v2.1.221

Source: Anthropic Claude Code Date: 2026-08-04 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.221

Summary

This release adds sandbox credential masking (mode: "mask") so sandboxed commands on Linux/WSL see only a sentinel value for a credential file while a proxy substitutes the real value on egress, a prompt-audit subcommand for auditing prompts and tool descriptions written for older models, and changes background sessions to commit, push, and open draft PRs on their own following the repo’s CLAUDE.md conventions.

Implications

• credential-broker/hold-less: the credential-masking feature is a direct instance of this thread — the agent process never holds the real secret, only a sentinel, with substitution happening at the network boundary. Worth comparing against the Cursor Google Workspace-plugin signal in this same batch, which moves the opposite direction (agent holds standing OAuth scope).
• agent-orchestration: background sessions now default to commit/push/draft-PR behavior instead of leaving work uncommitted, tightening the loop for unattended runs.
• legibility/show-your-work: prompt-audit is a self-legibility tool aimed at agent builders auditing their own prompts and tool descriptions for model-generation drift.