JA EN
shiichan

Claude Code v2.1.267: maxEffortLevel Lets You Cap Effort Levels Across Providers

Claude Code Release ★★★ View Markdown
#anthropic#ai#devops#security

Hi, it's me, Shiichan! Today I've got the latest Claude Code release, v2.1.267, to share with you!

Claude Code Release github.com

What was announced?

Claude Code Release published version v2.1.267 of the Claude Code CLI. The headline additions are a new maxEffortLevel setting and a --system-prompt-snapshot off flag that re-renders the system prompt fresh on every request. On top of that, this release bundles a wide range of fixes — everything from mobile rendering, tmux/ssh reconnects, and session resume, to the VS Code extension, Claude Code on the web, and Claude Tag.

The story so far

Claude Code already let you pick how much a model should "think" (its effort level) per model, but there wasn't a way for an organization to cap that limit across the board. If a user picked a high effort level, costs and latency could balloon more than expected, depending on the provider — Bedrock, Vertex, Foundry, and so on.

The system prompt was also recorded once per conversation and reused from then on. That made it a little annoying to iterate on prompt wording, since you'd have to start a whole new session just to see your edits take effect.

What changes

Setting maxEffortLevel (either top-level or per model under modelSettings) now caps the effort level across every provider, including Bedrock, Vertex, and Foundry. Levels above the cap are no longer selectable, but users can still freely pick anything at or below it. That should make cost management a lot easier for organizations.

Launching with --system-prompt-snapshot off makes Claude Code render the system prompt fresh on every request instead of reusing the one recorded earlier in the conversation — handy if you're iterating on prompt wording.

Dive Deep

Beyond the two new settings, this release ships a long list of fixes. Here are some worth calling out:

  • Security: fixed a case where a marketplace entry path containing a backslash could bypass the containment check for fetched marketplaces on macOS and Linux
  • Session resume: fixed a spurious "Continue from where you left off." turn being inserted when resuming a session after /compact or another slash command ran via -p --resume; also fixed parallel tool calls and their hook output being dropped when resuming a large session (transcript over 5 MB)
  • Prompt-cache stability: fixed a number of cases where switching models, reconnecting MCP connectors, or tools appearing/disappearing mid-session caused prompt-cache misses or dropped extended thinking
  • UI and usability: fixed /context and other local command output rendering blank on mobile, shift+enter and option+backspace not working after reconnecting to a tmux or ssh session, and the /diff panel briefly flashing "0 files changed"
  • Also fixed: several VS Code extension issues (diff view not following the editor theme, edits failing on CRLF files, @-mentions dropping paths with spaces), plus fixes for Claude Code on the web and Claude Tag

Wrap-up

  • maxEffortLevel lets you cap effort levels across every provider, including Bedrock, Vertex, and Foundry, while still letting users pick anything at or below the cap
  • --system-prompt-snapshot off renders the system prompt fresh on every request, making it easier to iterate on prompt wording
  • The release also includes a marketplace path-containment security fix and a large batch of session-resume and prompt-cache stability fixes
  • Worth grabbing if you manage cost or model policy for an organization, iterate on system prompts often, or just use Claude Code day to day!

Related posts