Skip to main content
Run the checks | LLMWhat each stage of the repository gate proves, and what it deliberately does not.LLMhow-tollmhow-todeveloperoperatorbuild-agent-systemsdeploy-operate-productsreference

Run the checks

task check

It runs four stages. None of them makes a paid provider call or provisions an external resource.

1. Rust​

cargo test --workspace --locked
cargo test -p b10x-llm-credentials --all-features --locked
cargo check -p b10x-llm-credentials --no-default-features --locked
cargo test -p b10x-llm-cost --all-features --locked
cargo check -p b10x-llm-cost --no-default-features --locked
cargo fmt --all --check
cargo clippy --workspace --all-targets --all-features --locked -- -D warnings

The two --no-default-features checks matter: they prove the optional file, keychain and SQLite dependencies really are optional, so a consumer that wants none of them pays for none of them.

task rust runs this stage alone, without requiring the planning tools.

2. Specification​

ess specify validate --path spec

The ESS domains under spec/domains/ give the repository's nouns typed homes. There are ten: accounting, budget, catalog, chat, hosting, inference, messages, responses, routing and secrets.

Nine of the ten have an authored scenario suite under contracts/. The exception is catalog, which describes declaration records rather than observations. One naming trap is worth knowing before you go looking: the suite directory is contracts/pricing/, but the domain it exercises is accounting, whose file is spec/domains/accounting.yaml. There is no pricing domain.

Every domain other than catalog describes adapter observations of real library calls. None of them is a claim that this repository publishes runtime events or runs a persistent job.

3. Conformance​

cargo run --locked -p b10x-llm-conformance -- check

The conformance runner checks generated-schema drift, then executes the authored scenarios against the real public library functions three times. Its report gate refuses missing coverage, failures, errors, unsupported observations and skips — a skipped scenario is a red verdict, not a caveat.

The adapter calls the real functions and exposes what they returned. It never reimplements the logic under test, never reads the suite and never branches on a scenario name.

Falsification is part of the evidence​

For each asserted behaviour, the production source is deliberately broken, a named scenario is shown to fail, and the source is restored byte for byte. Which mutation killed which scenario is recorded in docs/verification/*-falsification.json. A suite that still passes against a broken implementation is testing nothing, so this is checked rather than asserted.

4. Plan​

aep plan artifact validate

The AEP store under .engineering/planning/ owns lifecycle state. A story is not complete because the workspace builds.

What a green gate does not establish​

It does not establish live provider access, a working subscription presentation, OS keychain service availability, hosting-provider qualification or a released artifact. Local fixture evidence is local fixture evidence. See Limitations.