253 lines
10 KiB
Python
253 lines
10 KiB
Python
"""上下文构建器：负责为每次 LLM 调用组装完整消息上下文。
|
||
|
||
本模块主要做三件事：
|
||
1. 生成 system prompt（身份、运行时信息、bootstrap 文件、记忆、技能摘要）；
|
||
2. 将历史消息与当前用户输入拼接成模型可消费的 messages；
|
||
3. 在工具调用循环中追加 assistant/tool 消息，维持对话状态连续性。
|
||
"""
|
||
|
||
import base64
|
||
import mimetypes
|
||
import platform
|
||
from pathlib import Path
|
||
from typing import Any
|
||
|
||
from nanobot.agent.agent_registry import AgentRegistry
|
||
from nanobot.agent.memory import MemoryStore
|
||
from nanobot.agent.skills import SkillsLoader
|
||
|
||
|
||
class ContextBuilder:
|
||
"""
|
||
Agent 上下文装配器。
|
||
|
||
设计目标：
|
||
- 把“静态配置”（AGENTS/USER/TOOLS 等）与“动态上下文”（时间、会话、历史）统一拼装；
|
||
- 保持 prompt 结构稳定，降低模型行为波动；
|
||
- 让工具调用前后的消息追加逻辑集中在一个位置，便于维护。
|
||
"""
|
||
|
||
# bootstrap 文件按此顺序加载并拼接，顺序会影响最终提示词语义优先级。
|
||
BOOTSTRAP_FILES = ["AGENTS.md", "SOUL.md", "USER.md", "TOOLS.md", "IDENTITY.md"]
|
||
|
||
def __init__(
|
||
self,
|
||
workspace: Path,
|
||
skills_loader: SkillsLoader | None = None,
|
||
agent_registry: AgentRegistry | None = None,
|
||
):
|
||
self.workspace = workspace
|
||
# 记忆与技能都按 workspace 维度隔离，避免跨项目污染。
|
||
self.memory = MemoryStore(workspace)
|
||
# 若上层已构造好 SkillsLoader / AgentRegistry，则复用，避免重复扫描磁盘。
|
||
self.skills = skills_loader or SkillsLoader(workspace)
|
||
# agent_registry 可选：只有支持多 agent 委派时才会把可用 agent 摘要塞进 prompt。
|
||
self.agent_registry = agent_registry
|
||
|
||
def build_system_prompt(
|
||
self,
|
||
skill_names: list[str] | None = None,
|
||
execution_context: str | None = None,
|
||
) -> str:
|
||
"""构建 system prompt（身份 + 配置 + 记忆 + 技能信息）。"""
|
||
# skill_names 目前作为接口预留，便于未来按需只加载指定技能。
|
||
parts = []
|
||
|
||
# 1) 核心身份段：包含当前时间、系统环境、工作区路径等动态信息。
|
||
parts.append(self._get_identity())
|
||
|
||
# 2) workspace 里的 bootstrap 文件（若存在）按顺序拼接。
|
||
bootstrap = self._load_bootstrap_files()
|
||
if bootstrap:
|
||
parts.append(bootstrap)
|
||
|
||
# 3) 长期记忆上下文（来自 memory/MEMORY.md 等）。
|
||
memory = self.memory.get_memory_context()
|
||
if memory:
|
||
parts.append(f"# Memory\n\n{memory}")
|
||
|
||
# 4) 技能采用“渐进加载”策略。
|
||
# 4.1 always 技能：直接把完整内容塞进当前 prompt。
|
||
always_skills = self.skills.get_always_skills()
|
||
if always_skills:
|
||
always_content = self.skills.load_skills_for_context(always_skills)
|
||
if always_content:
|
||
parts.append(f"# Active Skills\n\n{always_content}")
|
||
|
||
# 4.2 可用技能：只放摘要，具体内容让 agent 运行时按需 read_file。
|
||
# 这样可以控制 token 体积，避免把所有技能全文塞入上下文。
|
||
skills_summary = self.skills.build_skills_summary()
|
||
if skills_summary:
|
||
parts.append(f"""# Skills
|
||
|
||
The following skills extend your capabilities. To use a skill, read its SKILL.md file using the read_file tool.
|
||
Skills with available="false" need dependencies installed first - you can try installing them with apt/brew.
|
||
|
||
{skills_summary}""")
|
||
|
||
if self.agent_registry:
|
||
# 把可委派 agent 目录加入 system prompt，模型才知道 `spawn` 能调用谁。
|
||
agents_summary = self.agent_registry.build_agents_summary()
|
||
if agents_summary:
|
||
parts.append(f"""# Available Agents
|
||
|
||
The following agents can be delegated to via the `spawn` tool.
|
||
Use `target` for a single agent and `targets` for a group.
|
||
|
||
{agents_summary}""")
|
||
|
||
if execution_context:
|
||
# `execution_context` 用于 cron / system task 这类“不是普通用户消息”的额外运行说明。
|
||
parts.append(f"# Execution Context\n\n{execution_context.strip()}")
|
||
|
||
# 各块之间用分隔线拼接，提升提示词可读性与结构稳定性。
|
||
return "\n\n---\n\n".join(parts)
|
||
|
||
def _get_identity(self) -> str:
|
||
"""生成核心身份段。"""
|
||
import time as _time
|
||
from datetime import datetime
|
||
# 时间与时区在 system prompt 中显式给出，减少模型对“当前时间”的猜测。
|
||
now = datetime.now().strftime("%Y-%m-%d %H:%M (%A)")
|
||
tz = _time.strftime("%Z") or "UTC"
|
||
# 固化绝对工作区路径，帮助模型生成更准确的文件操作指令。
|
||
workspace_path = str(self.workspace.expanduser().resolve())
|
||
# 运行时信息可帮助模型在跨平台命令选择时更稳健（如 macOS/Linux 差异）。
|
||
system = platform.system()
|
||
runtime = f"{'macOS' if system == 'Darwin' else system} {platform.machine()}, Python {platform.python_version()}"
|
||
|
||
return f"""# nanobot 🐈
|
||
|
||
You are nanobot, a helpful AI assistant.
|
||
|
||
## Current Time
|
||
{now} ({tz})
|
||
|
||
## Runtime
|
||
{runtime}
|
||
|
||
## Workspace
|
||
Your workspace is at: {workspace_path}
|
||
- Long-term memory: {workspace_path}/memory/MEMORY.md
|
||
- History log: {workspace_path}/memory/HISTORY.md (grep-searchable)
|
||
- Custom skills: {workspace_path}/skills/{{skill-name}}/SKILL.md
|
||
|
||
Reply directly with text for conversations. Only use the 'message' tool to send to a specific chat channel.
|
||
|
||
## Tool Call Guidelines
|
||
- Before calling tools, you may briefly state your intent (e.g. "Let me check that"), but NEVER predict or describe the expected result before receiving it.
|
||
- Before modifying a file, read it first to confirm its current content.
|
||
- Do not assume a file or directory exists — use list_dir or read_file to verify.
|
||
- After writing or editing a file, re-read it if accuracy matters.
|
||
- If a tool call fails, analyze the error before retrying with a different approach.
|
||
- Do not write directly into `{workspace_path}/skills`; new or updated skills must go through the review flow before activation.
|
||
|
||
## Memory
|
||
- Remember important facts: write to {workspace_path}/memory/MEMORY.md
|
||
- Recall past events: grep {workspace_path}/memory/HISTORY.md"""
|
||
|
||
def _load_bootstrap_files(self) -> str:
|
||
"""从 workspace 读取 bootstrap 文件并拼接。"""
|
||
parts = []
|
||
|
||
for filename in self.BOOTSTRAP_FILES:
|
||
file_path = self.workspace / filename
|
||
if file_path.exists():
|
||
# 缺失文件时静默跳过，保持默认可用。
|
||
content = file_path.read_text(encoding="utf-8")
|
||
parts.append(f"## {filename}\n\n{content}")
|
||
|
||
return "\n\n".join(parts) if parts else ""
|
||
|
||
def build_messages(
|
||
self,
|
||
history: list[dict[str, Any]],
|
||
current_message: str,
|
||
skill_names: list[str] | None = None,
|
||
execution_context: str | None = None,
|
||
media: list[str] | None = None,
|
||
channel: str | None = None,
|
||
chat_id: str | None = None,
|
||
) -> list[dict[str, Any]]:
|
||
"""构建一次 LLM 调用的完整 messages 数组。"""
|
||
messages = []
|
||
|
||
# 第 1 条固定是 system prompt。
|
||
system_prompt = self.build_system_prompt(skill_names, execution_context=execution_context)
|
||
if channel and chat_id:
|
||
# 把当前会话路由信息也写入系统提示，便于模型做跨渠道决策。
|
||
system_prompt += f"\n\n## Current Session\nChannel: {channel}\nChat ID: {chat_id}"
|
||
messages.append({"role": "system", "content": system_prompt})
|
||
|
||
# 追加历史消息（通常已由 SessionManager 做窗口与清洗）。
|
||
messages.extend(history)
|
||
|
||
# 追加当前用户输入；若带图片则转换为多模态 content 结构。
|
||
user_content = self._build_user_content(current_message, media)
|
||
messages.append({"role": "user", "content": user_content})
|
||
|
||
return messages
|
||
|
||
def _build_user_content(self, text: str, media: list[str] | None) -> str | list[dict[str, Any]]:
|
||
"""构建 user content，支持文本或“文本+图片”多模态格式。"""
|
||
# 无媒体时直接走纯文本，保持最简单路径。
|
||
if not media:
|
||
return text
|
||
|
||
images = []
|
||
for path in media:
|
||
p = Path(path)
|
||
mime, _ = mimetypes.guess_type(path)
|
||
# 仅接收本地图片文件，其他媒体类型暂不注入到模型内容。
|
||
if not p.is_file() or not mime or not mime.startswith("image/"):
|
||
continue
|
||
# 按 data URL 形式内联图片，兼容支持 image_url 的 provider 接口。
|
||
b64 = base64.b64encode(p.read_bytes()).decode()
|
||
images.append({"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}})
|
||
|
||
# 没有合法图片时回退纯文本，避免传空数组导致模型侧解析异常。
|
||
if not images:
|
||
return text
|
||
# 多模态结构中把图片放前、文本放后，便于模型先“看图”再读文字指令。
|
||
return images + [{"type": "text", "text": text}]
|
||
|
||
def add_tool_result(
|
||
self,
|
||
messages: list[dict[str, Any]],
|
||
tool_call_id: str,
|
||
tool_name: str,
|
||
result: str
|
||
) -> list[dict[str, Any]]:
|
||
"""把工具执行结果追加到 messages。"""
|
||
messages.append({
|
||
"role": "tool",
|
||
"tool_call_id": tool_call_id,
|
||
"name": tool_name,
|
||
"content": result
|
||
})
|
||
return messages
|
||
|
||
def add_assistant_message(
|
||
self,
|
||
messages: list[dict[str, Any]],
|
||
content: str | None,
|
||
tool_calls: list[dict[str, Any]] | None = None,
|
||
reasoning_content: str | None = None,
|
||
) -> list[dict[str, Any]]:
|
||
"""把 assistant 消息追加到 messages（可携带 tool_calls/reasoning）。"""
|
||
msg: dict[str, Any] = {"role": "assistant"}
|
||
|
||
# 始终写入 content 键：
|
||
# 部分 provider 在 key 缺失时会拒绝请求（即使值是 None 也要有该键）。
|
||
msg["content"] = content
|
||
|
||
if tool_calls:
|
||
msg["tool_calls"] = tool_calls
|
||
|
||
# reasoning_content 是“思考模型”专用字段，仅在有值时附加。
|
||
if reasoning_content is not None:
|
||
msg["reasoning_content"] = reasoning_content
|
||
|
||
messages.append(msg)
|
||
return messages
|