{"id":77010,"plugin_id":"plugins_6ab41ed5b8c4819199eb5bbf91ce25f5","kind":"skill","collection_source":"plugin_package","comparison_source":null,"observed_at":"2026-11-03T00:12:18.727Z","digest":"49e95946a5b2a11b180f1ae83688017324321249fcf08377a76991fdf6fe25ce","against":null,"payload":{"description":"Make code using LLM APIs, structured outputs, tool calls or retrieval safer, more reliable and testable.","included_files":[],"name":"harden-llm-integration","skill_md_contents":"---\nname: harden-llm-integration\ndescription: Make code using LLM APIs, structured outputs, tool calls or retrieval safer, more reliable and testable.\n---\n\n# Harden Llm Integration\n\nUse this skill only for its stated purpose. Follow the shared source-of-truth, permission and evidence rules in `../../references/workflow-contract.md`. Read `../../references/llm-integration-patterns.md` for detailed method when necessary.\n\n## Procedure\n1. Verify the actual model client/SDK version, contract and available API capabilities.\n2. Separate deterministic validation/state changes from model output; treat outputs and tool results as untrusted.\n3. Add typed outputs, bounded timeouts/retries where safe, cost/latency tracking and redaction.\n4. Test malformed outputs, rate limits, cancellation, unauthorized side effects and adversarial tool content.\n5. Record what was tested versus what requires a production evaluation.\nFor large RAG/agent architecture, route to that specialist domain when appropriate.\n\n## Handoff rule\nState actual artifacts or grounded code findings and validation status. Never invent file paths, executed tests, tool access, or completed subagent work. Keep output proportionate to the user request.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}