Image: Anthropic / github.com/anthropics/claude-code Claude Code 2.1.198: Background Agents Now Open Draft PRs Automatically, and Claude in Chrome Is GA
Claude Code 2.1.198 ships automatic commit, push, and draft PR creation for background agents, hooks for agent lifecycle events, and the general availability of Claude in Chrome.
Claude Code 2.1.198 released July 1 with a change that closes one of the more annoying loops in working with background agents: they now finish their work without stopping to ask permission.
Background agents commit and open draft PRs
When a background agent finishes code work in a worktree, it now automatically commits the changes, pushes to the remote, and opens a draft PR. Previously it would stop and prompt you to do this manually. The automatic behavior kicks in for agents launched from claude agents that are doing code work.
This is consistent with how most teams actually use background agents: you kick one off, go do something else, and expect to come back to a PR ready for review, not a session waiting on input.
Hooks for agent lifecycle events
The Notification hook now fires for two new events: agent_needs_input and agent_completed. If you’ve built any automation around Claude Code’s hook system, you can now react to when a background agent pauses for input or finishes without manually polling session status.
Claude in Chrome is generally available
Claude in Chrome, which lets Claude agents drive a browser session, moved from preview to general availability in this release. It was in limited testing for a while before broader rollout. If you’ve been waiting to use it in production setups, it’s now stable.
Transient network errors no longer abort sessions
Previously, a brief network drop mid-response could kill a session entirely. The fix: ECONNRESET errors and similar transient failures now retry with exponential backoff instead of aborting. On spotty connections this should keep sessions alive through momentary drops.
Other fixes
A recurring “image could not be processed” error that was generating extra token usage on unprocessable images is fixed. Background tasks getting stuck showing “Running” in the web, desktop, and VS Code task panels are fixed. The /agents setup wizard was removed; the recommendation now is to use direct prompts or edit .claude/agents/ directly.
The Explore agent now inherits the main session’s model (capped at Opus) rather than running on Haiku. For complex exploration tasks, that’s a noticeable capability upgrade.
Full changelog: github.com/anthropics/claude-code