(payload['messages'], tools=[]) except OpenAICompatError: return None return _clean_session_title(turn.content) def _clean_session_title(value: str) -> str | None: title = ' '.join(value.strip().strip('"\'“”‘’`').split()) title = title.rstrip('。.!！?？') if not title: return None return title[:24].rstrip() def _clean_manual_session_title(value: str) -> str | None: title = ' '.join(value.strip().strip('"\'“”‘’`').split()) if not title: return None return title[:80] def _write_session_metadata(path: Path, data: dict[str, Any]) -> None: try: path.write_text(json.dumps(data, ensure_ascii=False, indent=2), encoding='utf-8') except OSError: return def _session_json_path(directory: Path, session_id: str) -> Path: path = directory / session_id / 'session.json' if path.exists(): return path return directory / f'{session_id}.json' def _jupyter_binding_path(directory: Path, session_id: str) -> Path: return directory / session_id / JUPYTER_WORKSPACE_FILENAME def _save_jupyter_runtime_binding( directory: Path, runtime: JupyterRuntimeSession, ) -> None: path = _jupyter_binding_path(directory, runtime.binding.session_id) try: path.parent.mkdir(parents=True, exist_ok=True) _write_session_metadata(path, runtime.to_persisted_dict()) except OSError: return # ── Per-account remote-Jupyter workspace registry ───────────────────────────── # Saved at accounts//workspaces.json. Each entry stores enough to re-bind a # new chat without re-prompting (base_url + workspace_root + password). Cookies # expire and aren't reusable across sessions, so for one-click re-bind we store # the password — same plaintext-on-disk security posture as the per-session # jupyter_workspace.json that already keeps login cookies. def _jupyter_workspaces_registry_path(state: 'AgentState', account_id: str | None) -> Path: return state.account_paths(account_id)['base'] / JUPYTER_WORKSPACES_REGISTRY_FILENAME def _load_jupyter_workspaces_registry( state: 'AgentState', account_id: str | None ) -> list[dict[str, Any]]: path = _jupyter_workspaces_registry_path(state, account_id) try: payload = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return [] if not isinstance(payload, list): return [] out: list[dict[str, Any]] = [] for item in payload: if isinstance(item, dict) and isinstance(item.get('id'), str): out.append(item) return out def _save_jupyter_workspaces_registry( state: 'AgentState', account_id: str | None, entries: list[dict[str, Any]] ) -> None: path = _jupyter_workspaces_registry_path(state, account_id) try: path.parent.mkdir(parents=True, exist_ok=True) _write_session_metadata(path, entries) except OSError: return def _jupyter_workspace_registry_id(base_url: str, workspace_root: str) -> str: """Deterministic ID = sha1(base_url|workspace_root). Same (host, root) tuple de-duplicates rather than piling up entries on every successful bind.""" digest = hashlib.sha1( f'{base_url}|{workspace_root}'.encode('utf-8'), ).hexdigest() return digest[:16] def _public_jupyter_workspace_entry(entry: dict[str, Any]) -> dict[str, Any]: """Strip password before returning to the frontend.""" return { 'id': entry.get('id'), 'label': entry.get('label'), 'base_url': entry.get('base_url'), 'workspace_root': entry.get('workspace_root'), 'created_at': entry.get('created_at'), 'last_used_at': entry.get('last_used_at'), } def _register_jupyter_workspace( state: 'AgentState', account_id: str | None, *, base_url: str, workspace_root: str, password: str, label: str | None = None, ) -> dict[str, Any]: entries = _load_jupyter_workspaces_registry(state, account_id) workspace_id = _jupyter_workspace_registry_id(base_url, workspace_root) now = time.time() found: dict[str, Any] | None = None rest: list[dict[str, Any]] = [] for entry in entries: if entry.get('id') == workspace_id: found = entry else: rest.append(entry) if found is None: found = { 'id': workspace_id, 'label': label or base_url, 'base_url': base_url, 'workspace_root': workspace_root, 'created_at': now, } else: if label: found['label'] = label elif not found.get('label'): found['label'] = base_url found['base_url'] = base_url found['workspace_root'] = workspace_root found['password'] = password found['last_used_at'] = now rest.insert(0, found) # most-recently-used first _save_jupyter_workspaces_registry(state, account_id, rest) return found def _jupyter_runtime_for_session( state: AgentState, account_id: str | None, session_id: str, ) -> JupyterRuntimeSession | None: account_key = state._account_key(account_id) runtime = state.jupyter_runtime_manager.get(account_key, session_id) if runtime is not None: return runtime path = _jupyter_binding_path( state.account_paths(account_id)['sessions'], session_id, ) try: payload = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return None if not isinstance(payload, dict): return None try: return state.jupyter_runtime_manager.restore_session( account_id=account_key, session_id=session_id, payload=payload, ) except JupyterRuntimeError: return None def _feishu_paths(state: AgentState, account_id: str | None) -> dict[str, Path]: base = state.account_paths(account_id)['base'] / 'integrations' / 'feishu' paths = { 'root': base, 'home': base / 'home', 'config': base / 'config', 'npm_cache': base / 'npm-cache', } for path in paths.values(): path.mkdir(parents=True, exist_ok=True) return paths def _feishu_command() -> list[str]: return [ _feishu_npx_executable(), '-y', f'--registry={FEISHU_NPM_REGISTRY}', FEISHU_MCP_PACKAGE, ] def _feishu_npx_executable() -> str: explicit = os.environ.get('CLAW_NPX_BIN') if explicit and Path(explicit).expanduser().exists(): return str(Path(explicit).expanduser()) node_bin_dir = os.environ.get('CLAW_NODE_BIN_DIR') if node_bin_dir: candidate = Path(node_bin_dir).expanduser() / 'npx' if candidate.exists(): return str(candidate) return shutil.which('npx') or 'npx' def _feishu_env(paths: dict[str, Path]) -> dict[str, str]: env = { 'HOME': str(paths['home']), 'XDG_CONFIG_HOME': str(paths['config']), 'npm_config_cache': str(paths['npm_cache']), 'npm_config_update_notifier': 'false', 'NO_UPDATE_NOTIFIER': 'true', 'FEISHU_LOGIN_MODE': 'devicecode', } node_bin_dir = os.environ.get('CLAW_NODE_BIN_DIR') if node_bin_dir: env['PATH'] = f'{node_bin_dir}{os.pathsep}{os.environ.get("PATH", "")}' return env def _run_feishu_cli( paths: dict[str, Path], args: list[str], *, timeout_seconds: float, ) -> dict[str, Any]: env = os.environ.copy() env.update(_feishu_env(paths)) try: completed = subprocess.run( [*_feishu_command(), *args], stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, cwd=str(paths['root']), env=env, timeout=timeout_seconds, check=False, ) return { 'returncode': completed.returncode, 'output': completed.stdout.strip(), } except FileNotFoundError as exc: return { 'returncode': 127, 'output': f'缺少命令: {exc.filename}', } except subprocess.TimeoutExpired as exc: output = exc.stdout or '' return { 'returncode': 124, 'output': str(output).strip() or '飞书状态检查超时', } def _feishu_status_payload(state: AgentState, account_id: str | None) -> dict[str, Any]: paths = _feishu_paths(state, account_id) _reap_feishu_login(account_id) result = _run_feishu_cli(paths, ['status'], timeout_seconds=30) output = str(result.get('output') or '') lowered = output.lower() logged_in = result.get('returncode') == 0 and not any( marker in lowered for marker in ('not logged in', '未登录', 'no credentials') ) pending = _feishu_login_snapshot(account_id) payload: dict[str, Any] = { 'logged_in': logged_in, 'status': 'logged_in' if logged_in else 'not_logged_in', 'output': output, } if pending: payload['login'] = pending if not logged_in: payload['status'] = 'login_pending' if not logged_in and result.get('returncode') not in (0, 1): payload['status'] = 'error' payload['error'] = output or '无法获取飞书登录状态' return payload def _start_feishu_login(state: AgentState, account_id: str | None) -> dict[str, Any]: status = _feishu_status_payload(state, account_id) if status.get('logged_in'): return status key = _feishu_login_key(account_id) with _FEISHU_LOGIN_LOCK: existing = _FEISHU_LOGIN_PROCESSES.get(key) if existing is not None and existing.process.poll() is None: return { 'logged_in': False, 'status': 'login_pending', 'login': _feishu_login_snapshot_unlocked(existing), } _FEISHU_LOGIN_PROCESSES.pop(key, None) paths = _feishu_paths(state, account_id) env = os.environ.copy() env.update(_feishu_env(paths)) try: process = subprocess.Popen( [*_feishu_command(), 'login', '--device'], stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1, cwd=str(paths['root']), env=env, ) except FileNotFoundError as exc: raise HTTPException(status_code=500, detail=f'缺少命令: {exc.filename}') entry = FeishuLoginProcess(process=process, started_at=time.time()) with _FEISHU_LOGIN_LOCK: _FEISHU_LOGIN_PROCESSES[key] = entry reader = threading.Thread( target=_read_feishu_login_output, args=(key, entry), name=f'feishu-login-{key}', daemon=True, ) reader.start() deadline = time.time() + 5.0 while time.time() < deadline: snapshot = _feishu_login_snapshot(account_id) if snapshot and snapshot.get('login_url'): return { 'logged_in': False, 'status': 'login_pending', 'login': snapshot, } if process.poll() is not None: break time.sleep(0.1) snapshot = _feishu_login_snapshot(account_id) return { 'logged_in': False, 'status': 'login_pending' if process.poll() is None else 'login_failed', 'login': snapshot, } def _stop_feishu_login(account_id: str | None) -> None: key = _feishu_login_key(account_id) with _FEISHU_LOGIN_LOCK: entry = _FEISHU_LOGIN_PROCESSES.pop(key, None) if entry is None: return process = entry.process if process.poll() is None: process.terminate() try: process.wait(timeout=2.0) except subprocess.TimeoutExpired: process.kill() process.wait(timeout=2.0) def _reap_feishu_login(account_id: str | None) -> None: key = _feishu_login_key(account_id) with _FEISHU_LOGIN_LOCK: entry = _FEISHU_LOGIN_PROCESSES.get(key) if entry is not None and entry.process.poll() is not None: _FEISHU_LOGIN_PROCESSES.pop(key, None) def _read_feishu_login_output(key: str, entry: FeishuLoginProcess) -> None: stream = entry.process.stdout if stream is None: return for line in stream: with _FEISHU_LOGIN_LOCK: current = _FEISHU_LOGIN_PROCESSES.get(key) if current is not entry: return entry.output.append(line.rstrip()) text = '\n'.join(entry.output) login_url, user_code = _parse_feishu_login_details(text) if login_url: entry.login_url = login_url if user_code: entry.user_code = user_code def _feishu_login_snapshot(account_id: str | None) -> dict[str, Any] | None: key = _feishu_login_key(account_id) with _FEISHU_LOGIN_LOCK: entry = _FEISHU_LOGIN_PROCESSES.get(key) if entry is None: return None return _feishu_login_snapshot_unlocked(entry) def _feishu_login_snapshot_unlocked(entry: FeishuLoginProcess) -> dict[str, Any]: output = '\n'.join(entry.output[-20:]) if not entry.login_url or not entry.user_code: login_url, user_code = _parse_feishu_login_details(output) entry.login_url = entry.login_url or login_url entry.user_code = entry.user_code or user_code return { 'running': entry.process.poll() is None, 'started_at': entry.started_at, 'elapsed_seconds': max(0, int(time.time() - entry.started_at)), 'login_url': entry.login_url, 'user_code': entry.user_code, 'output': output, } def _parse_feishu_login_details(text: str) -> tuple[str | None, str | None]: urls = re.findall(r'https?://[^\s]+', text) login_url = next((url.rstrip('.,;') for url in urls if 'feishu' in url.lower()), None) user_code = None match = re.search(r'user_code=([A-Za-z0-9_-]+)', login_url or '') if match: user_code = match.group(1) if not user_code: match = re.search(r'\b([A-Z0-9]{4}-[A-Z0-9]{4})\b', text) if match: user_code = match.group(1) return login_url, user_code def _feishu_login_key(account_id: str | None) -> str: return _safe_account_id(account_id) def _feishu_server_profile(paths: dict[str, Path]) -> MCPServerProfile: return MCPServerProfile( name=FEISHU_MCP_SERVER_NAME, source_manifest='builtin:feishu-mcp-pro', transport='stdio', command='npx', args=( '-y', f'--registry={FEISHU_NPM_REGISTRY}', FEISHU_MCP_PACKAGE, ), env=_feishu_env(paths), cwd=str(paths['root']), description='账号隔离的飞书 MCP Pro。', ) def _record_feishu_online_doc( state: AgentState, account_id: str | None, *, file_path: Path, map_key: str | None = None, title: str, url: str, kind: str = 'doc', ) -> None: paths = _feishu_paths(state, account_id) map_path = paths['root'] / FEISHU_ONLINE_DOCS_FILENAME now = int(time.time()) clean_url = _clean_url(url) with _FEISHU_DOC_MAP_LOCK: try: payload = json.loads(map_path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): payload = {} files = payload.get('files') if isinstance(payload, dict) else None if not isinstance(files, dict): files = {} key = map_key or str(file_path) previous = files.get(key) created_at = ( previous.get('created_at') if isinstance(previous, dict) and isinstance(previous.get('created_at'), int) else now ) files[key] = { 'url': clean_url, 'title': title, 'kind': kind, 'file_path': str(file_path), 'source_path': key, 'created_at': created_at, 'updated_at': now, } map_path.write_text( json.dumps({'files': files}, ensure_ascii=False, indent=2), encoding='utf-8', ) def _resolve_feishu_source_file( state: AgentState, account_id: str | None, raw_path: str, ) -> tuple[Path, str | None]: if raw_path.startswith('jupyter://'): return _materialize_jupyter_file_for_feishu(state, account_id, raw_path), raw_path return _resolve_account_file(state, account_id, raw_path), None def _materialize_jupyter_file_for_feishu( state: AgentState, account_id: str | None, raw_uri: str, ) -> Path: session_id, remote_path = _parse_jupyter_file_uri(raw_uri) runtime = _jupyter_runtime_for_session(state, account_id, session_id) if runtime is None: raise HTTPException(status_code=404, detail='Jupyter runtime is not connected') try: info = runtime.file_info(remote_path) filename = _safe_uploaded_filename(str(info.get('name') or 'remote-file')) suffix = Path(filename).suffix.lower() max_bytes = ( FEISHU_REMOTE_SHEET_MAX_BYTES if suffix in FEISHU_SPREADSHEET_SUFFIXES else FEISHU_REMOTE_DOC_MAX_BYTES ) size = info.get('size') if isinstance(size, int) and size > max_bytes: raise HTTPException( status_code=413, detail=( f'远端文件过大，当前在线文档转换上限为 ' f'{_format_bytes(max_bytes)}，请先在 Jupyter 侧裁剪或抽样后再转换。' ), ) response, stream_filename = runtime.open_file_stream(remote_path) except JupyterRuntimeError as exc: raise HTTPException(status_code=400, detail=str(exc)) from exc filename = filename or _safe_uploaded_filename(stream_filename) target_dir = ( state.account_paths(account_id)['base'] / 'integrations' / 'feishu' / 'remote-files' / session_id ) target_dir.mkdir(parents=True, exist_ok=True) target = target_dir / filename written = 0 try: with target.open('wb') as handle: for chunk in response.iter_content(chunk_size=JUPYTER_STREAM_CHUNK_BYTES): if not chunk: continue written += len(chunk) if written > max_bytes: handle.close() target.unlink(missing_ok=True) raise HTTPException( status_code=413, detail=( f'远端文件过大，当前在线文档转换上限为 ' f'{_format_bytes(max_bytes)}，请先在 Jupyter 侧裁剪或抽样后再转换。' ), ) handle.write(chunk) finally: response.close() return target def _parse_jupyter_file_uri(raw_uri: str) -> tuple[str, str]: parsed = urlparse(raw_uri) session_id = _safe_session_id(parsed.netloc) remote_path = unquote(parsed.path or '') if not session_id or not remote_path.startswith('/'): raise HTTPException(status_code=400, detail='Invalid Jupyter file URI') return session_id, remote_path def _safe_uploaded_filename(filename: str) -> str: cleaned = re.sub(r'[\x00-\x1f/\\]+', '_', filename).strip(' ._') return cleaned[:180] or 'remote-file' def _format_bytes(value: int) -> str: if value >= 1024 * 1024 * 1024: return f'{value / (1024 * 1024 * 1024):.1f}GB' if value >= 1024 * 1024: return f'{value / (1024 * 1024):.0f}MB' if value >= 1024: return f'{value / 1024:.0f}KB' return f'{value}B' def _resolve_account_file(state: AgentState, account_id: str | None, raw_path: str) -> Path: path = Path(raw_path).expanduser().resolve() account_base = state.account_paths(account_id)['base'].resolve() try: path.relative_to(account_base) except ValueError: raise HTTPException(status_code=403, detail='文件不在当前账号目录内') if not path.exists(): raise HTTPException(status_code=404, detail='文件不存在') if not path.is_file(): raise HTTPException(status_code=400, detail='目标不是文件') return path def _clean_feishu_doc_title(value: str) -> str: title = re.sub(r'[\r\n\t/\\]+', ' ', value).strip() return title[:80].strip() or 'Claw 生成文档' def _create_feishu_online_spreadsheet( runtime: MCPRuntime, file_path: Path, *, title: str, folder_token: str | None = None, ) -> tuple[str, dict[str, Any]]: sheets = _load_file_as_feishu_sheets(file_path) create_params: dict[str, Any] = {'title': title} if folder_token and folder_token.strip(): create_params['folder_token'] = folder_token.strip() rendered, metadata = _call_feishu_mcp_tool_checked( runtime, 'sheet_ops', arguments={'action': 'create', 'params': create_params}, max_chars=API_TOOL_CONTENT_MAX_CHARS, timeout_seconds=60.0, ) spreadsheet_token = _extract_feishu_spreadsheet_token(rendered) if not spreadsheet_token: raise DataAgentInputError('飞书表格已创建，但没有解析到 spreadsheet token') for index, sheet_data in enumerate(sheets): rows = sheet_data['rows'] if not rows: continue if index == 0: # 新建飞书表格默认会带一个名为 Sheet1 的工作表。 # 不从 spreadsheet meta 里取顶层 title，避免把文件名误当成工作表名。 sheet_ref = 'Sheet1' else: sheet_ref = _create_feishu_worksheet( runtime, spreadsheet_token, title=str(sheet_data.get('title') or f'Sheet{index + 1}'), index=index, ) _write_feishu_sheet_rows(runtime, spreadsheet_token, sheet_ref, rows) return rendered, metadata def _call_feishu_mcp_tool_checked( runtime: MCPRuntime, tool_name: str, *, arguments: dict[str, Any], max_chars: int, timeout_seconds: float, ) -> tuple[str, dict[str, Any]]: rendered, metadata = runtime.call_tool( tool_name, arguments=arguments, server_name=FEISHU_MCP_SERVER_NAME, max_chars=max_chars, timeout_seconds=timeout_seconds, ) error_message = _extract_feishu_mcp_error(rendered) if metadata.get('is_error') or error_message: raise DataAgentInputError(error_message or f'飞书 MCP 工具 {tool_name} 调用失败') return rendered, metadata def _extract_feishu_mcp_error(text: str) -> str | None: payload = _parse_first_json_object(text) if not isinstance(payload, dict): return None error_value = payload.get('error') if not error_value: return None suggestion = payload.get('suggestion') code = payload.get('code') parts = [str(error_value)] if code is not None: parts.append(f'code={code}') if suggestion: parts.append(str(suggestion)) return '；'.join(parts) def _load_file_as_feishu_sheets(file_path: Path) -> list[dict[str, Any]]: suffix = file_path.suffix.lower() if suffix == '.csv': rows = _load_csv_rows_for_feishu(file_path) return [{'title': _clean_feishu_sheet_title(file_path.stem) or 'Sheet1', 'rows': rows}] if suffix == '.xlsx': return _load_xlsx_rows_for_feishu(file_path) raise DataAgentInputError(f'不支持转为飞书表格的文件类型: {suffix}') def _load_csv_rows_for_feishu(file_path: Path) -> list[list[str]]: rows: list[list[str]] = [] with file_path.open('r', encoding='utf-8-sig', errors='replace', newline='') as handle: reader = csv.reader(handle) for row_index, row in enumerate(reader): if row_index >= FEISHU_SPREADSHEET_MAX_ROWS: break rows.append(_clean_feishu_sheet_row(row)) normalized_rows = _normalize_feishu_sheet_rows(rows) if not normalized_rows: raise DataAgentInputError('没有解析到可转成飞书表格的数据') return normalized_rows def _load_xlsx_rows_for_feishu(file_path: Path) -> list[dict[str, Any]]: try: import openpyxl # type: ignore[import-not-found] except ImportError as exc: raise DataAgentInputError('解析 xlsx 需要 openpyxl') from exc workbook = openpyxl.load_workbook(file_path, data_only=True, read_only=True) sheets: list[dict[str, Any]] = [] for index, sheet in enumerate(workbook.worksheets): if hasattr(sheet, 'reset_dimensions'): sheet.reset_dimensions() rows: list[list[str]] = [] for row_index, row in enumerate(sheet.iter_rows(values_only=True)): if row_index >= FEISHU_SPREADSHEET_MAX_ROWS: break rows.append(_clean_feishu_sheet_row(row)) normalized_rows = _normalize_feishu_sheet_rows(rows) if normalized_rows: sheets.append( { 'title': _clean_feishu_sheet_title(str(sheet.title)) or f'Sheet{index + 1}', 'rows': normalized_rows, } ) if not sheets: raise DataAgentInputError('没有解析到可转成飞书表格的数据') return sheets def _clean_feishu_sheet_row(row: Any) -> list[str]: if not isinstance(row, (list, tuple)): return [] return [_clean_feishu_sheet_cell(cell) for cell in row[:FEISHU_SPREADSHEET_MAX_COLS]] def _clean_feishu_sheet_cell(value: Any) -> str: if value is None: return '' text = str(value) if len(text) > FEISHU_SPREADSHEET_MAX_CELL_CHARS: return text[:FEISHU_SPREADSHEET_MAX_CELL_CHARS] return text def _normalize_feishu_sheet_rows(rows: list[list[str]]) -> list[list[str]]: while rows and not any(cell for cell in rows[-1]): rows.pop() if not rows: return [] width = max((len(row) for row in rows), default=0) width = min(max(width, 1), FEISHU_SPREADSHEET_MAX_COLS) return [row[:width] + [''] * max(0, width - len(row)) for row in rows] def _clean_feishu_sheet_title(value: str) -> str: title = re.sub(r'[\r\n\t/\\?*\[\]:]+', ' ', value).strip() return title[:80].strip() def _extract_feishu_spreadsheet_token(text: str) -> str | None: url = _extract_first_url(text) candidates = [item for item in (url, text) if item] for candidate in candidates: match = re.search(r'/sheets/([A-Za-z0-9]+)', candidate) if match: return match.group(1) payload = _parse_first_json_object(text) token = _find_first_string_value( payload, {'spreadsheet_token', 'spreadsheetToken', 'token'}, ) return token.strip() if token else None def _discover_feishu_sheet_refs(runtime: MCPRuntime, spreadsheet_token: str) -> list[str]: try: rendered, _metadata = _call_feishu_mcp_tool_checked( runtime, 'sheet_ops', arguments={'action': 'meta', 'params': {'url_or_token': spreadsheet_token}}, max_chars=API_TOOL_CONTENT_MAX_CHARS, timeout_seconds=60.0, ) except Exception: return [] refs = _extract_feishu_sheet_refs(rendered) return refs def _extract_feishu_sheet_refs(text: str) -> list[str]: payload = _parse_first_json_object(text) refs: list[str] = [] def visit(value: Any) -> None: if isinstance(value, dict): raw_id = value.get('sheet_id') or value.get('sheetId') raw_title = value.get('title') or value.get('name') if isinstance(raw_id, str) and raw_id.strip(): refs.append(raw_id.strip()) elif ( isinstance(raw_title, str) and raw_title.strip() and ('index' in value or 'row_count' in value or 'col_count' in value) ): refs.append(raw_title.strip()) for child in value.values(): visit(child) elif isinstance(value, list): for child in value: visit(child) visit(payload) return _dedupe_strings([ref for ref in refs if ref]) def _dedupe_strings(values: list[str]) -> list[str]: seen: set[str] = set() result: list[str] = [] for value in values: if value in seen: continue seen.add(value) result.append(value) return result def _create_feishu_worksheet( runtime: MCPRuntime, spreadsheet_token: str, *, title: str, index: int, ) -> str: sheet_title = _clean_feishu_sheet_title(title) or f'Sheet{index + 1}' rendered, _metadata = _call_feishu_mcp_tool_checked( runtime, 'sheet_ops', arguments={ 'action': 'add_sheet', 'params': { 'spreadsheet_token': spreadsheet_token, 'title': sheet_title, 'index': index, }, }, max_chars=API_TOOL_CONTENT_MAX_CHARS, timeout_seconds=60.0, ) refs = _extract_feishu_sheet_refs(rendered) return refs[0] if refs else sheet_title def _write_feishu_sheet_rows( runtime: MCPRuntime, spreadsheet_token: str, sheet_ref: str, rows: list[list[str]], ) -> None: if not rows: return width = max(len(row) for row in rows) end_column = _spreadsheet_column_name(width) for start in range(0, len(rows), FEISHU_SPREADSHEET_WRITE_BATCH_ROWS): chunk = rows[start : start + FEISHU_SPREADSHEET_WRITE_BATCH_ROWS] start_row = start + 1 end_row = start + len(chunk) _call_feishu_mcp_tool_checked( runtime, 'sheet_ops', arguments={ 'action': 'write', 'params': { 'spreadsheet_token': spreadsheet_token, 'range': f'{sheet_ref}!A{start_row}:{end_column}{end_row}', 'values': json.dumps(chunk, ensure_ascii=False), }, }, max_chars=API_TOOL_CONTENT_MAX_CHARS, timeout_seconds=60.0, ) def _spreadsheet_column_name(index: int) -> str: if index <= 0: raise ValueError('index must be positive') name = '' current = index while current: current, remainder = divmod(current - 1, 26) name = chr(ord('A') + remainder) + name return name def _parse_first_json_object(text: str) -> Any: stripped = text.strip() candidates = [stripped] json_match = re.search(r'(\{.*\})', stripped, flags=re.DOTALL) if json_match: candidates.append(json_match.group(1)) for candidate in candidates: try: return json.loads(candidate) except json.JSONDecodeError: continue return None def _find_first_string_value(value: Any, keys: set[str]) -> str | None: if isinstance(value, dict): for key in keys: item = value.get(key) if isinstance(item, str) and item.strip(): return item for child in value.values(): found = _find_first_string_value(child, keys) if found: return found elif isinstance(value, list): for child in value: found = _find_first_string_value(child, keys) if found: return found return None def _convert_file_to_feishu_markdown(file_path: Path, *, title: str) -> str: suffix = file_path.suffix.lower() if suffix == '.md': body = file_path.read_text(encoding='utf-8', errors='replace') return _limit_feishu_markdown(f'# {title}\n\n> 来源文件：{file_path.name}\n\n{body}') if suffix == '.txt': body = file_path.read_text(encoding='utf-8', errors='replace') return _limit_feishu_markdown(f'# {title}\n\n> 来源文件：{file_path.name}\n\n{body}') loaded = load_input_sources( file_path.parent, [file_path.name], max_files=1, max_paragraphs_per_file=200, max_tables_per_file=30, max_rows_per_table=120, max_cell_chars=500, ) sources = loaded.get('sources') if not isinstance(sources, list) or not sources: raise DataAgentInputError('没有解析到可转成在线文档的内容') source = sources[0] lines = [ f'# {title}', '', f'> 来源文件：{file_path.name}', f'> 文件类型：{suffix.lstrip(".")}', ] warnings = source.get('warnings') if isinstance(warnings, list) and warnings: lines.append(f'> 解析提示：{"；".join(str(item) for item in warnings)}') for paragraph in source.get('paragraphs', []): if not isinstance(paragraph, dict): continue text = str(paragraph.get('text') or '').strip() if text: lines.extend(['', text]) for table in source.get('tables', []): if not isinstance(table, dict): continue rows = table.get('rows') if not isinstance(rows, list) or not rows: continue title_text = str(table.get('title') or '表格').strip() or '表格' lines.extend(['', f'## {title_text}', '']) lines.append(_render_markdown_table(rows)) row_count = table.get('row_count') if isinstance(row_count, int) and row_count > len(rows): lines.append(f'\n> 仅展示前 {len(rows)} 行，原表约 {row_count} 行。') markdown = '\n'.join(lines).strip() if not markdown: raise DataAgentInputError('没有解析到可转成在线文档的内容') return _limit_feishu_markdown(markdown) def _render_markdown_table(raw_rows: list[Any]) -> str: rows = [ [str(cell) for cell in row] for row in raw_rows if isinstance(row, list) ] if not rows: return '' width = max(len(row) for row in rows) normalized = [row + [''] * (width - len(row)) for row in rows] header = normalized[0] if any(cell.strip() for cell in normalized[0]) else [ f'列{index + 1}' for index in range(width) ] body = normalized[1:] if header is normalized[0] else normalized lines = [ '| ' + ' | '.join(_escape_markdown_table_cell(cell) for cell in header) + ' |', '| ' + ' | '.join('---' for _ in range(width)) + ' |', ] for row in body: lines.append('| ' + ' | '.join(_escape_markdown_table_cell(cell) for cell in row) + ' |') return '\n'.join(lines) def _escape_markdown_table_cell(value: str) -> str: return value.replace('\\', '\\\\').replace('|', '\\|').replace('\n', ' ').strip() def _limit_feishu_markdown(markdown: str) -> str: if len(markdown) <= FEISHU_DOC_MARKDOWN_MAX_CHARS: return markdown suffix = '\n\n> 内容较长，已截断后写入在线文档。' return markdown[: FEISHU_DOC_MARKDOWN_MAX_CHARS - len(suffix)].rstrip() + suffix def _extract_first_url(text: str) -> str | None: match = re.search(r'https?://[^\s<>\]\)\"\'“”‘’]+', text) return _clean_url(match.group(0)) if match else None def _clean_url(value: str) -> str: return value.strip().rstrip('.,;，。；、"\'“”‘’') def _safe_account_id(account_id: str | None) -> str: normalized = re.sub(r'[^a-zA-Z0-9._-]+', '_', (account_id or '').strip()) return normalized[:80] or 'default' _XIAOMI_EMAIL_RE = re.compile(r'^([\w.-]+)@xiaomi\.com$', re.IGNORECASE) _EMAIL_ACCOUNT_ID_RE = re.compile(r'^[a-z0-9._-]{2,40}$') def _account_id_from_email(email: str | None) -> str | None: """Validate a Xiaomi email and return its prefix as account_id. The prefix is lowercased and must satisfy the platform account_id rule: only lowercase letters / digits / dot / underscore / dash, length 2-40. Returns None if the email or prefix is invalid (caller must reject). """ if not isinstance(email, str): return None match = _XIAOMI_EMAIL_RE.match(email.strip()) if not match: return None prefix = match.group(1).lower() if not _EMAIL_ACCOUNT_ID_RE.match(prefix): return None return prefix def _admin_token() -> str: return os.environ.get('ZK_ADMIN_TOKEN') or 'admin' def _require_admin_token(token: str) -> None: if token != _admin_token(): raise HTTPException(status_code=401, detail='Admin unauthorized') def _accounts_root(state: AgentState) -> Path: return state.session_directory.parent / 'accounts' def _users_json_path(state: AgentState) -> Path: return _accounts_root(state) / 'users.json' def _auth_sessions_json_path(state: AgentState) -> Path: return _accounts_root(state) / 'auth_sessions.json' def _load_users_file(state: AgentState) -> dict[str, Any]: path = _users_json_path(state) try: payload = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): payload = {'users': []} if not isinstance(payload, dict) or not isinstance(payload.get('users'), list): return {'users': []} return payload def _save_users_file(state: AgentState, payload: dict[str, Any]) -> None: path = _users_json_path(state) path.parent.mkdir(parents=True, exist_ok=True) path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + '\n', encoding='utf-8') def _hash_admin_created_password(password: str = '123456') -> str: salt = os.urandom(16).hex() digest = hashlib.scrypt( password.encode('utf-8'), salt=salt.encode('utf-8'), n=16384, r=8, p=1, dklen=64, ).hex() return f'{salt}:{digest}' def _build_admin_summary(state: AgentState) -> dict[str, Any]: accounts = _list_admin_accounts(state) totals = { 'accounts': len(accounts), 'sessions': sum(item['session_count'] for item in accounts), 'tool_calls': sum(item['tool_calls'] for item in accounts), 'tokens': sum(item['total_tokens'] for item in accounts), } return { 'totals': totals, 'memory_queue': state.memory_manager.queue_snapshot(), 'accounts': accounts[:20], } def _list_admin_accounts(state: AgentState) -> list[dict[str, Any]]: users_payload = _load_users_file(state) known_users = { str(item.get('id') or item.get('username') or '').strip() for item in users_payload.get('users', []) if isinstance(item, dict) } accounts_root = _accounts_root(state) directory_users = ( { path.name for path in accounts_root.iterdir() if path.is_dir() and path.name not in {'__pycache__'} } if accounts_root.exists() else set() ) account_ids = sorted({item for item in known_users | directory_users if item}) result: list[dict[str, Any]] = [] for account_id in account_ids: base = accounts_root / account_id sessions_dir = base / 'sessions' session_files = _iter_session_files(sessions_dir) session_count = 0 tool_calls = 0 total_tokens = 0 input_tokens = 0 output_tokens = 0 reasoning_tokens = 0 latest_mtime = 0.0 models: dict[str, int] = {} for path in session_files: try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): continue session_count += 1 tool_calls += int(data.get('tool_calls') or 0) usage = data.get('usage') if isinstance(data.get('usage'), dict) else {} input_tokens += int(usage.get('input_tokens') or 0) output_tokens += int(usage.get('output_tokens') or 0) reasoning_tokens += int(usage.get('reasoning_tokens') or 0) total_tokens += int( usage.get('total_tokens') or (usage.get('input_tokens') or 0) + (usage.get('output_tokens') or 0) ) model_config = data.get('model_config') if isinstance(data.get('model_config'), dict) else {} model = str(model_config.get('model') or data.get('model') or 'unknown') models[model] = models.get(model, 0) + 1 latest_mtime = max(latest_mtime, _session_mtime(path)) memory_dir = base / 'memory' result.append( { 'account_id': account_id, 'registered': account_id in known_users, 'session_count': session_count, 'tool_calls': tool_calls, 'input_tokens': input_tokens, 'output_tokens': output_tokens, 'reasoning_tokens': reasoning_tokens, 'total_tokens': total_tokens, 'average_tokens_per_session': ( round(total_tokens / session_count, 1) if session_count else 0 ), 'average_tool_calls_per_session': ( round(tool_calls / session_count, 1) if session_count else 0 ), 'latest_session_at': latest_mtime, 'models': models, 'user_memory_lines': _count_optional_lines(memory_dir / USER_MEMORY_FILENAME), 'skill_memory_count': len(list((memory_dir / SKILL_MEMORY_DIRNAME).glob('*.md'))) if (memory_dir / SKILL_MEMORY_DIRNAME).exists() else 0, } ) return sorted(result, key=lambda item: item['latest_session_at'], reverse=True) def _admin_create_account(state: AgentState, account_id: str) -> dict[str, Any]: safe_id = _safe_account_id(account_id) if safe_id == 'default': raise HTTPException(status_code=400, detail='账号名不合法') users = _load_users_file(state) user_rows = users.setdefault('users', []) if any( isinstance(item, dict) and str(item.get('id') or item.get('username') or '').strip() == safe_id for item in user_rows ): raise HTTPException(status_code=400, detail='账号已存在') user_rows.append( { 'id': safe_id, 'username': safe_id, 'passwordHash': _hash_admin_created_password(), 'createdAt': datetime_utc_iso(), } ) _save_users_file(state, users) base = _accounts_root(state) / safe_id (base / 'sessions').mkdir(parents=True, exist_ok=True) state.memory_manager.ensure_account(safe_id) return {'account_id': safe_id, 'created': True, 'initial_password': '123456'} def _admin_delete_account(state: AgentState, account_id: str) -> dict[str, Any]: safe_id = _safe_account_id(account_id) users = _load_users_file(state) users['users'] = [ item for item in users.get('users', []) if not ( isinstance(item, dict) and str(item.get('id') or item.get('username') or '').strip() == safe_id ) ] _save_users_file(state, users) sessions_path = _auth_sessions_json_path(state) try: auth_sessions = json.loads(sessions_path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): auth_sessions = {'sessions': {}} if isinstance(auth_sessions.get('sessions'), dict): auth_sessions['sessions'] = { token: session for token, session in auth_sessions['sessions'].items() if not (isinstance(session, dict) and session.get('accountId') == safe_id) } sessions_path.write_text( json.dumps(auth_sessions, ensure_ascii=False, indent=2) + '\n', encoding='utf-8', ) base = _accounts_root(state) / safe_id if base.exists(): shutil.rmtree(base) state._clear_agents_for_account(safe_id) return {'account_id': safe_id, 'deleted': True} def _count_optional_lines(path: Path) -> int: try: return len(path.read_text(encoding='utf-8').splitlines()) except OSError: return 0 def datetime_utc_iso() -> str: return time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime()) def _is_internal_message(content: str) -> bool: return content.lstrip().startswith('') def _strip_session_context(content: str) -> str: marker = '\n\n[当前会话目录]' if marker in content: return content.split(marker, 1)[0].strip() marker = '\n[当前会话目录]' if marker in content: return content.split(marker, 1)[0].strip() return content.strip() def _extract_used_skill_names(transcript: Any) -> tuple[str, ...]: """从 transcript 中尽量提取实际调用过的 Skill 名称。""" names: list[str] = [] if not isinstance(transcript, (list, tuple)): return () for entry in transcript: if not isinstance(entry, dict): continue tool_calls = entry.get('tool_calls') if not isinstance(tool_calls, list): continue for call in tool_calls: if not isinstance(call, dict): continue function = call.get('function') if isinstance(call.get('function'), dict) else {} tool_name = call.get('name') or function.get('name') if str(tool_name).lower() != 'skill': continue arguments = call.get('arguments') or function.get('arguments') if isinstance(arguments, str): try: arguments = json.loads(arguments) except json.JSONDecodeError: arguments = {} if not isinstance(arguments, dict): continue for key in ('skill', 'skill_name', 'name'): value = arguments.get(key) if isinstance(value, str) and value.strip(): names.append(value.strip()) break return tuple(dict.fromkeys(names)) def _iter_session_files(directory: Path) -> list[Path]: if not directory.exists(): return [] by_session_id: dict[str, Path] = {} for path in directory.glob('*.json'): by_session_id[_session_id_from_path(path)] = path for path in directory.glob('*/session.json'): # 新目录格式包含 input/output/scratchpad，优先级高于旧的平铺 json。 by_session_id[_session_id_from_path(path)] = path return list(by_session_id.values()) def _delete_session_files(directory: Path, session_id: str) -> bool: deleted = False # 新目录格式：每个 session 独立目录，里面包含 session/input/output。 nested = directory / session_id if nested.exists(): shutil.rmtree(nested) deleted = True # 兼容早期平铺 session json，避免历史列表删除后旧数据又出现。 legacy = directory / f'{session_id}.json' if legacy.exists(): legacy.unlink() deleted = True return deleted def _update_session_title(directory: Path, session_id: str, title: str) -> bool: path = _session_json_path(directory, session_id) if not path.exists(): return False try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return False data['title'] = title data['title_source'] = 'manual' data['title_updated_at'] = int(time.time()) _write_session_metadata(path, data) return True def _update_session_training(directory: Path, session_id: str, is_training: bool) -> bool: path = _session_json_path(directory, session_id) if not path.exists(): return False try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return False data['is_training'] = is_training _write_session_metadata(path, data) return True # --------------------------------------------------------------------------- # Pipeline state watcher: agent declares 「watch」 entries in program-state.jsonl; # a backend scanner picks them up and polls the target file/dir on its own, # writing complete/failed entries when the watcher resolves. This decouples UI # state from agent's discipline (or lack thereof). # --------------------------------------------------------------------------- def _now_iso_local() -> str: from datetime import datetime, timezone return datetime.now(timezone.utc).astimezone().isoformat(timespec='seconds') def _watcher_append(path: Path, entry: dict[str, Any]) -> None: try: path.parent.mkdir(parents=True, exist_ok=True) with path.open('a', encoding='utf-8') as fp: fp.write(json.dumps(entry, ensure_ascii=False) + '\n') except OSError: return class _WatcherManager: def __init__(self) -> None: self._tasks: dict[tuple[str, str], asyncio.Task[None]] = {} def is_watching(self, session_id: str, step: str) -> bool: return (session_id, step) in self._tasks def register_file_exists( self, *, session_id: str, state_path: Path, step: str, target: Path, interval: float, timeout: float, run_id: str, agent_state: 'AgentState | None' = None, account_id: str | None = None, ) -> None: key = (session_id, step) if key in self._tasks: return try: loop = asyncio.get_running_loop() except RuntimeError: return task = loop.create_task( self._poll_file_exists( key=key, state_path=state_path, step=step, target=target, interval=max(2.0, interval), timeout=max(60.0, timeout), run_id=run_id, agent_state=agent_state, account_id=account_id, session_id=session_id, ) ) self._tasks[key] = task async def _poll_file_exists( self, *, key: tuple[str, str], state_path: Path, step: str, target: Path, interval: float, timeout: float, run_id: str, agent_state: 'AgentState | None' = None, account_id: str | None = None, session_id: str | None = None, ) -> None: try: mode = 'remote' if ( agent_state is not None and account_id is not None and session_id is not None and _jupyter_runtime_for_session(agent_state, account_id, session_id) is not None ) else 'local' await self._write_log( state_path, run_id, f'watcher 启动 step={step} target={target.name} 周期={int(interval)}s 超时={int(timeout)}s mode={mode}', agent_state=agent_state, account_id=account_id, session_id=session_id, ) start = time.monotonic() while True: exists = await self._check_target_exists( target, agent_state=agent_state, account_id=account_id, session_id=session_id, ) if exists: await self._write_step( state_path, step=step, status='complete', run_id=run_id, agent_state=agent_state, account_id=account_id, session_id=session_id, ) await self._write_log( state_path, run_id, f'✅ watcher 检测到 {target.name} 落盘 → step {step} complete', agent_state=agent_state, account_id=account_id, session_id=session_id, ) # R{n>=1} 链路：SFT 训练完后 submit_sft_via_cml.sh 内置 nohup # watcher 自动起 R{n} 评测，agent 不在回路中，没人写 # cml=running，UI 会从 sft=complete 直接跳到 cml=complete， # 中间 ~20min 评测期看起来像"流程卡死"。检测到同 session # 已注册 cml watch（说明 agent 提交 SFT 时预注册了下游评测 # 的 watch）就顺手补写 cml=running 同 run_id。 if ( step == 'sft' and session_id is not None and self.is_watching(session_id, 'cml') ): await self._write_step( state_path, step='cml', status='running', run_id=run_id, agent_state=agent_state, account_id=account_id, session_id=session_id, ) await self._write_log( state_path, run_id, '↪ sft 完成自动衔接 → step cml running ' '(submit_sft_via_cml.sh 内置 watcher 已触发 R 评测)', agent_state=agent_state, account_id=account_id, session_id=session_id, ) return if time.monotonic() - start > timeout: await self._write_step( state_path, step=step, status='failed', run_id=run_id, error=f'watcher 超时 {int(timeout)}s, target 未出现: {target}', agent_state=agent_state, account_id=account_id, session_id=session_id, ) await self._write_log( state_path, run_id, f'⚠️ watcher 超时 step={step} 未在 {int(timeout)}s 内看到 {target.name}', agent_state=agent_state, account_id=account_id, session_id=session_id, ) return await asyncio.sleep(interval) except asyncio.CancelledError: return except Exception as exc: # noqa: BLE001 await self._write_log( state_path, run_id, f'⚠️ watcher 异常 step={step}: {exc}', agent_state=agent_state, account_id=account_id, session_id=session_id, ) finally: self._tasks.pop(key, None) async def _write_step( self, state_path: Path, *, step: str, status: str, run_id: str | None = None, error: str | None = None, agent_state: 'AgentState | None' = None, account_id: str | None = None, session_id: str | None = None, ) -> None: entry: dict[str, Any] = { 'step': step, 'status': status, 'ts': _now_iso_local(), } if run_id: entry['run_id'] = run_id if error: entry['error'] = error await self._append_state_entry( state_path, entry, agent_state=agent_state, account_id=account_id, session_id=session_id, ) async def _write_log( self, state_path: Path, run_id: str, text: str, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, session_id: str | None = None, ) -> None: entry = { 'log': { 'ts': time.strftime('%H:%M:%S'), 'iter': run_id, 'text': text, } } await self._append_state_entry( state_path, entry, agent_state=agent_state, account_id=account_id, session_id=session_id, ) def cancel_all(self) -> None: for task in self._tasks.values(): task.cancel() self._tasks.clear() async def _append_state_entry( self, state_path: Path, entry: dict[str, Any], *, agent_state: 'AgentState | None' = None, account_id: str | None = None, session_id: str | None = None, ) -> None: """Append a state entry. Prefers remote (so sync doesn't overwrite our writes); falls back to local when no Jupyter binding exists.""" line = json.dumps(entry, ensure_ascii=False) + '\n' if ( agent_state is not None and account_id is not None and session_id is not None ): runtime = _jupyter_runtime_for_session( agent_state, account_id, session_id ) if runtime is not None: remote_dir = f'{runtime.binding.workspace_cwd}/output' remote_path = f'{remote_dir}/program-state.jsonl' cmd = ( f'mkdir -p {shlex.quote(remote_dir)} && ' f'printf %s {shlex.quote(line)} >> {shlex.quote(remote_path)}' ) try: await asyncio.to_thread( runtime.run_command, cmd, timeout_seconds=10.0, max_output_chars=64, ) return except Exception: # noqa: BLE001 pass # Local fallback await asyncio.to_thread(_watcher_append, state_path, entry) async def _check_target_exists( self, target: Path, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, session_id: str | None = None, ) -> bool: """File-existence check that prefers remote Jupyter when bound (so targets on /mnt/* mounts visible to the remote workspace are reachable).""" if ( agent_state is not None and account_id is not None and session_id is not None ): runtime = _jupyter_runtime_for_session( agent_state, account_id, session_id ) if runtime is not None: cmd = ( f'test -f {shlex.quote(str(target))} ' f'&& echo OK || echo MISSING' ) try: result = await asyncio.to_thread( runtime.run_command, cmd, timeout_seconds=10.0, max_output_chars=64, ) except Exception: # noqa: BLE001 return False return bool(result and result.stdout.strip().endswith('OK')) # Fallback: local file system try: return target.is_file() except OSError: return False @staticmethod def _append_step(state_path: Path, **fields: Any) -> None: entry: dict[str, Any] = {**fields, 'ts': _now_iso_local()} _watcher_append(state_path, entry) @staticmethod def _append_log(state_path: Path, run_id: str, text: str) -> None: entry = { 'log': { 'ts': time.strftime('%H:%M:%S'), 'iter': run_id, 'text': text, } } _watcher_append(state_path, entry) _watcher_manager = _WatcherManager() class _BashBackgroundManager: """Polls remote/local bg bash tasks; finalizes them and (optionally) fires an auto-resume turn when the agent has paused with the previous run already completed. Lifecycle mirrors `_WatcherManager`: one asyncio.Task per task_id, written to/read from `BashBgStore` so polling can resume after a server restart. """ POLL_INTERVAL_SECONDS = 5.0 MAX_TASK_LIFETIME_SECONDS = 6 * 3600.0 REMOTE_PROBE_TIMEOUT_SECONDS = 10.0 OUTPUT_PREVIEW_BYTES = 4096 def __init__(self) -> None: self._tasks: dict[str, asyncio.Task[None]] = {} self._chat_runner: 'Callable[[ChatRequest], dict[str, Any]] | None' = None def set_chat_runner( self, runner: 'Callable[[ChatRequest], dict[str, Any]]', ) -> None: self._chat_runner = runner async def register(self, state: 'AgentState', spec: BgTaskSpec) -> None: state.bash_bg_store.record_start(spec) if spec.task_id in self._tasks: return try: loop = asyncio.get_running_loop() except RuntimeError: return self._tasks[spec.task_id] = loop.create_task( self._poll_alive(state, spec) ) async def recover_from_store(self, state: 'AgentState') -> None: for row in state.bash_bg_store.list_active(): spec = self._spec_from_row(row) await self.register(state, spec) async def cancel(self, state: 'AgentState', task_id: str) -> bool: row = state.bash_bg_store.get(task_id) if row is None or row['status'] != 'running': return False await self._kill_remote_or_local(state, row) state.bash_bg_store.mark_killed(task_id) existing = self._tasks.pop(task_id, None) if existing is not None: existing.cancel() return True def status_query( self, state: 'AgentState', task_id: str ) -> 'BgTaskStatus | None': row = state.bash_bg_store.get(task_id) if row is None: return None preview = self._read_output_preview_sync(state, row) return BgTaskStatus( task_id=row['task_id'], status=row['status'], pid=row['pid'], started_at=row['started_at'], finished_at=row['finished_at'], exit_code=row['exit_code'], output_path=row['output_path'], output_preview=preview, auto_resumed=bool(row['auto_resumed']), ) def cancel_all(self) -> None: for task in list(self._tasks.values()): task.cancel() self._tasks.clear() async def _poll_alive(self, state: 'AgentState', spec: BgTaskSpec) -> None: try: start = time.monotonic() while True: alive = await self._check_pid_alive(state, spec) if not alive: await self._finalize(state, spec) return if time.monotonic() - start > self.MAX_TASK_LIFETIME_SECONDS: state.bash_bg_store.mark_completed( spec.task_id, exit_code=124, ) return await asyncio.sleep(self.POLL_INTERVAL_SECONDS) except asyncio.CancelledError: return except Exception as exc: # noqa: BLE001 print( f'[bg-bash] poll error task={spec.task_id}: {exc}', flush=True, ) finally: self._tasks.pop(spec.task_id, None) async def _check_pid_alive( self, state: 'AgentState', spec: BgTaskSpec ) -> bool: runtime = _jupyter_runtime_for_session( state, spec.account_key or None, spec.session_id, ) if runtime is not None: cmd = ( f'kill -0 {spec.pid} 2>/dev/null && echo alive || echo dead' ) try: result = await asyncio.to_thread( runtime.run_command, cmd, timeout_seconds=self.REMOTE_PROBE_TIMEOUT_SECONDS, max_output_chars=64, cancel_event=None, ) except Exception: # noqa: BLE001 # Treat probe failure as "still running" — don't finalize on # transient runtime errors. The 6h lifetime cap is the backstop. return True return 'alive' in result.stdout # Local fallback. Note: the spawned child is detached via # start_new_session=True but the FastAPI process never reaps it, so # after exit it becomes a zombie and kill(pid, 0) keeps reporting it # alive. The MAX_TASK_LIFETIME_SECONDS cap is the eventual backstop; # production runs use jupyter_runtime where the child belongs to a # different process tree. try: os.kill(spec.pid, 0) return True except (ProcessLookupError, PermissionError): return False except OSError: return True async def _finalize(self, state: 'AgentState', spec: BgTaskSpec) -> None: exit_code = await self._read_exit_code(state, spec) state.bash_bg_store.mark_completed(spec.task_id, exit_code=exit_code) if not spec.wait_for_completion: return await self._maybe_auto_resume(state, spec, exit_code) async def _read_exit_code( self, state: 'AgentState', spec: BgTaskSpec ) -> int: runtime = _jupyter_runtime_for_session( state, spec.account_key or None, spec.session_id, ) if runtime is not None: cmd = f'cat {shlex.quote(spec.exit_code_path)} 2>/dev/null || echo' try: result = await asyncio.to_thread( runtime.run_command, cmd, timeout_seconds=self.REMOTE_PROBE_TIMEOUT_SECONDS, max_output_chars=64, cancel_event=None, ) text = result.stdout.strip() return int(text) if text else -1 except (ValueError, Exception): # noqa: BLE001 return -1 try: text = Path(spec.exit_code_path).read_text( encoding='utf-8', ).strip() return int(text) if text else -1 except (OSError, ValueError): return -1 async def _read_output_preview_async( self, state: 'AgentState', spec_row: 'dict[str, Any] | BgTaskSpec' ) -> str: if isinstance(spec_row, BgTaskSpec): account = spec_row.account_key session = spec_row.session_id output_path = spec_row.output_path else: account = spec_row['account_key'] session = spec_row['session_id'] output_path = spec_row['output_path'] runtime = _jupyter_runtime_for_session(state, account or None, session) if runtime is not None: cmd = ( f'tail -c {self.OUTPUT_PREVIEW_BYTES} ' f'{shlex.quote(output_path)} 2>/dev/null || true' ) try: result = await asyncio.to_thread( runtime.run_command, cmd, timeout_seconds=self.REMOTE_PROBE_TIMEOUT_SECONDS, max_output_chars=self.OUTPUT_PREVIEW_BYTES + 256, cancel_event=None, ) return result.stdout.strip() except Exception: # noqa: BLE001 return '(读取远端输出失败)' try: text = Path(output_path).read_text( encoding='utf-8', errors='replace', ) return text[-self.OUTPUT_PREVIEW_BYTES:].strip() except OSError: return '(无法读取输出文件)' def _read_output_preview_sync( self, state: 'AgentState', row: dict[str, Any] ) -> str: runtime = _jupyter_runtime_for_session( state, row['account_key'] or None, row['session_id'], ) if runtime is not None: cmd = ( f'tail -c {self.OUTPUT_PREVIEW_BYTES} ' f'{shlex.quote(row["output_path"])} 2>/dev/null || true' ) try: result = runtime.run_command( cmd, timeout_seconds=self.REMOTE_PROBE_TIMEOUT_SECONDS, max_output_chars=self.OUTPUT_PREVIEW_BYTES + 256, cancel_event=None, ) return result.stdout.strip() except Exception: # noqa: BLE001 return '(读取远端输出失败)' try: text = Path(row['output_path']).read_text( encoding='utf-8', errors='replace', ) return text[-self.OUTPUT_PREVIEW_BYTES:].strip() except OSError: return '(无法读取输出文件)' async def _kill_remote_or_local( self, state: 'AgentState', row: dict[str, Any] ) -> None: runtime = _jupyter_runtime_for_session( state, row['account_key'] or None, row['session_id'], ) pid = int(row['pid']) if runtime is not None: cmd = ( f'kill -- -$(ps -o pgid= -p {pid} 2>/dev/null ' f'| tr -d " ") 2>/dev/null; ' f'kill {pid} 2>/dev/null; true' ) try: await asyncio.to_thread( runtime.run_command, cmd, timeout_seconds=self.REMOTE_PROBE_TIMEOUT_SECONDS, max_output_chars=64, cancel_event=None, ) except Exception: # noqa: BLE001 pass return try: os.killpg(os.getpgid(pid), signal.SIGTERM) except (ProcessLookupError, PermissionError, OSError): try: os.kill(pid, signal.SIGTERM) except (ProcessLookupError, PermissionError, OSError): pass async def _maybe_auto_resume( self, state: 'AgentState', spec: BgTaskSpec, exit_code: int, ) -> None: runner = self._chat_runner if runner is None: return account_key = spec.account_key or state._account_key(None) session_id = spec.session_id if not session_id: return snapshot = state.run_state_store.snapshot_latest( account_key, session_id, ) if not snapshot or snapshot.get('status') != 'completed': return run_lock = state.run_lock_for(account_key, session_id) if not run_lock.acquire(blocking=False): return run_lock.release() if not state.bash_bg_store.mark_auto_resumed(spec.task_id): return output_preview = await self._read_output_preview_async(state, spec) prompt = ( f'[system] 后台任务 {spec.task_id} 已结束，' f'exit_code={exit_code}\n' f'命令：{spec.command}\n' f'输出预览（最后 {self.OUTPUT_PREVIEW_BYTES} 字节）：\n' f'{output_preview}' ) request = ChatRequest( prompt=prompt, account_id=spec.account_key or None, session_id=session_id, resume_session_id=session_id, ) loop = asyncio.get_running_loop() loop.run_in_executor(None, _safe_run_chat_payload, runner, request) @staticmethod def _spec_from_row(row: dict[str, Any]) -> BgTaskSpec: return BgTaskSpec( task_id=row['task_id'], account_key=row['account_key'] or '', session_id=row['session_id'] or '', run_id=row['run_id'] or '', pid=int(row['pid']), task_dir=row['task_dir'], output_path=row['output_path'], pid_path=row['pid_path'], exit_code_path=row['exit_code_path'], command=row['command'], started_at=float(row['started_at']), wait_for_completion=bool(row['wait_for_completion']), ) _bash_bg_manager = _BashBackgroundManager() def _safe_run_chat_payload( runner: 'Callable[[ChatRequest], dict[str, Any]]', request: ChatRequest, ) -> None: """Fire-and-forget wrapper around the chat payload runner for auto-resume. Runs in the default thread pool. Any exception is logged but swallowed — the bg task is already finalized in SQLite, so the user can still bash_status / 继续 manually if the auto-resume fails. """ try: runner(request) except Exception as exc: # noqa: BLE001 print( f'[bg-bash] auto-resume failed session={request.session_id}: {exc}', flush=True, ) def _sync_remote_program_state( state: AgentState, account_id: str, session_id: str, local_state_path: Path, ) -> None: """If a Jupyter binding exists, cat the remote program-state.jsonl into local. Backend reads only the local file; this keeps both in sync so agent writes on remote (where its bash runs) are visible to UI/watcher logic.""" runtime = _jupyter_runtime_for_session(state, account_id, session_id) if runtime is None: return remote_state_path = ( f'{runtime.binding.workspace_cwd}/output/program-state.jsonl' ) try: result = runtime.run_command( f'test -f {shlex.quote(remote_state_path)} && ' f'cat {shlex.quote(remote_state_path)} || true', timeout_seconds=10.0, max_output_chars=200_000, ) except Exception: # noqa: BLE001 return if not result or result.exit_code != 0: return new_content = result.stdout or '' if not new_content.strip(): return # Skip rewrite if local already matches remote (avoid touching mtime, which # SSE uses to detect changes). try: if local_state_path.is_file() and ( local_state_path.read_text(encoding='utf-8') == new_content ): return except OSError: pass try: local_state_path.parent.mkdir(parents=True, exist_ok=True) tmp = local_state_path.with_suffix('.jsonl.sync-tmp') tmp.write_text(new_content, encoding='utf-8') tmp.replace(local_state_path) except OSError: return def _scan_for_watchers(state: AgentState) -> None: """Synchronous scan body — used as fallback when no event loop is available. Prefer _scan_for_watchers_async in normal operation so blocking calls (remote sync, LLM) run in worker threads instead of the loop.""" accounts_root = state.session_directory.parent / 'accounts' if not accounts_root.is_dir(): return for account_dir in accounts_root.iterdir(): if not account_dir.is_dir(): continue account_id = account_dir.name sessions_dir = account_dir / 'sessions' if not sessions_dir.is_dir(): continue for session_dir in sessions_dir.iterdir(): session_id = session_dir.name state_path = session_dir / 'output' / 'program-state.jsonl' try: _sync_remote_program_state( state, account_id, session_id, state_path ) except Exception as exc: # noqa: BLE001 print( f'[scanner] remote sync failed for {account_id}/{session_id}: {exc}', flush=True, ) if not state_path.is_file(): continue entries, _ = _read_program_state(state_path) last_status_per_step: dict[str, str] = {} for entry in entries: step = entry.get('step') status = entry.get('status') if isinstance(step, str) and step and isinstance(status, str): last_status_per_step[step] = status # See note in _scan_for_watchers_async: only the latest watch # entry per step is current. latest_watch_per_step: dict[str, dict[str, Any]] = {} latest_running_run_id: dict[str, str] = {} for entry in entries: watch = entry.get('watch') if isinstance(watch, dict): w_step = watch.get('step') if isinstance(w_step, str) and w_step: latest_watch_per_step[w_step] = watch continue e_step = entry.get('step') e_status = entry.get('status') e_run = entry.get('run_id') if ( isinstance(e_step, str) and e_step and e_status == 'running' and isinstance(e_run, str) and e_run ): latest_running_run_id[e_step] = e_run for step, watch in latest_watch_per_step.items(): if last_status_per_step.get(step) in { 'complete', 'failed', 'cancelled', }: continue if _watcher_manager.is_watching(session_id, step): continue kind = watch.get('kind', 'file_exists') if kind != 'file_exists': continue path_str = watch.get('path') if not isinstance(path_str, str) or not path_str: continue try: interval = float(watch.get('interval', 30) or 30) timeout = float(watch.get('timeout', 3600) or 3600) except (TypeError, ValueError): interval = 30.0 timeout = 3600.0 run_id = str(watch.get('run_id') or '?') running_run_id = latest_running_run_id.get(step) if running_run_id and run_id != running_run_id: continue _watcher_manager.register_file_exists( session_id=session_id, state_path=state_path, step=step, target=Path(path_str), interval=interval, timeout=timeout, run_id=run_id, agent_state=state, account_id=account_id, ) async def _scan_for_watchers_async(state: AgentState) -> None: """Async-friendly scan: runs blocking I/O (remote sync, LLM target_set extraction) in worker threads so the event loop stays responsive for SSE.""" accounts_root = state.session_directory.parent / 'accounts' if not accounts_root.is_dir(): return try: account_dirs = list(accounts_root.iterdir()) except OSError: return for account_dir in account_dirs: if not account_dir.is_dir(): continue account_id = account_dir.name sessions_dir = account_dir / 'sessions' if not sessions_dir.is_dir(): continue try: session_dirs = list(sessions_dir.iterdir()) except OSError: continue for session_dir in session_dirs: session_id = session_dir.name state_path = session_dir / 'output' / 'program-state.jsonl' # Remote sync runs HTTP/WS — blocking. Off the loop. try: await asyncio.to_thread( _sync_remote_program_state, state, account_id, session_id, state_path, ) except Exception as exc: # noqa: BLE001 print( f'[scanner] remote sync failed for {account_id}/{session_id}: {exc}', flush=True, ) if not state_path.is_file(): continue try: entries, _ = await asyncio.to_thread( _read_program_state, state_path ) except Exception: # noqa: BLE001 continue # Refresh LLM-derived target_set cache (also off the loop). trigger_text = await asyncio.to_thread( _extract_trigger_from_session, sessions_dir, session_id ) if trigger_text: try: await asyncio.to_thread( _refresh_target_set_cache_blocking, state.model_config_for(account_id), trigger_text, ) except Exception as exc: # noqa: BLE001 print( f'[scanner] target_set LLM refresh failed: {exc}', flush=True, ) # Watch registration is fast (just spawns asyncio tasks); stay on loop. last_status_per_step: dict[str, str] = {} for entry in entries: step = entry.get('step') status = entry.get('status') if isinstance(step, str) and step and isinstance(status, str): last_status_per_step[step] = status # Only the LAST watch entry per step represents the current round's # intent. Earlier watch entries from previous rounds must NOT be # re-registered when a new round resets the step to running, or # the resulting watcher will fire with a stale run_id and the UI # will never see a complete tagged with the current round. latest_watch_per_step: dict[str, dict[str, Any]] = {} latest_running_run_id: dict[str, str] = {} for entry in entries: watch = entry.get('watch') if isinstance(watch, dict): w_step = watch.get('step') if isinstance(w_step, str) and w_step: latest_watch_per_step[w_step] = watch continue e_step = entry.get('step') e_status = entry.get('status') e_run = entry.get('run_id') if ( isinstance(e_step, str) and e_step and e_status == 'running' and isinstance(e_run, str) and e_run ): latest_running_run_id[e_step] = e_run for step, watch in latest_watch_per_step.items(): if last_status_per_step.get(step) in { 'complete', 'failed', 'cancelled', }: continue if _watcher_manager.is_watching(session_id, step): continue kind = watch.get('kind', 'file_exists') if kind != 'file_exists': continue path_str = watch.get('path') if not isinstance(path_str, str) or not path_str: continue try: interval_s = float(watch.get('interval', 30) or 30) timeout_s = float(watch.get('timeout', 3600) or 3600) except (TypeError, ValueError): interval_s = 30.0 timeout_s = 3600.0 run_id = str(watch.get('run_id') or '?') # If a newer round has written `step=, status=running` # but the matching watch entry hasn't landed yet, skip and # wait for the next scan rather than spawning a watcher tied # to the previous round's run_id. running_run_id = latest_running_run_id.get(step) if running_run_id and run_id != running_run_id: continue _watcher_manager.register_file_exists( session_id=session_id, state_path=state_path, step=step, target=Path(path_str), interval=interval_s, timeout=timeout_s, run_id=run_id, agent_state=state, account_id=account_id, ) async def _watcher_scanner_loop( state: AgentState, interval: float = 5.0 ) -> None: """Periodically scan all session program-state.jsonl files for new watch declarations and spawn watchers for them. Blocking calls run in threads.""" while True: try: await _scan_for_watchers_async(state) except asyncio.CancelledError: return except Exception as exc: # noqa: BLE001 print(f'[watcher-scanner] scan error: {exc}', flush=True) try: await asyncio.sleep(interval) except asyncio.CancelledError: return def _session_state_path(sessions_dir: Path, session_id: str | None) -> Path | None: """Resolve //output/program-state.jsonl. Returns None if unsafe id.""" safe_id = _safe_session_id(session_id) if safe_id is None: return None return sessions_dir / safe_id / 'output' / 'program-state.jsonl' def _read_program_state(path: Path | None) -> tuple[list[dict[str, Any]], float | None]: """Returns (entries, mtime_seconds). Empty if path is None / missing / unreadable.""" if path is None: return [], None try: if not path.is_file(): return [], None st = path.stat() text = path.read_text(encoding='utf-8', errors='replace') except (FileNotFoundError, OSError, PermissionError): return [], None entries: list[dict[str, Any]] = [] for raw in text.splitlines(): line = raw.strip() if not line or line.startswith('#'): continue try: obj = json.loads(line) except json.JSONDecodeError: continue if isinstance(obj, dict): entries.append(obj) return entries, st.st_mtime def _normalize_step_token(value: str) -> str: return re.sub(r'[\s_\-/.]+', '', value.strip().lower()) def _step_matches_card(step: str, card: dict[str, Any]) -> bool: s = _normalize_step_token(step) if not s: return False key = _normalize_step_token(str(card.get('key') or '')) if s == key: return True title = _normalize_step_token(str(card.get('title') or '')) if title and (s in title or title.startswith(s) or s in key): return True return False # Round-section pipeline data model. # # Old: 3 fixed phases (eval / intervene / train) with cards dynamically injected. # New: per-round numbered sections (R0·Baseline → R1·Train → R1·Analysis → ...), # with optional HiTL gates inserted between rounds when agent writes # step:"human-check" or step:"human-review". # # Each step belongs to a "section type" (baseline / train / analysis): # - R0 has only one section: R0·Baseline (cml/gold-drift/dist-analysis/report/hypothesis/log) # - R{n>=1} has two sections: R{n}·Train (augment/verify/sft) # R{n}·Analysis (cml/gold-drift/dist-analysis/report/hypothesis/log) # # 'hypothesis' is classified as 'analysis' (not 'train'): forming the next-round # hypothesis is the conclusion of the current round's analysis cycle, not the # start of the next round's training. Putting it in analysis makes the gate # (Human Check / Review) — which sorts after all sections of run_id=R{n} but # before R{n+1}·Train — block training kickoff cleanly. # # 'next-round' is a boundary marker — does not produce a card. # 'human-check' / 'human-review' produce gate items, not cards. # step → kind: 'analysis' | 'train' | 'gate' | 'boundary' _STEP_KIND: dict[str, str] = { 'cml': 'analysis', 'gold-drift': 'analysis', 'dist-analysis': 'analysis', 'hypothesis': 'analysis', 'log': 'analysis', 'augment': 'train', 'verify': 'train', 'sft': 'train', 'human-check': 'gate', 'human-review': 'gate', 'next-round': 'boundary', } _GATE_TITLES: dict[str, dict[str, str]] = { 'human-check': { 'title': '人工确认', 'description': '已生成本轮假设与候选，等待你点头进入下一轮训练', }, 'human-review': { 'title': '人工复核', 'description': '触发了 §4.0.1 / Gold-drift 规则，需要你逐条审一下再继续', }, } # (step, section_type) → display metadata. section_type is 'baseline' for R0 # steps, 'train'/'analysis' for R{n>=1}. _CARD_META: dict[tuple[str, str], dict[str, Any]] = { # R0·Baseline ('cml', 'baseline'): {'icon': 'bar-chart', 'title': 'Baseline 评测', 'subtitle': 'CML workflow metric_diff'}, ('gold-drift', 'baseline'): {'icon': 'alert-triangle', 'title': 'Gold Drift 检查', 'subtitle': 'drift_.json'}, ('dist-analysis', 'baseline'): {'icon': 'trending-up', 'title': '分层结果分析 & 报告', 'subtitle': '根因归类 + workflow.md'}, ('hypothesis', 'baseline'): {'icon': 'dna', 'title': '形成假设', 'subtitle': 'iteration_log hypothesis 字段'}, ('log', 'baseline'): {'icon': 'clipboard-list', 'title': '记录迭代日志', 'subtitle': 'iteration_log.jsonl'}, # R{n}·Train ('augment', 'train'): {'icon': 'flask', 'title': '数据增强', 'subtitle': 'augment_.jsonl'}, ('verify', 'train'): {'icon': 'shield', 'title': '修改返回验证', 'subtitle': 'sanity + 人审'}, ('sft', 'train'): {'icon': 'graduation-cap','title': 'SFT 训练', 'subtitle': 'submit_sft_via_cml.sh'}, # R{n}·Analysis ('cml', 'analysis'): {'icon': 'bar-chart', 'title': 'CML 评测', 'subtitle': 'workflow metric_diff'}, ('gold-drift', 'analysis'): {'icon': 'alert-triangle', 'title': 'Gold Drift 检查', 'subtitle': 'drift_.json'}, ('dist-analysis', 'analysis'): {'icon': 'trending-up', 'title': '分层结果分析 & 报告', 'subtitle': '根因归类 + workflow.md'}, ('hypothesis', 'analysis'): {'icon': 'dna', 'title': '形成假设', 'subtitle': 'iteration_log hypothesis 字段'}, ('log', 'analysis'): {'icon': 'clipboard-list', 'title': '记录迭代日志', 'subtitle': 'iteration_log.jsonl'}, } # Order within a section. _CARD_ORDER: dict[tuple[str, str], int] = { # baseline ('cml', 'baseline'): 0, ('gold-drift', 'baseline'): 1, ('dist-analysis', 'baseline'): 2, ('hypothesis', 'baseline'): 3, ('log', 'baseline'): 4, # train ('augment', 'train'): 0, ('verify', 'train'): 1, ('sft', 'train'): 2, # analysis ('cml', 'analysis'): 0, ('gold-drift', 'analysis'): 1, ('dist-analysis', 'analysis'): 2, ('hypothesis', 'analysis'): 3, ('log', 'analysis'): 4, } _SECTION_LABELS: dict[str, str] = { 'baseline': 'Baseline', 'train': 'Train', 'analysis': 'Analysis', } _SECTION_DESCRIPTIONS: dict[str, str] = { 'baseline': 'Establish baseline performance and metrics', 'train': 'Form hypothesis & retrain on augmented data', 'analysis': 'Evaluate new checkpoint & diagnose regression', } def _parse_run_index(run_id: str | None) -> int | None: """'R0' → 0, 'R12' → 12, anything else → None.""" if not isinstance(run_id, str): return None m = re.match(r'^[Rr](\d+)$', run_id.strip()) if not m: return None try: return int(m.group(1)) except ValueError: return None def _infer_run_id( entry: dict[str, Any], iteration_log_count: int, *, same_step_hint: str | None = None, ) -> str | None: """Resolve which round this state entry belongs to. 1. explicit entry.run_id wins — UNLESS it's semantically invalid (a train step tagged R0; R0 is baseline-only by design). In that case we treat the explicit value as a typo and fall through to inference. 2. else: same_step_hint (run_id from nearest same-step entry) — handles stateless writes like the watcher's complete/failed entries that lack run_id but should obviously inherit the round of the agent-tagged cml/running line they accompany. 3. else: derive from kind + iteration_log line count - analysis steps: evaluating R{count}'s output → R{count} - train steps: preparing R{count+1} → R{count+1} - gates / boundary: cannot infer alone — caller back-fills from neighbor """ step = entry.get('step') kind = _STEP_KIND.get(step) if isinstance(step, str) else None explicit = entry.get('run_id') if isinstance(explicit, str) and explicit.strip(): explicit_idx = _parse_run_index(explicit) if explicit_idx is not None: # R0+train is a common agent typo: hypothesis/augment/verify/sft are # "preparing R1", but agents sometimes tag them with the round they # see in the report. We rewrite to R1 specifically (NOT current-time # inference) so retroactive iteration_log changes don't shift the # cards into R{n>=2}. if kind == 'train' and explicit_idx == 0: return 'R1' return explicit.strip() if not isinstance(step, str): return None if same_step_hint: return same_step_hint if kind == 'analysis': return f'R{iteration_log_count}' if kind == 'train': return f'R{max(iteration_log_count, 0) + 1}' return None def _build_same_step_hints( state_entries: list[dict[str, Any]], ) -> list[str | None]: """For each entry, find the nearest same-step entry (by index distance) that carries an explicit, semantically-valid run_id, and return its run_id. Used as a fallback for stateless writes (e.g. watcher's complete line lacks run_id but the agent's cml/running line right above has it). Returns a parallel list aligned to state_entries; None for entries with no usable neighbor. """ explicit_by_step: dict[str, list[tuple[int, str]]] = {} for idx, entry in enumerate(state_entries): step = entry.get('step') if not isinstance(step, str) or not step: continue kind = _STEP_KIND.get(step) if kind in (None, 'boundary', 'gate'): continue rid = entry.get('run_id') if not (isinstance(rid, str) and rid.strip()): continue parsed = _parse_run_index(rid) if parsed is None: continue # Apply same R0+train typo rewrite as _infer_run_id so the hint stays # consistent with what _infer_run_id would have returned for that # entry on its own. canonical = 'R1' if (kind == 'train' and parsed == 0) else rid.strip() explicit_by_step.setdefault(step, []).append((idx, canonical)) hints: list[str | None] = [None] * len(state_entries) for idx, entry in enumerate(state_entries): step = entry.get('step') if not isinstance(step, str) or not step: continue candidates = explicit_by_step.get(step) if not candidates: continue best = min(candidates, key=lambda ix_rid: abs(ix_rid[0] - idx)) hints[idx] = best[1] return hints def _section_type_for(step: str, run_index: int) -> str | None: """Which section a card belongs to within its round. Train steps under R0 are filtered out by `_infer_run_id` (R0 has no train phase). If one slips through here, we drop it rather than aliasing to baseline — silently mixing train cards into a baseline section is what caused hypothesis to render under R0·Baseline. """ kind = _STEP_KIND.get(step) if kind == 'analysis': return 'baseline' if run_index == 0 else 'analysis' if kind == 'train': return 'train' if run_index >= 1 else None return None def _derive_section_status(card_statuses: list[str]) -> str: if not card_statuses: return 'pending' if all(s == 'complete' for s in card_statuses): return 'complete' if 'failed' in card_statuses and not any(s in ('running', 'complete') for s in card_statuses): return 'failed' if 'running' in card_statuses or 'complete' in card_statuses: return 'running' return 'pending' def _build_pipeline_items( state_entries: list[dict[str, Any]], iteration_log_count: int, ) -> list[dict[str, Any]]: """Assemble the items[] payload from raw state entries. Returns an ordered list of {type:'round', ...} and {type:'gate', ...} items in display order. Each item shows agent-written progress only — sections not yet touched do not appear. """ # 1. Walk entries: keep latest per step (per round-namespaced key). # We namespace by run_id so the same step name on different rounds # doesn't overwrite each other. latest_step_by_key: dict[str, dict[str, Any]] = {} latest_gate_by_key: dict[str, dict[str, Any]] = {} section_first_ts: dict[tuple[str, str], str] = {} # (run_id, section_type) → earliest ts last_non_gate_run_id: str | None = None gate_position_index: dict[str, int] = {} # gate_key → position in entries (for ordering) same_step_hints = _build_same_step_hints(state_entries) for idx, entry in enumerate(state_entries): step = entry.get('step') if not isinstance(step, str) or not step: continue kind = _STEP_KIND.get(step) if kind == 'boundary': continue if kind == 'gate': run_id = entry.get('run_id') if not (isinstance(run_id, str) and run_id.strip()): run_id = last_non_gate_run_id if not run_id: continue run_id = run_id.strip() gate_key = f'{run_id}:{step}' existing = latest_gate_by_key.get(gate_key) if existing is None: latest_gate_by_key[gate_key] = { **entry, '_run_id': run_id, '_first_ts': entry.get('ts', ''), } gate_position_index[gate_key] = idx else: existing.update({k: v for k, v in entry.items() if k != 'ts'}) # keep first ts for ordering continue # analysis or train run_id = _infer_run_id( entry, iteration_log_count, same_step_hint=same_step_hints[idx] ) if not run_id: continue last_non_gate_run_id = run_id run_index = _parse_run_index(run_id) if run_index is None: continue section_type = _section_type_for(step, run_index) if section_type is None: continue card_key = f'{run_id}:{step}' ts = entry.get('ts', '') if card_key not in latest_step_by_key: latest_step_by_key[card_key] = { **entry, '_run_id': run_id, '_section_type': section_type, '_first_ts': ts, } else: existing = latest_step_by_key[card_key] for k, v in entry.items(): existing[k] = v section_key = (run_id, section_type) if section_key not in section_first_ts: section_first_ts[section_key] = ts # 2. Group cards by section. sections: dict[tuple[str, str], dict[str, Any]] = {} for card_key, entry in latest_step_by_key.items(): run_id = entry['_run_id'] section_type = entry['_section_type'] section_key = (run_id, section_type) if section_key not in sections: sections[section_key] = { 'run_id': run_id, 'section_type': section_type, 'cards': [], 'first_ts': section_first_ts.get(section_key, ''), } step = entry['step'] meta = _CARD_META.get((step, section_type), { 'icon': 'clock', 'title': step, 'subtitle': '', }) order = _CARD_ORDER.get((step, section_type), 999) card: dict[str, Any] = { 'key': card_key, 'step': step, 'icon': meta['icon'], 'title': meta['title'], 'subtitle': meta['subtitle'], 'status': entry.get('status') or 'pending', '_order': order, } if 'progress' in entry: try: value = float(entry['progress']) if value > 1: value /= 100 card['progress'] = max(0.0, min(1.0, value)) except (TypeError, ValueError): pass sections[section_key]['cards'].append(card) # 3. Sort cards within each section. for sec in sections.values(): sec['cards'].sort(key=lambda c: c.get('_order', 999)) for c in sec['cards']: c.pop('_order', None) # 4. Build round items in display order. # Order: by run_index ascending, then within same round: train before analysis. section_order = {'baseline': 0, 'train': 0, 'analysis': 1} section_items: list[dict[str, Any]] = [] for (run_id, section_type), sec in sections.items(): run_index = _parse_run_index(run_id) if run_index is None: continue statuses = [c['status'] for c in sec['cards']] round_status = _derive_section_status(statuses) label = f'{run_id} · {_SECTION_LABELS[section_type]}' section_items.append({ 'type': 'round', 'run_id': run_id, 'run_index': run_index, 'section_type': section_type, 'label': label, 'description': _SECTION_DESCRIPTIONS[section_type], 'status': round_status, 'cards': sec['cards'], '_order_key': (run_index, section_order.get(section_type, 9), sec.get('first_ts', '')), }) # 5. Build gate items. gate_items: list[dict[str, Any]] = [] for gate_key, entry in latest_gate_by_key.items(): run_id = entry['_run_id'] run_index = _parse_run_index(run_id) if run_index is None: continue step = entry['step'] meta_g = _GATE_TITLES.get(step, {'title': step, 'description': ''}) # Place gate after the latest section of this run_id (so use a high # secondary order). Use position index for tie-break across same round. raw_status = entry.get('status') or 'pending' # Gate running → 'waiting': the agent has surfaced the checkpoint and # ended its turn. There's no work in flight on the agent side; we're # waiting on the user. The frontend renders 'waiting' with a distinct # amber pill ("等待人工确认") instead of the spinning blue "IN PROGRESS" # used for actual running steps. gate_status = 'waiting' if raw_status == 'running' else raw_status # Structured fields for the gate body — agent should prefer these so # the UI can render labeled rows (现状 / 提议 / 请选). `reason` is kept # as a free-text fallback for entries that haven't been migrated. gate_items.append({ 'type': 'gate', 'key': gate_key, 'run_id': run_id, 'run_index': run_index, 'gate_kind': step, 'title': meta_g['title'], 'description': meta_g['description'], 'status': gate_status, 'reason': entry.get('reason'), 'summary': entry.get('summary'), 'proposal': entry.get('proposal'), 'ask': entry.get('ask'), '_order_key': (run_index, 99, gate_position_index.get(gate_key, 0)), }) # 6. Filter: drop empty/pending sections (no cards or all pending). Running # and complete sections both render — RoundSection's status pill conveys # the difference, so the transition complete→running on the same numbered # box is in-place rather than a separate chip strip popping in/out. section_items = [ s for s in section_items if s['status'] in ('running', 'complete', 'failed') ] # 7. Merge and sort all items by (run_index, section_order, ts/position). all_items = section_items + gate_items all_items.sort(key=lambda x: x['_order_key']) # 8. Attach 1-based numeric index for display ("01", "02", ...) after # filtering so numbering is dense (no gaps from hidden running sections). for i, item in enumerate(all_items, start=1): item['index'] = i item.pop('_order_key', None) return all_items def _compute_in_flight( state_entries: list[dict[str, Any]], iteration_log_count: int, ) -> dict[str, Any] | None: """Render the in-progress phase as an inline chip row. The wrapping section only appears once the whole phase is COMPLETE. Until then we still want the user to see what's been done in the current phase: every step that has reached `complete` plus the one currently `running`. Pending steps stay hidden — they appear when their turn comes and replace nothing. Returns a single object describing the in-progress phase + its progressively-filling cards, in canonical step order. None when no phase is currently running (everything settled, or the latest signal is a gate/boundary). """ same_step_hints = _build_same_step_hints(state_entries) target_run_id: str | None = None target_section: str | None = None seen_keys: set[str] = set() for idx in range(len(state_entries) - 1, -1, -1): entry = state_entries[idx] if entry.get('kpi') or entry.get('log'): continue step = entry.get('step') if not isinstance(step, str) or not step: continue kind = _STEP_KIND.get(step) if kind in ('boundary', 'gate'): continue run_id = _infer_run_id( entry, iteration_log_count, same_step_hint=same_step_hints[idx] ) if not run_id: continue key = f'{run_id}:{step}' if key in seen_keys: continue seen_keys.add(key) if entry.get('status') != 'running': continue run_index = _parse_run_index(run_id) if run_index is None: continue section_type = _section_type_for(step, run_index) if section_type is None: continue target_run_id = run_id target_section = section_type break if not target_run_id or not target_section: return None latest_per_step: dict[str, dict[str, Any]] = {} for idx, entry in enumerate(state_entries): if entry.get('kpi') or entry.get('log'): continue step = entry.get('step') if not isinstance(step, str) or not step: continue kind = _STEP_KIND.get(step) if kind in ('boundary', 'gate'): continue run_id = _infer_run_id( entry, iteration_log_count, same_step_hint=same_step_hints[idx] ) if run_id != target_run_id: continue run_index = _parse_run_index(run_id) if run_index is None: continue section_type = _section_type_for(step, run_index) if section_type != target_section: continue latest_per_step[step] = entry cards: list[dict[str, Any]] = [] for step, entry in latest_per_step.items(): status = entry.get('status') if status not in ('complete', 'running'): continue meta = _CARD_META.get((step, target_section), { 'icon': 'clock', 'title': step, 'subtitle': '', }) progress: float | None = None if 'progress' in entry: try: value = float(entry['progress']) if value > 1: value /= 100 progress = max(0.0, min(1.0, value)) except (TypeError, ValueError): pass cards.append({ 'key': f'{target_run_id}:{step}', 'step': step, 'icon': meta['icon'], 'title': meta['title'], 'subtitle': meta['subtitle'], 'status': status, 'progress': progress, }) if not cards: return None cards.sort(key=lambda c: _CARD_ORDER.get((c['step'], target_section), 999)) return { 'run_id': target_run_id, 'section_type': target_section, 'section_label': f'{target_run_id} · {_SECTION_LABELS[target_section]}', 'cards': cards, } def _apply_program_state( payload: dict[str, Any], state_entries: list[dict[str, Any]], *, iteration_log_count: int = 0, session_failure_status: str | None = None, run_active: bool | None = None, session_id: str | None = None, ) -> None: """Mutate payload in place. Builds items[] from state entries and appends log lines. KPI entries (legacy `{kpi:..,value:..}`) are ignored — KPIs are computed by the backend in _compute_kpis.""" # Append logs first (they are independent of items[]). appended_logs: list[dict[str, Any]] = [] for entry in state_entries: if entry.get('kpi'): continue if entry.get('log'): log_obj = entry.get('log') if isinstance(log_obj, dict): appended_logs.append(log_obj) elif isinstance(log_obj, str): appended_logs.append({'text': log_obj, 'ts': entry.get('ts', '')}) if appended_logs: existing = payload.setdefault('logs', []) existing.extend(appended_logs) # Build items (already filtered to complete sections + visible gates). items = _build_pipeline_items(state_entries, iteration_log_count) payload['items'] = items payload['in_flight'] = _compute_in_flight(state_entries, iteration_log_count) # Derive overall status from RAW entries — items[] is filtered to complete # sections only, so deriving from items would falsely report 'complete' # whenever a running section is hidden. latest_per_step: dict[str, str] = {} for entry in state_entries: if entry.get('kpi') or entry.get('log'): continue step = entry.get('step') status = entry.get('status') if isinstance(step, str) and step and isinstance(status, str) and status: # boundary steps don't count toward overall progress if _STEP_KIND.get(step) == 'boundary': continue latest_per_step[step] = status non_gate_running = any( status == 'running' and _STEP_KIND.get(step) != 'gate' for step, status in latest_per_step.items() ) gate_running = any( status == 'running' and _STEP_KIND.get(step) == 'gate' for step, status in latest_per_step.items() ) statuses = list(latest_per_step.values()) if not statuses: payload['status']['state'] = 'pending' elif non_gate_running: payload['status']['state'] = 'running' elif gate_running: # Only HiTL gates are 'running' — agent has handed control back to the # user. Surface as 'waiting' so the header pill says "等待人工" instead # of the spinning "Running" badge. payload['status']['state'] = 'waiting' elif all(s == 'complete' for s in statuses): payload['status']['state'] = 'complete' elif 'complete' in statuses: payload['status']['state'] = 'running' elif 'failed' in statuses: payload['status']['state'] = 'failed' else: payload['status']['state'] = 'pending' # Orphan detection: agent wrote `step:running` then turn ended cleanly # without a closing `complete`/`failed` entry — run_state_store says the # run is no longer active, but the pipeline still claims `running`. Fold # this into session_failure_status='failed' so the same red-failed visual # applies (per user UX choice). Note: only triggers when there's at least # one explicit running step — purely-pending state stays pending. # # Exception: if every running step has an active watcher polling for it, # it's NOT an orphan — agent legitimately ended its turn and is waiting # for a remote file to land via the watcher. The watcher will write the # closing entry when it triggers (file_exists / timeout / cancel). # # Gates (human-check / human-review) are also never orphans: agent writes # `step:running` to surface the gate card, ends the turn, and waits for # the user to reply. There's no watcher and no remote file — the closing # `complete` entry comes from the next user-driven turn. running_steps = { step for step, status in latest_per_step.items() if status == 'running' and _STEP_KIND.get(step) != 'gate' } all_watched = bool(running_steps) and all( _watcher_manager.is_watching(session_id, step) for step in running_steps if session_id ) if ( run_active is False and session_failure_status is None and payload['status']['state'] == 'running' and running_steps and not all_watched ): session_failure_status = 'failed' # Session-level failure (agent process died / cancelled) overrides the # state derived from program-state.jsonl. The state file can't write its # own tombstone when the writer is the dead process — we cross-reference # session.json to surface the failure on the monitor panel. if session_failure_status in ('failed', 'cancelled', 'interrupted'): # Cancelled keeps a softer 'cancelled' label so the UI can distinguish # user-stop from crash; everything else maps to 'failed' (red). target_state = 'cancelled' if session_failure_status == 'cancelled' else 'failed' if payload['status']['state'] != 'complete': payload['status']['state'] = target_state in_flight = payload.get('in_flight') if isinstance(in_flight, dict): cards = in_flight.get('cards') if isinstance(cards, list): for card in cards: if isinstance(card, dict) and card.get('status') == 'running': card['status'] = target_state # Same treatment for any running cards inside items[] — running # round sections also get flipped, and the section status follows. for item in payload.get('items', []): if not isinstance(item, dict): continue if item.get('type') == 'round': section_cards = item.get('cards', []) flipped = False if isinstance(section_cards, list): for card in section_cards: if isinstance(card, dict) and card.get('status') == 'running': card['status'] = target_state flipped = True if flipped or item.get('status') == 'running': item['status'] = target_state elif item.get('type') == 'gate' and item.get('status') == 'running': item['status'] = target_state def _autoresearch_root() -> Path: return Path(os.environ.get('AUTORESEARCH_ROOT', '/mnt/wangsenhao/autoresearch-zk')) def _chat_root_for_runtime( runtime: 'JupyterRuntimeSession | None', ) -> str | None: """Resolve the per-chat autoresearch root as a REMOTE path string. Returns the NFS path under /mnt/wangsenhao/autoresearch-zk-users/// so jupyter pod and SFT training pod both read/write the same directory. Returns None when no runtime is bound — caller decides whether to fall back to the legacy global path.""" if runtime is None: return None return runtime.chat_workspace_root def _run_history_dir() -> Path: return Path( os.environ.get( 'RUN_HISTORY_DIR', '/mnt/xiaoai-zk-model-train-tj5/workflow5', ) ) def _skill_config_yaml_path() -> Path: # backend/api/server.py → ../../skills/model-iteration/assets/config.yaml return ( Path(__file__).resolve().parent.parent.parent / 'skills' / 'model-iteration' / 'assets' / 'config.yaml' ) def _read_workflow_version() -> str | None: path = _skill_config_yaml_path() try: text = path.read_text(encoding='utf-8') except (FileNotFoundError, OSError): return None # Tiny parser to avoid pulling pyyaml just for two lines. in_cml = False for raw in text.splitlines(): stripped = raw.rstrip() if not stripped or stripped.lstrip().startswith('#'): continue if not stripped.startswith(' '): in_cml = stripped.split(':', 1)[0].strip() == 'cml_eval' continue if in_cml: inner = stripped.strip() if inner.startswith('version:'): value = inner.split(':', 1)[1].strip().strip('"\'') return value or None return None def _scan_run_history_max() -> int | None: root = _run_history_dir() try: if not root.is_dir(): return None except OSError: return None best = -1 for entry in root.iterdir(): if not entry.is_dir(): continue m = re.match(r'^workflow(\d+)$', entry.name) if m: try: value = int(m.group(1)) except ValueError: continue if value > best: best = value return best if best >= 0 else None _ITER_COUNT_CACHE: dict[str, int | None] = {} def _iter_count_cache_key(session_id: str | None, account_id: str | None) -> str: return f'{account_id or "_"}|{session_id or "_"}' def _read_iteration_log_count( session_id: str | None = None, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, ) -> int | None: """Count non-empty lines in iteration_log.jsonl. Chat-root lives under the remote workspace_cwd, so reads MUST go through the bound jupyter runtime — the round-trip is multi-second. To keep the SSE pipeline loop responsive, this function returns the last cached value when one exists; only a cold cache pays the synchronous remote cost. The cache is refreshed in the background by `_iter_count_refresh_loop`, which the SSE handler spawns alongside its event loop. """ cache_key = _iter_count_cache_key(session_id, account_id) if cache_key in _ITER_COUNT_CACHE: return _ITER_COUNT_CACHE[cache_key] value = _read_iteration_log_count_uncached( session_id, agent_state=agent_state, account_id=account_id, ) _ITER_COUNT_CACHE[cache_key] = value return value async def _iter_count_refresh_loop( session_id: str | None, *, agent_state: 'AgentState | None', account_id: str | None, interval_seconds: float = 3.0, ) -> None: """Background task that re-reads iteration_log_count every `interval_seconds` and updates the in-memory cache. Cancelled when the caller (SSE handler) exits.""" cache_key = _iter_count_cache_key(session_id, account_id) while True: try: value = await asyncio.to_thread( _read_iteration_log_count_uncached, session_id, agent_state=agent_state, account_id=account_id, ) _ITER_COUNT_CACHE[cache_key] = value except Exception: # noqa: BLE001 pass await asyncio.sleep(interval_seconds) def _read_iteration_log_count_uncached( session_id: str | None = None, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, ) -> int | None: runtime = ( _jupyter_runtime_for_session(agent_state, account_id, session_id) if session_id and agent_state is not None else None ) remote_root = _chat_root_for_runtime(runtime) if runtime is not None and remote_root is not None: remote_path = f'{remote_root}/results/iteration_log.jsonl' cmd = ( f'test -f {shlex.quote(remote_path)} ' f'&& grep -c "[^[:space:]]" {shlex.quote(remote_path)} ' f'|| echo __MISSING__' ) try: result = runtime.run_command( cmd, timeout_seconds=10.0, max_output_chars=2000, ) except Exception: # noqa: BLE001 return None stdout = (result.stdout or '').strip() if result is not None else '' if stdout and stdout != '__MISSING__': try: return int(stdout.splitlines()[-1]) except (ValueError, IndexError): return None return None legacy = _autoresearch_root() / 'results' / 'iteration_log.jsonl' try: if legacy.is_file(): with legacy.open('r', encoding='utf-8') as fp: return sum(1 for line in fp if line.strip()) except (OSError, UnicodeDecodeError): pass return None _METRICS_HISTORY_CACHE: dict[str, dict[str, Any] | None] = {} def _read_iteration_log_metrics_history( session_id: str | None = None, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, ) -> dict[str, Any] | None: """Cached read of full iteration_log.jsonl as metrics time-series.""" cache_key = _iter_count_cache_key(session_id, account_id) if cache_key in _METRICS_HISTORY_CACHE: return _METRICS_HISTORY_CACHE[cache_key] value = _read_iteration_log_metrics_history_uncached( session_id, agent_state=agent_state, account_id=account_id, ) _METRICS_HISTORY_CACHE[cache_key] = value return value async def _metrics_history_refresh_loop( session_id: str | None, *, agent_state: 'AgentState | None', account_id: str | None, interval_seconds: float = 3.0, ) -> None: cache_key = _iter_count_cache_key(session_id, account_id) while True: try: value = await asyncio.to_thread( _read_iteration_log_metrics_history_uncached, session_id, agent_state=agent_state, account_id=account_id, ) _METRICS_HISTORY_CACHE[cache_key] = value except Exception: # noqa: BLE001 pass await asyncio.sleep(interval_seconds) def _read_iteration_log_metrics_history_uncached( session_id: str | None = None, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, ) -> dict[str, Any] | None: """Parse iteration_log.jsonl into {'metrics': [keys], 'rounds': [...]}. Each entry's `results` dict contributes one round point; metric key union is taken across all rounds (sorted by first-appearance for stable column order). Missing values produce nulls so frontend line charts break the series naturally.""" runtime = ( _jupyter_runtime_for_session(agent_state, account_id, session_id) if session_id and agent_state is not None else None ) text: str | None = None if runtime is not None: remote_root = _chat_root_for_runtime(runtime) if remote_root is not None: remote_path = f'{remote_root}/results/iteration_log.jsonl' cmd = ( f'test -f {shlex.quote(remote_path)} ' f'&& cat {shlex.quote(remote_path)} ' f'|| echo __MISSING__' ) try: result = runtime.run_command( cmd, timeout_seconds=10.0, max_output_chars=400_000, ) except Exception: # noqa: BLE001 return None stdout = (result.stdout or '').strip() if result is not None else '' if stdout and stdout != '__MISSING__': text = stdout if text is None: legacy = _autoresearch_root() / 'results' / 'iteration_log.jsonl' try: if legacy.is_file(): text = legacy.read_text(encoding='utf-8', errors='replace') except OSError: return None if not text: return None metrics_order: list[str] = [] metrics_seen: set[str] = set() rounds: list[dict[str, Any]] = [] for line in text.splitlines(): line = line.strip() if not line: continue try: obj = json.loads(line) except (ValueError, TypeError): continue if not isinstance(obj, dict): continue results = obj.get('results') if not isinstance(results, dict): continue values: dict[str, float] = {} for k, v in results.items(): if not isinstance(k, str): continue try: fv = float(v) except (TypeError, ValueError): continue values[k] = fv if k not in metrics_seen: metrics_seen.add(k) metrics_order.append(k) if not values: continue iteration = obj.get('iteration') if isinstance(iteration, str) and iteration.lstrip('-').isdigit(): iteration = int(iteration) if not isinstance(iteration, int): iteration = len(rounds) run_dic = obj.get('runDic') or obj.get('run_dic') if isinstance(run_dic, str) and run_dic.isdigit(): run_dic = int(run_dic) if not isinstance(run_dic, int): run_dic = 0 timestamp = obj.get('timestamp') if not isinstance(timestamp, str): timestamp = '' rounds.append({ 'iteration': iteration, 'runDic': run_dic, 'label': f'R{iteration}', 'timestamp': timestamp, 'values': values, }) if not rounds: return None return {'metrics': metrics_order, 'rounds': rounds} def _read_iteration_log_h1_label( runtime: 'JupyterRuntimeSession | None', run_dic: int | None, *, max_chars: int = 32, ) -> str | None: """读取 iteration_log.jsonl，提取与本轮 dist-analysis 候选集对应的 H1 假设短标签。 优先匹配 runDic 一致的 entry，否则取最新一条；从 ``hypothesis`` / ``next_hypothesis`` 字段切第一句并截断。读不到或无 iteration_log 返回 None （调用方据此决定是否在标题里加括号）。 """ if runtime is None: return None chat_root = _chat_root_for_runtime(runtime) if not chat_root: return None remote_path = f'{chat_root}/results/iteration_log.jsonl' cmd = ( f'test -f {shlex.quote(remote_path)} ' f'&& cat {shlex.quote(remote_path)} ' f'|| echo __MISSING__' ) try: result = runtime.run_command( cmd, timeout_seconds=10.0, max_output_chars=200_000 ) except Exception: # noqa: BLE001 return None stdout = (result.stdout or '').strip() if result is not None else '' if not stdout or stdout == '__MISSING__': return None entries: list[dict[str, Any]] = [] for line in stdout.splitlines(): line = line.strip() if not line: continue try: obj = json.loads(line) except (ValueError, TypeError): continue if isinstance(obj, dict): entries.append(obj) if not entries: return None chosen: dict[str, Any] | None = None if run_dic is not None: for entry in reversed(entries): rd = entry.get('runDic') or entry.get('run_dic') if isinstance(rd, str) and rd.isdigit(): rd = int(rd) if rd == run_dic: chosen = entry break if chosen is None: chosen = entries[-1] for field in ('hypothesis', 'next_hypothesis'): val = chosen.get(field) if isinstance(val, str) and val.strip(): label = val.strip() for sep in ('；', ';', '。', '\n'): if sep in label: label = label.split(sep, 1)[0].strip() break if len(label) > max_chars: label = label[: max_chars - 1] + '…' return label or None return None def _extract_trigger_from_state(state_entries: list[dict[str, Any]]) -> str | None: """Backwards compat: if agent wrote {kpi:'TRIGGER',value:...}, use that. Backend now owns KPI rendering, but TRIGGER is a passthrough of the user's own phrasing — fine to read from state.""" last: str | None = None for entry in state_entries: if entry.get('kpi') == 'TRIGGER': value = entry.get('value') if isinstance(value, str) and value.strip(): last = value return last def _extract_trigger_from_session( sessions_dir: Path, session_id: str | None ) -> str | None: """Return the user's first non-empty message that looks like a training trigger. Used as raw input for _parse_target_set.""" if not session_id: return None safe_id = _safe_session_id(session_id) if safe_id is None: return None path = sessions_dir / safe_id / 'session.json' try: data = json.loads(path.read_text(encoding='utf-8')) except (OSError, json.JSONDecodeError): return None messages = data.get('messages') if not isinstance(messages, list): return None for msg in messages: if not isinstance(msg, dict): continue if msg.get('role') != 'user': continue content = msg.get('content') text = content if isinstance(content, str) else '' if not text: continue # Strip blocks etc. cleaned = re.sub(r'<[^>]+>', ' ', text).strip() if not cleaned: continue if _looks_like_training_trigger(cleaned): return cleaned return None def _looks_like_training_trigger(text: str) -> bool: keywords = ( '训练', 'training', '评测', '迭代', '开始', '需求集合', '目标集合', '目标是', '针对', 'cml', '复杂导航', '.csv', ) lower = text.lower() return any(k.lower() in lower for k in keywords) _TARGET_SET_LLM_CACHE: dict[str, str | None] = {} def _llm_extract_target_set( model_config: ModelConfig | None, text: str ) -> str | None: """Use the configured LLM to extract a target set name from a user message. Cached by message hash so the same input doesn't repeatedly call the model.""" if not text or not model_config: return None text = text.strip() if len(text) > 4000: text = text[:4000] cache_key = hashlib.sha256(text.encode('utf-8')).hexdigest() if cache_key in _TARGET_SET_LLM_CACHE: return _TARGET_SET_LLM_CACHE[cache_key] # Tighten timeout for this short auxiliary call so SSE doesn't stall. cfg = replace(model_config, timeout_seconds=min(20.0, model_config.timeout_seconds)) try: from src.openai_compat import OpenAICompatClient, OpenAICompatError except ImportError: _TARGET_SET_LLM_CACHE[cache_key] = None return None client = OpenAICompatClient(cfg) system_prompt = ( '你是一个文本抽取器。任务：从用户的训练触发消息里抽出"目标需求集合"的名称。\n' '- 名称可能是 CSV 文件名、目录名、或一个标识词\n' '- 只返回名称字符串本身，不要任何前后缀、引号、说明、标点\n' '- 没有明确目标时返回大写 NONE' ) few_shot = [ {'role': 'user', 'content': '开始, icl_test'}, {'role': 'assistant', 'content': 'icl_test'}, {'role': 'user', 'content': '开始，复杂导航过召专项0511.csv'}, {'role': 'assistant', 'content': '复杂导航过召专项0511.csv'}, { 'role': 'user', 'content': '我要进行模型训练，目标是icl_test中的复杂导航过召专项0511.csv，基模在/mnt/zhangzhaowen/icl/v2/test/', }, {'role': 'assistant', 'content': '复杂导航过召专项0511.csv'}, {'role': 'user', 'content': '针对 dapan_test 跑一轮 SFT'}, {'role': 'assistant', 'content': 'dapan_test'}, {'role': 'user', 'content': '你好'}, {'role': 'assistant', 'content': 'NONE'}, ] messages: list[dict[str, Any]] = [{'role': 'system', 'content': system_prompt}] messages.extend(few_shot) messages.append({'role': 'user', 'content': text}) try: result = client.complete(messages, tools=[]) except OpenAICompatError: _TARGET_SET_LLM_CACHE[cache_key] = None return None except Exception: # noqa: BLE001 — keep KPI render robust to upstream LLM hiccups _TARGET_SET_LLM_CACHE[cache_key] = None return None answer = (getattr(result, 'content', '') or '').strip().strip('"\'`，,。()（） ') if not answer or answer.upper() == 'NONE' or len(answer) > 200: _TARGET_SET_LLM_CACHE[cache_key] = None return None # Single-line: take only the first line in case model added explanation answer = answer.split('\n', 1)[0].strip() if not answer or answer.upper() == 'NONE': _TARGET_SET_LLM_CACHE[cache_key] = None return None _TARGET_SET_LLM_CACHE[cache_key] = answer return answer def _parse_target_set( trigger: str | None, model_config: ModelConfig | None = None ) -> str | None: """Sync, non-blocking. Reads cache only — never makes the LLM call inline. Cache is filled by the watcher scanner background loop via _refresh_target_set_cache_blocking (which is run in a thread).""" if not trigger: return None text = trigger.strip() if not text: return None if len(text) > 4000: text = text[:4000] cache_key = hashlib.sha256(text.encode('utf-8')).hexdigest() return _TARGET_SET_LLM_CACHE.get(cache_key) def _refresh_target_set_cache_blocking( model_config: ModelConfig | None, trigger: str ) -> None: """Synchronous wrapper that calls the LLM and populates the cache. Designed to be invoked via asyncio.to_thread from a background task.""" if not model_config or not trigger: return text = trigger.strip() if not text: return if len(text) > 4000: text = text[:4000] cache_key = hashlib.sha256(text.encode('utf-8')).hexdigest() if cache_key in _TARGET_SET_LLM_CACHE: return _llm_extract_target_set(model_config, text) _METRIC_DIFF_CACHE: dict[int, dict[str, Any]] = {} _METRIC_DIFF_NEG_RETRY_AT: dict[int, float] = {} _METRIC_DIFF_NEG_TTL_SECONDS = 30.0 def _read_workflow_metric_diff( run_id: int, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, session_id: str | None = None, ) -> dict[str, Any] | None: """Read workflow/metric_diff/lark_template.json. metric_diff lives on /mnt/xiaoai-zk-model-train-tj5/... which is mounted on the remote jupyter workspace but typically not on the backend host. So we try local first, then fall back to `cat` over the session's jupyter runtime if available. lark_template.json is written once by cml workflow and never changes, so successful reads are cached forever; misses honor a 30s negative cache to avoid hammering the remote during the cml-running phase. """ cached = _METRIC_DIFF_CACHE.get(run_id) if cached is not None: return cached now = time.monotonic() next_retry = _METRIC_DIFF_NEG_RETRY_AT.get(run_id) if next_retry is not None and now < next_retry: return None relative = f'workflow{run_id}/metric_diff/lark_template.json' local_path = _run_history_dir() / relative try: if local_path.is_file(): data = json.loads(local_path.read_text(encoding='utf-8')) if isinstance(data, dict): _METRIC_DIFF_CACHE[run_id] = data _METRIC_DIFF_NEG_RETRY_AT.pop(run_id, None) return data except (OSError, json.JSONDecodeError): pass if agent_state is not None and session_id is not None: runtime = _jupyter_runtime_for_session(agent_state, account_id, session_id) if runtime is not None: remote_path = f'{_run_history_dir()}/{relative}' cmd = ( f'test -f {shlex.quote(remote_path)} ' f'&& cat {shlex.quote(remote_path)} || echo __MISSING__' ) try: result = runtime.run_command( cmd, timeout_seconds=10.0, max_output_chars=200_000, ) except Exception: # noqa: BLE001 _METRIC_DIFF_NEG_RETRY_AT[run_id] = now + _METRIC_DIFF_NEG_TTL_SECONDS return None stdout = (result.stdout or '').strip() if result is not None else '' if ( result is not None and result.exit_code == 0 and stdout and stdout != '__MISSING__' ): try: data = json.loads(stdout) if isinstance(data, dict): _METRIC_DIFF_CACHE[run_id] = data _METRIC_DIFF_NEG_RETRY_AT.pop(run_id, None) return data except json.JSONDecodeError: pass _METRIC_DIFF_NEG_RETRY_AT[run_id] = now + _METRIC_DIFF_NEG_TTL_SECONDS return None def _format_metric_value(metrics: dict[str, Any] | None, key: str) -> str: if not metrics: return '—' value = metrics.get(key) if value is None: return '—' if isinstance(value, (int, float)): if 0 <= value <= 1: return f'{value * 100:.2f}%' return f'{value:.2f}' return str(value) # lark_template.json 的指标都嵌在 eval_output_info 文本里，行格式： # 🟢 specific_test || icl_test: 86.11% (1873/2175) -> 86.11% (1873/2175) (+0.00%) # 我们要 "->" 后面那个值（after-migration pass rate）。 _EVAL_LINE_RE = re.compile( r'^\s*[🟢🔴]\s*(?P.+?):\s*[\d.]+%\s*\([^)]+\)\s*->\s*(?P[\d.]+)%' ) def _extract_eval_metric(eval_output_info: str | None, target_path: str) -> float | None: """Find the line whose 'path' (the bit between emoji and ":") equals target_path, return the post-value as a 0-100 float, or None. Match is exact-equality first, then falls back to path-tail equality — eval lines often carry full relative paths like `specific_test || data/specific_test_set/icl_test/.csv` while target_path is just `specific_test || .csv`.""" if not isinstance(eval_output_info, str) or not eval_output_info: return None target = target_path.strip() # Strip the optional ` || ` so we can suffix-compare just the path tail. target_tail = target.split('||', 1)[-1].strip() if '||' in target else target for line in eval_output_info.splitlines(): m = _EVAL_LINE_RE.match(line) if not m: continue path = m.group('path').strip() path_tail = path.split('||', 1)[-1].strip() if '||' in path else path matched = path == target or path_tail == target_tail or path_tail.endswith( f'/{target_tail}' ) if matched: try: return float(m.group('post')) except ValueError: return None return None def _enrich_metrics_from_eval_output( metrics: dict[str, Any] | None, target_set: str | None ) -> None: """Backfill structured pass-rate keys parsed out of eval_output_info text. Mutates metrics in place. No-op if metrics is None or already has the keys.""" if not metrics: return eval_text = metrics.get('eval_output_info') if not isinstance(eval_text, str): return # _format_metric_value 把 0..1 当作 fraction 格式化为 XX.XX%，所以这里 /100 入库。 # specific_test 大盘 if metrics.get('specific_test_pass_rate') is None: v = _extract_eval_metric(eval_text, 'specific_test') if v is not None: metrics['specific_test_pass_rate'] = v / 100.0 # 大盘 overrall || 车载（注意 lark template 里就是 "overrall" 拼写，不是 "overall"） if metrics.get('overall_car_pass_rate') is None: v = _extract_eval_metric(eval_text, 'overrall || 车载') if v is not None: metrics['overall_car_pass_rate'] = v / 100.0 # target set: specific_test || # Trigger 里带 .csv 后缀；lark template 偶尔写不带后缀的。两种都试。 if ( metrics.get('target_set_pass_rate') is None and target_set and target_set != '—' ): candidates = [target_set] if target_set.endswith('.csv'): candidates.append(target_set[:-4]) else: candidates.append(f'{target_set}.csv') for name in candidates: v = _extract_eval_metric(eval_text, f'specific_test || {name}') if v is not None: metrics['target_set_pass_rate'] = v / 100.0 break def _compute_step0_start_seconds( state_entries: list[dict[str, Any]], ) -> float | None: """Earliest cml running ts seen in state file.""" earliest: float | None = None for entry in state_entries: if entry.get('step') != 'cml': continue if entry.get('status') != 'running': continue ts = entry.get('ts') if not isinstance(ts, str): continue try: from datetime import datetime dt = datetime.fromisoformat(ts.replace('Z', '+00:00')) seconds = dt.timestamp() except (ValueError, OSError): continue if earliest is None or seconds < earliest: earliest = seconds return earliest def _format_elapsed(start_seconds: float | None) -> str: if start_seconds is None: return '00:00:00' delta = max(0.0, time.time() - start_seconds) total = int(delta) h, rem = divmod(total, 3600) m, s = divmod(rem, 60) return f'{h:02d}:{m:02d}:{s:02d}' def _compute_kpis( sessions_dir: Path, session_id: str | None, state_entries: list[dict[str, Any]], model_config: ModelConfig | None = None, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, iteration_log_count: int | None = None, metrics_history: dict[str, Any] | None = None, ) -> list[dict[str, Any]]: """Compute the 7 top KPIs from authoritative sources. Falls back to '—'. RUN ID prefers the runDic this session has already written into program-state.jsonl — that's the only source guaranteed to belong to THIS chat. Falls back to a global scan of RUN_HISTORY_DIR for sessions that haven't logged a runDic yet (e.g. baseline pre-cml). Without this priority, a concurrently-running chat that allocated a larger runDic on the shared /mnt/xiaoai-zk-model-train-tj5/workflow5/ would shadow this chat's value. Metric reads use the same jupyter runtime the watcher uses, so /mnt/* paths reachable only on the remote workspace still work. """ run_max = _resolve_run_dic_from_state(state_entries) if run_max is None: run_max = _scan_run_history_max() run_id_value = str(run_max) if run_max is not None else '—' version = _read_workflow_version() or '—' trigger = _extract_trigger_from_state(state_entries) or _extract_trigger_from_session( sessions_dir, session_id ) target_set = _parse_target_set(trigger, model_config=model_config) or '—' metrics = ( _read_workflow_metric_diff( run_max, agent_state=agent_state, account_id=account_id, session_id=session_id, ) if run_max is not None else None ) _enrich_metrics_from_eval_output(metrics, target_set) target_metric = _format_metric_value(metrics, 'target_set_pass_rate') overall_metric = _format_metric_value(metrics, 'overall_car_pass_rate') specific_metric = _format_metric_value(metrics, 'specific_test_pass_rate') # Round number = MAX numeric `iteration` field across iteration_log.jsonl # entries, NOT the line count. iteration_log.jsonl carries multiple rows # per round (separate hypothesis-only / results / final-summary entries), # so line count overcounts. Falls back to line-count for backwards # compatibility when metrics_history isn't available. max_iteration: int | None = None if metrics_history and isinstance(metrics_history.get('rounds'), list): for r in metrics_history['rounds']: it = r.get('iteration') if isinstance(r, dict) else None if isinstance(it, int): if max_iteration is None or it > max_iteration: max_iteration = it if max_iteration is not None: iteration = 'R0-baseline' if max_iteration == 0 else f'R{max_iteration}' else: if iteration_log_count is not None: iter_count = iteration_log_count else: iter_count = _read_iteration_log_count( session_id, agent_state=agent_state, account_id=account_id, ) if iter_count is None or iter_count == 0: iteration = 'R0-baseline' else: iteration = f'R{iter_count}' elapsed = _format_elapsed(_compute_step0_start_seconds(state_entries)) target_value = ( f'{target_set} · {target_metric}' if target_metric != '—' else target_set ) return [ {'label': 'RUN ID', 'value': run_id_value}, {'label': 'VERSION', 'value': version}, {'label': '目标集合', 'value': target_value}, {'label': '大盘车载', 'value': overall_metric}, {'label': 'SPECIFIC TEST', 'value': specific_metric}, {'label': 'ITERATION', 'value': iteration}, {'label': '耗时', 'value': elapsed, 'icon': 'clock'}, ] def _resolve_run_dic_from_state( state_entries: list[dict[str, Any]], ) -> int | None: """Extract runDic from program-state.jsonl. Priority: 1. Explicit ``runDic`` field on any entry (latest wins). 2. Fallback: scan string fields recursively for ``workflow`` / ``runDic=N`` and pick the **max** match. Path strings often contain both the global mount root (``/mnt/.../workflow5/...``) and the actual run number (``workflow17793/...``) — max picks the run number, not the fixed mount-dir digit. """ workflow_re = re.compile(r'workflow(\d+)') rundic_re = re.compile(r'runDic\s*[=:取]?\s*(\d+)') def _scan_strings(obj: Any) -> int | None: if isinstance(obj, str): candidates = [int(x) for x in workflow_re.findall(obj)] candidates += [int(x) for x in rundic_re.findall(obj)] return max(candidates) if candidates else None if isinstance(obj, dict): best: int | None = None for v in obj.values(): hit = _scan_strings(v) if hit is not None and (best is None or hit > best): best = hit return best if isinstance(obj, list): best = None for v in obj: hit = _scan_strings(v) if hit is not None and (best is None or hit > best): best = hit return best return None # 1. Explicit field — latest wins. for entry in reversed(state_entries): run_dic = entry.get('runDic') or entry.get('run_dic') or entry.get('rundic') if isinstance(run_dic, int): return run_dic if isinstance(run_dic, str) and run_dic.isdigit(): return int(run_dic) # 2. Fallback string scan — take max across all entries. best: int | None = None for entry in state_entries: hit = _scan_strings(entry) if hit is not None and (best is None or hit > best): best = hit return best def _resolve_run_dic_per_round( state_entries: list[dict[str, Any]], ) -> dict[str, int]: """Map each run_id (e.g., 'R2') to the runDic that round actually used. Same priority as ``_resolve_run_dic_from_state`` but bucketed per run_id so per-round step-detail views read the right artifact directory instead of always using the latest workflow id seen anywhere in state.""" workflow_re = re.compile(r'workflow(\d+)') rundic_re = re.compile(r'runDic\s*[=:取]?\s*(\d+)') def _scan(obj: Any) -> int | None: if isinstance(obj, str): cands = [int(x) for x in workflow_re.findall(obj)] cands += [int(x) for x in rundic_re.findall(obj)] return max(cands) if cands else None if isinstance(obj, dict): best: int | None = None for v in obj.values(): hit = _scan(v) if hit is not None and (best is None or hit > best): best = hit return best if isinstance(obj, list): best = None for v in obj: hit = _scan(v) if hit is not None and (best is None or hit > best): best = hit return best return None def _entry_run_id(entry: dict[str, Any]) -> str | None: # 顶层 run_id 优先；否则从 watch.run_id / log.iter 兜底，让 watcher/log # 这种 agent 不写顶层 run_id 但嵌套里带轮次信号的条目也能贡献 runDic。 rid = entry.get('run_id') if isinstance(rid, str) and rid.strip(): return rid.strip() watch = entry.get('watch') if isinstance(watch, dict): wrid = watch.get('run_id') if isinstance(wrid, str) and wrid.strip(): return wrid.strip() log = entry.get('log') if isinstance(log, dict): lit = log.get('iter') if isinstance(lit, str) and lit.strip(): return lit.strip() return None explicit: dict[str, int] = {} fallback: dict[str, int] = {} for entry in state_entries: run_id = _entry_run_id(entry) if not run_id: continue rd = entry.get('runDic') or entry.get('run_dic') or entry.get('rundic') if isinstance(rd, str) and rd.isdigit(): rd = int(rd) if isinstance(rd, int): explicit[run_id] = rd # latest wins continue hit = _scan(entry) if hit is not None: cur = fallback.get(run_id) if cur is None or hit > cur: fallback[run_id] = hit out = dict(fallback) out.update(explicit) return out def _step_artifact_paths( step: str, run_dic: int | None, runtime: 'JupyterRuntimeSession | None' = None, ) -> list[str]: """Map step key → ordered list of remote artifact paths to try. Per-chat artifacts (results/, ai-planning/, output/) live under the chat workspace root on shared NFS at `/mnt/wangsenhao/autoresearch-zk-users///`. Reads still go through the bound jupyter runtime since the path is only mounted there. metric_diff (cml step) stays on the global RUN_HISTORY_DIR mount that is shared across chats. When no runtime is bound, falls back to the legacy global AUTORESEARCH_ROOT so KPI fetches don't crash on unbound sessions. """ chat_root = _chat_root_for_runtime(runtime) or str(_autoresearch_root()) metric_root = os.environ.get( 'RUN_HISTORY_DIR', '/mnt/xiaoai-zk-model-train-tj5/workflow5' ) paths: list[str] = [] if step == 'cml': if run_dic is not None: paths.append( f'{metric_root}/workflow{run_dic}/metric_diff/lark_template.json' ) paths.append( f'{metric_root}/workflow{run_dic}/metric_diff/specific_comparison.csv' ) elif step == 'gold-drift': if run_dic is not None: paths.append(f'{chat_root}/results/gold_drift/drift_{run_dic}.json') elif step == 'dist-analysis': if run_dic is not None: paths.append(f'{chat_root}/results/workflow{run_dic}.md') paths.append( f'{chat_root}/output/relabel_candidates_{run_dic}.csv' ) # 后向兼容：旧 agent 把 csv 写到 jupyter pod 本地 $(pwd)/output/ if runtime is not None: paths.append( f'{runtime.binding.workspace_cwd}/output/relabel_candidates_{run_dic}.csv' ) elif step == 'hypothesis': paths.append(f'{chat_root}/results/iteration_log.jsonl') elif step == 'augment': if run_dic is not None: paths.append( f'{chat_root}/results/data_clean_{run_dic}/modified_samples.jsonl' ) paths.append( f'{chat_root}/results/augment_raw/augment_{run_dic}_raw.jsonl' ) paths.append( f'{chat_root}/ai-planning/data/train_set/zk_intent/augment_{run_dic}.jsonl' ) paths.append( f'{chat_root}/results/data_clean_{run_dic}/label_master_review.jsonl' ) elif step == 'verify': paths.append(f'{chat_root}/results/iteration_log.jsonl') elif step == 'sft': paths.append(f'{chat_root}/results/iteration_log.jsonl') elif step == 'log': paths.append(f'{chat_root}/results/iteration_log.jsonl') paths.append(f'{chat_root}/results/error_registry.jsonl') elif step == 'next-round': paths.append(f'{chat_root}/results/iteration_log.jsonl') return paths def _csv_to_table( text: str, max_rows: int | None ) -> tuple[list[str], list[list[str]], int]: """Parse CSV text into (columns, rows, total). Strips BOM, handles quoted multi-line cells. ``max_rows=None`` means no row cap.""" if text.startswith(''): text = text[1:] import io as _io reader = csv.reader(_io.StringIO(text)) columns: list[str] = [] rows: list[list[str]] = [] total = 0 for i, record in enumerate(reader): if i == 0: columns = [c.strip() for c in record] continue # 跳过整行全空的视觉分隔行（agent 写人审 CSV 时常插空行） if not any(cell.strip() for cell in record): continue total += 1 if max_rows is None or len(rows) < max_rows: row = [_truncate_cell(c) for c in record] if len(row) < len(columns): row += [''] * (len(columns) - len(row)) elif len(row) > len(columns): row = row[: len(columns)] rows.append(row) return columns, rows, total def _jsonl_to_table( text: str, max_rows: int | None ) -> tuple[list[str], list[list[str]], int]: """Parse jsonl into (columns, rows, total). Column order is the order keys appear in the first record, with new keys appended. ``max_rows=None`` means no row cap.""" columns: list[str] = [] seen: set[str] = set() sample_rows: list[dict[str, Any]] = [] total = 0 for line in text.splitlines(): stripped = line.strip() if not stripped: continue try: obj = json.loads(stripped) except json.JSONDecodeError: continue if not isinstance(obj, dict): continue total += 1 for k in obj.keys(): if k not in seen: seen.add(k) columns.append(k) if max_rows is None or len(sample_rows) < max_rows: sample_rows.append(obj) rows: list[list[str]] = [] for obj in sample_rows: row: list[str] = [] for col in columns: val = obj.get(col) if val is None: row.append('') elif isinstance(val, str): row.append(_truncate_cell(val)) elif isinstance(val, (int, float, bool)): row.append(str(val)) else: row.append(_truncate_cell(json.dumps(val, ensure_ascii=False))) rows.append(row) return columns, rows, total def _truncate_cell(text: str, limit: int = 4000) -> str: if len(text) <= limit: return text return text[:limit] + f'…（共 {len(text)} 字，已截断）' def _detect_artifact_format(path: str) -> str: lower = path.lower() if lower.endswith('.md'): return 'markdown' if lower.endswith('.jsonl'): return 'jsonl' if lower.endswith('.json'): return 'json' if lower.endswith('.csv'): return 'csv' return 'text' def _read_artifact( runtime: 'JupyterRuntimeSession | None', remote_path: str ) -> tuple[str | None, str | None]: """Read a remote artifact via Jupyter `cat`. Falls back to local Path read if no remote runtime (dev / unbound sessions). Returns (content, error).""" if runtime is not None: cmd = ( f'test -f {shlex.quote(remote_path)} && ' f'cat {shlex.quote(remote_path)} || echo __MISSING__' ) try: result = runtime.run_command( cmd, timeout_seconds=15.0, max_output_chars=400_000, ) except Exception as exc: # noqa: BLE001 return None, f'remote read failed: {exc}' if not result: return None, 'no result' out = result.stdout or '' # Trim our own marker if file missing if out.strip().endswith('__MISSING__'): return None, f'remote file not found: {remote_path}' return out, None try: p = Path(remote_path) if not p.is_file(): return None, f'file not found: {remote_path}' return p.read_text(encoding='utf-8', errors='replace'), None except OSError as exc: return None, f'local read failed: {exc}' def _hardcoded_autoresearch_pipeline( session_id: str | None, sessions_dir: Path | None = None, state_entries: list[dict[str, Any]] | None = None, model_config: ModelConfig | None = None, *, agent_state: 'AgentState | None' = None, account_id: str | None = None, iteration_log_count: int | None = None, metrics_history: dict[str, Any] | None = None, ) -> dict[str, Any]: """Canonical autoresearch model-iteration pipeline structure. Card / phase / log dynamics still come from program-state.jsonl, but the 7 top KPIs are now backend-owned (read from skills/config.yaml, iteration_log, run history dir, CML metric_diff, session state) — agent cannot override them via state file. """ now_ms = int(time.time() * 1000) if sessions_dir is not None: kpis = _compute_kpis( sessions_dir, session_id, state_entries or [], model_config=model_config, agent_state=agent_state, account_id=account_id, iteration_log_count=iteration_log_count, metrics_history=metrics_history, ) else: # Defensive default if caller didn't pass sessions_dir. kpis = [ {'label': 'RUN ID', 'value': '—'}, {'label': 'VERSION', 'value': '—'}, {'label': '目标集合', 'value': '—'}, {'label': '大盘车载', 'value': '—'}, {'label': 'SPECIFIC TEST', 'value': '—'}, {'label': 'ITERATION', 'value': 'R0-baseline'}, {'label': '耗时', 'value': '00:00:00', 'icon': 'clock'}, ] return { 'session_id': session_id or '', 'generated_at_ms': now_ms, 'available': True, 'status': {'mode': 'Model Iteration', 'state': 'pending'}, 'kpis': kpis, # items[] is heterogeneous: {type:'round',...} for R0·Baseline / # R{n}·Train / R{n}·Analysis, and {type:'gate',...} for HiTL gates. # Built dynamically from program-state.jsonl by _build_pipeline_items — # sections only appear when the agent has touched at least one step # in them. 'items': [], 'in_flight': None, 'logs': [], 'metrics_history': metrics_history, } def _session_mtime(path: Path) -> float: try: return path.stat().st_mtime except OSError: return 0.0 def _session_id_from_path(path: Path) -> str: if path.name == 'session.json': return path.parent.name return path.stem def _safe_session_id(session_id: str | None) -> str | None: if not session_id: return None normalized = re.sub(r'[^a-zA-Z0-9._-]+', '_', session_id.strip()) return normalized[:120] or None