diff --git a/agentbench/00/arithmetic/add-mult.answer b/agentbench/00/arithmetic/add-mult.answer
deleted file mode 100644
index 4099407..0000000
--- a/agentbench/00/arithmetic/add-mult.answer
+++ /dev/null
@@ -1 +0,0 @@
-23
diff --git a/agentbench/00/arithmetic/add-mult.question b/agentbench/00/arithmetic/add-mult.question
deleted file mode 100644
index b45851f..0000000
--- a/agentbench/00/arithmetic/add-mult.question
+++ /dev/null
@@ -1 +0,0 @@
-计算 3 + 4 × 5，把结果打印出来。
diff --git a/agentbench/00/arithmetic/int-div.answer b/agentbench/00/arithmetic/int-div.answer
deleted file mode 100644
index 8351c19..0000000
--- a/agentbench/00/arithmetic/int-div.answer
+++ /dev/null
@@ -1 +0,0 @@
-14
diff --git a/agentbench/00/arithmetic/int-div.question b/agentbench/00/arithmetic/int-div.question
deleted file mode 100644
index 9492762..0000000
--- a/agentbench/00/arithmetic/int-div.question
+++ /dev/null
@@ -1 +0,0 @@
-计算 100 ÷ 7 的整数商（整除结果），并打印。
diff --git a/agentbench/00/string/concat.answer b/agentbench/00/string/concat.answer
deleted file mode 100644
index 3b18e51..0000000
--- a/agentbench/00/string/concat.answer
+++ /dev/null
@@ -1 +0,0 @@
-hello world
diff --git a/agentbench/00/string/concat.question b/agentbench/00/string/concat.question
deleted file mode 100644
index 67f7527..0000000
--- a/agentbench/00/string/concat.question
+++ /dev/null
@@ -1 +0,0 @@
-把字符串 "hello" 和 " world" 拼接成一个字符串并打印。
diff --git a/agentbench/01/arithmetic/sum-1-100.answer b/agentbench/01/arithmetic/sum-1-100.answer
deleted file mode 100644
index c3ac783..0000000
--- a/agentbench/01/arithmetic/sum-1-100.answer
+++ /dev/null
@@ -1 +0,0 @@
-5050
diff --git a/agentbench/01/arithmetic/sum-1-100.question b/agentbench/01/arithmetic/sum-1-100.question
deleted file mode 100644
index c8d49f3..0000000
--- a/agentbench/01/arithmetic/sum-1-100.question
+++ /dev/null
@@ -1 +0,0 @@
-求 1 到 100 之间所有整数的和（含 1 和 100），打印结果。
diff --git a/agentbench/01/string/reverse.answer b/agentbench/01/string/reverse.answer
deleted file mode 100644
index 59d2d1c..0000000
--- a/agentbench/01/string/reverse.answer
+++ /dev/null
@@ -1 +0,0 @@
-cba
diff --git a/agentbench/01/string/reverse.question b/agentbench/01/string/reverse.question
deleted file mode 100644
index f428dc9..0000000
--- a/agentbench/01/string/reverse.question
+++ /dev/null
@@ -1 +0,0 @@
-把字符串 abc 反转后打印。
diff --git a/agentbench/01/tool-shell/echo.answer b/agentbench/01/tool-shell/echo.answer
deleted file mode 100644
index 6fc6dca..0000000
--- a/agentbench/01/tool-shell/echo.answer
+++ /dev/null
@@ -1 +0,0 @@
-byteseek
diff --git a/agentbench/01/tool-shell/echo.question b/agentbench/01/tool-shell/echo.question
deleted file mode 100644
index bfed943..0000000
--- a/agentbench/01/tool-shell/echo.question
+++ /dev/null
@@ -1 +0,0 @@
-用 shell 打印字符串 byteseek。
diff --git a/agentbench/02/array/max.answer b/agentbench/02/array/max.answer
deleted file mode 100644
index ec63514..0000000
--- a/agentbench/02/array/max.answer
+++ /dev/null
@@ -1 +0,0 @@
-9
diff --git a/agentbench/02/array/max.question b/agentbench/02/array/max.question
deleted file mode 100644
index ef06cdc..0000000
--- a/agentbench/02/array/max.question
+++ /dev/null
@@ -1 +0,0 @@
-给定数组 [7, 2, 9, 4]，找出其中的最大值并打印。
diff --git a/agentbench/02/tool-python/pow.answer b/agentbench/02/tool-python/pow.answer
deleted file mode 100644
index d7b1c44..0000000
--- a/agentbench/02/tool-python/pow.answer
+++ /dev/null
@@ -1 +0,0 @@
-1024
diff --git a/agentbench/02/tool-python/pow.question b/agentbench/02/tool-python/pow.question
deleted file mode 100644
index 19a82ca..0000000
--- a/agentbench/02/tool-python/pow.question
+++ /dev/null
@@ -1 +0,0 @@
-用 python 计算 2 的 10 次方并打印结果。
diff --git a/agentbench/02/tool-shell/kernel.answer b/agentbench/02/tool-shell/kernel.answer
deleted file mode 100644
index 3f255c9..0000000
--- a/agentbench/02/tool-shell/kernel.answer
+++ /dev/null
@@ -1 +0,0 @@
-期望(冒烟): 输出等于 `uname -s`（本平台为 Linux）
diff --git a/agentbench/02/tool-shell/kernel.question b/agentbench/02/tool-shell/kernel.question
deleted file mode 100644
index e009a9d..0000000
--- a/agentbench/02/tool-shell/kernel.question
+++ /dev/null
@@ -1 +0,0 @@
-用 shell 打印当前操作系统的内核名称（uname）。
diff --git a/agentbench/03/array/sum.answer b/agentbench/03/array/sum.answer
deleted file mode 100644
index 60d3b2f..0000000
--- a/agentbench/03/array/sum.answer
+++ /dev/null
@@ -1 +0,0 @@
-15
diff --git a/agentbench/03/array/sum.question b/agentbench/03/array/sum.question
deleted file mode 100644
index 86d925a..0000000
--- a/agentbench/03/array/sum.question
+++ /dev/null
@@ -1 +0,0 @@
-给定数组 [1, 2, 3, 4, 5]，求所有元素之和并打印。
diff --git a/agentbench/03/string/char-count.answer b/agentbench/03/string/char-count.answer
deleted file mode 100644
index 00750ed..0000000
--- a/agentbench/03/string/char-count.answer
+++ /dev/null
@@ -1 +0,0 @@
-3
diff --git a/agentbench/03/string/char-count.question b/agentbench/03/string/char-count.question
deleted file mode 100644
index da5a401..0000000
--- a/agentbench/03/string/char-count.question
+++ /dev/null
@@ -1 +0,0 @@
-统计字符串 banana 中字符 a 出现的次数并打印。
diff --git a/agentbench/03/tool-http/fetch-example.answer b/agentbench/03/tool-http/fetch-example.answer
deleted file mode 100644
index 82d7471..0000000
--- a/agentbench/03/tool-http/fetch-example.answer
+++ /dev/null
@@ -1 +0,0 @@
-期望(冒烟): 输出 yes（响应正文含 "Example Domain"）
diff --git a/agentbench/03/tool-http/fetch-example.question b/agentbench/03/tool-http/fetch-example.question
deleted file mode 100644
index 4c68318..0000000
--- a/agentbench/03/tool-http/fetch-example.question
+++ /dev/null
@@ -1 +0,0 @@
-用 http 抓取 https://example.com 的正文，判断其中是否包含 "Example Domain"，打印 yes 或 no。
diff --git a/agentbench/04/array/second-max.answer b/agentbench/04/array/second-max.answer
deleted file mode 100644
index 1e8b314..0000000
--- a/agentbench/04/array/second-max.answer
+++ /dev/null
@@ -1 +0,0 @@
-6
diff --git a/agentbench/04/array/second-max.question b/agentbench/04/array/second-max.question
deleted file mode 100644
index 82382ce..0000000
--- a/agentbench/04/array/second-max.question
+++ /dev/null
@@ -1 +0,0 @@
-给定数组 [3, 1, 4, 1, 5, 9, 2, 6]，找出第二大（严格小于最大值的最大元素）并打印。
diff --git a/agentbench/04/dict/read-member.answer b/agentbench/04/dict/read-member.answer
deleted file mode 100644
index 671a444..0000000
--- a/agentbench/04/dict/read-member.answer
+++ /dev/null
@@ -1 +0,0 @@
-kv
diff --git a/agentbench/04/dict/read-member.question b/agentbench/04/dict/read-member.question
deleted file mode 100644
index f683da1..0000000
--- a/agentbench/04/dict/read-member.question
+++ /dev/null
@@ -1 +0,0 @@
-创建一个成员为 name="kv"、ver=1 的结构，打印其 name 成员的值。
diff --git a/agentbench/04/string/palindrome.answer b/agentbench/04/string/palindrome.answer
deleted file mode 100644
index 7cfab5b..0000000
--- a/agentbench/04/string/palindrome.answer
+++ /dev/null
@@ -1 +0,0 @@
-yes
diff --git a/agentbench/04/string/palindrome.question b/agentbench/04/string/palindrome.question
deleted file mode 100644
index 76b3701..0000000
--- a/agentbench/04/string/palindrome.question
+++ /dev/null
@@ -1 +0,0 @@
-判断字符串 racecar 是否为回文（正反读相同），是打印 yes，否打印 no。
diff --git a/agentbench/05/dict/linked-list.answer b/agentbench/05/dict/linked-list.answer
deleted file mode 100644
index 00750ed..0000000
--- a/agentbench/05/dict/linked-list.answer
+++ /dev/null
@@ -1 +0,0 @@
-3
diff --git a/agentbench/05/dict/linked-list.question b/agentbench/05/dict/linked-list.question
deleted file mode 100644
index 7ec3ca6..0000000
--- a/agentbench/05/dict/linked-list.question
+++ /dev/null
@@ -1 +0,0 @@
-用路径指针构造两节点链表：/n1 的 val=1、next="/n2"，/n2 的 val=2、next=""。从 /n1 遍历到末尾，打印沿途所有 val 之和。
diff --git a/agentbench/05/kvspace/persist.answer b/agentbench/05/kvspace/persist.answer
deleted file mode 100644
index d81cc07..0000000
--- a/agentbench/05/kvspace/persist.answer
+++ /dev/null
@@ -1 +0,0 @@
-42
diff --git a/agentbench/05/kvspace/persist.question b/agentbench/05/kvspace/persist.question
deleted file mode 100644
index 0756b19..0000000
--- a/agentbench/05/kvspace/persist.question
+++ /dev/null
@@ -1 +0,0 @@
-把值 42 写入绝对路径 /bench/answer，再从同一路径读回并打印。
diff --git a/agentbench/05/reasoning/cats.answer b/agentbench/05/reasoning/cats.answer
deleted file mode 100644
index 00750ed..0000000
--- a/agentbench/05/reasoning/cats.answer
+++ /dev/null
@@ -1 +0,0 @@
-3
diff --git a/agentbench/05/reasoning/cats.question b/agentbench/05/reasoning/cats.question
deleted file mode 100644
index 30c0d5e..0000000
--- a/agentbench/05/reasoning/cats.question
+++ /dev/null
@@ -1 +0,0 @@
-四只猫 A/B/C/D 体重分别为 3/5/7/9 公斤（各不同）。已知：A 不是最轻的，B 比 C 重，D 是 9 公斤。问 C 多重（只打印数字）。
diff --git a/agentbench/06/array/binary-search.answer b/agentbench/06/array/binary-search.answer
deleted file mode 100644
index 00750ed..0000000
--- a/agentbench/06/array/binary-search.answer
+++ /dev/null
@@ -1 +0,0 @@
-3
diff --git a/agentbench/06/array/binary-search.question b/agentbench/06/array/binary-search.question
deleted file mode 100644
index da43b5a..0000000
--- a/agentbench/06/array/binary-search.question
+++ /dev/null
@@ -1 +0,0 @@
-在有序数组 [1, 3, 5, 7, 9, 11, 13] 中查找 7 的下标（0 起），打印该下标。
diff --git a/agentbench/06/kvspace/multi-write.answer b/agentbench/06/kvspace/multi-write.answer
deleted file mode 100644
index 1e8b314..0000000
--- a/agentbench/06/kvspace/multi-write.answer
+++ /dev/null
@@ -1 +0,0 @@
-6
diff --git a/agentbench/06/kvspace/multi-write.question b/agentbench/06/kvspace/multi-write.question
deleted file mode 100644
index fde13c7..0000000
--- a/agentbench/06/kvspace/multi-write.question
+++ /dev/null
@@ -1 +0,0 @@
-把三个键值对 a=1、b=2、c=3 写入 /bench/ 下，然后读回三个值，打印它们的和。
diff --git a/agentbench/06/planning/pipe-multitool.answer b/agentbench/06/planning/pipe-multitool.answer
deleted file mode 100644
index 59d2d1c..0000000
--- a/agentbench/06/planning/pipe-multitool.answer
+++ /dev/null
@@ -1 +0,0 @@
-cba
diff --git a/agentbench/06/planning/pipe-multitool.question b/agentbench/06/planning/pipe-multitool.question
deleted file mode 100644
index 5be2efe..0000000
--- a/agentbench/06/planning/pipe-multitool.question
+++ /dev/null
@@ -1 +0,0 @@
-先用 shell 打印字符串 abc，再用 python 把它反转，打印最终结果。要求经过 shell 与 python 两个工具。
diff --git a/agentbench/07/kvspace/nested-tree.answer b/agentbench/07/kvspace/nested-tree.answer
deleted file mode 100644
index 190a180..0000000
--- a/agentbench/07/kvspace/nested-tree.answer
+++ /dev/null
@@ -1 +0,0 @@
-123
diff --git a/agentbench/07/kvspace/nested-tree.question b/agentbench/07/kvspace/nested-tree.question
deleted file mode 100644
index 4b19d15..0000000
--- a/agentbench/07/kvspace/nested-tree.question
+++ /dev/null
@@ -1 +0,0 @@
-在 /bench/ 下写入嵌套结构 /bench/a/b/c=123，然后读回最深层 c 的值并打印。
diff --git a/agentbench/07/planning/fetch-parse.answer b/agentbench/07/planning/fetch-parse.answer
deleted file mode 100644
index ae90161..0000000
--- a/agentbench/07/planning/fetch-parse.answer
+++ /dev/null
@@ -1 +0,0 @@
-期望(冒烟): 输出 Sample Slide Show
diff --git a/agentbench/07/planning/fetch-parse.question b/agentbench/07/planning/fetch-parse.question
deleted file mode 100644
index 023c7cc..0000000
--- a/agentbench/07/planning/fetch-parse.question
+++ /dev/null
@@ -1 +0,0 @@
-用 http 抓取 https://httpbin.org/json 的正文，用 json 解析出顶层 slideshow.title 字段并打印。
diff --git a/agentbench/07/reasoning/cities.answer b/agentbench/07/reasoning/cities.answer
deleted file mode 100644
index 8bf3bb7..0000000
--- a/agentbench/07/reasoning/cities.answer
+++ /dev/null
@@ -1 +0,0 @@
-深圳
diff --git a/agentbench/07/reasoning/cities.question b/agentbench/07/reasoning/cities.question
deleted file mode 100644
index 9f99a60..0000000
--- a/agentbench/07/reasoning/cities.question
+++ /dev/null
@@ -1 +0,0 @@
-五个学生分别来自北京/上海/广州/深圳/杭州（各不同）。已知：A 不是北京也不是深圳，B 是广州，C 不是深圳也不是杭州，D 是北京。问 E 来自哪里（只打印城市名）。
diff --git a/agentbench/08/meta/self-gen-llm.answer b/agentbench/08/meta/self-gen-llm.answer
deleted file mode 100644
index afe4629..0000000
--- a/agentbench/08/meta/self-gen-llm.answer
+++ /dev/null
@@ -1 +0,0 @@
-期望(冒烟): 输出 42（需 DEEPSEEK_API_KEY）
diff --git a/agentbench/08/meta/self-gen-llm.question b/agentbench/08/meta/self-gen-llm.question
deleted file mode 100644
index af286c5..0000000
--- a/agentbench/08/meta/self-gen-llm.question
+++ /dev/null
@@ -1 +0,0 @@
-调用 llm.call 生成一段计算 6 × 7 并打印的程序，再执行它，打印最终结果。
diff --git a/agentbench/08/meta/self-inspect.answer b/agentbench/08/meta/self-inspect.answer
deleted file mode 100644
index a70062b..0000000
--- a/agentbench/08/meta/self-inspect.answer
+++ /dev/null
@@ -1 +0,0 @@
-期望(冒烟): 输出 yes（需完整 bootstrap，kvlangbrief 已播种）
diff --git a/agentbench/08/meta/self-inspect.question b/agentbench/08/meta/self-inspect.question
deleted file mode 100644
index 991adfc..0000000
--- a/agentbench/08/meta/self-inspect.question
+++ /dev/null
@@ -1 +0,0 @@
-读回 /lib/byteseek.kvlangbrief 的内容，判断其长度是否大于 100，打印 yes 或 no。
diff --git a/agentbench/08/planning/error-recovery.answer b/agentbench/08/planning/error-recovery.answer
deleted file mode 100644
index df881dd..0000000
--- a/agentbench/08/planning/error-recovery.answer
+++ /dev/null
@@ -1 +0,0 @@
-recovered
diff --git a/agentbench/08/planning/error-recovery.question b/agentbench/08/planning/error-recovery.question
deleted file mode 100644
index 3c35bd7..0000000
--- a/agentbench/08/planning/error-recovery.question
+++ /dev/null
@@ -1 +0,0 @@
-尝试用 shell 执行一个必然失败的命令 nonexistent_cmd_xyz，不因失败中断，最终打印 recovered。
diff --git a/agentbench/09/meta/memory-cross.answer b/agentbench/09/meta/memory-cross.answer
deleted file mode 100644
index 3042f8d..0000000
--- a/agentbench/09/meta/memory-cross.answer
+++ /dev/null
@@ -1 +0,0 @@
-最喜欢的颜色是蓝色
diff --git a/agentbench/09/meta/memory-cross.question b/agentbench/09/meta/memory-cross.question
deleted file mode 100644
index fc27465..0000000
--- a/agentbench/09/meta/memory-cross.question
+++ /dev/null
@@ -1 +0,0 @@
-把一段记忆「最喜欢的颜色是蓝色」写入 /mem/fav（模拟持久化记忆），再从 /mem/fav 读回并打印。
diff --git a/agentbench/09/planning/primes.answer b/agentbench/09/planning/primes.answer
deleted file mode 100644
index 3f204e1..0000000
--- a/agentbench/09/planning/primes.answer
+++ /dev/null
@@ -1 +0,0 @@
-2 3 5 7 11 13 17 19 23 29
diff --git a/agentbench/09/planning/primes.question b/agentbench/09/planning/primes.question
deleted file mode 100644
index 81a1ff3..0000000
--- a/agentbench/09/planning/primes.question
+++ /dev/null
@@ -1 +0,0 @@
-不借助网络，用 shell/python 组合求出前 10 个质数，按从小到大打印（空格分隔）。
diff --git a/agentbench/09/reasoning/mislabeled-boxes.answer b/agentbench/09/reasoning/mislabeled-boxes.answer
deleted file mode 100644
index 1ad95a3..0000000
--- a/agentbench/09/reasoning/mislabeled-boxes.answer
+++ /dev/null
@@ -1 +0,0 @@
-混装
diff --git a/agentbench/09/reasoning/mislabeled-boxes.question b/agentbench/09/reasoning/mislabeled-boxes.question
deleted file mode 100644
index 1e729b8..0000000
--- a/agentbench/09/reasoning/mislabeled-boxes.question
+++ /dev/null
@@ -1 +0,0 @@
-三个盒子分别装苹果、橘子、混装（苹果+橘子），但三个标签全贴错了。你只能从一个盒子里摸出一个水果（不能看里面）来判断三个盒子的真实内容。问应该摸贴哪个标签的盒子（打印：苹果 / 橘子 / 混装）。
diff --git a/agentbench/10/meta/open-goal.answer b/agentbench/10/meta/open-goal.answer
deleted file mode 100644
index 4c339d8..0000000
--- a/agentbench/10/meta/open-goal.answer
+++ /dev/null
@@ -1 +0,0 @@
-评分标准: 能力定义清晰；有可运行 kvlang 代码；输出可复现验证；改动落在 kvspace 树内可观测。
diff --git a/agentbench/10/meta/open-goal.question b/agentbench/10/meta/open-goal.question
deleted file mode 100644
index 3f47549..0000000
--- a/agentbench/10/meta/open-goal.question
+++ /dev/null
@@ -1 +0,0 @@
-定义一项你认为属于「智能」的新能力（如新工具、新记忆策略、新自改进协议），在 kvspace 中实现并演示它：说明思路、写出 kvlang 代码、给出可验证的输出。
diff --git a/agentbench/10/planning/full-autonomy.answer b/agentbench/10/planning/full-autonomy.answer
deleted file mode 100644
index 6c61782..0000000
--- a/agentbench/10/planning/full-autonomy.answer
+++ /dev/null
@@ -1 +0,0 @@
-评分标准: 读入→排序→持久化→读回全链路正确；/bench/sorted 最终状态与打印一致；步骤说明清晰。
diff --git a/agentbench/10/planning/full-autonomy.question b/agentbench/10/planning/full-autonomy.question
deleted file mode 100644
index 52562cf..0000000
--- a/agentbench/10/planning/full-autonomy.question
+++ /dev/null
@@ -1 +0,0 @@
-从零实现一个完整小任务：读入三个数、升序排序、把结果写入 /bench/sorted、再读回打印，全程用 kvspace 持久化，并说明每一步的作用。
diff --git a/agentbench/README.md b/agentbench/README.md
index 90ab6a9..bf75229 100644
--- a/agentbench/README.md
+++ b/agentbench/README.md
@@ -1,58 +1,105 @@
-# agentbench —— byteseek agent 能力测试基准
+# agentbench —— HumanEval / SWE-bench 基准入口
 
-对 corebrain（代码脑）做端到端能力测试：给定自然语言任务，agent 生成 kvlang 程序执行，
-结果经 `println` 落到终端。每题两个文件：`<name>.question`（任务）与 `<name>.answer`（基准答案）。
-
-## 目录结构
+agentbench 不再自带题目。它是一个**薄入口**：调用业界两套标准基准（Codex / Claude 评测用的
+HumanEval 系列与 SWE-bench 系列），把 byteseek 当作被测 agent 接进去。
 
 ```
 agentbench/
-  {难度00..10}/{类别}/{name}.question
-                        {name}.answer
+  bench        入口脚本（bash 壳，exec 到 BENCH_HOME 的 venv）
+  bench.py     驱动：题库 → agent → predictions → 官方评测
 ```
 
-难度在前、类别在后，两级子目录。难度用两位补零 `00`–`10`（共 11 级）。
+## 快速开始
+
+```bash
+agentbench/bench env                                   # 环境自检
+agentbench/bench list                                  # suite 与题量
+agentbench/bench all humaneval --agent gold            # 验管道（题库自带答案，不调 LLM）
+agentbench/bench run humaneval --agent byteseek --limit 10
+agentbench/bench eval humaneval
+```
 
-## 难度量表（0–10）
+## 命令
 
-| 难度 | 说明 |
+| 命令 | 作用 |
 |------|------|
-| 0 | 单步、无工具：直接算术 / 字符串拼接 |
-| 1 | 单步单工具（shell/python），或简单循环累加 |
-| 2 | 单工具带参数 / 单次算法（数组最值、幂运算） |
-| 3 | 单算法 / 字符串统计 / 网络冒烟 |
-| 4 | 数组算法 / 回文判断 / dict 成员读 |
-| 5 | dict 指针链表 / kvspace 读写 / 中等逻辑 |
-| 6 | 多工具组合 / kvspace 多键 / 二分查找 |
-| 7 | 抓取→解析→存储 / 嵌套状态树 / 多约束推理 |
-| 8 | 自省自改 / 错误恢复 / 嵌套代码生成（需 LLM） |
-| 9 | 跨会话记忆 / 开放分解 / 经典谜题 |
-| 10 | 开放目标，无固定解（按 rubric 评分） |
-
-## 类别
-
-| 类别 | 内容 |
+| `bench env` | 打印框架版本、题库路径与题量、docker、kvlang/kvspace/byteseek 就绪状态 |
+| `bench list` | 列出 suite 与题量 |
+| `bench run <suite>` | 生成 `predictions.jsonl`（agent 阶段） |
+| `bench eval <suite>` | 官方评测（eval 阶段）；默认取该 suite 最近一次 run 的 predictions |
+| `bench all <suite>` | run + eval |
+
+`run` 选项：`--agent gold|byteseek`、`--limit N`、`--offset N`、`--task id1,id2`、
+`--out DIR`、`--timeout S`（单题 agent 超时）、`--no-boot`。
+`eval` 选项：`--predictions FILE`、`--workers N`、`--k 1 10`（human-eval pass@k）、
+`--eval-timeout S`、`--dataset-name`（SWE-bench 数据集名）。
+
+退出码：`0` 通过 / `1` 未通过 / `2` 前置条件不满足（缺题库、缺 docker、byteseek 未就绪）/ `3` 参数错误。
+
+## suite 与评测后端
+
+| suite | 题量 | 评测 |
+|-------|------|------|
+| `humaneval` | 164 | 官方 `human_eval.evaluate_functional_correctness`（pass@k） |
+| `humaneval+` | 164 | `evalplus.evaluate`（HumanEvalPlus，base / plus 两档） |
+| `swebench-verified` | 500 | `swebench.harness.run_evaluation`（**需 Docker**） |
+| `swebench-lite` | 300 | 同上 |
+
+agent 阶段与评测阶段分离，与官方 harness 一致：`run` 只产 predictions，`eval` 才判分。
+因此可以本机产 predictions、拿到有 Docker 的机器上判分。
+
+## agent 契约
+
+- **`gold`**：直接用题库自带答案（`canonical_solution` / `patch`），用来验证管道本身，不调 LLM。
+- **`byteseek`**：每题生成一个 `.kv`，内容为 `llm·call(<题目提示>) -> e` 加 `byteseek·run(e)`，
+  由 `kvlang` 驱动并捕获 stdout。agent 必须把答案包在 `<<<ANSWER>>>` 与 `<<<END>>>` 之间；
+  未取到标记时只在日志里提示，prediction 记为空串（不伪造内容）。
+
+  HumanEval 题问「补全该 Python 函数」；SWE-bench 题问「按问题描述产出 unified diff」。
+  SWE-bench 默认只取 diff（不检出仓库）；需要 agent 直接改仓库时，在提示里给出检出目录。
+
+## 环境
+
+重依赖装在 `BENCH_HOME`（默认 `/home/peng.li24/benchmarks`，跨 pod 持久），**不进 git 仓库**：
+
+```
+$BENCH_HOME/
+  venv/                         human-eval + evalplus + swebench + datasets
+  data/human-eval/              HumanEval.jsonl.gz（openai/human-eval 官方题）
+  data/evalplus/                HumanEvalPlus.jsonl（evalplus release）
+  data/swe-bench/               SWE-bench_Verified.jsonl / SWE-bench_Lite.jsonl（HF princeton-nlp）
+  runs/<suite>-<agent>-<时间>/  tasks/ logs/ predictions.jsonl run.json
+  trash/                        旧 agentbench 题目（已被本入口取代）
+```
+
+重装：
+
+```bash
+python3 -m venv $BENCH_HOME/venv
+$BENCH_HOME/venv/bin/pip install human-eval evalplus swebench datasets huggingface_hub
+```
+
+## 已验证（2026-09-13）
+
+| 检查 | 结果 |
 |------|------|
-| arithmetic | 数值计算与算术表达式 |
-| string | 字符串拼接 / 反转 / 统计 |
-| array | 数组算法（最值 / 求和 / 查找 / 排序） |
-| dict | 键值结构与路径指针 |
-| tool-shell | shell 工具调用 |
-| tool-python | python 工具调用 |
-| tool-http | http 网络抓取 |
-| kvspace | 状态树读写与持久化 |
-| reasoning | 逻辑推理与约束满足 |
-| planning | 多步规划与多工具组合 |
-| meta | 记忆、自省、自改（高阶） |
-
-## 答案约定
-
-- 确定性题目：`.answer` 即 `println` 的期望输出（裸值，不含换行）。
-- 冒烟题目（需真实 LLM / 网络 / 完整 bootstrap）：`.answer` 以 `期望(冒烟):` 开头，给出可验证检查。
-- 开放题目（难度 10）：`.answer` 以 `评分标准:` 开头，给出 rubric，无单一标准输出。
-
-## 环境假设
-
-冒烟题依赖完整 byteseek runtime（已 bootstrap，`/lib/byteseek.kvlangbrief` 与
-`/byteseek/prompt/system` 已播种）、可用的 `DEEPSEEK_API_KEY`（嵌套代码生成题）与出网能力
-（http 题，走 `http_proxy`/`https_proxy` 环境变量）。
+| `run humaneval --agent gold` + `eval humaneval` | `pass@1 = 0.9939`（164 题，1 题官方 canonical 解过不了官方测试） |
+| `eval humaneval+`（同一批 gold 样本） | `base 0.9878 / plus 0.8537`（canonical 解在 extra tests 上的已知表现） |
+| `run swebench-lite --agent gold` | 产出 5 题标准格式 predictions |
+| `eval swebench-lite` | 无 docker → 退出码 2，并打印可在有 Docker 机器直接执行的命令 |
+| `run humaneval --agent byteseek` | **阻塞**：见下 |
+
+## 已知阻塞：byteseek lib 与 kvlang 版本漂移
+
+`KVLANG_LIB=lib kvlang` 当前有 3 个文件 layout 失败，byteseek 的 agent 链路因此不可用
+（`--agent byteseek` 会在引导后立刻报错退出，不再静默产出空答案）：
+
+| 文件 | 报错 | 原因 |
+|------|------|------|
+| `lib/byteseek/llm.kv` | `member write on undefined container "/tmp/esc"` | `/tmp/esc·sys` 用了 `·`（成员分隔符），新 kvlang 要求先显式声明容器；路径应写 `/tmp/esc/sys` |
+| `lib/byteseek/shell.kv` | `container literal {…} on "noenv" needs a map langtype` | 未标注 langtype 的 `{}` 字面量 |
+| `lib/byteseek/python.kv` | 同上（`a`） | 同上 |
+
+byteseek `deps.json` 锁的是 kvlang `v0.2.15`，而本机是 `v0.2.18-9`——`/usr/bin/kvlang` 与
+`kvlang/bin/kvlang` 两个二进制都同样失败。这与本次改造无关：修 lib 或对齐 deps 后，
+`--agent byteseek` 即可跑通。
diff --git a/agentbench/bench b/agentbench/bench
new file mode 100755
index 0000000..5dc58ae
--- /dev/null
+++ b/agentbench/bench
@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+# agentbench 入口：所有重依赖（human-eval / evalplus / swebench / datasets）在 BENCH_HOME 的 venv 里。
+set -euo pipefail
+BENCH_HOME=${BENCH_HOME:-/home/peng.li24/benchmarks}
+PY=$BENCH_HOME/venv/bin/python
+[ -x "$PY" ] || { echo "缺少 $PY（先按 agentbench/README.md 安装基准框架）" >&2; exit 2; }
+exec "$PY" "$(dirname "$(readlink -f "$0")")/bench.py" "$@"
diff --git a/agentbench/bench.py b/agentbench/bench.py
new file mode 100755
index 0000000..441d75d
--- /dev/null
+++ b/agentbench/bench.py
@@ -0,0 +1,368 @@
+#!/usr/bin/env python3
+"""agentbench —— HumanEval / SWE-bench 基准入口。
+
+  bench env                              环境自检（框架 / 题库 / byteseek / docker）
+  bench list                             列出 suite 与题量
+  bench run <suite> [选项]               生成 predictions（agent 阶段）
+  bench eval <suite> [选项]              官方评测（eval 阶段）
+  bench all <suite> [选项]               run + eval
+
+suite:  humaneval | humaneval+ | swebench-verified | swebench-lite
+agent:  gold（题库自带答案，验管道） | byteseek（驱动 byteseek 代码脑）
+
+退出码：0 通过 / 1 未通过 / 2 前置条件不满足（缺题库、缺 docker、byteseek 不可用）/ 3 参数错误。
+"""
+
+import argparse
+import gzip
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+BENCH_HOME = Path(os.environ.get("BENCH_HOME", "/home/peng.li24/benchmarks"))
+DATA = BENCH_HOME / "data"
+RUNS = BENCH_HOME / "runs"
+REPO = Path(__file__).resolve().parent.parent
+KVLANG = os.environ.get("KVLANG", "kvlang")
+KVSPACE = os.environ.get("KVSPACE", "redis://127.0.0.1:6379")
+
+BEGIN = "<<<ANSWER>>>"
+END = "<<<END>>>"
+
+SUITES = {
+    "humaneval": {"kind": "humaneval", "file": DATA / "human-eval/HumanEval.jsonl.gz"},
+    "humaneval+": {"kind": "humaneval", "file": DATA / "evalplus/HumanEvalPlus.jsonl"},
+    "swebench-verified": {"kind": "swebench", "file": DATA / "swe-bench/SWE-bench_Verified.jsonl"},
+    "swebench-lite": {"kind": "swebench", "file": DATA / "swe-bench/SWE-bench_Lite.jsonl"},
+}
+
+
+# ── 题库 ────────────────────────────────────────────────────────────────
+def load_tasks(suite):
+    spec = SUITES[suite]
+    path = spec["file"]
+    if not path.exists():
+        die(2, f"题库缺失：{path}（先跑 bench env 看安装状态）")
+    op = gzip.open if path.suffix == ".gz" else open
+    with op(path, "rt", encoding="utf-8") as fh:
+        return [json.loads(line) for line in fh if line.strip()]
+
+
+def task_id_of(task):
+    return task.get("task_id") or task["instance_id"]
+
+
+# ── agent 阶段 ──────────────────────────────────────────────────────────
+def kv_literal(text):
+    """Python 串 → kvlang 双引号字面量（kvlang 与主流语言一致的 \\n \\" \\\\ 转义）。"""
+    return json.dumps(text, ensure_ascii=False)
+
+
+def boot_byteseek():
+    env = dict(os.environ, KVLANG_LIB="lib", KVSPACE=KVSPACE)
+    r = subprocess.run([KVLANG], cwd=REPO, env=env, capture_output=True, text=True)
+    if r.returncode != 0:
+        die(2, f"byteseek 引导失败：\n{r.stdout}{r.stderr}")
+    if not kvspace_get("/lib/llm·call/[0,0]"):
+        die(2, "byteseek 未就绪：/lib/llm·call 缺失——lib/llm.kv 与当前 kvlang 版本不兼容"
+               "（`KVLANG_LIB=lib kvlang` 的完整报错见上）。修好 lib 再跑 --agent byteseek。")
+    return r.stdout + r.stderr
+
+
+def kvspace_get(key):
+    r = subprocess.run(["kvspace", "get", key], capture_output=True, text=True,
+                       env=dict(os.environ, KVSPACE=KVSPACE))
+    return "" if "(nil)" in r.stdout or r.returncode else r.stdout.strip()
+
+
+def prompt_for(kind, task, repo_dir=None):
+    if kind == "humaneval":
+        return (
+            "用 byteseek 完成下面的 Python 编程题：补全函数，把**完整可运行的 Python 实现**"
+            "打印出来（不要解释、不要额外文字）。\n"
+            f"输出必须严格包在 {BEGIN} 与 {END} 之间。\n\n"
+            f"题目：\n{task['prompt']}\n"
+        )
+    where = f"代码仓库已检出在 {repo_dir}，可用 shell 工具查阅与修改。\n" if repo_dir else ""
+    return (
+        "用 byteseek 修复下面的软件缺陷，输出一个可 patch 的 unified diff"
+        "（`git diff` 格式，不要解释）。\n"
+        f"输出必须严格包在 {BEGIN} 与 {END} 之间。\n{where}\n"
+        f"仓库：{task['repo']}\n基线 commit：{task['base_commit']}\n"
+        f"问题描述：\n{task['problem_statement']}\n"
+    )
+
+
+def extract(text):
+    m = re.search(re.escape(BEGIN) + r"\s*(.*?)\s*" + re.escape(END), text, re.S)
+    return m.group(1) if m else ""
+
+
+def strip_prompt(block, prompt):
+    """agent 若把原题函数签名一起打印了，剥掉前缀，只留 completion。"""
+    head = prompt.strip().splitlines()[0].strip() if prompt.strip() else ""
+    pos = block.find(head) if head else -1
+    return block[pos + len(head):] if pos >= 0 else block
+
+
+def run_byteseek(kind, task, outdir, timeout):
+    tid = task_id_of(task).replace("/", "_")
+    tasks_dir = outdir / "tasks"
+    tasks_dir.mkdir(parents=True, exist_ok=True)
+    kvfile = tasks_dir / f"{tid}.kv"
+    kvfile.write_text(
+        "rwfunc test() -> () {\n"
+        f"\tllm·call({kv_literal(prompt_for(kind, task))}) -> e\n"
+        "\tbyteseek·run(e)\n"
+        "}\n",
+        encoding="utf-8",
+    )
+    env = dict(os.environ, KVSPACE=KVSPACE)
+    try:
+        r = subprocess.run([KVLANG, str(kvfile)], cwd=REPO, env=env,
+                           capture_output=True, text=True, timeout=timeout)
+        raw = r.stdout + r.stderr
+    except subprocess.TimeoutExpired as e:
+        raw = (e.stdout or "") + (e.stderr or "") + f"\n[bench] 超时 {timeout}s"
+    (outdir / "logs").mkdir(parents=True, exist_ok=True)
+    (outdir / "logs" / f"{tid}.log").write_text(raw, encoding="utf-8")
+    return extract(raw), raw
+
+
+def predict_gold(kind, task):
+    if kind == "humaneval":
+        sol = task["canonical_solution"]
+        return sol if task.get("prompt", "") in sol else task["prompt"] + sol
+    return task["patch"]
+
+
+def cmd_run(a):
+    kind = SUITES[a.suite]["kind"]
+    tasks = load_tasks(a.suite)
+    if a.task:
+        want = set(a.task.split(","))
+        tasks = [t for t in tasks if task_id_of(t) in want]
+    tasks = tasks[a.offset:]
+    if a.limit:
+        tasks = tasks[: a.limit]
+    if not tasks:
+        die(3, "没有选中任何题目")
+
+    outdir = Path(a.out) if a.out else RUNS / f"{a.suite}-{a.agent}-{time.strftime('%Y%m%d-%H%M%S')}"
+    outdir.mkdir(parents=True, exist_ok=True)
+    if a.agent == "byteseek" and not a.no_boot:
+        print("[bench] 引导 byteseek ...", file=sys.stderr)
+        boot_byteseek()
+
+    preds = []
+    for i, task in enumerate(tasks, 1):
+        tid = task_id_of(task)
+        if a.agent == "gold":
+            ans = predict_gold(kind, task)
+        else:
+            ans, raw = run_byteseek(kind, task, outdir, a.timeout)
+            if not ans:
+                print(f"[bench] {i}/{len(tasks)} {tid}: 未取到答案标记，原始输出见 logs/", file=sys.stderr)
+        if kind == "humaneval":
+            completion = strip_prompt(ans, task["prompt"])
+            preds.append({"task_id": tid, "completion": completion, "solution": task["prompt"] + completion})
+        else:
+            preds.append({"instance_id": tid, "model_patch": ans, "model_name_or_path": f"byteseek-{a.agent}"})
+        print(f"[bench] {i}/{len(tasks)} {tid}: {len(ans)} 字符", file=sys.stderr)
+
+    pred_file = outdir / "predictions.jsonl"
+    with pred_file.open("w", encoding="utf-8") as fh:
+        for p in preds:
+            fh.write(json.dumps(p, ensure_ascii=False) + "\n")
+    (outdir / "run.json").write_text(
+        json.dumps({"suite": a.suite, "agent": a.agent, "n": len(preds),
+                    "predictions": str(pred_file)}, ensure_ascii=False, indent=1))
+    print(pred_file)
+    return 0
+
+
+# ── eval 阶段 ───────────────────────────────────────────────────────────
+def cmd_eval(a):
+    kind = SUITES[a.suite]["kind"]
+    pred = Path(a.predictions) if a.predictions else latest_predictions(a.suite)
+    if not pred.exists():
+        die(2, f"predictions 不存在：{pred}（先 bench run）")
+    if kind == "humaneval":
+        return eval_humaneval(a.suite, pred, a)
+    return eval_swebench(a.suite, pred, a)
+
+
+def latest_predictions(suite):
+    cands = sorted(RUNS.glob(f"{suite}-*/predictions.jsonl"), key=lambda p: p.stat().st_mtime)
+    return cands[-1] if cands else RUNS / f"{suite}-none/predictions.jsonl"
+
+
+def eval_humaneval(suite, pred, a):
+    if suite == "humaneval":
+        from human_eval.evaluation import evaluate_functional_correctness
+        res = evaluate_functional_correctness(
+            str(pred), problem_file=problems_for(suite, pred), k=a.k,
+            n_workers=a.workers, timeout=a.eval_timeout)
+        print(json.dumps(res, ensure_ascii=False))
+        return 0 if res.get(f"pass@{a.k[0]}", 0) > 0 else 1
+    need = {t["task_id"] for t in load_tasks(suite)}
+    have = {json.loads(l)["task_id"] for l in pred.open(encoding="utf-8") if l.strip()}
+    if have != need:
+        die(2, f"evalplus 要求样本覆盖全部 {len(need)} 题，当前只有 {len(have & need)} 题"
+               f"（先 bench run {suite} --agent ... 跑全量）")
+    r = subprocess.run([sys.executable, "-m", "evalplus.evaluate", "--samples", str(pred),
+                        "--dataset", "humaneval", "--parallel", str(a.workers)],
+                       capture_output=True, text=True)
+    if r.returncode != 0:
+        print(r.stdout + r.stderr)
+        return r.returncode
+    res = Path(str(pred).replace(".jsonl", "_eval_results.json"))
+    if res.exists():
+        ev = json.loads(res.read_text())["eval"]
+        n = len(ev)
+        base = sum(1 for v in ev.values() if v[0]["base_status"] == "pass")
+        plus = sum(1 for v in ev.values() if v[0]["plus_status"] == "pass")
+        print(json.dumps({"n": n, "base_pass@1": round(base / n, 4), "plus_pass@1": round(plus / n, 4),
+                          "results": str(res)}, ensure_ascii=False))
+        return 0 if plus > 0 else 1
+    print(r.stdout + r.stderr)
+    return r.returncode
+
+
+def problems_for(suite, pred):
+    """human-eval 要求样本覆盖题库全部题目；跑子集时裁一份同名题库陪跑。"""
+    src = SUITES[suite]["file"]
+    ids = {json.loads(l)["task_id"] for l in pred.open(encoding="utf-8") if l.strip()}
+    rows = load_tasks(suite)
+    if {r["task_id"] for r in rows} == ids:
+        return str(src)
+    out = pred.parent / "problems.jsonl"
+    with out.open("w", encoding="utf-8") as fh:
+        for r in rows:
+            if r["task_id"] in ids:
+                fh.write(json.dumps(r, ensure_ascii=False) + "\n")
+    return str(out)
+
+
+def eval_swebench(suite, pred, a):
+    if not shutil.which("docker"):
+        print(
+            "SWE-bench 官方评测需要 Docker（每个实例在自己的容器里装环境、跑测试）。\n"
+            f"当前环境没有 docker。predictions 已就绪，可在有 Docker 的机器上执行：\n\n"
+            f"  python -m swebench.harness.run_evaluation \\\n"
+            f"    --dataset_name {a.dataset_name or 'princeton-nlp/SWE-bench_Lite'} \\\n"
+            f"    --predictions_path {pred} --max_workers {a.workers} \\\n"
+            f"    --run_id bench-{time.strftime('%Y%m%d-%H%M%S')} --cache_level env\n\n"
+            "或只做「能跑通」的轻量验证：byteseek 生成的 patch 非空即视为产出成功。",
+            file=sys.stderr)
+        return 2
+    from swebench.harness.run_evaluation import main as run_evaluation
+    run_id = f"bench-{time.strftime('%Y%m%d-%H%M%S')}"
+    sys.argv = ["run_evaluation", "--dataset_name", a.dataset_name or "princeton-nlp/SWE-bench_Lite",
+                "--predictions_path", str(pred), "--max_workers", str(a.workers),
+                "--run_id", run_id, "--cache_level", "env"]
+    run_evaluation()
+    return 0
+
+
+# ── 自检 ────────────────────────────────────────────────────────────────
+def cmd_env(_a):
+    def ver(mod):
+        try:
+            __import__(mod)
+        except Exception as e:
+            return f"缺失({e.__class__.__name__})"
+        try:
+            from importlib.metadata import version
+            return version(mod)
+        except Exception:
+            return "已装(版本未知)"
+
+    print(f"BENCH_HOME : {BENCH_HOME}")
+    print(f"python     : {sys.version.split()[0]} ({sys.executable})")
+    print(f"frameworks : human-eval {ver('human_eval')} | evalplus {ver('evalplus')} | "
+          f"swebench {ver('swebench')} | datasets {ver('datasets')}")
+    print(f"docker     : {shutil.which('docker') or '缺失（SWE-bench 官方评测不可用）'}")
+    print(f"kvlang     : {shutil.which(KVLANG) or '缺失'}   KVSPACE={KVSPACE}")
+    print(f"byteseek   : {REPO}")
+    for suite in SUITES:
+        path = SUITES[suite]["file"]
+        n = len(load_tasks(suite)) if path.exists() else 0
+        print(f"  {suite:<18} {n:>4} 题  {path if path.exists() else '缺失'}")
+    for key in ("/byteseek/llm.key", "/byteseek/prompt/system"):
+        r = subprocess.run(["kvspace", "get", key], capture_output=True, text=True,
+                           env=dict(os.environ, KVSPACE=KVSPACE))
+        ok = "(nil)" not in r.stdout
+        print(f"kvspace {key:<22} {'已种入' if ok else '未种入（先 KVLANG_LIB=lib kvlang）'}")
+    return 0
+
+
+def cmd_list(_a):
+    for suite, spec in SUITES.items():
+        n = len(load_tasks(suite)) if spec["file"].exists() else 0
+        print(f"{suite:<18} {spec['kind']:<10} {n:>4} 题")
+    return 0
+
+
+def die(code, msg):
+    print(f"[bench] {msg}", file=sys.stderr)
+    sys.exit(code)
+
+
+def main():
+    ap = argparse.ArgumentParser(prog="bench", description="HumanEval / SWE-bench 基准入口")
+    sub = ap.add_subparsers(dest="cmd", required=True)
+
+    sub.add_parser("env", help="环境自检").set_defaults(fn=cmd_env)
+    sub.add_parser("list", help="列出 suite 与题量").set_defaults(fn=cmd_list)
+
+    def add_common(p):
+        p.add_argument("suite", choices=list(SUITES))
+        p.add_argument("--predictions", help="predictions.jsonl（默认取最近一次 run）")
+        p.add_argument("--workers", type=int, default=8)
+        p.add_argument("--timeout", type=int, default=300)
+
+    r = sub.add_parser("run", help="生成 predictions")
+    add_common(r)
+    r.add_argument("--agent", choices=["gold", "byteseek"], default="byteseek")
+    r.add_argument("--limit", type=int, default=0, help="只跑前 N 题（0 = 全部）")
+    r.add_argument("--offset", type=int, default=0)
+    r.add_argument("--task", help="逗号分隔的 task_id / instance_id 白名单")
+    r.add_argument("--out", help="输出目录（默认 $BENCH_HOME/runs/<suite>-<agent>-<时间>）")
+    r.add_argument("--no-boot", action="store_true", help="跳过 byteseek 引导")
+    r.set_defaults(fn=cmd_run)
+
+    def add_eval_args(p):
+        p.add_argument("--k", type=int, nargs="+", default=[1], help="pass@k（humaneval）")
+        p.add_argument("--eval-timeout", type=float, default=10.0, help="单题执行超时秒数")
+        p.add_argument("--dataset-name", help="SWE-bench 数据集名（默认 SWE-bench_Lite）")
+
+    e = sub.add_parser("eval", help="官方评测")
+    add_common(e)
+    add_eval_args(e)
+    e.set_defaults(fn=cmd_eval)
+
+    x = sub.add_parser("all", help="run + eval")
+    add_common(x)
+    add_eval_args(x)
+    x.add_argument("--agent", choices=["gold", "byteseek"], default="byteseek")
+    x.add_argument("--limit", type=int, default=0)
+    x.add_argument("--task")
+    x.set_defaults(fn=None)
+
+    a = ap.parse_args()
+    if a.cmd == "all":
+        rc = cmd_run(argparse.Namespace(**{**vars(a), "offset": 0, "out": None, "no_boot": False}))
+        if rc:
+            return rc
+        return cmd_eval(a)
+    return a.fn(a)
+
+
+if __name__ == "__main__":
+    sys.exit(main())