Skip to content
Pioneer Page PP markPPPIONEERPAGE

Pages from the advancing edge of computing.

Foundation models

Claude Code v2.1.325 release notes detail zstd decoding fix and subagent model routing corrections

The latest Claude Code update addresses a security flaw in compressed response handling and resolves issues with subagent model sizing and file read errors.

Claude Code v2.1.325 release notes detail zstd decoding fix and subagent model routing corrections
Depth
1 outlet · newest source 213h ago

Key points

  • Subagents on different models now correctly handle file reads and MCP results instead of misrouting to the main session model.
  • Sessions on third-party gateways and Foundry no longer stop with context length errors on every turn.

The update also addresses multiple stability and usability issues across the tool's various features and integrations.

How we got here

Previous versions of the tool reportedly suffered from subagents being moved to the plan-mode model when the setting was opusplan or haiku. Users also experienced sessions stopping with context length errors on third-party gateways instead of compacting, and shell commands running in launch directories after the working directory was deleted.

What is new

The release fixes subagents on a different model sizing file reads and MCP results with the main session's model. It also adds a spinner tip in VS Code about re-running notebook cells and clarifies error messages when organization policies block requests or when sandboxed git operations fail on protected files.

Reading between the lines

The update includes numerous interface fixes, such as correcting the welcome header display when agent names contain control characters and fixing the /stats command which previously showed an active days total that was one day too high. It also resolves issues with the /insights report writing unescaped values into HTML.

The open questions

The release notes state a security issue was fixed but do not provide a CVE identifier, CVSS score, or technical details on the exploit vector. For model evaluators, the lack of specific metrics regarding the "zstd-compressed responses" flaw limits the ability to assess the severity or potential data exposure. It is unclear if this affects only local decoding or server-side transmission, and no independent audit confirms the scope of the vulnerability.

Why it matters

  • Evaluators can now test subagent model routing accuracy using the fixed version, ensuring distinct models are not misrouted to the main session model during file reads.
  • The zstd decoding fix may impact how compressed data integrity is assessed in security audits of the Claude Code interface.
  • Users on third-party gateways can verify that context length compaction works correctly, preventing session interruptions during long-running evaluations.
  • The updated error messages for policy blocks provide clearer signals for compliance testing in organizational deployments.

Topics: Claude Code, Anthropic, security, subagents, zstd

Updated

The strongest doubt

Counterpoint from the sources

The release notes do not specify whether the zstd fix requires user action or if it is automatic upon update.

For readers

Developers

  • Subagent model routing is corrected, ensuring accurate file reads and MCP results.
  • Shell commands now fail with a message if the working directory is deleted, preventing silent errors.

Researchers

  • The zstd decoding fix may impact studies on data compression security in AI interfaces.
  • Context length compaction on third-party gateways is now functional, aiding long-context research.

Policy

  • Error messages now clearly indicate when organization policies block requests, aiding compliance monitoring.
  • The /fast command now correctly reports if fast mode is disabled by organization settings.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

Does the v2.1.325 update include a CVE for the zstd security issue?

No, the release notes do not list a CVE identifier or specific technical details about the zstd decoding security issue.

How does this update affect subagent model sizing?

It fixes an issue where subagents on a different model were incorrectly sizing file reads and MCP results with the main session's model.

Are there changes to Jupyter notebook interactions?

Yes, a spinner tip was added in VS Code and similar terminals to suggest re-running a notebook cell after Claude has read or edited it.

Connected stories

This storyDOD Contractor Linked Personal AI Agent to Secure Pentagon System for Two MonthsDOD Contractor Linked Personal ABlackstone Predicts High Junk Rating for Anthropic Post-IPO, Contradicting Prior ForecastsBlackstone Predicts High Junk RaAnthropic IPO Eyes $64B, Debuts Claude Meetings, Sheets, WhiteboardAnthropic IPO Eyes $64B, Debuts Anthropic's FlashProve generates cryptographic proofs at 110 tokens per second on H100 GPUsAnthropic's FlashProve generates

Reading receipt

  • Subagents on different models now correctly handle file reads and MCP results instead of misrouting to the main session model.
  • Sessions on third-party gateways and Foundry no longer stop with context length errors on every turn.

Still open: What is the specific technical nature of the zstd decoding security issue?

Still open: Does the fix affect data transmitted over the network or only locally decoded responses?

Time spent: 0:00 · Depth: Full

Sources

  1. Claude Code releases

Related stories

Background reading