2026-05-06 · Anthropic

v2.1.129

agentsmodels

read at source ↗ github.com

v2.1.129

Source: Anthropic Claude Code Date: 2026-05-06 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.129

Summary

Claude Code v2.1.129 shipped 22 changes focused on plugin infrastructure, terminal compatibility, and bug-fix density. Notable additions: --plugin-url <url> flag for fetching plugin archives from a URL at session start; CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE for silent background upgrades on Homebrew/WinGet; skillOverrides setting now functional with three visibility modes (off, user-invocable-only, name-only); and a fix for the CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY flag that was auto-enabled in 2.1.126–2.1.128. Also fixed: OAuth refresh race on wake-from-sleep, 1-hour prompt cache TTL being silently downgraded to 5 minutes, and /context dumping ~1.6k tokens of ASCII grid into the conversation.

Implications

  • The --plugin-url flag and skillOverrides granularity together indicate the plugin system is maturing toward a proper package distribution model — plugins can now be fetched from arbitrary URLs and their surface area can be controlled per-deployment, which matters for enterprise configurations.
  • The prompt cache TTL fix (1-hour silently → 5-minute) is a correctness issue that would have caused invisible cost inflation for long-running sessions; worth auditing any deployments that assumed full TTL was being honored.
  • Gateway model discovery reverting to opt-in after being auto-enabled for three releases is a sign of the model-picker UX still stabilizing — the feature exists but isn’t production-ready for all deployment configurations yet.
  • Feeds the Claude Code platform maturity thread: each release is accumulating fixes across subagent panels, MCP metrics, OTel, and enterprise policy enforcement, indicating the surface area being actively tested at scale is widening.

← all signals