Fix vulns at the source.

Hiro Security Review — Clean ✅

Reviewed the diff at /var/folders/sy/.../hiro-diff-2yi0mr3y.patch.

My concerns are all addressed. Safe to ship.

What changed

1. backend/agents/scan_agent.py — adds agent_type="scan" to the streaming call so the scan agent gets the right toolset (e.g. github_analyze_repo).
2. backend/tests/agents/test_scan_agent.py — new test asserting run_scan builds the agent with agent_type="scan" and the right default_memory_source.
3. frontend/src/pages/TaskPage.tsx — pentest validation subtasks (parent_task_id && validation_phase) now stay on the unified TaskPage instead of routing to ScanPage; back-link label adjusted accordingly.

Security analysis

• agent_type="scan": hardcoded literal, not user-controlled. Can’t be flipped by another tenant or input. This adds a needed capability rather than relaxing a guard.
• Frontend routing change: only decides which React component renders. Tenant authz lives server-side on task fetch and is unchanged. isPentestValidationSubtask is derived from the task’s own metadata — if a tenant forges those fields on their own task they just get a different UI, no cross-tenant exposure.
• No new sinks: no dangerouslySetInnerHTML, no raw href={field}, the Link to={...parent_task_id} shape already existed. The added strings ('View pentest run' | 'View parent scan') are static literals.
• Test file: pytest-only, no prod surface.

One non-security nit

In test_scan_agent.py the patch target is agents.chat_agent.run_security_agent_stream. If scan_agent.py does from agents.chat_agent import run_security_agent_stream at module load, patching at that path won’t intercept the call — patch at the import site (agents.scan_agent.run_security_agent_stream) for the conventional shape.