
Claude Code v2.1.199 Major Updates
This page has been translated by machine translation. View original
This is Ishikawa from the Cloud Business Division. Claude Code v2.1.199 (released 2026-07-02) has been released. This release is primarily focused on bug fixes, centered on improving the stability of background agents and sub-agents, as well as reliability improvements around networking such as rate limiting, streaming, and SSL.
Normally, I would merge an update of this level with the next one, but since the behavior of Skills seemed to have clearly changed, I decided to write about it just in case. What do you all think?
Update Summary
v2.1.199 includes 24 changes. The majority are fixes related to background agent/sub-agent behavior and networking, and rather than notable new features, the content is centered on improving day-to-day stability when running multiple agents in parallel.
Notable Updates
- New Feature: When chaining slash skills like
/skill-a /skill-b do XYZ, up to 5 skills listed at the beginning are now loaded together, not just the first one. - Bug Fix: A critical bug on Linux where the background agent daemon was forcibly terminating itself and all running agents approximately every 50 seconds due to corrupted worker records has been fixed. Additionally, a problem where sub-agents were reporting API errors (such as usage limit reached) as successful results to the parent agent has also been fixed.
- Security: A problem where plan mode was not requesting confirmation for browser tool calls that modify state has been fixed, and only read-only
browser_batchcalls are now correctly auto-approved.
Update Details
New Features
- When chaining slash skills (e.g.,
/skill-a /skill-b do XYZ), up to 5 skills listed at the beginning are now all loaded, not just the first one. Previously, only the first one was loaded.
Improvements
- Temporary server-side rate limit errors (429) unrelated to usage limits are now automatically retried with backoff for subscribers, instead of failing the turn.
- The environment variable
CLAUDE_CODE_RETRY_WATCHDOGnow raises the default retry count for temporary non-capacity errors to 300, and removes the upper limit (15) ofCLAUDE_CODE_MAX_RETRIES. - In
claude agentssession rows, pull request links are now displayed in#Nformat without a redundant "PR" label.
Fixes
- Fixed a bug where the background agent daemon on Linux was forcibly terminating itself: When corrupted worker records were left behind due to abnormal termination, the daemon was forcibly terminating itself and all running agents approximately every 50 seconds.
- Fixed a bug where sub-agents were reporting API errors as successes: API errors such as usage limit reached were being treated as successful results, but errors will now be correctly reported to the parent agent.
- Fixed a bug where streaming responses were being discarded: When the API returned an overloaded/server error mid-stream after partial output, the entire response was being discarded, but partial responses are now retained and a notification is given that the response is incomplete.
- Fixed a bug where interrupted sub-agent work was being lost: Sub-agents interrupted by rate limits or server errors were silently failing without returning their partial work to the parent.
- Fixed a bug where SSL certificate errors were wasting retries: Errors such as TLS inspection proxies,
NODE_EXTRA_CA_CERTSnot being set, and expired certificates were consuming retries before showing how to address them. They now fail immediately and display hints for fixes. - Fixed a bug where corrupted configuration files were being discarded without recovery: Configuration files were being lost when resetting from the startup recovery dialog, but a backup is now taken before resetting.
- In addition, numerous minor bugs have been fixed, including cold start via SSH on macOS (a regression in 2.1.196), conflicts between
claude stopand restarts, stderr suppression forSessionStart/Setup/SubagentStarthooks, incorrect routing ofSendMessage, the reconnection page for Claude in Chrome, agent panel display, transcript bloat, and more.
Conclusion
v2.1.199 is a release centered on stability and reliability fixes that will have an impact on users who regularly use background agents and sub-agents. In particular, the bug where the daemon on Linux was periodically bringing down all agents, and the bug where sub-agent errors were being silently swallowed, had significant impact in environments running multiple agents in parallel. Why not update and give it a try?
References