菜单
Emit observability traces and decision logs that are safe to store and show. Use whenever an agent or loop logs its actions, reasoning, or tool calls — especially in production, shared logs, or anything a user might see. Trigger on any logging/tracing/audit step.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Capture a durable lesson every time the human corrects the agent, so the same mistake doesn't recur. Use immediately after any user correction, course-correction, or "no, do it this way". Also use at session start to load relevant past lessons. Trigger on corrections and on session start.
Turn a feature request or task into a written spec before any code is written. Use whenever the user asks to build, add, or implement something non-trivial (3+ steps or an architectural decision), or says "let's design", "write the spec", or "plan this". Produces the spec the human approves and the agent executes against.
Prove a change works before calling it done — build, test, lint, smoke-check, and report pass/fail. Use after any code change, before any commit/PR, and as the self-verify step inside an autonomous loop. Trigger whenever the agent is about to claim a task is complete.
Write a failing test first, then the minimum code to pass it. Use for any bug fix or new behavior.
Choose the cheapest sufficient retrieval strategy for a query before answering. Use whenever the system has multiple retrievers (vector, hybrid, graph, metadata, summary) and needs to decide how to fetch evidence for a question. Trigger on any retrieval/QA task where query types are mixed.
Keep a long agent session inside its context window without losing the plot. Use when a session is getting long, tool outputs are large, the agent starts "going off the rails" late in a task, or context usage is high. Trigger on big logs, long transcripts, or before a context-heavy step.
| name | safe-trace |
| description | Emit observability traces and decision logs that are safe to store and show. Use whenever an agent or loop logs its actions, reasoning, or tool calls — especially in production, shared logs, or anything a user might see. Trigger on any logging/tracing/audit step. |
Traces are how you debug and evaluate a non-deterministic system — but they must never leak private reasoning or secrets.
{"step": 3, "tool": "graph_search", "decision_reason": "impact question needs relationship traversal", "evidence_status": "sufficient", "tokens": 1840, "latency_ms": 220}
evals/ reads these traces out-of-band to score trajectory quality.