Report · marked 25 Aug 2026
vllm-project/vllm
Solid foundations, held back in a few specific places.
Python · 90,023 stars · 6,714 files · branch main
83
/100
A-
Instructions
27/27Whether the repo tells an agent how to behave before it starts guessing. Full marks.Setup
8/14Whether an agent can install the project and get it running without a human. Losing marks on deterministic install.Verification loop
24/25Whether an agent can check its own work. This is the category that most decides whether agent output is trustworthy. Losing marks on static type checking.Context economy
11/20Whether the repo fits in a context window, or fights it. Losing marks on file sizes fit in context and repository weight.Navigability
10/10How quickly anything new can orient itself. Full marks.
Higher is better. Accent marks what is costing you most.
Fix these first
01
Deterministic install
Commit the lockfile your package manager generates. Without it, an agent's install and your install are different builds, and 'works on my machine' becomes unfalsifiable.
02
File sizes fit in context
A single file that fills the context window forces an agent to work from fragments, and it will confidently edit code it never saw. Splitting the worst offenders pays for itself immediately.
03
Repository weight
Large binaries and vendored trees slow every operation an agent performs. Git LFS or a separate assets repo keeps the working tree navigable.
The full marking
Every deduction below names the file or setting it came from.
Instructions 27/27 A+
Whether the repo tells an agent how to behave before it starts guessing.
✓Agent instruction file
12/12
AGENTS.md is present, which is the format the most tools read.
Also found: CLAUDE.md
✓Instruction quality
12/12
Specific enough that an agent can act on it.
5,745 characters, a workable length
names the actual commands to run
uses headings, so an agent can skim it
includes copyable code blocks
✓README as an entry point
3/3
README has a clear getting-started section.
Setup 8/14 B-
Whether an agent can install the project and get it running without a human.
✗Deterministic install
0/6
No lockfile. An agent installing dependencies may not get what CI got.
Commit the lockfile your package manager generates. Without it, an agent's install and your install are different builds, and 'works on my machine' becomes unfalsifiable.
✓Discoverable commands
6/6
Commands are declared where an agent will look for them.
pyproject.toml scripts
scripts/ directory
–Environment config
n/a
Not applicable — this reads as a library, with no environment to configure.
✓Pinned runtime version
2/2
The language runtime version is pinned.
Found rust-toolchain.toml
–Reproducible environment
n/a
Not applicable — a library doesn't need a container to be worked on.
Verification loop 24/25 A+
Whether an agent can check its own work. This is the category that most decides whether agent output is trustworthy.
✓Tests exist
8/8
2075 test files against 4870 source files.
e.g. .buildkite/amd-disagg/run-slurm-disagg-test.sh, .buildkite/lm-eval-harness/test_lm_eval_correctness.py, .buildkite/performance-benchmarks/tests/genai-perf-tests.json
✓Test command is discoverable
7/7
An agent can find and run `pytest`.
✓Continuous integration
4/4
10 GitHub Actions workflows define what "passing" means.
✓Lint and format rules
3/3
pyproject.toml encodes the house style.
✗Static type checking
2/3
Type checking is configured, but not in strict mode.
A type checker gives an agent an error message instead of a runtime surprise. It's the second-fastest feedback loop after the compiler.
Context economy 11/20 B-
Whether the repo fits in a context window, or fights it.
✓No committed build output
6/6
No generated directories committed.
✓.gitignore hygiene
3/3
.gitignore covers 120 patterns.
✗File sizes fit in context
0/6
25 source files are over 100KB.
A single file that fills the context window forces an agent to work from fragments, and it will confidently edit code it never saw. Splitting the worst offenders pays for itself immediately.
vllm/entrypoints/serve/instrumentator/static/swagger-ui-bundle.js — 1475KB
vllm/v1/worker/gpu_model_runner.py — 335KB
rust/src/server/src/routes/tests.rs — 230KB
✗Repository weight
2/5
About 267MB checked out. Large enough that cloning and searching are both slow.
Large binaries and vendored trees slow every operation an agent performs. Git LFS or a separate assets repo keeps the working tree navigable.
Here is your AGENTS.md
Drafted from what is actually in this repository: the install
command from your lockfile, the commands you already declare, your real directory layout. Anything
marked TODO needs a person. Save it at the root as AGENTS.md.
# AGENTS.md A high-throughput and memory-efficient inference and serving engine for LLMs ## Setup ``` # TODO: the install command. No lockfile was found, so this could not be inferred. ``` ## Commands ``` pytest # run the test suite ruff check . # lint ``` ## Layout - `vllm/` 2234 source files - `tests/` 1762 source files - `rust/` 310 source files - `examples/` 166 source files - `csrc/` 135 source files - `benchmarks/` 108 source files ## Conventions - Tests live alongside the code they cover, following `.buildkite/amd-disagg/run-slurm-disagg-test.sh`. - CI defines what passing means. See `.github/workflows/add_label_automerge.yml`, and keep it green. - TODO: add the two or three conventions a newcomer always gets wrong here. ## Gotchas - `vllm/entrypoints/serve/instrumentator/static/swagger-ui-bundle.js` is 1475KB. It will not fit comfortably in context, so read it in parts. - No lockfile is committed, so an install here may not match what CI produced. --- Drafted by agentreadme.com from what is in this repository. Everything marked TODO needs a human. Check it in as AGENTS.md at the root.
Open the raw markdown
or
curl -o AGENTS.md agentreadme.com/draft/vllm-project/vllm.md
Show the mark
The badge re-checks daily, so it keeps up as the repository changes. Use mark again to force it now.
[](https://agentreadme.com/vllm-project/vllm)Other Python repositories, marked
Think this mark is wrong?
Every deduction above names the file it came from, so this can be settled by looking. If a check missed something, that is a rule worth fixing.