thomasmustier/pi-nes-sessions

--- pretty_name: coding agent session traces task_categories: - text-generation tags: - agent-traces - coding-agent - pi-share-hf language: - en - code license: other --- # Coding agent session traces for thomasmustier/pi-nes-sessions This dataset contains redacted coding agent session traces collected while working on [tmustier/pi-nes](https://github.com/tmustier/pi-nes). The traces were exported with [pi-share-hf](https://github.com/badlogic/pi-share-hf) from a local [pi](https://pi.dev) workspace and filtered to keep only sessions that passed deterministic redaction and LLM review. ## Data description Each `*.jsonl` file is a redacted pi session. Sessions are stored as JSON Lines files where each line is a structured session entry. Entries include session headers, user and assistant messages, tool results, model changes, thinking level changes, compaction summaries, branch summaries, and custom extension data. Pi session files are tree-structured via `id` and `parentId`, so a single session file may contain multiple branches of work. See the upstream session format documentation for the exact schema: - https://github.com/badlogic/pi-mono/blob/main/packages/coding-agent/docs/session.md Source git repo: [tmustier/pi-nes](https://github.com/tmustier/pi-nes) ## Redaction and review The data was processed with [pi-share-hf](https://github.com/badlogic/pi-share-hf) using deterministic secret redaction plus an LLM review step. Deterministic redaction targets exact known secrets and curated credential patterns. The LLM review decides whether a session is about the OSS project, whether it is fit to share publicly, and whether any sensitive content appears to have been missed. Embedded images may be preserved in the uploaded sessions unless the workspace was initialized with `--no-images`. ## Limitations This dataset is best-effort redacted. Coding agent transcripts can still contain sensitive or off-topic content, especially if a session mixed OSS work with unrelated private tasks. Use with appropriate caution.pretty_name: 编码智能体会话追踪（coding agent session traces） task_categories: - 文本生成 tags: - 智能体追踪（agent-traces） - 编码智能体（coding-agent） - pi-share-hf language: - 英语 - 代码 license: 其他 --- # 面向 thomasmustier/pi-nes-sessions 的编码智能体会话追踪数据集 本数据集收录了在开发 [tmustier/pi-nes](https://github.com/tmustier/pi-nes) 过程中收集的经过脱敏处理的编码智能体会话追踪数据。这些追踪数据通过 [pi-share-hf](https://github.com/badlogic/pi-share-hf) 从本地 [pi](https://pi.dev) 工作区导出，并经过筛选，仅保留通过确定性脱敏与大语言模型（LLM）审核的会话。 ## 数据说明 每个 `*.jsonl` 文件均为一条脱敏后的 pi 会话。会话以 JSON Lines 格式存储，每一行对应一条结构化会话条目。条目包含会话头部、用户与助手消息、工具调用结果、模型变更、思考层级变更、压缩摘要、分支摘要以及自定义扩展数据。 Pi 会话文件通过 `id` 与 `parentId` 构建树形结构，因此单个会话文件可包含多条工作分支。如需了解精确的架构（schema）规范，请参阅上游会话格式文档： - https://github.com/badlogic/pi-mono/blob/main/packages/coding-agent/docs/session.md 源 Git 仓库：[tmustier/pi-nes](https://github.com/tmustier/pi-nes) ## 脱敏与审核流程 本数据集通过 [pi-share-hf](https://github.com/badlogic/pi-share-hf) 进行处理，结合了确定性密钥脱敏与大语言模型审核两个步骤。确定性脱敏针对已知的敏感密钥与预设的凭证模式进行处理。大语言模型审核环节将判断会话是否与开源项目相关、是否适合公开发布，以及是否存在遗漏的敏感内容。 除非工作区初始化时使用了 `--no-images` 参数，上传的会话中可保留嵌入的图像。 ## 局限性说明 本数据集仅尽力完成脱敏处理。编码智能体的会话记录仍可能包含敏感或无关内容，尤其是当会话同时涉及开源项目工作与非相关私有任务时。使用时请谨慎评估。