Source Intelligence
Reading a new release v2.1.251 Analysing changes · 2/5 Sorting the findings · 1/4 steps 470 findings $18.91 so far

DisclaimerUnofficial, and not affiliated with Anthropic. Nearly all of this is read straight out of what ships: npm bundles, captured prompts, published docs. Anthropic's own notes go in verbatim, marked as theirs. The rest is my reading, and every entry carries the strings behind it. If one looks wrong, vote it down and say why.

All of v2.1.193 Home All releases olderv2.1.191 v2.1.195newer
Claude Code v2.1.193

REPL Context Poisoning Protection

What

When REPL sandbox code renders a global non-configurable (making it impossible to restore to a clean state), Claude Code now detects this, resets the VM context automatically, and shows a clear recovery message.

Details
  • A new VMContextPoisonedError class is thrown when this condition is detected.
  • The REPL context is reset and the user sees: "REPL context poisoned by non-configurable global; context reset" with a note that "global state (variables, registered tools) starts fresh."
  • Uses a WeakSet to identify boundary cap errors — the set is identity-based and not inspectable by attacker code, so a Proxy cannot fake a cap error to escape the walker.
Evidence

New error class (search for "REPL sandbox code made the global" or "VMContextPoisonedError")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

See this entry in the whole of v2.1.193 →