Claude Code fixes 60-plus bugs in 1M-context sessions

• Anthropic says Claude Code got another 60-plus reliability fixes this week, after 50-plus the week before, with the biggest changes aimed at long-running sessions and heavier terminal use, according to ClaudeDevs' rollout thread and the official Claude Code changelog.
• The stability work includes a fix for stdio MCP servers that could push memory past 10GB, plus cleaner recovery after wake-from-sleep and large claude -p stdin runs, as ClaudeDevs' stability post lines up with v2.1.132 and v2.1.129.
• The agent loop changes are mostly about making long contexts less wasteful: ClaudeDevs' agent loop post says sub-agent summaries now hit the prompt cache and 1M-context sessions use more of their window before throwing “Prompt is too long.”
• Auth got a practical fallback for WSL2, SSH, and containers, where users can paste the OAuth code directly into the terminal, while proxy and token-race fixes landed in the same release train, per ClaudeDevs' auth post and the GitHub releases.

You can jump straight to the full changelog, cross-check individual GitHub tags like v2.1.129 and v2.1.132, and Anthropic had already queued a “What’s new in Claude Code” session on its Code with Claude agenda before this cleanup run ClaudeDevs' agenda post.

Anthropic's most concrete bug fix is the memory leak. ClaudeDevs' stability post says a misbehaving stdio MCP server could drive memory past 10GB, and v2.1.132 documents the same 10GB-plus RSS issue in the release notes.

The same bucket also covers bigger piped inputs, cleaner wake-from-sleep recovery, and output that appears reliably after thinking completes. That is unglamorous release engineering, but it is the stuff that decides whether a long agent run feels solid or haunted.

The agent-loop fixes cluster around token efficiency and failure handling:

• Sub-agent summaries now hit the prompt cache.
• The opt-in 1-hour prompt cache now gets honored correctly.
• Parallel shell calls keep running if a read-only sibling fails.
• 1M-context sessions use more of the available window before hitting the length limit.

The cache item matters because v2.1.129 says the 1-hour TTL had been silently falling back to 5 minutes. For people running long coding sessions, that is a real cost and latency fix, not just a changelog footnote.

The new login path is aimed at the boring environments where local OAuth callbacks break: WSL2, SSH sessions, containers, slow proxies, and IPv6-only devcontainers. Anthropic says users can now paste the OAuth code directly into the terminal when localhost is unreachable ClaudeDevs' auth post.

The changelog adds two more backend fixes around the same pain point: OAuth refresh races after wake-from-sleep in v2.1.129, and refresh-token handling plus full proxy-aware MCP OAuth flow fixes in the Claude Code changelog.

MCP servers that connect but fail tools/list now retry once and show a clearer /mcp status, which matches the wording in v2.1.132. Anthropic also says tool results with images and structured content now preserve those images, which is a nice quality-of-life fix for anyone using visual tooling in the terminal ClaudeDevs' MCP post.

The last batch is pure terminal polish, and it was probably overdue. ClaudeDevs' rendering post says Claude Code fixed too-fast scrolling in Cursor, older VS Code, and JetBrains terminals, improved Japanese, Korean, and Chinese text rendering on Windows, and stopped /-prefixed pasted text from getting hijacked as commands.