Contributing

Checks

scripts/verify.sh must pass before every push. It runs:

• cargo fmt --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace --all-features
• cargo check -p typedlm --no-default-features — the core builds without HTTP, derive and eval
• cargo doc with warnings as errors
• cargo +1.85 check without optimize and cargo +1.88.0 check with it — the minimum supported Rust versions

CI runs the same on pull requests and pushes to main.

Live tests

Tests against real endpoints are #[ignore] and live in crates/typedlm/tests/live.rs:

TYPEDLM_LIVE_BASE_URL=http://localhost:11434/v1 TYPEDLM_LIVE_MODEL=qwen3:32b \
TYPEDLM_LIVE_DIALECT=ollama cargo test --test live -- --ignored --nocapture

TYPEDLM_LIVE_API_KEY is optional; TYPEDLM_LIVE_DIALECT is strict (default), generic or ollama; TYPEDLM_LIVE_STRATEGIES limits the strategies, e.g. tool_call,json_mode.

Conventions

• Dependencies: only what users barely pay for; write narrow pieces instead of adding heavy or rare crates. No thiserror, no async-trait. Keep syn on the major version serde’s and schemars’ derives use, so it is not compiled twice.
• Errors: own enums, #[non_exhaustive], Display and Error by hand. Display never prints model output.
• Public enums that may grow are #[non_exhaustive].
• Async: native async fn in traits; object safety through a separate Dyn… trait with a blanket impl.
• Runtime: the core does not depend on an async runtime; tokio only in the http feature and in tests.
• Names: acronyms as words (TypedLm); derive attributes under #[lm(...)]; descriptions from doc comments.
• Tracing: OpenTelemetry GenAI field names, otherwise typedlm.*; no inputs or answers without opt-in.
• Tests: offline and deterministic. Use the providers in typedlm::testing; HTTP tests run against a small server inside the test, not a mock crate.