Agents

How a Go security skill became an OWASP LLM review skill, then an iOS mobile security skill, by keeping the review method and adapting the risks to the product.

A practical starting setup for building production apps with coding agents across different stacks, issue trackers, and model families.

An old config backup lost a per-install channel record. Diagnose the selected target before overriding an update guard protecting newer database state.

Agent progress files can preserve unverified assumptions. Rebuild a small factual baseline, then check inherited claims against their original evidence.

Security transcripts mix observations with model guesses, and classifiers may stop the audit. Preserve provenance, checkpoints, and a provider-neutral handoff.

Keep Claude Code local, use SSH for remote commands, and submit long jobs to OpenPBS with explicit resources, paths, and completion checks.

A detached wrapper already owned the work. Launching another resumable driver raced in directory cleanup and raised OSError: Directory not empty.

Semicolon-separated build and launch commands ran a stale binary after nvcc failed. Check build status before execution and verify the runtime identity.

A scheduled agent left 67 failure dumps without a success receipt. Monitor verified outcomes and alert outside the agent when work stops.

An hourly agent kept hitting a known quota reset window. Persist a shared not-before deadline and defer model calls until a bounded recovery probe is due.