Strict sandbox and execution approvals
Anthropic altered the execution model for sandboxed Bash operations in Claude Code version 2.1.325. Sandboxed terminal commands now follow the standard interactive permission check rather than running automatically. Operators who want unprompted execution must explicitly enable auto-allow flags in their managed settings. The sandbox also prevents writes directed at agent definitions, skills, hooks, and configuration files within added folders, mirroring the existing restrictions of the primary project directory.
Anthropic also resolved permission validation bugs in command parsing. Compound Bash commands that previously avoided approval prompts now trigger permission requests. Furthermore, automated executions like loop routines and scheduled runs will stop to request permission when invoking any skill that specifies tool permissions or hooks, unless administrators pre-approve that skill by name.
Subagent modeling and environment boundaries
The release fixes inconsistencies in how subagents handle distinct model parameters. Previous versions inadvertently shifted subagents with independent model assignments onto planning models under specific configuration modes. Additionally, subagents assigned to alternate models previously relied on the main conversation model to calculate token sizing for file reads and tool outputs.
Repository isolation received further constraints during agent sessions. Agents operating inside isolated git worktrees are blocked from running background maintenance routines, bisect runs, and scalar setup commands. Untrusted external repositories also require manual confirmation before read-only git queries can proceed, unless the repository path is explicitly provided using the designated directory flag.
Our assessment
Teams running automated scripting or scheduled tasks with Claude Code should review whether their existing unattended workflows rely on sandboxed Bash commands executing without approval. If automated scripts stall after upgrading, engineers must evaluate whether enabling the explicit auto-allow sandbox flag adheres to internal safety guidelines, or whether critical skills should be individually added to the allowed permissions list instead.
Source: Anthropic original publication.
Read the original material
Anthropic (opens in a new tab)The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.





