[Release] anthropics/claude-code v2.1.92: v2.1.92

Claude Code v2.1.92 introduces enhanced security policies with remote settings refresh, an interactive Bedrock setup wizard for AWS integration, improved cost transparency with per-model breakdowns, and numerous bug fixes. Key improvements include better session naming with hostname prefixes, fixed subagent spawning issues, enhanced prompt cache feedback for Pro users, and performance optimizations for the Write tool. The release also streamlines the CLI by removing `/tag` and `/vim` commands while improving terminal compatibility and sandbox security.

Key Points

• •New `forceRemoteSettingsRefresh` policy enforces fail-closed security by blocking startup until remote managed settings are freshly fetched
• •Interactive Bedrock setup wizard guides AWS authentication, region configuration, credential verification, and model pinning from login screen
• •Per-model and cache-hit cost breakdown now available in `/cost` command for subscription users
• •Session names now default to hostname prefix (e.g., `myhost-graceful-unicorn`) with `--remote-control-session-name-prefix` override option
• •Pro users receive footer hints showing uncached token count when returning to sessions after prompt cache expiration
• •Fixed critical subagent spawning failure caused by tmux window killing/renumbering during long-running sessions
• •Resolved prompt-type Stop hooks failing on small fast model `ok:false` responses and restored `preventContinuation:true` semantics
• •Write tool diff computation improved by 60% for large files with special characters (tabs, `&`, `$`)
• •Fixed tool input validation for streaming array/object fields emitted as JSON-encoded strings
• •Linux sandbox now includes `apply-seccomp` helper in both npm and native builds for unix-socket blocking

Workflow Diagram

Concepts

Artifacts (2)

--remote-control-session-name-prefix <prefix>

forceRemoteSettingsRefresh: true
# When set, CLI blocks startup until remote managed settings are freshly fetched
# Exits if fetch fails (fail-closed security model)