A skill your agent uses when a batch of env-gated (ifdef MPKDSV3 / os.environ-controlled, default-OFF) MPK optimization levers needs to be consolidated into a single clean code path for a PR…
Install the "mpk-lever-cleanup" agent skill from https://github.com/mirage-project/mirage/tree/mpk/.claude/skills/mpk-lever-cleanup into .claude/skills/mpk-lever-cleanup/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "mpk-lever-cleanup", then confirm the skill loads.
Claude Code copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
Type this inside Codex. $skill-installer <name> installs a curated skill from openai/skills. The installer writes to $CODEX_HOME/skills (default ~/.codex/skills). Restart Codex if the skill does not show up.
skills CLI
$ npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a codex
Project install goes to .agents/skills/; add -g for ~/.codex/skills/.
Install the "mpk-lever-cleanup" agent skill from https://github.com/mirage-project/mirage/tree/mpk/.claude/skills/mpk-lever-cleanup into .agents/skills/mpk-lever-cleanup/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "mpk-lever-cleanup", then confirm the skill loads.
Codex copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
skills CLI
$ npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a cursor
Project install goes to .agents/skills/; add -g for ~/.cursor/skills/.
Install the "mpk-lever-cleanup" agent skill from https://github.com/mirage-project/mirage/tree/mpk/.claude/skills/mpk-lever-cleanup into .cursor/skills/mpk-lever-cleanup/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "mpk-lever-cleanup", then confirm the skill loads.
Cursor copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
--scope user (default) or --scope workspace; --path is the subfolder of the repo that holds the skill; --consent skips the security confirmation prompt.
skills CLI
$ npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a gemini-cli
Project install goes to .agents/skills/; add -g for ~/.gemini/skills/.
Install the "mpk-lever-cleanup" agent skill from https://github.com/mirage-project/mirage/tree/mpk/.claude/skills/mpk-lever-cleanup into .gemini/skills/mpk-lever-cleanup/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "mpk-lever-cleanup", then confirm the skill loads.
Gemini CLI copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
Installs for Copilot at project scope by default; add --scope user for a personal install. Preview a skill first with gh skill preview. Needs GitHub CLI 2.90.0 or later (public preview).
skills CLI
$ npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a github-copilot
Project install goes to .agents/skills/; add -g for ~/.copilot/skills/.
Install the "mpk-lever-cleanup" agent skill from https://github.com/mirage-project/mirage/tree/mpk/.claude/skills/mpk-lever-cleanup into .github/skills/mpk-lever-cleanup/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "mpk-lever-cleanup", then confirm the skill loads.
GitHub Copilot copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
skills CLI
$ npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a opencode
OpenCode documents no install command of its own. Project install goes to .agents/skills/; add -g for ~/.config/opencode/skills/.
Install the "mpk-lever-cleanup" agent skill from https://github.com/mirage-project/mirage/tree/mpk/.claude/skills/mpk-lever-cleanup into .opencode/skills/mpk-lever-cleanup/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "mpk-lever-cleanup", then confirm the skill loads.
OpenCode copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
Facts
Skill name
mpk-lever-cleanup
GitHub stars
2.5k
Token cost
~2.2k tokens
SKILL.md length
1,053 words
Files
1
Skills in repo
24
Repo updated
First seen
Licence
Apache-2.0
At a glance
A skill your agent uses when a batch of env-gated (ifdef MPKDSV3 / os.environ-controlled, default-OFF) MPK optimization levers needs to be consolidated into a single clean code path for a PR…
Works in 7 steps: Enumerate all gates → Classify every gate (VERIFY each one —… → Codex-vet the classification + refactor… → …
A batch of env-gated (ifdef MPKDSV3 / os.environ-controlled
SKILL.md covers Procedure (7 steps, in order) and Key pitfalls (all hit in…
Calls git
What it does
Mpk Lever Cleanup is an agent skill from mirage-project/mirage. Use when a batch of env-gated (ifdef MPKDSV3 / os.environ-controlled, default-OFF) MPK optimization levers needs to be consolidated into a single clean code path for a PR: hard-wire every winning lever as the default, delete the legacy else branches, remove the env vars that select new-vs-old logic, revert dead levers that measured KILL/NULL/regress, delete diagnostic probes, then commit one clean version. Applies to the wrap-up stage where the optimization work has settled and is being merged to mainline. Not…
Its SKILL.md is about 2.2k tokens, which your agent loads only when the skill is triggered. It is a single SKILL.md file with no bundled scripts.
It sits in Development, covering Code quality. The repository describes itself as: Mirage Persistent Kernel: Compiling LLMs into a MegaKernel. The licence is Apache-2.0.
When your agent uses it
A batch of env-gated (ifdef MPKDSV3 / os.environ-controlled
Delete the legacy else branches
Remove the env vars that select new-vs-old logic
Revert dead levers that measured KILL/NULL/regress
Example prompts
“/mpk-lever-cleanup”
Workflow steps
7 steps, taken from the step headings in SKILL.md.
Read from SKILL.md and the folder at commit f9eb70c. It shows what the files ask for, not the result of running them.
Tool permissions
Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.
From allowed-tools in the SKILL.md frontmatter.
Runs code
Shell commands in SKILL.md call:
git
From the folder's file list and the shell code blocks in SKILL.md.
Network
No URLs in SKILL.md. Its commands use git, which can reach the network depending on how they are called.
From URLs in SKILL.md, links to its own repository left out.
Credentials
Names no API keys, tokens, secrets or passwords.
From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.
Context cost
Mpk Lever Cleanup loads about 2.2k tokens when it runs. Until then it costs about 165 tokens; SKILL.md has 1,053 words of instructions outside code blocks.
Always· name and description, kept in context so the agent knows when to use it
~165
When it runs· the whole SKILL.md, loaded when a task matches
~2.2k
Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.
Safety
Auto-check passed
The automated check found no risky patterns in SKILL.md.
Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.
Download SKILL.mdSave it as .claude/skills/mpk-lever-cleanup/SKILL.md (or your agent's skills folder).
name
mpk-lever-cleanup
description
Use when a batch of env-gated (`#ifdef MPK_DSV3_*` / `os.environ`-controlled, default-OFF) MPK optimization levers needs to be consolidated into a single clean code path for a PR: hard-wire every winning lever as the default, delete the legacy `#else` branches, remove the env vars that select new-vs-old logic, revert dead levers that measured KILL/NULL/regress, delete diagnostic probes, then commit one clean version. Applies to the wrap-up stage where the optimization work has settled and is being merged to mainline. Not for the exploration phase (levers should stay env-gated default-OFF there) or for runtime/execution-model changes.
tags
mpk, cleanup, refactor, pr
related_skills
mpk-internals, add-mpk-task
version
1.0.0
MPK Lever Cleanup — consolidate env-gated optimizations into one clean path
During exploration, every MPK performance optimization is an env-gated,
default-OFF lever (#if MPK_DSV3_XXX + the #else legacy path +
os.environ.get(...) → -D injection in persistent_kernel.py). Once the work
has settled and is headed for mainline, the levers must be consolidated into a
single path: hard-wire the winners as the default, delete the legacy code,
remove the control variables. This skill is the complete procedure — plus the
pitfalls actually hit — from that wrap-up refactor.
Core mindset: this is a refactor that intentionally changes the default
build — the default path flips from "the safe legacy logic" to "the
optimized path". The usual "default build must stay byte-identical" commit
gate is therefore deliberately waived here; that is the whole point of
this commit.
List both the #ifdef gates in the .cuh files and the os.environ/-D
injections in persistent_kernel.py.
2. Classify every gate (VERIFY each one — never assume default-OFF)
Class
Action
How to decide
WIN
Hard-wire ON: remove the gate + delete the #else legacy + remove the env injection; keep the geometry guard (TP8/mbt/workers)
Lever already committed in git log + a WIN row in experiment_history
DEAD
Fully revert (delete all of its code; roll back any ABI it changed)
A KILL/NULL/REGRESS row in experiment_history
DIAGNOSTIC
Delete (probe/no-op/poison/xor — not a lever)
Name contains PROBE/NOOP/POISON/XOR; used only for measurement
ALREADY-ON
Confirm it is still present; keep it as the unconditional default (do not add a gate)
grep persistent_kernel.py: is it already injected via -D unconditionally?
LEAVE-UNTOUCHED
Do not touch
A fallback outside this path (e.g. the TP<8 ROUTER_GEMV), inert, or a generic non-DSv3 flag
⚠️ Classification must be verified, never recalled from memory. Pitfalls
actually hit: TOPK_PARALLEL, assumed default-OFF and slated for revert, is in
fact ON in every decode build (part of the current winning stack — KEEP);
ROUTER_GEMV, assumed obsolete, is in fact the router of the TP<8 fallback
(must not be deleted). Use grep -n on the injection condition in
persistent_kernel.py and check whether the gate is actually exercised at the
production geometry.
3. Codex-vet the classification + refactor plan
Hand the gate inventory + classification + goal to Codex (mcp__codex__codex)
for a multi-round discussion: validate the classification, agree on a safe
execution order, the correctness risks, how to verify the ABI-revert, and the
structural-vs-leaf distinction. Codex will catch conflicts in the
classification (see the step-2 pitfalls).
4. Freeze the reference (the correctness comparison baseline)
Before changing anything, run the current winning stack (all levers ON) and
record its output: e2e tpot + logits/prose (run it twice to get the A/A
nondeterminism envelope). The final "clean default build" is compared against
this winning-stack reference, not against the old safe default.
5. Execute in the safe order (build-check after every step)
First revert the dead levers that changed the ABI (most dangerous; do it
in isolation and verify). If a dead lever touched the
(num_in,num_out,TASK_ENUM,variant) tuple in graph.cc/task_register or
added a tensor, a single ABI mismatch = "Invalid global read" at runtime. If
those changes were never committed, simply
git checkout HEAD -- <producer files> to return to a clean ABI, then:bash
Delete the dead lever's consumer half at the same time (otherwise it is
a stale-env out-of-bounds landmine).
Delete the remaining dead levers + all diagnostic probes → build-check.
Hard-wire the LEAF wins (leaf optimizations with a clean #else):
remove the gate, delete the #else, keep the geometry guard → build-check.
Hard-wire the STRUCTURAL wins last (path-selectors that change the graph
shape / large control flow): delete the entire alternate path, keep the
TP8/mbt/workers guard → build-check + a --layers 0-3 in-MPK smoke.
(Structural gates are more dangerous than leaves — removing one deletes a
whole alternate code path. Do it after the tree has already shrunk.)
Show full SKILL.md (435 more words)Show less
6. Verify correctness (the default path's math has changed)
Token-identity cannot be used (DSv3 TP8 decode is FP-nondeterministic —
cross-CTA atomicAdd).
Use instead: the A/A envelope (winning stack compared against itself) +
clean default vs the winning-stack reference with the per-step
logit-cosine inside the envelope + stable top-k overlap + 512 tokens of
coherent prose + no NaN/Inf.
Perf smoke: e2e tpot should ≈ the winning-stack reference (confirms no
win was silently dropped).
TP8 JIT smoke: --layers 0-3 confirms the hard-wired megakernel really
instantiates and runs (the #else-deleted path only instantiates at
world_size==8).
Stage source files only (kernels/.cuh, builder.py, persistent_kernel.py,
task_register.cc); exclude.claude/, scratch/, experiment_history/,
CSVs/outputs/.pk_compile and other local artifacts.
Run mpk-commit-reviewer and tell it explicitly that the default-build
change is intentional (otherwise it will BLOCK per the standard gate); it
still checks staged-path hygiene, the allowed surface, the message, and the
correctness story.
The commit message lists everything: which levers were hard-wired (+ each
one's Δ), which dead levers were reverted, which diagnostics were deleted,
that the ABI was restored, the verification evidence, and an explicit
pre-merge gate (if TP8 runtime validation was blocked by box capacity,
write it into the message as a must-run item before merging). Include
Co-Authored-By.
Key pitfalls (all hit in practice)
The default build is intentionally NOT byte-identical — that is the goal,
not a bug; waive that one commit gate.
The ABI-revert is the most dangerous step — do it first and in
isolation, grep it clean, diff against the last clean commit, and delete the
consumer half together with it.
Structural wins go last — a path-selector is more dangerous than a
leaf-opt (removing it deletes an entire alternate path).
Classification must be verified — some gates are already unconditionally
ON, some are fallbacks; never assume "default-OFF".
Orphaned legacy functions — after deleting the #else call site, the
__device__ function definition may remain (nvcc elides it; harmless but
unclean); either delete it or flag it as a known nit in the PR.
The correctness gate = A/A envelope + coherence, not token-identity (the
path is FP-nondeterministic).
The TP8 runtime gate may be blocked by box capacity — a commit meant for
PR review may land with a documented pre-merge gate (a PR is review, not
auto-merge); never fabricate numbers.
Sub-agents can contradict each other about the same fact (e.g. "the
lever was deleted" vs "the lever was hard-wired") — verify yourself with
grep -c <the win's body symbol> that the win's body is still present
(zero refs to the macro ≠ the win was deleted; possibly only the gate was
removed and the body became unconditional).
Mpk Lever Cleanup next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.
Reviews WooCommerce code changes against the project's standards, flagging backend PHP architecture, naming, documentation, data integrity and testing violations.
Makes the agent pick the laziest solution that works: skip unneeded work, reuse what exists, prefer the standard library and platform features, and keep diffs small.
Guides refactoring in phases based on Martin Fowler's method: research, test coverage check, planning and small tested steps, with your approval at each phase.
Installs, updates or migrates the vendored anti-slop Oxlint plugin in a repository, keeping local rule changes and the plugin's license and provenance files.
Records a project's quality bar in CONSTRAINTS.md and watches diffs for signs an agent quietly weakened it, such as suppressions, skipped tests or lowered thresholds.
A skill your agent uses when the user wants to design or extend a FlashAttention-style forward kernel on B200/Blackwell, involving the two MMAs QKᵀ and PV, online softmax, S/P/O in TMEM, warp roles…
End-to-end pipeline for adding or porting a model to MPK Runtime-V2 — from a compute-graph spec (shapes + draw.io graph + HF checkpoint + TP/EP plan) to a working multi-GPU demo.
A skill your agent uses when a batch of env-gated (ifdef MPKDSV3 / os.environ-controlled, default-OFF) MPK optimization levers needs to be consolidated into a single clean code path for a PR…. Mpk Lever Cleanup is an agent skill from mirage-project/mirage.environ-controlled, default-OFF) MPK optimization levers needs to be consolidated into a single clean code path for a PR: hard-wire every winning lever as the default, delete the legacy else branches, remove the env vars that select new-vs-old logic, revert dead levers that measured KILL/NULL/regress, delete diagnostic probes, then commit one clean version.
When should I use Mpk Lever Cleanup?
Mpk Lever Cleanup fits situations like: A batch of env-gated (ifdef MPKDSV3 / os.environ-controlled; delete the legacy else branches; remove the env vars that select new-vs-old logic; revert dead levers that measured KILL/NULL/regress.
How do I install Mpk Lever Cleanup in Claude Code?
Run `npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a claude-code`. Or copy the skill folder (.claude/skills/mpk-lever-cleanup in mirage-project/mirage) into .claude/skills/mpk-lever-cleanup in your project. Claude Code loads it when a task matches its description.
How do I install Mpk Lever Cleanup in Codex?
Run `npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a codex`. Or copy the skill folder (.claude/skills/mpk-lever-cleanup in mirage-project/mirage) into .agents/skills/mpk-lever-cleanup in your project. Codex loads it when a task matches its description.
Can I use Mpk Lever Cleanup in Cursor, Gemini CLI or GitHub Copilot?
Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add mirage-project/mirage --skill mpk-lever-cleanup -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/mpk-lever-cleanup, .gemini/skills/mpk-lever-cleanup, .github/skills/mpk-lever-cleanup and .opencode/skills/mpk-lever-cleanup in your project.
What does Mpk Lever Cleanup need to run?
Going by SKILL.md and its folder, Mpk Lever Cleanup needs the command-line tools its instructions call (git).
Does Mpk Lever Cleanup access the network?
SKILL.md contains no URLs. Its commands use git, which can reach the network depending on how they are called. This is read from the text; nothing was executed.
Is Mpk Lever Cleanup safe to install?
Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.
What licence does Mpk Lever Cleanup use?
Mpk Lever Cleanup is published under the Apache-2.0 licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.
How many tokens does Mpk Lever Cleanup use?
About 2.2k tokens (SKILL.md is roughly 8.8k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.
What are the alternatives to Mpk Lever Cleanup?
Skills that share tags, products or a category with Mpk Lever Cleanup: WooCommerce Code Review (woocommerce/woocommerce, 11k stars), Ponytail Lazy Developer Mode (DietrichGebert/ponytail, 160k stars), Systematic Code Refactoring (luongnv89/claude-howto, 42k stars) and Install Anti-Slop Oxlint Rules (dmmulroy/anti-slop, 5.4k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.
Who maintains Mpk Lever Cleanup?
mirage-project (a GitHub organization) maintains it in mirage-project/mirage, which has 2,545 GitHub stars. The repository holds 24 skills in this directory. The repository was last updated on October 7, 2026.
Source: mirage-project/mirage on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.
