Menu
Review repository code changes with a findings-first focus on correctness, regressions, compatibility, and missing validation. Use for code review, risk review, or model adaptation review requests.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
将 xhquant_llm 模型家族迁移到 xh2modelzoo 的 LLMConverter 架构，并完成 converter/convert_config、architecture 注册、examples 导出脚本与 hmonnx 验证闭环。当用户要求把已有 xhquant_llm 模型同步到 xh2modelzoo、补齐 xh2modelzoo 缺失模型支持、排查迁移后导出或推理失败时使用。
Build or update example scripts for adapted LLM and VLM models, covering export, generate, and optional quant flows in the style used by this repository.
Review whether an adapted transformers LLM or VLM model is compatible with the xhmodel_merak architecture. Focus on wrapper registration, graph_forward legality, cache flow, and output contracts.
Convert a Hugging Face transformers text model or multimodal model into an xhquant-callable and executable graph by following the wrapper, graph_forward, frontend, HMONNX, and runtime-contract patterns used by qwen3_legacy, qwen3_vl, and qwen3_5.
| name | code-reviewer |
| description | Review repository code changes with a findings-first focus on correctness, regressions, compatibility, and missing validation. Use for code review, risk review, or model adaptation review requests. |
| license | Apache-2.0 |
| metadata | {"author":"Lihui","version":"3.0"} |
Use this skill when the user asks for code review, risk review, regression review, or quality review.
Review in this order:
xhmodel_merak/xh_llm/models, check registration, wrapper replacement, cache flow, and output semantics..github/skills/xhquant_llm_reviewer/SKILL.md.When the user asks to review a specific adapted model, default review targets are:
xhmodel_merak/xh_llm/models/<model_name>/examples_merak/llm/<model_name>/ if it existsWhen the user asks for a broad LLM review, enumerate model directories under xhmodel_merak/xh_llm/models/ and review them one by one instead of giving a vague aggregate answer.
Return findings first, ordered by severity.
For each finding include:
Then include:
If there are no findings, say that explicitly and still mention residual risks or unverified areas.