Borrowing it
Nothing to install: this file belongs to LBurny/symkit-mcp. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/LBurny/symkit-mcp/main/.agents/skills/sandbox-experiment/SKILL.mdgit clone --depth 1 https://github.com/LBurny/symkit-mcpWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/lburny/symkit-mcp/sandbox-experiment)<a href="https://agentmods.dev/skills/lburny/symkit-mcp/sandbox-experiment"><img src="https://agentmods.dev/badge/skills/lburny/symkit-mcp/sandbox-experiment/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/lburny/symkit-mcp/sandbox-experiment"><img src="https://agentmods.dev/badge/skills/lburny/symkit-mcp/sandbox-experiment.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00213 | $0.03379 |
| Opus 5.5 | $0.00085 | $0.01352 |
| Sonnet 5.5 | $0.00043 | $0.00676 |
| Haiku 4.5 | $0.00021 | $0.00338 |
Grade A, and why
sandbox-experiment scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 77 lines — stays where its author put it; the contents beside it link to each section on GitHub.
symkit-mcp 沙箱试验（黑箱轮）操作手册
黑箱原则：symkit-mcp 只通过它的 MCP 工具接口被评测。操作员（无头 Claude）和探针（stdio MCP client）从不读源码；源码侧的工作（根因、修复）由主控和修复代理在仓库里做。观察到的摩擦驱动框架改进——1.5.0→1.12.0 的十余轮实战（综合轮 + 量纲/Lean/算子面/python_exec 专项）都是这条闭环跑出来的。
六步闭环
| 步骤 | 关键产物 | 详参 |
|---|---|---|
| 0. 决策：轮型 / 目标 wheel / lane 划分 | 卡片清单 | lab-setup.md |
| 1. 搭实验台 | symkit-mcp-test-rNN/（venv + wheel + smoke 全绿） |
lab-setup.md |
| 2. 写试验卡 | tasks/task-NN-*.md |
lab-setup.md |
| 3. 跑卡 | runs/<run-id>/stream.jsonl + findings/task-NN.md |
probes.md |
| 4. 探针复现 | probe/audit*.py：每个缺陷坐实或驳回 |
probes.md |
| 5. 缺陷分析修复 | 仓库内 TDD 修复 + 回归测试 | findings-acceptance.md |
| 6. 验收闭环 | 重装 wheel + 探针/失败卡重跑 + 四门禁 + 文档同步 | findings-acceptance.md |
lab 目录布局（模板见 assets/）：
symkit-mcp-test-rNN/
├── .venv/ # lab 自己的解释器（wheel 装在这里）
├── dist/ # 本轮被测 wheel
├── data/<lane>/ # 每条 lane 独立 SYMKIT_DATA_DIR
├── .mcp.json → runs/<run-id>/.mcp.json # run_task.sh 每 run 生成，勿手写
├── tasks/task-NN-*.md # 试验卡
├── runs/<run-id>/ # stream.jsonl / stderr.log / 拷回的会话 JSON
├── findings/task-NN.md # 操作员逐卡报告
├── findings_rNNa/ # 重跑验收时备份的旧 findings
├── probe/ # smoke.py / regress.py / audit*.py / lean_real.py
├── analyze.py # stream.jsonl 汇总表
├── run_task.sh / run_suite.sh / run_task_lean.sh
└── runs/_analysis/ # ACCEPTANCE.md / suite-summary.json / 白盒证据
角色分工
- 主控（你）：定轮型与卡片、跑探针坐实根因后再派工、写死接口契约与互不相交的文件所有权、验收时亲自审 diff + 跑全量门禁。设计决策（根因、修法方向、语义红线）主控做完，代理只实现。
- lane 子代理（ZCode 子代理，每 lane 一个）：驱动无头 Claude 跑
run_suite.sh（run_in_background + 轮询 stream.jsonl 进度），收齐 findings 后照录汇报。派工单写死：lane 编号、卡号区间、data-dir、RUN_PREFIX、max-turns 预算、"只跑卡不改卡不改 harness"。 - 操作员（无头 Claude，被 lane 子代理驱动）：只拿试验卡黑箱操作，产物写 findings/。
- 修复代理：单 Wave 内 TDD 实现，聚焦测试（
--no-cov -p no:cacheprovider），越界需求停下报告。派工单里每个指控逐条对号其复现探针编号（r20 教训：不对号的指控会被代理自由发挥）。 - 并行上限 3 条 lane / 3 个子代理；子代理 Bash 上限 600 秒 → 必须 run_in_background + 轮询。
What ships with it
11 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- assets/analyze.py 5.5 KB runs code
- assets/probe_lean_real.py 2.9 KB runs code
- assets/probe_lean_regress.py 7.6 KB runs code
- assets/probe_smoke.py 1.8 KB runs code
- assets/run_suite.sh 1.0 KB runs code
- assets/run_task_lean.sh 1.7 KB runs code
- assets/run_task.sh 2.1 KB runs code
- assets/task-template.md 1.5 KB
- references/findings-acceptance.md 6.0 KB
- references/lab-setup.md 8.6 KB
- references/probes.md 8.5 KB
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 1mo ago Changed · +6 lines · +43 tokens per session ff81b43b29a0
- 1mo ago Changed · +1 lines b9320afa9ece
- 1mo ago First seen · 70 lines · 170 tokens per session scan A a92c06e98ecb
sandbox-experiment is a skill published in the GitHub repository LBurny/symkit-mcp (13 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 213 tokens to every session and 3,379 once invoked, about $0.0009 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-15.
Other skills, from other repositories
lean-probe
Fast Lean 4 proof verification for coding agents via the LeanProbe MCP server's warm REPL. Use when you need to know whether Lean 4 code compiles (no errors, no sorry), iterate on a declaration inside a Lake/Mathlib project, test a replacement proof candidate, or explore a goal tactic by tactic — far faster than lake…
api-testing
Testing patterns for MCP tool/resource handlers using createMockContext and Vitest. Covers mock context options, handler testing, McpError assertions, format testing, Vitest config setup, and test isolation conventions.
add-test
Scaffold a test file for an existing tool, resource, or service. Use when the user asks to add tests, improve coverage, or when a definition exists without a matching test file.
field-test
Exercise tools, resources, and prompts against a live HTTP server via MCP JSON-RPC over curl. Starts the server, surfaces the catalog, runs real and adversarial inputs, measures every call (bytes, token estimate, wall-clock) and weighs the catalog, renders app tools' views in a headless MCP Apps host, and produces a…
research-engineer
An uncompromising Academic Research Engineer. Operates with absolute scientific rigor, objective criticism, and zero flair. Focuses on theoretical correctness, formal verification, and optimal implementation across any required technology.
tika-eval-compare
Compare extracts from two Tika builds over a corpus to detect regressions in content, encoding, exceptions, and embedded-document handling. Use for "compare before/after extracts", "eval this change against the corpus".