feat(memory): implement stability stage

Co-authored-by: multica-agent <github@multica.ai>

apps/web/server/utils/ai-provider.ts

@@ -5,6 +5,14 @@ const CompletionResponseSchema = z.object({
   choices: z.array(z.object({
     message: z.object({ content: z.string().min(1) }),
   })).min(1),
+  usage: z.object({
+    prompt_tokens: z.number().nonnegative().nullish(),
+    completion_tokens: z.number().nonnegative().nullish(),
+    input_tokens: z.number().nonnegative().nullish(),
+    output_tokens: z.number().nonnegative().nullish(),
+    cost: z.number().nonnegative().nullish(),
+    total_cost: z.number().nonnegative().nullish(),
+  }).optional(),
 })
 
 export type AiProvider = z.infer<typeof ProviderSchema>
@@ -26,10 +34,18 @@ export interface JsonCompletionRequest {
   jsonSchema: Record<string, unknown>
 }
 
-interface JsonCompletion {
+export interface JsonCompletion {
   endpoint: string
   headers: Record<string, string>
   body: Record<string, unknown>
+  provider: AiProvider
+  model: string
+}
+
+export interface AiUsageSnapshot {
+  inputTokens: number
+  outputTokens: number
+  costUsd: number
 }
 
 const nonEmptyString = (value: unknown, name: string) => {
@@ -73,6 +89,8 @@ export function buildJsonCompletion(config: AiRuntimeConfig, request: JsonComple
   if (provider.provider === 'deepseek') {
     return {
       endpoint: provider.endpoint,
+      provider: provider.provider,
+      model: provider.model,
       headers: { Authorization: `Bearer ${provider.apiKey}`, 'Content-Type': 'application/json' },
       body: {
         ...common,
@@ -88,6 +106,8 @@ export function buildJsonCompletion(config: AiRuntimeConfig, request: JsonComple
 
   return {
     endpoint: provider.endpoint,
+    provider: provider.provider,
+    model: provider.model,
     headers: { Authorization: `Bearer ${provider.apiKey}`, 'Content-Type': 'application/json', 'X-Title': 'Dungeons & Ground' },
     body: {
       ...common,
@@ -101,3 +121,12 @@ export function parseJsonCompletion(payload: unknown): unknown {
   const completion = CompletionResponseSchema.parse(payload)
   return JSON.parse(completion.choices[0]!.message.content)
 }
+
+export function parseAiUsage(payload: unknown): AiUsageSnapshot {
+  const completion = CompletionResponseSchema.parse(payload)
+  return {
+    inputTokens: Math.max(0, Math.trunc(completion.usage?.prompt_tokens ?? completion.usage?.input_tokens ?? 0)),
+    outputTokens: Math.max(0, Math.trunc(completion.usage?.completion_tokens ?? completion.usage?.output_tokens ?? 0)),
+    costUsd: Math.max(0, completion.usage?.cost ?? completion.usage?.total_cost ?? 0),
+  }
+}