Running Claude Code Unattended? Three October Changes That Matter: Idle Compaction, Retry Stops and Auto-Mode Setup

Claude Code shipped a long run of releases in October 2026, and most of them are small. A few are not, if you leave sessions running without watching them: scheduled jobs, long refactors, overnight test loops. According to the third-party release tracker Havoptic, three of them target exactly that situation: idle compaction in v2.1.299 (October 14), better unattended retries and automatic stops at usage limits in v2.1.301 (October 15), and an /auto-mode-setup command in v2.1.304 (October 19). This lesson explains what each is for and how to set up a long-running session so that it stops cleanly instead of failing noisily or burning budget. If you already read our lesson on making a guard hook fail closed, this is the companion piece about the rest of an unattended setup.

What You'll Learn

What idle compaction, the unattended retry and usage-limit behavior, and /auto-mode-setup do according to the release notes; a checklist for configuring an unattended Claude Code session; and how to confirm each setting on your own installation instead of trusting a summary.

A caution on sources first

Havoptic is a third-party aggregator that summarizes release notes, and its own page has small inconsistencies (for example, its summary count does not match the list). Treat the one-line descriptions below as pointers, then confirm the exact behavior in claude --version, the in-app /release-notes and the official Claude Code documentation before you depend on any of them.

Change 1: idle compaction (v2.1.299, October 14)

Long conversations eventually have to be compacted: older turns are summarized so the session can continue within the context window. The release adds idle compaction for long conversations, and the tracker notes it can be switched off with the setting "idleCompaction": false.

Why it matters for unattended work: compaction that happens while a session sits idle means the next step starts from a smaller context rather than hitting the limit in the middle of a task. The trade-off is that compaction is lossy. A summary can drop a detail you needed, such as the exact failing test name or a decision you made an hour ago.

What to do: keep a short PLAN.md or NOTES.md in the repo that the agent updates as it goes, and tell it to re-read that file after any compaction. If your workflow depends on exact earlier output, test with idle compaction on and off and compare. The setting goes in your Claude Code settings file; check the settings documentation for the correct location for your scope (user, project or local).

Change 2: unattended retries and usage-limit stops (v2.1.301, October 15)

The tracker describes v2.1.301 as handling unattended retries better and stopping requests automatically when usage limits are hit. For a job nobody is watching, the failure modes are a loop that keeps retrying a request that can never succeed, or a session that quietly stalls at a limit.

What to do:

• Run unattended jobs with a clear stop condition in the prompt (for example, a maximum number of attempts at the same failing step) rather than relying on the tool alone.
• Have the job write its status to a file or log as it goes, so a stop at a usage limit leaves a record of how far it got.
• Schedule heavy jobs away from your interactive hours so the two do not compete for the same limits.

Change 3: /auto-mode-setup (v2.1.304, October 19)

The tracker says this release added /auto-mode-setup, which learns from your project to guide the setup of auto mode, and that v2.1.302 gave users more control over auto mode on their device. Auto mode is the setting where the agent proceeds without asking about every action, so the quality of your permission configuration decides how safe it is.

What to do: run the command in a throwaway branch first and read what it proposes before accepting it. Then compare the result against the permission rules you already have: anything that allows broad shell access or network calls deserves a second look. Pair it with fail-closed hooks for the actions you never want to run unreviewed.

Related: a safety telemetry event per stopped or prompted call (v2.1.307, October 21)

The same tracker lists a new telemetry event emitted for every tool call that was stopped or prompted. If your organization collects Claude Code telemetry, this gives you a count of how often guardrails fire. A sudden drop in prompted calls after you enable auto mode is worth investigating, because it can mean the agent is now allowed to do things it used to ask about.

An unattended-session checklist

1. Work on a branch or a clean worktree so a bad run is easy to discard.
2. Write the goal, the stop condition and the definition of done into a notes file the agent re-reads.
3. Decide on idle compaction deliberately instead of leaving it to chance.
4. Use fail-closed hooks for destructive commands and secret access.
5. Run /auto-mode-setup on a throwaway branch and review the result.
6. Log progress to a file and check it when the run ends or stops.
7. Review the diff yourself before merging. Unattended does not mean unreviewed.

Verify before you rely on it

Run claude --version to see which release you have, open /release-notes for the official wording, and test each setting on a small job. A five-minute dry run is cheaper than discovering at 3 a.m. that a setting behaves differently from the summary you read.

Conclusion

None of these October changes is dramatic on its own. Together they make long unattended runs more predictable: context that is managed while the session idles, requests that stop at limits instead of spinning, and a guided way to configure auto mode. The habit that keeps them safe is the same one as always: small scope, a written stop condition, fail-closed guards and a human review of the final diff. For the research behind safe agent workflows, see the security and prompt chapters in the Vibe Coding Ebook, and get new lessons like this in the Academy newsletter. Source for the release details: Havoptic's October 2026 Claude Code release list, read October 25, 2026; verify against the official changelog.