Claude Code v2.1.325 was released on November 13, 2026. According to the official release notes, this update addresses three reliability and security issues critical for automated harnesses, alongside a tightening of default sandbox permissions. This review is grounded directly in the release notes and repository changes.
Three Core Fixes
1. Missing Recent Conversation Turns on Session Recovery
During long-running development sessions, unexpected terminations—terminal exits, system reboots, or out-of-memory kills—previously left the session event log in an incomplete state. When resuming an interrupted session, the most recent turns and their associated tool outputs were often missing.
This caused resumed agents to re-attempt edits blindly or operate under stale assumptions about the repository state. Version 2.1.325 ensures event logs are flushed more deterministically, preserving the immediate pre-crash turns when sessions are recovered.
2. Subagent Model Context and MCP Sizing Miscalculations
When configuring subagents with specialized models different from the primary session (for example, pairing an Opus orchestrator with Sonnet or Haiku worker agents), file reads and Model Context Protocol (MCP) tool outputs executed by subagents were incorrectly sized against the parent session's model context limits.
Additionally, subagents could have their designated models unintentionally overridden by planning-mode fallbacks (opusplan). Version 2.1.325 isolates context window calculations so subagents operate strictly within their own configured model parameters, eliminating premature context compacting and false token-overflow errors.
3. Security Vulnerability in zstd Response Decoding
A dedicated security patch was shipped addressing how Zstandard (zstd) compressed API responses and stream chunks are decoded. While specific CVE identifiers were not disclosed in the public notes, the update hardens the decompression pipeline against malformed compressed payloads.
Sandbox and Permission Defaults
Beyond bug fixes, several default operational behaviors changed in v2.1.325:
- Sandboxed Bash Prompts by Default: On macOS and Linux, sandboxed Bash commands now route through standard confirmation prompts rather than running silently. To retain unattended execution in CI environments, set
autoAllowBashIfSandboxed: trueinside the sandbox configuration block of~/.claude/settings.json. - Guardrails on Loops and Scheduled Tasks: Executing skills that declare
allowed-toolsor hooks from scheduled tasks or/looprequires explicit pre-approval viapermissions.allowor interactive prompts. - Worktree and Repository Boundaries: Read-only git invocations in untrusted directories outside the primary project now require user confirmation, while unauthorized worktree mutations (
git worktree add,move,remove) are blocked inside sandboxed shells.
Harness Verification Matrix
| Area | Verification Step | Expected Behavior |
|---|---|---|
| Session Recovery | Terminate an active task process and run claude --resume |
Pre-crash turns and tool call outcomes are fully retained |
| Subagent Sizing | Run subagents with divergent model tiers against large MCP outputs | Sizing and compaction trigger according to the subagent model, not the parent |
| CI Pipelines | Verify sandbox.autoAllowBashIfSandboxed in automated environments |
Unattended runs do not block on interactive permission prompts |
Source: Claude Code v2.1.325 Release.