285 lines
12 KiB
Python
285 lines
12 KiB
Python
|
# ============================================================
|
|||
|
|
# 滚动摘要 · LLM 压缩生成方（moduleId: core-summarizer, 可重生 ✅）
|
|||
|
|
# 矩阵 64 行剩余项：滚动摘要的 LLM 压缩生成方接线 + 越权上下文测试。
|
|||
|
|
# - compress_scroll_summary：固定中文提示词 -> ModelProvider.chat_text 压缩；
|
|||
|
|
# provider 不可用/超时/异常/输出为空/超预算 -> 确定性回退
|
|||
|
|
# （deterministic_compress，复用 estimate_tokens 预算口径截断）。
|
|||
|
|
# - maybe_roll_session_summary：每轮对话后的预算门控接线（fail closed）。
|
|||
|
|
# 绝不编造：提示词限定只基于给定消息；任何异常输出都回退确定性截断；
|
|||
|
|
# 压缩结果经 scroll_summary 入库，sourceMessageIds/supersedes 可溯源。
|
|||
|
|
# ============================================================
|
|||
|
|
from __future__ import annotations
|
|||
|
|
|
|||
|
|
import inspect
|
|||
|
|
import re
|
|||
|
|
from dataclasses import dataclass, field
|
|||
|
|
from datetime import datetime
|
|||
|
|
from typing import Any
|
|||
|
|
|
|||
|
|
from server.agent_core.context_policy import (
|
|||
|
|
DEFAULT_WINDOW,
|
|||
|
|
ContextBudget,
|
|||
|
|
ContextBudgetExceeded,
|
|||
|
|
assemble_context,
|
|||
|
|
estimate_tokens,
|
|||
|
|
get_policy,
|
|||
|
|
scroll_summary,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 固定中文提示词：只允许压缩给定消息，禁止编造。
|
|||
|
|
SYSTEM_PROMPT = (
|
|||
|
|
"你是一名 APS 排产助手的长会话压缩器。你的唯一任务：把给定的对话消息压缩成"
|
|||
|
|
"一段结构化中文摘要。\n"
|
|||
|
|
"铁律：\n"
|
|||
|
|
"1. 只能使用给定对话里出现过的内容——事实、数字、决策、未决事项、验收标准；\n"
|
|||
|
|
"2. 禁止编造、补全或推断给定对话中不存在的信息；\n"
|
|||
|
|
"3. 输出纯文本摘要正文，不要输出 JSON、markdown 代码块、消息编号或任何解释；\n"
|
|||
|
|
"4. 必须保留：用户的验收标准与硬性要求、已做决策及理由、未决事项、关键数字、"
|
|||
|
|
"当前进度；\n"
|
|||
|
|
"5. 长度不得超过给定预算（中文 1 字 ≈ 1 token 估算），宁可保守删减细节，不可超长。"
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
|
|||
|
|
@dataclass(frozen=True)
|
|||
|
|
class SummaryCompression:
|
|||
|
|
"""压缩结果：摘要文本 + 被压缩消息 id 列表 + 生成方标记与回退原因。"""
|
|||
|
|
|
|||
|
|
text: str
|
|||
|
|
compressed_message_ids: list[str] = field(default_factory=list)
|
|||
|
|
mode: str = "deterministic" # "llm" | "deterministic"
|
|||
|
|
reason: str | None = None # 仅回退时记录原因（审计/测试可检查）
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _normalize_message(message: Any, idx: int) -> dict[str, Any] | None:
|
|||
|
|
"""与 context_policy._normalize_message 同口径：只保留 user/agent 且有文本的消息。"""
|
|||
|
|
if not isinstance(message, dict):
|
|||
|
|
return None
|
|||
|
|
role = str(message.get("role") or "")
|
|||
|
|
if role in ("assistant", "bot", "ai"):
|
|||
|
|
role = "agent"
|
|||
|
|
if role not in ("user", "agent"):
|
|||
|
|
return None
|
|||
|
|
text = str(message.get("text") or message.get("content") or "").strip()
|
|||
|
|
if not text:
|
|||
|
|
return None
|
|||
|
|
return {
|
|||
|
|
"messageId": str(message.get("messageId") or message.get("id") or f"m{idx}"),
|
|||
|
|
"role": role,
|
|||
|
|
"text": text,
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _normalize_messages(messages: list[Any]) -> list[dict[str, Any]]:
|
|||
|
|
return [
|
|||
|
|
item for item in (
|
|||
|
|
_normalize_message(message, idx)
|
|||
|
|
for idx, message in enumerate(messages)
|
|||
|
|
)
|
|||
|
|
if item is not None
|
|||
|
|
]
|
|||
|
|
|
|||
|
|
|
|||
|
|
def build_compression_prompt(messages: list[dict[str, Any]], budget_tokens: int) -> str:
|
|||
|
|
"""固定中文提示词模板：把给定消息原样呈现给 LLM，禁止引用外部信息。"""
|
|||
|
|
lines = [
|
|||
|
|
f"预算：摘要估算 token 不得超过 {budget_tokens}。",
|
|||
|
|
"",
|
|||
|
|
"需要压缩的对话（按时间顺序）：",
|
|||
|
|
"",
|
|||
|
|
]
|
|||
|
|
for idx, message in enumerate(messages, start=1):
|
|||
|
|
who = "用户" if message.get("role") == "user" else "助手"
|
|||
|
|
lines.append(f"[消息{idx}] {who}：{message.get('text')}")
|
|||
|
|
lines.extend(["", "请只输出压缩后的摘要正文："])
|
|||
|
|
return "\n".join(lines)
|
|||
|
|
|
|||
|
|
|
|||
|
|
def deterministic_compress(messages: list[Any], budget_tokens: int) -> SummaryCompression:
|
|||
|
|
"""确定性回退：复用 estimate_tokens 预算口径，从最旧开始保留直到超预算。
|
|||
|
|
|
|||
|
|
只使用给定消息内容（绝不引入外部信息）；输出可溯源到消息 id。
|
|||
|
|
"""
|
|||
|
|
if budget_tokens <= 0:
|
|||
|
|
raise ValueError("budget_tokens must be positive")
|
|||
|
|
normalized = _normalize_messages(messages)
|
|||
|
|
if not normalized:
|
|||
|
|
raise ValueError("no compressible messages")
|
|||
|
|
message_ids = [m["messageId"] for m in normalized]
|
|||
|
|
lines = [f"{'用户' if m['role'] == 'user' else '助手'}：{m['text']}" for m in normalized]
|
|||
|
|
kept: list[str] = []
|
|||
|
|
for line in lines:
|
|||
|
|
candidate = "\n".join([*kept, line])
|
|||
|
|
if estimate_tokens(candidate) > budget_tokens:
|
|||
|
|
break
|
|||
|
|
kept.append(line)
|
|||
|
|
text = "\n".join(kept).strip()
|
|||
|
|
if not text:
|
|||
|
|
# 单条都放不下：对最新一条做预算内强制截断，保证非空回退。
|
|||
|
|
text = lines[-1]
|
|||
|
|
while estimate_tokens(text) > budget_tokens and len(text) > 8:
|
|||
|
|
text = text[:-8].rstrip()
|
|||
|
|
return SummaryCompression(
|
|||
|
|
text=text,
|
|||
|
|
compressed_message_ids=message_ids,
|
|||
|
|
mode="deterministic",
|
|||
|
|
reason="budget-truncation",
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
|
|||
|
|
_FENCE_RE = re.compile(r"```(?:json|text)?\s*([\s\S]*?)```", re.IGNORECASE)
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _strip_fence(text: str) -> str:
|
|||
|
|
"""容错：剥离 markdown 代码块围栏（部分国产端点会在文本外套 ```）。"""
|
|||
|
|
match = _FENCE_RE.search(text)
|
|||
|
|
return (match.group(1) if match else text).strip()
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _chat_accepts_timeout(provider: Any) -> bool:
|
|||
|
|
try:
|
|||
|
|
return "timeout" in inspect.signature(provider.chat_text).parameters
|
|||
|
|
except (TypeError, ValueError):
|
|||
|
|
return True # 拿不到签名时按完整签名处理
|
|||
|
|
|
|||
|
|
|
|||
|
|
async def _provider_chat(provider: Any, system: str, user: str, timeout: float) -> str | None:
|
|||
|
|
if _chat_accepts_timeout(provider):
|
|||
|
|
return await provider.chat_text(system, user, timeout=timeout)
|
|||
|
|
return await provider.chat_text(system, user)
|
|||
|
|
|
|||
|
|
|
|||
|
|
async def compress_scroll_summary(
|
|||
|
|
messages: list[dict[str, Any]], *,
|
|||
|
|
provider: Any, budget_tokens: int = 600, timeout: float = 45.0,
|
|||
|
|
) -> SummaryCompression:
|
|||
|
|
"""LLM 压缩生成方：固定中文提示词 + provider.chat_text（ModelProvider 接口）。
|
|||
|
|
|
|||
|
|
- 只允许压缩给定消息内容：提示词限定 + 输出校验 + 异常回退，绝不编造。
|
|||
|
|
- provider 不可用（未启用/返回 None）/ 超时 / 异常 / 输出为空 / 输出超预算
|
|||
|
|
-> 确定性回退 deterministic_compress（复用 estimate_tokens 预算截断）。
|
|||
|
|
Returns:
|
|||
|
|
SummaryCompression{text, compressedMessageIds, mode, reason}
|
|||
|
|
Raises:
|
|||
|
|
ValueError: 无可压缩消息或 budget_tokens <= 0（fail closed）
|
|||
|
|
"""
|
|||
|
|
normalized = _normalize_messages(messages)
|
|||
|
|
if not normalized:
|
|||
|
|
raise ValueError("compress_scroll_summary: no compressible messages")
|
|||
|
|
if budget_tokens <= 0:
|
|||
|
|
raise ValueError("compress_scroll_summary: budget_tokens must be positive")
|
|||
|
|
message_ids = [m["messageId"] for m in normalized]
|
|||
|
|
|
|||
|
|
def fallback(reason: str) -> SummaryCompression:
|
|||
|
|
compressed = deterministic_compress(normalized, budget_tokens)
|
|||
|
|
return SummaryCompression(
|
|||
|
|
text=compressed.text,
|
|||
|
|
compressed_message_ids=message_ids,
|
|||
|
|
mode="deterministic",
|
|||
|
|
reason=reason,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
raw: str | None = None
|
|||
|
|
reason: str | None = None
|
|||
|
|
try:
|
|||
|
|
raw = await _provider_chat(
|
|||
|
|
provider,
|
|||
|
|
SYSTEM_PROMPT,
|
|||
|
|
build_compression_prompt(normalized, budget_tokens),
|
|||
|
|
timeout,
|
|||
|
|
)
|
|||
|
|
except Exception as exc: # noqa: BLE001 - provider 异常一律回退，绝不阻断
|
|||
|
|
reason = f"provider-error:{type(exc).__name__}"
|
|||
|
|
if raw is None:
|
|||
|
|
return fallback(reason or "provider-unavailable")
|
|||
|
|
text = _strip_fence(str(raw))
|
|||
|
|
if not text:
|
|||
|
|
return fallback("llm-empty-output")
|
|||
|
|
if estimate_tokens(text) > budget_tokens:
|
|||
|
|
return fallback("llm-over-budget")
|
|||
|
|
return SummaryCompression(
|
|||
|
|
text=text,
|
|||
|
|
compressed_message_ids=message_ids,
|
|||
|
|
mode="llm",
|
|||
|
|
reason=None,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
|
|||
|
|
async def maybe_roll_session_summary(
|
|||
|
|
store_data: dict[str, Any], session_id: str, *,
|
|||
|
|
provider: Any | None = None,
|
|||
|
|
budget: ContextBudget | None = None,
|
|||
|
|
window: int = DEFAULT_WINDOW,
|
|||
|
|
auto_summarize: bool | None = None,
|
|||
|
|
by: str | None = None,
|
|||
|
|
now: datetime | None = None,
|
|||
|
|
) -> dict[str, Any] | None:
|
|||
|
|
"""每轮对话后的滚动摘要钩子：预算门控 + LLM 压缩（fail closed）。
|
|||
|
|
|
|||
|
|
触发条件（预算门控）：
|
|||
|
|
- 会话原始消息（未钉住）估算 token 超过会话层预算（长会话），或
|
|||
|
|
- 当前会话层组装已超预算（layer == "session"）。
|
|||
|
|
满足时，把「滑出窗口且未被既有摘要覆盖」的消息交给 compress_scroll_summary
|
|||
|
|
（LLM 压缩；异常/超时/输出不合格 -> 确定性回退），再经 scroll_summary 入库
|
|||
|
|
（sourceMessageIds + supersedes 溯源链，compressedBy 标记生成方）。
|
|||
|
|
滚动后会话层仍超预算 -> ContextBudgetExceeded（可预测拒绝，fail closed）。
|
|||
|
|
|
|||
|
|
auto_summarize: None -> 读会话策略 contextPolicies[session].autoSummarize，
|
|||
|
|
缺省 True（长会话默认开启）；False -> 永不滚动。
|
|||
|
|
Returns:
|
|||
|
|
新摘要 dict（发生滚动）或 None（未触发/被禁用）。
|
|||
|
|
Raises:
|
|||
|
|
ContextBudgetExceeded: 滚动后仍超预算（可预测拒绝）
|
|||
|
|
ValueError: 会话/项目不可访问（越权 fail closed）
|
|||
|
|
"""
|
|||
|
|
budget = budget or ContextBudget()
|
|||
|
|
session_limit = budget.layer_limit("session")
|
|||
|
|
policy = get_policy(store_data, session_id)
|
|||
|
|
enabled = policy.get("autoSummarize", True) if auto_summarize is None else auto_summarize
|
|||
|
|
if not enabled:
|
|||
|
|
return None
|
|||
|
|
|
|||
|
|
normalized = _normalize_messages((store_data.get("messages") or {}).get(session_id) or [])
|
|||
|
|
pinned_ids = {str(item.get("messageId")) for item in policy.get("pinned") or []}
|
|||
|
|
raw_history = [m for m in normalized if m["messageId"] not in pinned_ids]
|
|||
|
|
if not raw_history:
|
|||
|
|
return None
|
|||
|
|
raw_tokens = estimate_tokens("\n".join(m["text"] for m in raw_history))
|
|||
|
|
|
|||
|
|
layer_over = False
|
|||
|
|
try:
|
|||
|
|
assemble_context(store_data, session_id, None, budget, window=window)
|
|||
|
|
except ContextBudgetExceeded as exc:
|
|||
|
|
if exc.layer != "session":
|
|||
|
|
raise
|
|||
|
|
layer_over = True
|
|||
|
|
if raw_tokens <= session_limit and not layer_over:
|
|||
|
|
return None
|
|||
|
|
|
|||
|
|
covered: set[str] = set()
|
|||
|
|
for entry in [policy.get("summary"), *(policy.get("summaries") or [])]:
|
|||
|
|
if isinstance(entry, dict):
|
|||
|
|
covered.update(str(sid) for sid in (entry.get("sourceMessageIds") or []))
|
|||
|
|
recent_ids = {m["messageId"] for m in normalized[-window:]}
|
|||
|
|
eligible = [
|
|||
|
|
m for m in raw_history
|
|||
|
|
if m["messageId"] not in recent_ids and m["messageId"] not in covered
|
|||
|
|
]
|
|||
|
|
if not eligible:
|
|||
|
|
if layer_over:
|
|||
|
|
# 无可压缩消息但会话层仍超预算：可预测拒绝（fail closed）。
|
|||
|
|
assemble_context(store_data, session_id, None, budget, window=window)
|
|||
|
|
return None
|
|||
|
|
|
|||
|
|
summary_budget = max(1, session_limit // 2)
|
|||
|
|
compression = await compress_scroll_summary(
|
|||
|
|
eligible, provider=provider, budget_tokens=summary_budget,
|
|||
|
|
)
|
|||
|
|
new_summary = scroll_summary(
|
|||
|
|
store_data, session_id, compression.text,
|
|||
|
|
source_message_ids=compression.compressed_message_ids,
|
|||
|
|
by=by or "planner", now=now, summarizer=compression.mode,
|
|||
|
|
)
|
|||
|
|
# 滚动后验证：会话层仍超预算 -> ContextBudgetExceeded（可预测拒绝）。
|
|||
|
|
assemble_context(store_data, session_id, None, budget, window=window)
|
|||
|
|
return new_summary
|