# Backend: `s3` (planned)

Status: **STUB** — not yet implemented. Tracked at [#962](https://git.integrolabs.net/roctinam/aiwg/issues/962).

## Planned design

S3-compatible object store, primarily intended for **bulk artifacts** (research corpus, media caches) rather than memory artifacts where latency and lack of full-text query are problematic.

When implemented:
- Built on `@aws-sdk/client-s3` v3
- Authenticates via the AWS default credential chain (env, instance profile, SSO, etc.) — never via `storage.config`
- Supports S3-compatible endpoints (Cloudflare R2, MinIO, Backblaze B2) via the `endpoint` field
- `read` via `GetObject`, `write` via `PutObject`, `list` via `ListObjectsV2`

## Anticipated configuration

```jsonc
{
  "version": "1",
  "backends": {
    "research": {
      "type": "s3",
      "bucket": "my-aiwg-archive",
      "prefix": "aiwg-research/",
      "region": "us-east-1",
      "endpoint": "https://s3.amazonaws.com"  // optional; for non-AWS endpoints
    }
  }
}
```

Credentials come from the AWS default credential chain. `storage.config` will never accept `accessKeyId` / `secretAccessKey` (the schema rejects them).

## Why deferred

S3 doesn't fit the memory-shaped subsystems (no full-text query, latency for many small reads/writes). It's slated for the bulk-artifact subsystems (research, media). Implementation is gated on phase-3 work.

## How to track / contribute

- Tracking issue: [#962](https://git.integrolabs.net/roctinam/aiwg/issues/962)
- Design: `.aiwg/architecture/storage-design.md` (§5.7)