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 agentmods add instructions/agjs/gb10-clock-cap/agents-mdgit clone --depth 1 https://github.com/agjs/gb10-clock-capWrote 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/instructions/agjs/gb10-clock-cap/agents-md)<a href="https://agentmods.dev/instructions/agjs/gb10-clock-cap/agents-md"><img src="https://agentmods.dev/badge/instructions/agjs/gb10-clock-cap/agents-md.svg" alt="Measured on agentmods" 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 | $0.01319 | $0.01319 |
| Opus 5 | $0.00660 | $0.00660 |
| Sonnet 5 | $0.00264 | $0.00264 |
| Haiku 4.5 | $0.00132 | $0.00132 |
Grade B, and why
gb10-clock-cap AGENTS.md scanned grade B with 2 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 4d 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.
Asks for rootmediumPrivilege escalation
A mod that escalates privileges can change anything on the machine, not only the project.
2. `sudo -n nvidia-smi` works on every host. Clock control needs root. Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
| `--model` | `curl $API/v1/models \| jq -r .data[0].id` | Every request 404s | How it starts
The opening of the file, as written. The whole thing — 105 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Agent contract
You are running a GPU clock-cap evaluation on one or more inference hosts. This file is the spec. Follow it in order. Do not improvise the measurement method; the thresholds and orderings here exist because the obvious alternatives give wrong answers.
One-command path
./run.sh --hosts "hostA hostB" --api-host hostA --model MODEL_NAME --mode full
Emits results/summary.json. Read verdict.recommend and stop. Everything below is for when you need to deviate.
Inputs you must resolve before running
| Input | How to get it | Fails if wrong |
|---|---|---|
--hosts |
SSH aliases for every node with a GPU under test | Sampler collects nothing |
--api-host |
The node serving the OpenAI-compatible API | Benchmarks cannot connect |
--model |
curl $API/v1/models | jq -r .data[0].id |
Every request 404s |
--cap |
Default 2200. Only change with a reason | — |
--prefill-tokens |
Mean real prompt size on this system. Default 33000 | Probe does not represent the workload |
Get the mean prompt size from the server, do not guess:
vllm:request_prompt_tokens_sum / vllm:request_prompt_tokens_count
Preflight gates
run.sh --mode preflight checks all of these and refuses to continue on failure:
- Every host reachable over SSH without a password prompt.
sudo -n nvidia-smiworks on every host. Clock control needs root.nvidia-smi -lgc 0,<cap>applies andnvidia-smi -rgcreverts. Some GPUs do not support this. Confirm the reported clock actually changes.- API responds and
--modelappears in/v1/models. stdbufpresent on every host.
Measurement rules
These are not stylistic preferences. Violating any one of them produces a confidently wrong number.
Normalise throttle time to load seconds, never to uptime. Read cumulative throttle counters from nvidia-smi -q -d PERFORMANCE. Divide by time under load, which for a vLLM server is vllm:e2e_request_latency_seconds_sum. On an idle-heavy host the uptime denominator understates throttling by roughly 100x.
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.
- 4d ago First seen · 105 lines · 1,319 tokens per session scan B 62f2b6145195
gb10-clock-cap AGENTS.md is an instructions file published in the GitHub repository agjs/gb10-clock-cap (48 stars, last pushed 1mo ago), licensed MIT. It adds 1,319 tokens to every session, about $0.0066 per session on Opus 5. A static security scan graded it B with 2 findings (asks for root, makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
param-decomp CLAUDE.md
Instructions for goodfire-ai/param-decomp, covering repository guidance, repository-wide constraints and development.
Score-CAM CLAUDE.md
Instructions for tabayashi0117/Score-CAM, covering claude.md, what this repository is, layout, non-negotiable rules and paper fidelity.
oci-agent CLAUDE.md
Instructions for Netflix-Skunkworks/oci-agent, covering observational causal inference (oci) agent and rules.
ai4science-studio CLAUDE.md
Instructions for AMDResearch/ai4science-studio, covering claude.md, repository purpose, agent entry points (read these first), directory layout and model slug rule.
math-modeling-skill-dify CLAUDE.md
Instructions for MrSGSA/math-modeling-skill-dify, a project described as: 可直接拖入工作目录的数学建模Agent Skill与Dify九库多模态MCP桥，支持Codex、Claude Code及其他Agent.
gpu-ai-skills CLAUDE.md
Instructions for intel/gpu-ai-skills, covering claude.md, what this repository is, repository structure, commands and validation (required before any skill change).