The Wire · Showcase
CLAUDE AGENT SDK DROPS THREE CRITICAL FIXES FOR ZOMBIE PROCESSES AND SILENT DATA LOSS
By RepoJournal · Filed · About Anthropic
The Python SDK shipped fixes for a cancellation bug that leaves subprocess children running after parent shutdown, a whitespace-stripping bug that corrupts JSON inside NDJSON payloads over 64KiB, and a security warning for shadowed tool permission callbacks.
Three separate pull requests hit the claude-agent-sdk-python repo addressing production hazards that have already hit real callers [1] [2] [3]. The first shields subprocess cleanup from cancellation errors, which means killed CLI children no longer linger as zombies when the surrounding task scope times out or fails [2]. The second patches a silent data corruption bug where NDJSON lines larger than one 64KiB read chunk lose whitespace wherever a read boundary lands, including inside JSON string values [3]. The third adds warnings when can_use_tool callbacks are shadowed by allowed_tools or bypassPermissions in the permission ladder, preventing security-critical tool restrictions from silently becoming dead code [1]. All three close genuine failure modes in production agent deployments. Plugin validation continues across both knowledge-work-plugins and claude-plugins-official, with five SHA bumps landing in the official repo and three in the knowledge-work repo, all pre-validated through the plugin validation workflow [4] [5] [6] [7].
Action items
- → Review and merge subprocess cancellation fix [ref:6] and whitespace corruption patch [ref:7] before next agent SDK deployment anthropics/claude-agent-sdk-python [immediate]
- → Audit any can_use_tool callbacks against allowed_tools and bypassPermissions configuration to catch shadowed security logic [ref:5] anthropics/claude-agent-sdk-python [plan]
- → Merge validated plugin bumps across knowledge-work-plugins and claude-plugins-official once scan jobs pass anthropics/claude-plugins-official [monitor]
References
- [1] Warn when can_use_tool is shadowed by allowed_tools or bypassPermissions ↗ anthropics/claude-agent-sdk-python
- [2] Shield subprocess cleanup from cancellation to stop zombie CLI children ↗ anthropics/claude-agent-sdk-python
- [3] Fix silent whitespace loss on NDJSON lines larger than 64KiB ↗ anthropics/claude-agent-sdk-python
- [4] bump(wix): 61f5c5f6 → e4a3c42c ↗ anthropics/knowledge-work-plugins
- [5] bump(huggingface-skills): 4948baec → 79512173 ↗ anthropics/claude-plugins-official
- [6] bump(hyperframes): 114d3191 → e8d19b1b ↗ anthropics/claude-plugins-official
- [7] bump(wix): 61f5c5f6 → e4a3c42c ↗ anthropics/claude-plugins-official
FAQ
- What changed in Anthropic on July 6, 2026?
- The Python SDK shipped fixes for a cancellation bug that leaves subprocess children running after parent shutdown, a whitespace-stripping bug that corrupts JSON inside NDJSON payloads over 64KiB, and a security warning for shadowed tool permission callbacks.
- What should Anthropic teams do about it?
- Review and merge subprocess cancellation fix [ref:6] and whitespace corruption patch [ref:7] before next agent SDK deployment • Audit any can_use_tool callbacks against allowed_tools and bypassPermissions configuration to catch shadowed security logic [ref:5] • Merge validated plugin bumps across knowledge-work-plugins and claude-plugins-official once scan jobs pass
- Which Anthropic repositories shipped on July 6, 2026?
- anthropics/claude-agent-sdk-python, anthropics/knowledge-work-plugins, anthropics/claude-plugins-official