Guidance for Claude, Codex and other coding agents working in the Vortex repository.
- When asked a question about the PR or codebase, especially via
/query, use the.agents/skills/queryskill. - When asked to investigate a CI failure, especially via
/ci-failure-analysis, use the.agents/skills/ci-failure-analysisskill.
Vortex is a Rust monorepo for columnar array processing, compression encodings, and file IO.
The workspace also contains Java bindings in java/, Python bindings in vortex-python/,
documentation in docs/, and benchmark tooling in vortex-bench/ and benchmarks/.
vortex-bufferdefines zero-copy alignedBuffer<T>andBufferMut<T>, guaranteed to be aligned toTor to a requested runtime alignment.vortex-array/src/dtypecontains theDTypelogical type system used throughout Vortex.vortex-arraycontains the coreArraytrait and the base encodings, including most Apache Arrow-style encodings.encodings/*contains more specialized compressed encodings.vortex-fileimplements file IO. It usesLayoutReaderfromvortex-layout.vortex-ioholds the core async and blocking IO traits, plus the genericobject_storeadapters that implement them.vortex-cloudholds the cloud object store integration: the URL-to-ObjectStoreregistry and the OpenDAL-backed services (cos://,oss://). Every binding resolves URLs through it.vortex-scan,vortex-session,vortex-datafusion, andvortex-duckdbcontain scan and execution integrations.vortex-pythoncontains Python bindings. RST-flavored project docs live indocs/.
Before changing files in a subtree, read the closest nested AGENTS.md. In particular:
.github/AGENTS.mdcovers workflows and other GitHub configuration.docs/AGENTS.mdcovers Sphinx documentation.vortex-python/AGENTS.mdcovers Python and PyO3 binding work.
Prefer narrow crate builds while iterating:
cargo build -p <crate-name>Use workspace-wide builds only when the change spans crate boundaries or before handing off a broad refactor:
cargo build --workspaceRun tests for the crate or binding you touched before broader checks:
cargo nextest run -p <crate-name>If cargo-nextest is not available, you can install it with:
cargo install --locked cargo-nextestFor Rust doc comments or crate documentation, run doctests for the affected crate:
cargo test --doc -p <crate-name>Run verification that matches the files changed. Do not run expensive Rust checks for changes that
only touch Markdown, agent configuration, comments outside Rust code, symlinks, or other metadata
with no Rust/API behavior impact. For docs/config-only changes, validate formatting by inspection
or with a targeted doc/config command, and verify symlink or path changes with ls, find, and
git status.
For Rust code, public API, feature flag, or generated-file changes, run these before stopping:
cargo +nightly fmt --all
cargo clippy --all-targets --all-featuresFor changed C++ and CUDA source or header files covered by CI (.cpp, .hpp, .cu, .cuh, and
.h files under lang/cpp, vortex-cuda, vortex-duckdb, and vortex-ffi), format with the
repository's .clang-format configuration and verify the result:
clang-format --style=file -i <changed-files>
clang-format --dry-run --Werror --style=file <changed-files>Pass only the files you changed; CI excludes vendored or generated CUDA and Arrow headers from its repository-wide check.
Notes:
- For
.github/changes, follow.github/AGENTS.mdand runyamllint --strict -c .yamllint.yamlon changed workflow files. - If cargo fails with exactly
sccache: error: Operation not permitted, rerun that command withRUSTC_WRAPPER=so rustc runs directly. Only do this for that exact error.
- When iterating on CI failures, fetch only failed job logs first:
gh run view <run-id> --job <job-id> --log-failed. - Run narrow local repro commands for the affected crate, test, docs target, or binding before running workspace-wide checks.
- If a
ghcommand fails witherror connecting to api.github.comin the sandbox, immediately rerun it with escalated network permissions instead of retrying in the sandbox. - Verify causation from logs, diffs, and local repros before attributing a failure to a PR.
- Follow
STYLE.mdfor Rust formatting, documentation, API, error-handling, import, and safety conventions. - Only write comments that explain non-obvious logic or important context. Do not comment self-explanatory code.
- Keep public APIs small and consistent with neighboring crates.
Avoid hidden-cost per-element accessors in hot loops, follow the performance guidance in
STYLE.md, and benchmark changes to hot paths.
- Strongly consider
rstestcases when parameterizing repetitive test logic. - Prefer test functions that return
VortexResult<()>and use?instead ofunwrap. - Prefer test module names
tests, nottest. - Use
assert_arrays_eq!for array comparisons instead of element-by-element assertions. - Keep tests concise and focused on behavior, edge cases, and regressions.
- If a bug fix is requested, add or identify a failing test first when practical. A test that passes before and after the fix does not prove the fix.
- If clippy lints in tests prohibit patterns that are acceptable only in test code, consider allowing the lint at the test module level.
- If an existing
foo.rsmodule needs many tests, promote it to a directory module:foo/mod.rsplusfoo/tests.rs, included fromfoo/mod.rsbehind the appropriate test configuration.
Check new and modified lines against this list before finishing:
- Running broad CI-style commands before trying a narrow local repro.
- Using
unwrap,expect, or panic-oriented assertions in tests whereVortexResult<()>and?would be clearer. - Comparing arrays element by element instead of using
assert_arrays_eq!. - Adding imports inside functions when module-level imports would work.
- Introducing
unsafewithout proving that safe Rust cannot express the same operation. - Updating expected test output to match buggy behavior without independently verifying the intended semantics.
- Silently reducing the scope of an approved plan when implementation is harder than expected.
- Calling a hidden-cost per-element accessor (
Validity::is_valid,scalar_at,BitBuffer:: valueaccumulation) inside a hot loop instead of materializing once.
When summarizing work, write valid Markdown that can be copied into GitHub. Include the checks you ran and call out any checks you could not run.
When creating a branch on the user's behalf, prefix its name with the user's established branch
prefix, not the agent's name (for example, do not use codex/ or claude/). Infer the user's
prefix from their existing branches when possible. Otherwise, use available identity context such
as whoami, Git configuration, or other information the user has provided. If the evidence is
ambiguous, ask the user instead of inventing a prefix.
All commits must be signed off by the committers in this form:
Signed-off-by: "COMMITTER" <COMMITTER_EMAIL>