Keep long Claude Code sessions going on their own,
have Claude remember what you taught it, and catch its usual slips.
English · 繁體中文 · Install · Features · Guide · Changelog
Long Claude Code sessions run into the same chores again and again. ctx-handoff takes care of them in the background, so in normal use you type no commands.
| When… | Without it | With ctx-handoff |
|---|---|---|
| the conversation is nearly full | You write a summary, /clear, paste it back |
A summary is written, the conversation is cleared, and a fresh one continues |
| you step away for an hour | Your next message rereads everything at full price | The cache is kept warm for up to about 4 hours |
| you repeat an instruction | The next conversation forgets it | It becomes a project note, and after 3 times a line in your repo |
you /clear or reopen Claude Code |
The new conversation starts blind | It's told where the last one stopped |
| Claude retries a failing command, or says "done" untested | You find out later | Claude gets a one-line nudge right away |
Requires Claude Code 2.1.287 or later. Tested up to 2.1.293.
claude plugin marketplace add cablate/ctx-handoff-mod
claude plugin install ctx-handoff@ctx-handoff-modStart a new session and type /handoff. You should see:
context 12034 / threshold 600000 (window 1000000)
Cache refresh on, refreshed 0/3 this idle period, timer not started
That's it. Update with claude plugin update ctx-handoff@ctx-handoff-mod, or turn on auto-update under Marketplaces in /plugin.
Note
Built for Claude subscriptions and long sessions (for example on a 1M-context model). On an API key, Bedrock or Vertex the cache lasts only 5 minutes, so run /handoff refresh off; everything else works. ctx-handoff is experimental, because Claude Code mods are still in early access.
ctx-handoff is a growing toolkit for long sessions: new features for working with Claude over hours land here.
At 600k tokens it waits for Claude and its subagents to finish, writes a summary that keeps verified work apart from unverified changes, lists what failed and the limits you set, ends with one next step, clears, and continues in a fresh conversation. Anything you type meanwhile is carried over.
Preferences, corrections and facts become one Markdown notes file per project, loaded into every new conversation. Routines you repeat become project skills. Only the new part of the conversation is read, by Sonnet 5.5 at low effort.
/handoff guard suggest turns a rule that keeps coming up into a check on tool calls, and nothing applies until you approve it. /handoff panel opens a panel above the prompt for guards, memories, rules, the latest notes update and every setting.
Everything is on the panel's Settings tab (/handoff panel, then 5), with where each value comes from and a reset button. The defaults are meant to be left alone. The guide lists every setting.
- No servers of its own. It runs on your Claude Code sign-in and sends nothing anywhere but Anthropic, like the conversation itself.
- Small, predictable costs. A handoff is one request; a cache refresh reads from cache at about a tenth of the input price; a notes update sends only the new part of the conversation. The nudges cost nothing.
- Plain files you own. Notes are Markdown in
~/.claude/projects/. Check what the mod may do withclaude plugin validate; the guide explains every entry.
- Guide: every feature in detail, settings, commands, permissions, supported environments, limitations, troubleshooting and uninstalling.
- Changelog: what changed in each version and why it's worth upgrading.
- Contributing: how to run it from source and send changes. Bug reports and ideas go in issues.
- Security: how to report a vulnerability privately.