-
Notifications
You must be signed in to change notification settings - Fork 185
Add CLAUDE.md agent file #676
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Open
jakecorrenti
wants to merge
1
commit into
containers:main
Choose a base branch
from
jakecorrenti:add-agent-file
base: main
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
Open
Changes from all commits
Commits
File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1 @@ | ||
| @AGENTS.md |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,149 @@ | ||
| # AGENTS.md | ||
|
|
||
| This file provides guidance to Agents when working with code in this repository. | ||
|
|
||
| ## What this project is | ||
|
|
||
| **libkrun** is a Rust dynamic library that provides virtualization-based process isolation via KVM (Linux) and HVF (macOS). It exposes a C API (`include/libkrun.h`) that lets callers run a process inside a lightweight VM with configurable vCPUs, RAM, virtio devices, and an embedded init binary. | ||
|
|
||
| ## Build commands | ||
|
|
||
| All builds go through the Makefile, which handles feature flags, platform detection, and sysroot management for cross-compilation. Direct `cargo build` skips that plumbing. | ||
|
|
||
| ```bash | ||
| # Release build (minimal — no optional devices) | ||
| make | ||
|
|
||
| # Release build with common optional devices | ||
| make BLK=1 NET=1 GPU=1 SND=1 INPUT=1 | ||
|
|
||
| # Debug build | ||
| make debug | ||
|
|
||
| # TEE variants (mutually exclusive with each other and GPU/SND/INPUT) | ||
| make SEV=1 # AMD SEV — produces libkrun-sev.so | ||
| make TDX=1 # Intel TDX — produces libkrun-tdx.so | ||
|
|
||
| # Other optional features | ||
| make VHOST_USER=1 | ||
| make VIRGL_RESOURCE_MAP2=1 | ||
|
|
||
| # Install to /usr/local (or PREFIX=...) | ||
| make install | ||
| make PREFIX=$HOME/.local install | ||
| ``` | ||
|
|
||
| The Makefile exports `CC_LINUX` for Rust build scripts that compile C code targeting Linux. On macOS it auto-downloads a Debian sysroot; on Linux it uses the host toolchain. | ||
|
|
||
| ## Lint and format | ||
|
|
||
| Clippy is run with `-D warnings` — zero warnings are allowed. The CI checks several feature combinations; you should too when touching device or feature-gated code: | ||
|
|
||
| ```bash | ||
| # Required before every PR — same as CI | ||
| cargo clippy --locked -- -D warnings | ||
| cargo clippy --locked --features amd-sev -- -D warnings | ||
| cargo clippy --locked --features tdx -- -D warnings | ||
| cargo clippy --locked --features net,blk,gpu,snd,input -- -D warnings | ||
|
|
||
| # Format check | ||
| cargo fmt -- --check | ||
| ``` | ||
|
|
||
| ## Tests | ||
|
|
||
| ```bash | ||
| # Unit tests (requires init/init to exist and KVM access on Linux) | ||
| touch init/init | ||
| cargo test | ||
|
|
||
| # Integration tests (builds and installs the library first) | ||
| make test | ||
| make test TEST=test_name # run a single integration test | ||
| make test BLK=1 # integration tests with blk feature | ||
| ``` | ||
|
|
||
| Integration tests live in `tests/` as a separate Cargo workspace. They require the library to be installed to a local prefix (`test-prefix/`) which `make test` handles automatically. | ||
|
|
||
| ## Crate architecture | ||
|
|
||
| The workspace (`Cargo.toml`) contains these crates under `src/`: | ||
|
|
||
| | Crate | Role | | ||
| |---|---| | ||
| | **libkrun** | C API surface. Manages `KrunContext` instances (one per VM), translates API calls into VMM configuration, and drives startup via `krun_start_enter()`. | | ||
| | **vmm** (krun-vmm) | VMM core: builds and runs the VM, owns the vCPU threads, wires devices to memory and IRQs. | | ||
| | **devices** (krun-devices) | All virtio device implementations: console, block, fs (virtiofs passthrough + read-only wrapper), net, gpu, sound, input, vsock. | | ||
| | **arch** (krun-arch) | Platform-specific VM setup: GDT/IDT (x86_64), FDT (aarch64/riscv), boot protocol. | | ||
| | **kernel** | Loads the kernel image and sets up the boot parameters passed to the VMM. | | ||
| | **hvf** | macOS HVF hypervisor bindings (Linux uses kvm-ioctls directly in vmm). | | ||
| | **cpuid** | x86_64 CPUID leaf manipulation for vCPU feature exposure. | | ||
| | **rutabaga_gfx** | Wraps virglrenderer for virtio-gpu. | | ||
| | **display** / **input** | Host-side display and input backends (used by the gpu feature). | | ||
| | **utils**, **polly**, **smbios**, **arch_gen** | Shared utilities, event loop, SMBIOS table generation, architecture codegen. | | ||
| | **aws_nitro** | AWS Nitro Enclave support. | | ||
|
|
||
| ### How a VM starts | ||
|
|
||
| 1. Caller invokes `krun_create_ctx()` → allocates a `KrunContext` | ||
| 2. Caller configures it (vCPUs, RAM, disks, network, exec path, etc.) | ||
| 3. `krun_start_enter()` calls into **vmm**, which: | ||
| - Loads the kernel via **kernel** | ||
| - Instantiates virtio devices from **devices** | ||
| - Starts vCPU threads (KVM ioctls on Linux, HVF on macOS) | ||
| - The guest boots, the C init binary runs as PID 1, reads `.krun_config.json` from the virtiofs overlay, and execs the workload | ||
|
|
||
| ### The init binary (`init/init.c`) | ||
|
|
||
| The guest PID-1 is a statically-linked C binary (`init/init.c`) compiled by `src/devices/build.rs` via `CC_LINUX`. The compiled binary path is set in the `KRUN_INIT_BINARY_PATH` env var at build time and embedded into the devices crate via `include_bytes!`. The passthrough fs backend exposes it as a virtual read-only file named `init.krun` (inode defined in `src/devices/src/virtio/fs/linux/passthrough.rs`) — the real host filesystem never sees it. | ||
|
|
||
| AWS Nitro uses a separate C init (`init/aws-nitro/`) built by the Makefile when `AWS_NITRO=1`. | ||
|
|
||
| ### Feature flags | ||
|
|
||
| Features are additive and controlled at the `libkrun` crate level. Each device feature (`blk`, `net`, `gpu`, `snd`, `input`) enables the corresponding code in both `devices` and `vmm`. The TEE variants (`amd-sev`, `tdx`) imply `blk` + `tee` and affect the soname of the output library. | ||
|
|
||
| ## Pull Request expectations | ||
| - New tests are added when necessary. | ||
| - Documentation is added or updated when necessary. | ||
| - Linting and formatting has been done. | ||
| - All of the tests pass. | ||
| - The commit structure follows what's described in [Commit structure](### Commit structure). | ||
|
|
||
| ## Code Quality | ||
|
|
||
| ### Commit structure | ||
|
|
||
| - Format: `<subsystem>: <title>` — e.g., `virtio/blk: add print_text() function` | ||
| - Commits must be self-contained, compile, and pass tests independently | ||
| - Sign all commits with `git commit -s` (DCO requirement) | ||
| - Agent attribution: `Claude Code: <model-name>` | ||
| - Commit messages should be concise and written in the imperative mood. Small, focused commits are preferred. | ||
|
|
||
| ### Rust coding style | ||
| - No error handling for impossible scenarios. | ||
| - Avoid checking for empty input when calling a function if the function already handles the base case well (e.g. empty input is noop). | ||
| - Use `use` imports instead of inline full paths. One level of qualitifation is fine when it clarifies what something is (e.g., `log::trace!`, `fs::read_to_string()`), but don't use longer paths like `std::process::Command::new(...)` or `crate::foo::bar::baz()` -- import with `use` instead. | ||
|
|
||
| Creating stubs to make code compile causes silent runtime failures: | ||
| - **EventFd**: Don't create new `EventFd::new()` - propagate existing ones (via `Arc<EventFd>` or `.try_clone()`). New `EventFd`s are never signaled -> deadlock. | ||
| - **Queues**: Don't create empty queues - pass existing ones through. Empty queues -> nothing received. | ||
| - **Files**: Don't create dummy files for `include_bytes!()`. Ask the user how to build/obtain them. | ||
|
|
||
| **Principle**: When something is wired into the system, propagate it. Don't create disconnected instances. If a build fails due to missing dependencies, ask the user rather than stubbing. | ||
|
|
||
| ## Platform support | ||
|
|
||
| ### Host platforms (where libkrun runs as a library) | ||
|
|
||
| - **Linux x86_64 / aarch64**: primary targets, full feature support (KVM) | ||
| - **Linux riscv64**: experimental (KVM) | ||
| - **macOS aarch64**: supported (HVF), GPU works | ||
| - **FreeBSD**: experimental cross-compilation target (`make BUILD_BSD_INIT=1`) | ||
|
|
||
| ### Guest platforms (what runs inside the VM) | ||
|
|
||
| The guest ISA always matches the host ISA — KVM and HVF use hardware virtualization, not emulation. | ||
|
|
||
| - **Linux** (x86_64, aarch64, riscv64): primary guest OS; the default init binary (`init/init.c`) is Linux-specific | ||
| - **FreeBSD** (same arch as host): experimental; requires building the FreeBSD init with `make BUILD_BSD_INIT=1` | ||
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So this section comes from my CLAUDE.md, but I am not sure if we should include it fully verbatim in the committed version.
This is also perhaps bit of workflow thing too. Though it's also a bit specific to things I found Claude (at least in the past) struggle with in libkrun. But if you test the code you submit you, you would notice your change doesn't actually work and just deadlocks immediately.
The more I think about this I realize there is sort of a tradeoff here. If we put a lot of these guidelines here it could be annoying for people that have their own guidelines like this. However it would be really beneficial for people with default agents configuration especially if they are inexperienced or don't notice the issues produced by the agent - I can see this improving the quality of random drive-by PRs.
Maybe a middle-ground could be to keep this section but shorten it?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@jakecorrenti Have you noticed Claude or other LLMs also struggle with this? If not maybe we should hold off on this for now?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I don't think I've run into that personally. The one thing I've been running into a lot is the agent basically liking to hear itself talk. i.e.