What your agent can do with Retain
Once your assistant is connected, it can do two kinds of things with Retain: run the slash commands you type, and read and write your Retain on your behalf while you work. This page lists both. If you have not connected yet, start with Connect Claude Code or Codex.
The slash commands you type
You type these in Claude Code. In Codex the same skills are named with a dollar sign, such as /retain-apply. Each one needs your agent to be working in a checkout of your repository.
- /retain:explore thinks a change through with you. Your agent reads the code in front of it and what your Retain knows about it: who asked for the change, what was promised, what was tried before. It draws diagrams and asks questions, and it writes nothing. When the thinking settles, it offers /retain:design once.
- /retain:design writes the change up as a build packet. You confirm the direction, the packet's name, the Thread it advances if any, and whether it builds on another packet. Your agent then writes the proposal and the design and sends them to Retain only when both pass Retain's checks.
- /retain:apply builds a packet. Give it the packet's name. Your agent pulls the packet, checks the design against the code, writes the requirements and tasks, sends them to Retain, and builds task by task on a branch of its own. Retain shows the ticked checklist and the pull request as it goes. Add --review to stop after the requirements and tasks are in Retain for you to read.
- /retain:next asks Retain which packet to build next. Retain picks from this repository's packets that are ready and not already being built, shows its reasons, and waits for you. On a yes your agent builds it in its own worktree. See Choose what to build next.
- /retain:sync sends Retain how your linked work has moved, then shows today's digest: what you owe, what is waiting on others, and what needs an answer. It asks before it links anything or marks anything finished. Passing morning runs it with no questions, for a scheduled task.
- /retain:curate-work goes through this repository's Working packets four at a time. For each it checks the code and the pull requests, recommends Archive, Drop, Finish or Keep, and does only what you pick. See The Work list.
The three that build or write packets also need the small command-line tool that the setup installs for your agent. You do not run it.
What your agent can read
You do not need a slash command for these. Ask in plain words, in Claude Code, Codex, Claude or ChatGPT.
- Meetings, notes and email you have kept. Your agent searches them and reads the one that matters, then answers from it. It names the source and date, and quotes a meeting with the speaker and the time.
- Threads. A Thread is what Retain keeps about one piece of ongoing work: its status, what is owed, who is waiting, and the evidence. Your agent can read one, or list the ones that need you. See Threads and what's owed.
- Designs and other documents. Ask for a design doc, a proposal or a notebook by what it is about. Your agent reads it and says whether it is a draft. A document is what Retain wrote with you, so for what was agreed your agent goes back to the meeting.
- Your tasks. Ask what is on your plate and it lists your open tasks, including the actions suggested on your Threads.
- Your email. Where email search is on for your account, your agent searches your connected Gmail by sender, date, subject or words, or by meaning, and reads the messages it relies on. It reads only: it does not send, delete or label mail, and it cannot read inside an attachment.
- Audio briefings. When a Thread has a short audio briefing, your agent offers the link so you can hear it yourself.
- Today's digest, as /retain:sync shows it.
Before it starts work linked to a Thread, your agent also checks what changed there since you last looked, such as a meeting that moved a date. If the plan disagrees with Retain, it tells you with the source and does not edit your files to match.
What your agent can write
Your agent writes to Retain only when you ask or confirm.
- A confirmed update on a Thread. A decision, something done or a result, recorded as your own move. Retain, not your agent, decides what that means for the Thread's status, so an update does not by itself close anything. See Closing the loop.
- A task, when you ask for one, added after it checks that the task is not already there.
- Notes you wrote or pasted, saved as something you said. Your agent does not save its own summaries as notes.
- A link between your work and a Thread. Your agent can link an OpenSpec change or a checklist on your computer to the Thread it advances. Retain then records its progress once a day. Only a title, a short summary and the count of done tasks cross, never file contents.
- A weekly check-off. Ask to go through what you owe. Your agent goes item by item, and records an outcome only for what you say is done.
- Packets, as above: the design, the requirements, the tasks and their ticks.
It never marks a Thread done on its own, and it says what Retain returned rather than what it expects. Progress on a file is not a decision: tick marks reach Retain through the link, not as updates.
Where each assistant fits
Claude Code and Codex can do everything above. The Claude and ChatGPT apps can search, read, list tasks and save confirmed updates, but they have no repository, so they cannot run the slash commands that explore, design or build. For the build flow end to end, see Explore, design, apply.
Need help?
Can't find what you're looking for? Contact our support team for a quick reply.
Contact Support