Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add HorizonRobotics/OE-Skills --skill j6-plugin-consistency-debuggit clone --depth 1 https://github.com/HorizonRobotics/OE-SkillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/horizonrobotics/oe-skills/j6-plugin-consistency-debug)<a href="https://agentmods.dev/skills/horizonrobotics/oe-skills/j6-plugin-consistency-debug"><img src="https://agentmods.dev/badge/skills/horizonrobotics/oe-skills/j6-plugin-consistency-debug/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/horizonrobotics/oe-skills/j6-plugin-consistency-debug"><img src="https://agentmods.dev/badge/skills/horizonrobotics/oe-skills/j6-plugin-consistency-debug.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00092 | $0.05331 |
| Opus 5 | $0.00046 | $0.02665 |
| Sonnet 5 | $0.00018 | $0.01066 |
| Haiku 4.5 | $0.00009 | $0.00533 |
Grade A, and why
j6-plugin-consistency-debug scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 454 lines — stays where its author put it; the contents beside it link to each section on GitHub.
J6 Horizon Plugin PyTorch 训练部署一致性问题定位
核心原则
训练侧与部署侧不追求比特一致。先用稳定评测精度、可复现 badcase、多帧可视化结果确认问题真实存在，再按 qat.pt -> qat.export.pt -> qat.bc -> quantized.bc -> hbm / 板端 分段定位。不要只凭单帧数值差异下结论。
执行前门禁
在开始执行任何分析命令、修改脚本或给出结论前，先确认信息是否足够。
必需信息
| 信息 | 为什么需要 |
|---|---|
| 当前异常现象 | 用来判断优先排查 export、convert 还是 compile / 部署 |
| 已验证正常和异常的产物 | 必须知道哪一段开始掉点 |
| 至少一个稳定可复现 badcase，或可复现实验数据集 | QuantAnalysis、逐层对比和敏感度分析都依赖稳定输入 |
| 使用平台 / march | J6E/M 才优先考虑高一致性 QAT 策略 |
| plugin、profiler、hbdk 版本 | 部分接口和一致性策略有版本前提 |
compile / 板端问题时额外必需：
| 信息 | 为什么需要 |
|---|---|
| 输入输出前后处理、layout、部署适配说明 | compile / 板端问题经常来自用户侧适配差异 |
如果上述信息缺失且会影响下一步判断，先反问用户。不要在缺少 badcase、模型阶段结果或平台信息时直接跑完整链路或武断归因。
推荐信息
qat.pt/ QAT 模型、qat.export.pt、qat.bc、quantized.bc、hbm的可用路径。example_inputs或构造方式。- 评测脚本、后处理脚本、板端推理脚本。
model_check_result.txt、fx_graph.txt、compare_per_layer_out.txt、output_xxx_sensitive_ops.txt/.pt。- 用户已经尝试过的修复或规避方式。
先判断问题落在哪一段
这是最关键的一步。 错误的阶段判断会导致后续所有分析方向偏离。请严格按下面的现象-阶段映射表定位，不要凭直觉跳过。
| 现象 | 优先阶段 | 下一步 |
|---|---|---|
qat.pt 正常，qat.bc 异常 |
export | 先看 qat.export.pt 是否已异常，再比较 qat.export.pt 与 qat.bc |
qat.export.pt 已异常 |
pre_export / 查表转定点 | 用 QuantAnalysis(qat_pt, qat_export_pt, "pre_export") 定位 LUT / 图改写误差 |
qat.export.pt 正常，qat.bc 异常 |
export / HBIR 表达 | 用 QuantAnalysis(qat_export_pt, qat_bc, "export") 定位导出差异 |
qat.bc 正常，quantized.bc 异常 |
convert | 用 QuantAnalysis(qat_bc, quantized_bc, "convert") 做 badcase 和逐层对比 |
quantized.bc 正常，hbm 或板端异常 |
compile / 部署适配 | 优先检查输入输出、插入节点、前后处理、layout、padding、板端接入 |
| 不确定从哪开始 | 基线确认 | 先验证 quantized.bc 精度，再决定是否继续看 compile / 板端 |
正确 API 用法
fake quant 状态
普通导出和一致性分析前，QAT 模型应处于 eval + validation 状态：
from horizon_plugin_pytorch.quantization import FakeQuantState, set_fake_quantize
qat_pt.eval()
set_fake_quantize(qat_pt, FakeQuantState.VALIDATION)
FakeQuantState._FLOAT 只适合作为诊断手段，用来关闭伪量化验证 bc 推理流程或输入输出适配是否明显有问题；不要把它当成正常导出状态。
export 与 pre_export
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 454 lines · 92 tokens per session scan A e10c334c1f44
j6-plugin-consistency-debug is a skill published in the GitHub repository HorizonRobotics/OE-Skills (19 stars, last pushed 2mo ago), licensed Apache-2.0. It adds 92 tokens to every session and 5,331 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
debug-optimize-lcp
Guides debugging and optimizing Largest Contentful Paint (LCP) using Chrome DevTools MCP tools. Use this skill whenever the user asks about LCP performance, slow page loads, Core Web Vitals optimization, or wants to understand why their page's main content takes too long to appear. Also use when the user mentions…
systematic-debugging
Use when debugging a failing test, build error, or runtime issue that isn't immediately obvious. Guides a 4-phase root cause analysis instead of random fix attempts.
diagnose
Trace from a reproduced symptom to the source code that causes it. Pin the specific file and approximate line, rate confidence in the cause and clarity of the fix independently, and always propose a concrete fix.
repro-admin
Reproduce an EmDash admin UI bug. Attach a container, start the demo dev server, drive the admin with agent-browser using the dev-bypass session, and capture the reproduction as screenshots plus a replayable transcript.
log-error-digest
Analyze log files to troubleshoot errors, identify peak error periods, and produce error clustering, frequency statistics, and time distribution reports. Supports JSON, syslog, and Nginx formats with automatic detection. Use when a user uploads a .log file and asks to analyze errors, find patterns, debug issues, or…
byted-util-volcengine-detect-retry
An orchestration workflow for Volcengine Cloud Detect, a service that checks websites or network endpoints from test locations.