arle
AGENTS.md19 starsChanged 54 days agoMITView on GitHub
Assisting ckl. Project caveats and hard gates only — generic Rust / CUDA / Metal / git knowledge is intentionally absent, and so is anything you can read off the file tree. Match the surrounding code's idiom, naming, and comment density rather than a style rulebook. Load on demand, not upfront: AGENTS.md is canonical; CLAUDE.md is a symlink to it. ARLE is a Rust-native, device-neutral inference runtime with an integrated local agent and On-Policy Distillation (OPD) workflows. No PyTorch, no…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 33,315 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository6 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- arle-upstream-runtime-scanSkill
- find-simplificationsSkill
- kernel-optimizationSkill
- tmux-agent-controlSkill
- understand-until-simpleSkill
The file
We split this AGENTS.md into the rules it gives an agent
12 rules. About 0 are common, 10 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈10
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 54
No Metal unit test loads a model, so there is no unit-test opt-out to name.
- Very specific: line 65
Backend isolation (CRITICAL): Never cfg-leak backend types into cross-backend modules — everything above the seam (infer-core / -server / …
- Very specific: line 72
Run it before pushing any infer-cuda / cuda-kernels / cuda-gated cli edit.
- Distinctive: line 82
Fast path only: If it only works on the eager/un-captured path, it isn't done — no per-step readback/sync in the hot loop.
- Distinctive: line 95
No dense jargon, no hedging, no restating the question.
- Distinctive: line 103
Never delete content outside the stated scope.
- Distinctive: line 116
Tests: minimal and end-to-end: Default is no new test. Add one only when the change carries logic that can silently break (branch, parser, …
- Distinctive: line 133
Commit only your own files by explicit path.
- Distinctive: line 134
After git mv + edits, re-check git status — the fmt step can de-stage renames.
- Distinctive: line 148
Standard terms: Use the established term for a concept.
- Distinctive: line 157
Table cells may be descriptive: Not every entry needs a number or conclusion — an entry may describe only what happened.
- Distinctive: line 159
Unknown cause: Write "cause unknown"; do not append unverified explanations.
≈ “Common” is an estimate: at least 1 in 5 of the 42,283 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
ARLE — Agent Contract
Assisting ckl. Project caveats and hard gates only — generic Rust / CUDA / Metal / git knowledge is intentionally absent, and so is anything you can read off the file tree. Match the surrounding code's idiom, naming, and comment density rather than a style rulebook.
Load on demand, not upfront:
| When | Read |
| Evidence bar, decomposition, distilled lessons | docs/agent-method.md |
| Any bench or trace | docs/bench-and-trace-spec.md |
| Where code lives / execution paths | docs/codebase-map.md, docs/architecture.md |
Editing crates/{cuda-kernels,mlx-sys}/ | that crate's AGENTS.md |
| Backend / model / quant support level | docs/support-matrix.md |
| Env vars, SM tier policy | docs/environment.md |
| Session start | docs/index.md |
AGENTS.md is canonical; CLAUDE.md is a symlink to it.
Project shape
ARLE is a Rust-native, device-neutral inference runtime with an integrated
local agent and On-Policy Distillation (OPD) workflows. No PyTorch, no Python
on the hot path.
Two backends plug into one seam (infer_seam::{BackendExecutor, KvPool}, two
host-only traits): CUDA continuous batching (cudarc + vendored FlashMLA /
DeepGEMM / DeepEP + TileLang AOT + native CUDA C) and Metal (crates/mlx-sys
C++ bridge, packed varlen decode). One infer_core::Engine<E, K> drives both —
the seam is the engine's cost contract (submit/poll + StepLimits +
explicit capability accessors): a new backend implements the core loop and
declares its capabilities, and a family whose decode is not submit-poll
shaped forks the loop (precedent: diffusion_executor.rs) rather than
bending the trait.
Non-obvious ownership:
infer-*owns serving/runtime truth. There is no monolithicinfer/crate; any doc or command referencinginfer/or-p inferis stale.infer-api(LoadedInferenceEngine) is the single programmatic entry point;arleis the CLI entry point.- OPD training lives in the separate
arle-opdrepo and consumes arle's crates. - CUDA kernels: adopt-official-first (
vendor/), hand-rolled atcrates/cuda-kernels/csrc/only for the genuine gap.
Metal canonical model — globally unified:
mlx-community/Qwen3.6-35B-A3B-4bit (MoE, ~19 GB, HF-cached) — the default for
every Metal serve, scripts/bench_*.sh, smoke, and Metal perf claim; detects
MoE regressions a dense model cannot. CUDA benches keep their own defaults.
No Metal unit test loads a model, so there is no unit-test opt-out to name.
- Auto-wired-limit (always-on): the Metal executor pins weights via
mlx::set_wired_limitat construction (infer-metal/src/wired_limit.rs) — c=1 p99 86→15 ms on Qwen3.6. MLX_MAX_OPS_PER_BUFFER/MLX_MAX_MB_PER_BUFFERare not defaults — a Qwen3.5-dense tune, wash-or-loss on Qwen3.6 MoE. Per-workload matched-A/B only.
Hard gates
Backend isolation (CRITICAL). Never cfg-leak backend types into
cross-backend modules — everything above the seam (infer-core / -server /
-api) stays device-neutral; backend types live only in infer-cuda /
infer-metal. CUDA stubs on other targets: todo!("GPU required: ..."). Mac
pre-push lint without nvcc (CI Lint mirror; CUDARC_CUDA_VERSION skips the
toolkit probe):
CUDARC_CUDA_VERSION=12080 cargo clippy -p infer-api --release --no-default-features --features cuda,no-cuda,nccl,deepep --lib -- -D warnings.
Run it before pushing any infer-cuda / cuda-kernels / cuda-gated cli
edit; metal,no-cuda and plain cargo check never see those lints.
Perf claims cite a matched A/B. A runtime perf claim cites a matched A/B run (same machine, same build profile, ≥3 trials) in the PR description. The bench parameters and A/B contract live in the bench spec.
GPU kernel work ships a measured before/after — ncu (CUDA) or Xcode Metal
capture / MLX instruments (Metal).
Fast path only. If it only works on the eager/un-captured path, it isn't done — no per-step readback/sync in the hot loop; capture at sync points, rebuild transient state on restore.
Correctness parity = the correct-inference gate (byte-identity is not
required, due to MoE non-determinism): scripts/needle_gate.py + scripts/lever_gate.sh, needle
ladder ×3 same-config vs the baseline envelope. Default flips additionally need a
wall-clock perf license.
No half-states. Finish a refactor unit or revert it; never leave parallel old+new paths in the tree.
Use plain language. Say the finding in plain words first, numbers second. No dense jargon, no hedging, no restating the question. If a sentence needs a glossary, rewrite it.
Approach-first for >3 files or architectural decisions — outline, then execute. Wait for the user ONLY when there is a real tradeoff to adjudicate (two viable paths with different costs). No tradeoff → nothing to decide → don't ask. Adopting the SOTA/industry-standard approach is never a decision point — just execute it. Never delete content outside the stated scope; inside it, prefer deletion-style refactors (collapse duplicates, converge on one flow) over layering adapters.
Working rules
Phases (non-trivial tasks): Explore until you can name every file you will
touch → Plan (accepted in writing; >5 files or irreversible → stop and flag) →
Implement (compiles, simplify pass on the diff) → Verify (cargo test
--workspace, cargo clippy -- -D warnings).
Trivial → Implement + Verify.
Tests: minimal and end-to-end. Default is no new test. Add one only when the change carries logic that can silently break (branch, parser, quant, rollback, sampling, security), and then the smallest end-to-end gate that fails when it breaks — not a per-function suite.
Delegation. Claude does direction, docs, planning, and integration;
general-purpose subagents execute; Explore maps; Plan handles >5-file
plans; review is codex review --uncommitted at Bash, backgrounded and tee'd.
Independent tasks go out in one message, in parallel. Two failed subagent
attempts → hand-write the diff or re-brief a fresh agent with what was tried.
Never codex:codex-rescue / mcp__openmax__execute_with_codex for
execution — both hang.
Git. Commitizen <type>(<scope>): <subject>, scopes metal cuda
scheduler qwen3 qwen35 http kv-tier docs. Small tranches, each
self-contained, simplify pass first.
Commit only your own files by explicit path; never git stash others' work.
After git mv + edits, re-check git status — the fmt step can de-stage renames.
Code layout gotchas. Flat modules, no mod.rs —
crates/cuda-kernels/src/ffi.rs declares a
sibling #[path = "ffi/attention.rs"] pub mod attention;. Weights are &self
(immutable, pool-shared); per-request mutable state lives in the State
associated type. Comments carry the non-obvious why in ≤1 line, in English,
never the what and never which task added it; issue numbers only when naming a
specific bug. If the code already reads clearly, leave it bare — no comment.
Writing (reports, PR descriptions, docs)
- Standard terms. Use the established term for a concept; do not coin a new one.
- No metaphors. Name the referent directly; any phrase that needs the reader to infer what it stands for gets rewritten.
- Neutral headers. Table headers and category names are neutral nouns (Problem, Phenomenon, Impact, Result). Keep section status labels consistent (e.g. Completed / Confirmed / Below baseline / Pending).
- No "X, not Y" sentences. State the fact directly without contrastive framing.
- Table cells may be descriptive. Not every entry needs a number or conclusion — an entry may describe only what happened.
- Unknown cause. Write "cause unknown"; do not append unverified explanations. Once a mechanism is confirmed, that content may stay.
- No colloquialisms. Formal, precise wording only.
- No personification. Systems, processes, and data do not act or feel.
Build & run
Iteration builds use --profile release-fast (cu=16, no LTO); perf numbers and
shipped artifacts use --release (cu=1, thin LTO). A release-fast binary is
not a valid bench source: pod.sh build defaults to release-fast, and
pod-remote-run.sh refuses a bench-labelled run (label prefix bench) whose
build receipt profile is not release. Debug GPU builds are unusably slow.
CUDA_HOME=/usr/local/cuda cargo build --release --features cuda # CUDA (Linux+NVIDIA)
cargo build --release --no-default-features --features metal,no-cuda # Metal (Apple Silicon)
cargo build --release --no-default-features --features cpu,no-cuda # portable / CI smoke
# Multi-GPU features stack: nccl (implies cuda) → deepep (implies nccl)
# CI test lanes use --profile release-fast (cu=16, no LTO); the release profile's
# cu=1 cold build OOMs 16 GB CI runners:
cargo test -p arle --profile release-fast --no-default-features --features cpu,no-cuda,cli
cargo test -p cli --release --no-default-features --features metal,no-cuda
cargo test -p kv-native-sys --profile release-fastWeekly: cargo sweep --time 30.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

