Claude Code v2.1.86 Released — Session ID Headers, VCS Fixes, 20+ Bug Fixes
Anthropic shipped Claude Code 2.1.86 tonight, and the release note that will get the most quiet celebration is one most users will never notice: a new X-Claude-Code-Session-Id header on every API request. For teams running Claude Code behind a proxy or observability layer — LiteLLM, Portkey, enterprise OpenRouter setups — this means you can now group all tool calls from a single session without parsing request bodies. It is a small addition that unlocks a lot of operational visibility at scale.
The version also delivers meaningful cost improvements for Bedrock, Vertex, and Foundry users. By removing dynamic content from tool descriptions, the prompt cache hit rate gets a significant boost — cache stays warm across requests rather than busting on every invocation. For teams running high-volume Claude Code workflows on cloud-managed inference, this directly reduces per-request cost. On the bug fix side, the marketplace plugin scripts are back (broken since 2.1.83), the --resume flag no longer throws cryptic errors on sessions from before 2.1.85, and Write/Edit/Read now correctly handle files outside the project root — a fix that anyone using conditional skills or editing ~/.claude/CLAUDE.md will appreciate immediately.
Rounding out the release: VCS hygiene improvements exclude .jj (Jujutsu) and .sl (Sapling) metadata directories from search and autocomplete, the /feedback command no longer crashes on long sessions with large transcripts, and the statusline correctly shows your own session's model when running multiple Claude Code instances simultaneously. Twenty-plus fixes in a single release, with enterprise infrastructure and cost efficiency as the clear theme.