v2.1.196

Source: Anthropic Claude Code Date: 2026-06-29 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.196

Summary

Claude Code v2.1.196 is one of the larger point releases in recent history. The headlining changes: organization admins can now set a default model from the console (surfaced as “Org default” in /model), the streaming idle watchdog goes on by default for all providers (kills and retries any stream silent for 5 minutes), and a security fix prevents .mcp.json servers from being spawned via a repo’s own committed settings.json — untrusted workspaces now show these servers as pending approval rather than auto-starting them. The background-agent reliability work continues: workers killed by daemon restarts auto-resume, and a particularly sharp fix corrects a bug where waking a background job would permanently delete its conversation and re-run the original prompt from scratch.

Implications

• Agent autonomy. The background-agent fixes here collectively push toward “set it and forget it” reliability: auto-resume after daemon restarts, correct wake behavior, and cleaner status reporting in the agents view. Each patch reduces the failure modes that require human intervention to recover from.
• Dev tooling. The org-default-model feature is the first administrative control that lets an organization standardize model choice across all Claude Code users — a meaningful enterprise capability that shifts model selection from per-user preference to org policy. The /deep-research verifier fix (misreporting failures as “all claims refuted”) is also worth noting: a research tool that cried wolf on evidence quality was quietly undermining one of the more powerful workflows in the tool.
• Standards / security. The .mcp.json + settings.json trust fence tightening is a small but meaningful hardening: repos can no longer bootstrap their own MCP server approval through a committed config, closing a self-approval loop in shared workspaces.