perf(skills): reduce common-path context loading

scripts/gen_skill_shortcut_sections.py

@@ -1,13 +1,15 @@
 #!/usr/bin/env python3
-"""Generate shortcut sections for DWS skills.
+"""Generate shortcut discovery sections for DWS skills.
 
 The skill should teach agents which high-level shortcut entries are available
-in the public catalog and Runtime Schema. Leaf Schema publishes the Agent
-contract, while leaf `--help` remains the source of truth for accepted flags.
+without forcing large product catalogs into every common task. Leaf Schema
+publishes the Agent contract, while leaf `--help` remains the source of truth
+for accepted flags.
 """
 
 from __future__ import annotations
 
+import argparse
 import json
 import os
 import sys
@@ -47,6 +49,13 @@ MONO_END = "<!-- VISIBLE_SHORTCUTS_OVERVIEW_END -->"
 PRODUCT_START = "<!-- VISIBLE_SHORTCUTS_START -->"
 PRODUCT_END = "<!-- VISIBLE_SHORTCUTS_END -->"
 
+# Large, high-frequency product skills should route known intents directly and
+# keep their full shortcut inventory in Runtime Catalog/Schema. Add services
+# here only after verifying that the product skill has its own reviewed routing
+# section and intent table; compacting a sparse skill without an alternative
+# route would make its shortcuts harder to discover.
+COMPACT_PRODUCT_SERVICES = {"chat"}
+
 
 def md_escape(value: Any) -> str:
     text = str(value or "")
@@ -90,21 +99,24 @@ def mono_overview(items: list[dict[str, Any]]) -> str:
         path = SERVICE_TO_SKILL.get(service)
         skill = "—"
         if path:
-            skill = next((part for part in path.parts if part.startswith("dingtalk-")), path.parent.name)
-        rows.append(f"| `{md_escape(service)}` | {count} | `{md_escape(skill)}` | `dws shortcut list --service {md_escape(service)} --format json` |")
+            skill = next((part for part in reversed(path.parts) if part.startswith("dingtalk-")), path.parent.name)
+        rows.append(f"| `{md_escape(service)}` | {count} | `{md_escape(skill)}` |")
     body = "\n".join(rows)
     return f"""{MONO_START}
 ## Shortcut 总览
 
-下面统计当前公开 catalog 中的 shortcut。mono 模式不展开 200+ 行明细，避免 skill 过重；需要执行时先按产品路由，再用 `dws shortcut list --service <service> --format json` 读取参数、约束、风险和示例，最后用 `dws <service> +<shortcut> --help` 核对当前 Cobra flags。multi 模式的各产品 skill 会展开该产品的 shortcut 表。
+下面只统计当前公开 catalog 中的 shortcut，不展开完整明细。已知意图应先按产品 Skill、意图表或任务 reference 选择唯一命令；命令已选中时直接执行，只在参数或安全语义不确定时读取 leaf Schema，在当前 Cobra flags 不确定时读取 leaf Help。仅当现有路由和 reference 都无法定位低频能力时，才用 `dws shortcut list --service <service> --format json` 做最后回退；不要为已知高频意图加载完整产品 Catalog。
 
-| 服务 | shortcut 数 | multi skill | 发现命令 |
-|---|---:|---|---|
+| 服务 | shortcut 数 | multi skill |
+|---|---:|---|
 {body}
 {MONO_END}"""
 
 
 def product_section(service: str, rows: list[dict[str, Any]]) -> str:
+    if service in COMPACT_PRODUCT_SERVICES:
+        return compact_product_section(service, rows)
+
     table = []
     for item in rows:
         table.append(
@@ -114,7 +126,7 @@ def product_section(service: str, rows: list[dict[str, Any]]) -> str:
     return f"""{PRODUCT_START}
 ## Shortcuts（无专用脚本/recipe 时优先）
 
-以下 shortcut 同时进入公开 catalog 与 Runtime Schema。先按本 skill 的意图表、脚本和 recipe 路由：存在精确覆盖该场景的专用脚本/recipe 时按其执行；否则用户意图命中时，shortcut 优先于手写原子命令。用 leaf Schema（例如 `dws schema --cli-path "{service} +<shortcut>" --format json`）读取 Agent 选择、参数、约束、风险和确认语义；用 `dws shortcut list --service {service} --format json` 批量发现；最后以 `dws {service} <shortcut> --help` 核对当前 Cobra flags。
+以下 shortcut 同时进入公开 catalog 与 Runtime Schema。先按本 skill 的意图表、脚本和 recipe 路由：存在精确覆盖该场景的专用脚本/recipe 时按其执行；否则用户意图命中时，shortcut 优先于手写原子命令。命令已选中时直接执行；只在参数或安全语义不确定时读取 leaf Schema（例如 `dws schema --cli-path "{service} +<shortcut>" --format json`），在当前 Cobra flags 不确定时读取 `dws {service} <shortcut> --help`。仅当现有路由和 reference 都无法定位低频能力时，才用 `dws shortcut list --service {service} --format json` 批量发现。
 
 | Shortcut | 风险 | 适用场景 |
 |---|---|---|
@@ -122,17 +134,38 @@ def product_section(service: str, rows: list[dict[str, Any]]) -> str:
 {PRODUCT_END}"""
 
 
-def update_mono(items: list[dict[str, Any]]) -> None:
+def compact_product_section(service: str, rows: list[dict[str, Any]]) -> str:
+    return f"""{PRODUCT_START}
+## Shortcut 发现（按需）
+
+`{md_escape(service)}` 当前有 {len(rows)} 条公开 shortcut，完整清单保留在 Runtime Catalog 与 Schema，不在高频产品根 Skill 中重复展开。已知意图直接使用下方的优先路由、意图表或任务 reference；命令已选中时直接执行，只在参数/安全语义不确定时读取 leaf Schema，在当前 Cobra flags 不确定时读取 leaf Help。
+
+仅当现有路由和 reference 都无法定位低频能力时，才执行 `dws shortcut list --service {md_escape(service)} --format json` 做最后回退；不要为已知高频意图加载完整 Shortcut Catalog 或产品级 Schema。
+{PRODUCT_END}"""
+
+
+def apply_update(path: Path, text: str, updated: str, check: bool) -> bool:
+    if updated == text:
+        return False
+    if check:
+        print(f"generated skill drift: {path.relative_to(ROOT)}", file=sys.stderr)
+    else:
+        path.write_text(updated, encoding="utf-8")
+    return True
+
+
+def update_mono(items: list[dict[str, Any]], check: bool) -> list[Path]:
     text = MONO_SKILL.read_text(encoding="utf-8")
     block = mono_overview(items)
     updated = replace_block(text, MONO_START, MONO_END, block, "## 产品总览")
-    MONO_SKILL.write_text(updated, encoding="utf-8")
+    return [MONO_SKILL] if apply_update(MONO_SKILL, text, updated, check) else []
 
 
-def update_product_skills(items: list[dict[str, Any]]) -> None:
+def update_product_skills(items: list[dict[str, Any]], check: bool) -> list[Path]:
     by_service: dict[str, list[dict[str, Any]]] = defaultdict(list)
     for item in items:
         by_service[item["service"]].append(item)
+    changed = []
     for service, path in SERVICE_TO_SKILL.items():
         if service not in by_service:
             continue
@@ -142,13 +175,26 @@ def update_product_skills(items: list[dict[str, Any]]) -> None:
         block = product_section(service, by_service[service])
         anchor = "## 概念地图" if service == "devapp" else "## 意图表"
         updated = replace_block(text, PRODUCT_START, PRODUCT_END, block, anchor)
-        path.write_text(updated, encoding="utf-8")
+        if apply_update(path, text, updated, check):
+            changed.append(path)
+    return changed
 
 
 def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--check",
+        action="store_true",
+        help="verify generated skill sections are current without rewriting files",
+    )
+    args = parser.parse_args()
+
     items = collect_visible()
-    update_mono(items)
-    update_product_skills(items)
+    changed = update_mono(items, args.check)
+    changed.extend(update_product_skills(items, args.check))
+    if args.check and changed:
+        print("run: python3 scripts/gen_skill_shortcut_sections.py", file=sys.stderr)
+        return 1
     print(f"visible_shortcuts={len(items)} services={len(set(item['service'] for item in items))}")
     return 0
 

scripts/policy/check-skill-context-budget.sh

@@ -0,0 +1,41 @@
+#!/bin/sh
+set -eu
+
+ROOT="$(CDPATH= cd -- "$(dirname -- "$0")/../.." && pwd)"
+cd "$ROOT"
+
+python3 scripts/gen_skill_shortcut_sections.py --check
+
+chat_skill="skills/multi/dingtalk-chat/SKILL.md"
+mono_skill="skills/mono/SKILL.md"
+chat_max_bytes=14000
+
+chat_bytes="$(wc -c < "$chat_skill" | tr -d ' ')"
+if [ "$chat_bytes" -gt "$chat_max_bytes" ]; then
+	printf '%s\n' \
+		"skill context budget exceeded: $chat_skill is ${chat_bytes} bytes (max ${chat_max_bytes})" >&2
+	exit 1
+fi
+
+shortcut_rows="$(
+	awk '
+		/<!-- VISIBLE_SHORTCUTS_START -->/ { in_block = 1; next }
+		/<!-- VISIBLE_SHORTCUTS_END -->/ { in_block = 0 }
+		in_block && /^\| `dws chat \+/ { count++ }
+		END { print count + 0 }
+	' "$chat_skill"
+)"
+if [ "$shortcut_rows" -ne 0 ]; then
+	printf '%s\n' \
+		"skill context budget exceeded: $chat_skill re-expanded $shortcut_rows shortcut rows" >&2
+	exit 1
+fi
+
+if grep -Fq "充分阅读产品参考文件" "$mono_skill"; then
+	printf '%s\n' \
+		"skill context budget regression: $mono_skill requires full product-reference loading" >&2
+	exit 1
+fi
+
+printf '%s\n' \
+	"skill context budget: ok (chat_bytes=$chat_bytes max=$chat_max_bytes shortcut_rows=$shortcut_rows)"