dstack-beads-project-alignment-review

Project alignment review

Tier 1 is read-only for repository source.

1. Initialize or inspect the audit:

   "{baseDir}/../../bin/dstack" ctl alignment initialize --title "<title>" \
     --target-branch <branch> --scope "<scope>"
   

1. Compare current specifications, durable docs, architecture patterns, code, tests, Beads work, and delivery evidence. Treat the current repository and Beads records as the review inputs; do not persist a Git baseline.

2. Decide bounded corrective outcomes, acceptance criteria, priorities, and real dependencies. Create code/test corrections with alignment add-correction; do not create documentation or reconciliation tasks because landing is the sole final reconciliation.

3. Create temporary PLAN.json. Use a canonical JSON plan helper exposed by the current CLI when available; otherwise write the file directly. Markdown record scaffolds are not authoritative plan input.

4. Populate exactly dstack.alignment-plan/v2:

   • schema: dstack.alignment-plan/v2;
   • scope;
   • findings: {title, evidence, rationale} objects;
   • accepted_corrections: {title, description, acceptance, priority, depends_on} objects whose dependency values are accepted correction titles;
   • rejected_corrections: {title, rationale} objects;
   • validation_expectations: strings;
   • documentation_impact: exactly end_user_operator, developer_reviewer, and future_auditor string arrays;
   • deferred_findings: {title, rationale} objects; and
   • accepted_risks: {title, rationale} objects.

   Include every field, use [] for an empty collection, and add no other keys.

5. Finish the canonical plan and stop with the human gate open:

   "{baseDir}/../../bin/dstack" ctl alignment finish-plan AUDIT --plan-file PLAN.json
   

   The controller validates, canonicalizes, stores, rereads, and binds the plan before any authorization state closes.

Return findings, correction graph, decisions required, and /project-alignment-execute <audit>.