diff --git a/README.md b/README.md index 28ed1a7..9cd58a9 100644 --- a/README.md +++ b/README.md @@ -57,6 +57,17 @@ And one thing at the end that is meant to be *withheld*: **"done"** has a defini clean log is not one of them. Anything short of all six is a checkpoint, and should be reported as a checkpoint with what remains. +**What the split costs is not measured, and the claim should not be overstated.** `SKILL.md` loads in +full; a reference loads when the symptom index or `references/routing.md` sends you to one. That is +usually described as "cheaper context", and only half of it has been measured here: the always-loaded +body has a measured size and a gate that fails when it grows (`check_budget.py`). The other half has +not — how many references a real question pulls in, and whether fetching them one at a time ends up +costing more than the alternative would have. A question that needs four references is not obviously +cheaper than one longer file would have been, and every reference also carries its own entry in the +index that has to be read to find it. Treat the split's economy as `unknown` -- nothing here has +measured it -- and treat the body budget as `observed`. The second is the one that compounds on every +task, which is why it is the one gated. + If you are an agent reading this: the cheapest possible first command is `python skills/apk-reverse/scripts/doctor.py`. It tells you which of these tools exist here, which scripts can actually run, and whether something in the environment is already poisoning your diff --git a/README.zh-CN.md b/README.zh-CN.md index e0ed8ad..e9291f7 100644 --- a/README.zh-CN.md +++ b/README.zh-CN.md @@ -41,6 +41,8 @@ 另外还有一样在结尾、应当被**扣住**的东西:**「done」**有定义(六项)。一份干净的日志不在其中。少于六项全部达成,就只是一个检查点,应当作为检查点汇报,并说明还剩什么。 +**这套切分的成本没有被测量过,所以不要把它说大了。** `SKILL.md` 全文加载;某个 reference 则是在症状索引或 `references/routing.md` 把你送过去时才加载。这通常被描述为「更省上下文」,但这里只测了其中一半:常驻加载的那部分有实测体积,也有门禁在它膨胀时失败(`check_budget.py`)。另一半没有测 —— 一个真实问题会牵出几个 reference,以及一次取一个是否反而比另一种做法更贵。一个要读四个 reference 才能回答的问题,并不显然比一个更长的文件更便宜,而且每个 reference 还要连带为找到它所读的索引行。把切分带来的经济性当作 `unknown`(这里没有任何东西测量过它),把主体预算当作 `observed`。后者才是每个任务都会累积的那一半,也正是被门禁管住的那一半。 + 如果你是正在读这段的 agent:成本最低的第一条命令是 `python skills/apk-reverse/scripts/doctor.py`。它会告诉你这里存在哪些工具、哪些脚本真的能跑,以及环境里是否已经有什么东西在污染你的测量结果。 ## 它擅长什么