byebai13/raolu-chengjing

Raolu-chengjing is a skill for AI coding agents, designed to work inside OpenAI Codex. Its purpose is to help an agent system decide what to do with the detours it takes during tasks. When an AI agent wants to use a particular tool, library, or workflow but the environment does not have it, the agent improvises a workaround and finishes the task anyway. The problem is that the same workaround often repeats the next time a similar task comes up. This skill provides a lightweight review process for examining those detours and asking: should this one be fixed permanently, or is it a one-off? To do that review, the skill reconstructs what the agent originally wanted to do, identifies the missing layer (whether a system tool, a credential, a runtime check, or a local data source was absent), and looks at how much time or quality the detour cost. It also asks whether the same gap is likely to show up again in high-value workflows. Based on that analysis, it picks the smallest durable fix: installing a tool, adding a preflight check to an existing skill, adding runtime detection, writing a runbook, or noting it as a new skill candidate. Creating a new skill is treated as a last resort, used only when the pattern is independent, stable, and recurring, and nothing existing comes close to covering it. Most detours do not need that level of response. The repository contains the skill definition file, an OpenAI agent metadata file, and three reference documents covering how to read a trace, how to pick the smallest fix, and how to verify a capability improvement. There is also one worked example using a PDF processing workflow. The code has an MIT license and is written to support both English and Chinese users.