The next session knows what the last one did.

A project where your AI has a name, and where a new session can say that name and what its last checkpoint recorded. [1] If you go further, you also get a handoff at the end of a session, and a script that checks what the handoff points at before the next session takes over. [1]

It picks up where rung 3 left off: the same named AI, now with a record that outlives the chat.

• Claude Code, version 2.1.275 or later, for the one-line install. [1]
• A git project. Outside a git repository the plugin says so and stops. [1] [3]
• git and a POSIX shell. On Windows that means Git for Windows. [1]
• Nothing else to install. There is no account, server, database or network call. [1]

In a Claude Code session, type this: [1]

/plugin install succession --marketplace thecontinuitylayer/succession

The README says that line comes from Claude Code's documentation and has not been run by the plugin's authors yet. The two lines below have been run. Use them in your shell if the first fails or your Claude Code is older: [1]

claude plugin marketplace add thecontinuitylayer/succession
claude plugin install succession@succession-marketplace

Once the plugin is installed, the next new session in a git project starts with a name, such as Sextant. It keeps nothing else: no charter, no log, no checkpoints, no handoffs. The session says its name and gives one line on how to turn handoffs on. The name stays for later sessions. /succession:seat --name Compass changes it. [1]

Stop here if you like. This already ends in the shown result below. The README says --name has not yet been run by a model in a real session; only its scripts are tested. [1]

1. 3Turn handoffs on. Open the project in Claude Code and type /succession:seat. Your AI shows you a one-page charter, proposes a name such as Sextant-001, and tells you what to send to confirm it (for example /succession:seat Sextant). The name you already have is kept and gets a number. [1] [3]
2. 4Work, and checkpoint. Type /succession:checkpoint when something is worth recording. [1]
3. 5Close. At the end of the day type /succession:close. Your AI writes a handoff and a short note of lessons, and the seat is empty. [1]
4. 6Next time. Type /succession:seat again. A script checks the previous handoff and shows a count, such as “9 claims, 4 pointers verified, 2 failed, 3 not checkable.” A failed claim does not stop you; it is shown to you and recorded. [1]

If a session ends without a close, you do not lose the seat. The next session start tells your AI its name and its last checkpoint, and its first reply opens with a line like “I'm Sextant-001. My last checkpoint says a.txt and b.txt are committed.” [1]

• A checkpoint is a dated note of what was done and where the proof is in git. [1] [3]
• A handoff is the note written at the close. Each claim in it is one line with a pointer: a commit, a file in a commit, or a log entry. [1]
• Everything is plain Markdown and JSON in a .succession/ folder in your project. [1]

It checks pointers. “Pointer verified” means the commit, the file or the log entry exists. It does not mean the sentence beside it is true. The README gives an example from its own tests: a model wrote “made before this tenure” beside a commit made during it. The pointer was verified and the sentence was false. [1]

It does not run your tests or any command. It does not judge claims it cannot check; it counts them as “not checkable” so you can decide what to look at. Work not yet committed has nothing to point to. [1]

The log is hash-chained, so editing a line is detected. That is evidence of tampering, not protection from it. [1]

The charter says what your AI may do without asking. Since 0.9.3, a charter already in the repository counts only if it is the plugin's template word for word, or you accepted its exact text in this clone. Otherwise the scripts and the session start say NOT in force, and the template applies. To accept, your AI shows you the whole file, and on your yes runs log-append.sh accept <sha256>. Under Claude Code's default permission mode, it asks you before that runs. A single line recording the acceptance is written in your project's .git/ folder, the one file the plugin writes outside .succession/. [2] [1]

Whatever a charter says, your AI asks first before it pushes, merges, tags, publishes or deploys; deletes files or branches, rewrites history or forces anything; changes dependencies; or touches secrets, credentials, billing or anything outside the repository. A handoff authorizes nothing. [1] [2]

The README gives uninstall, not a pause switch: [1]

claude plugin uninstall succession@succession-marketplace

Delete the project's .succession/ folder to remove its record. If you accepted a charter, the one acceptance line stays in .git/succession/; the README names it and does not say to delete it. [1]