feat(v2): T4 Workspace 交流文本协议（schema/校验/渲染/rollup）
This commit is contained in:
tzt
+67
-3
@@ -9,10 +9,18 @@ system:
|
||||
version: 0.1.0
|
||||
|
||||
router:
|
||||
low_confidence_threshold: 0.60 # 分类置信度低于此值 -> 直接走大模型
|
||||
judge_fallback_threshold: 0.70 # Judge 质量分低于此值 -> 升级大模型
|
||||
low_confidence_threshold: 0.60 # 分类置信度低于此值 -> 直接走最后处理者
|
||||
judge_fallback_threshold: 0.70 # Judge 质量分低于此值 -> 升级最后处理者
|
||||
default_temperature: 0.2
|
||||
|
||||
# 专家系统内核执行配置（L0 默认零参数）
|
||||
execution:
|
||||
mode: rule # rule（默认，L0 零参数）| hybrid
|
||||
planner: rule # rule（规则拆解）| hf（可选小模型拆解）
|
||||
expert_backend: rule # rule（规则执行器）| hf | api（本地小模型按需加载）
|
||||
model_level: L0 # L0 纯规则 | L1 分类/Planner增强 | L2 领域生成
|
||||
max_plan_depth: 3 # 任务拆解深度上限
|
||||
|
||||
classifier:
|
||||
type: rule # rule（零依赖）| hf（transformers）
|
||||
model: Qwen/Qwen3-0.6B
|
||||
@@ -23,17 +31,31 @@ domains:
|
||||
- math
|
||||
- legal
|
||||
- medical
|
||||
- finance
|
||||
- life
|
||||
- education
|
||||
- general
|
||||
|
||||
# 两级路由：大领域分组（用户接口指定 group → 组内路由模型 → 组内专业小模型）
|
||||
# 组内路由模型只识别本组领域，体积约为统一路由模型的 1/4
|
||||
domain_groups:
|
||||
tech: [code, math]
|
||||
professional: [legal, medical, finance]
|
||||
lifestyle: [life, education]
|
||||
general: [general]
|
||||
|
||||
experts:
|
||||
code: { type: mock, model: Qwen/Qwen2.5-Coder-7B-Instruct }
|
||||
math: { type: mock, model: Qwen/Qwen3-4B-Instruct }
|
||||
legal: { type: mock, model: Qwen/Qwen3-4B-Instruct }
|
||||
medical: { type: mock, model: Qwen/Qwen3-4B-Instruct }
|
||||
finance: { type: mock, model: Qwen/Qwen3-4B-Instruct }
|
||||
life: { type: mock, model: Qwen/Qwen3-1.7B-Instruct }
|
||||
education: { type: mock, model: Qwen/Qwen3-1.7B-Instruct }
|
||||
general: { type: mock, model: Qwen/Qwen3-1.7B-Instruct }
|
||||
|
||||
fallback:
|
||||
type: mock # mock | api（OpenAI 兼容，如 DeepSeek）
|
||||
type: mock # none（降级模板）| mock | local（本地≤8B 按需加载）| api
|
||||
model: deepseek-chat
|
||||
base_url: https://api.deepseek.com/v1
|
||||
api_key_env: DEEPSEEK_API_KEY
|
||||
@@ -47,3 +69,45 @@ cache:
|
||||
semantic_enabled: true # 语义缓存（字符 n-gram 相似度，零依赖）
|
||||
similarity_threshold: 0.88
|
||||
promote_frequency: 5 # 命中 N 次后提升为精确缓存
|
||||
|
||||
# ============================================================
|
||||
# v2：端云协同 LLM 协作系统（《实现方案_v2》）配置段
|
||||
# v1 段（上方）保留，供 legacy 路由（POST /chat/legacy）使用。
|
||||
# ============================================================
|
||||
|
||||
runtime:
|
||||
llama_server:
|
||||
binary: bin/llama-server.exe # 捆绑上游 release，不改源码（D1）
|
||||
model: models/qwen3.5-4b-q4_k_m.gguf
|
||||
port: 8901
|
||||
hw_profile: auto # auto | gpu12 | gpu8 | cpu
|
||||
extra_args: ["-fa", "-ctk", "q8_0", "-ctv", "q8_0", "--cache-reuse", "256"]
|
||||
tiers: # 三档硬件模板（保守默认，可手动覆盖）
|
||||
gpu12: {ngl: 99, ctx: 32768}
|
||||
gpu8: {ngl: 14, ctx: 16384}
|
||||
cpu: {ngl: 0, ctx: 8192}
|
||||
|
||||
architect: # 大模型（API）
|
||||
model: deepseek-chat
|
||||
base_url: https://api.deepseek.com/v1
|
||||
api_key_env: DEEPSEEK_API_KEY
|
||||
temperature: 0.2
|
||||
timeout_s: 60
|
||||
|
||||
worker: # 小模型（本地）
|
||||
backend: llama_server
|
||||
temperature: 0.3
|
||||
max_fix_attempts: 2
|
||||
per_step_timeout_s: 300
|
||||
|
||||
pipeline:
|
||||
fast_path: true
|
||||
rounds_cap: 6
|
||||
api_token_cap: 8000
|
||||
breach_policy: architect_do # architect_do | local_only
|
||||
|
||||
review:
|
||||
queue_db: data/review.sqlite3
|
||||
sample_rate: 0.10 # 随机抽样送审
|
||||
force_tags: [safety] # brief.tags 命中即强制送审
|
||||
|
||||
|
||||
@@ -0,0 +1,516 @@
|
||||
"""交流文本（Workspace）—— 端云协同 LLM 协作系统的核心协议（零依赖）。
|
||||
|
||||
大模型（Architect）与小模型（Worker）互不共享内部状态，只通过这份
|
||||
schema 约束的结构化 JSON 共享工作区交接（类比前后端通过 API 契约协作）。
|
||||
|
||||
本模块实现（对齐《实现方案_v2》第 4 节）：
|
||||
- WORKSPACE_SCHEMA：draft-07 风格 schema 常量（文档/校验依据）
|
||||
- validate()：结构 + 字段长度校验（写入前必过，D2/D9）
|
||||
- 锚点寻址：a://<file>#L<start>-<end>（引用工件片段，替代全文复制）
|
||||
- 双渲染函数：render_for_architect（≤1200 token）、render_for_worker（≤8K token）
|
||||
- rollup()：已完成步骤折叠为 archive 摘要行；超限压缩（只减不删，4.5）
|
||||
- 状态机：draft -> in_progress -> reviewing -> done / escalated / failed
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import copy
|
||||
import json
|
||||
import re
|
||||
from pathlib import Path
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
VERSION = "1.0"
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 字段长度上限（同时是 rollup 依据，4.2）
|
||||
# ---------------------------------------------------------------------------
|
||||
LIMITS = {
|
||||
"goal": 500, # brief.goal 字数
|
||||
"task": 300, # brief.plan[].task 字数
|
||||
"summary": 200, # progress[].summary 字数
|
||||
"issue_text": 300, # issues[].observed/expected/tried/ask 字数
|
||||
"reply": 600, # decisions[].reply 字数
|
||||
"archive": 160, # archive[] 每条字数
|
||||
"constraints": 8, # brief.constraints 上限条数
|
||||
"plan_steps": 5, # brief.plan 上限步数
|
||||
"acceptance": 20, # brief.acceptance 上限条数
|
||||
"query_truncate": 200, # render_for_architect 中 query 截断
|
||||
}
|
||||
|
||||
# 允许的领域标签（4.2 brief.tags；仅用于安全标记与验证接地，不做路由 D3）
|
||||
ALLOWED_TAGS = {"code", "math", "legal", "medical", "finance",
|
||||
"life", "education", "general", "safety"}
|
||||
|
||||
STATUS_FLOW = {
|
||||
"draft": {"in_progress"},
|
||||
"in_progress": {"reviewing", "escalated", "failed", "in_progress"},
|
||||
"reviewing": {"done", "in_progress", "failed"},
|
||||
"escalated": {"reviewing", "done", "failed"},
|
||||
"done": set(),
|
||||
"failed": set(),
|
||||
}
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# JSON Schema（draft-07 风格，draft-07 依赖内嵌；供校验与文档参考）
|
||||
# ---------------------------------------------------------------------------
|
||||
WORKSPACE_SCHEMA: Dict[str, Any] = {
|
||||
"$schema": "http://json-schema.org/draft-07/schema#",
|
||||
"title": "Communication Workspace",
|
||||
"type": "object",
|
||||
"required": ["version", "request_id", "query", "meta"],
|
||||
"additionalProperties": False,
|
||||
"properties": {
|
||||
"version": {"const": VERSION},
|
||||
"request_id": {"type": "string", "minLength": 1},
|
||||
"query": {"type": "string", "minLength": 1},
|
||||
"meta": {
|
||||
"type": "object",
|
||||
"required": ["status", "round", "budget"],
|
||||
"properties": {
|
||||
"status": {"enum": ["draft", "in_progress", "reviewing", "escalated", "done", "failed"]},
|
||||
"round": {"type": "integer", "minimum": 0},
|
||||
"budget": {
|
||||
"type": "object",
|
||||
"required": ["api_input_tokens", "api_output_tokens", "api_token_cap", "rounds_cap"],
|
||||
"properties": {
|
||||
"api_input_tokens": {"type": "integer", "minimum": 0},
|
||||
"api_output_tokens": {"type": "integer", "minimum": 0},
|
||||
"api_token_cap": {"type": "integer", "minimum": 1},
|
||||
"rounds_cap": {"type": "integer", "minimum": 1},
|
||||
},
|
||||
},
|
||||
},
|
||||
},
|
||||
"brief": {
|
||||
"type": "object",
|
||||
"required": ["goal", "constraints", "tags", "acceptance", "plan"],
|
||||
"properties": {
|
||||
"goal": {"type": "string"},
|
||||
"constraints": {"type": "array", "items": {"type": "string"}},
|
||||
"tags": {"type": "array", "items": {"type": "string"}},
|
||||
"acceptance": {"type": "array", "items": {"type": "object"}},
|
||||
"plan": {"type": "array", "items": {"type": "object"}},
|
||||
},
|
||||
},
|
||||
"progress": {"type": "array", "items": {"type": "object"}},
|
||||
"issues": {"type": "array", "items": {"type": "object"}},
|
||||
"decisions": {"type": "array", "items": {"type": "object"}},
|
||||
"archive": {"type": "array", "items": {"type": "string"}},
|
||||
},
|
||||
}
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 工具
|
||||
# ---------------------------------------------------------------------------
|
||||
_TOKEN_PER_CHAR_ZH = 1 / 1.6 # 中文约 1.6 字/token
|
||||
_TOKEN_PER_CHAR_EN = 1 / 4.0 # 英文约 4 字/token
|
||||
|
||||
|
||||
def estimate_tokens(text: str) -> int:
|
||||
"""粗略 token 估算（中英混合，用于渲染预算校验）。"""
|
||||
if not text:
|
||||
return 0
|
||||
zh = sum(1 for ch in text if "\u4e00" <= ch <= "\u9fff")
|
||||
en = len(text) - zh
|
||||
return max(1, int(zh * _TOKEN_PER_CHAR_ZH + en * _TOKEN_PER_CHAR_EN))
|
||||
|
||||
|
||||
def build_anchor(filename: str, start: int = 1, end: Optional[int] = None) -> str:
|
||||
"""构造锚点：a://<file>#L<start>-<end>；end 缺省仅 L<start>。"""
|
||||
if end is None:
|
||||
return f"a://{filename}#L{start}"
|
||||
return f"a://{filename}#L{start}-{end}"
|
||||
|
||||
|
||||
_ANCHOR_RE = re.compile(r"^a://(?P<file>[^#]+?)(?:#L(?P<start>\d+)(?:-(?P<end>\d+))?)?$")
|
||||
|
||||
|
||||
def parse_anchor(anchor: str) -> Optional[Dict[str, Any]]:
|
||||
"""解析锚点为 {file, start, end}；非法返回 None。"""
|
||||
m = _ANCHOR_RE.match(anchor)
|
||||
if not m:
|
||||
return None
|
||||
start = int(m.group("start")) if m.group("start") else 1
|
||||
end = int(m.group("end")) if m.group("end") else start
|
||||
return {"file": m.group("file"), "start": start, "end": end}
|
||||
|
||||
|
||||
def _clip(text: str, limit: int) -> str:
|
||||
"""按字数截断（中文按字符）。"""
|
||||
if len(text) <= limit:
|
||||
return text
|
||||
return text[:limit] + "…"
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 校验
|
||||
# ---------------------------------------------------------------------------
|
||||
def validate(ws: Dict[str, Any]) -> List[str]:
|
||||
"""校验 workspace 结构 + 字段长度。返回错误列表（空 = 合法）。"""
|
||||
errors: List[str] = []
|
||||
if not isinstance(ws, dict):
|
||||
return ["workspace 必须是 object"]
|
||||
if ws.get("version") != VERSION:
|
||||
errors.append(f"version 必须是 {VERSION}")
|
||||
if not isinstance(ws.get("request_id"), str) or not ws["request_id"]:
|
||||
errors.append("request_id 必须是非空字符串")
|
||||
if not isinstance(ws.get("query"), str) or not ws["query"]:
|
||||
errors.append("query 必须是非空字符串")
|
||||
|
||||
meta = ws.get("meta")
|
||||
if not isinstance(meta, dict):
|
||||
errors.append("meta 必须是 object")
|
||||
else:
|
||||
if meta.get("status") not in STATUS_FLOW:
|
||||
errors.append(f"meta.status 非法: {meta.get('status')}")
|
||||
budget = meta.get("budget")
|
||||
if not isinstance(budget, dict):
|
||||
errors.append("meta.budget 必须是 object")
|
||||
else:
|
||||
for k in ("api_input_tokens", "api_output_tokens", "api_token_cap", "rounds_cap"):
|
||||
if not isinstance(budget.get(k), int) or budget.get(k) < 0:
|
||||
errors.append(f"meta.budget.{k} 必须是非负整数")
|
||||
|
||||
brief = ws.get("brief")
|
||||
if brief is not None:
|
||||
if not isinstance(brief, dict):
|
||||
errors.append("brief 必须是 object")
|
||||
else:
|
||||
if not isinstance(brief.get("goal"), str):
|
||||
errors.append("brief.goal 必须是字符串")
|
||||
elif len(brief["goal"]) > LIMITS["goal"]:
|
||||
errors.append(f"brief.goal 超长（>{LIMITS['goal']}字）")
|
||||
if not isinstance(brief.get("constraints"), list):
|
||||
errors.append("brief.constraints 必须是数组")
|
||||
elif len(brief["constraints"]) > LIMITS["constraints"]:
|
||||
errors.append(f"brief.constraints 超过 {LIMITS['constraints']} 条")
|
||||
if not isinstance(brief.get("tags"), list):
|
||||
errors.append("brief.tags 必须是数组")
|
||||
for t in brief.get("tags", []) or []:
|
||||
if t not in ALLOWED_TAGS:
|
||||
errors.append(f"brief.tags 含非法标签: {t}")
|
||||
acc = brief.get("acceptance")
|
||||
if not isinstance(acc, list) or len(acc) > LIMITS["acceptance"]:
|
||||
errors.append(f"brief.acceptance 需为 ≤{LIMITS['acceptance']} 的数组")
|
||||
plan = brief.get("plan")
|
||||
if not isinstance(plan, list) or len(plan) > LIMITS["plan_steps"]:
|
||||
errors.append(f"brief.plan 需为 ≤{LIMITS['plan_steps']} 步的数组")
|
||||
else:
|
||||
ids = [p.get("id") for p in plan if isinstance(p, dict)]
|
||||
if len(set(ids)) != len(ids):
|
||||
errors.append("brief.plan 存在重复 step id")
|
||||
for p in plan:
|
||||
if not isinstance(p, dict):
|
||||
errors.append("brief.plan 元素必须是 object")
|
||||
continue
|
||||
if not isinstance(p.get("task"), str):
|
||||
errors.append(f"brief.plan[{p.get('id')}].task 必须是字符串")
|
||||
elif len(p["task"]) > LIMITS["task"]:
|
||||
errors.append(f"brief.plan[{p.get('id')}].task 超长（>{LIMITS['task']}字）")
|
||||
|
||||
for i, entry in enumerate(ws.get("progress", []) or []):
|
||||
if not isinstance(entry, dict):
|
||||
errors.append(f"progress[{i}] 必须是 object"); continue
|
||||
if entry.get("status") not in ("done", "failed", "blocked"):
|
||||
errors.append(f"progress[{i}].status 非法")
|
||||
if not isinstance(entry.get("summary"), str) or len(entry["summary"]) > LIMITS["summary"]:
|
||||
errors.append(f"progress[{i}].summary 非法或超长")
|
||||
|
||||
for i, entry in enumerate(ws.get("issues", []) or []):
|
||||
if not isinstance(entry, dict):
|
||||
errors.append(f"issues[{i}] 必须是 object"); continue
|
||||
for k in ("observed", "expected", "tried", "ask"):
|
||||
if isinstance(entry.get(k), str) and len(entry[k]) > LIMITS["issue_text"]:
|
||||
errors.append(f"issues[{i}].{k} 超长（>{LIMITS['issue_text']}字）")
|
||||
|
||||
for i, entry in enumerate(ws.get("decisions", []) or []):
|
||||
if not isinstance(entry, dict):
|
||||
errors.append(f"decisions[{i}] 必须是 object"); continue
|
||||
if isinstance(entry.get("reply"), str) and len(entry["reply"]) > LIMITS["reply"]:
|
||||
errors.append(f"decisions[{i}].reply 超长（>{LIMITS['reply']}字）")
|
||||
|
||||
for i, line in enumerate(ws.get("archive", []) or []):
|
||||
if not isinstance(line, str) or len(line) > LIMITS["archive"]:
|
||||
errors.append(f"archive[{i}] 非法或超长")
|
||||
return errors
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Workspace
|
||||
# ---------------------------------------------------------------------------
|
||||
class Workspace:
|
||||
"""交流文本对象：持有状态、执行写入前校验、渲染、rollup、持久化。"""
|
||||
|
||||
def __init__(self, data: Dict[str, Any]):
|
||||
errors = validate(data)
|
||||
if errors:
|
||||
raise ValueError("workspace 校验失败: " + "; ".join(errors[:5]))
|
||||
self._data = data
|
||||
self._brief_locked = False
|
||||
|
||||
# ---------- 构造 ----------
|
||||
@classmethod
|
||||
def new(cls, request_id: str, query: str,
|
||||
api_token_cap: int = 8000, rounds_cap: int = 6) -> "Workspace":
|
||||
data = {
|
||||
"version": VERSION,
|
||||
"request_id": request_id,
|
||||
"query": query,
|
||||
"meta": {
|
||||
"status": "draft",
|
||||
"round": 0,
|
||||
"budget": {
|
||||
"api_input_tokens": 0,
|
||||
"api_output_tokens": 0,
|
||||
"api_token_cap": api_token_cap,
|
||||
"rounds_cap": rounds_cap,
|
||||
},
|
||||
},
|
||||
"brief": None,
|
||||
"progress": [],
|
||||
"issues": [],
|
||||
"decisions": [],
|
||||
"archive": [],
|
||||
}
|
||||
return cls(data)
|
||||
|
||||
# ---------- 访问 ----------
|
||||
@property
|
||||
def request_id(self) -> str:
|
||||
return self._data["request_id"]
|
||||
|
||||
@property
|
||||
def status(self) -> str:
|
||||
return self._data["meta"]["status"]
|
||||
|
||||
@property
|
||||
def data(self) -> Dict[str, Any]:
|
||||
return copy.deepcopy(self._data)
|
||||
|
||||
def get(self, key: str, default: Any = None) -> Any:
|
||||
return self._data.get(key, default)
|
||||
|
||||
def __getitem__(self, key: str) -> Any:
|
||||
return self._data[key]
|
||||
|
||||
def meta(self) -> Dict[str, Any]:
|
||||
return self._data["meta"]
|
||||
|
||||
def budget(self) -> Dict[str, int]:
|
||||
return self._data["meta"]["budget"]
|
||||
|
||||
# ---------- 写入（均先校验） ----------
|
||||
def _commit(self, data: Dict[str, Any]) -> None:
|
||||
errors = validate(data)
|
||||
if errors:
|
||||
raise ValueError("写入校验失败: " + "; ".join(errors[:5]))
|
||||
self._data = data
|
||||
|
||||
def transition(self, new_status: str) -> None:
|
||||
cur = self.status
|
||||
if new_status == cur:
|
||||
return
|
||||
if new_status not in STATUS_FLOW.get(cur, set()):
|
||||
raise ValueError(f"非法状态迁移: {cur} -> {new_status}")
|
||||
self._data["meta"]["status"] = new_status
|
||||
|
||||
def apply_brief(self, brief: Dict[str, Any]) -> None:
|
||||
"""写入 brief（写一次后锁定，D2：brief 恒定位于文档前部，prefix cache 友好）。"""
|
||||
if self._data.get("brief") is not None or self._brief_locked:
|
||||
raise ValueError("brief 已写入，不可重复")
|
||||
new = copy.deepcopy(self._data)
|
||||
new["brief"] = brief
|
||||
new["meta"]["status"] = "in_progress"
|
||||
self._commit(new)
|
||||
self._brief_locked = True
|
||||
|
||||
def add_progress(self, step: str, status: str, summary: str,
|
||||
artifact: Optional[str] = None) -> None:
|
||||
new = copy.deepcopy(self._data)
|
||||
entry: Dict[str, Any] = {"step": step, "status": status, "summary": summary}
|
||||
if artifact:
|
||||
entry["artifact"] = artifact
|
||||
new["progress"].append(entry)
|
||||
self._commit(new)
|
||||
|
||||
def add_issue(self, step: str, anchor: str, observed: str, expected: str,
|
||||
tried: str, ask: str) -> str:
|
||||
new = copy.deepcopy(self._data)
|
||||
iid = f"i{len(new['issues']) + 1}"
|
||||
entry = {
|
||||
"id": iid, "step": step, "anchor": anchor,
|
||||
"observed": observed, "expected": expected,
|
||||
"tried": tried, "ask": ask,
|
||||
}
|
||||
new["issues"].append(entry)
|
||||
self._commit(new)
|
||||
return iid
|
||||
|
||||
def add_decision(self, ref: str, reply: str,
|
||||
patch_plan: Optional[List[Dict[str, str]]] = None) -> None:
|
||||
new = copy.deepcopy(self._data)
|
||||
new["decisions"].append({
|
||||
"ref": ref, "reply": reply,
|
||||
"patch_plan": patch_plan or [],
|
||||
})
|
||||
self._commit(new)
|
||||
|
||||
def mark_round(self) -> None:
|
||||
self._data["meta"]["round"] += 1
|
||||
|
||||
def add_budget(self, input_tokens: int = 0, output_tokens: int = 0) -> None:
|
||||
b = self._data["meta"]["budget"]
|
||||
b["api_input_tokens"] += int(input_tokens)
|
||||
b["api_output_tokens"] += int(output_tokens)
|
||||
self._commit(self._data)
|
||||
|
||||
def exhausted(self) -> bool:
|
||||
"""预算熔断判定：API token 或回合任一触顶（D6）。"""
|
||||
b = self._data["meta"]["budget"]
|
||||
used = b["api_input_tokens"] + b["api_output_tokens"]
|
||||
if b["api_token_cap"] and used >= b["api_token_cap"]:
|
||||
return True
|
||||
if b["rounds_cap"] and self._data["meta"]["round"] >= b["rounds_cap"]:
|
||||
return True
|
||||
return False
|
||||
|
||||
# ---------- rollup（4.5） ----------
|
||||
def rollup(self) -> int:
|
||||
"""把 done 的 progress 折叠为 archive 摘要行，并清理解析完成的问题。
|
||||
|
||||
只减不删 archive 历史；progress 中 done 的条目折叠后移除（保留 failed/blocked）。
|
||||
返回本次折叠的条目数。
|
||||
"""
|
||||
folded = 0
|
||||
new_progress: List[Dict[str, Any]] = []
|
||||
for entry in self._data.get("progress", []):
|
||||
if entry.get("status") == "done" and entry.get("step"):
|
||||
line = _clip(f"{entry['step']}: {entry.get('summary', '')}", LIMITS["archive"])
|
||||
if line not in self._data["archive"]:
|
||||
self._data["archive"].append(line)
|
||||
folded += 1
|
||||
else:
|
||||
new_progress.append(entry)
|
||||
resolved = {d.get("ref") for d in self._data.get("decisions", [])}
|
||||
kept_issues: List[Dict[str, Any]] = []
|
||||
for iss in self._data.get("issues", []):
|
||||
if iss.get("id") in resolved:
|
||||
line = _clip(f"{iss['id']}: {iss.get('expected', '')[:60]}", LIMITS["archive"])
|
||||
if line not in self._data["archive"]:
|
||||
self._data["archive"].append(line)
|
||||
else:
|
||||
kept_issues.append(iss)
|
||||
self._data["issues"] = kept_issues
|
||||
self._data["progress"] = new_progress
|
||||
return folded
|
||||
|
||||
# ---------- 渲染 ----------
|
||||
def render_for_architect(self) -> str:
|
||||
"""渲染 Architect 输入（D7）：meta+query(截断)+全部 issues+最近3条 decisions
|
||||
+最近回合 progress 摘要。目标 ≤1200 token。"""
|
||||
d = self._data
|
||||
parts: List[str] = []
|
||||
m = d["meta"]
|
||||
parts.append("== meta ==")
|
||||
parts.append(f"status={m['status']} round={m['round']} "
|
||||
f"budget={json.dumps(m['budget'], ensure_ascii=False)}")
|
||||
parts.append("== query ==")
|
||||
parts.append(_clip(d["query"], LIMITS["query_truncate"]))
|
||||
if d.get("brief"):
|
||||
b = d["brief"]
|
||||
parts.append("== brief(锁定) ==")
|
||||
parts.append(f"goal: {_clip(b['goal'], 120)}")
|
||||
parts.append(f"plan: {[p['id'] for p in b.get('plan', [])]}")
|
||||
parts.append(f"acceptance: {[a.get('id') for a in b.get('acceptance', [])]}")
|
||||
parts.append("== issues ==")
|
||||
for iss in d.get("issues", []):
|
||||
parts.append(f"{iss['id']} step={iss.get('step')} anchor={iss.get('anchor')} "
|
||||
f"ask={_clip(iss.get('ask', ''), 80)}")
|
||||
parts.append("== 最近 3 条 decisions ==")
|
||||
for dec in d.get("decisions", [])[-3:]:
|
||||
parts.append(f"ref={dec.get('ref')} reply={_clip(dec.get('reply', ''), 80)}")
|
||||
parts.append("== progress 摘要 ==")
|
||||
for p in d.get("progress", [])[-5:]:
|
||||
parts.append(f"{p.get('step')} [{p.get('status')}] {_clip(p.get('summary', ''), 40)}")
|
||||
parts.append("== archive ==")
|
||||
for line in d.get("archive", [])[-8:]:
|
||||
parts.append(line)
|
||||
# token 预算：超限先截断最旧 archive（已只保留 3 条 decisions）
|
||||
out = "\n".join(parts)
|
||||
while estimate_tokens(out) > 1200 and len(d.get("archive", [])) > 4:
|
||||
d = copy.deepcopy(d)
|
||||
d["archive"] = d["archive"][4:]
|
||||
out = "\n".join(_rerender(self, d))
|
||||
return out
|
||||
|
||||
def render_for_worker(self, step_id: str,
|
||||
artifact_text: Optional[str] = None) -> str:
|
||||
"""渲染 Worker 输入（4.4）：brief 全文+该 step 定义+依赖 step 的 archive 摘要行
|
||||
+该 step 现有工件全文+验收标准。目标 ≤8K token。"""
|
||||
d = self._data
|
||||
b = d.get("brief")
|
||||
parts: List[str] = []
|
||||
if b:
|
||||
parts.append("== 任务目标 (goal) ==")
|
||||
parts.append(b["goal"])
|
||||
parts.append("== 约束 (constraints) ==")
|
||||
parts.extend(f"- {c}" for c in b.get("constraints", []))
|
||||
parts.append("== 全部步骤 (plan) ==")
|
||||
for p in b.get("plan", []):
|
||||
mark = " <-- 当前步" if p.get("id") == step_id else ""
|
||||
parts.append(f"{p['id']}: {p.get('task', '')}{mark}")
|
||||
parts.append(f" done_criteria: {p.get('done_criteria', '')}")
|
||||
parts.append("== 依赖步摘要 (archive) ==")
|
||||
for line in d.get("archive", [])[-6:]:
|
||||
parts.append(line)
|
||||
if artifact_text:
|
||||
parts.append(f"== 当前步已有工件（{step_id}） ==")
|
||||
parts.append(artifact_text)
|
||||
parts.append("== 验收标准 ==")
|
||||
if b:
|
||||
for a in b.get("acceptance", []):
|
||||
parts.append(f"- {a.get('id')}: {a.get('check', '')} "
|
||||
f"(machine_checkable={a.get('machine_checkable', False)})")
|
||||
parts.append("== 要求 ==")
|
||||
parts.append("请实现当前步，并用可执行验证/事实对照/结构检查自验证；"
|
||||
"通过则写 progress(done)，失败自修 ≤2 次，仍失败则写 issue。")
|
||||
return "\n".join(parts)
|
||||
|
||||
# ---------- 持久化 ----------
|
||||
def save(self, path: Path) -> None:
|
||||
path = Path(path)
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
with open(path, "w", encoding="utf-8") as f:
|
||||
json.dump(self._data, f, ensure_ascii=False, indent=2)
|
||||
|
||||
@classmethod
|
||||
def load(cls, path: Path) -> "Workspace":
|
||||
with open(path, "r", encoding="utf-8") as f:
|
||||
data = json.load(f)
|
||||
return cls(data)
|
||||
|
||||
def to_dict(self) -> Dict[str, Any]:
|
||||
return self.data
|
||||
|
||||
|
||||
def _rerender(ws: "Workspace", d: Dict[str, Any]) -> List[str]:
|
||||
"""用裁剪后的数据重建 Architect 渲染（供超限压缩内部用）。"""
|
||||
parts: List[str] = []
|
||||
m = d["meta"]
|
||||
parts.append("== meta ==")
|
||||
parts.append(f"status={m['status']} round={m['round']}")
|
||||
parts.append("== query ==")
|
||||
parts.append(_clip(d["query"], LIMITS["query_truncate"]))
|
||||
parts.append("== issues ==")
|
||||
for iss in d.get("issues", []):
|
||||
parts.append(f"{iss['id']} step={iss.get('step')} ask={_clip(iss.get('ask', ''), 80)}")
|
||||
parts.append("== decisions(最近3) ==")
|
||||
for dec in d.get("decisions", [])[-3:]:
|
||||
parts.append(f"ref={dec.get('ref')} reply={_clip(dec.get('reply', ''), 80)}")
|
||||
parts.append("== progress ==")
|
||||
for p in d.get("progress", [])[-5:]:
|
||||
parts.append(f"{p.get('step')} [{p.get('status')}] {_clip(p.get('summary', ''), 40)}")
|
||||
parts.append("== archive ==")
|
||||
for line in d.get("archive", [])[-6:]:
|
||||
parts.append(line)
|
||||
return parts
|
||||
@@ -0,0 +1,231 @@
|
||||
"""T4 交流文本 Workspace 单测（封闭，纯逻辑）。"""
|
||||
import json
|
||||
|
||||
import pytest
|
||||
|
||||
from router_system.workspace import (
|
||||
LIMITS,
|
||||
STATUS_FLOW,
|
||||
Workspace,
|
||||
build_anchor,
|
||||
estimate_tokens,
|
||||
parse_anchor,
|
||||
validate,
|
||||
)
|
||||
|
||||
ALLOWED_BRIEF = {
|
||||
"goal": "用 Python 实现快速排序并解释复杂度",
|
||||
"constraints": ["必须用标准库", "时间复杂度 O(n log n)"],
|
||||
"tags": ["code"],
|
||||
"acceptance": [{"id": "a1", "check": "排序结果正确", "machine_checkable": True}],
|
||||
"plan": [
|
||||
{"id": "s1", "task": "实现快排主函数", "deps": [], "done_criteria": "可运行"},
|
||||
{"id": "s2", "task": "补充单元测试", "deps": ["s1"], "done_criteria": "3/3 通过"},
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def _ws(**kw):
|
||||
return Workspace.new(request_id=kw.get("request_id", "abc123def456"),
|
||||
query=kw.get("query", "写个快排"),
|
||||
api_token_cap=kw.get("api_token_cap", 8000),
|
||||
rounds_cap=kw.get("rounds_cap", 6))
|
||||
|
||||
|
||||
# ---------- 结构/校验 ----------
|
||||
def test_new_has_valid_defaults():
|
||||
ws = _ws()
|
||||
assert ws.status == "draft"
|
||||
assert validate(ws.data) == []
|
||||
b = ws.budget()
|
||||
assert b["api_token_cap"] == 8000 and b["rounds_cap"] == 6
|
||||
|
||||
|
||||
def test_brief_write_once():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
assert ws.status == "in_progress"
|
||||
with pytest.raises(ValueError):
|
||||
ws.apply_brief(ALLOWED_BRIEF) # 二次写入被拒
|
||||
|
||||
|
||||
def test_apply_brief_missing_field_rejected():
|
||||
ws = _ws()
|
||||
bad = dict(ALLOWED_BRIEF)
|
||||
del bad["plan"]
|
||||
with pytest.raises(ValueError):
|
||||
ws.apply_brief(bad)
|
||||
|
||||
|
||||
def test_apply_brief_invalid_tag_rejected():
|
||||
ws = _ws()
|
||||
bad = dict(ALLOWED_BRIEF, tags=["hacker"])
|
||||
with pytest.raises(ValueError):
|
||||
ws.apply_brief(bad)
|
||||
|
||||
|
||||
def test_apply_brief_overlong_goal_rejected():
|
||||
ws = _ws()
|
||||
bad = dict(ALLOWED_BRIEF, goal="长" * (LIMITS["goal"] + 1))
|
||||
with pytest.raises(ValueError):
|
||||
ws.apply_brief(bad)
|
||||
|
||||
|
||||
def test_apply_brief_too_many_plan_steps_rejected():
|
||||
ws = _ws()
|
||||
plan = [{"id": f"s{i}", "task": "t", "deps": [], "done_criteria": "c"}
|
||||
for i in range(LIMITS["plan_steps"] + 1)]
|
||||
bad = dict(ALLOWED_BRIEF, plan=plan)
|
||||
with pytest.raises(ValueError):
|
||||
ws.apply_brief(bad)
|
||||
|
||||
|
||||
# ---------- 状态机 ----------
|
||||
def test_status_transitions():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.transition("reviewing")
|
||||
ws.transition("done")
|
||||
assert ws.status == "done"
|
||||
with pytest.raises(ValueError):
|
||||
ws.transition("in_progress") # done 之后不允许再回
|
||||
|
||||
|
||||
def test_illegal_transition():
|
||||
ws = _ws()
|
||||
with pytest.raises(ValueError):
|
||||
ws.transition("done") # draft 不能直接到 done
|
||||
|
||||
|
||||
# ---------- 写入 + 预算 ----------
|
||||
def test_progress_and_issues_and_decisions():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.add_progress("s1", "done", "实现完成", artifact="a://s1_main.py")
|
||||
iid = ws.add_issue("s2", "a://s2_tests.py#L12-18",
|
||||
"测试失败", "3/3 通过", "修了两次", "请裁决")
|
||||
assert iid == "i1"
|
||||
ws.add_decision(iid, "改用排序后断言", [{"id": "s2", "task": "修测试"}])
|
||||
assert len(ws["progress"]) == 1
|
||||
assert len(ws["issues"]) == 1
|
||||
assert len(ws["decisions"]) == 1
|
||||
|
||||
|
||||
def test_budget_accumulation_and_exhaustion():
|
||||
ws = _ws(api_token_cap=10)
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.add_budget(input_tokens=6, output_tokens=2)
|
||||
assert ws.exhausted() is False
|
||||
ws.add_budget(input_tokens=3) # 累计 11 > 10
|
||||
assert ws.exhausted() is True
|
||||
|
||||
|
||||
def test_round_exhaustion():
|
||||
ws = _ws(rounds_cap=2)
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.mark_round()
|
||||
ws.mark_round()
|
||||
assert ws.exhausted() is True
|
||||
|
||||
|
||||
def test_issue_text_overlong_rejected():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
with pytest.raises(ValueError):
|
||||
ws.add_issue("s1", "a://f.py", "观" * (LIMITS["issue_text"] + 1), "e", "t", "q")
|
||||
|
||||
|
||||
# ---------- rollup ----------
|
||||
def test_rollup_folds_done_progress_and_resolved_issues():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.add_progress("s1", "done", "快排实现完成", "a://s1_main.py")
|
||||
ws.add_progress("s2", "failed", "测试未过")
|
||||
iid = ws.add_issue("s2", "a://s2_tests.py#L1", "失败", "通过", "试过", "问")
|
||||
ws.add_decision(iid, "已修复")
|
||||
n = ws.rollup()
|
||||
assert n == 1 # 折叠 1 条 done progress
|
||||
assert "s1" in "".join(ws["archive"])
|
||||
# s2 failed 保留；已 resolve 的 issue 从 issues 移除
|
||||
steps = [p["step"] for p in ws["progress"]]
|
||||
assert steps == ["s2"]
|
||||
assert len(ws["issues"]) == 0
|
||||
|
||||
|
||||
def test_rollup_keeps_unresolved_issue():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.add_progress("s1", "done", "ok", "a://s1.py")
|
||||
ws.add_issue("s2", "a://s2.py#L1", "obs", "exp", "try", "ask") # 无 decision
|
||||
ws.rollup()
|
||||
assert len(ws["issues"]) == 1
|
||||
|
||||
|
||||
# ---------- 渲染 ----------
|
||||
def test_render_for_architect_within_budget():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
for i in range(6):
|
||||
ws.add_progress(f"s{i}", "done", f"步骤{i}完成" * 3)
|
||||
ws.add_issue(f"s{i}", f"a://f{i}.py#L1", "obs", "exp", "try", "ask" * 10)
|
||||
ws.add_decision(f"i{i+1}", "裁决" * 20)
|
||||
out = ws.render_for_architect()
|
||||
assert estimate_tokens(out) <= 1200
|
||||
assert "== meta ==" in out
|
||||
|
||||
|
||||
def test_render_for_worker_has_step_and_criteria():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
out = ws.render_for_worker("s2", artifact_text="def test(): pass")
|
||||
assert "当前步" in out
|
||||
assert "s2" in out
|
||||
assert "def test(): pass" in out
|
||||
assert "验收标准" in out and "a1" in out
|
||||
|
||||
|
||||
def test_render_for_worker_marks_current_step():
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
out = ws.render_for_worker("s1")
|
||||
assert "<-- 当前步" in out
|
||||
|
||||
|
||||
# ---------- 锚点 ----------
|
||||
def test_anchor_build_and_parse():
|
||||
a = build_anchor("s1_main.py", 12, 18)
|
||||
assert a == "a://s1_main.py#L12-18"
|
||||
p = parse_anchor(a)
|
||||
assert p == {"file": "s1_main.py", "start": 12, "end": 18}
|
||||
# 无行号
|
||||
assert parse_anchor("a://f.py") == {"file": "f.py", "start": 1, "end": 1}
|
||||
assert parse_anchor("not-an-anchor") is None
|
||||
|
||||
|
||||
# ---------- 持久化 ----------
|
||||
def test_save_load_roundtrip(tmp_path):
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.add_progress("s1", "done", "ok")
|
||||
path = tmp_path / "ws.json"
|
||||
ws.save(path)
|
||||
loaded = Workspace.load(path)
|
||||
assert loaded.to_dict() == ws.to_dict()
|
||||
assert loaded.request_id == ws.request_id
|
||||
|
||||
|
||||
def test_json_serializable(tmp_path):
|
||||
ws = _ws()
|
||||
ws.apply_brief(ALLOWED_BRIEF)
|
||||
ws.add_issue("s1", "a://f.py#L1", "obs", "exp", "try", "ask")
|
||||
# 应能被 json 直接序列化（中文 ensure_ascii=False）
|
||||
s = json.dumps(ws.to_dict(), ensure_ascii=False)
|
||||
assert json.loads(s)["request_id"] == ws.request_id
|
||||
|
||||
|
||||
def test_validate_rejects_non_object():
|
||||
assert validate("nope") != []
|
||||
|
||||
|
||||
def test_status_flow_enumeration():
|
||||
assert set(STATUS_FLOW.keys()) == {"draft", "in_progress", "reviewing", "escalated", "done", "failed"}
|
||||
+1
-26
@@ -70,7 +70,7 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
|
||||
| T1 | 环境与基线确认（126 测试全绿；README 环境备忘） | ✅ 完成 | （并入 T2 commit） |
|
||||
| T2 | 运维层：hw_profile + llama_server 进程管理 | ✅ 完成 | T2 |
|
||||
| T3 | ArchitectClient（DeepSeek API，JSON 约束） | ⬜ | |
|
||||
| T4 | Workspace（交流文本 schema/校验/渲染/rollup） | ⬜ | |
|
||||
| T4 | Workspace（交流文本 schema/校验/渲染/rollup） | ✅ 完成 | T4 |
|
||||
| T5 | WorkerLoop + 接地验证 | ⬜ | |
|
||||
| T6 | CollaborativePipeline 编排 | ⬜ | |
|
||||
| T7 | 网关扩展（/chat 切 v2，/chat/legacy） | ⬜ | |
|
||||
@@ -81,28 +81,3 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
|
||||
| T12 | 实验脚本 bench_tokens.py + 数据集 | ⬜ | |
|
||||
| T13 | E1–E5 跑数到 research/v2_experiments/ | ⬜ | |
|
||||
| T14 | 文档收口（README v2 改写） | ⬜ | |
|
||||
|
||||
|
||||
---
|
||||
|
||||
## 五、v2 任务登记（端云协同 LLM 协作系统，见《实现方案_v2_端云协同LLM协作系统.md》）
|
||||
|
||||
> 每个任务一个 commit（`feat(v2): Tn 描述`），交付含封闭单测；v1 的 126 项测试保持全绿。
|
||||
|
||||
| T | 内容 | 状态 | commit |
|
||||
|---|------|------|--------|
|
||||
| T1 | 环境与基线确认（126 测试全绿；README 环境备忘） | ✅ 完成 | （并入 T2 commit） |
|
||||
| T2 | 运维层：hw_profile + llama_server 进程管理 | ✅ 完成 | T2 |
|
||||
| T3 | ArchitectClient（DeepSeek API，JSON 约束） | ⬜ | |
|
||||
| T4 | Workspace（交流文本 schema/校验/渲染/rollup） | ⬜ | |
|
||||
| T5 | WorkerLoop + 接地验证 | ⬜ | |
|
||||
| T6 | CollaborativePipeline 编排 | ⬜ | |
|
||||
| T7 | 网关扩展（/chat 切 v2，/chat/legacy） | ⬜ | |
|
||||
| T8 | 人工检验队列 ReviewQueue | ⬜ | |
|
||||
| T9 | token 计量与账单 | ⬜ | |
|
||||
| T10 | rollup + prefix cache 调优 | ⬜ | |
|
||||
| T11 | 打包分发 setup_runtime.py | ⬜ | |
|
||||
| T12 | 实验脚本 bench_tokens.py + 数据集 | ⬜ | |
|
||||
| T13 | E1–E5 跑数到 research/v2_experiments/ | ⬜ | |
|
||||
| T14 | 文档收口（README v2 改写） | ⬜ | |
|
||||
|
||||
|
||||
Reference in New Issue
Block a user