TenantAtlas/.specify/memory/constitution.md

<!--
Sync Impact Report

- Version change: 1.4.0 → 1.5.0
- Modified principles:
  - Tenant Isolation is Non-negotiable (added deny-as-not-found requirement)
- Added sections:
  - RBAC Standard (RBAC-001..RBAC-009)
- Removed sections: None
- Templates requiring updates:
  - ✅ .specify/templates/plan-template.md
  - ✅ .specify/templates/spec-template.md
  - ✅ .specify/templates/tasks-template.md
  - N/A: .specify/templates/commands/ (directory not present)
- Follow-up TODOs: None
-->

# TenantPilot Constitution

## Core Principles

### Inventory-first, Snapshots-second
- All modules MUST operate primarily on Inventory as “last observed” state.
- Inventory is the source of truth for what TenantPilot last observed; Microsoft Intune remains the external truth.
- Snapshots/Backups MUST be explicit actions (manual or scheduled) and MUST remain immutable.

### Read/Write Separation by Default
- Analysis, reporting, and monitoring features MUST be read-only by default.
- Any write/change function (restore, remediation, promotion) MUST include preview/dry-run, explicit confirmation, audit logging, and tests.
- High-risk policy types default to `preview-only` restore unless explicitly enabled by a feature spec + tests.

### Single Contract Path to Graph
- All Microsoft Graph calls MUST go through `GraphClientInterface`.
- Object types and endpoints MUST be modeled first in the contract registry (`config/graph_contracts.php`).
- Feature code MUST NOT hardcode “quick endpoints” or bypass contracts.
- Unknown/missing policy types MUST fail safe (preview-only / no Graph calls) rather than guessing endpoints.

### Deterministic Capabilities
- Backup/restore/risk/support flags MUST be derived deterministically from config/contracts via a Capabilities Resolver.
- The resolver output MUST be programmatically testable (snapshot/golden tests) so config changes cannot silently break behavior.

### Tenant Isolation is Non-negotiable
- Every read/write MUST be tenant-scoped.
- Cross-tenant views (MSP/Platform) MUST be explicit, access-checked, and aggregation-based (no ID-based shortcuts).
- Prefer least-privilege roles/scopes; surface warnings when higher privileges are selected.
- A non-member attempting to access a tenant route MUST be deny-as-not-found (404).

### RBAC Standard (RBAC-001)

RBAC-001 Two Planes
- The platform MUST maintain two strictly separated authorization planes:
  - Tenant plane (`/admin/t/{tenant}`): authenticated Entra users (`users`), authorization is tenant-scoped.
  - Platform plane (`/system`): authenticated platform users (`platform_users`), authorization is platform-scoped.
- Cross-plane access MUST be deny-as-not-found (404) (not 403) to avoid route enumeration.

RBAC-002 Capabilities-first Authorization
- Feature code MUST NOT check raw roles directly (e.g. string role comparisons).
- Feature code MUST check capabilities via Gates/Policies only.
- A canonical capability registry MUST exist as the single source of truth (e.g. `TenantCapabilities` / `PlatformCapabilities`).
- Role → capability mapping MUST reference only registry entries.

RBAC-003 Least Privilege Role Semantics
- Tenant roles MUST follow least-privilege semantics:
  - Readonly: view-only; MUST NOT start operations and MUST NOT mutate data.
  - Operator: MAY start allowed tenant operations; MUST NOT manage credentials, settings, members, or perform destructive actions.
  - Manager: MAY manage tenant configuration and start operations; MUST NOT manage tenant memberships (Owner-only).
  - Owner: MAY manage memberships and all tenant configuration; Owner-only “danger zone” actions MUST remain Owner-only.

RBAC-004 UI is not Security
- Hiding UI elements is NOT sufficient.
- Every mutation endpoint and action MUST enforce authorization server-side (Policy/Gate).

RBAC-005 Destructive Actions Gate
- All destructive actions (delete / force delete / irreversible operations) MUST:
  - require an explicit confirmation (e.g., `requiresConfirmation()` or equivalent),
  - be protected by a Policy/Gate,
  - have at least one regression test asserting the action is forbidden for non-authorized roles.

RBAC-006 Membership Safety Rule
- The system MUST prevent removing or demoting the last remaining Owner of a tenant.

RBAC-007 Tenant Isolation
- All tenant-plane queries MUST be tenant-scoped.
- A non-member attempting to access a tenant route MUST be deny-as-not-found (404).

RBAC-008 Auditing
- All access-control relevant changes MUST write `AuditLog` entries with stable action IDs, including:
  - membership add / role change / remove
  - provider credential rotation / connection disable
  - break-glass enter / exit / expire (platform plane)
- `AuditLog` entries MUST be redacted (no secrets/tokens, minimal identity fields).

RBAC-009 Testability Gate
- Any new feature that introduces or changes authorization MUST include:
  - at least one positive test (authorized user can do it),
  - at least one negative test (unauthorized user cannot do it),
  - and MUST NOT introduce role-string checks outside the central mapping/registry.

### Operations / Run Observability Standard
- Every long-running or operationally relevant action MUST be observable, deduplicated, and auditable via Monitoring → Operations.
- An action MUST create/reuse a canonical `OperationRun` and execute asynchronously when any of the following applies:
  1. It can take > 2 seconds under normal conditions.
  2. It performs remote/external calls (e.g., Microsoft Graph).
  3. It is queued or scheduled.
  4. It is operationally relevant for troubleshooting/audit (“what ran, who started it, did it succeed, what failed?”).
- Actions that are DB-only and typically complete in < 2 seconds MAY skip `OperationRun`.
- OPS-EX-AUTH-001 — Auth Handshake Exception:
  - OIDC/SAML login handshakes MAY perform synchronous outbound HTTP (e.g., token exchange) without an `OperationRun`.
  - Rationale: interactive, session-critical, and not a tenant-operational “background job”.
  - Guardrail: outbound HTTP for auth handshakes is allowed only on `/auth/*` endpoints and MUST NOT occur on Monitoring/Operations pages.
- If an action is security-relevant or affects operational behavior (e.g., “Ignore policy”), it MUST write an `AuditLog` entry
  including actor, tenant, action, target, before/after, and timestamp.
- The `OperationRun` record is the canonical source of truth for Monitoring (status, timestamps, counts, failures),
  even if implemented by multiple jobs/steps (“umbrella run”).
- “Single-row” runs MUST still use consistent counters (e.g., `total=1`, `processed=0|1`) and outcome derived from success/failure.
- Monitoring pages MUST be DB-only at render time (no external calls).
- Start surfaces MUST NOT perform remote work inline; they only: authorize, create/reuse run (dedupe), enqueue work,
  confirm + “View run”.
- Active-run dedupe MUST be enforced at DB level (partial unique index/constraint for active states).
- Failures MUST be stored as stable reason codes + sanitized messages; never persist secrets/tokens/PII/raw payload dumps
  in failures or notifications.
- Graph calls are allowed only via explicit user interaction and only when delegated auth is present; never as a render side-effect (restore group mapping is intentionally DB-only).
- Monitoring → Operations is reserved for `OperationRun`-tracked operations.
- Scheduled/queued operations MUST use locks + idempotency (no duplicates).
- Graph throttling and transient failures MUST be handled with backoff + jitter (e.g., 429/503).

### Data Minimization & Safe Logging
- Inventory MUST store only metadata + whitelisted `meta_jsonb`.
- Payload-heavy content belongs in immutable snapshots/backup storage, not Inventory.
- Logs MUST not contain secrets/tokens; monitoring MUST rely on run records + error codes (not log parsing).

### Badge Semantics Are Centralized (BADGE-001)
- Status-like badges (status/outcome/severity/risk/availability/boolean signals) MUST render via `BadgeCatalog` / `BadgeRenderer`.
- Filament resources/pages/widgets/views MUST NOT introduce ad-hoc status-like badge mappings (use a `BadgeDomain` instead).
- Introducing or changing a status-like value MUST include updating the relevant badge mapper and adding/updating tests for the mapping.
- Tag/category chips (e.g., type/platform/environment) are not status-like and are not governed by BADGE-001.

### Spec-First Workflow
- For any feature that changes runtime behavior, include or update `specs/<NNN>-<slug>/` with `spec.md`, `plan.md`, `tasks.md`, and `checklists/requirements.md`.
- New work branches from `dev` using `feat/<NNN>-<slug>` (spec + code in the same PR).

## Quality Gates
- Changes MUST be programmatically tested (Pest) and run via targeted `php artisan test ...`.
- Run `./vendor/bin/pint --dirty` before finalizing.

## Governance

### Scope & Compliance
- This constitution applies across the repo. Feature specs may add stricter constraints but not weaker ones.
- Restore semantics changes require: spec update, checklist update (if applicable), and tests proving safety.

### Amendment Procedure
- Propose changes as a PR that updates `.specify/memory/constitution.md`.
- The PR MUST include a short rationale and list of impacted templates/specs.
- Amendments MUST update **Last Amended** date.

### Versioning Policy (SemVer)
- **PATCH**: clarifications/typos/non-semantic refinements.
- **MINOR**: new principle/section or materially expanded guidance.
- **MAJOR**: removing/redefining principles in a backward-incompatible way.

**Version**: 1.5.0 | **Ratified**: 2026-01-03 | **Last Amended**: 2026-01-27