modbison/docs/planning/trust_safety.md

# Trust & Safety Operations — Plan

> **Status:** Planned
> **Owner:** Platform team
> **Feature ID:** `trust_safety`
> **Godot version:** 4.6.2 stable
> **Backend:** FastAPI + SQLAlchemy (async) + PostgreSQL (+ Redis available, unused)
> **Related:** `content_moderation.md` (M5 — content rating/flagging at upload time),
> `social.md` (chat + report-message), `app/dependencies.py` (JWT auth)

## 1. Overview

This feature is the **enforcement workflow** for content/behavior that is
**already live** and has been reported by a real user. M5
(`content_moderation.md`) covers rating/flagging at authoring/upload time; this
feature covers what happens **after** something is live and someone reports it:

1. **Reporting** — report a user, an experience (world/asset), or a chat
   message with a reason and optional detail.
2. **Moderation queue** — staff review incoming reports, see report + target
   context, and resolve them.
3. **Ban / suspension** — apply a temporary suspension or permanent ban to a
   user, with enforcement at the auth boundary.
4. **Appeals** — a sanctioned user can appeal; staff can uphold or overturn.

The userbase being 18+ does **not** shrink the moderation need — harassment,
hate speech, spam, griefing, and CSAM/policy violations still require real
enforcement. Safety is a **blocking requirement**, not an add-on.

## 2. Existing constraints (do not re-build)

- **Auth:** JWT bearer; `get_current_user` / `CurrentUser` in
  `app/dependencies.py`. Every new endpoint reuses it.
- **DB:** async SQLAlchemy + Alembic. Current head is `0002_add_date_of_birth_to_users`.
- **Models:** `User`, `World` (owner_id), `Asset` (owner_id, is_public). There
  is no `role` field yet, and no chat/message table yet (social.md is planned,
  not built).
- **Redis:** configured (`REDIS_URL`) but unused; not required by this feature.
- **Client:** Godot 4.6.2, minimal `main_menu.gd`; no report/chat UI yet.

## 3. Key design decisions

| Concern | Decision |
|---|---|
| Report target addressing | Generic `target_type` + `target_id` (string) so user/world/asset/message all flow through one report path; router resolves target existence via a small registry |
| Report reasons | Fixed enum `reason_category` (harassment, hate_speech, sexual_content, violence, spam, self_harm, other) + free-text `details` |
| Staff roles | Add `role` enum on `User` (`user` / `moderator` / `admin`); `RequireStaff` dependency gates the queue/action/appeal-review endpoints |
| Enforcement | Denormalized `banned_until` / `suspended_until` timestamps on `User` (kept in sync by moderation actions); `get_current_user` rejects sanctioned users with `403` so enforcement applies to **all** existing endpoints |
| Ban vs. suspension | Suspension = temporary (`expires_at`); ban = permanent (`expires_at IS NULL`). Both are rows in `moderation_actions` for auditability |
| Appeals | Appeal references the specific `moderation_action`; approving an appeal revokes the action and clears the user's sanction |
| Audit | Every moderation action is an append-only row with `moderator_id`, `reason`, `created_at`, `revoked_at`; reports record `resolved_by` + `resolution` |

## 4. Data model (target shape)

```
users
  + role             ENUM(user, moderator, admin)  default 'user'
  + banned_until     TIMESTAMPTZ  NULL
  + suspended_until  TIMESTAMPTZ  NULL

reports
  id, reporter_id FK users, target_type ENUM(user,world,asset,chat_message),
  target_id VARCHAR(64), reason_category ENUM(...), details TEXT,
  status ENUM(open,in_review,resolved,dismissed) default 'open',
  resolved_by FK users NULL, resolution ENUM(...) NULL,
  created_at, updated_at, resolved_at

moderation_actions
  id, moderator_id FK users, target_user_id FK users,
  target_type ENUM(user,world,asset,chat_message), target_id VARCHAR(64),
  action ENUM(warn,suspend,ban,restrict,remove,restore),
  reason TEXT, report_id FK reports NULL,
  expires_at TIMESTAMPTZ NULL, revoked_at TIMESTAMPTZ NULL, created_at

appeals
  id, moderation_action_id FK moderation_actions,
  user_id FK users, reason TEXT,
  status ENUM(pending,approved,rejected) default 'pending',
  reviewed_by FK users NULL, reviewed_at TIMESTAMPTZ NULL, created_at
```

## 5. User stories

### US-1 Reporting
As a player, I can report a user, an experience (world/asset), or a chat
message, choosing a reason and adding optional detail.

**Acceptance criteria**
- Report requires a valid target of the matching type (404 otherwise) and a
  reason from the fixed category set.
- Duplicate reports from the same reporter for the same target are coalesced
  (409) rather than creating unbounded duplicates.
- A report cannot be filed by a banned/suspended user (enforced by existing auth).
- Each report stores reporter, target, reason, and created_at; status starts `open`.

### US-2 Moderation queue & review
As a moderator, I can list open reports (filterable by type/status/reason),
view a report with its target, and resolve it with an action.

**Acceptance criteria**
- Queue is staff-only (`RequireStaff`); non-staff get 403.
- List supports status/type/reason filters + pagination (offset/limit).
- Resolving sets `status`, `resolution`, `resolved_by`, `resolved_at` and
  (where applicable) records a `moderation_actions` row.

### US-3 Ban / suspension enforcement
As a moderator, I can suspend a user temporarily or ban them permanently, and
the sanction blocks the user from authenticated endpoints immediately.

**Acceptance criteria**
- Suspension sets `suspended_until`; ban sets `banned_until` (permanent). Warn
  records an action without blocking.
- `get_current_user` returns `403` for any user past a live sanction.
- Expired suspensions no longer block (time-based, no cron required).
- Removing a ban/suspension is itself a recorded `restore` action.

### US-4 Appeals
As a sanctioned user, I can appeal an active ban/suspension; as a moderator I
can approve (overturn) or reject (uphold) it.

**Acceptance criteria**
- A user can submit exactly one open appeal per sanction.
- Appeal is tied to the specific `moderation_action`.
- Approving revokes the action and clears the user's sanction; rejecting
  upholds it. Both record reviewer + timestamp.

## 6. Task breakdown

Right-sized to ~1–3 files / one coherent piece of functionality. Task IDs are
dependency order.

| # | Task | Story | Depends on |
|---|------|-------|-----------|
| 1 | Moderation data model + Alembic migration (reports, moderation_actions, appeals, user `role`/`banned_until`/`suspended_until`) | US-1..4 | — |
| 2 | Staff authorization + sanction enforcement dependency (`RequireStaff`; extend `get_current_user` to reject sanctioned users) | US-2, US-3 | 1 |
| 3 | Report submission endpoints (report user / experience / chat message) | US-1 | 1 |
| 4 | Moderation queue + review endpoints (list, detail, resolve→action) | US-2 | 1, 2 |
| 5 | Ban/suspend/restore action application + user-sanction sync | US-3 | 2, 4 |
| 6 | Appeals endpoints (submit, list, review → uphold/overturn) | US-4 | 1, 5 |
| 7 | Integration tests: report → review → ban → appeal → overturn | US-1..4 | 3, 4, 5, 6 |

## 7. Phasing

- **Phase 1 (data + enforcement core):** tasks 1–2. Land the schema and make
  sanctioned users actually unable to act — highest value.
- **Phase 2 (report → review → action):** tasks 3–5. The moderation loop.
- **Phase 3 (appeals):** task 6.
- **Phase 4 (verification):** task 7 integration tests.

## 8. Out of scope for this feature

- **Client (Godot) report/appeal UI** — no chat/social UI exists yet; report
  endpoints are consumed once `social.md` US-6 lands. Client is a follow-up.
- Automated content scanning / hash detection (covered by M5 /
  `content_moderation.md`).
- Third-party moderation/CSAM-scanning vendors.
- Rate limiting of report submission (Redis is available; add a token-bucket
  dependency later if abuse is observed).
- Voice chat moderation.
- Per-jurisdiction legal-takedown workflows (DMCA etc.).