Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
67 lines (49 loc) · 2.41 KB

CONTRIBUTING.md

File metadata and controls

67 lines (49 loc) · 2.41 KB
Raw

Contributing

Thanks for looking at Auto. CLAUDE.md records the engineering norms we held to; read it before a first change. This file is the short version of the norms that gate a merge.

Build gates

CI runs these on every pull request, and they must be green before merge:

cargo fmt --all --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace

clippy -D warnings means warnings fail the build. No unsafe without a justification comment above it (the workspace denies unsafe_code).

Toolchain is pinned: rust 1.96.1 (rust-toolchain.toml) and flatc 25.12.19 exactly (it must match the flatbuffers crate; the IR build fails on a mismatch). See the README build section for how the build resolves flatc.

End-to-end loops

Changes to the pipeline should keep the end-to-end scripts green. They record, compile, run, and prove the negative paths (emit blocked on a wrong module, guard abstention, deopt and recompile, registry tamper refused):

cargo build -p auto-cli
bash evals/toy-agent/e2e.sh
PYTHON_BIN=python bash evals/distill-agent/e2e.sh

CI runs the full set (serve-proxy, pipeline-agent, daemon, tool-agent, registry-remote) in .github/workflows/ci.yml. None of them touch the network, and none spend money.

No unverified claims

Honesty is load-bearing here.

  • Manifests report measured numbers or null. Never fabricated, never rounded up.
  • A parity or performance claim carries an eval-run id. A failing or unmeasurable contract blocks emit; that is the point of the gate.
  • No pass merges without differential tests against the reference interpreter.
  • A stub is labeled stub in code, in --help, and in docs. No mock pretends to be the compiler.

Decisions and scope

  • Irreversible decisions get a numbered, terse ADR under spec/adr/, with alternatives listed. Anything under spec/ is written for external readers.
  • Adjacent gaps you notice but do not fix belong in spec/adr/open-questions.md, not in scope creep.

Commits and pull requests

  • Granular commits with honest messages. One logical change per commit.
  • Never push to main directly. Open a pull request; keep CI green.
  • Frontier API spend has a hard per-session cap and is logged to a ledger. We built this under limited resources, so paid OpenAI usage stays capped.

By contributing you agree that your contributions are licensed under the Apache License 2.0 (see LICENSE).