Claude Code: Background Agents Now Survive Process Restarts and Auto-Updates
Claude CodeClaude Code 2.1.196 delivers a major reliability upgrade for background agents: long-running commands and workflows now persist through process stops, restarts, and auto-updates, including on Windows where background shells are handed off instead of killed outright. Workers interrupted by a daemon restart automatically resume from their last checkpoint the next time the agents view opens. Anthropic also enabled the streaming idle watchdog by default across all providers, protecting active sessions from silent stream stalls that previously required manual intervention to resolve.
Key Takeaways
- Background agents no longer lose work when Claude Code's process is stopped, restarted, or auto-updated: a persistent pain point for long-running autonomous coding tasks is now resolved.
- Windows users benefit from a shell hand-off model, where background shells are passed to a persistent process instead of being killed when the main Claude Code process exits.
- Auto-resume on daemon restart means workers interrupted by crashes or updates automatically continue from their last checkpoint when the agents view is reopened, with no manual intervention needed.
- The streaming idle watchdog is now on by default, aborting and retrying stalled response streams after 5 minutes of inactivity across all providers, including Bedrock and Vertex.
- Sessions now receive readable names at startup, improving discoverability when multiple background agents run in parallel and reducing the need to manually rename sessions.
- Clickable file attachments let users Cmd/Ctrl-click any attached file in chat to open its location in Finder or Explorer, reducing context-switching friction during file-heavy sessions.
Sources & Mentions
5 external resources covering this update
Background Agents Get a Reliability Overhaul
Claude Code 2.1.196, released June 29, 2026, resolves one of the most frustrating aspects of long-running agentic workflows: losing progress when something interrupts the session. With this release, background agents and their associated shell commands are now resilient to the full range of process interruptions: intentional stops, daemon restarts, and software auto-updates alike.
Long-Running Commands Now Survive Interruptions
Previously, if Claude Code's process was stopped or updated mid-task, background agents working on extended commands (file operations, build pipelines, test suites) would lose their state entirely, forcing users to restart from scratch. In 2.1.196, Anthropic changed this so that background commands and workflows survive the session's process being stopped, restarted, or auto-updated. On Windows specifically, background shells are now handed off to a persistent process rather than being killed when the main process exits.
This is particularly meaningful for users running extended autonomous coding sessions, CI integration workflows, or any multi-step background task that spans more than a few minutes. The improvement eliminates the class of frustrating half-finished agent runs that left codebases in an intermediate state.
Auto-Resume After Daemon Restarts
When a Claude Code daemon is killed by a system update, crash, or manual intervention, background agent workers now automatically resume from where they left off the next time the agents view is opened. There is no need to manually restart tasks or reconstruct lost context. Agents pick up at their last checkpoint transparently, reducing the operational overhead of long autonomous sessions.
Streaming Idle Watchdog Now On by Default
As a complementary reliability improvement, the streaming idle watchdog is now enabled by default for all providers. When an active response stream produces no events for five minutes, Claude Code automatically aborts the stalled request and retries. Previously this behavior was opt-in. Users who prefer to disable the watchdog can set CLAUDE_ENABLE_STREAM_WATCHDOG=0 in their environment. This change reduces the occurrence of silent session freezes during long agentic tasks, particularly when working with slower or heavily loaded model providers.
Additional Session Quality-of-Life Improvements
Claude Code 2.1.196 also ships two UX improvements to session management. Sessions now receive readable default names at startup, making it easier to identify and message individual conversations when multiple agents are running in parallel. File attachments in chat are now clickable: Cmd-click on macOS (Ctrl-click on Windows and Linux) opens the file's location in Finder or Explorer directly from the interface.