v2.1.208

Source: Anthropic Claude Code Date: 2026-07-14 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.208

Summary

Claude Code v2.1.208 is a large maintenance release: a new opt-in screen reader mode (--ax-screen-reader), a CLAUDE_CODE_PROCESS_WRAPPER setting that routes every self-spawned Claude Code process through a required corporate launcher executable, and a long list of fixes to background-agent/daemon reliability, memory leaks (MCP stderr buffering, LSP documents, headless tool-result payloads), and performance (cached tool-pool assembly gives up to 7x faster tool rounds at high tool counts; transcript size cut up to 79x in edit-heavy sessions). It also extends the “prompt before catastrophic rm -rf” guard to commands hidden behind $(…)/backticks/<(…) substitution, closing an escape hatch in --dangerously-skip-permissions and auto mode.

Implications

• dev tooling / coding agents. CLAUDE_CODE_PROCESS_WRAPPER is another enterprise control point in the same vein as recent releases: it lets an org intercept every process the agent spawns, not just gate its actions. Continues the pattern of the harness treating the agent as a process to be sandboxed/instrumented rather than a user to be trusted on the honor system.
• The substitution-hiding fix for destructive commands is a real security tightening, not cosmetic — $(rm -rf ~) previously slipped past the same prompt that catches the plain form.
• Otherwise mostly reliability/performance debt paydown (memory leaks, background-daemon edge cases, Bedrock auth) — solid incremental hardening rather than a new capability, consistent with the maintenance cadence between the more consequential 2.1.20x releases.