@cannbot-plugin/cannbot-harness 0.0.0-stage → 0.1.4

package/assets/ops-direct-invoke/skills/ops-spec-gen/scripts/generate_spec.py

@@ -0,0 +1,1571 @@
+#!/usr/bin/env python3
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+
+"""Interactive algorithm-spec.yaml generator.
+
+Reads registries/*.yaml to get controlled enums (26 categories, 27 paradigms,
+category->required-paradigms map), prompts the user for the minimum information
+needed, and writes ops/<op>/spec.yaml using templates/spec.yaml.tmpl.
+
+Non-interactive use:
+    python3 generate_spec.py \
+        --op-name softmax --category reduction_composite \
+        --paradigms Reduction,NumericalStable,FusedComposite \
+        --inputs x:float16,float32,bfloat16 \
+        --outputs y \
+        --output-dir /path/to/ops/softmax
+
+After generation, run validate_spec.py to catch the {{TODO}} placeholders that
+must be filled in by hand (formula, reference_oracle.api, supported_combinations,
+boundary_conditions, etc.).
+"""
+
+from __future__ import annotations
+
+import argparse
+import re
+import sys
+from dataclasses import dataclass, field
+from pathlib import Path
+
+import yaml
+
+SKILL_DIR = Path(__file__).resolve().parent.parent
+REGISTRIES = SKILL_DIR / "registries"
+TEMPLATE = SKILL_DIR / "templates" / "spec.yaml.tmpl"
+
+
+def _load(name: str) -> dict:
+    with open(REGISTRIES / name, "r", encoding="utf-8") as f:
+        return yaml.safe_load(f)
+
+
+def load_registries() -> dict:
+    import json as _json
+    schema = _json.loads((SKILL_DIR / "schemas" / "op-spec.json").read_text(encoding="utf-8"))
+    return {
+        "categories": _load("category_enum.yaml")["categories"],
+        "paradigms": [p["name"] for p in _load("paradigm_enum.yaml")["paradigms"]],
+        "category_map": _load("category_paradigm_map.yaml"),
+        "chips": _load("chip_registry.yaml")["chips"],
+        "layouts": schema.get("$defs", {}).get("layout", {}).get("enum", []),
+    }
+
+
+def _chip_dtype_table(chips: list[dict]) -> dict[str, set[str]]:
+    """chip_id → supported_dtypes set（从 registry 抽出来的查表）。"""
+    return {c["id"]: set(c.get("supported_dtypes") or []) for c in chips}
+
+
+def narrow_chips_by_dtypes(chips: list[dict], dtypes: set[str]) -> list[str]:
+    """返回能覆盖所有 dtypes 的芯片 id 列表；dtypes 为空时返回全部 chip id。"""
+    if not dtypes:
+        return [c["id"] for c in chips]
+    out: list[str] = []
+    for c in chips:
+        supported = set(c.get("supported_dtypes") or [])
+        if dtypes.issubset(supported):
+            out.append(c["id"])
+    return out
+
+
+def _load_tolerance_defaults() -> dict:
+    """Read per-dtype tolerance defaults from the shared registry."""
+    return _load("tolerance_defaults.yaml")["defaults"]
+
+
+@dataclass
+class TensorSpec:
+    name: str
+    dtype_set: list[str]
+    rank_range: tuple[int, int] = (0, 8)
+    data_distribution: str = ""
+    role: str | None = None
+    optional: bool = False
+    list_length: dict | None = None
+
+
+@dataclass
+class OutputSpec:
+    name: str
+    role: str = "tensor"
+    optional: bool = False
+    list_length: dict | None = None
+
+
+@dataclass
+class GenInput:
+    op_name: str
+    description: str
+    category: str
+    paradigms: list[str]
+    inputs: list[TensorSpec]
+    outputs: list[str | OutputSpec]
+    promotion: str = "same_as_first_input"
+    broadcast_kind: str = "numpy"
+    accumulation_order: str = "none"
+    numerical_stability_required: bool = False
+    supported_chips: list[str] = None  # type: ignore[assignment]
+    axis_source: str = "attribute"
+    paradigm_groups_mode: str = ""  # "" | "combination" | "fusion"
+    format_variants: list[dict] = field(default_factory=list)
+    attributes: list[dict] = field(default_factory=list)
+    semantic_cases: list[dict] = field(default_factory=list)
+    layout_contract: dict | None = None
+
+    def __post_init__(self):
+        if self.supported_chips is None:
+            self.supported_chips = []
+
+
+def _default_input_role(role: str | None, paradigms: list[str], index: int) -> str:
+    """Resolve the implicit role used by both interactive and CLI rendering."""
+    if role:
+        return role
+    return "state" if "Stateful" in paradigms and index == 0 else "tensor"
+
+
+def _filter_elementwise(paradigms: list[str]) -> list[str]:
+    """Elementwise 与其他范式组合时，只保留非 Elementwise 的。
+
+    [Elementwise, Reduction] → [Reduction]
+    [Elementwise] → [Elementwise]（仅 Elementwise 时保留）
+    """
+    if "Elementwise" in paradigms and len(paradigms) > 1:
+        return [p for p in paradigms if p != "Elementwise"]
+    return paradigms
+
+
+_PARADIGM_TO_CATEGORY: dict[str, str] = {
+    "Broadcast": "Broadcast",
+    "Reduction": "Reduction",
+    "Contraction": "Contraction",
+    "ArgReduce": "ArgReduce",
+    "LayoutTransform": "LayoutTransform",
+    "Padding": "Padding",
+    "IndexGather": "IndexGather",
+    "ScatterUpdate": "ScatterUpdate",
+    "AtomicUpdate": "AtomicUpdate",
+    "Interpolation": "Interpolation",
+    "MaskPredicate": "MaskPredicate",
+    "SortSelect": "SortSelect",
+    "Histogram": "Histogram",
+    "SlidingWindow": "SlidingWindow",
+    "ControlFlow": "ControlFlow",
+    "Recurrence": "Recurrence",
+    "Spectral": "Spectral",
+    "VariableOutput": "VariableOutput",
+    "RandomSampling": "RandomSampling",
+    "Stateful": "Stateful",
+    "Sparse": "Sparse",
+}
+
+
+def _resolve_category(category: str, paradigms: list[str]) -> str:
+    """当 category=Elementwise 但存在其他 paradigm 时，用主 paradigm 替换 category。
+
+    [Elementwise, Reduction] → category=Reduction
+    [Elementwise]             → category=Elementwise（保持不变）
+    """
+    if category == "Elementwise" and paradigms:
+        main = paradigms[0]
+        return _PARADIGM_TO_CATEGORY.get(main, category)
+    return category
+
+
+def _is_pure_reduction(category: str, paradigms: list[str]) -> bool:
+    return category == "Reduction" and "Reduction" in paradigms and "Recurrence" not in paradigms
+
+
+def parse_format_variants(raw: str, valid_layouts: list[str] | None = None) -> list[dict]:
+    """Parse 'FORMAT:RANK:AXES;...' into list of variant dicts.
+
+    Example: 'NCHW:4:0,2,3;NHWC:4:0,1,2;NCDHW:5:0,2,3,4'
+    → [{"format": "NCHW", "rank": [4], "reduction_axes": [0,2,3]}, ...]
+
+    If valid_layouts is provided, format names are validated against it.
+    """
+    if not raw.strip():
+        return []
+    variants = []
+    for part in raw.split(";"):
+        part = part.strip()
+        if not part:
+            continue
+        tokens = part.split(":")
+        if len(tokens) != 3:
+            raise SystemExit(
+                f"--format-variants 格式错误：'{part}'，"
+                "期望 FORMAT:RANK:AXES (如 NCHW:4:0,2,3)")
+        fmt, rank_str, axes_str = tokens
+        fmt = fmt.strip()
+        if valid_layouts and fmt not in valid_layouts:
+            raise SystemExit(
+                f"--format-variants 格式错误：'{fmt}' 不是有效的 layout 值，"
+                f"有效值: {valid_layouts}")
+        try:
+            ranks = [int(r.strip()) for r in rank_str.split(",") if r.strip()]
+            axes = [int(a.strip()) for a in axes_str.split(",") if a.strip()]
+        except ValueError:
+            raise SystemExit(
+                f"--format-variants 格式错误：'{part}'，"
+                "RANK 和 AXES 必须是逗号分隔的整数")
+        variants.append({
+            "format": fmt.strip(),
+            "rank": ranks,
+            "reduction_axes": axes,
+        })
+    return variants
+
+
+def parse_tensor_arg(s: str) -> TensorSpec:
+    """Parse 'x:float16,float32,bfloat16' or 'x'."""
+    if ":" in s:
+        name, dtypes = s.split(":", 1)
+        dtype_set = [d.strip() for d in dtypes.split(",") if d.strip()]
+    else:
+        name, dtype_set = s, ["float32"]
+    if not re.match(r"^[a-z][a-zA-Z0-9_]*$", name):
+        raise ValueError(f"Invalid tensor name: {name}")
+    return TensorSpec(name=name, dtype_set=dtype_set)
+
+
+def _normalize_outputs(outputs: list[str | OutputSpec]) -> list[OutputSpec]:
+    """Preserve the historical list[str] caller interface while supporting roles."""
+    return [item if isinstance(item, OutputSpec) else OutputSpec(name=item) for item in outputs]
+
+
+def _output_names(outputs: list[str | OutputSpec]) -> list[str]:
+    return [item.name if isinstance(item, OutputSpec) else item for item in outputs]
+
+
+def _render_mapping(lines: list[str], key: str, value: dict, indent: str) -> None:
+    """Render a small YAML mapping without introducing a second serializer path."""
+    lines.append(f"{indent}{key}:")
+    dumped = yaml.safe_dump(value, allow_unicode=True, default_flow_style=False, sort_keys=False).rstrip()
+    lines.extend(f"{indent}  {line}" for line in dumped.splitlines())
+
+
+def prompt(msg: str, default: str = "", allow_empty: bool = False) -> str:
+    """Read a line from stdin; type 'q' or Ctrl-D / Ctrl-C to quit.
+
+    allow_empty=True 用于"空回车 = 结束 / 跳过"的交互场景（如列表收集），
+    此时不再循环要求重新输入。
+    """
+    while True:
+        try:
+            ans = input(f"{msg} [{default}]: " if default else f"{msg}: ").strip()
+        except (EOFError, KeyboardInterrupt):
+            print("\n  ↷ 用户取消生成（EOF / Ctrl-C）")
+            raise SystemExit(0)
+        if ans.lower() in ("q", "quit", "exit"):
+            print("  ↷ 用户取消生成（输入 q）")
+            raise SystemExit(0)
+        if ans:
+            return ans
+        if default:
+            return default
+        if allow_empty:
+            return ""
+
+
+def prompt_choice(msg: str, choices: list[str], default: str = "") -> str:
+    """Pick one from a list. Type 'q' / Ctrl-D / Ctrl-C to quit."""
+    print(f"{msg}")
+    for i, c in enumerate(choices, 1):
+        marker = "*" if c == default else " "
+        print(f"  {marker} {i:2d}. {c}")
+    print("  (输入 q / Ctrl-D 退出生成)")
+    while True:
+        try:
+            ans = input(f"  selection [{default or '?'}]: ").strip()
+        except (EOFError, KeyboardInterrupt):
+            print("\n  ↷ 用户取消生成")
+            raise SystemExit(0)
+        if ans.lower() in ("q", "quit", "exit"):
+            print("  ↷ 用户取消生成")
+            raise SystemExit(0)
+        if not ans and default:
+            return default
+        if ans.isdigit() and 1 <= int(ans) <= len(choices):
+            return choices[int(ans) - 1]
+        if ans in choices:
+            return ans
+        print(f"  ✗ 无效输入 {ans!r}；请输入 1-{len(choices)} 的编号或选项名")
+
+
+def prompt_multi_choice(msg: str, choices: list[str], defaults: list[str] | None = None) -> list[str]:
+    """Pick one or more from a list (comma-separated indices or names).
+
+    空回车 → 返回 defaults（若提供）或全选。q / Ctrl-D 退出生成。
+    """
+    print(f"{msg}")
+    default_set = set(defaults or [])
+    for i, c in enumerate(choices, 1):
+        marker = "*" if c in default_set else " "
+        print(f"  {marker} {i:2d}. {c}")
+    hint = "逗号分隔编号或名称；空回车=默认；q 退出"
+    print(f"  ({hint})")
+    while True:
+        try:
+            ans = input(f"  selection [{','.join(defaults) if defaults else 'all'}]: ").strip()
+        except (EOFError, KeyboardInterrupt):
+            print("\n  ↷ 用户取消生成")
+            raise SystemExit(0)
+        if ans.lower() in ("q", "quit", "exit"):
+            print("  ↷ 用户取消生成")
+            raise SystemExit(0)
+        if not ans:
+            return list(defaults) if defaults else list(choices)
+        picked: list[str] = []
+        bad: list[str] = []
+        for tok in [t.strip() for t in ans.split(",") if t.strip()]:
+            if tok.isdigit() and 1 <= int(tok) <= len(choices):
+                c = choices[int(tok) - 1]
+            elif tok in choices:
+                c = tok
+            else:
+                bad.append(tok)
+                continue
+            if c not in picked:
+                picked.append(c)
+        if bad:
+            print(f"  ✗ 无效项: {bad}；请重新输入")
+            continue
+        if not picked:
+            print(f"  ✗ 至少选一项")
+            continue
+        return picked
+
+
+def prompt_name_list(msg: str) -> list[str]:
+    """Collect a comma-separated list; schema/stage 2 validates declared names."""
+    raw = prompt(msg, default="", allow_empty=True)
+    return [name.strip() for name in raw.split(",") if name.strip()]
+
+
+def _prompt_non_negative_int(msg: str, default: str) -> int:
+    """Prompt until a decimal non-negative integer is entered."""
+    while True:
+        raw = prompt(msg, default=default)
+        try:
+            value = int(raw)
+        except ValueError:
+            print("  [error] value must be a non-negative integer")
+            continue
+        if value < 0:
+            print("  [error] value must be a non-negative integer")
+            continue
+        return value
+
+
+def collect_list_length(label: str) -> dict:
+    """Collect the outer-length contract for one TensorList in interactive mode."""
+    kind = prompt_choice(
+        f"  {label}.list_length.kind",
+        ["unconstrained", "fixed", "range", "same_as", "expression"],
+        default="unconstrained",
+    )
+    if kind == "fixed":
+        return {
+            "kind": kind,
+            "value": _prompt_non_negative_int(
+                "    value (non-negative integer)", default="0",
+            ),
+        }
+    if kind == "range":
+        while True:
+            minimum = _prompt_non_negative_int(
+                "    min (non-negative integer)", default="0",
+            )
+            maximum = _prompt_non_negative_int(
+                "    max (non-negative integer)", default="1",
+            )
+            if minimum <= maximum:
+                return {"kind": kind, "min": minimum, "max": maximum}
+            print("  [error] min must be less than or equal to max")
+    if kind == "same_as":
+        return {"kind": kind, "ref": prompt("    ref (input.<name> or output.<name>)")}
+    if kind == "expression":
+        return {"kind": kind, "value": prompt("    expression")}
+    return {"kind": kind}
+
+
+def collect_conditional_attributes() -> list[dict]:
+    """Collect existing `attributes[]` entries needed by conditional signatures."""
+    attributes: list[dict] = []
+    if prompt_choice("是否补充用于条件签名的 attributes", ["no", "yes"], default="no") != "yes":
+        return attributes
+    print("  仅填写条件签名实际依赖的属性；自动注入属性无需重复填写。")
+    while True:
+        name = prompt(f"Attribute #{len(attributes) + 1} name (blank to finish)", default="", allow_empty=True)
+        if not name:
+            break
+        attr_type = prompt_choice(
+            "  type",
+            ["bool", "enum", "int64", "float32", "string", "list[int64]"],
+            default="bool",
+        )
+        default_text = prompt(
+            "  default (YAML scalar/list)", default="false" if attr_type == "bool" else "", allow_empty=True,
+        )
+        try:
+            default = yaml.safe_load(default_text)
+        except yaml.YAMLError as exc:
+            print(f"  ✗ default 不是合法 YAML 值: {exc}")
+            continue
+        attributes.append({
+            "name": name,
+            "type": attr_type,
+            "default": default,
+            "semantics": prompt("  semantics"),
+        })
+    return attributes
+
+
+def collect_semantic_cases(inputs: list[TensorSpec], outputs: list[OutputSpec]) -> list[dict]:
+    """Collect only presence contracts; formula/shape/dtype stay in their existing fields."""
+    if not any(item.optional for item in [*inputs, *outputs]):
+        return []
+    if prompt_choice("是否声明 optional I/O 的 semantic_cases", ["no", "yes"], default="no") != "yes":
+        return []
+    cases: list[dict] = []
+    print("  when 可使用 attr.<name>、input.<name>.dtype、input.<name>.is_present。")
+    while True:
+        case_id = prompt(f"Case #{len(cases) + 1} id (blank to finish)", default="", allow_empty=True)
+        if not case_id:
+            break
+        cases.append({
+            "id": case_id,
+            "when": prompt("  when"),
+            "inputs": {
+                "required": prompt_name_list("  required inputs (comma-separated, blank for none)"),
+                "forbidden": prompt_name_list("  forbidden inputs (comma-separated, blank for none)"),
+            },
+            "outputs": {
+                "present": prompt_name_list("  present outputs (comma-separated, blank for none)"),
+                "absent": prompt_name_list("  absent outputs (comma-separated, blank for none)"),
+            },
+        })
+    return cases
+
+
+def collect_layout_contract(inputs: list[TensorSpec], outputs: list[OutputSpec]) -> dict | None:
+    """Collect an optional logical-layout contract without adding CLI parameters.
+
+    Physical format stays in the existing per-I/O ``layout`` field.  This
+    collector records only the observable logical axis order for single-Tensor
+    inputs and outputs in each conditional layout variant.
+    """
+    eligible = [
+        # TensorSpec.role=None is the legacy/default construction path used by
+        # the non-interactive generator; build_inputs_block renders it as
+        # role=tensor.  Treat it identically here.
+        *[f"input.{item.name}" for item in inputs if item.role in (None, "tensor")],
+        *[f"output.{item.name}" for item in outputs if item.role == "tensor"],
+    ]
+    if not eligible:
+        return None
+    if prompt_choice(
+        "是否声明 layout_contract（逻辑布局变体及各 Tensor 的轴顺序）",
+        ["no", "yes"], default="no",
+    ) != "yes":
+        return None
+
+    variants: list[dict] = []
+    print("  when 可使用 attr.<name>、input.<name>.dtype、input.<name>.is_present；留空表示无条件。")
+    print(f"  可声明的单 Tensor 引用: {', '.join(eligible)}")
+    while True:
+        variant_id = prompt(
+            f"Layout variant #{len(variants) + 1} id (blank to finish)",
+            default="", allow_empty=True,
+        )
+        if not variant_id:
+            if not variants:
+                print("  至少需要 1 个 layout variant 或选 no")
+                continue
+            break
+        variant: dict = {"id": variant_id, "tensors": []}
+        when = prompt("  when (blank for unconditional)", default="", allow_empty=True)
+        if when:
+            variant["when"] = when
+        while True:
+            ref = prompt(
+                "  tensor ref (input.<name> or output.<name>; blank to finish this variant)",
+                default="", allow_empty=True,
+            )
+            if not ref:
+                if not variant["tensors"]:
+                    print("    至少需要 1 个 Tensor 轴表")
+                    continue
+                break
+            axes = prompt("    logical_axes (comma-separated, e.g. B,S,N,D)")
+            variant["tensors"].append({
+                "ref": ref,
+                "logical_axes": [axis.strip() for axis in axes.split(",") if axis.strip()],
+            })
+        variants.append(variant)
+    return {"variants": variants}
+
+
+def interactive_collect(reg: dict) -> GenInput:
+    print("\n=== algorithm-spec.yaml interactive generator ===\n")
+    op_name = prompt("Operator name (lowercase, e.g. softmax)")
+    if not re.match(r"^[a-z][a-z0-9_]*$", op_name):
+        raise SystemExit(f"Invalid op name: {op_name}")
+    description = prompt("One-line description")
+
+    category = prompt_choice("Pick category (single-select, 26 controlled values):",
+                             reg["categories"])
+    required = reg["category_map"]["category_requires_paradigms"].get(category, [])
+    print(f"\n  category={category} requires paradigm(s): {required}")
+    extra = prompt(
+        "Add extra paradigms (comma-separated, blank for none)",
+        default="", allow_empty=True).strip()
+    paradigms = list(required)
+    if extra:
+        for p in [x.strip() for x in extra.split(",") if x.strip()]:
+            if p not in reg["paradigms"]:
+                raise SystemExit(f"Unknown paradigm: {p}. Allowed: {reg['paradigms']}")
+            if p not in paradigms:
+                paradigms.append(p)
+    paradigm_groups_mode = ""
+    if len(paradigms) >= 2:
+        paradigm_groups_mode = prompt_choice(
+            "paradigm_groups (范式之间的关系)",
+            ["none", "combination", "fusion"],
+            default="none")
+        if paradigm_groups_mode == "none":
+            paradigm_groups_mode = ""
+
+    # combination 模式下不过滤 Elementwise（Elementwise 是独立的范式分支）
+    if paradigm_groups_mode != "combination":
+        paradigms = _filter_elementwise(paradigms)
+    category = _resolve_category(category, paradigms)
+
+    axis_source = "attribute"
+    if "Reduction" in paradigms and "Recurrence" not in paradigms:
+        axis_source = prompt_choice(
+            "axis_source (how the reduction axis is specified)",
+            ["attribute", "input_tensor", "fixed", "implicit_all"],
+            default="attribute")
+
+    # format_variants: 多数据排布格式
+    format_variants: list[dict] = []
+    if "Reduction" in paradigms and "Recurrence" not in paradigms:
+        fv_answer = prompt_choice(
+            "是否声明 format_variants（归约轴因数据排布格式而异）",
+            ["no", "yes"],
+            default="no")
+        if fv_answer == "yes":
+            valid_layouts = reg.get("layouts", [])
+            print(f"  有效 layout 值: {valid_layouts}")
+            while True:
+                fmt = prompt(
+                    f"  format #{len(format_variants)+1} name (blank to finish)",
+                    default="", allow_empty=True).strip()
+                if not fmt:
+                    if not format_variants:
+                        print("  至少需要 1 个 format variant 或选 no")
+                        continue
+                    break
+                if valid_layouts and fmt not in valid_layouts:
+                    print(f"  ✗ '{fmt}' 不是有效 layout 值；有效值: {valid_layouts}")
+                    continue
+                rank_str = prompt(f"    rank (comma-separated, e.g. 4 or 4,5)",
+                                  default="4")
+                axes_str = prompt(f"    reduction_axes (comma-separated, e.g. 0,2,3)",
+                                  default="0")
+                try:
+                    ranks = [int(r.strip()) for r in rank_str.split(",")]
+                    axes = [int(a.strip()) for a in axes_str.split(",")]
+                except ValueError:
+                    print("  ✗ rank 和 axes 必须是逗号分隔的整数")
+                    continue
+                format_variants.append({
+                    "format": fmt,
+                    "rank": ranks,
+                    "reduction_axes": axes,
+                })
+
+    print("\n--- Inputs ---")
+    inputs: list[TensorSpec] = []
+    while True:
+        name = prompt(f"Input #{len(inputs)+1} name (blank to finish)", default="", allow_empty=True)
+        if not name:
+            if not inputs:
+                print("  must have ≥ 1 input")
+                continue
+            break
+        dtypes = prompt("  dtype_set (comma-separated)", default="float16,float32,bfloat16")
+        role = prompt_choice(
+            "  role",
+            ["tensor", "tensor_list", "scalar", "attribute_alias", "state"],
+            default=_default_input_role(None, paradigms, len(inputs)),
+        )
+        optional = prompt_choice("  optional", ["false", "true"], default="false") == "true"
+        list_length = collect_list_length(name) if role == "tensor_list" else None
+        inputs.append(TensorSpec(
+            name=name, dtype_set=[d.strip() for d in dtypes.split(",")],
+            role=role, optional=optional, list_length=list_length,
+        ))
+
+    print("\n--- Outputs ---")
+    outputs: list[OutputSpec] = []
+    while True:
+        name = prompt(f"Output #{len(outputs)+1} name (blank to finish)", default="", allow_empty=True)
+        if not name:
+            if not outputs:
+                print("  must have ≥ 1 output")
+                continue
+            break
+        role = prompt_choice("  role", ["tensor", "tensor_list"], default="tensor")
+        optional = prompt_choice("  optional", ["false", "true"], default="false") == "true"
+        list_length = collect_list_length(name) if role == "tensor_list" else None
+        outputs.append(OutputSpec(name=name, role=role, optional=optional, list_length=list_length))
+
+    attributes = collect_conditional_attributes()
+    semantic_cases = collect_semantic_cases(inputs, outputs)
+    layout_contract = collect_layout_contract(inputs, outputs)
+
+    promotion = prompt_choice(
+        "dtype_policy.promotion",
+        ["same_as_first_input", "numpy_promote", "fixed", "upcast_for_accum"],
+        default="same_as_first_input")
+    broadcast_kind = prompt_choice(
+        "broadcast.kind",
+        ["none", "numpy", "explicit"],
+        default="numpy")
+    accumulation_order = prompt_choice(
+        "determinism.accumulation_order",
+        ["none", "stable_in_axis", "tree", "sequential"],
+        default="none")
+
+    # supported_chips: 按 inputs.dtype_set 收窄候选
+    declared_dtypes: set[str] = set()
+    for t in inputs:
+        declared_dtypes.update(t.dtype_set)
+    candidates = narrow_chips_by_dtypes(reg["chips"], declared_dtypes)
+    if not candidates:
+        # 没有任一芯片覆盖全部 dtype（典型：dtype_set 横跨 fp4 + 老芯片）
+        print(f"\n  ⚠ 没有芯片同时支持所有 dtype {sorted(declared_dtypes)}；")
+        print(f"    请手动从全集挑选并在生成后修正 dtype_set / supported_combinations")
+        candidates = [c["id"] for c in reg["chips"]]
+    supported_chips = prompt_multi_choice(
+        f"\nsupported_chips（按 dtype_set {sorted(declared_dtypes)} 自动收窄；多选）",
+        candidates,
+        defaults=candidates,
+    )
+
+    return GenInput(
+        op_name=op_name,
+        description=description,
+        category=category,
+        paradigms=paradigms,
+        inputs=inputs,
+        outputs=outputs,
+        promotion=promotion,
+        broadcast_kind=broadcast_kind,
+        accumulation_order=accumulation_order,
+        numerical_stability_required=("NumericalStable" in paradigms),
+        supported_chips=supported_chips,
+        axis_source=axis_source,
+        paradigm_groups_mode=paradigm_groups_mode,
+        format_variants=format_variants,
+        attributes=attributes,
+        semantic_cases=semantic_cases,
+        layout_contract=layout_contract,
+    )
+
+
+def build_attributes_block(category: str, paradigms: list[str], inputs: list[TensorSpec],
+                           axis_source: str = "attribute", explicit_attributes: list[dict] | None = None) -> str:
+    """Inject minimum required attrs per paradigm injection table.
+
+    Multiple injectors may apply (e.g. Quantization + RandomSampling). Order
+    matters only cosmetically — collected lines are joined.
+
+    axis_source controls how the reduction axis is specified:
+      - attribute:      dim attribute (default)
+      - input_tensor:   axis is a tensor input (already in inputs)
+      - fixed:          axis is hardcoded (no dim attribute)
+      - implicit_all:   no axis parameter, reduces all axes
+    """
+    blocks: list[str] = []
+    injected_names: set[str] = set()
+
+    if "Reduction" in paradigms and "Recurrence" not in paradigms:
+        if axis_source == "attribute":
+            _axis_input_names = {inp.name for inp in inputs
+                                 if inp.name in ("axes", "axis", "dim")
+                                 and any(d in ("int32", "int64") for d in inp.dtype_set)}
+            if not _axis_input_names:
+                target = inputs[0].name
+                blocks.append(
+                    "  - name: dim\n"
+                    "    type: list[int64]\n"
+                    "    default: []\n"
+                    '    semantics: "归约轴列表；为空时的语义由算子定义"\n'
+                    "    machine_constraint:\n"
+                    "      kind: list_of_int_in_range_relative_to_rank\n"
+                    f"      target: {target}\n"
+                    f'      lower_inclusive: "-rank({target})"\n'
+                    f'      upper_exclusive: "rank({target})"'
+                )
+                injected_names.add("dim")
+        if _is_pure_reduction(category, paradigms) and axis_source != "implicit_all":
+            blocks.append(
+                "  - name: keep_dims\n"
+                "    type: bool\n"
+                "    default: false\n"
+                '    semantics: "是否保留归约轴；保留时归约轴长度为 1"\n'
+                "    machine_constraint: {kind: bool}"
+            )
+            injected_names.add("keep_dims")
+
+    if "Quantization" in paradigms:
+        blocks.append(
+            "  - name: scale\n    type: float32\n    default: 1.0\n"
+            '    semantics: "量化缩放因子"\n'
+            "    machine_constraint: {kind: float_in_range}\n"
+            "  - name: zero_point\n    type: int64\n    default: 0\n"
+            '    semantics: "量化零点"\n'
+            "    machine_constraint: {kind: int_in_range}"
+        )
+        injected_names.update({"scale", "zero_point"})
+
+    if "RandomSampling" in paradigms:
+        blocks.append(
+            "  - name: seed\n    type: int64\n    default: 0\n"
+            '    semantics: "随机数种子；固定 seed 保证 bitwise_reproducible"\n'
+            "    machine_constraint: {kind: int_in_range}"
+        )
+        injected_names.add("seed")
+
+    explicit_attributes = explicit_attributes or []
+    explicit_names = [item.get("name") for item in explicit_attributes]
+    duplicates = {name for name in explicit_names if explicit_names.count(name) > 1}
+    duplicates |= injected_names & set(explicit_names)
+    if duplicates:
+        raise ValueError(f"attributes 与自动注入属性重名: {sorted(duplicates)}")
+    explicit_block = ""
+    if explicit_attributes:
+        dumped = yaml.safe_dump(explicit_attributes, allow_unicode=True, default_flow_style=False, sort_keys=False).rstrip()
+        explicit_block = "\n".join(f"  {line}" for line in dumped.splitlines())
+    return "\n".join(part for part in ("\n".join(blocks), explicit_block) if part) or "  []"
+
+
+def build_op_block(op_name: str, description: str, category: str,
+                   paradigms: list[str],
+                   supported_chips: list[str] | None = None,
+                   error_codes: list[str] | None = None) -> str:
+    """Build the `op:` field.
+
+    发射规则：
+      * supported_chips 非空 → 写入 platform_constraints.supported_chips
+      * CollectiveCommunication 范式 → 追加 requires_hccl: true
+      * error_codes 非空 → 写入 op.error_codes（语义类别）
+    """
+    parts = [
+        f"  name: {op_name}",
+        f'  version: "1.0"',
+        f'  description: "{description}"',
+        f"  category: {category}",
+        f"  paradigms: [{', '.join(paradigms)}]",
+    ]
+    if error_codes:
+        parts.append("  error_codes:")
+        for ec in error_codes:
+            parts.append(f"    - {ec}")
+    chips = supported_chips or []
+    needs_hccl = "CollectiveCommunication" in paradigms
+    if chips or needs_hccl:
+        parts.append("  platform_constraints:")
+        if chips:
+            parts.append(f"    supported_chips: [{', '.join(chips)}]")
+        if needs_hccl:
+            parts.append("    requires_hccl: true")
+    return "\n".join(parts)
+
+
+_ACCUMULATION_CATEGORIES = {
+    "Reduction", "ReductionComposite",
+    "Recurrence", "AtomicUpdate", "Histogram",
+}
+_INTEGER_DTYPES = {"int4", "int8", "int16", "int32", "int64",
+                   "uint1", "uint4", "uint8", "uint16", "uint32", "uint64", "bool"}
+
+
+def build_inputs_block(inputs: list[TensorSpec], paradigms: list[str], category: str) -> str:
+    lines = []
+    is_pure_reduction = _is_pure_reduction(category, paradigms)
+    is_accumulation = category in _ACCUMULATION_CATEGORIES
+    for i, inp in enumerate(inputs):
+        # 折叠维名直接用 input 全名，避免多 input 同首字母冲突（e.g. xa/xb 都生成 "...x"）
+        folded_name = inp.name.lower()
+        # Stateful: 第一个 input 自动设为 role: state（占位；用户可改）
+        role = _default_input_role(inp.role, paradigms, i)
+        lines.append(f"  - name: {inp.name}")
+        lines.append(f"    role: {role}")
+        lines.append(f"    dtype_set: [{', '.join(inp.dtype_set)}]")
+        rank_min = max(inp.rank_range[0], 1) if is_pure_reduction and i == 0 else inp.rank_range[0]
+        lines.append(f"    rank_range: [{rank_min}, {inp.rank_range[1]}]")
+        lines.append(f"    layout: ND")
+        # 累加类算子的 tensor 数据输入（非整型索引/轴）自动注入 data_distribution: normal
+        # 整型 tensor（axis/index）和非累加类不注入
+        is_integer_tensor = role == "tensor" and set(inp.dtype_set).issubset(_INTEGER_DTYPES)
+        if is_accumulation and role == "tensor" and not is_integer_tensor:
+            lines.append(f"    data_distribution: normal")
+        elif inp.data_distribution:
+            lines.append(f"    data_distribution: {inp.data_distribution}")
+        lines.append(f"    shape:")
+        if is_pure_reduction and i == 0:
+            lines.append(f'      symbolic: ["...{folded_name}", "R"]')
+        else:
+            lines.append(f'      symbolic: ["...{folded_name}"]')
+        if inp.optional:
+            lines.append("    optional: true")
+        if inp.list_length is not None:
+            _render_mapping(lines, "list_length", inp.list_length, "    ")
+    return "\n".join(lines)
+
+
+def build_outputs_block(outputs: list[str | OutputSpec], first_input: str, category: str, paradigms: list[str],
+                        inputs: list[TensorSpec], axis_source: str = "attribute") -> str:
+    """生成 outputs 段。
+
+    - VariableOutput（nonzero / unique）→ data_dependent
+    - Reduction / ReductionComposite / ArgReduce：
+      - axis_source=input_tensor → data_dependent（运行时值决定输出 shape）
+      - axis_source=attribute → numpy_expr + np.reduce_shape（attribute 值可静态求值）
+      - axis_source=fixed → numpy_expr + np.reduce_shape（固定 axis 值）
+      - axis_source=implicit_all → numpy_expr + scalar output（reduce 所有轴）
+    - 其他 → numpy_expr（占位，TODO 手填）
+    """
+    lines = []
+    is_variable_output = category == "VariableOutput"
+    is_reduction_like = category in ("Reduction", "ReductionComposite", "ArgReduce")
+    has_axis_input = any(
+        inp.name in ("axes", "axis", "dim")
+        and any(d in ("int32", "int64") for d in inp.dtype_set)
+        for inp in inputs
+    )
+    for output in _normalize_outputs(outputs):
+        out = output.name
+        lines.append(f"  - name: {out}")
+        if output.role != "tensor":
+            lines.append(f"    role: {output.role}")
+        if output.optional:
+            lines.append("    optional: true")
+        if output.list_length is not None:
+            _render_mapping(lines, "list_length", output.list_length, "    ")
+        if is_variable_output:
+            lines.append(f"    shape_rule_kind: data_dependent")
+            lines.append(f"    data_dependent_shape: true")
+            lines.append(f"    shape_rule_description: |")
+            lines.append(f"      # TODO: 描述 {out}.shape 与 {first_input} 实际值的关系")
+            lines.append(f"      # 例如：{out}.shape = (K, rank({first_input}))，K = count({first_input} != 0)")
+            lines.append(f"    shape_bounds:")
+            lines.append(f"      max_elements: \"prod({first_input}.shape)\"  # TODO 上界")
+            lines.append(f"      output_rank: 2  # TODO")
+            lines.append(f"    dtype_rule_kind: numpy_expr")
+            lines.append(f"    dtype_rule: |")
+            lines.append(f"      {out}.dtype = np.int64  # TODO: VariableOutput 通常输出索引")
+        elif is_reduction_like and (axis_source == "input_tensor"
+                                    or (axis_source == "attribute" and has_axis_input)):
+            lines.append(f"    shape_rule_kind: data_dependent")
+            lines.append(f"    data_dependent_shape: true")
+            lines.append(f"    shape_rule_description: |")
+            lines.append(f"      # keep_dims=true:  {out}.shape = {first_input}.shape 归约轴置 1")
+            lines.append(f"      # keep_dims=false: {out}.shape = {first_input}.shape 去除归约轴")
+            lines.append(f"    shape_bounds:")
+            lines.append(f"      max_elements: \"prod({first_input}.shape)\"")
+            lines.append(f"      output_rank: 8  # TODO: 替换为实际最大 rank")
+            lines.append(f"    dtype_rule_kind: numpy_expr")
+            lines.append(f"    dtype_rule: |")
+            lines.append(f"      {out}.dtype = {first_input}.dtype  # TODO")
+        elif _is_pure_reduction(category, paradigms) and axis_source == "fixed":
+            lines.append(f"    shape_rule_kind: numpy_expr")
+            lines.append(f"    shape_rule: |")
+            lines.append(
+                f"      {out}.shape = np.reduce_shape({first_input}.shape, axis=-1, keepdims=keep_dims)"
+            )
+            lines.append(f"    dtype_rule_kind: numpy_expr")
+            lines.append(f"    dtype_rule: |")
+            lines.append(f"      {out}.dtype = {first_input}.dtype  # TODO")
+        elif _is_pure_reduction(category, paradigms) and axis_source == "implicit_all":
+            lines.append(f"    shape_rule_kind: numpy_expr")
+            lines.append(f"    shape_rule: |")
+            lines.append(f"      {out}.shape = ()  # reduce all axes → scalar")
+            lines.append(f"    dtype_rule_kind: numpy_expr")
+            lines.append(f"    dtype_rule: |")
+            lines.append(f"      {out}.dtype = {first_input}.dtype  # TODO")
+        else:
+            lines.append(f"    shape_rule_kind: numpy_expr")
+            lines.append(f"    shape_rule: |")
+            if _is_pure_reduction(category, paradigms):
+                lines.append(
+                    f"      {out}.shape = np.reduce_shape({first_input}.shape, axis=tuple(dim), keepdims=keep_dims)"
+                )
+            else:
+                lines.append(f"      {out}.shape = {first_input}.shape  # TODO: numpy 表达式")
+            lines.append(f"    dtype_rule_kind: numpy_expr")
+            lines.append(f"    dtype_rule: |")
+            lines.append(f"      {out}.dtype = {first_input}.dtype  # TODO")
+        lines.append(f"    layout: ND")
+        lines.append(f"    aliasing: none")
+    return "\n".join(lines)
+
+
+def build_semantic_cases_block(cases: list[dict]) -> str:
+    if not cases:
+        return ""
+    dumped = yaml.safe_dump(cases, allow_unicode=True, default_flow_style=False, sort_keys=False).rstrip()
+    lines = ["semantic_cases:"]
+    lines.extend(f"  {line}" for line in dumped.splitlines())
+    return "\n".join(lines)
+
+
+def build_layout_contract_block(contract: dict | None) -> str:
+    """Render the optional top-level logical-layout contract."""
+    if not contract:
+        return ""
+    dumped = yaml.safe_dump(contract, allow_unicode=True, default_flow_style=False, sort_keys=False).rstrip()
+    lines = ["layout_contract:"]
+    # The root key was emitted above; insert only its mapping contents.
+    lines.extend(f"  {line}" for line in dumped.splitlines())
+    return "\n".join(lines)
+
+
+def build_shape_symbols_block(category: str, paradigms: list[str]) -> str:
+    if _is_pure_reduction(category, paradigms):
+        return "{R: {kind: dim, range: [0, 9223372036854775807]}}"
+    return "{}"
+
+
+def build_reduction_block(axis_source: str) -> str:
+    """Generate the top-level `reduction:` block for non-attribute axis sources.
+
+    Only emitted when axis_source is fixed or implicit_all.
+    For attribute/input_tensor modes, returns empty string.
+    """
+    if axis_source == "fixed":
+        return (
+            "reduction:\n"
+            "  axis_source: fixed\n"
+            "  fixed_value: -1  # 替换为实际固定归约轴；同步修改 shape_rule 中的 axis 值"
+        )
+    if axis_source == "implicit_all":
+        return (
+            "reduction:\n"
+            "  axis_source: implicit_all"
+        )
+    return ""
+
+
+def build_supported_combinations_block(inputs: list[TensorSpec],
+                                        outputs: list[str],
+                                        promotion: str,
+                                        axis_source: str = "attribute") -> str:
+    """生成 dtype_policy.supported_combinations。
+
+    策略：
+      * 同 dtype 直传 — 取所有输入 dtype_set 交集（每个 dtype 一行）
+      * promotion=numpy_promote：在交集行外加几行 promote 组合（取 fp16+fp32 / bf16+fp32 这种典型对）
+      * 其他 promotion：仅同 dtype 直传，开发者后续补充
+      * axis_source=input_tensor：axis 输入不参与 dtype 交集计算，单独用其 dtype
+    """
+    if not inputs or not outputs:
+        return "    # TODO: supported_combinations 至少 1 行\n    - {inputs: {x: float32}, outputs: {y: float32}}"
+
+    axis_input_names = set()
+    if axis_source == "input_tensor":
+        axis_input_names = {inp.name for inp in inputs
+                           if inp.name in ("axes", "axis", "dim")
+                           and any(d in ("int32", "int64") for d in inp.dtype_set)}
+
+    data_inputs = [inp for inp in inputs if inp.name not in axis_input_names]
+    if not data_inputs:
+        data_inputs = inputs
+
+    dtype_intersection = set(data_inputs[0].dtype_set)
+    for inp in data_inputs[1:]:
+        dtype_intersection &= set(inp.dtype_set)
+    if not dtype_intersection:
+        dtype_intersection = {data_inputs[0].dtype_set[0]}
+
+    lines = []
+    for dtype in sorted(dtype_intersection):
+        in_parts = []
+        for inp in inputs:
+            if inp.name in axis_input_names:
+                in_parts.append(f"{inp.name}: {inp.dtype_set[0]}")
+            else:
+                in_parts.append(f"{inp.name}: {dtype}")
+        in_part = ", ".join(in_parts)
+        out_part = ", ".join(f"{out}: {dtype}" for out in outputs)
+        lines.append(f"    - {{inputs: {{{in_part}}}, outputs: {{{out_part}}}}}")
+
+    if promotion == "numpy_promote" and len(data_inputs) >= 2:
+        mid = len(data_inputs) // 2 or 1
+        for narrow, wide in (("float16", "float32"), ("bfloat16", "float32")):
+            assigned = [narrow if i < mid else wide for i in range(len(data_inputs))]
+            if all(d in inp.dtype_set for d, inp in zip(assigned, data_inputs)) and narrow != wide:
+                in_parts = []
+                for inp in inputs:
+                    if inp.name in axis_input_names:
+                        in_parts.append(f"{inp.name}: {inp.dtype_set[0]}")
+                    else:
+                        idx = data_inputs.index(inp)
+                        in_parts.append(f"{inp.name}: {assigned[idx]}")
+                in_part = ", ".join(in_parts)
+                out_part = ", ".join(f"{out}: {wide}" for out in outputs)
+                lines.append(f"    - {{inputs: {{{in_part}}}, outputs: {{{out_part}}}}}")
+
+    return "\n".join(lines)
+
+
+# Per-dtype 容差默认值从 registries/tolerance_defaults.yaml 加载
+# 这样 generator 与 stage 7 校验器共享同一份数据，避免漂移
+def _format_tol_value(v) -> str:
+    """Format a numeric value for inline yaml output (preserve scientific notation)."""
+    if isinstance(v, (int, bool)) and not isinstance(v, bool) or v == 0:
+        return "0"
+    if isinstance(v, float):
+        if v == 0:
+            return "0"
+        return f"{v:.1e}"
+    return str(v)
+
+
+def build_per_dtype_tolerance_block(inputs: list[TensorSpec], outputs: list[str],
+                                     promotion: str,
+                                     axis_source: str = "attribute",
+                                     category: str = "",
+                                     op_name: str = "") -> str:
+    """生成 numerical_tolerance.per_dtype，覆盖 supported_combinations 中所有出现的 output dtype。"""
+    defaults = _load_tolerance_defaults()
+
+    reduction_categories = {"Reduction", "ReductionComposite"}
+    is_reduction = category in reduction_categories
+    is_exact_reduce = is_reduction and op_name.endswith(("max", "min"))
+
+    fp32_default = defaults.get("float32", {"rtol": 1.0e-5, "atol": 1.0e-5, "metric": "max_relative"})
+    
+    axis_input_names = set()
+    if axis_source == "input_tensor":
+        axis_input_names = {inp.name for inp in inputs
+                           if inp.name in ("axes", "axis", "dim")
+                           and any(d in ("int32", "int64") for d in inp.dtype_set)}
+
+    data_inputs = [inp for inp in inputs if inp.name not in axis_input_names]
+    if not data_inputs:
+        data_inputs = inputs
+    
+    dtype_intersection = set(data_inputs[0].dtype_set) if data_inputs else {"float32"}
+    for inp in (data_inputs[1:] if data_inputs else []):
+        dtype_intersection &= set(inp.dtype_set)
+    if not dtype_intersection:
+        dtype_intersection = {data_inputs[0].dtype_set[0]} if data_inputs else {"float32"}
+    out_dtypes = set(dtype_intersection)
+    if promotion == "numpy_promote" and len(data_inputs) >= 2:
+        for narrow, wide in (("float16", "float32"), ("bfloat16", "float32")):
+            mid = len(data_inputs) // 2 or 1
+            assigned = [narrow if i < mid else wide for i in range(len(data_inputs))]
+            if all(d in inp.dtype_set for d, inp in zip(assigned, data_inputs)):
+                out_dtypes.add(wide)
+
+    lines = []
+    for dtype in sorted(out_dtypes):
+        entry = defaults.get(dtype) or {"rtol": 1.0e-5, "atol": 1.0e-5, "metric": "max_relative"}
+        if is_reduction and not is_exact_reduce and entry["metric"] == "bitwise_equal":
+            entry = dict(fp32_default)
+        rtol = _format_tol_value(entry["rtol"])
+        atol = _format_tol_value(entry["atol"])
+        metric = entry["metric"]
+        lines.append(f"    {dtype}: {{rtol: {rtol}, atol: {atol}, metric: {metric}}}")
+    return "\n".join(lines)
+
+
+def build_accumulator_block(category: str, paradigms: list[str]) -> str:
+    triggers = {"Contraction", "ReductionComposite"}
+    if category in triggers or (
+        "NumericalStable" in paradigms and "Reduction" in paradigms
+    ):
+        return "  accumulator_dtype: float32"
+    return ""
+
+
+def build_composition_block(paradigms: list[str], inputs: list[TensorSpec],
+                            outputs: list[str]) -> str:
+    if "FusedComposite" not in paradigms:
+        return ""
+    in_names = ",".join(inp.name for inp in inputs)
+    out_name = outputs[0]
+    return f"""
+  composition:
+    primitives:
+      # TODO: 至少 2 条原语；op 必须在 PRIMITIVE_WHITELIST
+      - id: prim1
+        op: elementwise_unary
+        inputs: [{in_names}]
+        outputs: [tmp]
+      - id: prim2
+        op: elementwise_unary
+        inputs: [tmp]
+        outputs: [{out_name}]
+    dataflow:
+      intermediates: [tmp]
+      no_leak: true
+      fusable_groups:
+        - [prim1, prim2]"""
+
+
+# stage 6 要求 — 按 op.paradigms 注入对应必含 case 的占位
+# 关键词必须能命中 registries/boundary_min_cases.yaml 中对应 paradigm 的 match_any
+# 字段，否则生成的骨架会在 stage 6 直接 FAIL（违背"骨架可用"承诺）。
+
+_PARADIGM_BOUNDARY_TEMPLATES = {
+    "Reduction": [
+        '  - case: "reduce 轴长度为 1"\n    synthesize: { __FIRST_INPUT__.shape: "[1]" }\n    machine_check: {kind: matches_oracle}',
+        '  - case: "rank=0 标量输入"\n    synthesize: { __FIRST_INPUT__.shape: "[]" }\n    machine_check: {kind: matches_oracle}',
+        '  - case: "空 Tensor（A 轴=0，输出为空）"\n    synthesize: { __FIRST_INPUT__.shape: "[0, 4]" }\n    machine_check: {kind: returns_empty}',
+        '  - case: "reduce 轴 R=0（归约轴为空，输出=A 轴 shape）"\n    synthesize: { __FIRST_INPUT__.shape: "[4, 0]" }\n    machine_check: {kind: matches_oracle}',
+    ],
+    "SlidingWindow": [
+        '  - case: "stride > kernel"\n    synthesize: { __FIRST_INPUT__.shape: "[1, 1, 8, 8]" }\n    machine_check: {kind: matches_oracle}',
+    ],
+    "Padding": [
+        '  - case: "零 padding 退化"\n    synthesize: { __FIRST_INPUT__.shape: "[2, 3]" }\n    machine_check: {kind: matches_oracle}',
+        '  - case: "超大 padding"\n    synthesize: { __FIRST_INPUT__.shape: "[2, 3]" }\n    machine_check: {kind: matches_oracle}',
+    ],
+    "IndexGather": [
+        '  - case: "索引越界（负向）"\n    synthesize: { __FIRST_INPUT__.shape: "[4, 8]" }\n    machine_check: {kind: raises_error, error_type: attribute_value_out_of_range}',
+    ],
+    "ScatterUpdate": [
+        '  - case: "索引越界"\n    synthesize: { __FIRST_INPUT__.shape: "[4, 8]" }\n    machine_check: {kind: raises_error, error_type: attribute_value_out_of_range}',
+    ],
+    "AtomicUpdate": [
+        '  - case: "索引越界"\n    synthesize: { __FIRST_INPUT__.shape: "[4, 8]" }\n    machine_check: {kind: raises_error, error_type: attribute_value_out_of_range}',
+    ],
+    "SortSelect": [
+        '  - case: "k=0"\n    synthesize: { __FIRST_INPUT__.shape: "[8]" }\n    machine_check: {kind: matches_oracle}',
+        '  - case: "k=N"\n    synthesize: { __FIRST_INPUT__.shape: "[8]" }\n    machine_check: {kind: matches_oracle}',
+    ],
+    "Histogram": [
+        '  - case: "所有值落同一 bin"\n    synthesize: { __FIRST_INPUT__.shape: "[16]" }\n    machine_check: {kind: matches_oracle}',
+    ],
+    "Spectral": [
+        '  - case: "长度=2 幂次"\n    synthesize: { __FIRST_INPUT__.shape: "[16]" }\n    machine_check: {kind: matches_oracle}',
+        '  - case: "非 2 幂次长度"\n    synthesize: { __FIRST_INPUT__.shape: "[15]" }\n    machine_check: {kind: matches_oracle}',
+    ],
+    "Stateful": [
+        '  - case: "跨调用状态保持（first call vs second call）"\n    synthesize: { __FIRST_INPUT__.shape: "[8]" }\n    machine_check: {kind: matches_oracle}',
+    ],
+    "Broadcast": [
+        '  - case: "含 size-1 维的广播"\n    synthesize: { __FIRST_INPUT__.shape: "[1, 4]" }\n    machine_check: {kind: matches_oracle}',
+        '  - case: "广播不兼容 → 报错"\n    synthesize: { __FIRST_INPUT__.shape: "[2, 3]" }\n    machine_check: {kind: raises_error, error_type: shape_mismatch}',
+    ],
+}
+
+
+_PURE_REDUCTION_BOUNDARY_TEMPLATES = [
+    '  - case: "reduce 轴长度为 1"\n'
+    '    synthesize: { __FIRST_INPUT__.shape: "[1]", attr.dim: [-1], attr.keep_dims: false }\n'
+    '    machine_check: {kind: matches_oracle}',
+    '  - case: "空 Tensor（A 轴=0，输出为空）"\n'
+    '    synthesize: { __FIRST_INPUT__.shape: "[0, 4]", attr.dim: [-1], attr.keep_dims: false }\n'
+    '    machine_check: {kind: returns_empty}',
+    '  - case: "reduce 轴 R=0（归约轴为空，输出=A 轴 shape）"\n'
+    '    synthesize: { __FIRST_INPUT__.shape: "[4, 0]", attr.dim: [-1], attr.keep_dims: false }\n'
+    '    machine_check: {kind: matches_oracle}',
+]
+
+
+_PARADIGM_EXTREME_TEMPLATES = {
+    "NumericalStable": [
+        '  - case: "fp16 上溢边界"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: "all_same(60000.0)", target: __FIRST_INPUT__}\n      dtype: float16\n    machine_check: {kind: matches_oracle}',
+    ],
+    "IndexGather": [
+        '  - case: "全相同索引"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: all_zero, target: __FIRST_INPUT__}\n    machine_check: {kind: matches_oracle}',
+    ],
+    "AtomicUpdate": [
+        '  - case: "索引冲突"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: all_zero, target: __FIRST_INPUT__}\n    machine_check: {kind: matches_oracle}',
+    ],
+    "MaskPredicate": [
+        '  - case: "全 True mask"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: all_zero, target: __FIRST_INPUT__}\n    machine_check: {kind: matches_oracle}',
+        '  - case: "全 False mask"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: all_zero, target: __FIRST_INPUT__}\n    machine_check: {kind: matches_oracle}',
+    ],
+    "ArgReduce": [
+        '  - case: "tie / 等值"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: all_zero, target: __FIRST_INPUT__}\n    machine_check: {kind: matches_oracle}',
+    ],
+    "RandomSampling": [
+        '  - case: "deterministic 固定 seed"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: all_zero, target: __FIRST_INPUT__}\n    machine_check: {kind: matches_oracle}',
+    ],
+    "Quantization": [
+        '  - case: "全零输入（zero_point 行为）"\n    synthesize:\n      shapes: { __FIRST_INPUT__: "[8]" }\n      patterns:\n        - {pattern: all_zero, target: __FIRST_INPUT__}\n    machine_check: {kind: matches_oracle}',
+    ],
+}
+
+
+def build_default_boundary_case(category: str, paradigms: list[str], first_input: str,
+                                axis_source: str = "attribute") -> str:
+    if _is_pure_reduction(category, paradigms):
+        if axis_source == "fixed":
+            return (
+                '  - case: "rank=0 标量输入 → 归约轴越界"\n'
+                f'    synthesize: {{ {first_input}.shape: "[]", attr.keep_dims: false }}\n'
+                "    machine_check: {kind: raises_error, error_type: attribute_value_out_of_range}"
+            )
+        if axis_source == "implicit_all":
+            return ""
+        if axis_source == "input_tensor":
+            return (
+                '  - case: "rank=0 标量输入"\n'
+                f'    synthesize: {{ {first_input}.shape: "[]", attr.keep_dims: false }}\n'
+                "    machine_check: {kind: matches_oracle}"
+            )
+        return (
+            '  - case: "rank=0 标量输入 → 归约轴越界"\n'
+            f'    synthesize: {{ {first_input}.shape: "[]", attr.dim: [-1], attr.keep_dims: false }}\n'
+            "    machine_check: {kind: raises_error, error_type: attribute_value_out_of_range}"
+        )
+
+    return (
+        '  - case: "rank=0 标量输入"\n'
+        f'    synthesize: {{ {first_input}.shape: "[]" }}\n'
+        "    machine_check: {kind: matches_oracle}"
+    )
+
+
+def build_extra_boundary_cases(category: str, paradigms: list[str], first_input: str,
+                               axis_source: str = "attribute") -> str:
+    """生成 paradigm 必含的 boundary case 占位，避免 stage 6 FAIL。"""
+    lines: list[str] = []
+    for p in paradigms:
+        if p == "Reduction" and _is_pure_reduction(category, paradigms):
+            if axis_source == "implicit_all":
+                templates = _PARADIGM_BOUNDARY_TEMPLATES.get(p, [])
+            elif axis_source in ("fixed", "input_tensor"):
+                templates = [
+                    t.replace("attr.dim: [-1], ", "").replace(", attr.dim: [-1]", "")
+                    for t in _PURE_REDUCTION_BOUNDARY_TEMPLATES
+                ]
+            else:
+                templates = _PURE_REDUCTION_BOUNDARY_TEMPLATES
+        else:
+            templates = _PARADIGM_BOUNDARY_TEMPLATES.get(p, [])
+        for tmpl in templates:
+            lines.append(tmpl.replace("__FIRST_INPUT__", first_input))
+    return "\n".join(lines)
+
+
+def build_extra_extreme_cases(paradigms: list[str], first_input: str) -> str:
+    """生成 paradigm 必含的 extreme case 占位，避免 stage 6 FAIL。"""
+    lines: list[str] = []
+    for p in paradigms:
+        for tmpl in _PARADIGM_EXTREME_TEMPLATES.get(p, []):
+            lines.append(tmpl.replace("__FIRST_INPUT__", first_input))
+    return "\n".join(lines)
+
+
+def infer_error_codes(paradigms: list[str], attributes_block: str) -> list[str]:
+    """按 paradigm + 注入的 attributes 推断该算子的默认 error_codes 集合。
+
+    最小集：null_input（任何算子都可能传 nullptr）。其余按结构特征加入：
+      * Broadcast / Reduction / Contraction / IndexGather / LayoutTransform
+        → 输入 shape 之间存在硬约束，加 shape_mismatch
+      * attributes 段已注入（非 "[]"）→ 有可越界的属性，加 attribute_value_out_of_range
+    """
+    codes: list[str] = ["null_input"]
+    shape_paradigms = {
+        "Broadcast", "Reduction", "Contraction", "IndexGather",
+        "ScatterUpdate", "LayoutTransform", "Padding", "SortSelect",
+        "ArgReduce",
+    }
+    if any(p in shape_paradigms for p in paradigms):
+        codes.append("shape_mismatch")
+    if attributes_block.strip() and attributes_block.strip() != "[]":
+        codes.append("attribute_value_out_of_range")
+    return codes
+
+
+def build_raises_error_boundary_case(error_codes: list[str], first_input: str) -> str:
+    """挑一条最贴合 error_codes 的 raises_error 占位 case。
+
+    优先级：shape_mismatch > attribute_value_out_of_range > null_input。
+    """
+    if "shape_mismatch" in error_codes:
+        return (
+            '  - case: "输入 shape 不满足算子约束 → 报错（占位，请按算子改具体形状）"\n'
+            f'    synthesize: {{ {first_input}.shape: "[2, 3]" }}\n'
+            "    machine_check: {kind: raises_error, error_type: shape_mismatch}"
+        )
+    if "attribute_value_out_of_range" in error_codes:
+        return (
+            '  - case: "属性取值越界 → 报错（占位，请按算子改具体属性值）"\n'
+            f'    synthesize: {{ {first_input}.shape: "[2, 3]" }}\n'
+            "    machine_check: {kind: raises_error, error_type: attribute_value_out_of_range}"
+        )
+    return (
+        '  - case: "必填输入为 null → 报错（占位，请按算子改实际入参）"\n'
+        f'    synthesize: {{ {first_input}.shape: "[2, 3]" }}\n'
+        "    machine_check: {kind: raises_error, error_type: null_input}"
+    )
+
+
+def build_paradigm_groups_block(paradigms: list[str], mode: str) -> str:
+    """生成 paradigm_groups 段（横向组合 / 纵向融合）。
+
+    mode:
+      * "combination" — 多范式择一，由属性值决定激活哪个。
+        为每个 paradigm 生成一条 combination 组（TODO 待填 switch / when）。
+      * "fusion" — 多范式串联执行。
+        生成一条 fusion 组包含所有基础范式。
+      * "" — 不生成 paradigm_groups。
+    """
+    if not mode or not paradigms:
+        return ""
+    if mode == "combination":
+        lines = ["  paradigm_groups:"]
+        for p in paradigms:
+            lines.append(f"    - kind: combination")
+            lines.append(f"      paradigms: [{p}]")
+            lines.append(f"      switch: TODO  # 属性名（如 reduction）")
+            lines.append(f"      when: TODO    # 属性值（如 none），最后一个组用 \"*\" 表示兜底")
+        return "\n".join(lines)
+    if mode == "fusion":
+        return (
+            "  paradigm_groups:\n"
+            f"    - kind: fusion\n"
+            f"      paradigms: [{', '.join(paradigms)}]"
+        )
+    return ""
+
+
+def build_format_variants_block(variants: list[dict]) -> str:
+    """生成 format_variants 段。
+
+    每个 variant 包含 format / rank / reduction_axes / oracle_kwargs（可选）。
+    渲染为 YAML 缩进文本，插入 math_semantics 末尾。
+    """
+    if not variants:
+        return ""
+    lines = ["  format_variants:"]
+    for v in variants:
+        fmt = v.get("format", "NCHW")
+        rank = v.get("rank", [])
+        axes = v.get("reduction_axes", [])
+        lines.append(f"    - format: {fmt}")
+        lines.append(f"      rank: {rank}")
+        lines.append(f"      reduction_axes: {axes}")
+        kw = v.get("oracle_kwargs")
+        if kw:
+            lines.append(f"      oracle_kwargs: {kw}")
+    return "\n".join(lines)
+
+
+def render(gi: GenInput) -> str:
+    text = TEMPLATE.read_text(encoding="utf-8")
+    first_input = gi.inputs[0].name
+    output_names = _output_names(gi.outputs)
+    first_output = output_names[0]
+
+    techniques = "[]"
+    if gi.numerical_stability_required:
+        techniques = (
+            "\n    - name: max_subtraction\n"
+            '      applies_to: "TODO"\n'
+            '      rationale: "TODO"\n'
+            "      anti_pattern_id: AP-004"
+        )
+
+    # 范式驱动的 boundary/extreme 占位 case
+    # 让生成器产出的骨架自身就过 stage 6（不需要用户额外想关键词）
+    default_boundary_case = build_default_boundary_case(gi.category, gi.paradigms, first_input, gi.axis_source)
+    extra_boundary_cases = build_extra_boundary_cases(gi.category, gi.paradigms, first_input, gi.axis_source)
+    extra_extreme_cases = build_extra_extreme_cases(gi.paradigms, first_input)
+
+    # 推断错误码集合并补一条 raises_error 占位 case，让 stage 2 J 规则有可校验对象
+    attributes_block = build_attributes_block(
+        gi.category, gi.paradigms, gi.inputs, gi.axis_source, gi.attributes,
+    )
+    error_codes = infer_error_codes(gi.paradigms, attributes_block)
+    raises_error_case = build_raises_error_boundary_case(error_codes, first_input)
+    extra_boundary_cases = (extra_boundary_cases + "\n" + raises_error_case
+                            if extra_boundary_cases.strip()
+                            else raises_error_case)
+
+    replacements = {
+        "{{op_name}}": gi.op_name,
+        "{{description}}": gi.description,
+        "{{category}}": gi.category,
+        "{{paradigms_csv}}": ", ".join(gi.paradigms),
+        "{{op_block}}": build_op_block(gi.op_name, gi.description,
+                                        gi.category, gi.paradigms,
+                                        supported_chips=gi.supported_chips,
+                                        error_codes=error_codes),
+        "{{paradigm_groups_block}}": build_paradigm_groups_block(
+            gi.paradigms, gi.paradigm_groups_mode),
+        "{{inputs_block}}": build_inputs_block(gi.inputs, gi.paradigms, gi.category),
+        "{{attributes_block}}": attributes_block,
+        "{{outputs_block}}": build_outputs_block(gi.outputs, first_input, gi.category, gi.paradigms, gi.inputs, gi.axis_source),
+        "{{semantic_cases_block}}": build_semantic_cases_block(gi.semantic_cases),
+        "{{layout_contract_block}}": build_layout_contract_block(gi.layout_contract),
+        "{{reduction_block}}": build_reduction_block(gi.axis_source),
+        "{{shape_symbols}}": build_shape_symbols_block(gi.category, gi.paradigms),
+        "{{first_input}}": first_input,
+        "{{first_output}}": first_output,
+        "{{promotion}}": gi.promotion,
+        "{{broadcast_kind}}": gi.broadcast_kind,
+        "{{default_boundary_case}}": default_boundary_case,
+        "{{accumulation_order}}": gi.accumulation_order,
+        "{{numerical_stability_required}}": str(gi.numerical_stability_required).lower(),
+        "{{numerical_stability_techniques}}": techniques,
+        "{{accumulator_block}}": build_accumulator_block(gi.category, gi.paradigms),
+        "{{composition_block}}": build_composition_block(gi.paradigms, gi.inputs, output_names),
+        "{{format_variants_block}}": build_format_variants_block(gi.format_variants),
+        "{{supported_combinations_block}}": build_supported_combinations_block(
+            gi.inputs, output_names, gi.promotion, gi.axis_source),
+        "{{per_dtype_tolerance_block}}": build_per_dtype_tolerance_block(
+            gi.inputs, output_names, gi.promotion, gi.axis_source, gi.category, gi.op_name),
+        "{{extra_boundary_cases}}": extra_boundary_cases,
+        "{{extra_extreme_cases}}": extra_extreme_cases,
+    }
+    for k, v in replacements.items():
+        text = text.replace(k, v)
+    return text
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description="Generate spec.yaml skeleton.")
+    ap.add_argument("--op-name")
+    ap.add_argument("--category")
+    ap.add_argument("--paradigms", help="Comma-separated PascalCase, e.g. Reduction,NumericalStable")
+    ap.add_argument("--inputs", help="Comma-separated 'name:dtype1,dtype2'", default="")
+    ap.add_argument("--outputs", help="Comma-separated names", default="")
+    ap.add_argument("--description", default=None,
+                    help="算子一句话描述；省略时生成 'TODO: describe <op_name>'")
+    ap.add_argument("--promotion", default="same_as_first_input")
+    ap.add_argument("--broadcast-kind", dest="broadcast_kind", default="numpy")
+    ap.add_argument("--accumulation-order", dest="accumulation_order", default="none")
+    ap.add_argument("--chips", default="",
+                    help="Comma-separated chip ids（如 Ascend910B,Ascend950PR）；"
+                         "省略时按 inputs.dtype_set 自动收窄到全部兼容芯片")
+    ap.add_argument("--axis-source", dest="axis_source", default="attribute",
+                    choices=["attribute", "input_tensor", "fixed", "implicit_all"],
+                    help="How the reduction axis is specified (only for Reduction paradigms): "
+                         "attribute (dim attr), input_tensor (axis as tensor input), "
+                         "fixed (hardcoded axis), implicit_all (reduce all axes)")
+    ap.add_argument("--paradigm-groups", dest="paradigm_groups", default="",
+                    choices=["", "combination", "fusion"],
+                    help="范式组合模式: combination（横向组合，属性切换范式）"
+                         "或 fusion（纵向融合，范式串联执行）。"
+                         "combination 模式下 Elementwise 不会被过滤。")
+    ap.add_argument("--format-variants", dest="format_variants", default="",
+                    help="数据排布格式变体，格式: 'FORMAT:RANK:AXES;...' "
+                         "(如 'NCHW:4:0,2,3;NHWC:4:0,1,2;NCDHW:5:0,2,3,4')")
+    ap.add_argument("--output-dir", required=True,
+                    help="Directory to write spec.yaml into (created if absent)")
+    ap.add_argument("--force", action="store_true",
+                    help="Overwrite existing spec.yaml")
+    args = ap.parse_args()
+
+    reg = load_registries()
+    non_interactive = args.op_name and args.category and args.paradigms
+
+    if non_interactive:
+        paradigms = [p.strip() for p in args.paradigms.split(",") if p.strip()]
+        pg_mode = args.paradigm_groups or ""
+        # combination 模式下不过滤 Elementwise（Elementwise 是独立的范式分支）
+        if pg_mode != "combination":
+            paradigms = _filter_elementwise(paradigms)
+        category = _resolve_category(args.category, paradigms)
+        if args.inputs:
+            inputs = [parse_tensor_arg(s) for s in args.inputs.split(";") if s.strip()]
+            outputs = [s.strip() for s in args.outputs.split(",") if s.strip()] or ["y"]
+            attributes = []
+            semantic_cases = []
+        else:
+            inputs = [TensorSpec(name="x", dtype_set=["float32"])]
+            outputs = [s.strip() for s in args.outputs.split(",") if s.strip()] or ["y"]
+            attributes = []
+            semantic_cases = []
+
+        # supported_chips: CLI 显式给 → 直接用；否则按 dtype_set 自动收窄
+        declared_dtypes: set[str] = set()
+        for t in inputs:
+            declared_dtypes.update(t.dtype_set)
+        all_chip_ids = [c["id"] for c in reg["chips"]]
+        if args.chips:
+            supported_chips = [c.strip() for c in args.chips.split(",") if c.strip()]
+            unknown = [c for c in supported_chips if c not in all_chip_ids]
+            if unknown:
+                raise SystemExit(
+                    f"Unknown chip id(s): {unknown}. Allowed: {all_chip_ids}"
+                )
+        else:
+            supported_chips = narrow_chips_by_dtypes(reg["chips"], declared_dtypes)
+            if not supported_chips:
+                raise SystemExit(
+                    f"No chip in registry supports all of dtype_set "
+                    f"{sorted(declared_dtypes)}; pass --chips explicitly."
+                )
+
+        gi = GenInput(
+            op_name=args.op_name,
+            description=args.description or f"TODO: describe {args.op_name} operator",
+            category=category,
+            paradigms=paradigms,
+            inputs=inputs,
+            outputs=outputs,
+            promotion=args.promotion,
+            broadcast_kind=args.broadcast_kind,
+            accumulation_order=args.accumulation_order,
+            numerical_stability_required=("NumericalStable" in paradigms),
+            supported_chips=supported_chips,
+            axis_source=args.axis_source,
+            paradigm_groups_mode=pg_mode,
+            format_variants=parse_format_variants(args.format_variants, reg.get("layouts")),
+            attributes=attributes,
+            semantic_cases=semantic_cases,
+            layout_contract=None,
+        )
+        if args.axis_source == "attribute" and "Reduction" in paradigms:
+            print(
+                "⚠ WARNING: axis_source=attribute 将注入 dim/keep_dims attribute。"
+                "如果 REQUIREMENTS.md 中算子接口描述无归约轴相关参数，"
+                "请使用 --axis-source fixed 或 --axis-source implicit_all",
+                file=sys.stderr,
+            )
+        if gi.category not in reg["categories"]:
+            raise SystemExit(f"Unknown category: {gi.category}")
+        for p in gi.paradigms:
+            if p not in reg["paradigms"]:
+                raise SystemExit(f"Unknown paradigm: {p}")
+    else:
+        gi = interactive_collect(reg)
+
+    out_dir = Path(args.output_dir).resolve()
+    out_dir.mkdir(parents=True, exist_ok=True)
+    out_path = out_dir / "spec.yaml"
+    if out_path.exists() and not args.force:
+        raise SystemExit(f"{out_path} exists. Use --force to overwrite.")
+
+    out_path.write_text(render(gi), encoding="utf-8")
+    print(f"\n✓ Wrote {out_path}")
+    print(f"  category={gi.category}  paradigms={gi.paradigms}")
+    print("\nNext steps:")
+    print("  1. Fill in TODOs (formula / oracle.api / supported_combinations / boundary cases)")
+    print(f"  2. Validate: python3 {SKILL_DIR}/scripts/validate_spec.py {out_path}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())