Claude Code: Stalled Stream Auto-Recovery and Session Reliability Improvements
Claude CodeClaude Code v2.1.105 addresses a long-standing reliability issue where stalled API streams caused sessions to hang indefinitely. Streams now automatically abort after five minutes of inactivity and retry using a non-streaming fallback, ending the need for manual kills. The release also improves network error messaging, file write display for long single-line outputs, and /doctor diagnostics layout.
Sources & Mentions
4 external resources covering this update
Stalled API Streams Now Recover Automatically
One of the most persistent frustrations reported by Claude Code users has been sessions that silently freeze when the API streaming connection stalls mid-response. Previously, a dropped or unresponsive SSE stream would leave Claude Code in an indefinite waiting state — the only remedy being a manual process kill. Multiple GitHub issues documented the pattern: streaming starts normally, the first chunk arrives within seconds, then the connection goes silent for anywhere from one to thirty minutes before eventually timing out.
Claude Code v2.1.105 fixes this at the source. Stalled API streams now abort automatically after five minutes of no data, and the client immediately retries using a non-streaming request instead of hanging. This means users no longer need to interrupt and restart long-running tasks due to transient connection issues. The fix complements an earlier improvement in v2.1.98 that addressed similar timeout behavior for 429 rate-limit retries.
Network Error Messages Now Point Toward Resolution
Alongside the stream recovery fix, v2.1.105 improves how network errors are surfaced. When a connection error occurs, Claude Code now immediately shows a retry message — replacing the silent spinner that previously left users wondering whether the tool had frozen or was actively retrying. This is a small but meaningful change for anyone working in environments with intermittent connectivity.
File Write Display Handles Long Single-Line Outputs
Long single-line writes — such as minified JSON blobs or base64-encoded data — previously caused the file write display to paginate across many screens, making it difficult to see what actually changed. Claude Code v2.1.105 now truncates these outputs in the UI, keeping the write display readable without losing the underlying data.
/doctor Gets a Status Icon Layout
The /doctor diagnostic command received a layout update in this release, adding status icons for each check and flagging items that Claude Code can fix automatically. This makes it easier to triage environment configuration issues without reading through dense text output.
Skill Description Handling Expanded
Skill descriptions are now allowed up to 1,536 characters (up from 250), with a startup warning when descriptions are truncated. This gives plugin and skill authors more room to provide context that helps Claude select and invoke skills accurately.
Bug Fixes Across Terminal, Rendering, and Agent Workflows
The release includes more than thirty bug fixes addressing a wide range of edge cases:
- Fixed garbled Bash output with clickable file links
- Fixed alt+enter and Ctrl+J newline insertion in various terminals
- Fixed queued user prompts disappearing in focus mode
- Fixed one-shot scheduled tasks re-firing repeatedly
- Fixed inbound channel notifications being silently dropped
- Fixed MCP tools missing on the first turn of headless sessions
- Fixed 429 rate-limit errors showing raw JSON instead of a readable message
- Fixed crash on resume with malformed text blocks
- Fixed washed-out 16-color palette over SSH and mosh connections
- Fixed stdio MCP servers with malformed output hanging sessions
- Fixed marketplace auto-update breaking the official marketplace