Skip to content

Latest commit

 

History

History
75 lines (59 loc) · 3.69 KB

RELEASE_RUNBOOK.md

File metadata and controls

75 lines (59 loc) · 3.69 KB

Release runbook

Prepare the candidate

  1. Select one roadmap milestone; record exact APIs, errors, exclusions, resource bounds, corpus entries and test commands in its scope manifest. Update the requirement coverage rows with actual evidence links.
  2. Run cargo xtask freshness, local checks and applicable compiler/platform and assurance gates. Unavailable evidence does not count as a pass.
  3. Update status, README, changelog and release-notes/vX.Y.Z.md.
  4. Before publication, verify all package archives in dependency order in an isolated staging environment. Record graph/SBOM, compiler/profile/corpus and grammar/Unicode revisions, checksums and relevant resource/performance evidence. The foundation local gate verifies core; full-family archive rehearsal is still planned and must complete before the first family publication.

One report committed with the candidate

Follow Brynja's reporting workflow. Use one permanent report at security/pentest/vX.Y.Z[-rc.N].md, starting from the template. Record Version, Status, Open-Findings, Retest, Date, Tester and Scope, followed by methods, findings, remediation, retest results and limitations. No separate commit-hash field is required.

Perform the candidate pentest, fix findings and retest. Commit the final report with the candidate. If CI requires a subsequent change, review/retest it and commit that change with the corresponding report update, then rerun CI. Do not invent a passing assessment or describe automated checks as independent review. A report is evidence from its author; the gate validates its state and Git relationship, not the quality or truth of that assessment.

For a non-release status check:

scripts/release/validate-current-pentest.sh

Missing evidence leaves release blocked. A committed report may say Status: RETEST REQUIRED, Open-Findings: 0, Retest: PENDING while awaiting retest; this never passes the required release gate. Keep ordinary implementation CI free of report freshness requirements; it runs the gate's regression tests.

Release and tag check

After the final report is committed, require:

scripts/release/validate-current-pentest.sh --required

The gate selects the current Cargo version and requires PASS/PASS with zero open findings, a regular committed report, a clean checkout, a current candidate/report update and an absent release tag. The direct equivalent for a named candidate is scripts/release/validate-release-readiness.sh vX.Y.Z.

The release workflow fetches full history and runs this required check. GitHub CodeQL uses Default setup only. The user confirms that CI and CodeQL are green and authorizes tagging. Then create a signed annotated tag with subject Synir vX.Y.Z (lowercase synir is also accepted). Verify before publication:

SYNIR_RELEASE_PUBLISH_TAG=vX.Y.Z scripts/release/validate-release-readiness.sh vX.Y.Z

This context checks the signature, subject and direct HEAD commit target of the existing tag. The gate itself never creates/pushes tags or publishes packages. Publish reviewed packages in order: core, host, macros, facade; retain notes, checksums and provenance, and verify installed registry contents. Keep publishing credentials out of untrusted pull-request jobs. Local development commits remain allowed without a pentest; release tags and publication require the passing gate.

Synir keeps pentests for every release version, including patches and RCs. The adaptation does not add Brynja's checkpoint cadence or exceptional deferral flow. 1.0 also requires complete capability evidence, independent review/remediation, platform runtime acceptance and the admitted production candidate.