# Claude Code Settings Not Saving? Check the File Before Changing Permissions

Claude Code 2.1.326 exposes failed settings writes. Separate an unsaved change from invalid JSON or a policy override without weakening your setup.

By Paras Madan | 2026-11-17


A permission choice that works once but disappears tomorrow is not necessarily a model problem. Before changing your prompts or approving broader access, establish whether the configuration was saved, accepted and used. Those are three different checks.

Claude Code 2.1.326, released on November 16, makes the first check easier. Changes made through `/permissions`, a main-conversation permission prompt or `/add-dir` previously could appear saved even when the settings file was unwritable. The terminal now warns and identifies that file. A malformed settings file saved during a session also produces a named-file warning rather than being silently skipped. These are documented fixes, not results from a hands-on test. [Anthropic's release notes](https://github.com/anthropics/claude-code/releases/tag/v2.1.326)

## Record the symptom before trying a fix

Start with a short reproduction note: what you changed, where you changed it, whether it worked immediately, and what happened in a new session. Include the project folder and whether you used the terminal or an editor integration. This keeps a one-project configuration problem from turning into an unnecessary machine-wide reset.

Check the installed version with `claude --version`. That command reports the version; it does not install an update. If your organization controls releases, give the version to the administrator rather than replacing its installation. The [CLI reference](https://code.claude.com/docs/en/cli-reference) documents the available commands and flags.

For example, “an approved command asks again after restarting in the same folder” is useful evidence. “Claude forgot my settings” leaves too many explanations open. Avoid repeating an approval just to make a warning disappear.

## If a warning names an unwritable file

Preserve the existing contents before attempting a correction. Do not delete the configuration directory or recursively loosen file permissions as a shortcut. On a work machine, send the administrator the named path and warning, not a copied configuration containing credentials.

Success means more than seeing a menu selection change. Repeat the narrow test in a fresh session from the same project and confirm the intended behavior remains. Keep other variables unchanged while doing this.

## If the file saved but the behavior did not change

Run `/status` to inspect loaded settings sources and `claude doctor` to find rejected entries. The status list confirms sources, not which source supplied every individual key. Settings files require strict JSON. A trailing comma or comment can break parsing. [Settings documentation](https://code.claude.com/docs/en/settings)

Check whether another settings source governs the key. Managed policy requires an administrator, not a competing local override. Consult the key's documentation before editing again. Change one thing at a time so you can tell which correction helped.

## Stop once you have a narrow explanation

Do not use a real deployment, deletion or production-data operation as the test case. Choose a harmless action that demonstrates the same issue. Keep any support example small enough that someone else can understand the failure without receiving your repository or account configuration.

Anthropic's [troubleshooting guide](https://code.claude.com/docs/en/troubleshooting) directs configuration issues to configuration debugging and offers `/doctor` for broader checks. Review any proposed fix before accepting it. Escalate with the version, surface, warning and minimal reproduction rather than a full environment dump.

The useful outcome is a specific conclusion: the file was unwritable, an entry was rejected, or another setting controlled the behavior. That is a better basis for a repair than granting Claude Code more access and hoping the symptom stops.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help engineering teams make AI development workflows easier to operate and debug.

[Book a 30-minute call](https://cal.com/paras-madan/30min)
