Claude Code v2.1.299 auto-summarizes idle chats over 200,000 tokens

The three releases span v2.1.299 through v2.1.301, published October 14–15, 2026 UTC, and the release notes add up to 329 items, with 190 in v2.1.299, 88 in v2.1.300, and 51 in v2.1.301. The idle summarization is the headline change, but the same version also alters what becomes invisible after a summary: skills can drop out of Claude's view and be restored with /reload-skills, and subagent names assigned before the last summary must be addressed by agent ID. A second summary trigger appears in v2.1.301 when switching to a smaller-context model, where the outgoing model summarizes the overflow first, and that summary may consume usage credits unless the version's fix applies.

The permission changes sit alongside the summarization work. In v2.1.299, rm commands assembled inside scripts Claude Code cannot read are no longer confirmed by default, and MCP permission rules no longer let mcp__a cover a__b or a_. v2.1.301 also fixes WorktreeCreate hooks that deleted a hook-created worktree at exit without confirmation even when unpushed commits might exist.

The piece's reading is that Claude Code is shifting management of conversation length and cache lifetime from the user to the tool itself. That reading rests on the idle summary in v2.1.299, the cache-write measurement by lifetime in v2.1.300, and the model-switch summary in v2.1.301; it would be falsified if the next three releases add nothing on summarization or caching. For context, the piece compared trigger settings as of October 17, 2026: Codex uses a token threshold via model_auto_compact_token_limit, and Aider uses a history token cap via --max-chat-history-tokens; neither setting list showed an idle-time trigger, while Gemini CLI was unconfirmed.