generated: '2026-08-22' method: searched source: https://github.com/HorizonRobotics/OE-Skills note: 'Provider-published Agent Skills, saved as an index rather than authored by API Evangelist. OE-Skills is Horizon Robotics'' own Apache-2.0 repository of structured skill packs for AI coding agents (Claude Code, Codex, Cursor) that operate the Horizon OpenExplorer toolchain end to end: HBDK compilation, HMCT/Plugin PTQ+QAT quantization, UCP on-board inference, performance and accuracy evaluation, and LLM compression. `horizon-router` is the top-level entry point and dispatches to a sub-skill by matching the user request against each skill''s `description` field in skill-index.json. The repository also ships the install path for the companion remote MCP server at https://mcp.oe.horizon.auto/mcp.' repository: url: https://github.com/HorizonRobotics/OE-Skills owner: HorizonRobotics license: Apache-2.0 default_branch: main pack_version: 0.2.0 last_push: '2026-07-01' stars: 14 install: instruction: 'Install this skill: https://github.com/HorizonRobotics/OE-Skills/blob/main/agent-setup.md' setup_doc: skills/horizon-robotics-oe-agent-setup.md skills_root: .horizon/skills targets: - Claude Code - Codex - Cursor modules: - id: hbdk title: HBDK skill_count: 2 - id: plugin title: Horizon Plugin skill_count: 7 - id: hmct title: HMCT / Quantization skill_count: 1 - id: ucp title: UCP / Runtime skill_count: 6 - id: horizon_tc_ui title: Horizon TC UI skill_count: 2 - id: llm title: LLM / 量化与压缩 skill_count: 4 skill_count: 29 skill_md_files_in_repo: 37 skills: - name: hb-analyzer-performance module: horizon_tc_ui version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon_tc_ui/hb-analyzer-performance/SKILL.md description: 使用 hb_analyzer 分析 J5/J6 模型性能。当用户提到模型文件（.onnx/.bc/.hbm）并想了解性能、延时、带宽、BPU利用率、瓶颈、实测、对比时触发。常见场景：评测模型、检查性能、找瓶颈、板端测试、分析 perf JSON。关键词："性能"、"延时"、"带宽"、"BPU"、"瓶颈"、"太慢"、"实测"、"对比"、"评测"、"hb_analyzer"。即使用户没明确说"分析"，只要提到模型+性能相关词就应该触发。 - name: horizon-tc-ui module: horizon_tc_ui version: 0.1.1 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon_tc_ui/horizon-tc-ui/SKILL.md description: OpenExplorer 工具链 CLI 包的集成 Skill。适用于处理 hb_compile、hb_model_info、hb_verifier、hb_analyzer、hb_config_generator、hb_eval_preprocess 等工具相关问题，包括 YAML 配置编写、ONNX/Caffe 模型编译、PTQ 量化、HBIR/HBM 产物生成、模型信息查看、精度验证、性能分析、评估数据预处理和板端部署验证。当用户提到 OpenExplorer、horizon_tc_ui、工具链、编译模型、生成 YAML、量化、calibration、HBIR、HBM、march、fast-perf、check 模式、板端推理、定点/浮点 BC 模型判断等场景时，应优先使用本 Skill。 - name: hbdk-manual module: hbdk version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/hbdk/hbdk-manual/SKILL.md description: HBDK4模型编译工具的使用指南索引，按任务场景组织 - name: j6-hbdk-compile module: hbdk version: 1.1.8 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/hbdk/j6-hbdk-compile/SKILL.md description: 通用模型编译 skill。基于 YAML 配置文件驱动，支持 ONNX 和 BC 模型格式。支持从 PTQ config 自动提取输入预处理参数。工作流：生成配置 -> 默认必须经用户确认 -> 再执行编译与报告。Use when the user mentions 模型编译、compile、导出 hbm、生成上板产物、pyramid 输入、resizer 输入。 - name: hmct-workflow module: hmct version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/hmct/SKILL.md description: "HMCT 模型转换与精度调优总入口。根据用户意图自动路由：\n (1) 提供了校准数据 → 调用 reference/run_build.py build 执行完整量化构建；\n\ \ (2) 未提供校准数据 → 调用 reference/run_build.py check 使用随机数据快速验证转换流程；\n (3) 用户希望进行精度调优 → 转交 j6-hmct-cosine-similarity-tuning\ \ SKILL 执行多阶段调优；\n (4) 用户希望进行单项精度 debug 分析（节点灵敏度、数据分布、累积误差等）→ 调用 hmct-debugger CLI 执行对应分析工具。\n 当用户提示词中出现\ \ HMCT、模型转换、模型量化、PTQ、精度调优、cosine similarity、节点灵敏度、数据分布、累积误差、debug 等关键词时应触发此 Skill。" - name: j6-hmct-cosine-similarity-tuning module: hmct version: 2.1.1 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/hmct/j6-hmct-cosine-similarity-tuning/SKILL.md description: "HMCT PTQ 精度调优工作流。自动按 INT8→INT16→dual-int16→FP16 逐级验证上限，\n 结合节点敏感度分析渐进回退，找到满足 Cosine Similarity\ \ 阈值（默认 >=0.99）的\n 最小混精度 quant_config，产出达标配置与调优报告。\n 触发关键词：HMCT 精度调优、PTQ 调优、量化精度、cosine similarity\ \ 不达标、混合精度。" - name: j6-plugin-adaptation module: plugin version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/plugin/j6-plugin-adaptation/SKILL.md description: 为用户的浮点 PyTorch 模型适配地平线机器人公司的 QAT 工具 `horizon_plugin_pytorch`。这是一个编排型 skill，必须按固定顺序依次调用：`j6-plugin-set-march` → `j6-plugin-insert-quant-dequant` → `j6-plugin-dynamic-block` → `j6-plugin-prepare` → `j6-plugin-set-fake-quantize`。 - name: j6-plugin-export module: plugin version: 1.0.1 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/plugin/j6-plugin-export/SKILL.md description: 使用 horizon_plugin_pytorch 的 hbdk4.export 将 QAT 模型导出为 HBIR IR 模块。以独立导出脚本的形式执行，不在训练或评测脚本中添加导出逻辑。 - name: j6-plugin-graph-diff module: plugin version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/plugin/j6-plugin-graph-diff/SKILL.md description: 对比两份 FX Graph 计算图文件，精确定位计算图结构和算子参数差异，并根据 diff 报告在原始模型代码中找到对应的差异位置。 - name: j6-plugin-hbdk-generating module: plugin version: 1.2.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/plugin/j6-plugin-hbdk-generating/SKILL.md description: 为基础网络结构生成从量化到编译的完整流程代码（set_march → Quant/DeQuant → 量化配置 → prepare → 校准 → QAT → export → convert → remove_io_op → statistics → compile HBM）。当用户需要同时覆盖量化和编译多个步骤时触发，如"帮我写量化编译全流程代码"、"Conv+BN+ReLU 量化部署"、"地平线量化编译"、"基础结构量化到 HBM"。即使用户没有明确说"全流程"，只要涉及从量化到编译的多个步骤都应触发。如果用户只需要量化或只需要编译，应路由到对应子 skill。关键词："量化编译"、"量化部署"、"全流程"、"set_march 到 HBM"、"地平线量化"、"Horizon 量化编译"、"基础结构量化"、"QAT 量化编译"。 - name: j6-plugin-model-check-result module: plugin version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/plugin/j6-plugin-model-check-result/SKILL.md description: 在 horizon_plugin_pytorch 量化配置检查中，读取并分析 prepare 产出的 model_check_result.txt，定位模型结构与 qconfig 问题（算子融合、共享模块、qconfig 异常、averaging_constant 非默认值、算子输入浮点检查）。 - name: j6-plugin-consistency-debug module: plugin version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/plugin/j6-plugin-consistency-debug/SKILL.md description: 定位 Horizon Plugin PyTorch 训练-部署一致性问题。当用户报告 QAT 模型训练侧精度正常，但 export/convert/compile/HBM 板端出现精度下降或掉点时触发。按 qat.pt → qat.export.pt → qat.bc → quantized.bc → hbm 分段定位，使用 QuantAnalysis 做逐层对比和敏感度分析。触发关键词："训练部署一致性"、"export 掉点"、"convert 掉点"、"compile 掉点"、"BC 精度偏差"、"HBM 掉点"、"板端精度下降"、"quantized.bc 不一致"、"pre_export 误差"、"一致性 debug"。注意：仅处理训练侧正常但部署侧掉点；若 calibration/QAT 本身精度就不达标，应路由到 j6-plugin-precision-tuning。bypass 模式下参见 bypass.md。 - name: j6-plugin-precision-tuning module: plugin version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/plugin/j6-plugin-precision-tuning/SKILL.md description: floatvscalib debug 结果解读与量化误差分析（截断误差与舍入误差分类、逐算子误差归因、fix-scale 建议、compare_per_layer_out 分析）。也处理 Horizon Plugin PyTorch 侧精度调优：calibration 后精度不达标、QAT 训练 loss 不收敛或精度崩溃、int8/int16/fp16 混合精度调优。聚焦 PyTorch 侧训练/校准精度，不处理 export/convert/compile 阶段的部署一致性问题。触发关键词："floatvscalib"、"debug 结果解读"、"截断误差"、"舍入误差"、"compare_per_layer_out"、"fixscale"、"fix-scale"、"calibration 精度"、"QAT 精度"、"校准掉点"、"QAT loss 不收敛"、"混合精度"、"int16"、"fp16"、"sensitivity"、"敏感度"、"QconfigSetter"、"FixedScaleObserver"、"HistogramObserver"、"精度调优"。注意：若训练侧精度正常但 export/convert/compile/HBM 掉点，应路由到 j6-plugin-consistency-debug。bypass 模式下参见 bypass.md。 - name: j6-ucp-hbm-infer module: ucp version: 1.1.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/ucp/j6-ucp-hbm-infer/SKILL.md description: Generate X86-side Python client code for hbm_infer, the Python SDK that connects to BPU boards via gRPC to deploy and run HBM model inference. Use this skill when the user wants to write Python code using hbm_infer client APIs — including HbmRpcSession (standard/flexible mode), HTensor for transmission optimization, output_config for output filtering, and related utilities like get_input_info, get_output_info, get_profile, compress_option, core_id. Triggers on writing or debugging Python inference scripts, optimizing data transfer between X86 and board, or any question about the hbm_infer Python client API. - name: j6-ucp-infer-generating module: ucp version: 1.1.1 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/ucp/j6-ucp-infer-generating/SKILL.md description: 提供基于 UCP 的模型推理 C++ 代码生成能力。当用户询问 UCP/DNN 推理接口怎么用、某个 API 的参数含义、如何加载模型、如何分配 tensor 内存、如何提交推理任务、Cache 同步怎么做，或要求只生成推理流程中某个模块的代码时使用。 - name: j6-ucp-perfetto-trace-analysis module: ucp version: 1.1.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/ucp/j6-ucp-perfetto-trace-analysis/SKILL.md description: Use this skill whenever the user wants to quickly locate, diagnose, or analyze likely performance bottlenecks in a UCP model inference Perfetto trace. Trigger on requests about UCP inference traces, `.pftrace` files, inference latency, slow inference, pipeline stalls, CPU/BPU gaps, OpInfer delays, dispatch or response delay, low effective occupancy, critical-path investigation, or trace-based performance triage in the UCP inference path. This skill is specialized for UCP / UCP-like inference-chain performance analysis and fast bottleneck localization, not generic Perfetto profiling unrelated to UCP inference. - name: horizon-router module: horizon-router version: 0.2.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon-router/SKILL.md description: OpenExplorer 工具链入口 Skill，处理 PTQ/QAT 量化编译、板端部署、性能精度评估等请求，并将任务路由到对应的 Horizon 子 Skill。 - name: board-detection module: horizon-router version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon-router/board-detection/SKILL.md description: 板卡硬件平台检测 Skill。当任务涉及板端运行、板端推理、远端 HBM、性能压测或 BPU 实测，且 .horizon/.env.board 不存在或不完整时触发。自动完成板卡 IP 获取、SSH 架构探测、平台信息写入。 - name: oe-package-detection module: horizon-router version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon-router/oe-package-detection/SKILL.md description: OE 包环境检测 Skill。当任务涉及普通 PTQ/QAT 量化、编译、部署等工具链操作，且 .horizon/.env.oe-package 不存在时触发。自动完成 OE 包路径定位、版本采集、本地环境匹配检查、GPU/CPU Docker 判定，并将结果写入 .env 文件。LLM 相关任务（llm_compression、LightCompress、LLM 量化/推理）请路由到 oe-llm-package-detection。 - name: oe-package-install module: horizon-router version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon-router/oe-package-install/SKILL.md description: OE 包本地安装 Skill。在 oe-package-detection 完成后触发，询问用户是否本地安装。若用户同意，则检测本地 Python / CUDA / PyTorch 版本，匹配 OE 包内 whl，创建 venv 全量安装，更新 .env 为 EXECUTION_MODE=local。 - name: oe-llm-package-detection module: horizon-router version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon-router/oe-llm-package-detection/SKILL.md description: OE-LLM 包环境检测 Skill。仅用于 LightCompress 等依赖 OE-LLM 包的任务。当任务涉及 LightCompress 量化实验（RTN/GPTQ/AWQ/SmoothQuant PPL 评测），且 .horizon/.env.oe-llm-package 不存在时触发。注意：llm_compression 独立包不走此流程，普通 PTQ/QAT 量化请路由到 oe-package-detection。 - name: oe-llm-package-install module: horizon-router version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/horizon-router/oe-llm-package-install/SKILL.md description: OE-LLM 包本地安装 Skill。在 oe-llm-package-detection 完成后触发，询问用户是否本地安装。若用户同意，则检测本地 Python / CUDA / PyTorch 版本，匹配 OE-LLM 包内 whl，创建 venv 全量安装，更新 .env 为 EXECUTION_MODE=local。 - name: j6-ucp-model-perf-eval module: ucp version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/ucp/j6-ucp-model-perf-eval/SKILL.md description: 自动化 hrt_model_exec perf 板端性能评测。触发条件：模型性能测试、benchmark、性能评估、板端测试、thread_num/core_id 参数扫描、吞吐量/延迟对比、远程部署 hrt_model_exec 运行 perf。 - name: j6-ucp-perfetto-trace-catcher module: ucp version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/ucp/j6-ucp-perfetto-trace-catcher/SKILL.md description: Capture UCP (Horizon Robotics inference SDK) Perfetto traces from a J6 development board. Use this skill whenever the user wants to grab UCP traces, capture Perfetto traces on a Horizon dev board, debug UCP scheduling issues with trace data, or pull .pftrace files from a remote board. Also trigger when the user mentions UCP performance analysis, UCP trace, hrt_model_exec tracing, or BPU trace capture on J6/S1000 platforms. Do NOT use for DSP trace or Chrome trace. - name: j6-board-monitor module: ucp version: 1.0.0 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/ucp/j6-board-monitor/SKILL.md description: J6 开发板资源监控与推理期间资源采集。当用户需要监控 BPU 占用率、DDR 带宽、内存使用时触发。支持三种场景：(1) Scenario A：CV 模型在指定帧率推理期间同步采集 BPU/DDR/内存数据；(2) Scenario B：独立监控板端硬件资源（无推理负载）；(3) Scenario C：LLM 模型循环推理期间同步采集资源数据，使用 simple_demo_request 保持模型持续运行。关键词：BPU 监控、DDR 带宽、内存使用、资源监控、LLM 推理监控、simple_demo_request、hrt_ucp_monitor、hrut_ddr、板端资源评估。也适用于 LLM 模型板端推理期间的资源监控（配合循环推理脚本）。注意：不要用 hbm_infer/gRPC 做高频推理监控（通信开销太大）。 - name: lightcompress-batch-quantize module: llm version: 2.0.3 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/llm/lightcompress/lightcompress-batch-quantize/SKILL.md description: 批量执行 LightCompress 量化实验并生成对比表。当用户要求对多个模型、多种量化方法或多种配置进行批量量化实验时触发。编排型 skill，循环调用 lightcompress-quant-explore。触发关键词：批量量化、批量实验、多模型量化、量化对比、LightCompress 批量、跑一组实验。 - name: lightcompress-quant-explore module: llm version: 2.0.3 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/llm/lightcompress/lightcompress-quant-explore/SKILL.md description: 执行单个 LightCompress 量化实验并生成精度报告。当用户要求运行量化实验、量化精度测试或 PPL 评估时触发。触发关键词：量化实验、LightCompress、PPL 评估、量化精度测试、RTN、GPTQ、SmoothQuant、AWQ、跑个实验、量化探索。支持模型：Qwen、InternVL、DeepSeek、Gemma4 等。 - name: llmcompression-add-model module: llm version: 2.0.3 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/llm/llm_compression/llmcompression-add-model/SKILL.md description: 为 llm_compression 框架新增 LLM/VLM 模型支持。当用户需要在 llm_compression/models/ 中接入新模型时触发。触发关键词：新增模型、接入模型、add model、llm_compression 模型、模型适配、模型注册、模型集成。注意：llm_compression 是独立于 OE 的包，使用独立的 conda 环境。日常操作（校准/编译/评测）请路由到 llmcompression-operations。 - name: llmcompression-operations module: llm version: 2.0.2 skill_file: https://github.com/HorizonRobotics/OE-Skills/blob/main/horizon/skills/llm/llm_compression/llmcompression-operations/SKILL.md description: llm_compression 日常操作 Skill。覆盖校准(calib.sh)、GPU精度评测(torch_eval.sh)、HBM编译(compile.sh)、板端精度评测(hbm_rpc_eval.sh)、量化分析(quant_analysis.sh)的标准用法。触发关键词：校准、calib、torch_eval、compile、hbm_rpc_eval、quant_analysis、LLM编译、LLM量化评测、板端推理评测。注意：llm_compression 是独立于 OE 的工具包，不走 OE/OE-LLM 包检查流程。新增模型支持请路由到 llmcompression-add-model。 files: - horizon-robotics-oe-skill-index.json - horizon-robotics-oe-agent-setup.md x-evidence: fetched: '2026-08-22' urls: - url: https://api.github.com/repos/HorizonRobotics/OE-Skills status: 200 - url: https://raw.githubusercontent.com/HorizonRobotics/OE-Skills/main/horizon/skill-index.json status: 200 - url: https://raw.githubusercontent.com/HorizonRobotics/OE-Skills/main/agent-setup.md status: 200 - url: https://raw.githubusercontent.com/HorizonRobotics/OE-Skills/main/README.en.md status: 200