llama.cpp fixed a tool-call parser memory-safety bug as Pi pod tightened credential handling

TOP 3

1. llama.cpp published a tool-call parser memory-safety fix.
2. Pi pod released changes to credential removal and ticket handling.
3. A local media-search app for macOS surfaced on Show HN.

Agentic AI & Agent Skills

• Pi pod v0.1.9: The October 5 release updates the self-hosted Pi coding-agent sandbox. Its merged change description says removing a provider credential now removes its leased copy from running pods, session tickets are stored as digests, and credential-bearing query values are redacted from request logs. The maintainer also reports image lookup falling from 34.6–35.3 s to 0.33–0.35 s in a disposable-droplet test; that is a project test, not a general performance guarantee.

Tools & Open Source

• llama.cpp parser fix: The October 5 b11393 pre-release clears a stale current_tool pointer after a pending tool call is reset. The release notes say a subsequent TOOL_ID after TOOL_CLOSE could otherwise cause a use-after-free and a second free of the ID buffer. The notes do not establish exploitability; this is a memory-safety fix in a pre-release.
• SCM / Screen Memories: An October 4 Show HN post surfaced a local-first macOS media-search project. Its README describes vision search over photos and video scenes, separate OCR and Whisper dialogue search, and optional local chat over extracted evidence. These are the project’s stated capabilities, not independent test results.