The AI chat assistant scoped tool queries by company but ignored the
per-user Bouncer abilities the rest of the app enforces, so any `use ai`
holder could read customers, invoices, payments, and company financials
their role couldn't otherwise see. Each AiTool now declares a required
ability (entity-aligned); the registry hides unauthorized tools from the
model and refuses to execute them as a backstop.
Separately, admin/owner-supplied URLs were fetched server-side with no
guard against private/reserved targets (SSRF): the AI base URL, the
CurrencyConverter "DEDICATED" exchange-rate URL, and S3/Spaces file-disk
endpoints. A shared PrivateNetworkGuard now backs a PublicHttpUrl
validation rule (save-time) and runtime guards in each driver.
- AiTool::requiredAbility() + mapping across all 12 tools
- AiToolRegistry filters schemas() by ability and re-checks in execute()
- PrivateNetworkGuard / BlockedUrlException / PublicHttpUrl rule (new)
- Rule wired into AI config (service + 3 controllers), exchange-rate,
and file-disk endpoints; runtime guards in OpenRouterDriver,
CurrencyConverterDriver, and FileDiskService
- Tests for ability filtering, the guard, the rule, and 422 rejections
Adds three new read-only tools the chat LLM can call to answer
"who/what did the most X" questions that previously fell through
the cracks:
- rank_top_customers — ranks customers by invoiced_total, paid_total,
invoice_count, or outstanding_balance over a named time period
- rank_top_items — ranks catalog items by quantity_sold or revenue
- rank_expense_categories — ranks expense categories by total spend
All three share a new ResolvesPeriod trait that centralizes the
period-name → [start, end] logic. GetCompanyStatsTool is refactored
onto the same trait (identical public schema — the 'all_time' option
is only exposed on the new ranking tools, where an unbounded window
makes sense; stats over "all time" collapses every record into one
giant bucket and is rarely useful).
Each tool follows the existing pattern: snake_case name, one-sentence
description tuned for LLM tool selection, JSON-schema parameters
with injected company scoping (never trusting LLM-supplied company
IDs), and JSON-encodable output. outstanding_balance on the customer
tool explicitly ignores the period param since it's a current-state
snapshot.
Multi-company scoping tests lock down the session-authoritative
boundary on every new tool. Per-metric ordering tests verify the
aggregate queries actually rank correctly, and an ad-hoc-item
exclusion test verifies rank_top_items skips invoice lines where
item_id is null (free-typed entries that have no catalog row to
rank by id).
15 new tests added (tests/Feature/Ai/Tools/); test suite grows from
398 to 413 passing. LLM tool count goes from 9 to 12 — the model
will discover the new tools automatically via the function-calling
schema with no prompt changes required.
Extracts the two inline LLM prompts (AiAssistantService's chat system
prompt and AiTextGenerationService's writing preamble) out of PHP
heredocs and into plain-markdown template files under resources/ai/
prompts/. Each file can now be edited without opening a service
class, without wrestling with PHP string interpolation, and with
proper markdown syntax highlighting in editors.
A tiny PromptLoader helper at app/Support/Ai/PromptLoader.php reads
the file and does {{placeholder}} substitution via strtr() — no
Blade, because Blade's {{ $var }} HTML-escapes ampersands and quotes,
which is wrong for LLM prompts (a company called "Smith & Co" would
be sent as "Smith & Co"). Missing templates throw RuntimeException
so they fail loud during development.
Pure refactor: no prompt wording changes. Existing AI feature tests
(AiChatFlowTest, AiGenerationTest) pass unchanged — they assert on
message structure via ScriptedAiDriver, not on prompt content. Three
new unit tests in PromptLoaderTest lock the helper's contract:
placeholder substitution, no-var loading, missing-file error.
The assistant's output looked perfectly structured — headings, bold,
lists — but had zero visual cues. Sonnet 4.6 defaults to restrained
prose and won't decorate without a nudge. Users reviewing a list of
invoices or an overdue summary benefit from status icons (✅ paid,
⚠️ overdue, 📝 draft) scanning across many records.
Adds an emoji guideline to buildSystemPrompt() with a vocabulary
mapped to the domain model's statuses (paid / partially paid /
overdue / draft / sent / viewed / declined) plus totals, dates,
stats and tips. Capped at one emoji per bullet so responses don't
turn into decoration soup.
Also expands the Markdown rule to explicitly mention headings, bold,
bullet lists and tables — the model was already producing these but
the prompt only vaguely said "Format in Markdown", which left tables
off the table (pun intended) when comparing multiple records.
Sets the default AI chat model to anthropic/claude-sonnet-4.6 and the
default text-generation (WYSIWYG writing) model to anthropic/claude-
haiku-4.5 across all three layers where defaults live: the backend
hydrateDefaults() fallback in AiConfigurationService, the frontend
createDefaults() in AiConfigurationForm, and the docblock example in
AiTextGenerationService.
Refreshes the DriverRegistryProvider suggested-model list to only
include recent models from Anthropic (Sonnet 4.6, Haiku 4.5, Opus
4.6), OpenAI (GPT-5.4, GPT-5.4 mini), Google (Gemini 3.1 Pro preview,
Gemini 3.1 Flash Lite preview) and Z.AI (GLM 5.1, GLM 4.7 Flash).
Drops GPT-4o, Claude 3.5, Gemini 1.5 and Llama 3.3.
The underlying config still accepts any OpenRouter model ID, so the
suggested list is purely a UX surface — existing companies with a
custom ai_chat_model retain their value untouched.
Third and final phase of the AI feature. A SparklesIcon button is added to every Tiptap WYSIWYG editor (invoice notes, email body compose, note templates — ~6 places where RichEditor is used) that opens a modal with a prompt input, optional 'use current content as context' toggle, preview area, and Insert / Replace / Regenerate actions.
**Backend (thin)** — AiTextGenerationService is stateless: resolve config → check text_generation_enabled → instantiate driver → call textCompletion() with a system-prompt-wrapped user instruction. The system prompt is terse and opinionated: 'Return only the requested text. No preamble, no explanation, no markdown code fences.' When context is provided, it's included as a separate framed block ('Context (current content the user is working with):') so the model knows it's operating on existing copy.
**GenerationController** — POST /api/v1/ai/generate with {prompt, context?}. Validates prompt required (max 4000 chars) and context optional (max 20000 chars). Rate-limited via the same 'ai' RateLimiter from Phase 2 (30/min per user/company). Gated by 'use ai' Bouncer ability + AiConfigurationService resolution. Returns {text} on success or {error, message} with 422 on any AiException.
**Frontend modal (AiTextGenerationModal.vue)** — mounted globally in CompanyLayout when bootstrap reports ai.enabled && text_generation_enabled. Uses the existing modalStore pattern: self-registers on componentName='AiTextGenerationModal'. Modal state includes prompt, useContext toggle, generatedText preview. Callers (currently RichEditor) pass onInsert/onReplace callbacks via modalStore.data; the modal invokes them with the final text and closes — it knows nothing about tiptap or ProseMirror.
**RichEditor integration** — the Sparkles toolbar button is pushed onto the existing editorButtons ref at setup time, gated on globalStore.ai.enabled && text_generation_enabled. The button opens the modal with the editor's current getHTML() as context and callbacks that use the tiptap chain API: insertContent for Insert, selectAll().deleteSelection().insertContent for Replace. No reactivity on the flag check — it's set once at bootstrap and doesn't change during a session.
**Tests** (7 new) — AiGenerationTest with a dedicated TextGenDriver test double that tracks the exact prompt passed to textCompletion(). Covers: happy path, context inclusion/omission, AI globally disabled rejection, text_generation role disabled rejection, prompt/context length validation, response whitespace trimming.
395 tests pass (was 388, +7 new). Pint clean. npm run build clean. The AI feature is now complete end-to-end: provider configuration (Phase 1), chat assistant with DB tool-calling (Phase 2), and text generation popup (Phase 3).
Second phase of the AI feature. Users can now open a slide-in chat drawer from the SiteHeader and ask natural-language questions about their company's invoices, customers, payments, and expenses. The LLM invokes pre-defined read-only tool functions (scoped to the current company at execute time) to fetch data and synthesize answers.
**Database** — new ai_conversations and ai_messages tables. Messages are stored in OpenAI's chat format so AiAssistantService serializes a conversation into an API request with zero translation. Columns: role, content, tool_call_id, tool_calls JSON, model, tokens_in, tokens_out. Conversations are scoped (company_id, user_id) — one user's chats are invisible to everyone else, even inside the same company. Foreign-key cascade deletes.
**Tool infrastructure** — AiTool abstract base + AiToolRegistry singleton (registered in a new AiServiceProvider). The base class enforces the security rule: every tool's execute() receives companyId and userId as injected parameters; tools' JSON schemas NEVER include a company_id field. An LLM physically cannot pass a company_id and escape tenancy. Modules can register their own tools by resolving the registry from their own ServiceProvider::boot().
**Nine built-in tools**: search_invoices, get_invoice, search_customers, get_customer, list_recent_payments, list_overdue_invoices, get_company_stats (aggregates for named periods), search_items, list_expense_categories. All read-only; no mutations. Each returns JSON-encodable data the LLM can parse.
**AiAssistantService orchestration loop** — the heart of Phase 2. Flow: persist user message → build payload from system prompt + recent history (40-message window) + new user message → call driver.chatCompletion with tools → if tool_calls, execute each one via the registry (with injected scope), persist tool result, loop → if plain text, persist and return. Hard cap at 5 iterations to prevent runaway LLMs. System prompt pins the assistant to this company's data and forbids mutation.
**Controllers + policy + rate limit** — POST /api/v1/ai/chat runs the orchestration loop. GET/PATCH/DELETE /api/v1/ai/conversations for CRUD. AiConversationPolicy enforces user_id+company_id match on every action. A new 'ai' RateLimiter in RouteServiceProvider throttles to 30 req/min per (user, company). New 'use ai' Gate defined in AppServiceProvider returns true for any authenticated user — the per-company kill-switch still goes through AiConfigurationService::resolveForCompany.
**Frontend** — new features/company/ai/ folder with a Pinia store (ai-chat.store.ts) holding drawer state, current conversation, messages, and loading flags. AiChatDrawer.vue is a slide-in panel teleported to <body>, mounted globally in CompanyLayout.vue when bootstrap reports ai.enabled && ai.chat_enabled. Sub-components: AiChatMessage (user bubbles vs assistant bubbles), AiChatMessageInput (Enter submits, Shift+Enter newline), AiChatConversationList (sidebar with 'new chat' button, rename, delete). A SparklesIcon button in SiteHeader toggles the drawer.
**Driver test double** — tests use a ScriptedAiDriver registered via AiDriverFactory::register('scripted', ...) that returns pre-queued AiChatResponse objects. Feature tests cover: happy path (new conversation + message persistence), tool-call loop (multi-round-trip with search_invoices), runaway-loop cap, driver-throws path, ai_enabled=NO rejection, chat role disabled rejection, per-user conversation visibility, cross-user policy enforcement, cascade delete.
388 tests pass (was 372, +16 new). Pint clean. npm run build clean. Phase 3 (WYSIWYG text generation popup) is the remaining follow-up.
