Visitor Pool Security

The visitor pool provides anonymous, sandboxed access to SciTeX Hub for users who have not signed up. Four pre-allocated accounts (visitor-001 to visitor-004) are shared in rotation.

This page covers the security measures added on 2026-03-24 to prevent chat history and project files from leaking between consecutive visitors on the same slot.

Background: the fix

Before the fix, a visitor who reused a recycled slot could see the previous visitor’s chat sessions in the UI. The root cause was that ChatSession records were not deleted on deallocation.

Three layers of protection were added.

Layer 1 — `@reset_workspace_after`

Applied to PoolAllocator.deallocate_visitor().

After the slot is marked inactive the decorator calls WorkspaceManager.reset_visitor_workspace(user) immediately. This is the primary cleanup path for normal session endings (expiry, sign-up, or explicit logout).

# apps/infra/project_app/services/visitor_pool/decorators.py

@reset_workspace_after
def deallocate_visitor(cls, session): ...

Layer 2 — `@ensure_clean_workspace`

Applied to PoolAllocator._try_allocate_slot().

Before handing a slot to a new visitor the decorator checks whether the slot was previously used and, if so, calls the same workspace reset. This catches edge cases where Layer 1 did not run: server crash, Docker restart, idle-timeout expiry, or NAS reboot.

# apps/infra/project_app/services/visitor_pool/decorators.py

@ensure_clean_workspace
def _try_allocate_slot(cls, visitor_num, session, pool_size): ...

Layer 3 — `reset_visitor_pool` on container restart

deployment/docker/docker_prod/entrypoint.sh runs the management command in the background immediately after gunicorn binds:

python manage.py reset_visitor_pool   # wipe all slots
python manage.py create_visitor_pool  # ensure accounts exist

No state survives a container bounce.

What gets cleared

WorkspaceManager.reset_visitor_workspace(user) deletes and re-creates the following for each visitor account:

Project DB row — deleted then re-created from the scitex_minimal template via scitex.template.clone_template().
Project filesystem — shutil.rmtree + fresh template clone.
ChatSession / ChatMessage — deleted by _clear_visitor_data().
LLMUsageLog — deleted by _clear_visitor_data().
VisitorAllocation — marked is_active=False.

Pool lifecycle diagram

Container start
    → reset_visitor_pool        # hard reset all 4 slots
    → create_visitor_pool       # ensure accounts/projects exist

Visitor arrives
    → @ensure_clean_workspace   # layer 2 safety net
    → allocate_visitor()        # DB lock prevents race conditions

Visitor uses platform (~1 hour)

Session ends
    → deallocate_visitor()
    → @reset_workspace_after    # layer 1 immediate cleanup

Operations

# Reset all slots (run automatically on restart)
python manage.py reset_visitor_pool

# Free expired allocations only (no workspace wipe)
python manage.py reset_visitor_pool --free-expired

# Pool health (free / total)
make status

Visitor Pool Security

Background: the fix

Layer 1 — @reset_workspace_after

Layer 2 — @ensure_clean_workspace

Layer 3 — reset_visitor_pool on container restart