emmy: Skill for Claude Code

.agents/skills/tune-kernels/SKILL.md

tune-kernels is a skill for Claude Code, Codex from cloudrift-ai/emmy. It costs 75 tokens per session (4,943 once invoked), scanned A, original, Apache-2.0.

A workflow for improving machine-learning kernels, which are small performance-critical pieces of computation, for a model or compiler test set. It searches and records hardware-specific configurations and measurements.

In plain words
What is it for?
Use it to tune kernels, seed searches with proposed settings, compare search strategies, refresh GPU-specific goldens, and document per-kernel findings.
Why use it?
It helps diagnose slow or failing kernels and compare guided search with other search methods while keeping experimental and reviewed deployment records separate.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one. Also seen: positional $N argument; installed under .agents/ (shared by several agents).

This is cloudrift-ai/emmy's own configuration. It tells Claude Code and Codex how to work on emmy itself, so it is not a mod to install elsewhere. Copy it as a starting point and replace the rules that are about this project. Everything emmy configures →

Reuse

Borrowing it

Nothing to install: this file belongs to cloudrift-ai/emmy. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.

Copy the file
curl -O https://raw.githubusercontent.com/cloudrift-ai/emmy/main/.agents/skills/tune-kernels/SKILL.md
Clone the repo
git clone --depth 1 https://github.com/cloudrift-ai/emmy

Made for: Claude Code, Codex.

Wrote 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.

agentmods badge for tune-kernels

README.md
[![agentmods](https://agentmods.dev/badge/skills/cloudrift-ai/emmy/tune-kernels.svg)](https://agentmods.dev/skills/cloudrift-ai/emmy/tune-kernels)
Your own site
<a href="https://agentmods.dev/skills/cloudrift-ai/emmy/tune-kernels"><img src="https://agentmods.dev/badge/skills/cloudrift-ai/emmy/tune-kernels.svg" alt="Measured on agentmods" height="20"></a>
Per session 75 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 4,943 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5.1 $0.00075 $0.04943
Opus 5 $0.00037 $0.02472
Sonnet 5 $0.00015 $0.00989
Haiku 4.5 $0.00007 $0.00494

Measured yesterday against content hash b7ef6d12045b, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-06, from the pricing page.

Security

Grade A, and why

tune-kernels 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 yesterday.

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.

.agents/skills/tune-kernels/SKILL.md · 355 lines

How it starts

The opening of the file, as written. The whole thing — 355 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Tune Emmy kernels

Use one golden YAML format throughout the workflow. Keep its two trust levels separate:

  • A working golden is an untracked experiment artifact. Each structural target contains a realizations array; each realization has named dimension bindings, explicit registered input pins, and optional knob proposals, measurements, or ranking feedback written by emmy tune.
  • A canonical golden is reviewed deploy evidence under recipes/<model>/golden/, with one file per exact GPU model and compute capability. Every realization requires an explicit knobs mapping (empty for a forkless anchor) and paired positive deployable emmy_us / reference_us measurements with a named reference_backend. Never write search feedback into a canonical file. Model-agnostic hardware goldens remain under emmy/compiler/pipeline/search/goldens/.

pins and knobs are both registered knob mappings with different times of application: pins constrains candidate enumeration, while knobs records the winner measured inside that regime. FAST_MATH has no special YAML field; write it under pins exactly like any other input knob.

emmy tune --golden PATH rejects canonical repository paths because it updates its input. Copy a canonical file to a fresh _tune/<run>/working.yaml first. Do not commit a trace-created working golden automatically; leave that decision to the author or agent after validation.

--golden PATH is one flag on run, compile, tune, serve and eval golden: the golden YAML whose measured rows are the golden evidence that command deploys from, instead of the repository's per-card goldens. There is one deploy mechanism, the measured-evidence pick: a golden row is a measured row in the same index the tune DB feeds, never a separate authoritative tier. --realization NAME (run, compile, tune) selects one realization by exact name or an unambiguous substring; --strict-evidence (run, compile, serve) raises instead of deploying a prediction when a kernel has no measured evidence.

Read the full file on GitHub · 355 lines

Changes

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.

  1. yesterday Changed · +14 lines b7ef6d12045b
  2. 6d ago First seen · 341 lines · 75 tokens per session scan A 874c8a36bf9b

Subscribe to this mod's changes

tune-kernels is a skill published in the GitHub repository cloudrift-ai/emmy (80 stars, last pushed today), licensed Apache-2.0. It adds 75 tokens to every session and 4,943 once invoked, about $0.0004 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.

Related

Other skills, from other repositories