Claude Code claims debug logs after 3-minute hang reports

• petergyang's hang report showed Claude Code sitting for 3 minutes with no clear status, and bcherny's reply said Anthropic is working on better responsiveness plus self-serve debug logs.
• The promise landed in the middle of a broader reliability push: ClaudeDevs' update thread said the team shipped 50-plus fixes the prior week and 60-plus more this week, including long-session and agent-loop changes.
• According to ClaudeDevs' stability notes, one recent bug let memory grow past 10GB when an stdio MCP server wrote junk to stdout, while ClaudeDevs' agent loop notes said 1M-context sessions now use their full window before throwing a prompt-length error.
• A separate desktop change in om_patel5's screenshot now exposes context-window usage directly, including message tokens, deferred MCP tools, autocompact buffer, and remaining free space.

You can check Anthropic's official changelog, see the 3-minute hang screenshot, and inspect the new context-window panel that breaks a session into messages, system tools, deferred MCP tools, and autocompact buffer.

The immediate news here is small but useful: a public complaint about silent hangs got a same-day engineering response.

In that thread, Peter Yang showed Claude Code stalled for 3 minutes while a PR-sized change sat in limbo. Boris Cherny, Claude Code lead at Anthropic, replied in bcherny's reply that the team wants the product to feel more responsive and plans to add debug logs so users can inspect what caused a hang themselves.

That is a narrow promise, but it points at a real product gap: the failure mode in petergyang's screenshot is not just latency, it is missing visibility.

Anthropic spent the last two weeks shipping bugfixes across the exact surfaces people keep complaining about: long-running sessions, auth, MCP, and terminal rendering.

The concrete items from the thread and official changelog are more interesting than the fix count:

• Stability: ClaudeDevs' stability notes said claude -p now accepts more than 10MB of piped stdin, resumes cleanly after a Mac wakes from sleep, and no longer lets one MCP stdout bug blow memory past 10GB.
• Agent loop: ClaudeDevs' agent loop notes said sub-agent summaries now hit the prompt cache, parallel shell calls survive a read-only sibling failure, and 1M-context sessions wait longer before hitting Prompt is too long.
• Auth: ClaudeDevs' auth notes added terminal code pasting for WSL2, SSH, and container setups where localhost auth breaks.
• MCP and rendering: ClaudeDevs' MCP notes and ClaudeDevs' rendering notes covered retry behavior in /mcp, image-preserving tool results, terminal scroll fixes, and better Windows CJK rendering.

Those fixes do not erase the complaints. levelsio's complaint described Claude Code as slow and risky after it took down a site, and om_patel5's Reddit repost amplified a longer complaint about instruction drift, slower turns, and auto-commit behavior.

The other useful change is a new Claude Desktop panel that finally surfaces how much context a conversation has burned.

The panel in om_patel5's screenshot does more than show one percentage. It breaks usage into:

• Messages
• System tools
• System prompt
• Skills
• MCP tools, deferred
• System tools, deferred
• Autocompact buffer
• Free space

It also shows plan usage on separate timers, including a 5-hour limit, weekly all-model usage, and model-specific buckets. For people stretching long chats, that is the first visible hint of when context loss is about to become a product behavior instead of a mystery.