Scheduled tasks, background processes, and the promise rule定时任务、后台进程，与“承诺规则”
Scheduled work runs in one of two modes, and the choice between them is a cost decision.定时工作以两种模式之一运行，二者之间的选择是一个成本决策。
A task is a full agent run every tick: it can reason, use tools, and respond to what it finds — capable, but each tick costs an LLM call. A command is a direct script invocation: near-free, and exactly right for deterministic work like fetching and pushing a file.task 是每个 tick 一次完整的 Agent 运行：能推理、能用工具、能应对所见——能力强，但每个 tick 都要花一次 LLM 调用。command 是直接调用脚本：几乎免费，正适合取文件、推文件这类确定性工作。
The scheduler runs on UTC, and the script is the source of truth — the schedule only points at it. Data is always fetched inside the script, never read from memory or pasted in, because LLMs hallucinate numbers and a stale number is worse than no number.调度器按 UTC 运行，脚本是真相源——调度只指向它。数据始终在脚本内部抓取，绝不从记忆里读或粘贴进来，因为 LLM 会幻觉出数字，而过期数字比没有数字更糟。
No recurring job activates without the user seeing what it will cost.任何周期性任务在激活之前，用户都必须先看到它的成本。
The estimate is concrete: runs per day times cost per run, shown both as a daily and a monthly figure. High-frequency polling is not refused — it is priced. If a job polls every minute, it gets a cost line like anything else, and a cheaper variant is offered: a coarser interval, or a command mode, or fewer steps per run.预估是具体的：每天运行次数乘以每次运行成本，分别给出每日与每月的数字。高频轮询不会被拒绝——只会被定价。若某任务每分钟轮询一次，它和其他任务一样会得到一行成本，并会被提供一个更便宜的方案：更粗的间隔、改用 command 模式，或每次运行更少的步骤。
Pricing first, then permission. The user decides; the system's job is to make the decision honest. A job is created in a "dry-run" state if that is what the user asked for, and it stays there until the estimate has been seen.先定价，再授权。决定权在用户；系统的职责是让这个决定诚实。若用户要求，任务会以「试运行」状态创建，在预估被看到之前一直停留于此。
Long work runs in a background session or a detached process — and a hard rule governs how a turn may end.长任务在后台会话或分离进程中运行——而一条硬性规则约束着一轮对话如何收尾。
The rule: never end a turn on a bare promise. A time-bound promise must register a scheduled task in the same turn. A completion-bound promise must spawn a watcher that notices when the job finishes. Nothing wakes the agent otherwise — so an unregistered promise is a lie.规则是：绝不在空头承诺上结束一轮。带时间约束的承诺必须在同一轮注册一个定时任务；带完成约束的承诺必须派生一个能察觉任务结束的 watcher。除此之外没有任何东西会唤醒 Agent——所以未注册的承诺就是谎言。
A background job that produces artifacts is not done when the files exist.产出文件的后台任务，在文件存在的那一刻并没有完成。
It is done when the artifacts reach the user — pushed to a channel, added to the site, written into a report. Integration is scripted into the job itself, so delivery cannot be forgotten; the last step of the job is the delivery. If that final step fails, the job fails — a produced-but-undelivered artifact is still an unfinished job.当产物真正送达用户才算完成——推到频道、加入站点、写进报告。集成被写进任务脚本本身，因此交付不可能被遗忘；任务的最后一步就是交付。若最后一步失败，任务就是失败——产出了却未交付，仍然算未完成。
“What's running?” aggregates three surfaces, so the answer is always one screen away.「现在在跑什么？」汇总三个层面，答案始终就在一个界面之外。
The three surfaces are registered scheduled jobs, live background processes, and auto-started preview services. Cost is attributed per caller — job:, chat:, preview: — which makes credit drain diagnosable instead of mysterious, including the resident services that never appear in any task list.三个层面是：已注册的定时任务、存活的后台进程、自动启动的预览服务。成本按调用方归属——job:、chat:、preview:——这让额度流失可诊断而非神秘，包括那些从不出现在任何任务清单里的常驻服务。