Imported from vanyastaff/nebula (tools/xtask/AGENTS.md). Install upstream with npx skills add vanyastaff/nebula --skill xtask. Copyright stays with the author.

AGENTS.md — nebula-xtask

Read the repository-root AGENTS.md first.

Purpose

Repository-only automation that derives CI package plans from Cargo metadata and validates versioned post-selection quality-gate policy. It is not a product crate and does not participate in the product layer map.

Invariants

• No dependency may be a nebula-* product package.
• A successful ci-plan writes one compact, deterministic, versioned JSON plan. --help and --version are successful human-readable stdout; invalid CLI usage uses Clap's stderr and exit code. Planner failures emit no partial stdout.
• Workspace members, package ownership, dependency edges, and declared features come from Cargo metadata. Never add a hardcoded package list or infer a Cargo package name from a directory name.
• Diff uncertainty widens to the full workspace. Invalid nonempty Git revisions remain hard errors so configuration failures are visible.
• package.metadata.nebula.ci.test-features affects tests only; it never changes check, documentation, or dependency resolution policy.
• Consumers may name packages in an independent, documented gate policy only after plan selection. The current no-default-feature policy names nebula-resilience, nebula-log, nebula-expression, nebula-credential, nebula-resource, and nebula-storage; they never influence selector membership.
• north-star-gates validate reads the repository-owned registry, evidence schema, canonical multi-run evidence, and checked-in workflow job IDs only. It is post-selection gate policy: it never selects packages or changes the ci-plan result.
• North Star registry v1 rejects passed; its checked-in state is only the baseline. The runtime-authority verifier accepts a complete immutable artifact inventory only when runner provenance and every semantic threshold match, then emits the deterministic effective partial state for each covered gate. A failed verification emits no effective-state result.
• runtime-repair-red verify is expected-failure evidence policy, not a test runner. It accepts only the raw nextest test-failure exit, exact manifest identities, ordinary failures, and exact reason markers from bounded JUnit. It never accepts ignored, skipped, retried, timed-out, sentinel, or synthetic production evidence.

Verification

cargo nextest run -p nebula-xtask
cargo clippy -p nebula-xtask --all-targets -- -D warnings
cargo xtask ci-plan full | jq .
cargo xtask north-star-gates validate
cargo xtask runtime-repair-red validate-manifest