Claude Code 2.1.267 caps effort level across providers

Claude Code 2.1.267 adds maxEffortLevel, a setting that caps the effort level on every provider, including Bedrock, Vertex and Foundry, while leaving users free to choose a lower level. The release notes on Github list it alongside a new --system-prompt-snapshot off flag and a long run of prompt-cache and session-resume fixes.

At a glance

• A marketplace entry path containing a backslash could bypass the containment check for fetched marketplaces on macOS and Linux, and that check now holds; unreadable managed allowlists admit nothing.
• The new --system-prompt-snapshot off flag renders the system prompt fresh on every request instead of reusing the prompt recorded in the conversation, which suits work on prompt text itself.
• Ten fixes address prompt-cache misses and dropped extended thinking in resumed sessions, on model switches with /model, and when MCP connectors reconnect or re-send tools mid-conversation.

The effort cap reads as an administrative control rather than a user feature: it sets a ceiling rather than a default, holds on Bedrock, Vertex and Foundry alike, and only ratchets downward. The rest of the release is dominated by prompt-cache accounting, which is where cost and latency in long Claude Code sessions actually live. Fixing the cases where a reconnecting MCP server rewrites the tool block likely matters more to daily bills than any new flag.

maxEffortLevel sits at the top level or per model under modelSettings

The setting caps the effort level on every provider, including Bedrock, Vertex and Foundry, and users can still select a lower level within that ceiling. It can be declared once at the top level of settings or attached to an individual model under modelSettings.

A related fix restores effort: frontmatter on custom commands, skills and subagents, which was ignored on models whose default effort is still pinned: Opus 4.7, Opus 4.8 and Fable 5. Separately, --system-prompt-snapshot off renders the system prompt fresh on every request instead of reusing the conversation's recorded prompt. The flag is aimed at iterating on prompt text.

Ten fixes target prompt-cache misses and dropped extended thinking

Several concern resumed sessions: rewriting the inline tool set when an MCP connector reconnects at a different moment, re-rendering tool descriptions rather than replaying the recorded ones, and rewriting earlier MCP tool announcements before their connectors reconnect. Resuming a print-mode conversation interactively also leaves the system prompt prefix intact.

Mid-session changes are handled too: a tool that disappears when an MCP server disconnects, or an upgrade removes it, previously rewrote the tool list and discarded earlier thinking, and a background worker forked from a conversation added EnterWorktree to the tool block. Switching models with /model re-sent every tool definition; commit and PR attribution text now arrives as a conversation note that updates on model changes.

Other resume work is separate from the cache: sessions with transcripts over 5 MB no longer drop parallel tool calls and their hook output, and resuming after /compact under -p --resume stops inserting a spurious "Continue from where you left off." turn.

Unreadable managed allowlists now admit nothing instead of everything

Managed allowedHttpHookUrls, httpHookAllowedEnvVars and allowedChannelPlugins now fail closed when unreadable. A marketplace entry path containing a backslash could bypass the containment check for fetched marketplaces on macOS and Linux, and Cowork scheduled tasks in the cloud failed at startup for organizations whose managed settings require sandboxing.

On machines whose managed settings require Claude apps gateway sign-in, Esc now closes the /login dialog. Expired AWS or Google Cloud credentials under a host app such as Claude Desktop retried ten times with a generic "request failed" before the re-authenticate error surfaced.

Gateway upstreams configured with forward_user_identity return a 429 as-is to the developer whose email was forwarded instead of failing over to the next upstream, so the proxy's per-user limits hold. On self-hosted runners, --use-anthropic-git-proxy is now reported to the server at registration, and each session still cloning through the legacy git proxy prints a warning.

VS Code, web and Claude Tag

In the VS Code extension, the extension host hung at 100% CPU when forking, editing an earlier message or rewinding a conversation whose transcript contains a cyclic parent link; accepting an edit on a file with CRLF line endings now works.

Claude Code on the web refreshes an expired GitHub Enterprise Server token automatically, and Claude Tag now says when an organization has run out of usage credits instead of reporting an invalid request. The release notes give no schedule beyond the version itself.