Connect and collect your selection

Connect the account and check the assignment

Outlook in Codex. Use Part 4's file route for this collection. The Outlook Email connection can return email text in a search result and does not expose the conversation identifiers this preview needs. It cannot provide the body-free, whole-conversation preview described in Part 3. Choose the messages yourself in Outlook and save only the records you approve as plain text. Do not connect a new service, forward your inbox or grant sending access to get around that limit. Week 3 has a separate Outlook drafts-only exercise; it does not change this collection boundary.

Connect the account yourself in your agent app. For Claude Code, add it in Claude’s Settings, then Connectors, and follow the sign-in steps there. For Codex, connect it in Codex’s own settings. Do not give an agent your login.

You will see a permission screen before anything connects. An email connector usually asks to read your mail and also to send, draft and organize it. That's normal, and you don't need to switch anything off in your app. The agents are scoped to reading for this lesson: when Gigawatt assigns the connection, the collecting agent gets only the search and read tools, and in Claude Code every tool that could send, reply or delete is set to ask you first. In Codex that limit comes from the agents’ instructions rather than an app setting, so read any approval Codex asks for before you allow it. If you're not ready to connect yet, go to Part 4.

Connecting the account does not decide which agent should use it. Think back to Module 1: the registry's connector entry is a label, not a switch.

Chief should name the agent responsible for the collection and why its job needs the source. Gigawatt checks that the intended agent can use the prepared reading connection. You review any proposed assignment before it is applied. You do not edit the registry yourself.

After the assignment is applied, Gigawatt reads the record back to you: one line per agent and connector, from your own record of who holds what (work/course/staff/connectors.yaml). Check that the agent you agreed on is the one listed. In Codex, connectors are set for the chat. Whether a Codex agent is limited to its own connectors has not been tested yet, so treat every agent as able to reach what the chat can until it is tested. The record shows your choice for Claude Code.

After you say yes and Gigawatt applies the assignment, close your agent app, open it again and resume this lesson. Agents pick up their tools when a session starts, so the reading agent can't use the new connection until you restart. Without the restart, its first test read fails even though the connection is fine.

Not every agent gets every tool. An agent whose job doesn’t need a tool shouldn’t be able to reach it. Read the job line and ask: whose job needs this source? Cipher will later work from your approved evidence; it does not need your whole inbox or drive to map this process.

Do this now: check one relationship

Read the intended agent's job line beside the proposed source access. Write one sentence saying why that agent needs it. If the name or responsibility is unclear, ask Chief to resolve it before collection.

Preview the selection

Start with the source list, not a request to “collect my email” or “collect my drive.” Have Chief arrange a preview limited to the process records you identified.

Before the preview search runs, agree the search or the named threads, files or records, the date range, and the most items to collect. The preview itself looks into your source, so what it looks at must be allowed too. Give your dates as the first and last day you want. The collection counts days on UTC, the world's standard clock, so a message sent near midnight where you are may count toward the day before or after. Chief shows you the exact range it will use, so you can check it rather than work it out yourself.

Your session, the conversation you started this lesson in, prepares the preview and shows it to you with an approval line. Nothing is collected until you repeat that line exactly. After you do, the same session collects only what the preview showed. Gigawatt's part ends once the connection works.

For email connections that support this preview, the first look at your mailbox is a short list. Each thread gets a number and shows its dates, how many messages it has, and the web domain of each sender (the part after the @, such as example.com). It doesn't show who received the messages, the subjects or any of the text, so nothing inside an email reaches your session before you approve. If you want a thread left out, tell Chief its number. The saved plan records the thread's ID, not its subject, because your workbench keeps a history of every version of its files, and a subject line saved there would stay in that history.

Before you approve, open the listed threads yourself in your own Gmail, not through an agent. To find a listed thread, search Gmail for the sender's domain and dates, for example from:example.com after:2026/06/01 before:2026/06/08 (a made-up example), and match the dates and message count. Check each thread for passwords, login codes or keys that someone sent, and leave out any thread that has them, by its number. This is your check to make, because once an agent reads a message, the text has already gone to the model. As a backstop, the collection holds back any message that looks like it carries a password or key: it saves nothing from it, and Chief tells you which thread was held so you can leave it out and collect again.

Then check the list that comes back before any message bodies, files or attachments are collected. The preview should show:

• The source account or selected export files, described without credentials.
• The query or exact records and date range used.
• The number of matching items and the maximum number proposed for collection, distinguishing messages, threads, and files where relevant.
• Which selected records belong to each of your three process instances.
• Any attachments, by type and size, and whether each one can be brought in.
• Where the selected copies will be saved, which agent will read them, and the model-processing and recording scope.

Here is a fictional preview: three named kickoff threads within an agreed date range, containing eight selected messages, one of which lists a signed-brief attachment (a PDF). That is three process instances, not ten. If one thread includes unrelated work, narrow the selection before confirming it. On a drive or an export, the same preview names each file with its size and date, and a spreadsheet or CSV export shows how many rows it holds, so you can check it the same way.

One email collection holds up to 25 messages, counted across every selected thread. That cap belongs to the import, not to the exercise. The exercise asks for three completed examples of one process, however many messages each one takes, and 25 messages is usually enough for three. If your three examples still don't fit, the preview tells you which thread was cut. Then choose: leave that example out and pick a shorter one, narrow the dates, or, if the preview used a lower limit, ask for a new one with a higher limit. Don't split one selection into several collections just to fit it, because the limit is there to keep each collection small enough for you to check.

What happens to attachments. The collection saves each message and lists its attachments by type and size; it does not bring the attachments in. Claude's Gmail connector can list an attachment but cannot fetch it. Fetching one through Codex's Gmail connector has not been tested for this course yet, so do not count on it. When an attachment matters to your process, save it yourself and hand it over on the file route in Part 4, as a text copy where you can. Sallee's card for that message records the attachment and whether it was brought in, and an answer that might depend on it says so.

A missing result is a reason to correct the query or use a specific export. It is not permission to search or import the whole inbox or drive.

Do this now: decide what to collect

Compare the actual preview with your source list. Remove unrelated items and resolve any surprising count. Confirm the exact selection, including what it lists for attachments, only when it matches what you intend.

If the preview changes in a way that matters, review the new selection before collection goes ahead.

The file route: save the records yourself

Saving the records as files works just as well as connecting. It suits you if your IT policy doesn't allow connecting a work account, if you'd rather not connect one yet, or if your records are easier to save than to search. You can always connect later.

Pick whichever is easiest for you:

• Save the emails. In Gmail, open a message, choose the three-dot menu, then Download message. Save each message from your three instances.
• Copy and paste. Open each email, copy its text, and paste it into a plain text file, one file per message or thread. Keep the date, who sent it, and the order of the replies.
• Use reports and documents. A report or CSV download from your booking or invoicing tool, or the documents from the job itself, often show the process just as well.

Save only the records for your three instances. If a report has rows for other work, keep only the rows for these three. Take out anything that has to stay out before you hand the files over.

These file types work: plain text (.txt or .md), Word documents (.docx), saved emails (.eml, which is what Gmail's Download message gives you), web pages (.html), and CSV or JSON files. A saved .eml lists its attachments but does not extract them, so save an attachment you need as its own file. A PDF, an Outlook .msg file or any other type is kept as a copy, but its text isn't read, so copy the text into a plain text file instead. On Microsoft 365, a document collected through the connector has its text read only when it is plain text, Markdown or CSV; for a Word document, save it and use this file route, which reads .docx. If your email is in Outlook, the simplest route is to copy each message into a text file as described above.

Have Chief show the exact file list and save location before import. Confirm those files, then inspect the collection receipt just as you would on the connection route.

A file that was listed but could not be read does not count as processed evidence. Do not replace it with an agent's guess about its contents.

Read the collection receipt

After collection, open the receipt: the written record of what the job actually did.

Compare it with the preview you confirmed. Look for the records collected, their saved locations, source references, and any items skipped, unsupported, or failed. An attachment should have its own result; the parent message's success does not prove the attachment arrived.

If only part of the selection succeeded, keep that partial result visible. Ask Gigawatt to inspect the failure before retrying. Check what's already saved first, so a retry doesn't quietly create duplicates or collect more than you chose.

These are now collected sources. They are not yet approved knowledge. The next lesson uses the receipt to show Sallee exactly what may be processed for your review.

If you get stuck

A smaller step: Preview one instance and its necessary records first. Once that selection is correct, add the other two within the same agreed process scope.

What you’ll keep

Your access decision, confirmed selection, actual collection receipt, and the selected records at the reported locations. Keep failed or missing items visible and record the next recovery action if there is one.

Before you finish

Check the decisions and the result:

• I checked that model training is off for the account my agent app uses, or that my plan doesn't train on my work.
• I answered the four access questions and either connected my account or saved the records as files myself.
• I can name the intended agent and explain why its job needs the source.
• I confirmed a specific collection, and the receipt shows what arrived and what did not.

You know which of the three instances each record supports. You have not approved any card or index entry yet.

Words you now own

• Reading only: access without sending or changing source material.
• A selection: the specific records this job may collect.
• The file route: saving the selected records yourself and handing over those files, an equal choice to connecting.
• A receipt: the record of what arrived and what did not.

Next: Lesson 4, process the collected material through Cairns.