v2.5.0 — Generic file intake (upload → extract → suggest → pre-fill)

Backward-compatible feature release. Opt-in; defaults unchanged. 2215 tests passing.

File intake — upload → extract → suggest create → pre-fill

A generic, entity-agnostic document pipeline built on FileAnalysisService.

analyze_file tool (#46) — extract a stored upload's text (PDF / Word / txt / csv / xlsx via DocumentService) and suggest the create action it implies. Driven by config keyword patterns → any create_<entity>, so the same mechanism handles invoices, customers, products, or any registered resource — not invoice-specific.

• Secure: sandboxed to file_analysis.base_path (path traversal / out-of-sandbox / absolute paths rejected — no arbitrary file reads), extension allowlist, size cap.
• Suggestions filtered to create tools that actually exist (validate_actions).

Field pre-fill + vision (#47)

• StructuredFileExtractor extracts each suggested create action's own declared fields (its tool parameters) into a ready payload — generic, so a new AiResource/create tool gets upload-to-prefill with zero extra code. Returns {action_id: 'create_invoice', prefill: {customer_name, customer_email, items: [...]}}.
• Images / scans (jpg/png/…) are OCR'd by a vision model (vision_model) and run the same suggest + pre-fill path — a photographed invoice behaves like a PDF.

'file_analysis' => [
    'enabled' => true,
    'prefill' => true, // also extract fields to pre-fill the create
    'base_path' => storage_path('app/uploads'),
    'keyword_suggestions' => [
        ['pattern' => '/\binvoice\b/i',  'action_id' => 'create_invoice'],
        ['pattern' => '/\bcustomer\b/i', 'action_id' => 'create_customer'],
    ],
],

Upgrade notes

No breaking changes. The feature is opt-in (ai-engine.file_analysis.enabled / .prefill). Model + vision calls hit a live provider at runtime; in tests they're mocked. See docs/agent-resources.mdx → "File intake".