[joaoguimaraespro]

What and why

Adds a spec-driven workflow so features can be built by AI agents with a human at the gates:

• Spec Kit 1.2.2 (Claude Code default, Codex, Cursor): /speckit-specify → /speckit-plan → /speckit-tasks produce specs/<NNN-feature>/{spec,plan,tasks}.md.
• Constitution (.specify/memory/constitution.md) derived from CONTRIBUTING.md and the ADRs: read-only brokers, decimal money, test-first, module boundaries, no personal data.
• AGENTS.md for every agent, and CLAUDE.md importing it.
• Subagents in .claude/agents/, adapted from ECC (MIT) to this stack: tdd-guide, csharp-reviewer, typescript-reviewer, security-reviewer, build-error-resolver.
• .agent-loop.env: the local verification gate for the agent loop (one fresh agent per task; a task counts as done only when this gate passes).
• style(web): three files that had drifted from Prettier, so the gate starts green.

No application code changes.

How it was tested

• Verification gate run on this branch: dotnet build, every backend test project except Integration.Tests (Architecture 19, Finance 110, Investments 51, Reporting 86, Integrations 20 — all passing), npm test (21 files passing) and npm run lint:format (clean).
• Agent loop exercised with a stub agent: refuses main and dirty trees, ticks and commits on green, stops and reports after repeated failures.
• gitleaks on the staged files: no findings.

Checklist

• dotnet test and npm test pass locally (integration tests run in CI)
• Architecture tests untouched
• UI strings added to both en.json and pt-PT.json — n/a, no UI strings
• Docs (and an ADR, for design changes) updated — workflow documented in AGENTS.md; an ADR in docs/adr/ if you want the process recorded here too
• No real financial data, secrets or personal paths in code, tests or screenshots