Files
quality_analyse/scripts/nao_quality_runtime.py
T

2119 lines
108 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""Run an isolated, delegated-auth NAO quality conversation as an NDJSON worker.
The worker keeps NAO's native engine and MCP protocol intact. The quality MCP
uses DelegatedApi, which can only call the local, actor-scoped runtime-query API.
Credentials are read from stdin and kept in a mode-0600 file outside the NAO
project; they are never passed to the model or written to worker output.
"""
from __future__ import annotations
import argparse
import base64
import contextlib
import fcntl
import hashlib
import http.cookiejar
import json
import math
import os
import re
import secrets
import shutil
import signal
import socket
import sqlite3
import stat
import subprocess
import sys
import threading
import time
import uuid
from dataclasses import dataclass
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Callable, Iterator
from urllib.error import HTTPError, URLError
from urllib.parse import parse_qs, quote, urlencode, urlsplit
from urllib.request import HTTPCookieProcessor, Request, build_opener
ROOT = Path(__file__).resolve().parents[1]
USERS_ROOT = ROOT / ".runtime" / "nao-users"
TRIAL = ROOT / ".runtime" / "nao-trial"
VENV = TRIAL / "venv"
PYTHON = VENV / "bin" / "python"
NAO_SERVER = VENV / "lib" / "python3.11" / "site-packages" / "nao_core" / "bin" / "nao-chat-server"
FASTAPI_MAIN = VENV / "lib" / "python3.11" / "site-packages" / "nao_core" / "bin" / "fastapi" / "main.py"
TEMPLATES = ROOT / "ai" / "nao-quality"
RUNTIME_QUERY_URL = "http://127.0.0.1:5281/api/analysis/runtime/query"
MAX_INSTANCES = 2
TURN_TIMEOUT_SECONDS = 600
SSE_IDLE_TIMEOUT_SECONDS = 60
NATIVE_STOP_TIMEOUT_SECONDS = 4.0
CHAT_ID_WAIT_SECONDS = 0.75
MAX_QUESTION_CHARS = 20_000
MAX_CONTEXT_CHARS = 32_000
MAX_HTML_BYTES = 12 * 1024 * 1024
_UUID_RE = re.compile(r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-8][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$")
_ALLOWED_PATHS = {
"GET": {
"/api/releases", "/api/options", "/api/metrics",
"/api/analysis/tools/catalog", "/api/analytics/top",
},
"POST": {"/api/analysis/tools/details", "/api/analysis/defects", "/api/analysis/investigate"},
}
class RuntimeErrorSafe(RuntimeError):
"""An error safe to return without exposing credentials or request bodies."""
def _require_private_regular_file(path: Path, *, missing_ok: bool = True) -> bool:
if path.is_symlink():
raise RuntimeErrorSafe("私有运行文件路径无效。")
try:
info = path.stat(follow_symlinks=False)
except FileNotFoundError:
if missing_ok:
return False
raise RuntimeErrorSafe("私有运行文件不可用。") from None
except OSError:
raise RuntimeErrorSafe("私有运行文件不可用。") from None
if not stat.S_ISREG(info.st_mode) or info.st_mode & 0o077:
raise RuntimeErrorSafe("私有运行文件权限无效。")
return True
def _uuid(value: Any, field: str) -> str:
if not isinstance(value, str) or not _UUID_RE.fullmatch(value):
raise ValueError(f"{field} 必须是 UUID。")
return str(uuid.UUID(value))
def _mode_private(path: Path, mode: int = 0o600) -> None:
path.chmod(mode)
def _atomic_json(path: Path, value: Any, *, mode: int = 0o600) -> None:
path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
temp = path.with_name(f".{path.name}.{os.getpid()}.{secrets.token_hex(5)}.tmp")
fd = os.open(temp, os.O_WRONLY | os.O_CREAT | os.O_EXCL, mode)
try:
with os.fdopen(fd, "w", encoding="utf-8") as stream:
json.dump(value, stream, ensure_ascii=False, separators=(",", ":"), allow_nan=False)
stream.write("\n")
stream.flush()
os.fsync(stream.fileno())
os.replace(temp, path)
_mode_private(path, mode)
finally:
temp.unlink(missing_ok=True)
def _atomic_bytes(path: Path, value: bytes, *, mode: int = 0o600) -> None:
"""Atomically replace a private file without ever following its final component."""
path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
if path.is_symlink():
raise RuntimeErrorSafe("NAO project 文件路径无效。")
temp = path.with_name(f".{path.name}.{os.getpid()}.{secrets.token_hex(5)}.tmp")
fd = os.open(temp, os.O_WRONLY | os.O_CREAT | os.O_EXCL | getattr(os, "O_NOFOLLOW", 0), mode)
try:
with os.fdopen(fd, "wb") as stream:
stream.write(value)
stream.flush()
os.fsync(stream.fileno())
os.replace(temp, path)
finally:
temp.unlink(missing_ok=True)
def _read_json(path: Path) -> Any:
with path.open("r", encoding="utf-8") as stream:
return json.load(stream)
def _safe_runtime_root(runtime_root: Path) -> Path:
runtime_root = Path(runtime_root).absolute()
if runtime_root.is_symlink():
raise ValueError("运行目录无效。")
root = runtime_root.resolve()
if root != USERS_ROOT.resolve() and not root.name.startswith("test-nao-users-"):
raise ValueError("运行目录无效。")
return root
def _runtime_identity(runtime_root: Path) -> tuple[str, str | None]:
"""Return (actor_id, session_id) for either a new session root or a legacy actor root."""
root = Path(runtime_root).absolute()
if root.parent.name == "sessions":
actor_id = _uuid(root.parent.parent.name, "actorId")
return actor_id, _uuid(root.name, "sessionId")
return _uuid(root.name, "actorId"), None
def _reject_symlink_components(path: Path, stop_at: Path) -> None:
"""Reject symlinks in a runtime path between path and its configured root."""
absolute = Path(os.path.abspath(path))
stop = Path(os.path.abspath(stop_at))
try:
relative = absolute.relative_to(stop)
except ValueError:
raise ValueError("运行路径越界。") from None
current = stop
if current.is_symlink():
raise ValueError("运行路径无效。")
for part in relative.parts:
current = current / part
if current.is_symlink():
raise ValueError("运行路径无效。")
@dataclass(frozen=True)
class AccessSnapshot:
actor_id: str
token: str
release_id: str
query_url: str
def load_access_file(
access_path: str | Path,
*,
runtime_root: Path = USERS_ROOT,
expected_actor_id: str | None = None,
expected_session_id: str | None = None,
) -> AccessSnapshot:
"""Read and validate a private access file without following path escapes.
Session-bound callers must provide both identity fields. The actor-root form
remains readable only for old installations and existing offline callers.
"""
root = _safe_runtime_root(Path(runtime_root))
absolute = Path(os.path.abspath(Path(access_path)))
_reject_symlink_components(absolute, root)
if expected_session_id is not None:
if expected_actor_id is None:
raise ValueError("会话访问凭据必须绑定 actorId。")
actor_id = _uuid(expected_actor_id, "actorId")
session_id = _uuid(expected_session_id, "sessionId")
session_root = root / actor_id / "sessions" / session_id
if absolute != session_root / "access.json":
raise ValueError("访问凭据文件路径与当前会话不匹配。")
else:
actor_dir = absolute.parent
if actor_dir.parent != root:
raise ValueError("访问凭据文件路径越界。")
actor_id = _uuid(actor_dir.name, "actorId")
session_id = None
if expected_actor_id is not None and actor_id != _uuid(expected_actor_id, "actorId"):
raise ValueError("访问凭据与当前 actor 不匹配。")
if absolute.name != "access.json" or not absolute.parent.is_dir():
raise ValueError("访问凭据文件路径无效。")
private_dirs = [root, root / actor_id]
if session_id is not None:
private_dirs.extend((root / actor_id / "sessions", root / actor_id / "sessions" / session_id))
for directory in private_dirs:
try:
directory_info = directory.stat(follow_symlinks=False)
except OSError:
raise ValueError("访问凭据目录不可用。") from None
if not stat.S_ISDIR(directory_info.st_mode) or directory_info.st_mode & 0o077:
raise ValueError("访问凭据目录权限无效。")
try:
fd = os.open(absolute, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0))
with os.fdopen(fd, "r", encoding="utf-8") as stream:
info = os.fstat(stream.fileno())
if not stat.S_ISREG(info.st_mode) or info.st_mode & 0o077:
raise ValueError("访问凭据文件权限无效。")
raw = json.load(stream)
except ValueError:
raise
except (OSError, json.JSONDecodeError):
raise ValueError("访问凭据文件内容无效。") from None
if not isinstance(raw, dict):
raise ValueError("访问凭据文件内容无效。")
if session_id is not None:
if raw.get("actorId") != actor_id or raw.get("sessionId") != session_id:
raise ValueError("访问凭据与当前会话不匹配。")
elif raw.get("actorId") not in (None, actor_id):
raise ValueError("访问凭据与当前 actor 不匹配。")
token = raw.get("token")
release_id = raw.get("releaseId")
query_url = raw.get("queryUrl")
if not isinstance(token, str) or not token.strip() or len(token) > 4096 or "\n" in token or "\r" in token:
raise ValueError("访问凭据不完整。")
release_id = _uuid(release_id, "releaseId")
parsed = urlsplit(query_url if isinstance(query_url, str) else "")
if (parsed.scheme != "http" or parsed.hostname != "127.0.0.1" or parsed.port != 5281
or parsed.path != "/api/analysis/runtime/query" or parsed.username or parsed.password
or parsed.query or parsed.fragment):
raise ValueError("质量查询地址必须是固定的本机接口。")
return AccessSnapshot(actor_id, token, release_id, RUNTIME_QUERY_URL)
def load_access_snapshot(
access_path: str | Path,
*,
runtime_root: Path = USERS_ROOT,
expected_actor_id: str,
expected_session_id: str,
) -> AccessSnapshot:
"""Strict loader for credentials owned by one actor/session pair."""
return load_access_file(access_path, runtime_root=runtime_root,
expected_actor_id=expected_actor_id,
expected_session_id=expected_session_id)
def _check_release_pin(path: str, payload: Any, release_id: str) -> None:
query = parse_qs(urlsplit(path).query, keep_blank_values=True, strict_parsing=False)
values = query.get("releaseId", [])
if any(value != release_id for value in values):
raise ValueError("请求的 releaseId 与本轮固定版本不一致。")
def walk(value: Any) -> Iterator[Any]:
if isinstance(value, dict):
for key, child in value.items():
if key == "releaseId":
yield child
yield from walk(child)
elif isinstance(value, list):
for child in value:
yield from walk(child)
if any(value != release_id for value in walk(payload)):
raise ValueError("请求的 releaseId 与本轮固定版本不一致。")
def _redact_credentials(value: Any, token: str) -> Any:
if isinstance(value, dict):
sensitive_keys = {"token", "accesstoken", "authorization", "bearer"}
return {
key: _redact_credentials(child, token)
for key, child in value.items()
if not isinstance(key, str) or key.casefold() not in sensitive_keys
}
if isinstance(value, list):
return [_redact_credentials(child, token) for child in value]
if isinstance(value, str):
return value.replace(token, "[REDACTED]")
return value
class DelegatedApi:
"""QualityTools API adapter. Each call reloads actor credentials and delegates."""
def __init__(
self,
access_path: str | Path,
*,
runtime_root: Path = USERS_ROOT,
expected_actor_id: str | None = None,
expected_session_id: str | None = None,
opener: Any | None = None,
timeout: float = 180,
) -> None:
self.access_path = Path(access_path)
self.runtime_root = Path(runtime_root)
self.expected_actor_id = expected_actor_id
self.expected_session_id = expected_session_id
self.opener = opener or build_opener()
self.timeout = timeout
def login(self) -> "DelegatedApi":
"""Validate the delegated file only; never logs into the NAO admin API."""
load_access_file(
self.access_path, runtime_root=self.runtime_root,
expected_actor_id=self.expected_actor_id,
expected_session_id=self.expected_session_id,
)
return self
def call(self, path: str, payload: dict[str, Any] | None = None, method: str | None = None) -> dict[str, Any]:
verb = (method or ("POST" if payload is not None else "GET")).upper()
if verb not in _ALLOWED_PATHS or not isinstance(path, str):
raise ValueError("质量查询接口不在允许范围内。")
parsed_path = urlsplit(path)
if (not parsed_path.path.startswith("/api/") or parsed_path.scheme or parsed_path.netloc
or parsed_path.fragment or parsed_path.path not in _ALLOWED_PATHS[verb]):
raise ValueError("质量查询接口不在允许范围内。")
snapshot = load_access_file(
self.access_path, runtime_root=self.runtime_root,
expected_actor_id=self.expected_actor_id,
expected_session_id=self.expected_session_id,
)
_check_release_pin(path, payload, snapshot.release_id)
if payload is not None and not isinstance(payload, dict):
raise ValueError("质量查询参数格式无效。")
wrapper = {"token": snapshot.token, "method": verb, "path": path, "payload": payload}
request = Request(
snapshot.query_url,
data=json.dumps(wrapper, ensure_ascii=False, separators=(",", ":"), allow_nan=False).encode("utf-8"),
headers={"Accept": "application/json", "Content-Type": "application/json"},
method="POST",
)
try:
with self.opener.open(request, timeout=self.timeout) as response:
raw = response.read(8 * 1024 * 1024 + 1)
except HTTPError as error:
# Runtime API may return a safe, actor-scoped detail (for example 403/409).
try:
detail = json.loads(error.read(64 * 1024).decode("utf-8", "replace"))
except Exception:
detail = None
safe_detail = detail.get("detail") if isinstance(detail, dict) else None
if isinstance(safe_detail, str) and len(safe_detail) <= 600:
raise RuntimeErrorSafe(safe_detail.replace(snapshot.token, "[REDACTED]")) from None
raise RuntimeErrorSafe(f"质量数据分析平台查询暂不可用(HTTP {error.code})。") from None
except (URLError, OSError, TimeoutError):
raise RuntimeErrorSafe("质量数据分析平台查询暂不可用,请确认本机质量服务后重试。") from None
if len(raw) > 8 * 1024 * 1024:
raise RuntimeErrorSafe("质量数据分析平台查询响应超过允许大小。")
try:
result = json.loads(raw.decode("utf-8")) if raw else {}
except (UnicodeDecodeError, json.JSONDecodeError):
raise RuntimeErrorSafe("质量数据分析平台查询返回格式无效。") from None
if not isinstance(result, dict):
raise RuntimeErrorSafe("质量数据分析平台查询返回格式无效。")
return _redact_credentials(result, snapshot.token)
def public_tool_call(tool: str, raw_input: Any) -> dict[str, Any]:
"""Operation trace only; never publish model reasoning, file contents or credentials."""
value = raw_input if isinstance(raw_input, dict) else {}
if tool == "mcp_call" and value.get("tool") == "quality_data":
args = value.get("arguments") if isinstance(value.get("arguments"), dict) else {}
labels = {"catalog": "确认可用数据", "query": "查询质量指标", "compare": "比较期间变化",
"top": "查询重点对象排名", "changes": "追查变化较大对象", "defects": "分析具体缺陷",
"details": "核查来源明细", "investigate": "核对全量明细与原表"}
allowed = {"mode", "scope", "period", "comparisonPeriod", "filters", "view", "stage", "dimension",
"defect", "category", "limit", "sort", "action", "offset", "recordFilter", "sourceRevisionId"}
nested = {"year", "month", "quarter", "department", "group", "product", "projectGroup", "includeUnassigned"}
public_input = {}
for key, item in args.items():
if key not in allowed:
continue
if isinstance(item, dict):
public_input[key] = {k: v for k, v in item.items() if k in nested and isinstance(v, (str, int, float, bool))}
elif isinstance(item, (str, int, float, bool)):
public_input[key] = item[:500] if isinstance(item, str) else item
return {"label": labels.get(args.get("mode"), "查询质量数据"), "input": public_input}
if tool in ("read", "list", "grep", "glob", "search"):
return {"label": "读取分析资料" if tool == "read" else "查找分析资料"}
if "story" in tool:
return {"label": "生成或更新分析报告"}
return {"label": {"python": "执行数据分析", "sql": "查询分析数据"}.get(tool, "执行分析工具")}
def public_progress_text(value: str) -> str:
value = re.sub(r"(?i)(?:bearer\s+)[A-Za-z0-9._~+/-]+", "Bearer [REDACTED]", value)
return re.sub(r"(?i)(api[_-]?key|password|token|secret|authorization)(\s*[=::]\s*)[^\s,,;;]+", r"\1\2[REDACTED]", value)[:8000]
def public_evidence_checks(output: Any) -> dict[str, str]:
"""Only verified tool outcomes update coverage warnings; never assistant text."""
if isinstance(output,str):
try: output=json.loads(output)
except ValueError:return {}
if not isinstance(output,dict):return {}
if isinstance(output.get("content"),list):
checks={}
for part in output["content"]:
if isinstance(part,dict) and isinstance(part.get("text"),str):checks.update(public_evidence_checks(part["text"]))
return checks
checks={}
def add(rows,period,proof=None):
for row in rows if isinstance(rows,list) else []:
if not isinstance(row,dict):continue
key=hashlib.sha256(json.dumps([output.get("releaseId"),output.get("stage"),period,output.get("filters"),row.get("category")],ensure_ascii=False,sort_keys=True).encode()).hexdigest()[:16]
if output.get("kind")=="quality_investigation":
expected=(row.get("expectedTotalsOnly") or {}).get("rows",0)>0 and row.get("unexpectedDifferenceRows")==0 and row.get("requiresFurtherInvestigation") is False
status="unresolved" if proof is not None and (proof.get("complete") is not True or proof.get("allSelectedValuesMatchStored") is not True) else "expected" if expected else "unresolved" if row.get("requiresFurtherInvestigation") else "checked"
else:
status="expected" if row.get("reconciliationStatus")=="EXPECTED_TOTALS_ONLY" else "unresolved" if row.get("reconciliationStatus")!="RECONCILED" else "checked"
checks[key]=status
if output.get("kind")=="quality_investigation":
for name,period in (("current",output.get("period")),("comparison",output.get("comparisonPeriod"))):
side=output.get(name)
if isinstance(side,dict):add(side.get("categories"),period,side.get("sourceVerification"))
elif output.get("kind")=="quality_fine_defects":
add(output.get("categoryReconciliation"),output.get("period"));add(output.get("comparisonCategoryReconciliation"),output.get("comparisonPeriod"))
return checks
def public_tool_failure(output: Any) -> str:
if isinstance(output,str):
try:output=json.loads(output)
except ValueError:return "工具调用未成功,未提供可展示的错误原因"
if isinstance(output,dict):
error=output.get("error")
if isinstance(error,dict) and error.get("code")=="invalid_or_unavailable" and isinstance(error.get("message"),str):
return "查询未成功:"+public_progress_text(error["message"])[:450]
if isinstance(output.get("structuredContent"),dict):return public_tool_failure(output["structuredContent"])
for part in output.get("content",[]) if isinstance(output.get("content"),list) else []:
if isinstance(part,dict) and isinstance(part.get("text"),str):return public_tool_failure(part["text"])
return "工具调用未成功,未提供可展示的错误原因"
def public_tool_result(output: Any) -> str:
"""Summarize only verified response structure; never forward raw tool payloads."""
if isinstance(output, str):
try: output = json.loads(output)
except ValueError: return "工具已返回;未提供可展示的结构化摘要"
if not isinstance(output, dict): return "工具已返回结果"
content = output.get("content")
if isinstance(content, list):
for part in content:
if isinstance(part, dict) and isinstance(part.get("text"), str):
try:
parsed = json.loads(part["text"])
if isinstance(parsed, dict): return public_tool_result(parsed)
except ValueError: pass
parts = []
if output.get("kind") == "quality_investigation":
current = output.get("current", {})
parts.append(f"全量统计 {current.get('totalRows', 0)} 行")
proof = current.get("sourceVerification") or {}
if proof: parts.append(f"原表核验 {proof.get('verifiedRows', 0)}/{proof.get('expectedRows', 0)} 行,差异 {proof.get('mismatchRows', 0)} 行")
for key in ("evidenceRef", "evidenceId"):
ref = output.get(key)
if isinstance(ref, str) and re.fullmatch(r"[A-Za-z0-9_.:/#-]{1,160}", ref):
parts.append("证据:" + ref); break
if isinstance(output.get("rows"), list): parts.append(f"返回 {len(output['rows'])} 条结果")
if isinstance(output.get("sourceRows"), int): parts.append(f"来源 {output['sourceRows']} 行")
coverage = output.get("coverage")
if isinstance(coverage, dict) and coverage.get("available") is False: parts.append("当前来源覆盖不足")
reconciliation = [row for key in ("categoryReconciliation", "comparisonCategoryReconciliation")
for row in (output[key] if isinstance(output.get(key), list) else [])]
if reconciliation and any(isinstance(row, dict) and row.get("reconciliationStatus") not in ("RECONCILED", "EXPECTED_TOTALS_ONLY") for row in reconciliation):
parts.append("细分与正式大类未对账:仅作为变化线索,不能据此判定正式报废主因")
if output.get("rowsTruncated") is True or output.get("truncated") is True or output.get("pairedTruncated") is True: parts.append("仅覆盖返回集合,不代表全量排名")
def number(value: Any) -> float | None:
if isinstance(value, (int, float)) and not isinstance(value, bool):
return float(value) if math.isfinite(value) else None
return None
def metric_summary(snapshot: Any, prefix: str = "") -> None:
if not isinstance(snapshot, dict): return
rate = number(snapshot.get("compositeRate"))
if rate is not None: parts.append(f"{prefix}综合合格率 {rate * 100:.2f}%")
for stage, label in (("black", "黑片"), ("bright", "亮片")):
metric = snapshot.get(stage)
if not isinstance(metric, dict): continue
weight = number(metric.get("scrapWeight"))
if weight is not None: parts.append(f"{prefix}{label}报废 {weight:,.1f} kg")
metric_summary(output.get("summary"))
metric_summary(output.get("current"), "本期")
metric_summary(output.get("comparison"), "对照期")
rows = output.get("rows")
if isinstance(rows, list) and rows and isinstance(rows[0], dict):
label, weight = rows[0].get("label"), number(rows[0].get("weight"))
if isinstance(label, str) and weight is not None:
prefix = "来源细分" if output.get("grain") == "source_defect_weight" else "返回项"
parts.append(f"首条{prefix}:{public_progress_text(label)[:160]} {weight:,.1f} kg")
return ";".join(parts) or "工具已返回结果;可继续核查"
def tool_operation_key(tool: str, raw_input: Any, *, complete_defects: bool = False) -> str:
value = json.loads(json.dumps(raw_input)) if isinstance(raw_input, dict) else {}
if tool == "read" and set(value) == {"path"} and value["path"] in ("/RULES.md", "/agent/mcps/quality/quality_data.json"):
# The native provider sometimes uses `path` before correcting it to read's
# real `file_path` parameter. Only the same known static file can recover it.
value = {"file_path": value["path"]}
if tool == "mcp_call" and value.get("tool") == "quality_data":
args = value.get("arguments")
if isinstance(args,dict) and args.get("scope")=="overall" and isinstance(args.get("filters"),dict) and args["filters"].get("group"):
# Rejected conflicting scope is recovered only by the successful
# same group query after the redundant overall scope is removed.
args.pop("scope")
if isinstance(args, dict) and args.get("mode") == "compare" and args.get("view") == "overall":
# `compare` already operates on the selected overall scope; an explicit
# overall view is the same investigation when the model corrects it.
args.pop("view")
if isinstance(args,dict) and args.get("mode")=="defects" and args.get("dimension")=="defect":
if args.get("defect")=="all":args.pop("defect")
if complete_defects:args.pop("limit",None)
encoded = json.dumps([tool, value], sort_keys=True, ensure_ascii=False)
return hashlib.sha256(encoded.encode()).hexdigest()
def complete_defect_result(output: Any) -> bool:
if isinstance(output,str):
try:output=json.loads(output)
except ValueError:return False
if not isinstance(output,dict):return False
if isinstance(output.get("content"),list):
return any(complete_defect_result(part.get("text")) for part in output["content"] if isinstance(part,dict))
return (output.get("kind")=="quality_fine_defects" and output.get("rankingComplete") is True
and output.get("truncated") is False and output.get("pairedTruncated",False) is False)
class ConversationAccumulator:
"""Keep provider-emitted reasoning separate from answers and tool payloads."""
def __init__(self, on_answer: Callable[[str], None] | None = None) -> None:
self.blocks: dict[str, str] = {}
self.block_order: list[str] = []
self.active_text_id = "text"
self.finish_reasons: list[str] = []
self.tool_calls: list[dict[str, Any]] = []
self.tool_outputs: dict[str, Any] = {}
self.tool_error = False
self.unattributed_tool_error = False
self.tool_operation_keys: dict[int, str] = {}
self.complete_defect_keys: dict[int, str] = {}
self.error_event = False
self.on_answer = on_answer
self.chat_id: str | None = None
self.done_received = False
self.reasoning_blocks: dict[str, str] = {}
self.reasoning_truncated = False
self.timeline: list[dict[str, Any]] = []
self.timeline_text: dict[str, dict[str, Any]] = {}
self.timeline_segments: dict[int, str] = {}
def record_event(self, kind: str, **fields: Any) -> dict[str, Any]:
entry = {"sequence": (self.timeline[-1]["sequence"] if self.timeline else 0) + 1, "kind": kind, "at": datetime.now(timezone.utc).isoformat(), **fields}
if len(self.timeline) < 3000:
self.timeline.append(entry)
return entry
def feed(self, event: dict[str, Any]) -> None:
kind = event.get("type")
if kind == "data-newChat":
value = (event.get("data") or {}).get("id") if isinstance(event.get("data"), dict) else None
if isinstance(value, str) and value and len(value) <= 256:
self.chat_id = value
elif kind == "reasoning-delta":
block_id = event.get("id") if isinstance(event.get("id"), str) else "reasoning"
delta = event.get("delta")
if isinstance(delta, str):
remaining = max(0, 200_000 - sum(len(value) for value in self.reasoning_blocks.values()))
self.reasoning_blocks[block_id] = self.reasoning_blocks.get(block_id, "") + delta[:remaining]
self.reasoning_truncated = self.reasoning_truncated or len(delta) > remaining
elif kind == "text-start":
block_id = event.get("id")
if isinstance(block_id, str) and block_id:
self.active_text_id = block_id
if block_id not in self.blocks:
self.blocks[block_id] = ""
self.block_order.append(block_id)
elif kind == "text-delta":
block_id = event.get("id") if isinstance(event.get("id"), str) else self.active_text_id
self.active_text_id = block_id
if block_id not in self.blocks:
self.blocks[block_id] = ""
self.block_order.append(block_id)
delta = event.get("delta")
if isinstance(delta, str):
self.blocks[block_id] += delta
segment = self.timeline_text.get(block_id)
if segment is None or not self.timeline or self.timeline[-1] is not segment:
segment = self.record_event("assistant", text="")
self.timeline_text[block_id] = segment
sequence = segment["sequence"]
self.timeline_segments[sequence] = self.timeline_segments.get(sequence, "") + delta
segment["text"] = public_progress_text(self.timeline_segments[sequence])
if self.on_answer:
self.on_answer(self.blocks[block_id])
elif kind == "finish":
reason = event.get("finishReason")
if isinstance(reason, str):
self.finish_reasons.append(reason)
elif kind in ("tool-input-available", "tool-input-error"):
tool_name = event.get("toolName")
if isinstance(tool_name, str):
call_id = event.get("toolCallId")
if isinstance(call_id, str) and any(c.get("toolCallId") == call_id for c in self.tool_calls): return
call = {"tool": tool_name, "status": "started", "startedAt": datetime.now(timezone.utc).isoformat(),
**public_tool_call(tool_name, event.get("input"))}
if isinstance(call_id, str):
call["toolCallId"] = call_id
self.tool_calls.append(call)
self.record_event("tool_start", tool=tool_name, toolCallId=call.get("toolCallId", ""), label=call["label"], input=call.get("input", {}), status="started")
if isinstance(event.get("input"), dict):
self.tool_operation_keys[id(call)] = tool_operation_key(tool_name, event["input"])
self.complete_defect_keys[id(call)] = tool_operation_key(tool_name,event["input"],complete_defects=True)
if kind == "tool-input-error":
self.tool_error = True
call.update(status="failed",finishedAt=datetime.now(timezone.utc).isoformat(),outputSummary="工具参数未通过校验")
self.record_event("tool_result",tool=tool_name,toolCallId=call.get("toolCallId", ""),label=call["label"],status="failed",text=call["outputSummary"])
elif kind in ("tool-output-available", "tool-result"):
call_id = event.get("toolCallId")
output = event.get("output")
if isinstance(call_id, str):
self.tool_outputs[call_id] = output
output_failed = _contains_error_flag(output)
matched = False
if output_failed:
self.tool_error = True
tool_name = event.get("toolName")
call_id = event.get("toolCallId")
if isinstance(tool_name, str) or isinstance(call_id, str):
for item in reversed(self.tool_calls):
matches = (item.get("toolCallId") == call_id if isinstance(call_id, str)
else item["tool"] == tool_name)
if matches and item["status"] == "started":
matched = True
item["status"] = "failed" if _contains_error_flag(output) else "completed"
item["finishedAt"] = datetime.now(timezone.utc).isoformat()
item["outputSummary"] = public_tool_failure(output) if item["status"] == "failed" else public_tool_result(output)
self.record_event("tool_result", tool=item["tool"], toolCallId=item.get("toolCallId", ""), label=item["label"], status=item["status"], text=item["outputSummary"], evidenceChecks=public_evidence_checks(output) if not output_failed else {})
if not output_failed:
key = self.tool_operation_keys.get(id(item))
complete_key=self.complete_defect_keys.get(id(item)) if complete_defect_result(output) else None
for previous in self.tool_calls:
if previous is item:
break
same=(key is not None and self.tool_operation_keys.get(id(previous))==key)
complete_same=(complete_key is not None and self.complete_defect_keys.get(id(previous))==complete_key)
if previous["status"] == "failed" and (same or complete_same):
previous["recovered"] = True
previous["outputSummary"] = "调整调用后已重试成功"
break
if output_failed and not matched:
self.unattributed_tool_error = True
elif kind in ("tool-output-error", "error"):
self.tool_error = True
matched = False
if kind == "tool-output-error":
call_id = event.get("toolCallId")
if isinstance(call_id, str):
for item in reversed(self.tool_calls):
if item.get("toolCallId") == call_id and item["status"] in ("started", "failed"):
matched = True
if item["status"] == "failed": break
item["status"] = "failed"
item["finishedAt"] = datetime.now(timezone.utc).isoformat()
item["outputSummary"] = "本次调用未成功"
self.record_event("tool_result", tool=item["tool"], toolCallId=call_id, label=item["label"], status="failed", text="本次调用未成功")
break
if not matched:
self.unattributed_tool_error = True
if kind == "error":
self.error_event = True
elif kind == "[DONE]":
self.done_received = True
@property
def model_reasoning(self) -> str:
return "\n\n".join(self.reasoning_blocks.values())
@property
def answer(self) -> str:
for block_id in reversed(self.block_order):
value = self.blocks.get(block_id, "").strip()
if value:
return value
return ""
def status(self, *, cancel_cause: str | None = None, stop_succeeded: bool | None = None,
transport_failed: bool = False) -> tuple[str, str | None]:
if cancel_cause == "user":
return "CANCELLED", "user_cancel"
if cancel_cause == "timeout":
return ("PARTIAL" if self.answer else "FAILED", "timeout")
if cancel_cause == "transport":
if any(reason == "length" for reason in self.finish_reasons):
return "PARTIAL", "length"
return ("PARTIAL" if self.answer else "FAILED", "transport_or_stream_error")
if any(reason == "length" for reason in self.finish_reasons):
return "PARTIAL", "length"
if transport_failed or self.error_event or not self.finish_reasons:
return ("PARTIAL" if self.answer else "FAILED", "transport_or_stream_error")
if self.tool_error and (self.unattributed_tool_error or any(call["status"] == "failed" and not call.get("recovered") for call in self.tool_calls)):
return ("PARTIAL" if self.answer else "FAILED", "mcp_error")
if self.answer and (self.done_received or "stop" in self.finish_reasons):
return "COMPLETED", "stop"
return ("PARTIAL" if self.answer else "FAILED", "incomplete")
def _contains_error_flag(value: Any) -> bool:
if isinstance(value, dict):
if value.get("isError") is True or value.get("error"):
return True
return any(_contains_error_flag(item) for item in value.values())
if isinstance(value, list):
return any(_contains_error_flag(item) for item in value)
if isinstance(value, str) and len(value) <= 64 * 1024:
stripped = value.strip()
if stripped.startswith(("{", "[")):
try:
return _contains_error_flag(json.loads(stripped))
except json.JSONDecodeError:
return False
return False
class NativeStopController:
"""Stop the native NAO chat before closing its SSE stream on cancel/timeout."""
def __init__(self, stop_call: Callable[[str | None], bool], timeout: float) -> None:
self.stop_call = stop_call
self.timeout = timeout
self.finished = threading.Event()
self._trigger = threading.Event()
self._condition = threading.Condition()
self._chat_id: str | None = None
self._response: Any | None = None
self.cause: str | None = None
self.stop_succeeded: bool | None = None
self.thread = threading.Thread(target=self._watch, name="nao-native-stop", daemon=True)
def start(self) -> None:
self.thread.start()
def set_response(self, response: Any) -> None:
with self._condition:
self._response = response
def set_chat_id(self, chat_id: str | None) -> None:
if not chat_id:
return
with self._condition:
self._chat_id = chat_id
self._condition.notify_all()
def request_cancel(self, cause: str = "user") -> None:
if cause not in ("user", "timeout", "transport"):
return
if not self._trigger.is_set():
self.cause = cause
self._trigger.set()
with self._condition:
self._condition.notify_all()
def _watch(self) -> None:
deadline = time.monotonic() + self.timeout
while not self.finished.is_set():
remaining = deadline - time.monotonic()
if remaining <= 0:
self.request_cancel("timeout")
break
if self._trigger.wait(min(0.1, remaining)):
break
if self.finished.is_set():
return
self.cause = self.cause or "timeout"
deadline = time.monotonic() + CHAT_ID_WAIT_SECONDS
chat_id = None
with self._condition:
while not self.finished.is_set() and not self._chat_id and time.monotonic() < deadline:
self._condition.wait(timeout=min(0.25, max(0.01, deadline - time.monotonic())))
chat_id = self._chat_id
response = self._response
try:
self.stop_succeeded = bool(self.stop_call(chat_id))
except Exception:
self.stop_succeeded = False
if response is not None:
with contextlib.suppress(Exception):
response.close()
def close(self) -> None:
if self._trigger.is_set():
if self.thread.is_alive():
self.thread.join(timeout=7)
if self.thread.is_alive():
self.finished.set()
with contextlib.suppress(Exception):
if self._response is not None:
self._response.close()
self.thread.join(timeout=0.2)
else:
self.finished.set()
with self._condition:
self._condition.notify_all()
if self.thread.is_alive():
self.thread.join(timeout=0.2)
class SessionState:
"""Per-session mapping to native chat IDs, with read-only legacy fallback."""
def __init__(self, runtime_root: Path, actor_id: str, session_id: str | None = None) -> None:
self.actor_id = _uuid(actor_id, "actorId")
actual_actor, path_session = _runtime_identity(runtime_root)
if actual_actor != self.actor_id:
raise ValueError("会话状态与当前 actor 不匹配。")
self.session_id = _uuid(session_id, "sessionId") if session_id is not None else path_session
if path_session is not None and self.session_id != path_session:
raise ValueError("会话状态与当前 session 不匹配。")
self.runtime_root = Path(runtime_root)
self.path = (self.runtime_root / "private" / "native-chat.json" if self.session_id
else self.runtime_root / "private" / "sessions.json")
def get_chat_id(self, session_id: str) -> str | None:
if self.session_id and session_id != self.session_id:
return None
if self.path.is_symlink():
return None
try:
value = _read_json(self.path)
except (OSError, json.JSONDecodeError):
value = None
if self.session_id:
if (isinstance(value, dict) and value.get("actorId") == self.actor_id
and value.get("sessionId") == self.session_id):
result = value.get("nativeChatId")
if isinstance(result, str) and result:
return result
return None
if not isinstance(value, dict) or value.get("actorId") != self.actor_id:
return None
sessions = value.get("sessions")
result = sessions.get(session_id) if isinstance(sessions, dict) else None
return result if isinstance(result, str) and result else None
def set_chat_id(self, session_id: str, chat_id: str) -> None:
session_id = _uuid(session_id, "sessionId")
if self.session_id:
if session_id != self.session_id or not isinstance(chat_id, str) or not chat_id:
raise ValueError("会话状态与当前 session 不匹配。")
self.path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
if self.path.is_symlink():
raise RuntimeErrorSafe("会话状态路径无效。")
_atomic_json(self.path, {"actorId": self.actor_id, "sessionId": session_id,
"nativeChatId": chat_id})
return
self.path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
try:
value = _read_json(self.path)
except (OSError, json.JSONDecodeError):
value = {}
if not isinstance(value, dict) or value.get("actorId") != self.actor_id:
value = {"actorId": self.actor_id, "sessions": {}}
sessions = value.setdefault("sessions", {})
sessions[session_id] = chat_id
_atomic_json(self.path, value)
def _model_settings() -> tuple[str, str, str]:
try:
from nao_model_settings import model_settings
base, key, model = model_settings()
except Exception:
raise RuntimeErrorSafe("所选模型配置不可用。") from None
if not all(isinstance(item, str) and item for item in (base, key, model)):
raise RuntimeErrorSafe("所选模型配置不完整。")
return base, key, model
def _config_fingerprint(model_base: str, model_key: str, model_name: str) -> str:
paths = [TEMPLATES / "RULES.md", TEMPLATES / "agent" / "prompts" / "system.md",
ROOT / "scripts" / "nao_quality_runtime.py", ROOT / "scripts" / "nao_quality_runtime_mcp.py",
ROOT / "scripts" / "nao_quality_tools.py", ROOT / "scripts" / "nao_quality_mcp.py",
ROOT / "scripts" / "nao_model_settings.py", ROOT / "scripts" / "nao_evidence_finalizer.py", ROOT / "scripts" / "nao_evidence_report.py"]
digest = hashlib.sha256()
for path in paths:
digest.update(str(path.relative_to(ROOT)).encode())
digest.update(path.read_bytes())
# The one-way fingerprint detects selected model changes without saving its key.
digest.update(model_base.encode())
digest.update(model_name.encode())
digest.update(hashlib.sha256(model_key.encode()).digest())
try:
sys.path.insert(0, str(ROOT / "scripts"))
from nao_quality_tools import tool_definition
digest.update(json.dumps(tool_definition(), ensure_ascii=False, sort_keys=True, separators=(",", ":")).encode())
except Exception:
raise RuntimeErrorSafe("质量工具定义无法读取。") from None
return digest.hexdigest()
def _write_private_state(actor_root: Path) -> dict[str, str]:
private_path = actor_root / "private" / "credentials.json"
if (actor_root.is_symlink() or private_path.parent.is_symlink()):
raise RuntimeErrorSafe("私有运行目录路径无效。")
private_path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
os.chmod(private_path.parent, 0o700)
if private_path.exists():
_require_private_regular_file(private_path, missing_ok=False)
try:
value = _read_json(private_path)
if all(isinstance(value.get(key), str) and value[key] for key in ("authSecret", "email", "password")):
return value
except (OSError, json.JSONDecodeError, AttributeError):
pass
private = {
"authSecret": secrets.token_urlsafe(48),
"email": f"nao-{uuid.uuid4().hex}@localhost.invalid",
"password": secrets.token_urlsafe(32),
}
_atomic_json(private_path, private)
return private
def _write_project(actor_root: Path, session_id: str, model_name: str, *,
refresh_templates: bool, refresh_mcp: bool, backup_schema: bool,
expected_actor_id: str | None = None) -> None:
session_id = _uuid(session_id, "sessionId")
path_actor_id, path_session_id = _runtime_identity(actor_root)
actor_id = _uuid(expected_actor_id, "actorId") if expected_actor_id is not None else path_actor_id
if actor_id != path_actor_id or (path_session_id is not None and path_session_id != session_id):
raise RuntimeErrorSafe("NAO project 与当前 actor/session 不匹配。")
project = actor_root / "project"
evidence_dir = actor_root / "evidence" if path_session_id else actor_root / session_id / "evidence"
session_state_path = actor_root / "private" / "current-session.json"
for candidate in (project, project / "agent", project / "agent" / "prompts",
project / "RULES.md", project / "nao_config.yaml",
project / "agent" / "prompts" / "system.md",
project / "agent" / "mcps", project / "agent" / "mcps" / "mcp.json",
actor_root / "private", session_state_path, evidence_dir):
_reject_symlink_components(candidate, actor_root)
project.mkdir(parents=True, exist_ok=True, mode=0o700)
(project / "agent" / "prompts").mkdir(parents=True, exist_ok=True, mode=0o700)
(project / "agent" / "mcps").mkdir(parents=True, exist_ok=True, mode=0o700)
for candidate in (project / "RULES.md", project / "nao_config.yaml",
project / "agent" / "prompts" / "system.md",
project / "agent" / "mcps" / "mcp.json"):
_reject_symlink_components(candidate, actor_root)
if refresh_templates:
old_schema = project / "agent" / "mcps" / "quality"
if backup_schema and old_schema.exists():
backup = actor_root / "backups" / datetime.now().strftime("%Y%m%d-%H%M%S-%f") / "generatedspec"
backup.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
shutil.move(str(old_schema), str(backup))
for relative in (Path("RULES.md"), Path("agent/prompts/system.md")):
source = TEMPLATES / relative
target = project / relative
if source.is_symlink() or not source.is_file():
raise RuntimeErrorSafe("NAO project 模板不可用。")
_atomic_bytes(target, source.read_bytes())
config = (
"project_name: quality-center\n"
"databases: []\n"
"llm:\n"
f" annotation_model: openaiCompatible:{model_name}\n"
" providers:\n"
" - provider: openaiCompatible\n"
" api_key: \"{{ env('NAO_MODEL_KEY') }}\"\n"
" base_url: \"{{ env('NAO_MODEL_BASE') }}\"\n"
" models:\n"
f" - id: {json.dumps(model_name, ensure_ascii=False)}\n"
" default: true\n"
" settings:\n"
f" maxOutputTokens: {32768 if model_name == 'mimo-v2.6-flash' else 8192}\n"
" temperature: 0.2\n"
" reasoningEffort: low\n"
)
config_path = project / "nao_config.yaml"
_atomic_bytes(config_path, config.encode("utf-8"))
evidence_dir.mkdir(parents=True, exist_ok=True, mode=0o700)
mcp = {
"mcpServers": {
"quality": {
"transport": "stdio",
"command": str(PYTHON),
"args": [str(ROOT / "scripts" / "nao_quality_runtime_mcp.py")],
"env": {
"QUALITY_AGENT_ACCESS_FILE": str(actor_root / "access.json"),
"QUALITY_AGENT_EVIDENCE_DIR": str(evidence_dir),
"QUALITY_AGENT_SESSION_STATE_FILE": str(session_state_path),
"QUALITY_AGENT_ACTOR_ID": actor_id,
"QUALITY_AGENT_SESSION_ID": session_id,
"QUALITY_AGENT_USERS_ROOT": str(USERS_ROOT.resolve()),
"PYTHONPATH": str(ROOT / "scripts"),
},
}
}
}
mcp_path = project / "agent" / "mcps" / "mcp.json"
if refresh_mcp or not mcp_path.exists():
_atomic_json(mcp_path, mcp)
_atomic_json(session_state_path, {"actorId": actor_id, "sessionId": session_id})
os.chmod(project, 0o700)
def _available_port() -> int:
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
sock.bind(("127.0.0.1", 0))
return int(sock.getsockname()[1])
def _ps_value(pid: int, field: str) -> str | None:
try:
result = subprocess.run(["ps", "-p", str(pid), "-o", f"{field}="], capture_output=True,
text=True, timeout=3, check=False)
except (OSError, subprocess.TimeoutExpired):
return None
if result.returncode != 0:
return None
return result.stdout.strip() or None
def _process_record(pid: int, kind: str, port: int) -> dict[str, Any]:
started = _ps_value(pid, "lstart")
if not started:
raise RuntimeErrorSafe("NAO 本地进程启动失败。")
return {"pid": pid, "pgid": pid, "lstart": started, "kind": kind, "port": port}
def _open_server_log(path: Path, runtime_root: Path) -> int:
fd: int | None = None
try:
_reject_symlink_components(path, runtime_root)
fd = os.open(path, os.O_WRONLY | os.O_CREAT | os.O_APPEND
| getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_NONBLOCK", 0), 0o600)
info = os.fstat(fd)
if not stat.S_ISREG(info.st_mode):
os.close(fd)
fd = None
raise RuntimeErrorSafe("NAO 日志文件路径无效。")
os.fchmod(fd, 0o600)
result = fd
fd = None
return result
except RuntimeErrorSafe:
raise
except (OSError, ValueError):
raise RuntimeErrorSafe("NAO 日志文件路径无效。") from None
finally:
if fd is not None:
with contextlib.suppress(OSError):
os.close(fd)
def _process_matches(record: dict[str, Any], actor_root: Path) -> bool:
try:
pid = int(record["pid"])
if int(record["pgid"]) != pid or os.getpgid(pid) != pid:
return False
except (KeyError, TypeError, ValueError, ProcessLookupError, PermissionError):
return False
start = _ps_value(pid, "lstart")
command = _ps_value(pid, "command") or ""
if not start or start != record.get("lstart"):
return False
kind = record.get("kind")
if kind == "fastapi":
return str(PYTHON) in command and str(FASTAPI_MAIN) in command
if kind == "server":
return (str(NAO_SERVER) in command and "serve" in command
and f"--port {int(record['port'])}" in command
and "--host 127.0.0.1" in command)
return False
def _stop_instance(actor_root: Path, *, force: bool = False, grace_seconds: float | None = None) -> bool:
state_path = actor_root / "instance.json"
if state_path.is_symlink():
raise RuntimeErrorSafe("NAO 实例状态路径无效,已拒绝停止该进程。")
try:
state = _read_json(state_path)
except (OSError, json.JSONDecodeError):
return False
actor_id, session_id = _runtime_identity(actor_root)
if not RuntimeManager._instance_identity_matches(state, actor_root, actor_id, session_id):
raise RuntimeErrorSafe("NAO 实例身份校验失败,已拒绝停止该进程。")
processes = state.get("processes", []) if isinstance(state, dict) else []
owned: list[dict[str, Any]] = []
for record in processes:
if not isinstance(record, dict):
continue
pid = record.get("pid")
if not isinstance(pid, int) or _ps_value(pid, "command") is None:
continue
if not _process_matches(record, actor_root):
raise RuntimeErrorSafe("NAO 进程身份校验失败,已拒绝停止该进程。")
owned.append(record)
for record in owned:
with contextlib.suppress(ProcessLookupError):
os.killpg(record["pgid"], signal.SIGTERM)
grace = grace_seconds if grace_seconds is not None else (8 if force else 5)
deadline = time.monotonic() + grace
while time.monotonic() < deadline and any(_process_matches(r, actor_root) for r in owned):
time.sleep(0.1)
still_owned = [r for r in owned if _process_matches(r, actor_root)]
for record in still_owned:
if not _process_matches(record, actor_root):
continue
with contextlib.suppress(ProcessLookupError):
os.killpg(record["pgid"], signal.SIGKILL)
state_path.unlink(missing_ok=True)
return bool(owned)
@contextlib.contextmanager
def _file_lock(path: Path, *, blocking: bool) -> Iterator[int]:
path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
if path.is_symlink():
raise RuntimeErrorSafe("运行锁路径无效。")
fd = os.open(path, os.O_CREAT | os.O_RDWR | getattr(os, "O_NOFOLLOW", 0), 0o600)
os.fchmod(fd, 0o600)
try:
try:
fcntl.flock(fd, fcntl.LOCK_EX | (0 if blocking else fcntl.LOCK_NB))
except BlockingIOError:
yield -1
else:
yield fd
finally:
if fd >= 0:
with contextlib.suppress(OSError):
fcntl.flock(fd, fcntl.LOCK_UN)
os.close(fd)
def _actor_dirs(users_root: Path = USERS_ROOT) -> list[Path]:
if not users_root.exists() or users_root.is_symlink():
return []
return [path for path in users_root.iterdir()
if not path.is_symlink() and path.is_dir() and _UUID_RE.fullmatch(path.name)]
def _runtime_roots(users_root: Path = USERS_ROOT) -> list[Path]:
"""Enumerate legacy actor roots and isolated actor/session runtime roots."""
roots: list[Path] = []
for actor_root in _actor_dirs(users_root):
# Keep old state readable and stoppable; never migrate or delete its history.
legacy_instance = actor_root / "instance.json"
if not legacy_instance.is_symlink() and legacy_instance.is_file():
roots.append(actor_root)
sessions_root = actor_root / "sessions"
if sessions_root.is_symlink() or not sessions_root.is_dir():
continue
for session_root in sessions_root.iterdir():
if (not session_root.is_symlink() and session_root.is_dir()
and _UUID_RE.fullmatch(session_root.name)
and not (session_root / "instance.json").is_symlink()):
roots.append(session_root)
return roots
def _rebind_legacy_project_path(database: sqlite3.Connection, chat_id: str,
legacy_project: Path, session_project: Path) -> None:
"""Rebind the copied NAO project row while preserving its id and chat links."""
tables = {row[0] for row in database.execute(
"SELECT name FROM sqlite_master WHERE type = 'table'")}
if "project" not in tables:
return
project_columns = {row[1] for row in database.execute('PRAGMA table_info("project")')}
if not {"id", "path"}.issubset(project_columns):
raise RuntimeErrorSafe("旧质量会话 project schema 无法安全迁移,历史数据已保留。")
old_path = str(legacy_project.absolute())
new_path = str(session_project.absolute())
chat_project_id: str | None = None
if "chat" in tables:
chat_columns = {row[1] for row in database.execute('PRAGMA table_info("chat")')}
if {"id", "project_id"}.issubset(chat_columns):
chat_row = database.execute(
'SELECT "project_id" FROM "chat" WHERE "id" = ?', (chat_id,)
).fetchone()
if chat_row is not None:
chat_project_id = chat_row[0]
if chat_project_id is not None:
project_row = database.execute(
'SELECT "path" FROM "project" WHERE "id" = ?', (chat_project_id,)
).fetchone()
if project_row is None:
raise RuntimeErrorSafe("旧质量会话 chat/project 关系无法验证,历史数据已保留。")
current_path = project_row[0]
if current_path == new_path:
return
if current_path not in (None, "", old_path):
raise RuntimeErrorSafe("旧质量会话 project 路径与 actor 不匹配,历史数据已保留。")
updated = database.execute(
'UPDATE "project" SET "path" = ? WHERE "id" = ? AND "path" IS ?',
(new_path, chat_project_id, current_path),
)
if updated.rowcount != 1:
raise RuntimeErrorSafe("旧质量会话 project 路径无法安全迁移,历史数据已保留。")
return
# Older NAO snapshots can predate chat.project_id. In that case only move
# the unique project that points at this actor's legacy project directory.
rows = database.execute(
'SELECT "id" FROM "project" WHERE "path" = ?', (old_path,)
).fetchall()
if not rows:
return
if len(rows) != 1:
raise RuntimeErrorSafe("旧质量会话 project 路径不唯一,历史数据已保留。")
updated = database.execute(
'UPDATE "project" SET "path" = ? WHERE "id" = ? AND "path" = ?',
(new_path, rows[0][0], old_path),
)
if updated.rowcount != 1:
raise RuntimeErrorSafe("旧质量会话 project 路径无法安全迁移,历史数据已保留。")
class RuntimeManager:
def __init__(self, users_root: Path = USERS_ROOT) -> None:
users_root = Path(users_root).absolute()
if users_root.is_symlink():
raise ValueError("运行目录无效。")
self.users_root = users_root.resolve()
def _actor_root(self, actor_id: str) -> Path:
actor_id = _uuid(actor_id, "actorId")
self.users_root.mkdir(parents=True, exist_ok=True, mode=0o700)
os.chmod(self.users_root, 0o700)
path = self.users_root / actor_id
if path.is_symlink():
raise RuntimeErrorSafe("actor 运行目录无效。")
path.mkdir(parents=True, exist_ok=True, mode=0o700)
os.chmod(path, 0o700)
return path
def _session_root(self, actor_id: str, session_id: str) -> Path:
actor_id = _uuid(actor_id, "actorId")
session_id = _uuid(session_id, "sessionId")
actor_root = self._actor_root(actor_id)
sessions_root = actor_root / "sessions"
if sessions_root.is_symlink():
raise RuntimeErrorSafe("session 运行目录无效。")
sessions_root.mkdir(mode=0o700, exist_ok=True)
os.chmod(sessions_root, 0o700)
session_root = sessions_root / session_id
if session_root.is_symlink():
raise RuntimeErrorSafe("session 运行目录无效。")
session_root.mkdir(mode=0o700, exist_ok=True)
os.chmod(session_root, 0o700)
return session_root
def acquire_actor(self, actor_id: str, session_id: str | None = None) -> tuple[Path, int]:
# The no-session form is retained for legacy callers. Actual analysis
# turns always pass Java's session UUID and receive their own lock.
actor_root = self._actor_root(actor_id) if session_id is None else self._session_root(actor_id, session_id)
lock_path = actor_root / "active.lock"
if lock_path.is_symlink():
raise RuntimeErrorSafe("运行锁路径无效。")
fd = os.open(lock_path, os.O_CREAT | os.O_RDWR | getattr(os, "O_NOFOLLOW", 0), 0o600)
os.fchmod(fd, 0o600)
try:
fcntl.flock(fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
except BlockingIOError:
os.close(fd)
raise RuntimeErrorSafe("该会话正处理另一轮分析,请稍后重试。") from None
return actor_root, fd
def release_actor(self, fd: int) -> None:
with contextlib.suppress(OSError):
fcntl.flock(fd, fcntl.LOCK_UN)
os.close(fd)
@staticmethod
def _current_session(actor_root: Path, expected_actor_id: str | None = None,
expected_session_id: str | None = None) -> str | None:
try:
path = actor_root / "private" / "current-session.json"
if path.is_symlink():
return None
value = _read_json(path)
actor_id, root_session_id = _runtime_identity(actor_root)
expected_actor = _uuid(expected_actor_id, "actorId") if expected_actor_id else actor_id
session_id = _uuid(value.get("sessionId"), "sessionId") if isinstance(value, dict) else None
if (isinstance(value, dict) and value.get("actorId") == expected_actor == actor_id
and (root_session_id is None or session_id == root_session_id)
and (expected_session_id is None or session_id == _uuid(expected_session_id, "sessionId"))):
return session_id
except (OSError, json.JSONDecodeError, ValueError):
return None
return None
def ensure_instance(self, actor_root: Path, session_id: str,
expected_actor_id: str | None = None) -> tuple[str, str, str, str]:
session_id = _uuid(session_id, "sessionId")
actor_id, root_session_id = _runtime_identity(actor_root)
if expected_actor_id is not None and _uuid(expected_actor_id, "actorId") != actor_id:
raise RuntimeErrorSafe("NAO runtime 与当前 actor 不匹配。")
if root_session_id is not None and root_session_id != session_id:
raise RuntimeErrorSafe("NAO runtime 与当前 session 不匹配。")
if root_session_id is not None:
self._migrate_legacy_session(actor_id, session_id, actor_root)
model_base, model_key, model_name = _model_settings()
fingerprint = _config_fingerprint(model_base, model_key, model_name)
_write_private_state(actor_root)
state_path = actor_root / "instance.json"
state = None
try:
state = _read_json(state_path)
except (OSError, json.JSONDecodeError):
pass
current_session = self._current_session(actor_root, actor_id, session_id)
project = actor_root / "project"
project_ready = all((project / relative).is_file() and not (project / relative).is_symlink()
for relative in ("RULES.md", "nao_config.yaml", "agent/prompts/system.md", "agent/mcps/mcp.json"))
refresh_templates = not project_ready or not isinstance(state, dict) or state.get("fingerprint") != fingerprint
refresh_mcp = not project_ready or current_session != session_id
backup_schema = isinstance(state, dict) and state.get("fingerprint") != fingerprint
if isinstance(state, dict):
procs = state.get("processes", [])
alive = (self._instance_identity_matches(state, actor_root, actor_id, root_session_id)
and all(isinstance(r, dict) and _process_matches(r, actor_root) for r in procs)
and len(procs) == 2)
if alive and state.get("fingerprint") == fingerprint and not refresh_mcp:
if self._wait_server(int(state["serverPort"]), timeout=0.5):
return model_base, model_key, model_name, f"http://127.0.0.1:{int(state['serverPort'])}"
# Stop only processes that still match the recorded PID, start time and process group.
_stop_instance(actor_root)
self._with_registry_lock(actor_root, session_id, model_base, model_key, model_name, fingerprint,
refresh_templates=refresh_templates, refresh_mcp=refresh_mcp,
backup_schema=backup_schema, expected_actor_id=actor_id)
state = _read_json(state_path)
return model_base, model_key, model_name, f"http://127.0.0.1:{int(state['serverPort'])}"
def _migrate_legacy_session(self, actor_id: str, session_id: str, session_root: Path) -> bool:
"""Import one old actor-scoped chat/database snapshot into its session root."""
legacy_root = self.users_root / _uuid(actor_id, "actorId")
legacy_private = legacy_root / "private"
mapping_path = legacy_private / "sessions.json"
if legacy_root.is_symlink() or legacy_private.is_symlink() or mapping_path.is_symlink():
raise RuntimeErrorSafe("旧质量会话记录路径无效,历史数据已保留。")
if not mapping_path.is_file():
return False
try:
mapping = _read_json(mapping_path)
except (OSError, json.JSONDecodeError):
raise RuntimeErrorSafe("旧质量会话记录无法读取,历史数据已保留。") from None
if not isinstance(mapping, dict) or mapping.get("actorId") != actor_id:
raise RuntimeErrorSafe("旧质量会话 actor 校验失败,历史数据已保留。")
sessions = mapping.get("sessions")
chat_id = sessions.get(session_id) if isinstance(sessions, dict) else None
if not isinstance(chat_id, str) or not chat_id:
return False
private_root = session_root / "private"
private_root.mkdir(parents=True, exist_ok=True, mode=0o700)
os.chmod(private_root, 0o700)
target_db = private_root / "nao.sqlite"
target_chat = private_root / "native-chat.json"
_reject_symlink_components(target_db, self.users_root)
_reject_symlink_components(target_chat, self.users_root)
if target_db.exists() and target_chat.exists():
_require_private_regular_file(target_db, missing_ok=False)
_require_private_regular_file(target_chat, missing_ok=False)
try:
current_chat = _read_json(target_chat)
except (OSError, json.JSONDecodeError):
raise RuntimeErrorSafe("当前质量会话记录无法验证,历史数据已保留。") from None
if (not isinstance(current_chat, dict) or current_chat.get("actorId") != actor_id
or current_chat.get("sessionId") != session_id):
raise RuntimeErrorSafe("当前质量会话 actor/session 校验失败。")
return False
if target_db.exists() or target_chat.exists():
raise RuntimeErrorSafe("旧质量会话导入状态不完整,历史数据已保留。")
source_db = legacy_private / "nao.sqlite"
source_credentials = legacy_private / "credentials.json"
if (source_db.is_symlink() or not source_db.is_file()
or source_credentials.is_symlink() or not source_credentials.is_file()):
raise RuntimeErrorSafe("旧质量会话数据库或本地账号凭据不可用,历史数据已保留。")
try:
credentials = _read_json(source_credentials)
except (OSError, json.JSONDecodeError):
raise RuntimeErrorSafe("旧质量会话本地账号凭据不可用,历史数据已保留。") from None
if not isinstance(credentials, dict) or not all(
isinstance(credentials.get(key), str) and credentials[key]
for key in ("authSecret", "email", "password")):
raise RuntimeErrorSafe("旧质量会话本地账号凭据不完整,历史数据已保留。")
_require_private_regular_file(source_db, missing_ok=False)
_require_private_regular_file(source_credentials, missing_ok=False)
temporary_db = private_root / f".nao.sqlite.{secrets.token_hex(6)}.tmp"
try:
try:
_reject_symlink_components(legacy_root / "project", self.users_root)
_reject_symlink_components(session_root / "project", self.users_root)
except ValueError:
raise RuntimeErrorSafe("旧质量会话 project 路径无效,历史数据已保留。") from None
source_uri = "file:" + quote(str(source_db.absolute()), safe="/") + "?mode=ro"
source = sqlite3.connect(source_uri, uri=True, timeout=5)
target = sqlite3.connect(temporary_db, timeout=5)
try:
source.backup(target)
_rebind_legacy_project_path(target, chat_id,
legacy_root / "project", session_root / "project")
target.commit()
finally:
target.close()
source.close()
os.chmod(temporary_db, 0o600)
os.replace(temporary_db, target_db)
_atomic_json(private_root / "credentials.json", credentials)
_atomic_json(target_chat, {"actorId": actor_id, "sessionId": session_id,
"nativeChatId": chat_id})
_atomic_json(private_root / "legacy-import.json", {
"actorId": actor_id, "sessionId": session_id,
"importedAt": datetime.now(timezone.utc).isoformat(),
})
return True
except RuntimeErrorSafe:
temporary_db.unlink(missing_ok=True)
target_db.unlink(missing_ok=True)
raise
except (OSError, sqlite3.Error):
temporary_db.unlink(missing_ok=True)
target_db.unlink(missing_ok=True)
raise RuntimeErrorSafe("旧质量会话数据库无法安全导入,历史数据已保留。") from None
def _with_registry_lock(self, actor_root: Path, session_id: str, model_base: str, model_key: str,
model_name: str, fingerprint: str, *, refresh_templates: bool,
refresh_mcp: bool, backup_schema: bool,
expected_actor_id: str | None = None) -> None:
actor_id, session_root_id = _runtime_identity(actor_root)
actor_id = _uuid(expected_actor_id, "actorId") if expected_actor_id is not None else actor_id
if actor_id != _runtime_identity(actor_root)[0] or (session_root_id is not None and session_root_id != session_id):
raise RuntimeErrorSafe("NAO runtime 与当前 actor/session 不匹配。")
with _file_lock(self.users_root / "registry.lock", blocking=True):
active = []
for other in _runtime_roots(self.users_root):
try:
state = _read_json(other / "instance.json")
except (OSError, json.JSONDecodeError):
continue
procs = state.get("processes", []) if isinstance(state, dict) else []
other_actor_id, other_session_id = _runtime_identity(other)
if (self._instance_identity_matches(state, other, other_actor_id, other_session_id)
and len(procs) == 2
and all(isinstance(r, dict) and _process_matches(r, other) for r in procs)):
active.append(other)
if actor_root not in active and len(active) >= MAX_INSTANCES:
evicted = False
for other in active:
if other == actor_root:
continue
with _file_lock(other / "active.lock", blocking=False) as lock_fd:
if lock_fd < 0:
continue
_stop_instance(other)
evicted = True
break
if not evicted:
raise RuntimeErrorSafe("当前已有两个质量分析实例正在使用,暂时无法启动新实例。")
_write_project(actor_root, session_id, model_name, refresh_templates=refresh_templates,
refresh_mcp=refresh_mcp, backup_schema=backup_schema,
expected_actor_id=actor_id)
self._start(actor_root, model_base, model_key, model_name, fingerprint)
@staticmethod
def _instance_identity_matches(state: Any, runtime_root: Path,
actor_id: str, session_id: str | None) -> bool:
if not isinstance(state, dict) or state.get("actorId") != actor_id:
return False
if session_id is None:
return state.get("sessionId") in (None, "")
return state.get("sessionId") == session_id
def _start(self, actor_root: Path, model_base: str, model_key: str, model_name: str, fingerprint: str,
*, expected_actor_id: str | None = None, session_id: str | None = None) -> None:
actor_id, path_session_id = _runtime_identity(actor_root)
actor_id = _uuid(expected_actor_id, "actorId") if expected_actor_id is not None else actor_id
session_id = path_session_id or (_uuid(session_id, "sessionId") if session_id else None)
if actor_id != _runtime_identity(actor_root)[0] or path_session_id not in (None, session_id):
raise RuntimeErrorSafe("NAO runtime 与当前 actor/session 不匹配。")
private = _write_private_state(actor_root)
project = actor_root / "project"
logs = actor_root / "logs"
try:
_reject_symlink_components(logs, actor_root)
except ValueError:
raise RuntimeErrorSafe("NAO 日志路径无效。")
logs.mkdir(parents=True, exist_ok=True, mode=0o700)
os.chmod(logs, 0o700)
fastapi_port = _available_port()
server_port = _available_port()
while server_port == fastapi_port:
server_port = _available_port()
env = os.environ.copy()
env.update({
"NAO_MODEL_BASE": model_base,
"NAO_MODEL_KEY": model_key,
"NAO_MODEL_NAME": model_name,
"NAO_DEFAULT_PROJECT_PATH": str(project),
"DB_URI": "sqlite:" + str(actor_root / "private" / "nao.sqlite"),
"BETTER_AUTH_URL": f"http://127.0.0.1:{server_port}",
"BETTER_AUTH_SECRET": private["authSecret"],
"FASTAPI_PORT": str(fastapi_port),
"PORT": str(fastapi_port),
"SERVER_PORT": str(server_port),
"ENABLE_USER_SIGNUP": "false",
"MODE": "prod",
"NAO_CORE_VERSION": "0.3.26",
"POSTHOG_DISABLED": "true",
"DO_NOT_TRACK": "1",
"OTEL_SDK_DISABLED": "true",
"ANONYMIZED_TELEMETRY": "False",
})
for name in ("nao.sqlite",):
path = actor_root / "private" / name
if path.is_symlink():
raise RuntimeErrorSafe("NAO 数据库路径无效。")
path.touch(mode=0o600, exist_ok=True)
path.chmod(0o600)
processes: list[dict[str, Any]] = []
log_path = logs / "server.log"
fd = _open_server_log(log_path, actor_root)
try:
with os.fdopen(fd, "ab") as log:
fastapi = subprocess.Popen(
[str(PYTHON), str(FASTAPI_MAIN)], cwd=project, env=env,
stdin=subprocess.DEVNULL, stdout=log, stderr=log, start_new_session=True,
)
processes.append(_process_record(fastapi.pid, "fastapi", fastapi_port))
server = subprocess.Popen(
[str(NAO_SERVER), "serve", "--port", str(server_port), "--host", "127.0.0.1"],
cwd=NAO_SERVER.parent, env=env, stdin=subprocess.DEVNULL,
stdout=log, stderr=log, start_new_session=True,
)
processes.append(_process_record(server.pid, "server", server_port))
except Exception:
for record in processes:
if _process_matches(record, actor_root):
with contextlib.suppress(ProcessLookupError):
os.killpg(record["pgid"], signal.SIGTERM)
raise RuntimeErrorSafe("NAO 本地运行实例启动失败。") from None
state = {
"actorId": actor_id, "sessionId": session_id,
"serverPort": server_port, "fastapiPort": fastapi_port,
"fingerprint": fingerprint, "processes": processes,
"startedAt": datetime.now(timezone.utc).isoformat(),
}
_atomic_json(actor_root / "instance.json", state)
if not self._wait_server(server_port, timeout=30):
_stop_instance(actor_root)
raise RuntimeErrorSafe("NAO 本地服务未能就绪。")
@staticmethod
def _wait_server(port: int, timeout: float) -> bool:
deadline = time.monotonic() + timeout
while time.monotonic() < deadline:
try:
with socket.create_connection(("127.0.0.1", port), timeout=0.25):
return True
except OSError:
time.sleep(0.1)
return False
def stop_idle_all(self) -> int:
stopped = 0
self.users_root.mkdir(parents=True, exist_ok=True, mode=0o700)
with _file_lock(self.users_root / "registry.lock", blocking=True):
for actor_root in _runtime_roots(self.users_root):
with _file_lock(actor_root / "active.lock", blocking=False) as lock_fd:
if lock_fd < 0:
continue
if _stop_instance(actor_root):
stopped += 1
(actor_root / "access.json").unlink(missing_ok=True)
return stopped
def write_access_file(actor_root: Path, token: str, release_id: str, *,
expected_actor_id: str | None = None,
expected_session_id: str | None = None) -> Path:
path_actor_id, path_session_id = _runtime_identity(actor_root)
actor_id = _uuid(expected_actor_id, "actorId") if expected_actor_id is not None else path_actor_id
session_id = (_uuid(expected_session_id, "sessionId") if expected_session_id is not None
else path_session_id)
if actor_id != path_actor_id or (path_session_id is not None and session_id != path_session_id):
raise RuntimeErrorSafe("访问凭据与当前 actor/session 不匹配。")
access_path = actor_root / "access.json"
value = {"token": token, "releaseId": _uuid(release_id, "releaseId"),
"queryUrl": RUNTIME_QUERY_URL, "actorId": actor_id}
if session_id is not None:
value["sessionId"] = session_id
_atomic_json(access_path, value)
return access_path
def _make_opener(actor_root: Path, base_url: str) -> Any:
cookie_path = actor_root / "private" / "cookies.txt"
if actor_root.is_symlink() or cookie_path.parent.is_symlink():
raise RuntimeErrorSafe("NAO 本地会话凭据路径无效。")
cookie_path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
if cookie_path.exists():
_require_private_regular_file(cookie_path, missing_ok=False)
jar = http.cookiejar.LWPCookieJar(str(cookie_path))
if cookie_path.exists():
try:
jar.load(ignore_discard=True, ignore_expires=True)
except (OSError, http.cookiejar.LoadError):
raise RuntimeErrorSafe("NAO 本地会话凭据不可用。") from None
opener = build_opener(HTTPCookieProcessor(jar))
opener._nao_cookie_jar = jar
opener._nao_cookie_path = cookie_path
opener._nao_base_url = base_url
return opener
def _save_cookies(client: Any) -> None:
jar = getattr(client, "_nao_cookie_jar", None)
path = getattr(client, "_nao_cookie_path", None)
if jar is not None and path is not None:
if Path(path).is_symlink():
raise RuntimeErrorSafe("NAO 本地会话凭据路径无效。")
jar.save(ignore_discard=True, ignore_expires=True)
Path(path).chmod(0o600)
def _http_json(client: Any, request: Request, timeout: float = 20) -> tuple[int, Any]:
try:
with client.open(request, timeout=timeout) as response:
body = response.read(1024 * 1024)
return response.status, json.loads(body.decode("utf-8")) if body else {}
except HTTPError as error:
try:
data = json.loads(error.read(1024 * 1024).decode("utf-8", "replace"))
except Exception:
data = None
return error.code, data
except (URLError, OSError, TimeoutError):
raise RuntimeErrorSafe("NAO 本地服务暂不可用。") from None
def _public_config(client: Any, base_url: str) -> dict[str, Any]:
url = base_url + "/api/trpc/system.getPublicConfig?input=" + quote('{"json":null}')
request = Request(url, headers={"Accept": "application/json", "Origin": base_url})
status, value = _http_json(client, request)
result = value.get("result", {}).get("data", {}).get("json") if isinstance(value, dict) else None
if status != 200 or not isinstance(result, dict) or not isinstance(result.get("enableUserSignup"), bool):
raise RuntimeErrorSafe("NAO 本地注册状态无法验证。")
return result
def _authenticate(actor_root: Path, base_url: str) -> Any:
private = _write_private_state(actor_root)
client = _make_opener(actor_root, base_url)
config = _public_config(client, base_url)
body = {"email": private["email"], "password": private["password"], "name": "Quality Actor"}
if config["enableUserSignup"]:
endpoint = "sign-up/email"
else:
endpoint = "sign-in/email"
request = Request(base_url + "/api/auth/" + endpoint,
data=json.dumps(body, separators=(",", ":")).encode("utf-8"),
headers={"Content-Type": "application/json", "Origin": base_url,
"Accept": "application/json"})
status, _ = _http_json(client, request)
if status >= 400 and endpoint == "sign-up/email":
# A concurrent/bootstrap edge may have created the one permitted account.
request = Request(base_url + "/api/auth/sign-in/email",
data=json.dumps(body, separators=(",", ":")).encode("utf-8"),
headers={"Content-Type": "application/json", "Origin": base_url,
"Accept": "application/json"})
status, _ = _http_json(client, request)
if status >= 400:
raise RuntimeErrorSafe("NAO 本地账号认证失败。")
_save_cookies(client)
config = _public_config(client, base_url)
if config.get("enableUserSignup") is not False:
raise RuntimeErrorSafe("NAO 本地注册状态未关闭,已停止本轮分析。")
return client
def _stop_native(client: Any, base_url: str, chat_id: str) -> bool:
body = json.dumps({"json": {"chatId": chat_id}}, separators=(",", ":")).encode("utf-8")
request = Request(base_url + "/api/trpc/chat.stop", data=body,
headers={"Content-Type": "application/json", "Origin": base_url,
"Accept": "application/json"}, method="POST")
try:
with client.open(request, timeout=NATIVE_STOP_TIMEOUT_SECONDS) as response:
response.read(1024 * 1024)
return 200 <= response.status < 300
except Exception:
return False
def _story_files(client: Any, base_url: str, chat_id: str | None,
tool_calls: list[dict[str, str]], tool_outputs: dict[str, Any]) -> tuple[str | None, str | None]:
if not chat_id:
return None, None
# Later story calls may replace an earlier report. Follow invocation order,
# newest first, and return the matching markdown/HTML pair from one call.
for story in reversed(tool_calls):
call_id = story.get("toolCallId")
if story.get("tool") != "story" or not isinstance(call_id, str):
continue
output = tool_outputs.get(call_id)
if not isinstance(output, dict) or output.get("success") is not True:
continue
markdown = output.get("code") if isinstance(output.get("code"), str) else None
slug = output.get("id")
if not isinstance(slug, str) or not re.fullmatch(r"[A-Za-z0-9_-]{1,128}", slug):
return markdown, None
query = {"json": {"chatId": chat_id, "storySlug": slug, "format": "html"}}
url = base_url + "/api/trpc/story.download?input=" + quote(json.dumps(query, separators=(",", ":")))
request = Request(url, headers={"Accept": "application/json", "Origin": base_url})
try:
with client.open(request, timeout=30) as response:
envelope = json.loads(response.read(MAX_HTML_BYTES * 2).decode("utf-8"))
export = envelope.get("result", {}).get("data", {}).get("json", {})
encoded = export.get("data") if isinstance(export, dict) else None
if not isinstance(encoded, str) or len(encoded) > MAX_HTML_BYTES * 2:
return markdown, None
html = base64.b64decode(encoded, validate=True)
if len(html) > MAX_HTML_BYTES:
return markdown, None
return markdown, html.decode("utf-8", "replace")
except Exception:
return markdown, None
return None, None
def _sse_request(
*, client: Any, base_url: str, model: str, question: str, release_id: str,
context: Any, chat_id: str | None, session_state: SessionState, session_id: str,
actor_root: Path, emit: Callable[[dict[str, Any]], None], cancel_event: threading.Event | None = None,
timeout: float = TURN_TIMEOUT_SECONDS,
) -> tuple[ConversationAccumulator, str, str | None, bool | None]:
context_text = ""
if context not in (None, "", [], {}):
context_text = "\n会话上下文(仅用于理解本轮追问,不改变数据权限):\n" + json.dumps(
context, ensure_ascii=False, separators=(",", ":"), allow_nan=False,
)
prompt = (f"【仅内部查询约束,禁止向用户复述】本轮数据查询使用 releaseId={release_id},后续所有数据查询必须沿用该标识。"
"如果当前数据下的证据不足,请如实说明,不要切换数据。用户可见的欢迎语、回答和报告只写业务问题、期间、范围、指标、结果与必要局限;不展示 releaseId、发布版本号、Story ID/version,也不说‘固定版本’。不要展示‘原始明细不发AI’等内部流程提示。"
+ "\n【平台配图约束】图表由平台从 quality_data 的真实结果自动绘制。只查询必要的正式统计;禁止为图表调用 execute_sql、数据库或 story,不探测图表服务,不声称前端图表不可用。拿到所需数据即完成文字答复。"
+ context_text + "\n\n用户问题:\n" + question)
body: dict[str, Any] = {
"message": {"text": prompt},
"model": {"provider": "openaiCompatible", "modelId": model},
"timezone": "Asia/Shanghai",
}
if chat_id:
body["chatId"] = chat_id
request = Request(base_url + "/api/agent/", data=json.dumps(body, ensure_ascii=False).encode("utf-8"),
headers={"Content-Type": "application/json", "Origin": base_url,
"Accept": "text/event-stream"}, method="POST")
def stop_or_shutdown(value: str | None) -> bool:
stopped = bool(value and _stop_native(client, base_url, value))
if stopped:
return True
try:
# A lost SSE transport must not leave a native model task running.
return bool(_stop_instance(actor_root, grace_seconds=1.0))
except Exception:
return False
stopper = NativeStopController(stop_or_shutdown, timeout)
bridge: threading.Thread | None = None
bridge_done = threading.Event()
if cancel_event is not None:
def bridge_cancel() -> None:
while not bridge_done.wait(0.05):
if cancel_event.is_set():
stopper.request_cancel("user")
return
bridge = threading.Thread(target=bridge_cancel, name="nao-cancel-bridge", daemon=True)
bridge.start()
current_answer = {"value": "", "last_emit": 0.0}
answer_lock = threading.Lock()
def publish_answer(value: str) -> None:
with answer_lock:
current_answer["value"] = value
now = time.monotonic()
should_emit = now - current_answer["last_emit"] >= 0.35
if should_emit:
current_answer["last_emit"] = now
if should_emit:
emit({"type": "answer", "answer": value})
accumulator = ConversationAccumulator(publish_answer)
last_reasoning_emit = 0.0
accumulator.chat_id = chat_id
stopper.set_chat_id(chat_id)
stopper.start()
transport_failed = False
try:
# urllib's timeout also applies to each blocking SSE read. Keep a generous
# idle interval; the independent watchdog enforces the 10-minute turn cap.
response = client.open(request, timeout=SSE_IDLE_TIMEOUT_SECONDS)
stopper.set_response(response)
with response:
for raw in response:
if stopper.cause:
break
line = raw.decode("utf-8", "replace").strip()
if not line.startswith("data:"):
continue
data = line[5:].strip()
if data == "[DONE]":
accumulator.done_received = True
break
try:
event = json.loads(data)
except json.JSONDecodeError:
continue
if not isinstance(event, dict):
continue
previous_chat = accumulator.chat_id
accumulator.feed(event)
if accumulator.chat_id and accumulator.chat_id != previous_chat:
session_state.set_chat_id(session_id, accumulator.chat_id)
stopper.set_chat_id(accumulator.chat_id)
kind = event.get("type")
if kind in ("text-delta", "text-end", "tool-input-available", "tool-output-available", "tool-result", "tool-output-error"):
now = time.monotonic()
if kind != "text-delta" or now - last_reasoning_emit >= 0.35:
last_reasoning_emit = now
latest = accumulator.tool_calls[-1] if accumulator.tool_calls else {}
emit({"type": "progress", "message": latest.get("label", "正在整理分析说明…"),
"toolCount": len(accumulator.tool_calls), "toolCalls": accumulator.tool_calls,
"timeline": accumulator.timeline})
elif kind == "reasoning-delta" and time.monotonic() - last_reasoning_emit >= 15:
last_reasoning_emit = time.monotonic()
emit({"type":"progress","message":"模型正在整理已返回的查询证据…","toolCount":len(accumulator.tool_calls),"toolCalls":accumulator.tool_calls,"timeline":accumulator.timeline})
# A clean native finish or [DONE] closes the task. A plain EOF can leave
# the upstream model task alive, so stop it before returning to Java.
if not stopper.cause and not accumulator.done_received and "stop" not in accumulator.finish_reasons:
stopper.request_cancel("transport")
except Exception:
transport_failed = True
stopper.request_cancel("transport")
finally:
stopper.close()
if bridge and bridge.is_alive():
bridge_done.set()
bridge.join(timeout=0.1)
cancel_cause = stopper.cause
status, stop_reason = accumulator.status(cancel_cause=cancel_cause,
stop_succeeded=stopper.stop_succeeded,
transport_failed=transport_failed)
# status and stop reason are attached by caller; keep return transport-neutral.
if status == "COMPLETED":
# The Java result mapper uses the native finish reason to recognize a full turn.
stop_reason = stop_reason or "stop"
elif status == "CANCELLED":
stop_reason = "用户停止" if cancel_cause == "user" else "超时停止"
return accumulator, status, stop_reason, stopper.stop_succeeded
def _normalize_request(raw: Any) -> dict[str, Any]:
if not isinstance(raw, dict):
raise ValueError("请求必须是 JSON 对象。")
turn_id = _uuid(raw.get("turnId"), "turnId")
session_id = _uuid(raw.get("sessionId"), "sessionId")
actor_id = _uuid(raw.get("actorId"), "actorId")
release_id = _uuid(raw.get("releaseId"), "releaseId")
token = raw.get("accessToken")
question = raw.get("question")
if not isinstance(token, str) or not token.strip() or len(token) > 4096 or "\n" in token or "\r" in token:
raise ValueError("accessToken 无效。")
if not isinstance(question, str) or not question.strip() or len(question) > MAX_QUESTION_CHARS:
raise ValueError("question 无效或超过长度限制。")
context = raw.get("context")
try:
context_size = len(json.dumps(context, ensure_ascii=False, separators=(",", ":"))) if context is not None else 0
except (TypeError, ValueError):
raise ValueError("context 必须是 JSON 数据。") from None
if context_size > MAX_CONTEXT_CHARS:
raise ValueError("context 超过允许大小。")
return {"turnId": turn_id, "sessionId": session_id, "actorId": actor_id,
"releaseId": release_id, "accessToken": token, "question": question.strip(), "context": context}
def _emit(stream: Any, lock: threading.Lock, value: dict[str, Any]) -> None:
with lock:
stream.write(json.dumps(value, ensure_ascii=False, separators=(",", ":"), allow_nan=False) + "\n")
stream.flush()
def _input_cancel_listener(stream: Any, on_cancel: Callable[[], None]) -> threading.Thread:
def listen() -> None:
for line in stream:
try:
value = json.loads(line)
except json.JSONDecodeError:
continue
if isinstance(value, dict) and value.get("type") == "cancel":
on_cancel()
return
thread = threading.Thread(target=listen, name="nao-worker-cancel-input", daemon=True)
thread.start()
return thread
def run_turn(payload: dict[str, Any], *, stdout: Any = sys.stdout, stdin: Any = sys.stdin) -> dict[str, Any]:
output_lock = threading.Lock()
emit = lambda value: _emit(stdout, output_lock, value)
actor_root: Path | None = None
actor_fd: int | None = None
access_path: Path | None = None
try:
request = _normalize_request(payload)
emit({"type": "progress", "message": "正在准备独立质量分析环境…"})
manager = RuntimeManager()
actor_root, actor_fd = manager.acquire_actor(request["actorId"], request["sessionId"])
access_path = write_access_file(actor_root, request["accessToken"], request["releaseId"],
expected_actor_id=request["actorId"],
expected_session_id=request["sessionId"])
del request["accessToken"]
model_base, model_key, model_name, base_url = manager.ensure_instance(
actor_root, request["sessionId"], expected_actor_id=request["actorId"])
# The model process receives only NAO's configured model credential. The delegated
# user token remains in access.json, read only by the MCP shim.
selected_connection = (model_base, model_key, model_name)
del model_base, model_key
client = _authenticate(actor_root, base_url)
state = SessionState(actor_root, request["actorId"], request["sessionId"])
native_chat_id = state.get_chat_id(request["sessionId"])
emit({"type": "progress", "message": "正在查询质量数据分析平台并整理结论…"})
cancel_event = threading.Event()
def request_cancel() -> None:
cancel_event.set()
_input_cancel_listener(stdin, request_cancel)
prior_evidence_files = frozenset(path.name for path in (actor_root / "evidence").glob("nao-quality-*.json"))
visible_answer_path = actor_root / "private" / "last-visible-answer.json"
if native_chat_id and _require_private_regular_file(visible_answer_path):
previous_visible = _read_json(visible_answer_path)
if (isinstance(previous_visible, dict) and previous_visible.get("releaseId") == request["releaseId"]
and previous_visible.get("sessionId") == request["sessionId"]
and isinstance(previous_visible.get("answer"), str)):
context = {**(request["context"] if isinstance(request["context"], dict) else {}),
"lastVisibleAnswer": previous_visible["answer"]}
if len(json.dumps(context, ensure_ascii=False)) <= MAX_CONTEXT_CHARS:
request["context"] = context
accumulator, status, stop_reason, stop_succeeded = _sse_request(
client=client, base_url=base_url, model=model_name, question=request["question"],
release_id=request["releaseId"], context=request["context"], chat_id=native_chat_id,
session_state=state, session_id=request["sessionId"], emit=emit,
actor_root=actor_root, cancel_event=cancel_event,
timeout=480 if model_name == "mimo-v2.6-flash" else TURN_TIMEOUT_SECONDS,
)
from nao_evidence_finalizer import eligible, evidence_summaries, synthesize, coverage_sufficient, FinalizationUnavailable
finalized = False
evidence_completion = None
if eligible(model_name, status, stop_reason, accumulator.tool_calls, cancel_event.is_set(), stop_succeeded):
final_call = None
try:
native_answer = accumulator.answer
native_completed = status == "COMPLETED"
evidence = evidence_summaries(actor_root / "evidence", request["releaseId"], exclude_files=prior_evidence_files)
if native_completed and not any(d["result"].get("kind")=="quality_comparison" for d in evidence):
raise FinalizationUnavailable("本轮原生答复已完成,无期间比较需要整理")
# An administrator may switch profiles while this turn runs. Its
# one synthesis uses the original connection, just like native tools.
selected_base, selected_key, selected_name = selected_connection
if cancel_event.is_set(): raise FinalizationUnavailable("已取消")
final_call = {"tool":"model.finalize", "toolCallId":"finalize-"+request["turnId"], "label":"整理已查询证据", "status":"started", "startedAt":datetime.now(timezone.utc).isoformat()}
accumulator.tool_calls.append(final_call)
accumulator.record_event("tool_start",tool=final_call["tool"],toolCallId=final_call["toolCallId"],label=final_call["label"],status="started")
emit({"type":"progress","message":"查询证据已返回,正在进行一次有限的最终答复整理…","toolCalls":accumulator.tool_calls,"toolCount":len(accumulator.tool_calls),"timeline":accumulator.timeline})
final = synthesize(base_url=selected_base,api_key=selected_key,model=selected_name,question=request["question"],evidence=evidence,cancel_event=cancel_event)
del selected_key
if cancel_event.is_set(): raise FinalizationUnavailable("已取消")
final_call.update(status="completed",finishedAt=datetime.now(timezone.utc).isoformat(),outputSummary="已按本轮真实查询证据生成最终答复")
accumulator.record_event("tool_result",tool=final_call["tool"],toolCallId=final_call["toolCallId"],label=final_call["label"],status="completed",text=final_call["outputSummary"])
accumulator.blocks["evidence-final"] = final["answer"]
accumulator.block_order.append("evidence-final")
# Keep the final answer once. The preceding native draft is not a
# second business answer after the verified report replaces it.
accumulator.timeline = [entry for entry in accumulator.timeline
if not (entry.get("kind") == "assistant" and native_answer
and entry.get("text") == public_progress_text(native_answer))]
accumulator.record_event("assistant",text=public_progress_text(final["answer"]))
server_coverage = coverage_sufficient(evidence,request["question"])
status = "COMPLETED" if final["evidenceSufficient"] and server_coverage else "PARTIAL"
completion = {"modelDeclaredEvidenceSufficient":final.get("modelDeclaredEvidenceSufficient", final["evidenceSufficient"]),
"serverCoverageSufficient":server_coverage,"missingEvidence":final["missingEvidence"],
"completionAdjusted":final.get("completionAdjusted",False)}
_atomic_json(actor_root / "finalization.json", {"turnId":request["turnId"],"status":status,"nativeStopReason":stop_reason,"model":final["model"],"requestParameters":{"max_completion_tokens":4096,"thinking":{"type":"disabled"},"stream":False},"evidenceRefs":[d["evidenceRef"] for d in evidence],"usage":final["usage"],"finishReason":final["finishReason"],**completion})
evidence_completion = completion
stop_reason, finalized = "evidence_finalized" if status=="COMPLETED" else "evidence_incomplete", True
accumulator.finish_reasons.append("stop")
except (FinalizationUnavailable, OSError, ValueError) as error:
_atomic_json(actor_root / "finalization.json", {"turnId":request["turnId"],"status":"CANCELLED" if cancel_event.is_set() else "COMPLETED" if status=="COMPLETED" else "PARTIAL","nativeStopReason":stop_reason,"reason":str(error) if isinstance(error,FinalizationUnavailable) else "证据整理不可用"})
if final_call is not None:
final_call.update(status="failed",finishedAt=datetime.now(timezone.utc).isoformat(),outputSummary="有限收尾未完成;查询证据和过程已保留")
accumulator.record_event("tool_result",tool=final_call["tool"],toolCallId=final_call["toolCallId"],label=final_call["label"],status="failed",text=final_call["outputSummary"])
if cancel_event.is_set(): status,stop_reason="CANCELLED","用户停止"
answer = accumulator.answer
if answer:
emit({"type": "answer", "answer": answer})
if status == "COMPLETED":
report_markdown, report_html = _story_files(client, base_url, accumulator.chat_id,
accumulator.tool_calls, accumulator.tool_outputs)
else:
report_markdown, report_html = None, None
if status == "COMPLETED" and not answer:
status, stop_reason = "FAILED", "empty_answer"
if status == "COMPLETED" and answer:
_atomic_json(visible_answer_path, {"releaseId":request["releaseId"],"sessionId":request["sessionId"],
"turnId":request["turnId"],"answer":answer})
result = {
"type": "result", "status": status, "engine": "NAO", "answer": answer,
"nativeChatId": accumulator.chat_id,
"completionMethod": "bounded_evidence_finalization" if finalized else "native",
"timeline": accumulator.timeline,
"toolCalls": accumulator.tool_calls,
"stopReason": stop_reason,
"nativeStopSucceeded": stop_succeeded,
"finishReason": accumulator.finish_reasons[-1] if accumulator.finish_reasons else None,
}
if finalized:
result["evidenceCompletion"] = evidence_completion
if answer and status in ("COMPLETED", "PARTIAL") and not cancel_event.is_set():
from analysis_visual_evidence import visual_result
result.update(visual_result(actor_root / "evidence", request["releaseId"],
request["question"], request["turnId"],
exclude_files=prior_evidence_files, cancel_event=cancel_event))
if report_markdown is not None:
result["reportMarkdown"] = report_markdown
if report_html is not None:
result["reportHtml"] = report_html
emit(result)
return result
except RuntimeErrorSafe as error:
result = {"type": "result", "status": "FAILED", "engine": "NAO", "answer": "",
"toolCalls": [], "stopReason": "runtime_error", "message": str(error)}
emit(result)
return result
except (ValueError, OSError, RuntimeError):
result = {"type": "result", "status": "FAILED", "engine": "NAO", "answer": "",
"toolCalls": [], "stopReason": "runtime_error", "message": "本轮分析无法启动,请检查本机运行状态后重试。"}
emit(result)
return result
finally:
if access_path is not None:
access_path.unlink(missing_ok=True)
if actor_fd is not None:
RuntimeManager().release_actor(actor_fd)
def _read_initial_request(stdin: Any) -> dict[str, Any]:
line = stdin.readline()
if not line:
raise ValueError("缺少首行请求。")
try:
value = json.loads(line)
except json.JSONDecodeError:
raise ValueError("首行请求不是有效 JSON。") from None
if not isinstance(value, dict) or value.get("type") == "cancel":
raise ValueError("首行必须是请求对象。")
return value
def main(argv: list[str] | None = None) -> int:
os.umask(0o077)
parser = argparse.ArgumentParser(add_help=True)
parser.add_argument("--stop-all", action="store_true", help="仅停止已验证且空闲的NAO质量实例")
args = parser.parse_args(argv)
if args.stop_all:
count = RuntimeManager().stop_idle_all()
sys.stdout.write(json.dumps({"stopped": count}, separators=(",", ":")) + "\n")
return 0
try:
request = _read_initial_request(sys.stdin)
run_turn(request)
return 0
except ValueError as error:
_emit(sys.stdout, threading.Lock(), {"type": "result", "status": "FAILED", "engine": "NAO",
"answer": "", "toolCalls": [], "stopReason": "invalid_request",
"message": str(error)})
return 2
except Exception:
_emit(sys.stdout, threading.Lock(), {"type": "result", "status": "FAILED", "engine": "NAO",
"answer": "", "toolCalls": [], "stopReason": "runtime_error",
"message": "本轮分析无法启动,请检查本机运行状态后重试。"})
return 1
if __name__ == "__main__":
raise SystemExit(main())