# Slash commands — /every > **What this chapter covers** > What the `/` command does, why it doesn't send, and why "keep trying until it's > done" became the default behaviour instead of a command. ## There is only one Type `/` in the box and the list appears. There is one entry in it today. | Command | What it does | |---|---| | `/every ` | Turns what you are typing into a **repeat** | > **There used to be a `/goal` alongside it.** It set "retry by yourself until this > condition is met" — and that behaviour became **the default for every > conversation**, so the command was removed. See "Retrying is the default" below > for what runs in its place. ## /every — make it a repeat ``` /every daily summarise yesterday's sales into a table ``` The words it understands as cycles come from the **same table** as the [scheduled screen](/guide/cn-queue). ``` hourly · daily · evening · weekly · friday · monthly ``` Korean words work too. For any other cycle, set a cron expression on the scheduled screen. ### Why it doesn't send `/every` **registers and stops there.** > There is no reason to also do "every morning, do X" once right now. Sending it > means only the author gets the first run twice. A toast tells you where it was created. ### Why it lives in the message box A request worth repeating is usually **the one that just worked**. And that request's settings are already chosen in the box in front of you. ``` Already chosen right now: agent · model · group · withheld tools · drive scope → /every carries them over. The only thing left to pick is the cycle ``` Send people to a blank form on the scheduled screen and they have to pick all five again — so nobody does. ## Retrying is the default — there is nothing to switch on This is what changed from earlier versions of this manual. > **Every conversation now keeps going until the result is reported.** There is no > command and no screen to enable it. ### How the loop turns When an agent finishes it **reports its own verdict** through `report_result`. That report decides whether the loop turns again. | Report | What happens next | |---|---| | `completed` | **Ends immediately.** A conversation that finishes in one shot never loops | | `failed` + one line on what to do next | The system appends that line and runs **one more turn** | | `blocked` (a human is needed) | Stops and asks for a person | ``` Default attempts 3 Hard cap 5 ``` ### Why the model doesn't write its own next prompt If the model writes its whole next instruction, **the judge and the subject become the same party.** It breaks in three ways in practice. | Failure | How it looks | |---|---| | Self-confidence | Writes `failed` and then "almost there, wrap it up" — circling the same spot | | Intent drift | The original request falls out of the history; a few turns later it is doing something else | | Constraint loss | Summarising quietly drops conditions like "expired coupons use this wording" | So the model owns **exactly one line — "what to do next"** — while the original request and completion condition are re-attached **verbatim by the system** on every turn. That is what stops the bottom two rows. This is the product doing "restate the important constraints" from [when context gets pushed out](/guide/ai-context-overflow) on your behalf. ### So what do you do instead? The command is gone, so **the completion condition goes inside the request.** ``` ✗ (before) /goal until every test passes ✓ (now) Fix the tests on the login screen. Done when: the tests for wrong password / no such account / locked account all pass. ``` > **A completion condition has to be machine-checkable.** "Until it looks good" is > not a condition. Same standard as the completion condition in > [the four parts of a good instruction](/guide/ai-intent-context) — you now write > it **in the body** rather than in a separate command. ## Typos just become messages Mistype `/evrey daily …` and it is not recognised as a command — **it goes to the model as text**, which reads it as an instruction. Picking from the list is safer. **The removed `/goal` behaves the same way.** If it is still in your fingers, `/goal ...` raises no error — the sentence simply goes to the model, which is more confusing, not less, because it does not look like nothing happened. Put the completion condition in the body instead. ## Common misunderstandings ### "Where do I edit what /every created?" **On the [scheduled screen](/guide/cn-queue).** Once created it is identical to one made there — editable, deletable, pausable. ### "Won't the retry loop run forever?" **No.** Three attempts by default, five at most. And the moment the model reports `completed` it stops, so short conversations never loop at all. ### "What if I want to turn retrying off?" **There is no off switch on screen.** That is the point of making it the default — it used to be off, so any conversation nobody switched it on for (which was nearly all of them) ended right where the model stopped halfway. The only signal a person got was a notification, and a notification tells you what finished; it cannot ask for it again. When calling from a program, setting `repeatMaxAttempts` to `0` **explicitly** in the [API](/guide/cn-api) turns it off. --- ## Check yourself **1. Why doesn't `/every` send?**
Answer **Because it is work for later.** Sending it once more now would give the author the first run twice. `/every` only registers, and a toast tells you where.
**2. Why is `/every` in the message box rather than the scheduled screen?**
Answer **Agent, model, group, tools and drive scope are already chosen there.** Making people pick them again in a blank form means nobody does it.
**3. What do you do to get "keep trying until it's done"?**
Answer **Nothing.** Retrying is the default for every conversation (3 attempts, 5 max). The only job left is to **write the completion condition into the request body**, and that condition has to be machine-checkable.
**4. What happens if you mistype a command?**
Answer **It goes out as a message.** The model reads it as an instruction, so picking from the list is safer. The removed `/goal` also becomes a plain message rather than an error.
--- Next, building agents → [What agents are](/guide/cn-agents-intro)