What happened
Claude Code shipped version 2.1.208, a maintenance release with more substance than most patch notes. It fixes a bug that made long sessions look full when they weren’t, closes a gap in how the CLI catches destructive commands, and adds a screen reader mode for accessibility.
The headline fix: after the CLI auto-updates, the context window (the amount of conversation Claude can hold in memory at once) would briefly reset its display to 200,000 tokens. If you were resuming a long session, this created a false “100% context used” warning, even though your actual usage hadn’t changed. Several other reliability fixes landed alongside it, plus real performance gains for anyone running lots of MCP tools in one session.
Why it matters
A false “context full” warning during a long session is more than cosmetic. It can push you to compact or start a new session when you didn’t need to, losing conversation history you were relying on. This fix means the number you see after an update actually reflects reality.
The CLI also closed a safety gap. Claude Code already prompts before running catastrophic commands like rm -rf ~, even when you’re running with --dangerously-skip-permissions or in auto mode. Until this release, that check could be bypassed by wrapping the command in $(...), backticks, or <(...), a pattern that hides the actual command from a naive text match. Now those obfuscated forms trigger the same prompt as the plain version.
Performance also improved for heavier setups. Sessions with many MCP tools (tools from Model Context Protocol servers, which let Claude call external services) see up to 7x faster tool rounds, because the CLI now caches how it assembles the tool list instead of rebuilding it every time. Transcript files in edit-heavy sessions shrink by up to 79x, and several memory leaks that built up over long sessions, like MCP server logs and open LSP documents, are now capped instead of growing unbounded.
Who should care
Anyone running long Claude Code sessions, especially with background agents or forked conversations, will notice the context and memory fixes directly. If you run with --dangerously-skip-permissions for speed, the closed rm -rf loophole matters to you specifically, since that’s exactly the mode where a hidden destructive command would have run without a prompt.
Vim users get a small but real quality-of-life add: vimInsertModeRemaps lets you map two-key sequences like jj to Escape. Screen reader users get an opt-in plain-text rendering mode via claude --ax-screen-reader.
What builders should do next
Update to 2.1.208 and, if you rely on background or forked sessions, watch for the context indicator staying accurate through your next auto-update instead of falsely spiking to 100%.
If you use --dangerously-skip-permissions or run Claude Code unattended, this release is a reminder to review what that flag actually protects against. It’s a smaller list of unchecked commands than it was last week, but it’s still not zero, so treat the flag as a speed trade-off, not a safety guarantee.
If you’re running MCP servers and noticed tool calls feeling sluggish, this update is worth the upgrade on performance alone. The gains scale with how many tools you have configured, so the more MCP servers you’re running, the more you’ll feel it.
End of article