Intake / architecture
Architect records scope, risks and the smallest safe design.
Independent concept for discussion.
No live execution or connected systems.
Claude Code prepares a scoped change. Specialist agents supply evidence. People review it; independent CI decides whether it can ship.
Architect records scope, risks and the smallest safe design.
A reusable skill turns intent into an explicit plan.
Implementation, tests and security review in scoped contexts.
Diff, test evidence and unresolved risks become PR #142.
Alex reviews only after independent PR checks pass.
Scope the change before tools can act.
CLAUDE.md
Project instructions loaded into Claude’s context. Guidance, not enforceable access control.
.claude/skills/change-plan/SKILL.md
On-demand process, not an always-running agent. This example is manually invoked with /change-plan.
.claude/agents/test-engineer.md
Separate context and scoped tools. Delegation uses the description; capabilities remain subject to permissions.
.claude/settings.json
Illustrative schema. Hook scripts must be supplied and reviewed; local hooks are NOT a security boundary.
Allowlisted servers, scoped identities and tool permissions. No servers connected in this concept. Never grant arbitrary remote shell or production writes.
View independent controlsLocal hooks add early feedback, not a security perimeter. Branch protection, CI scans and isolated identity independently enforce delivery.
ENG-104 opened. Synthetic release; no live execution.