Claude Code v2.1.267: Prompt Cache Fixes, Effort Caps, and Security Patches

Anthropic released Claude Code v2.1.267 on September 9th. This patch focuses on prompt-cache reliability, model effort caps, and a security fix for marketplace paths on macOS and Linux. If you're hitting cache misses or using MCP servers with dynamic tools, this update is worth reading.

New Settings and Flags

• maxEffortLevel: a new top-level or per-model setting (under modelSettings) that caps the effort level on every provider, including Bedrock, Vertex, and Foundry. Users can still select a lower level, but can't exceed the cap.
• --system-prompt-snapshot off: renders the system prompt fresh on each request instead of reusing the conversation's recorded prompt — useful when iterating on prompt text.

Prompt Cache and Session Reliability

The release fixes multiple scenarios that caused prompt-cache misses or dropped context:

• A background worker forked from a conversation no longer adds EnterWorktree to the tool block mid-session.
• Mid-session MCP and plugin tools are now deferred in sessions without ToolSearch, preserving cache reuse.
• Switching models with /model no longer re-sends every tool definition (previous cache miss). Commit and PR attribution now arrives as a conversation note that updates on model changes.
• Resumed sessions no longer rewrite the inline tool set when an MCP connector reconnects at a different time.
• Resumed sessions now replay recorded tool descriptions instead of re-rendering them on the first turn.
• Prompt-cache misses and dropped extended thinking are fixed when a claude.ai connector's tools change between a session and its resume.
• Resumed sessions no longer rewrite earlier MCP tool announcements before connectors reconnect.
• A -p conversation resumed interactively no longer breaks cache by changing the system prompt prefix.
• Earlier reasoning is no longer dropped when an MCP server re-sends a tool the model already loaded.
• If a tool disappears mid-conversation (due to MCP disconnect or upgrade), the tool list is no longer rewritten, discarding earlier thinking.

Security and Auth Fixes

• Fixed a macOS/Linux containment bypass where a marketplace entry path containing a backslash could escape the check for fetched marketplaces.
• Managed settings for allowedHttpHookUrls, httpHookAllowedEnvVars, and allowedChannelPlugins now admit nothing when unreadable, instead of everything.
• Expired AWS/Google Cloud credentials no longer retry ten times with generic errors before showing the re-authenticate prompt.
• /login Esc behavior is corrected for machines requiring gateway sign-in.
• claude remote-control now re-registers when the server credential expires, avoiding dropped sessions.

Other Fixes

• /context and other local commands render properly on mobile clients.
• Shift+Enter and option+backspace work again after resuming a tmux/ssh session in agent view.
• The dim last-prompt header shows when scrolling up in fullscreen mode.
• Artifact publishes retry once after a dropped connection instead of reporting an unknown outcome.
• Artifact publish errors now include line/column when the page file isn't valid UTF-8 or contains U+FFFD.
• effort: frontmatter on custom commands, skills, and subagents is now honored on models with pinned default effort (Opus 4.7, Opus 4.8, Fable 5).
• The /diff panel no longer flashes “0 files changed” before settling, and its empty state is centered.

Who Should Update

Teams heavily using MCP servers, resuming long sessions, or managing remote control sessions will see immediate improvements in reliability and cache efficiency. The marketplace path fix is important for security-sensitive environments.