| """格式校验器（方案 §13 第1步）：技能条目与 curator 编辑操作的机械校验。 |
| |
| 校验的是"可执行性与可归因性"，不是内容好坏： |
| · plan 的工具名/参数必须在 tool_schema 内且取值合法 |
| · 数值不得埋在自然语言字段（trigger/note 里出现 k=8、±10s 这类 → 拒） |
| · trigger 不得是无区分度写法（"任何问题"/"always"） |
| · target_buckets 必须是合法桶 |
| · plan 步骤数 ≤ MAX_PLAN_STEPS，代价 ≤ 预算 |
| · stop_condition / anti_pattern 必须是机器可判定结构 |
| · plan 与 stop_condition.max_tool_calls 不得自相矛盾（旧版 stop-loss 与 global-then-local 冲突） |
| """ |
|
|
| import re |
| from typing import Any, Dict, List, Tuple |
|
|
| from .tool_schema import TOOL_SCHEMA, MAX_PLAN_STEPS, DEFAULT_BUDGET, plan_cost |
| from .skill_schema import VALID_BUCKETS, FAILURE_MODES |
|
|
| |
| _NUM_IN_TEXT = re.compile(r"(k\s*=\s*\d+|top-?\d+|±\s*\d+\s*s|\b\d+\s*(?:s|sec|秒|frames?|帧)\b" |
| r"|\b\d+\s*次\b|max\s*\d+)", re.I) |
| _VAGUE_TRIGGER = re.compile(r"(any question|always applicable|任何问题|始终适用|all questions)", re.I) |
| |
| RUNTIME_PARAMS = {"query", "object", "t_start", "t_end", "t_min", "t_max"} |
| _PLACEHOLDER = re.compile(r"^\$(now|hit\.t|anchor\.t_start|anchor\.t_end)([+-]\d+(\.\d+)?)?$") |
|
|
|
|
| def validate_skill(sk: Dict[str, Any], budget: int = DEFAULT_BUDGET) -> Tuple[bool, List[str]]: |
| errs: List[str] = [] |
| |
| for f in ("skill_id", "trigger", "target_buckets", "plan", "params", |
| "stop_condition", "anti_pattern", "stats"): |
| if f not in sk: |
| errs.append(f"缺字段 {f}") |
| if errs: |
| return False, errs |
|
|
| |
| if _VAGUE_TRIGGER.search(sk["trigger"]): |
| errs.append("trigger 无区分度（禁'任何问题/始终适用'——全局预算属系统配置，不是技能）") |
| if len(sk["trigger"].strip()) < 12: |
| errs.append("trigger 过短，无法判定适用性") |
|
|
| |
| for field, text in [("trigger", sk["trigger"])] + \ |
| [(f"plan[{i}].note", s.get("note", "")) for i, s in enumerate(sk["plan"])]: |
| m = _NUM_IN_TEXT.search(str(text)) |
| if m: |
| errs.append(f"{field} 中埋了数值参数 '{m.group(0)}' —— 必须移到 params") |
|
|
| |
| if not sk["target_buckets"]: |
| errs.append("target_buckets 为空（注入靠桶精确匹配，空则永不命中）") |
| for b in sk["target_buckets"]: |
| if b not in VALID_BUCKETS: |
| errs.append(f"非法桶 {b}") |
|
|
| |
| if not isinstance(sk["plan"], list) or not sk["plan"]: |
| errs.append("plan 必须是非空步骤列表") |
| elif len(sk["plan"]) > MAX_PLAN_STEPS: |
| errs.append(f"plan 步骤 {len(sk['plan'])} > 上限 {MAX_PLAN_STEPS}") |
| else: |
| for i, st in enumerate(sk["plan"]): |
| tool = st.get("tool") |
| if tool not in TOOL_SCHEMA: |
| errs.append(f"plan[{i}] 未知工具 {tool}") |
| continue |
| spec = TOOL_SCHEMA[tool]["params"] |
| for p, v in (st.get("params") or {}).items(): |
| if p not in spec: |
| errs.append(f"plan[{i}] {tool} 无参数 {p}") |
| continue |
| d = spec[p] |
| if isinstance(v, str) and v.startswith("$"): |
| continue |
| if "range" in d: |
| try: |
| if not (d["range"][0] <= float(v) <= d["range"][1]): |
| errs.append(f"plan[{i}] {tool}.{p}={v} 越界 {d['range']}") |
| except (TypeError, ValueError): |
| errs.append(f"plan[{i}] {tool}.{p}={v} 非数值") |
| if "choices" in d and v not in d["choices"]: |
| errs.append(f"plan[{i}] {tool}.{p}={v} 不在 {d['choices']}") |
| |
| |
| for p, d in spec.items(): |
| if not d.get("required") or p in RUNTIME_PARAMS: |
| continue |
| if p not in (st.get("params") or {}): |
| errs.append(f"plan[{i}] {tool} 缺必填参数 {p}") |
| for p, v_ in (st.get("params") or {}).items(): |
| if isinstance(v_, str) and v_.startswith("$") and not _PLACEHOLDER.match(v_): |
| errs.append(f"plan[{i}] {tool}.{p} 占位符 '{v_}' 非法" |
| f"（合法: $now / $hit.t±N / $anchor.t_start±N / $anchor.t_end±N）") |
|
|
| |
| sc = sk.get("stop_condition") or {} |
| if "max_tool_calls" not in sc: |
| errs.append("stop_condition 缺 max_tool_calls（必须可机械判定）") |
| elif isinstance(sk.get("plan"), list) and len(sk["plan"]) > int(sc["max_tool_calls"]): |
| errs.append(f"自相矛盾: plan {len(sk['plan'])} 步 > max_tool_calls {sc['max_tool_calls']}") |
|
|
| |
| ap = sk.get("anti_pattern") or {} |
| if "forbidden_sequence" not in ap: |
| errs.append("anti_pattern 必须含 forbidden_sequence（列表形式，可机械判定）") |
| else: |
| for seq in ap["forbidden_sequence"]: |
| if not isinstance(seq, list) or any(t not in TOOL_SCHEMA for t in seq): |
| errs.append(f"anti_pattern 序列 {seq} 含未知工具或非列表") |
|
|
| |
| if isinstance(sk.get("plan"), list): |
| c = plan_cost(sk["plan"]) |
| if c > budget: |
| errs.append(f"plan 代价 {c} > 预算 {budget}") |
| return not errs, errs |
|
|
|
|
| def validate_edit(op: Dict[str, Any], lib: Dict[str, Any]) -> Tuple[bool, List[str]]: |
| """校验 curator 的编辑操作（方案 §6.2 的四种顶层算子）。""" |
| errs = [] |
| t = op.get("op") |
| if t not in ("add", "modify", "delete", "merge"): |
| return False, [f"未知算子 {t}"] |
| if not op.get("failure_mode") or op["failure_mode"] not in FAILURE_MODES: |
| errs.append(f"failure_mode 必须来自枚举，收到 {op.get('failure_mode')}") |
| if op.get("failure_mode") in ("bucket_misclassified", "low_adherence") and t in ("add", "delete"): |
| errs.append(f"{op['failure_mode']} 的修复对象不是技能内容，不应 {t}") |
| if not op.get("evidence_rollouts"): |
| errs.append("缺 evidence_rollouts（必须指明依据哪些 rollout 对照）") |
| ids = {s["skill_id"] for s in lib.get("skills", [])} |
| if t == "add": |
| ok, e = validate_skill(op.get("skill") or {}) |
| errs += [f"新技能: {x}" for x in e] |
| elif t == "modify": |
| if op.get("skill_id") not in ids: |
| errs.append(f"目标技能不存在 {op.get('skill_id')}") |
| if not op.get("changes"): |
| errs.append("modify 缺 changes") |
| elif t == "delete": |
| if op.get("skill_id") not in ids: |
| errs.append(f"目标技能不存在 {op.get('skill_id')}") |
| elif t == "merge": |
| if len(op.get("skill_ids") or []) < 2 or not set(op["skill_ids"]) <= ids: |
| errs.append("merge 需 ≥2 个存在的 skill_id") |
| ok, e = validate_skill(op.get("skill") or {}) |
| errs += [f"合并结果: {x}" for x in e] |
| return not errs, errs |
|
|
