Permissions and safety

Agent safety and permissions

ForgeLore lets agents perform normal product actions when their token, organization role, and repository access allow it. High-impact actions must be explicit, dry-run first, and human-confirmed.

Default stance

Agents can operate the product. They cannot silently perform irreversible or account-level destructive actions.

Allowed with scope

Read repository state, list activity, inspect locks, create summaries, mint connect tokens, and manage file locks.

Scoped by token

AI Access tokens grant repository access and scopes such as `repo:read`, `locks:write`, and `connect-token:create`.

Confirm before impact

Force unlocks, permanent deletes, subscription changes, payment entry, provider secrets, and infrastructure changes require human confirmation.

Recover safely

Every structured error should tell the agent whether a retry is safe and what to do next.

Human-confirmed actions

Do not let an autonomous agent continue without explicit confirmation for these actions:

  • Entering payment details or changing a subscription.
  • Deleting an organization or permanently deleting a repository.
  • Force unlocking assets owned by another user.
  • Exporting provider secrets, private keys, bearer tokens, or checkout URLs.
  • Changing production infrastructure or provider configuration.
  • Running destructive bulk actions across repositories, users, locks, or tokens.

Write-operation checklist

  1. Inspect the capability manifest and required scopes.
  2. Run a dry-run when the operation supports it.
  3. Use a stable idempotency key for any retryable write.
  4. Check repository locks before editing Unreal binary assets.
  5. Stop on `safeToRetry: false` and report the recovery fields.