Hour four. The agent has quietly stopped being good.

A context window is finite. Everything competes for it: your conversation, the files that were read, the command output, the instruction files, the skills. When it gets close to full, the older conversation is compressed into a summary.

The vendor documents which way the loss goes:

"Your requests and key code snippets are preserved; detailed instructions from early in the conversation may be lost."

code.claude.com, How Claude Code works

Early and detailed. That is the description of every rule you set up at the start of a long session.

The habit that works is unpopular, because it sounds like giving up: stop trying to make the session last, and make restarting cheap.

1. Write the state down before you need it

Every half hour or so, put the current state somewhere outside the chat. What is decided, what is in flight, what is next. People do this with a SESSION.md they keep updating. It looks like overhead until the first time it saves you two hours.

2. Compact on purpose, with a focus

You do not have to wait for the automatic pass to guess what matters. Running compact with instructions, such as focusing on the auth work, keeps what you choose rather than what it guesses. Clearing between unrelated tasks is the same idea, applied earlier.

3. Put the rules where they come back

This one is free and most people do not know it. A rule in your project-root instructions file is re-injected after compaction. The same rule in a nested folder or behind a file pattern is not, until a matching file is read again. If a rule must survive, move it to the root file. That is straight from the table above.

4. Restart earlier than feels necessary

The moment you notice repetition is already late. A fresh session that opens with a written state beats a four-hour session running on a summary of itself.

Mission HQ makes step one and step four automatic. What it changes is the cost of restarting, and it changes it a lot.

The work state is never in the conversation

Tasks, the decisions behind them, the warnings and the next step are records on a server. Compaction summarises a conversation. It cannot reach a record, because the record was never in the window.

Restarting costs one call

A new session asks for the handoff and gets back the current state, the single next action, and the warnings. That is the whole restart. No archaeology through a summary, no re-explaining the architecture to a fresh window.

The rules arrive with the task, not at the top of the session

Conventions and past lessons are delivered when a task of that kind starts, which means they arrive fresh rather than sitting at the far end of a long conversation waiting to be summarised.

Nothing in flight depends on the session ending well

Checkpoints are posted on every prompt and at the end of every turn. The 80-per-cent-done task that compaction wipes from the conversation is still a task with a status.

Be clear about the shape of it, because it decides whether this is useful to you. What changes is the cost of a restart. A session that has been running for four hours still fills up, still compacts, still drifts. That is the context window doing what a context window does.

What changes is that a restart stops being expensive. When the state is a record instead of a memory of a conversation, ending a rotted session is a decision you can make in a second, several times a day, without losing anything. The reason people push a bad session to hour six is that starting over used to cost them the morning.

So the advice on this page is the same advice with the price removed: restart early, restart often, and pick up exactly where you were.