3. Keeping a session healthy

/btw has no tools: it answers only from what's already in context. Press f in its overlay to fork the question into a subagent that does have tools. Claude Code also shows a recap on its own when you return after three or more minutes away, once the session has at least three turns.

• Back in full: system prompt and output style, project-root CLAUDE.md and unscoped rules (re-read from disk), auto memory, skill descriptions, a fresh git status, and the plan from plan mode.
• Back partly: up to 5 recently edited or read files (a file over 5,000 tokens returns as a path only), and invoked skill bodies capped at 5,000 tokens each and 25,000 in total, oldest dropped first.
• Back later: path-scoped rules and subfolder CLAUDE.md files, the next time Claude touches a matching file.
• Folded into the summary: the conversation itself. Your requests and key code stay, but detailed instructions from early on may not.
• Still running: background commands and background subagents.

Two levers: a # Compact instructions section in CLAUDE.md tells every compaction what to keep, and a SessionStart hook that matches the compact source runs again afterwards and can put context back.

On by default (autoCompactEnabled: true). On Opus 4.7 and later, Sonnet 5 and later, Haiku 5.5 and the Fable models, it triggers at about 967K tokens. Opus 4.6 and Sonnet 4.6 without the [1m] variant, and any model with CLAUDE_CODE_DISABLE_1M_CONTEXT=1, compact at the 200K boundary.

To compact earlier, set the auto-compact window, anywhere from 100K to 1M tokens. Strongest first:

1. CLAUDE_CODE_AUTO_COMPACT_WINDOW, for scripts and cloud environments.
2. claude --autocompact 500k, for one launch.
3. /autocompact 500k, saved per model in your user settings (v2.1.221 or later). /autocompact auto returns to the tuned window.
4. "autoCompactWindow": 200000 in a settings file, for every model.

A bare 200 means 200,000 in /autocompact, but 200 tokens in the environment variable. Whatever you set is capped at the model's real window.

Leave a big conversation for about 54 minutes and Claude Code may compact it while the one-hour prompt cache is still warm, so your next message sends a summary instead of reprocessing the whole conversation cold. It needs all of: an interactive terminal session, auto-compact on, the one-hour cache, feature flags reachable, and at least 200,000 tokens of context. Turn it off with "idleCompaction": false (v2.1.290 or later) or the row in /config.

The rule from the best-practices guide: if you've corrected Claude more than twice on the same issue, the context is full of failed approaches. /clear and start again with a better prompt.

/compact sends a separate request that rereads the conversation. While the cache is warm that costs a fraction of its size. After the cache expires it's the most expensive move you can make, which is why compacting a resumed old session hurts. /clear costs nothing. /rewind truncates back to a prefix that's already cached, so abandoning a dead end is cheaper by rewinding than by compacting. The main conversation's cache lasts 1 hour on a Claude plan within its included usage, and 5 minutes with an API key, usage credits, or a cloud provider.

• autoCompactEnabled: false or DISABLE_AUTO_COMPACT=1: no automatic or idle compaction, but /compact still works.
• DISABLE_COMPACT=1: no compaction at all, including /compact.
• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: compact at a lower percentage of the window. It can't raise the trigger.
• autoCompactWindow in a subagent's frontmatter: that subagent compacts earlier.
• "Autocompact is thrashing" means one file or output keeps refilling the window right after each summary, so Claude Code stopped retrying. Read the file in pieces, /compact with a focus that drops it, move the work to a subagent, or /clear.