AXME Code · 5 min read

I Went on Vacation. My Agent Came Back to an Empty Project.

Five days off. Session handoff gone. Half a week of decisions forgotten. Here's the 'what was I doing yesterday' problem and why session close is a ritual that matters.

I took five days off. Not a big trip, just a long weekend stretched by two national holidays. When I came back and opened Claude Code on the project I’d been deep in for the previous two weeks, the agent greeted me with the same energy it had when I first installed the tool.

“How can I help you today?”

I sat there for a full minute trying to remember what I had been working on when I stopped. I knew I had been refactoring the cost-tracking module. I knew there was an open question about whether to keep cents as integers or switch to Decimal. I knew I had left a failing test that was not actually a bug but a stale fixture. I did not remember which branch any of this was on.

I opened the repo and stared at git branch. Thirty feature branches, six of them touched in the last two weeks, none of them self-explanatory. I did not know which one I was in the middle of.

The shape of the problem

This is not Claude Code’s fault specifically. It is the shape of working with any tool that doesn’t write down what was happening when you stopped.

Here’s what I normally do when I stop a session mid-task:

  1. Commit whatever is ready, leave the rest as work-in-progress.
  2. Push the branch (sometimes).
  3. Close the terminal.
  4. Go do something else.

Tomorrow, next week, after vacation, I come back and try to reconstruct the state. Sometimes I left a half-written commit message. Sometimes I left a note in my head. Most of the time I just remember.

“Most of the time I just remember” works if your absence is less than a day. It breaks at day two. It shatters at day five. After a week, the reconstruction cost is longer than the context loss. You spend an hour figuring out what you were doing, which is an hour of pure pain because you’re not doing the work, you’re figuring out what the work was.

After vacation, this cost compounds with rust. You forgot the shape of the problem, the names of the fields, the reason a particular test was failing. Tools that work well on a fresh brain feel dumb when your brain is rusty. You don’t need your tools to be dumber when you’re trying to restart.

What I actually needed

I needed three things:

  1. A handoff note. One paragraph, human-written, saying what I was doing, what’s in progress, what the open question was.
  2. A branch indicator. Which branch I was on when I stopped.
  3. A list of open threads. Not just the in-progress work, but the things I was planning to look at next. The stuff in my head that I knew I was going to forget.

None of this is revolutionary. This is literally the summary I would write in a handoff doc if I were going on parental leave for three months. The weird thing is I don’t write it when I stop for five days. I should. I don’t.

The reason I don’t is that writing a handoff every time I stop a session is 90 seconds of friction, and most of the time I’m not going to need the handoff, so I skip it.

That calculation is wrong. It optimizes for the most common case (coming back tomorrow, fine) and ignores the expensive case (coming back in a week, hour of lost work). I should be writing the handoff every time specifically because I don’t know in advance which case this will be.

The session-close ritual

Here is what I do now, enforced by AXME Code.

When I say “end session,” a background audit agent reads the full session transcript. It extracts:

  • New memories: things I learned during the session that should carry forward
  • New decisions: architectural choices with rationale, tagged by enforce level
  • New safety rules: patterns I explicitly marked as never-do
  • A handoff: one paragraph summary of where the session stopped, what’s in progress, what’s next

All of that gets written to .axme-code/ in the repo. On the next session start, the agent reads it and begins with full context.

The “handoff” part is the important one for the vacation case. When I come back after five days, the first thing I see in the new session is:

## Previous Session Handoff

Stopped at: mid-refactor of cost-tracking module

In progress:
- Converting cents from int to Decimal across services
- Left failing test at test_cost_rollup_weekly.py:42 (stale fixture, not a bug)

Next steps:
- Finish Decimal migration in reporting module
- Update axme-admin CLI to format Decimal output
- Investigate whether we need a Decimal wrapper for external API

Open question: Should historical records in BigQuery also migrate to Decimal?

I read this and I am back in the context in fifteen seconds. Not an hour. Fifteen seconds.

The cost of the ritual

The ritual costs me about two minutes at session close, which I don’t even pay because the audit agent is automatic. I type “end session,” the agent runs the audit in the background, I go do whatever I was going to do next, and the handoff is waiting for me tomorrow or next Tuesday or whenever.

The only friction is remembering to end the session properly. For a while I was just closing the terminal, which skips the audit. I broke that habit within a week because the cost of not doing it (starting the next session with no handoff) was immediate and obvious.

What nobody talks about

There’s a second-order benefit to this ritual that I did not expect. Writing the handoff changes how I work during the session.

When I know the session is going to end with a handoff, I work more linearly. I don’t leave four half-done things at once. I try to get one thing to a stopping point before I start the next. Because if the audit agent has to write “you were in the middle of three different things,” the handoff is useless and I will pay for that tomorrow.

The ritual isn’t just a way to save state. It is a gentle pressure to work in a way that leaves a clean state.

The one-line version

If you do one thing after reading this, end your session with a written handoff. Open a text file, type three lines about what you were doing, what’s in progress, what’s next. Commit it. Go.

Even if the tool you use does not automate this, the discipline alone is worth ten times what it costs. Vacation is not the worst case. The worst case is when you don’t go on vacation and you come back tomorrow thinking “I’ll remember,” and you don’t, and you lose half a morning figuring out what yesterday-you was doing.

I spent a full hour reconstructing context after that five-day vacation. I do not have that hour to waste twice a year. I wrote a tool that makes sure I don’t.

More on AXME Code