Harry Xin
HARRY XIN / SELECTED WORK
9 projects. Different questions. One curious practice.9 个项目，不同的问题，同一份好奇。
An agent-loop runner built around a mission, a backlog and a journal: a way to carry useful state from one iteration into the next.围绕目标、待办与日志组织的 agent 循环工具，把一次迭代中的有效信息带到下一次。
Define the task and the stopping point.写清任务，也写清何时结束。
Local state illustration · no agent or background job runs本地状态示意 · 不启动 agent 或后台任务
02 / THE JOURNAL02 / 持续留下的记录
Use “Next step” above. Watch the example leave a trail.点击上方“下一步”，看示例如何留下记录。
journal.ndjsonLOCAL EXAMPLE03 / STATE ON DISK03 / 文件里的状态
mission.mdA mission file records the task, completion criteria and boundaries. A backlog keeps remaining work explicit.目标文件记录任务、完成条件和边界；待办列表明确剩余工作。
backlog.jsonBuild the next prompt from the mission, backlog and journal, then parse a structured report from the agent.结合目标、待办与日志构建下一次提示，并解析 agent 的结构化报告。
journal.ndjsonAppend the result and learning to the journal. Updated state informs the next iteration or a configured stopping condition.把结果与经验追加到日志，再根据更新后的状态继续，或在配置的停止条件下结束。
The repository documents mission.md, backlog.json and journal.ndjson, with a dry-run mode for inspecting the assembled prompt before executing an agent.仓库说明了 mission.md、backlog.json 和 journal.ndjson 的职责，并提供 dry-run，在实际执行前检查组装后的提示。