Luigit
repositories / smith

smith

There are many coding harnesses - but this one is fast

owned by admin

docs/dependencies.md

Raw
Rendered preview

Dependency reports

Luci job dependencies is configured for 04:00 UTC daily, publishing successful runs under smith/dependencies/<revision>/ after deployment.

Published evidence contains the report, audit, and all three resolution logs with explicit exit-status files.

For a current local report, run cargo x deps; cargo x deps --offline uses cached registry evidence only and cannot establish upstream freshness.

Downstream resolution checks run with cargo x deps --verify locked, cargo x deps --verify fresh-compatible, or cargo x deps --verify direct-minimum; append -p smith-core to scope any check to one runtime library, and --target <triple> to check a supported target (either option may come first).

locked checks selected workspace libraries against checkout Cargo.lock; without -p, it selects exactly the eight runtime libraries. Consumer modes create isolated path-based consumers under OS temporary space, deliberately generate consumer lockfiles, then check --locked, leaving checkout lock untouched. --target applies only to Cargo check, never lockfile generation; output names actual checked target.

fresh-compatible proves Cargo can resolve current compatible requirements; direct-minimum asks installed nightly Cargo to resolve each direct registry dependency from its original requirement using -Zdirect-minimal-versions, leaving transitive engine resolution to Cargo.

direct-minimum requires an installed nightly Rust toolchain; the command never installs one, and checks the resulting lockfile with stable Cargo.

Without --target, resolution checks explicitly compile the Cargo host target, overriding ambient CARGO_BUILD_TARGET; retaining target-specific declarations does not prove other targets. Choose an installed target supported by the selected dependencies and toolchain; the command does not install targets. Foreign Linux GNU/musl targets use rustup run stable cargo-zigbuild check, supplying Zig's C compiler, archiver, and sysroot without system package installs. Local cross checks require cargo-zigbuild and Zig on PATH; missing tools fail explicitly. Native host checks remain ordinary stable Cargo checks.

Luci job dependency-targets is configured for 04:30 UTC daily after deployment, publishing successful evidence under smith/dependency-targets/<revision>/<target>/. All three modes check all eight runtime libraries on Linux ARM64 using a digest-pinned Zig-enabled image, verified as non-root. No runtime apt-get, GCC package installation, or custom Containerfile support is required. Windows/macOS CI targets are deferred until supported runners/tooling exist; earlier core-only probes do not establish full-target compatibility. Build outputs are not cached wholesale: full target/ exceeds hosted cache budget. Cross-target Cargo checks prove compilation, not binary linking, runtime behavior, or frontend usability. Failure artifacts retain each attempted mode's log and exit status.

Windows ARM64 minimum regression requires fs2 0.4.3: older fs2 releases use winapi 0.2, which lacks ARM64 pointer types.

Host resolution does not verify guest tooling; guest build and component loading remain separate gates. The Responses codec guest uses the existing serde_json and wit-bindgen release set; native state/partition properties use proptest, allocation bounds use smith-alloc. It is built separately by the guest registry, tested through the sealed component host, and serves explicitly selected OpenAI subscription profiles without API-key billing fallback.

The OAuth guest declares public-client policy through secret-proxy.authorize; smith-harness owns callback, bounded HTTP, JWT verification, and atomic profile storage. The host crypto facade owns ring = 0.17.14 and base64 = 0.23.1; no token response crosses WIT. Synthetic binary tests cover login, metadata listing, renewable credential-source binding, reauthorization, serialized refresh, logout, and subscription dispatch. Live subscription entitlement remains unverified.

Failed CI runs retain target/dependencies/ as failure evidence, but do not publish a successful site report.

# Dependency reports

Luci job `dependencies` is configured for 04:00 UTC daily, publishing successful runs under `smith/dependencies/<revision>/` after deployment.

Published evidence contains the report, audit, and all three resolution logs with explicit exit-status files.

For a current local report, run `cargo x deps`; `cargo x deps --offline` uses cached registry evidence only and cannot establish upstream freshness.

Downstream resolution checks run with `cargo x deps --verify locked`, `cargo x deps --verify fresh-compatible`, or `cargo x deps --verify direct-minimum`; append `-p smith-core` to scope any check to one runtime library, and `--target <triple>` to check a supported target (either option may come first).

`locked` checks selected workspace libraries against checkout `Cargo.lock`; without `-p`, it selects exactly the eight runtime libraries.
Consumer modes create isolated path-based consumers under OS temporary space, deliberately generate consumer lockfiles, then check `--locked`, leaving checkout lock untouched.
`--target` applies only to Cargo check, never lockfile generation; output names actual checked target.

`fresh-compatible` proves Cargo can resolve current compatible requirements; `direct-minimum` asks installed nightly Cargo to resolve each direct registry dependency from its original requirement using `-Zdirect-minimal-versions`, leaving transitive engine resolution to Cargo.

`direct-minimum` requires an installed nightly Rust toolchain; the command never installs one, and checks the resulting lockfile with stable Cargo.

Without `--target`, resolution checks explicitly compile the Cargo host target, overriding ambient `CARGO_BUILD_TARGET`; retaining target-specific declarations does not prove other targets.
Choose an installed target supported by the selected dependencies and toolchain; the command does not install targets.
Foreign Linux GNU/musl targets use `rustup run stable cargo-zigbuild check`, supplying Zig's C compiler, archiver, and sysroot without system package installs.
Local cross checks require `cargo-zigbuild` and Zig on PATH; missing tools fail explicitly.
Native host checks remain ordinary stable Cargo checks.

Luci job `dependency-targets` is configured for 04:30 UTC daily after deployment, publishing successful evidence under `smith/dependency-targets/<revision>/<target>/`.
All three modes check all eight runtime libraries on Linux ARM64 using a digest-pinned Zig-enabled image, verified as non-root.
No runtime `apt-get`, GCC package installation, or custom Containerfile support is required.
Windows/macOS CI targets are deferred until supported runners/tooling exist; earlier core-only probes do not establish full-target compatibility.
Build outputs are not cached wholesale: full `target/` exceeds hosted cache budget.
Cross-target Cargo checks prove compilation, not binary linking, runtime behavior, or frontend usability.
Failure artifacts retain each attempted mode's log and exit status.

Windows ARM64 minimum regression requires fs2 0.4.3: older fs2 releases use winapi 0.2, which lacks ARM64 pointer types.

Host resolution does not verify guest tooling; guest build and component loading remain separate gates.
The Responses codec guest uses the existing `serde_json` and `wit-bindgen` release set; native state/partition properties use `proptest`, allocation bounds use `smith-alloc`.
It is built separately by the guest registry, tested through the sealed component host, and serves explicitly selected OpenAI subscription profiles without API-key billing fallback.

The OAuth guest declares public-client policy through `secret-proxy.authorize`; `smith-harness` owns callback, bounded HTTP, JWT verification, and atomic profile storage.
The host crypto facade owns `ring = 0.17.14` and `base64 = 0.23.1`; no token response crosses WIT.
Synthetic binary tests cover login, metadata listing, renewable credential-source binding, reauthorization, serialized refresh, logout, and subscription dispatch.
Live subscription entitlement remains unverified.

Failed CI runs retain `target/dependencies/` as failure evidence, but do not publish a successful site report.