feat: AI chat tool-use (Claude function calling)

The Asistent AI chat can now query the tenant DB directly through Claude
tool use. AiToolExecutor exposes 5 read-only tools (search_clients,
get_vehicle, find_parts, recent_workorders, low_stock_parts) all scoped
to the current tenant via BelongsToTenant.

AiAssistantService::callClaude loops on stop_reason=tool_use up to 5 rounds:
- normalize message history to content blocks
- send `tools` definitions + messages to Anthropic API
- on tool_use: execute each tool, append tool_results as user turn, recall
- on end_turn: collect text + cumulative token counts + tool-call audit in
  AiMessage.meta.tools

Single-shot helpers (suggestDiagnosis, suggestPrice, vinRecommendations,
suggestParts) are unchanged — only the conversational chat gets tool-use.

Tests (3 new):
- two-round tool_use → execute → final text; verify 5 tools sent both rounds;
  cumulative tokens
- executor finds part by brand
- unknown tool name returns error blob

Full suite: 109 passed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

app/Services/Ai/AiAssistantService.php

@@ -117,36 +117,90 @@ TXT;
     protected function callClaude(string $key, AiChat $chat, string $msg, Company $company): array
     {
         $messages = $this->historyMessages($chat);
-        // Anthropic requires alternating user/assistant; system is separate.
         $messages = array_values(array_filter($messages, fn ($m) => in_array($m['role'], ['user', 'assistant'], true)));
 
-        $r = Http::withHeaders([
-                'x-api-key' => $key,
-                'anthropic-version' => '2023-06-01',
-                'content-type' => 'application/json',
-            ])
-            ->timeout(60)
-            ->post('https://api.anthropic.com/v1/messages', [
-                'model' => 'claude-sonnet-4-5',
-                'max_tokens' => 1024,
-                'system' => $this->buildSystemPrompt($company),
-                'messages' => $messages,
-            ]);
+        // Normalize history to the structured content-block form (Claude tool-use
+        // requires content blocks for the assistant turn that emitted tool_use).
+        foreach ($messages as &$m) {
+            if (is_string($m['content'])) {
+                $m['content'] = [['type' => 'text', 'text' => $m['content']]];
+            }
+        }
+        unset($m);
 
-        if (! $r->successful()) {
-            return ['❌ ' . ($r->json('error.message') ?? 'Anthropic API error ' . $r->status()), ['status' => $r->status()]];
+        $headers = [
+            'x-api-key' => $key,
+            'anthropic-version' => '2023-06-01',
+            'content-type' => 'application/json',
+        ];
+        $system = $this->buildSystemPrompt($company);
+        $tools = AiToolExecutor::TOOLS;
+        $executor = app(AiToolExecutor::class);
+
+        $tokensIn = 0;
+        $tokensOut = 0;
+        $toolCalls = [];
+        $finalText = '';
+        $model = null;
+
+        // Loop on tool_use up to 5 rounds, then bail out with whatever text we have.
+        for ($round = 0; $round < 5; $round++) {
+            $r = Http::withHeaders($headers)->timeout(60)->post(
+                'https://api.anthropic.com/v1/messages',
+                [
+                    'model' => 'claude-sonnet-4-5',
+                    'max_tokens' => 1024,
+                    'system' => $system,
+                    'tools' => $tools,
+                    'messages' => $messages,
+                ]
+            );
+
+            if (! $r->successful()) {
+                return ['❌ ' . ($r->json('error.message') ?? 'Anthropic API error ' . $r->status()), ['status' => $r->status()]];
+            }
+
+            $body = $r->json();
+            $model = $body['model'] ?? $model;
+            $tokensIn += (int) ($body['usage']['input_tokens'] ?? 0);
+            $tokensOut += (int) ($body['usage']['output_tokens'] ?? 0);
+
+            $blocks = $body['content'] ?? [];
+            $finalText = collect($blocks)->where('type', 'text')->pluck('text')->implode("\n");
+
+            if (($body['stop_reason'] ?? null) !== 'tool_use') {
+                break;
+            }
+
+            // Append the assistant turn (with the tool_use block) to history.
+            $messages[] = ['role' => 'assistant', 'content' => $blocks];
+
+            // Execute every tool_use block and build the user reply with tool_results.
+            $toolResults = [];
+            foreach ($blocks as $b) {
+                if (($b['type'] ?? '') !== 'tool_use') continue;
+                $name = $b['name'];
+                $input = (array) ($b['input'] ?? []);
+                try {
+                    $out = $executor->execute($name, $input);
+                } catch (\Throwable $e) {
+                    $out = ['error' => $e->getMessage()];
+                }
+                $toolCalls[] = ['name' => $name, 'input' => $input];
+                $toolResults[] = [
+                    'type' => 'tool_result',
+                    'tool_use_id' => $b['id'],
+                    'content' => json_encode($out, JSON_UNESCAPED_UNICODE),
+                ];
+            }
+            $messages[] = ['role' => 'user', 'content' => $toolResults];
         }
 
-        $body = $r->json();
-        $text = collect($body['content'] ?? [])
-            ->where('type', 'text')
-            ->pluck('text')
-            ->implode("\n");
-
-        return [$text ?: '(răspuns gol)', [
-            'model' => $body['model'] ?? null,
-            'tokens_in' => $body['usage']['input_tokens'] ?? null,
-            'tokens_out' => $body['usage']['output_tokens'] ?? null,
+        return [$finalText ?: '(răspuns gol)', [
+            'model' => $model,
+            'tokens_in' => $tokensIn,
+            'tokens_out' => $tokensOut,
+            'tools' => $toolCalls,
         ]];
     }