菜单
CRG-first codebase navigation skill for all agents. Load when architecture must be traced, execution flow understood, impact radius determined, or entry points located before any implementation. Always queries code-review-graph before opening files — prevents blind file reads.
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
Lazy skill loading router for the Unity DOTS agent team. Use this to determine which Unity-Skills modules to load before spawning agents. Never load all modules — select only what the task requires.
Adaptive pipeline triage skill. Classify task complexity, blast radius, and domain in ≤8 file reads, then emit schema-valid workspace/triage.json that drives all downstream phase composition and agent selection.
Lightweight verification agent for small and medium complexity tasks. Mechanically runs the deterministic verification bundle from impl_result.json and emits verification_result.json. Does not design tests or edit code — replaces always-on tester for tiny/small/medium pipelines.
Structured protocol for using agentmemory as a recall layer during codebase investigation. Load when any agent would otherwise start with broad Read/Grep/Glob exploration. agentmemory accelerates orientation — it is NOT a substitute for reading current source files.
First skill using a duplicate name for testing duplicate detection.
Second skill using the same duplicate name for testing duplicate detection.
| name | codebase-understanding |
| description | CRG-first codebase navigation skill for all agents. Load when architecture must be traced, execution flow understood, impact radius determined, or entry points located before any implementation. Always queries code-review-graph before opening files — prevents blind file reads. |
| use-when | Load for any agent that needs to understand an unfamiliar system before working on it. Load at the start of bug investigation, feature design, or refactor planning. Load whenever an agent would otherwise start with broad Read/Grep/Glob. |
| do-not-use-when | Do not load when the task is a trivial change to a well-known file that requires no architecture investigation. Not needed for a quick comment fix or rename. |
| platforms | ["claude-code","codex","copilot","cursor","windsurf"] |
| task-categories | ["navigation","analysis","investigation","architecture"] |
| metadata | {"source":"internal","version":"1.0.0","tier":1} |
Query code-review-graph before reading any file. Never guess from filenames. Never grep blindly.
| Task | Agent | First CRG Call |
|---|---|---|
| Understand system architecture | architecture-agent | get_architecture_overview |
| Read an unfamiliar feature | codebase-reader | get_minimal_context |
| Investigate a bug | bug-investigation | trace_execution_flow |
| Plan a refactor | refactor-agent | get_impact_radius |
| Implement a new feature | feature-dev-agent | find similar feature |
| Tool | Use Case |
|---|---|
get_architecture_overview | Full system map before any analysis |
trace_execution_flow | Follow what runs in what order |
get_minimal_context | Orient around an unfamiliar symbol |
find_entry_points | Where does this feature start? |
list_related_symbols | What symbols participate in this feature? |
trace_callers_callees | Who calls this? What does it call? |
get_impact_radius | What breaks if I change this? |
trace_dependencies | What does this depend on? |
identify_shared_symbols | What is used by many callers? |
identify_extension_points | Where does new code attach? |
Maximum 8 files per investigation, each justified by graph evidence.
State "Running without CRG evidence" once. Use targeted Grep on known symbols. Read only files directly implicated. State your reasoning for each file opened.
## Summary
<one paragraph>
## Execution Flow
<numbered steps>
## Relevant Systems
<list with ownership>
## Root Cause / Behavior
<precise statement>
## Impact Radius
<what would change if this is modified>
## Recommended Next Step
<single concrete action>