Anthropic released Claude Code v2.1.97 on April 8, a major update to its terminal-based coding agent that addresses usability, security, and stability issues affecting both individual developers and enterprise deployments. The changelog lists over 40 changes spanning new features, hardened permissions, and bug fixes.

Focus View and Terminal Noise

The headline feature is Focus view (Ctrl+O), available in NO_FLICKER mode, which condenses tool output to a single line with edit diffstats rather than streaming the full content of every file read and tool call. For developers running agents on complex tasks that touch dozens of files, this reduces the terminal noise to what matters: the prompt, a summary of each tool action, and the final response.

The release also adds a refreshable status line that re-runs a custom status command every N seconds and a running subagent indicator in the /agents panel.

Bash Permission Hardening

The security changes focus on the Bash tool, the component that executes shell commands during agent sessions. According to the release notes, Anthropic tightened checks around environment variable prefixes and network redirects, reducing false permission prompts on common commands while restricting actual risk vectors. A separate fix addressed --dangerously-skip-permissions being silently downgraded after approving a write to a protected path, and another resolved managed-settings allow rules persisting after an admin removed them until a process restart.

Sandbox handling also improved: sandbox.network.allowMachLookup now takes effect on macOS, and auto mode now auto-approves sandbox network access prompts.

MCP and Infrastructure Fixes

A notable infrastructure fix addresses MCP HTTP/SSE connections accumulating approximately 50 MB per hour of unreleased buffers when servers reconnect, per the changelog. For long-running sessions connected to multiple MCP servers, this memory leak could degrade performance or crash the process.

Other MCP fixes include OAuth token refresh failures with ADFS and similar identity providers, and 429 retry logic that previously burned all attempts in approximately 13 seconds when servers returned small Retry-After values. Exponential backoff now applies as a minimum.

Resume and Transcript Reliability

The release fixes a cluster of issues around session resume (--resume): file-edit diffs disappearing for files larger than 10KB, cache misses from attachment messages not being saved to transcripts, messages typed while Claude is working not being persisted, and several picker UI issues. Transcript accuracy improved with per-block entries now carrying final token usage instead of streaming placeholders.

Production Signal

The changes collectively address the friction points that distinguish experimental tools from production infrastructure. Focus view reduces cognitive load during long agent runs. Bash hardening reduces the security review burden for teams deploying coding agents. MCP memory leak fixes prevent resource exhaustion in persistent sessions. Resume reliability ensures session history survives restarts.

For teams evaluating Claude Code alongside Cursor 3 (which launched its own multi-agent parallel execution this week) or GitHub Copilot Workspace, the update narrows the gap on enterprise readiness.