Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Claude Code 2.1.327 lists broken setup files

Written November 18, 2026

Claude Code 2.1.327 lists broken agent, skill and command files in /status, reads MCP server JSON from stdin, and asks before /export overwrites.

Claude Code 2.1.327, released on November 17, 2026, is mostly a release about silence ending. The v2.1.327 release notes describe setup problems that used to fail quietly: an agent file with an error in it, a keybinding that never took, a hook matcher that matches nothing. In 2.1.327 those problems get named in one place. The same release moves MCP server tokens off the command line, and stops two operations from destroying work without asking first. If you maintain a Claude Code setup with custom agents, skills or commands, this is the release that tells you which parts of it are broken.

/status names the broken files

The headline change is a list in /status. It names broken agent, skill, command and output style files, so a file that fails to load is now listed instead of simply absent from your session. The list goes further than files. It also surfaces keybinding errors, an unavailable sandbox, ignored permission settings and hook matchers that never match. Each of those is a failure mode that previously left no trace: the permission setting you wrote was ignored, the hook you configured could never fire, and nothing on screen said so. Opening /status after updating is the fastest way to learn whether your setup has been carrying one of these quiet faults. Fixing the named file is still your job, but the release at least points at the right one.

MCP tokens stay off the command line

A new form of the add command, claude mcp add-json <name> -, reads the MCP server's JSON from stdin instead of taking it as a command-line argument. The notes give the reason directly: tokens can stay off the command line, where they would otherwise sit in shell history and in the argument lists other tools can read. Pipe the JSON in, and the secret travels through stdin instead. One limit applies: the stdin form is not available inside a Claude Code session. It is a shell command, run outside the session, which fits how most people add a server in the first place. If your current habit is pasting a token into an add-json call, this release gives you a cleaner habit to switch to.

Two operations now ask first

/export no longer overwrites an existing file without asking. The dialog now asks before replacing it, and the named form behaves differently again: /export <name> keeps the existing file and tells you to run /export instead. An export you made last week is no longer one careless command away from being replaced by today's session. A second confirmation lands on the GitHub side: a confirmation now comes before /install-github-app changes a repository, so that change is a decision you make, not a side effect you discover afterwards.

Settings explain themselves

Two smaller additions close related gaps. In /config, a note now appears when other settings override a row, so a value that looks set but never takes effect has an explanation on the same screen. And when a settings change from /config or a similar command fails to save, a line under the prompt now says so, instead of the save failing silently and leaving you to trust a change that never landed. A resumed session gets the same treatment: a line appears when its earlier messages cannot be loaded, rather than the session resuming short with no word about what is missing.

Remote work stops losing files

The Remote Control fix in this release is the kind that matters once and matters a lot: stopping the server in worktree mode no longer deletes a running session's git-ignored files, such as .env. Remote MCP servers get a patience fix alongside it. Servers that drop, for example on a VPN drop, are now retried in the background for about a day, instead of staying disconnected after a longer outage. The rest of the fix list is long; fair samples are headless sessions no longer losing stored messages on resume, Grep paging results now sorted by file path so pages stop repeating or skipping entries, WebFetch refusing a URL over 2,000 characters before any permission prompt, and only the first 100 @ mentions in a message being looked up as files.

What to do

Two gateway and admin items round out the release: a x-claude-code-managed-settings-checksum gateway hint header, and an attributionProject setting that names a session's billing group. Most users will never touch either. The renderer also gets its plain name back: /tui classic names the non-fullscreen renderer, and /tui default still works. The previous release in this line, which tightened the sandbox, is covered in the page linked below.

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Claude Code 2.1.326 tightens the sandbox