Claude Code 2.1.211 Fixes Prompt-Cache Billing Bug on Bedrock, Vertex

Claude Code 2.1.211 fixes a prompt-caching regression that hit Bedrock, Vertex, Mantle, and Foundry. Trailing system context was getting billed as fresh input tokens instead of hitting cache, per the v2.1.211 changelog.

That's the main cost lever in long agent sessions. Teams put stable content first in a system prompt and volatile content last specifically so the stable part hits cache. If the trailing context silently stopped caching on these gateways, the cost model broke with no error and no warning.

The changelog doesn't say how long the regression window lasted. It also doesn't list which prior versions were affected.

Anyone running agents on these gateways needs to check their upgrade history against their gateway's billing console. If your expected cache-hit rate was 80% and your spend doesn't reflect that, this is likely why.

Two other fixes shipped in the same cycle. Version 2.1.210 fixed isolation: 'worktree' subagents running git-mutating commands against the main repo instead of their isolated worktree. That defeats the entire purpose of the isolation flag. The same version fixed killed background sessions leaving a permanent git worktree lock behind.

Version 2.1.211 also fixed auto mode overriding a PreToolUse hook's ask decision for unsandboxed Bash. Permission hooks that looked like they were working were sometimes getting bypassed.

All three bugs share a pattern. None of them crashed or threw an error. Each just quietly stopped enforcing a guarantee it was supposed to enforce. Pull your billing for the affected period. Re-test your permission hooks now that ask actually floors the decision. Run a git-mutating command inside a worktree-isolated subagent and check where it lands.