evlog 2.12.0 → 2.14.0

package/README.md

@@ -1,3 +1,7 @@
+<p align="center">
+  <img src="https://raw.githubusercontent.com/HugoRCD/evlog/main/assets/evlog-banner.gif" width="100%" alt="evlog — Digging through logs is not observability. It's hope" />
+</p>
+
 # evlog
 
 [![npm version](https://img.shields.io/npm/v/evlog?color=black)](https://npmjs.com/package/evlog)
@@ -689,6 +693,62 @@ export default defineNitroPlugin((nitroApp) => {
 })
 ```
 
+## Audit Logs
+
+Audit logs are not a parallel system: they are a typed `audit` field on the wide event plus a few helpers. Add 1 enricher + 1 drain wrapper + `log.audit()` and you get tamper-evident, redact-aware, force-kept audit events through the same pipeline.
+
+```typescript
+// server/plugins/evlog.ts
+import { auditEnricher, auditOnly, signed } from 'evlog'
+import { createAxiomDrain } from 'evlog/axiom'
+import { createFsDrain } from 'evlog/fs'
+
+export default defineNitroPlugin((nitroApp) => {
+  const enrich = [auditEnricher({ tenantId: ctx => ctx.headers?.['x-tenant-id'] })]
+  const audits = auditOnly(signed(createFsDrain({ path: '.audit/' }), { strategy: 'hash-chain' }), { await: true })
+  const main = createAxiomDrain()
+
+  nitroApp.hooks.hook('evlog:enrich', async ctx => { for (const e of enrich) await e(ctx) })
+  nitroApp.hooks.hook('evlog:drain', async ctx => { await Promise.all([main(ctx), audits(ctx)]) })
+})
+```
+
+```typescript
+// server/api/invoice/[id]/refund.post.ts
+import { auditDiff } from 'evlog'
+
+export default defineEventHandler(async (event) => {
+  const log = useLogger(event)
+  const before = await db.invoice.get(id)
+  const after = await db.invoice.refund(id)
+
+  log.audit?.({
+    action: 'invoice.refund',
+    actor: { type: 'user', id: user.id, email: user.email },
+    target: { type: 'invoice', id: after.id },
+    outcome: 'success',
+    changes: auditDiff(before, after),
+  })
+})
+```
+
+| Symbol | Kind | Purpose |
+|--------|------|---------|
+| `log.audit(fields)` / `log.audit.deny(reason, fields)` | method | Sugar over `log.set({ audit })` + force-keep |
+| `audit(fields)` | function | Standalone for jobs / scripts |
+| `withAudit({ action, target })(fn)` | wrapper | Auto-emit success / failure / denied |
+| `defineAuditAction(name, opts?)` | factory | Typed action registry |
+| `auditDiff(before, after)` | helper | Redact-aware JSON Patch for `changes` |
+| `mockAudit()` | test util | Capture and assert audits in tests |
+| `auditEnricher({ tenantId? })` | enricher | Auto-fill `req`/`trace`/`ip`/`ua`/`tenantId` context |
+| `auditOnly(drain, { await? })` | wrapper | Routes only events with `event.audit` |
+| `signed(drain, { strategy: 'hmac' \| 'hash-chain', ... })` | wrapper | Tamper-evident integrity |
+| `auditRedactPreset` | preset | Strict PII for audit events |
+
+`AuditFields` is exported and merges with `BaseWideEvent` — augment it with `declare module` if you need extra typed fields. Audit events are always force-kept by tail sampling and get a deterministic `idempotencyKey` so retries are safe across drains.
+
+See [the Audit Logs guide](https://evlog.dev/logging/audit) for compliance, GDPR, and recipe details.
+
 ## AI SDK Integration
 
 Capture token usage, tool calls, model info, and streaming metrics from the [Vercel AI SDK](https://ai-sdk.dev) into wide events. Requires `ai >= 6.0.0`.
@@ -715,6 +775,23 @@ The middleware captures: `inputTokens`, `outputTokens`, `cacheReadTokens`, `reas
 
 For embeddings: `ai.captureEmbed({ usage })`.
 
+The same metadata is also exposed as a public API for custom analytics, billing, or user-facing dashboards:
+
+```typescript
+const ai = createAILogger(log, {
+  cost: { 'claude-sonnet-4.6': { input: 3, output: 15 } },
+})
+
+await generateText({ model: ai.wrap('anthropic/claude-sonnet-4.6'), prompt })
+
+const metadata = ai.getMetadata()       // structured snapshot (AIMetadata)
+const cost = ai.getEstimatedCost()      // dollars, or undefined
+
+ai.onUpdate((metadata) => {             // incremental updates per step
+  pushToClient({ tokens: metadata.totalTokens, cost: metadata.estimatedCost })
+})
+```
+
 ## Adapters
 
 Send your logs to external observability platforms with built-in adapters.
@@ -1062,6 +1139,40 @@ log.emit()                         // Emit final event
 log.getContext()                   // Get current context
 ```
 
+### Wide event lifecycle and `log.fork()`
+
+The framework emits **one wide event per HTTP request** when the response finishes (or on error). After `emit()` runs — including when head sampling drops the event (`emit()` returns `null`) — that logger instance is **sealed**: further `set`, `error`, `info`, and `warn` calls are ignored and emit a **`[evlog]` console warning** listing dropped keys. A second `emit()` is ignored with a warning. This avoids silent data loss when async work (unawaited promises, `setTimeout`, etc.) still resolves `useLogger()` to the same logger via `AsyncLocalStorage` after the response has already been logged.
+
+**`log.fork(label, fn)`** runs work under a **child** request logger: inside `fn`, `useLogger()` returns the child. When `fn` settles, the child emits its **own** wide event with `operation` set to `label` and `_parentRequestId` set to the parent’s `requestId` (query and dashboard correlation). The parent event may be emitted **before** the child event; they are two separate events ordered by time.
+
+`fork` is attached by integrations that use `AsyncLocalStorage` for `useLogger()`. Standalone `createLogger()` instances do not have `fork`.
+
+| Integration | `log.fork()` |
+|-------------|----------------|
+| Express, Fastify, NestJS, SvelteKit, React Router, Elysia | Yes |
+| Next.js `withEvlog` | Yes |
+| Hono (`c.get('log')` only) | Not yet |
+| Nitro / Nuxt `useLogger(event)` | Not yet — use post-emit warnings; see [Wide events](https://evlog.dev/logging/wide-events) |
+
+```typescript
+import { evlog, useLogger } from 'evlog/express'
+
+app.post('/checkout', (req, res) => {
+  const log = req.log
+  log.set({ order_dispatched: true })
+
+  log.fork!('process_order', async () => {
+    const childLog = useLogger()
+    childLog.set({ inventory_checked: true })
+    // child emits automatically when this async function completes
+  })
+
+  res.json({ ok: true })
+})
+```
+
+Use optional chaining if `fork` might be absent: `log.fork?.('task', async () => { ... })`.
+
 ### `initWorkersLogger(options?)`
 
 Initialize evlog for Cloudflare Workers (object logs + correct severity).

package/dist/adapters/axiom.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 //#region src/adapters/axiom.d.ts
 interface BaseAxiomConfig {
   /** Axiom dataset name */

package/dist/adapters/better-stack.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 //#region src/adapters/better-stack.d.ts
 interface BetterStackConfig {
   /** Better Stack source token */

package/dist/adapters/datadog.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 //#region src/adapters/datadog.d.ts
 interface DatadogConfig {
   /** Datadog API key with Logs intake permission */

package/dist/adapters/fs.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 //#region src/adapters/fs.d.ts
 interface FsConfig {
   /** Directory for log files */

package/dist/adapters/hyperdx.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 import { OTLPConfig } from "./otlp.mjs";
 
 //#region src/adapters/hyperdx.d.ts

package/dist/adapters/otlp.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 //#region src/adapters/otlp.d.ts
 interface OTLPConfig {
   /** OTLP HTTP endpoint (e.g., http://localhost:4318) */

package/dist/adapters/posthog.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 //#region src/adapters/posthog.d.ts
 interface PostHogConfig {
   /** PostHog project API key */

package/dist/adapters/sentry.d.mts

@@ -1,4 +1,4 @@
-import { E as WideEvent, r as DrainContext } from "../types-D5OwxZCw.mjs";
+import { F as DrainContext, it as WideEvent } from "../audit-mUutdf6A.mjs";
 //#region src/adapters/sentry.d.ts
 interface SentryConfig {
   /** Sentry DSN */

package/dist/ai/index.d.mts

@@ -1,4 +1,4 @@
-import { _ as RequestLogger } from "../types-D5OwxZCw.mjs";
+import { Y as RequestLogger } from "../audit-mUutdf6A.mjs";
 import { GatewayModelId, TelemetryIntegration } from "ai";
 import { LanguageModelV3, LanguageModelV3Middleware } from "@ai-sdk/provider";
 
@@ -86,13 +86,18 @@ interface AIEmbeddingData {
   count?: number;
 }
 /**
- * Shape of the `ai` field written to the wide event.
+ * Shape of the `ai` field written to the wide event, and the public
+ * snapshot returned by `AILogger.getMetadata()`.
+ *
+ * `model` and `provider` are populated after the first model call.
+ * They may be undefined when only `captureEmbed` has been called or
+ * before any AI activity has happened.
  */
 interface AIEventData {
   calls: number;
-  model: string;
+  model?: string;
   models?: string[];
-  provider: string;
+  provider?: string;
   inputTokens: number;
   outputTokens: number;
   totalTokens: number;
@@ -116,6 +121,18 @@ interface AIEventData {
   embedding?: AIEmbeddingData;
   estimatedCost?: number;
 }
+/**
+ * Public alias for the metadata snapshot returned by `AILogger.getMetadata()`.
+ *
+ * Mirrors the shape of the `ai` field on the wide event, so the same object
+ * can be persisted, surfaced to end-users, or compared across runs.
+ */
+type AIMetadata = AIEventData;
+/**
+ * Callback fired on every metadata update (per step, per embed, on error,
+ * and on integration completion). Receives a structured snapshot.
+ */
+type AIMetadataListener = (metadata: AIMetadata) => void;
 interface AILogger {
   /**
    * Wrap a language model with evlog middleware.
@@ -163,6 +180,89 @@ interface AILogger {
     dimensions?: number;
     count?: number;
   }) => void;
+  /**
+   * Get a snapshot of the current AI execution metadata.
+   *
+   * Returns the same structured object that is written to the `ai` field
+   * of the wide event. Safe to call at any time — including inside the
+   * AI SDK's `onFinish` callback, after `await generateText()`, or while a
+   * stream is in progress.
+   *
+   * The returned snapshot is a fresh copy: mutating it does not affect
+   * subsequent calls or the underlying state.
+   *
+   * @example Persist execution history after a run
+   * ```ts
+   * const ai = createAILogger(log, { cost: { ... } })
+   *
+   * await generateText({ model: ai.wrap('anthropic/claude-sonnet-4.6'), prompt })
+   *
+   * const metadata = ai.getMetadata()
+   * await db.insert('ai_runs', metadata)
+   * ```
+   *
+   * @example Surface usage to end-users in a streaming response
+   * ```ts
+   * const result = streamText({
+   *   model: ai.wrap('anthropic/claude-sonnet-4.6'),
+   *   messages,
+   *   onFinish: () => {
+   *     const { totalTokens, estimatedCost } = ai.getMetadata()
+   *     trackUsage(userId, { totalTokens, estimatedCost })
+   *   },
+   * })
+   * ```
+   */
+  getMetadata: () => AIMetadata;
+  /**
+   * Get the current estimated cost in dollars.
+   *
+   * Returns `undefined` if no `cost` map was provided to `createAILogger`,
+   * or if the model is not in the pricing map.
+   *
+   * Convenience for `getMetadata().estimatedCost`.
+   *
+   * @example
+   * ```ts
+   * const ai = createAILogger(log, {
+   *   cost: { 'claude-sonnet-4.6': { input: 3, output: 15 } },
+   * })
+   *
+   * await generateText({ model: ai.wrap('anthropic/claude-sonnet-4.6'), prompt })
+   *
+   * console.log(`Cost: $${ai.getEstimatedCost()?.toFixed(4)}`)
+   * ```
+   */
+  getEstimatedCost: () => number | undefined;
+  /**
+   * Subscribe to metadata updates.
+   *
+   * The callback fires every time the underlying state flushes — once per
+   * step (in multi-step agent runs), once per `captureEmbed` call, on
+   * model errors, and once on `createEvlogIntegration`'s `onFinish`.
+   *
+   * Each invocation receives a fresh snapshot (same shape as `getMetadata`).
+   * Returns an unsubscribe function.
+   *
+   * @example Stream incremental usage updates to the client
+   * ```ts
+   * const ai = createAILogger(log)
+   *
+   * ai.onUpdate((metadata) => {
+   *   pushToClient({ type: 'ai-progress', metadata })
+   * })
+   *
+   * const result = streamText({ model: ai.wrap('...'), messages })
+   * ```
+   *
+   * @example Cleanup
+   * ```ts
+   * const off = ai.onUpdate((metadata) => { ... })
+   * // later
+   * off()
+   * ```
+   */
+  onUpdate: (callback: AIMetadataListener) => () => void;
   /**
    * Internal accumulator state exposed for `createEvlogIntegration` to share.
    * @internal
@@ -252,6 +352,7 @@ interface AccumulatorState {
   totalDurationMs: number | undefined;
   embedding: AIEmbeddingData | undefined;
   costMap: Record<string, ModelCost> | undefined;
+  subscribers: Set<AIMetadataListener>;
   /** @internal Logger reference for integration flush */
   _log?: RequestLogger;
 }
@@ -300,5 +401,5 @@ interface AccumulatorState {
  */
 declare function createEvlogIntegration(logOrAi: RequestLogger | AILogger, options?: AILoggerOptions): TelemetryIntegration;
 //#endregion
-export { AIEmbeddingData, AIEventData, AILogger, AILoggerOptions, AIStepUsage, AIToolExecution, ModelCost, ToolInputsOptions, createAILogger, createAIMiddleware, createEvlogIntegration };
+export { AIEmbeddingData, AIEventData, AILogger, AILoggerOptions, AIMetadata, AIMetadataListener, AIStepUsage, AIToolExecution, ModelCost, ToolInputsOptions, createAILogger, createAIMiddleware, createEvlogIntegration };
 //# sourceMappingURL=index.d.mts.map

package/dist/ai/index.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.mts","names":[],"sources":["../../src/ai/index.ts"],"mappings":";;;;;;;AAQA;UAAiB,iBAAA;;;;;EAKf,SAAA;EAM6B;;;AAM/B;;EANE,SAAA,IAAa,KAAA,WAAgB,QAAA;AAAA;;AAc/B;;UARiB,SAAA;EACf,KAAA;EACA,MAAA;AAAA;;;;UAMe,eAAA;EA2Bf;;;;;AAMF;;;EAxBE,UAAA,aAAuB,iBAAA;EAyBvB;;;;;;AASF;;;;;;;;;;AAUA;EA1BE,IAAA,GAAO,MAAA,SAAe,SAAA;AAAA;;;;UAMP,WAAA;EACf,KAAA;EACA,WAAA;EACA,YAAA;EACA,SAAA;AAAA;;;;UAMe,eAAA;EACf,IAAA;EACA,UAAA;EACA,OAAA;EACA,KAAA;AAAA;;;;UAMe,eAAA;EACf,KAAA;EACA,MAAA;EACA,UAAA;EACA,KAAA;AAAA;;;;UAMe,WAAA;EACf,KAAA;EACA,KAAA;EACA,MAAA;EACA,QAAA;EACA,WAAA;EACA,YAAA;EACA,WAAA;EACA,eAAA;EACA,gBAAA;EACA,eAAA;EACA,YAAA;EACA,SAAA,cAAuB,KAAA;IAAQ,IAAA;IAAc,KAAA;EAAA;EAC7C,UAAA;EACA,KAAA;EACA,UAAA,GAAa,WAAA;EACb,cAAA;EACA,UAAA;EACA,eAAA;EACA,KAAA;EACA,KAAA,GAAQ,eAAA;EACR,eAAA;EACA,SAAA,GAAY,eAAA;EACZ,aAAA;AAAA;AAAA,UAGe,QAAA;EAqBR;;;;;;;;;;;;;AA+BR;;;;;;;EA/BC,IAAA,GAAO,KAAA,EAAO,eAAA,GAAkB,cAAA,KAAmB,eAAA;EAsCnD;;;AAyDF;;;;;;;;;;;;;;EA5EE,YAAA,GAAe,MAAA;IACb,KAAA;MAAS,MAAA;IAAA;IACT,KAAA;IACA,UAAA;IACA,KAAA;EAAA;EAsGmF;;;;EA/FrF,MAAA,EAAQ,gBAAA;AAAA;AAAA,UAGA,gBAAA;EACR,WAAA;EACA,YAAA;EACA,eAAA;EACA,gBAAA;EACA,eAAA;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;iBAyDc,kBAAA,CAAmB,GAAA,EAAK,aAAA,EAAe,OAAA,GAAU,eAAA,GAAkB,yBAAA;;;;;;;;;;;;;;;;;;;AAocnF;;;;;;;;iBAtagB,cAAA,CAAe,GAAA,EAAK,aAAA,EAAe,OAAA,GAAU,eAAA,GAAkB,QAAA;AAAA,UAgCrE,gBAAA;EACR,KAAA;EACA,KAAA;EACA,KAAA,EAAO,gBAAA;EACP,MAAA;EACA,YAAA;EACA,YAAA;EACA,iBAAA,EAAmB,KAAA;IAAQ,IAAA;IAAc,KAAA;EAAA;EACzC,UAAA,EAAY,WAAA;EACZ,gBAAA;EACA,kBAAA;EACA,cAAA;EACA,SAAA;EACA,cAAA;EACA,UAAA;EACA,iBAAA,EAAmB,iBAAA;EACnB,cAAA,EAAgB,eAAA;EAChB,mBAAA;EACA,eAAA;EACA,SAAA,EAAW,eAAA;EACX,OAAA,EAAS,MAAA,SAAe,SAAA;;EAExB,IAAA,GAAO,aAAA;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgXO,sBAAA,CACd,OAAA,EAAS,aAAA,GAAgB,QAAA,EACzB,OAAA,GAAU,eAAA,GACT,oBAAA"}
+{"version":3,"file":"index.d.mts","names":[],"sources":["../../src/ai/index.ts"],"mappings":";;;;;;;AAQA;UAAiB,iBAAA;;;;;EAKf,SAAA;EAM6B;;;AAM/B;;EANE,SAAA,IAAa,KAAA,WAAgB,QAAA;AAAA;;AAc/B;;UARiB,SAAA;EACf,KAAA;EACA,MAAA;AAAA;;;;UAMe,eAAA;EA2Bf;;;;;AAMF;;;EAxBE,UAAA,aAAuB,iBAAA;EAyBvB;;;;;;AASF;;;;;;;;;;AAUA;EA1BE,IAAA,GAAO,MAAA,SAAe,SAAA;AAAA;;;;UAMP,WAAA;EACf,KAAA;EACA,WAAA;EACA,YAAA;EACA,SAAA;AAAA;;;;UAMe,eAAA;EACf,IAAA;EACA,UAAA;EACA,OAAA;EACA,KAAA;AAAA;;;;UAMe,eAAA;EACf,KAAA;EACA,MAAA;EACA,UAAA;EACA,KAAA;AAAA;;;;;;;;;UAWe,WAAA;EACf,KAAA;EACA,KAAA;EACA,MAAA;EACA,QAAA;EACA,WAAA;EACA,YAAA;EACA,WAAA;EACA,eAAA;EACA,gBAAA;EACA,eAAA;EACA,YAAA;EACA,SAAA,cAAuB,KAAA;IAAQ,IAAA;IAAc,KAAA;EAAA;EAC7C,UAAA;EACA,KAAA;EACA,UAAA,GAAa,WAAA;EACb,cAAA;EACA,UAAA;EACA,eAAA;EACA,KAAA;EACA,KAAA,GAAQ,eAAA;EACR,eAAA;EACA,SAAA,GAAY,eAAA;EACZ,aAAA;AAAA;;;;;;;KASU,UAAA,GAAa,WAAA;;;;;KAMb,kBAAA,IAAsB,QAAA,EAAU,UAAA;AAAA,UAE3B,QAAA;EA0Cb;;;;;;;;;;;;;;AAgGH;;;;;;EArHC,IAAA,GAAO,KAAA,EAAO,eAAA,GAAkB,cAAA,KAAmB,eAAA;EA2HnD;;;;AA0DF;;;;;;;;;;;;;EAlKE,YAAA,GAAe,MAAA;IACb,KAAA;MAAS,MAAA;IAAA;IACT,KAAA;IACA,UAAA;IACA,KAAA;EAAA;EA4LmF;;;;;;;;;AAyCtF;;;;;;;;;;;;;;;;;;;;;;;;EAjMC,WAAA,QAAmB,UAAA;EA0MQ;;;;;;;;;;;;;;;;;;;EArL3B,gBAAA;EAkMwB;;;;;;;;AAqY1B;;;;;;;;;;;;;;;;;;;;EAziBE,QAAA,GAAW,QAAA,EAAU,kBAAA;;;;;EAMrB,MAAA,EAAQ,gBAAA;AAAA;AAAA,UAGA,gBAAA;EACR,WAAA;EACA,YAAA;EACA,eAAA;EACA,gBAAA;EACA,eAAA;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;iBAyDc,kBAAA,CAAmB,GAAA,EAAK,aAAA,EAAe,OAAA,GAAU,eAAA,GAAkB,yBAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8BnE,cAAA,CAAe,GAAA,EAAK,aAAA,EAAe,OAAA,GAAU,eAAA,GAAkB,QAAA;AAAA,UA2CrE,gBAAA;EACR,KAAA;EACA,KAAA;EACA,KAAA,EAAO,gBAAA;EACP,MAAA;EACA,YAAA;EACA,YAAA;EACA,iBAAA,EAAmB,KAAA;IAAQ,IAAA;IAAc,KAAA;EAAA;EACzC,UAAA,EAAY,WAAA;EACZ,gBAAA;EACA,kBAAA;EACA,cAAA;EACA,SAAA;EACA,cAAA;EACA,UAAA;EACA,iBAAA,EAAmB,iBAAA;EACnB,cAAA,EAAgB,eAAA;EAChB,mBAAA;EACA,eAAA;EACA,SAAA,EAAW,eAAA;EACX,OAAA,EAAS,MAAA,SAAe,SAAA;EACxB,WAAA,EAAa,GAAA,CAAI,kBAAA;;EAEjB,IAAA,GAAO,aAAA;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAkYO,sBAAA,CACd,OAAA,EAAS,aAAA,GAAgB,QAAA,EACzB,OAAA,GAAU,eAAA,GACT,oBAAA"}

package/dist/ai/index.mjs

@@ -98,6 +98,14 @@ function createAILogger(log, options) {
 			};
 			flushState(log, state);
 		},
+		getMetadata: () => buildMetadata(state),
+		getEstimatedCost: () => computeEstimatedCost(state),
+		onUpdate: (callback) => {
+			state.subscribers.add(callback);
+			return () => {
+				state.subscribers.delete(callback);
+			};
+		},
 		_state: state
 	};
 }
@@ -152,7 +160,8 @@ function createAccumulatorState(options) {
 		generationStartTime: void 0,
 		totalDurationMs: void 0,
 		embedding: void 0,
-		costMap: options?.cost
+		costMap: options?.cost,
+		subscribers: /* @__PURE__ */ new Set()
 	};
 }
 function computeEstimatedCost(state) {
@@ -164,7 +173,7 @@ function computeEstimatedCost(state) {
 	const total = state.usage.inputTokens / 1e6 * pricing.input + state.usage.outputTokens / 1e6 * pricing.output;
 	return total > 0 ? Math.round(total * 1e6) / 1e6 : void 0;
 }
-function flushState(log, state) {
+function buildMetadata(state) {
 	const uniqueModels = [...new Set(state.models)];
 	const lastModel = state.models[state.models.length - 1];
 	const data = {
@@ -180,12 +189,15 @@ function flushState(log, state) {
 	if (state.usage.cacheWriteTokens > 0) data.cacheWriteTokens = state.usage.cacheWriteTokens;
 	if (state.usage.reasoningTokens > 0) data.reasoningTokens = state.usage.reasoningTokens;
 	if (state.lastFinishReason) data.finishReason = state.lastFinishReason;
-	if (state.toolInputs && state.allToolCallInputs.length > 0) data.toolCalls = [...state.allToolCallInputs];
+	if (state.toolInputs && state.allToolCallInputs.length > 0) data.toolCalls = state.allToolCallInputs.map((t) => ({ ...t }));
 	else if (state.allToolCalls.length > 0) data.toolCalls = [...state.allToolCalls];
 	if (state.lastResponseId) data.responseId = state.lastResponseId;
 	if (state.steps > 1) {
 		data.steps = state.steps;
-		data.stepsUsage = [...state.stepsUsage];
+		data.stepsUsage = state.stepsUsage.map((s) => ({
+			...s,
+			...s.toolCalls ? { toolCalls: [...s.toolCalls] } : {}
+		}));
 	}
 	if (state.lastMsToFirstChunk !== void 0) data.msToFirstChunk = state.lastMsToFirstChunk;
 	if (state.lastMsToFinish !== void 0) {
@@ -193,12 +205,23 @@ function flushState(log, state) {
 		if (state.usage.outputTokens > 0 && state.lastMsToFinish > 0) data.tokensPerSecond = Math.round(state.usage.outputTokens / state.lastMsToFinish * 1e3);
 	}
 	if (state.lastError) data.error = state.lastError;
-	if (state.toolExecutions.length > 0) data.tools = [...state.toolExecutions];
+	if (state.toolExecutions.length > 0) data.tools = state.toolExecutions.map((t) => ({ ...t }));
 	if (state.totalDurationMs !== void 0) data.totalDurationMs = state.totalDurationMs;
 	if (state.embedding) data.embedding = { ...state.embedding };
 	const cost = computeEstimatedCost(state);
 	if (cost !== void 0) data.estimatedCost = cost;
+	return data;
+}
+function notifySubscribers(state, metadata) {
+	if (state.subscribers.size === 0) return;
+	for (const subscriber of state.subscribers) try {
+		subscriber(metadata);
+	} catch {}
+}
+function flushState(log, state) {
+	const data = buildMetadata(state);
 	log.set({ ai: data });
+	notifySubscribers(state, data);
 }
 function recordModel(state, provider, modelId, responseModelId) {
 	const resolved = resolveProviderAndModel(provider, responseModelId ?? modelId);

package/dist/ai/index.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","names":[],"sources":["../../src/ai/index.ts"],"sourcesContent":["import { gateway, wrapLanguageModel, bindTelemetryIntegration } from 'ai'\nimport type { GatewayModelId, TelemetryIntegration, OnStartEvent, OnToolCallFinishEvent, OnFinishEvent } from 'ai'\nimport type { LanguageModelV3, LanguageModelV3Middleware, LanguageModelV3StreamPart } from '@ai-sdk/provider'\nimport type { RequestLogger } from '../types'\n\n/**\n * Fine-grained control over tool call input capture.\n */\nexport interface ToolInputsOptions {\n  /**\n   * Max character length for the stringified input JSON.\n   * Inputs exceeding this limit are truncated with a `…` suffix.\n   */\n  maxLength?: number\n  /**\n   * Custom transform applied to each captured input before storing.\n   * Receives the parsed input and tool name; return value is stored.\n   * Runs before `maxLength` truncation.\n   */\n  transform?: (input: unknown, toolName: string) => unknown\n}\n\n/**\n * Pricing entry for a single model: cost per 1 million tokens in dollars.\n */\nexport interface ModelCost {\n  input: number\n  output: number\n}\n\n/**\n * Options for `createAILogger` and `createAIMiddleware`.\n */\nexport interface AILoggerOptions {\n  /**\n   * When enabled, `toolCalls` contains `{ name, input }` objects instead of plain tool name strings.\n   * Opt-in because inputs can be large and may contain sensitive data.\n   *\n   * - `true` — capture all inputs as-is\n   * - `{ maxLength, transform }` — capture with truncation or custom transform\n   * @default false\n   */\n  toolInputs?: boolean | ToolInputsOptions\n  /**\n   * Pricing map for estimating request cost.\n   * Keys are model IDs (e.g. `'claude-sonnet-4.6'`, `'gpt-4o'`), values are\n   * `{ input, output }` in dollars per 1M tokens.\n   *\n   * When provided, the wide event includes `ai.estimatedCost` (in dollars).\n   *\n   * @example\n   * ```ts\n   * const ai = createAILogger(log, {\n   *   cost: {\n   *     'claude-sonnet-4.6': { input: 3, output: 15 },\n   *     'gpt-4o': { input: 2.5, output: 10 },\n   *   },\n   * })\n   * ```\n   */\n  cost?: Record<string, ModelCost>\n}\n\n/**\n * Per-step token usage breakdown for multi-step agent runs.\n */\nexport interface AIStepUsage {\n  model: string\n  inputTokens: number\n  outputTokens: number\n  toolCalls?: string[]\n}\n\n/**\n * Tool execution detail captured via `TelemetryIntegration`.\n */\nexport interface AIToolExecution {\n  name: string\n  durationMs: number\n  success: boolean\n  error?: string\n}\n\n/**\n * Embedding metadata captured via `captureEmbed`.\n */\nexport interface AIEmbeddingData {\n  model?: string\n  tokens: number\n  dimensions?: number\n  count?: number\n}\n\n/**\n * Shape of the `ai` field written to the wide event.\n */\nexport interface AIEventData {\n  calls: number\n  model: string\n  models?: string[]\n  provider: string\n  inputTokens: number\n  outputTokens: number\n  totalTokens: number\n  cacheReadTokens?: number\n  cacheWriteTokens?: number\n  reasoningTokens?: number\n  finishReason?: string\n  toolCalls?: string[] | Array<{ name: string, input: unknown }>\n  responseId?: string\n  steps?: number\n  stepsUsage?: AIStepUsage[]\n  msToFirstChunk?: number\n  msToFinish?: number\n  tokensPerSecond?: number\n  error?: string\n  tools?: AIToolExecution[]\n  totalDurationMs?: number\n  embedding?: AIEmbeddingData\n  estimatedCost?: number\n}\n\nexport interface AILogger {\n  /**\n   * Wrap a language model with evlog middleware.\n   * All `generateText` and `streamText` calls using the wrapped model\n   * are captured automatically into the wide event.\n   *\n   * Accepts a `LanguageModelV3` object or a model string (e.g. `'anthropic/claude-sonnet-4.6'`).\n   * Strings are resolved via the AI SDK gateway.\n   *\n   * Also works with pre-wrapped models (e.g. from supermemory, guardrails):\n   * `ai.wrap(withSupermemory(base, orgId))` composes correctly.\n   *\n   * @example\n   * ```ts\n   * const ai = createAILogger(log)\n   * const model = ai.wrap('anthropic/claude-sonnet-4.6')\n   *\n   * // Also accepts a model object\n   * const model = ai.wrap(anthropic('claude-sonnet-4.6'))\n   * ```\n   */\n  wrap: (model: LanguageModelV3 | GatewayModelId) => LanguageModelV3\n\n  /**\n   * Manually capture token usage from an `embed()` or `embedMany()` result.\n   * Embedding models use a different type than language models, so they\n   * cannot be wrapped with middleware.\n   *\n   * @example\n   * ```ts\n   * const { embedding, usage } = await embed({ model: embeddingModel, value: query })\n   * ai.captureEmbed({ usage })\n   *\n   * // With model info (v2)\n   * ai.captureEmbed({ usage, model: 'text-embedding-3-small', dimensions: 1536 })\n   *\n   * // After embedMany\n   * ai.captureEmbed({ usage, count: texts.length })\n   * ```\n   */\n  captureEmbed: (result: {\n    usage: { tokens: number }\n    model?: string\n    dimensions?: number\n    count?: number\n  }) => void\n\n  /**\n   * Internal accumulator state exposed for `createEvlogIntegration` to share.\n   * @internal\n   */\n  _state: AccumulatorState\n}\n\ninterface UsageAccumulator {\n  inputTokens: number\n  outputTokens: number\n  cacheReadTokens: number\n  cacheWriteTokens: number\n  reasoningTokens: number\n}\n\nfunction addUsage(\n  acc: UsageAccumulator,\n  usage: {\n    inputTokens: { total: number | undefined, cacheRead?: number | undefined, cacheWrite?: number | undefined }\n    outputTokens: { total: number | undefined, reasoning?: number | undefined }\n  },\n): void {\n  acc.inputTokens += usage.inputTokens.total ?? 0\n  acc.outputTokens += usage.outputTokens.total ?? 0\n  acc.cacheReadTokens += usage.inputTokens.cacheRead ?? 0\n  acc.cacheWriteTokens += usage.inputTokens.cacheWrite ?? 0\n  acc.reasoningTokens += usage.outputTokens.reasoning ?? 0\n}\n\n/**\n * When using `gateway('google/gemini-3-flash')`, the model object has\n * `provider: 'gateway'` and `modelId: 'google/gemini-3-flash'`.\n * This extracts the real provider and model name from the modelId.\n */\nfunction resolveProviderAndModel(provider: string, modelId: string): { provider: string, model: string } {\n  if (provider !== 'gateway' || !modelId.includes('/')) {\n    return { provider, model: modelId }\n  }\n  const slashIndex = modelId.indexOf('/')\n  return {\n    provider: modelId.slice(0, slashIndex),\n    model: modelId.slice(slashIndex + 1),\n  }\n}\n\n/**\n * Create the evlog AI middleware that captures AI SDK data into a wide event.\n *\n * Use this when you need explicit middleware composition with other wrappers\n * (e.g. supermemory, guardrails). For most cases, use `createAILogger` instead.\n *\n * Note: `captureEmbed` is not available with the raw middleware — use\n * `createAILogger` if you need embedding capture.\n *\n * @example Nuxt API route with supermemory\n * ```ts\n * import { createAIMiddleware } from 'evlog/ai'\n * import { wrapLanguageModel } from 'ai'\n *\n * export default defineEventHandler(async (event) => {\n *   const log = useLogger(event)\n *\n *   const model = wrapLanguageModel({\n *     model: withSupermemory(base, orgId),\n *     middleware: [createAIMiddleware(log, { toolInputs: true })],\n *   })\n * })\n * ```\n */\nexport function createAIMiddleware(log: RequestLogger, options?: AILoggerOptions): LanguageModelV3Middleware {\n  return buildMiddleware(log, options)\n}\n\n/**\n * Create an AI logger that captures AI SDK data into the wide event.\n *\n * Uses model middleware (`wrapLanguageModel`) to transparently intercept\n * all LLM calls. `onFinish` and `onStepFinish` remain free for user code.\n *\n * @example\n * ```ts\n * import { createAILogger } from 'evlog/ai'\n *\n * const log = useLogger(event)\n * const ai = createAILogger(log)\n * const model = ai.wrap('anthropic/claude-sonnet-4.6')\n *\n * const result = streamText({\n *   model,\n *   messages,\n *   onFinish: ({ text }) => saveConversation(text),\n * })\n * ```\n *\n * @example Capture tool call inputs\n * ```ts\n * const ai = createAILogger(log, { toolInputs: true })\n * ```\n */\nexport function createAILogger(log: RequestLogger, options?: AILoggerOptions): AILogger {\n  const state = createAccumulatorState(options)\n  state._log = log\n  const middleware = buildMiddlewareFromState(log, state)\n\n  return {\n    wrap: (model: LanguageModelV3 | GatewayModelId) => {\n      const resolved = typeof model === 'string' ? gateway(model) : model\n      return wrapLanguageModel({ model: resolved, middleware })\n    },\n\n    captureEmbed: (result: {\n      usage: { tokens: number }\n      model?: string\n      dimensions?: number\n      count?: number\n    }) => {\n      state.calls++\n      state.usage.inputTokens += result.usage.tokens\n      state.embedding = {\n        tokens: (state.embedding?.tokens ?? 0) + result.usage.tokens,\n        ...(result.model ? { model: result.model } : state.embedding?.model ? { model: state.embedding.model } : {}),\n        ...(result.dimensions ? { dimensions: result.dimensions } : state.embedding?.dimensions ? { dimensions: state.embedding.dimensions } : {}),\n        ...(result.count ? { count: (state.embedding?.count ?? 0) + result.count } : state.embedding?.count ? { count: state.embedding.count } : {}),\n      }\n      flushState(log, state)\n    },\n\n    _state: state,\n  }\n}\n\ninterface AccumulatorState {\n  calls: number\n  steps: number\n  usage: UsageAccumulator\n  models: string[]\n  lastProvider: string | undefined\n  allToolCalls: string[]\n  allToolCallInputs: Array<{ name: string, input: unknown }>\n  stepsUsage: AIStepUsage[]\n  lastFinishReason: string | undefined\n  lastMsToFirstChunk: number | undefined\n  lastMsToFinish: number | undefined\n  lastError: string | undefined\n  lastResponseId: string | undefined\n  toolInputs: boolean\n  toolInputsOptions: ToolInputsOptions | undefined\n  toolExecutions: AIToolExecution[]\n  generationStartTime: number | undefined\n  totalDurationMs: number | undefined\n  embedding: AIEmbeddingData | undefined\n  costMap: Record<string, ModelCost> | undefined\n  /** @internal Logger reference for integration flush */\n  _log?: RequestLogger\n}\n\nfunction resolveToolInputs(raw?: boolean | ToolInputsOptions): { enabled: boolean, options: ToolInputsOptions | undefined } {\n  if (!raw) return { enabled: false, options: undefined }\n  if (raw === true) return { enabled: true, options: undefined }\n  return { enabled: true, options: raw }\n}\n\nfunction processToolInput(input: unknown, toolName: string, options: ToolInputsOptions | undefined): unknown {\n  let value = input\n  if (options?.transform) {\n    value = options.transform(value, toolName)\n  }\n  if (options?.maxLength) {\n    const str = typeof value === 'string' ? value : JSON.stringify(value)\n    if (str.length > options.maxLength) {\n      return `${str.slice(0, options.maxLength)}…`\n    }\n  }\n  return value\n}\n\nfunction createAccumulatorState(options?: AILoggerOptions): AccumulatorState {\n  const { enabled, options: captureOpts } = resolveToolInputs(options?.toolInputs)\n  return {\n    calls: 0,\n    steps: 0,\n    usage: {\n      inputTokens: 0,\n      outputTokens: 0,\n      cacheReadTokens: 0,\n      cacheWriteTokens: 0,\n      reasoningTokens: 0,\n    },\n    models: [],\n    lastProvider: undefined,\n    allToolCalls: [],\n    allToolCallInputs: [],\n    stepsUsage: [],\n    lastFinishReason: undefined,\n    lastMsToFirstChunk: undefined,\n    lastMsToFinish: undefined,\n    lastError: undefined,\n    lastResponseId: undefined,\n    toolInputs: enabled,\n    toolInputsOptions: captureOpts,\n    toolExecutions: [],\n    generationStartTime: undefined,\n    totalDurationMs: undefined,\n    embedding: undefined,\n    costMap: options?.cost,\n  }\n}\n\nfunction computeEstimatedCost(state: AccumulatorState): number | undefined {\n  if (!state.costMap) return undefined\n  const lastModel = state.models[state.models.length - 1]\n  if (!lastModel) return undefined\n  const pricing = state.costMap[lastModel]\n  if (!pricing) return undefined\n  const inputCost = (state.usage.inputTokens / 1_000_000) * pricing.input\n  const outputCost = (state.usage.outputTokens / 1_000_000) * pricing.output\n  const total = inputCost + outputCost\n  return total > 0 ? Math.round(total * 1_000_000) / 1_000_000 : undefined\n}\n\nfunction flushState(log: RequestLogger, state: AccumulatorState): void {\n  const uniqueModels = [...new Set(state.models)]\n  const lastModel = state.models[state.models.length - 1]\n\n  const data: Partial<AIEventData> & { calls: number, inputTokens: number, outputTokens: number, totalTokens: number } = {\n    calls: state.calls,\n    inputTokens: state.usage.inputTokens,\n    outputTokens: state.usage.outputTokens,\n    totalTokens: state.usage.inputTokens + state.usage.outputTokens,\n  }\n\n  if (lastModel) data.model = lastModel\n  if (state.lastProvider) data.provider = state.lastProvider\n  if (uniqueModels.length > 1) data.models = uniqueModels\n  if (state.usage.cacheReadTokens > 0) data.cacheReadTokens = state.usage.cacheReadTokens\n  if (state.usage.cacheWriteTokens > 0) data.cacheWriteTokens = state.usage.cacheWriteTokens\n  if (state.usage.reasoningTokens > 0) data.reasoningTokens = state.usage.reasoningTokens\n  if (state.lastFinishReason) data.finishReason = state.lastFinishReason\n  if (state.toolInputs && state.allToolCallInputs.length > 0) {\n    data.toolCalls = [...state.allToolCallInputs]\n  } else if (state.allToolCalls.length > 0) {\n    data.toolCalls = [...state.allToolCalls]\n  }\n  if (state.lastResponseId) data.responseId = state.lastResponseId\n  if (state.steps > 1) {\n    data.steps = state.steps\n    data.stepsUsage = [...state.stepsUsage]\n  }\n  if (state.lastMsToFirstChunk !== undefined) data.msToFirstChunk = state.lastMsToFirstChunk\n  if (state.lastMsToFinish !== undefined) {\n    data.msToFinish = state.lastMsToFinish\n    if (state.usage.outputTokens > 0 && state.lastMsToFinish > 0) {\n      data.tokensPerSecond = Math.round((state.usage.outputTokens / state.lastMsToFinish) * 1000)\n    }\n  }\n  if (state.lastError) data.error = state.lastError\n  if (state.toolExecutions.length > 0) data.tools = [...state.toolExecutions]\n  if (state.totalDurationMs !== undefined) data.totalDurationMs = state.totalDurationMs\n  if (state.embedding) data.embedding = { ...state.embedding }\n  const cost = computeEstimatedCost(state)\n  if (cost !== undefined) data.estimatedCost = cost\n\n  log.set({ ai: data } as Record<string, unknown>)\n}\n\nfunction recordModel(state: AccumulatorState, provider: string, modelId: string, responseModelId?: string): void {\n  const resolved = resolveProviderAndModel(provider, responseModelId ?? modelId)\n  state.models.push(resolved.model)\n  state.lastProvider = resolved.provider\n}\n\nfunction safeParseJSON(input: string): unknown {\n  try {\n    return JSON.parse(input)\n  } catch {\n    return input\n  }\n}\n\nfunction recordError(log: RequestLogger, state: AccumulatorState, model: { provider: string, modelId: string }, error: unknown): void {\n  state.calls++\n  state.steps++\n  recordModel(state, model.provider, model.modelId)\n  state.lastFinishReason = 'error'\n  state.lastError = error instanceof Error ? error.message : String(error)\n\n  const resolved = resolveProviderAndModel(model.provider, model.modelId)\n  state.stepsUsage.push({\n    model: resolved.model,\n    inputTokens: 0,\n    outputTokens: 0,\n  })\n\n  flushState(log, state)\n}\n\nfunction buildMiddleware(log: RequestLogger, options?: AILoggerOptions): LanguageModelV3Middleware {\n  const state = createAccumulatorState(options)\n  state._log = log\n  return buildMiddlewareFromState(log, state)\n}\n\nfunction buildMiddlewareFromState(log: RequestLogger, state: AccumulatorState): LanguageModelV3Middleware {\n  return {\n    specificationVersion: 'v3',\n    wrapGenerate: async ({ doGenerate, model }) => {\n      try {\n        const result = await doGenerate()\n\n        state.calls++\n        state.steps++\n        addUsage(state.usage, result.usage)\n        recordModel(state, model.provider, model.modelId, result.response?.modelId)\n        state.lastFinishReason = result.finishReason.unified\n\n        if (result.response?.id) {\n          state.lastResponseId = result.response.id\n        }\n\n        const stepToolCalls: string[] = []\n        for (const item of result.content) {\n          if (item.type === 'tool-call') {\n            state.allToolCalls.push(item.toolName)\n            stepToolCalls.push(item.toolName)\n            if (state.toolInputs) {\n              const raw = typeof item.input === 'string' ? safeParseJSON(item.input) : item.input\n              state.allToolCallInputs.push({\n                name: item.toolName,\n                input: processToolInput(raw, item.toolName, state.toolInputsOptions),\n              })\n            }\n          }\n        }\n\n        const resolvedModel = resolveProviderAndModel(model.provider, result.response?.modelId ?? model.modelId)\n        state.stepsUsage.push({\n          model: resolvedModel.model,\n          inputTokens: result.usage.inputTokens.total ?? 0,\n          outputTokens: result.usage.outputTokens.total ?? 0,\n          ...(stepToolCalls.length > 0 ? { toolCalls: stepToolCalls } : {}),\n        })\n\n        flushState(log, state)\n        return result\n      } catch (error) {\n        recordError(log, state, model, error)\n        throw error\n      }\n    },\n\n    wrapStream: async ({ doStream, model }) => {\n      const streamStart = Date.now()\n      let firstChunkTime: number | undefined\n\n      let streamUsage: UsageAccumulator | undefined\n      let streamFinishReason: string | undefined\n      let streamModelId: string | undefined\n      let streamResponseId: string | undefined\n      const streamToolCalls: string[] = []\n      const streamToolInputBuffers = new Map<string, { name: string, chunks: string[] }>()\n      let streamError: string | undefined\n\n      let doStreamResult: Awaited<ReturnType<typeof doStream>>\n      try {\n        doStreamResult = await doStream()\n      } catch (error) {\n        recordError(log, state, model, error)\n        throw error\n      }\n\n      const { stream, ...rest } = doStreamResult\n\n      const transformStream = new TransformStream<\n        LanguageModelV3StreamPart,\n        LanguageModelV3StreamPart\n      >({\n        transform(chunk, controller) {\n          if (!firstChunkTime && chunk.type === 'text-delta') {\n            firstChunkTime = Date.now()\n          }\n\n          if (chunk.type === 'tool-input-start') {\n            streamToolCalls.push(chunk.toolName)\n            if (state.toolInputs) {\n              streamToolInputBuffers.set(chunk.id, { name: chunk.toolName, chunks: [] })\n            }\n          }\n\n          if (chunk.type === 'tool-input-delta' && state.toolInputs) {\n            const buffer = streamToolInputBuffers.get(chunk.id)\n            if (buffer) {\n              buffer.chunks.push(chunk.delta)\n            }\n          }\n\n          if (chunk.type === 'tool-input-end' && state.toolInputs) {\n            const buffer = streamToolInputBuffers.get(chunk.id)\n            if (buffer) {\n              const raw = safeParseJSON(buffer.chunks.join(''))\n              state.allToolCallInputs.push({\n                name: buffer.name,\n                input: processToolInput(raw, buffer.name, state.toolInputsOptions),\n              })\n              streamToolInputBuffers.delete(chunk.id)\n            }\n          }\n\n          if (chunk.type === 'finish') {\n            streamUsage = {\n              inputTokens: chunk.usage.inputTokens.total ?? 0,\n              outputTokens: chunk.usage.outputTokens.total ?? 0,\n              cacheReadTokens: chunk.usage.inputTokens.cacheRead ?? 0,\n              cacheWriteTokens: chunk.usage.inputTokens.cacheWrite ?? 0,\n              reasoningTokens: chunk.usage.outputTokens.reasoning ?? 0,\n            }\n            streamFinishReason = chunk.finishReason.unified\n          }\n\n          if (chunk.type === 'response-metadata') {\n            if (chunk.modelId) streamModelId = chunk.modelId\n            if (chunk.id) streamResponseId = chunk.id\n          }\n\n          if (chunk.type === 'error') {\n            streamError = chunk.error instanceof Error ? chunk.error.message : String(chunk.error)\n          }\n\n          controller.enqueue(chunk)\n        },\n\n        flush() {\n          state.calls++\n          state.steps++\n\n          if (streamUsage) {\n            state.usage.inputTokens += streamUsage.inputTokens\n            state.usage.outputTokens += streamUsage.outputTokens\n            state.usage.cacheReadTokens += streamUsage.cacheReadTokens\n            state.usage.cacheWriteTokens += streamUsage.cacheWriteTokens\n            state.usage.reasoningTokens += streamUsage.reasoningTokens\n          }\n\n          recordModel(state, model.provider, model.modelId, streamModelId)\n          state.lastFinishReason = streamFinishReason\n\n          state.allToolCalls.push(...streamToolCalls)\n\n          if (streamResponseId) {\n            state.lastResponseId = streamResponseId\n          }\n\n          if (firstChunkTime) {\n            state.lastMsToFirstChunk = firstChunkTime - streamStart\n          }\n          state.lastMsToFinish = Date.now() - streamStart\n\n          if (streamError) state.lastError = streamError\n\n          const resolvedModel = resolveProviderAndModel(model.provider, streamModelId ?? model.modelId)\n          state.stepsUsage.push({\n            model: resolvedModel.model,\n            inputTokens: streamUsage?.inputTokens ?? 0,\n            outputTokens: streamUsage?.outputTokens ?? 0,\n            ...(streamToolCalls.length > 0 ? { toolCalls: [...streamToolCalls] } : {}),\n          })\n\n          flushState(log, state)\n        },\n      })\n\n      return {\n        stream: stream.pipeThrough(transformStream),\n        ...rest,\n      }\n    },\n  }\n}\n\n/**\n * Create an AI SDK `TelemetryIntegration` that captures tool execution\n * timing, errors, and total generation wall time into the wide event.\n *\n * Complements the middleware-based `createAILogger`: the middleware captures\n * token usage and streaming metrics at the model level, while the integration\n * captures application-level lifecycle events (tool execution, total duration).\n *\n * When passed an `AILogger`, shares its accumulator so both paths write to\n * the same `ai.*` field. Can also be used standalone with a `RequestLogger`.\n *\n * @example Combined with middleware (recommended)\n * ```ts\n * import { createAILogger, createEvlogIntegration } from 'evlog/ai'\n *\n * const log = useLogger(event)\n * const ai = createAILogger(log)\n *\n * const result = await generateText({\n *   model: ai.wrap('anthropic/claude-sonnet-4.6'),\n *   tools: { getWeather },\n *   experimental_telemetry: {\n *     isEnabled: true,\n *     integrations: [createEvlogIntegration(ai)],\n *   },\n * })\n * ```\n *\n * @example Standalone (no middleware wrapping)\n * ```ts\n * import { createEvlogIntegration } from 'evlog/ai'\n *\n * const integration = createEvlogIntegration(log)\n *\n * const result = await generateText({\n *   model: openai('gpt-4o'),\n *   experimental_telemetry: {\n *     isEnabled: true,\n *     integrations: [integration],\n *   },\n * })\n * ```\n */\nexport function createEvlogIntegration(\n  logOrAi: RequestLogger | AILogger,\n  options?: AILoggerOptions,\n): TelemetryIntegration {\n  let log: RequestLogger\n  let state: AccumulatorState\n\n  if ('_state' in logOrAi && logOrAi._state) {\n    state = logOrAi._state\n    log = state._log!\n  } else {\n    log = logOrAi as RequestLogger\n    state = createAccumulatorState(options)\n    state._log = log\n  }\n\n  class EvlogIntegration implements TelemetryIntegration {\n\n    onStart(_event: OnStartEvent) {\n      state.generationStartTime = Date.now()\n    }\n\n    onToolCallFinish(event: OnToolCallFinishEvent) {\n      const execution: AIToolExecution = {\n        name: (event.toolCall as { toolName: string }).toolName,\n        durationMs: event.durationMs,\n        success: event.success,\n      }\n      if (!event.success && event.error) {\n        execution.error = event.error instanceof Error ? event.error.message : String(event.error)\n      }\n      state.toolExecutions.push(execution)\n    }\n\n    onFinish(_event: OnFinishEvent) {\n      if (state.generationStartTime) {\n        state.totalDurationMs = Date.now() - state.generationStartTime\n      }\n      flushState(log, state)\n    }\n  \n  }\n\n  return bindTelemetryIntegration(new EvlogIntegration())\n}\n"],"mappings":";;AAwLA,SAAS,SACP,KACA,OAIM;AACN,KAAI,eAAe,MAAM,YAAY,SAAS;AAC9C,KAAI,gBAAgB,MAAM,aAAa,SAAS;AAChD,KAAI,mBAAmB,MAAM,YAAY,aAAa;AACtD,KAAI,oBAAoB,MAAM,YAAY,cAAc;AACxD,KAAI,mBAAmB,MAAM,aAAa,aAAa;;;;;;;AAQzD,SAAS,wBAAwB,UAAkB,SAAsD;AACvG,KAAI,aAAa,aAAa,CAAC,QAAQ,SAAS,IAAI,CAClD,QAAO;EAAE;EAAU,OAAO;EAAS;CAErC,MAAM,aAAa,QAAQ,QAAQ,IAAI;AACvC,QAAO;EACL,UAAU,QAAQ,MAAM,GAAG,WAAW;EACtC,OAAO,QAAQ,MAAM,aAAa,EAAE;EACrC;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BH,SAAgB,mBAAmB,KAAoB,SAAsD;AAC3G,QAAO,gBAAgB,KAAK,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BtC,SAAgB,eAAe,KAAoB,SAAqC;CACtF,MAAM,QAAQ,uBAAuB,QAAQ;AAC7C,OAAM,OAAO;CACb,MAAM,aAAa,yBAAyB,KAAK,MAAM;AAEvD,QAAO;EACL,OAAO,UAA4C;AAEjD,UAAO,kBAAkB;IAAE,OADV,OAAO,UAAU,WAAW,QAAQ,MAAM,GAAG;IAClB;IAAY,CAAC;;EAG3D,eAAe,WAKT;AACJ,SAAM;AACN,SAAM,MAAM,eAAe,OAAO,MAAM;AACxC,SAAM,YAAY;IAChB,SAAS,MAAM,WAAW,UAAU,KAAK,OAAO,MAAM;IACtD,GAAI,OAAO,QAAQ,EAAE,OAAO,OAAO,OAAO,GAAG,MAAM,WAAW,QAAQ,EAAE,OAAO,MAAM,UAAU,OAAO,GAAG,EAAE;IAC3G,GAAI,OAAO,aAAa,EAAE,YAAY,OAAO,YAAY,GAAG,MAAM,WAAW,aAAa,EAAE,YAAY,MAAM,UAAU,YAAY,GAAG,EAAE;IACzI,GAAI,OAAO,QAAQ,EAAE,QAAQ,MAAM,WAAW,SAAS,KAAK,OAAO,OAAO,GAAG,MAAM,WAAW,QAAQ,EAAE,OAAO,MAAM,UAAU,OAAO,GAAG,EAAE;IAC5I;AACD,cAAW,KAAK,MAAM;;EAGxB,QAAQ;EACT;;AA4BH,SAAS,kBAAkB,KAAiG;AAC1H,KAAI,CAAC,IAAK,QAAO;EAAE,SAAS;EAAO,SAAS,KAAA;EAAW;AACvD,KAAI,QAAQ,KAAM,QAAO;EAAE,SAAS;EAAM,SAAS,KAAA;EAAW;AAC9D,QAAO;EAAE,SAAS;EAAM,SAAS;EAAK;;AAGxC,SAAS,iBAAiB,OAAgB,UAAkB,SAAiD;CAC3G,IAAI,QAAQ;AACZ,KAAI,SAAS,UACX,SAAQ,QAAQ,UAAU,OAAO,SAAS;AAE5C,KAAI,SAAS,WAAW;EACtB,MAAM,MAAM,OAAO,UAAU,WAAW,QAAQ,KAAK,UAAU,MAAM;AACrE,MAAI,IAAI,SAAS,QAAQ,UACvB,QAAO,GAAG,IAAI,MAAM,GAAG,QAAQ,UAAU,CAAC;;AAG9C,QAAO;;AAGT,SAAS,uBAAuB,SAA6C;CAC3E,MAAM,EAAE,SAAS,SAAS,gBAAgB,kBAAkB,SAAS,WAAW;AAChF,QAAO;EACL,OAAO;EACP,OAAO;EACP,OAAO;GACL,aAAa;GACb,cAAc;GACd,iBAAiB;GACjB,kBAAkB;GAClB,iBAAiB;GAClB;EACD,QAAQ,EAAE;EACV,cAAc,KAAA;EACd,cAAc,EAAE;EAChB,mBAAmB,EAAE;EACrB,YAAY,EAAE;EACd,kBAAkB,KAAA;EAClB,oBAAoB,KAAA;EACpB,gBAAgB,KAAA;EAChB,WAAW,KAAA;EACX,gBAAgB,KAAA;EAChB,YAAY;EACZ,mBAAmB;EACnB,gBAAgB,EAAE;EAClB,qBAAqB,KAAA;EACrB,iBAAiB,KAAA;EACjB,WAAW,KAAA;EACX,SAAS,SAAS;EACnB;;AAGH,SAAS,qBAAqB,OAA6C;AACzE,KAAI,CAAC,MAAM,QAAS,QAAO,KAAA;CAC3B,MAAM,YAAY,MAAM,OAAO,MAAM,OAAO,SAAS;AACrD,KAAI,CAAC,UAAW,QAAO,KAAA;CACvB,MAAM,UAAU,MAAM,QAAQ;AAC9B,KAAI,CAAC,QAAS,QAAO,KAAA;CAGrB,MAAM,QAFa,MAAM,MAAM,cAAc,MAAa,QAAQ,QAC9C,MAAM,MAAM,eAAe,MAAa,QAAQ;AAEpE,QAAO,QAAQ,IAAI,KAAK,MAAM,QAAQ,IAAU,GAAG,MAAY,KAAA;;AAGjE,SAAS,WAAW,KAAoB,OAA+B;CACrE,MAAM,eAAe,CAAC,GAAG,IAAI,IAAI,MAAM,OAAO,CAAC;CAC/C,MAAM,YAAY,MAAM,OAAO,MAAM,OAAO,SAAS;CAErD,MAAM,OAAiH;EACrH,OAAO,MAAM;EACb,aAAa,MAAM,MAAM;EACzB,cAAc,MAAM,MAAM;EAC1B,aAAa,MAAM,MAAM,cAAc,MAAM,MAAM;EACpD;AAED,KAAI,UAAW,MAAK,QAAQ;AAC5B,KAAI,MAAM,aAAc,MAAK,WAAW,MAAM;AAC9C,KAAI,aAAa,SAAS,EAAG,MAAK,SAAS;AAC3C,KAAI,MAAM,MAAM,kBAAkB,EAAG,MAAK,kBAAkB,MAAM,MAAM;AACxE,KAAI,MAAM,MAAM,mBAAmB,EAAG,MAAK,mBAAmB,MAAM,MAAM;AAC1E,KAAI,MAAM,MAAM,kBAAkB,EAAG,MAAK,kBAAkB,MAAM,MAAM;AACxE,KAAI,MAAM,iBAAkB,MAAK,eAAe,MAAM;AACtD,KAAI,MAAM,cAAc,MAAM,kBAAkB,SAAS,EACvD,MAAK,YAAY,CAAC,GAAG,MAAM,kBAAkB;UACpC,MAAM,aAAa,SAAS,EACrC,MAAK,YAAY,CAAC,GAAG,MAAM,aAAa;AAE1C,KAAI,MAAM,eAAgB,MAAK,aAAa,MAAM;AAClD,KAAI,MAAM,QAAQ,GAAG;AACnB,OAAK,QAAQ,MAAM;AACnB,OAAK,aAAa,CAAC,GAAG,MAAM,WAAW;;AAEzC,KAAI,MAAM,uBAAuB,KAAA,EAAW,MAAK,iBAAiB,MAAM;AACxE,KAAI,MAAM,mBAAmB,KAAA,GAAW;AACtC,OAAK,aAAa,MAAM;AACxB,MAAI,MAAM,MAAM,eAAe,KAAK,MAAM,iBAAiB,EACzD,MAAK,kBAAkB,KAAK,MAAO,MAAM,MAAM,eAAe,MAAM,iBAAkB,IAAK;;AAG/F,KAAI,MAAM,UAAW,MAAK,QAAQ,MAAM;AACxC,KAAI,MAAM,eAAe,SAAS,EAAG,MAAK,QAAQ,CAAC,GAAG,MAAM,eAAe;AAC3E,KAAI,MAAM,oBAAoB,KAAA,EAAW,MAAK,kBAAkB,MAAM;AACtE,KAAI,MAAM,UAAW,MAAK,YAAY,EAAE,GAAG,MAAM,WAAW;CAC5D,MAAM,OAAO,qBAAqB,MAAM;AACxC,KAAI,SAAS,KAAA,EAAW,MAAK,gBAAgB;AAE7C,KAAI,IAAI,EAAE,IAAI,MAAM,CAA4B;;AAGlD,SAAS,YAAY,OAAyB,UAAkB,SAAiB,iBAAgC;CAC/G,MAAM,WAAW,wBAAwB,UAAU,mBAAmB,QAAQ;AAC9E,OAAM,OAAO,KAAK,SAAS,MAAM;AACjC,OAAM,eAAe,SAAS;;AAGhC,SAAS,cAAc,OAAwB;AAC7C,KAAI;AACF,SAAO,KAAK,MAAM,MAAM;SAClB;AACN,SAAO;;;AAIX,SAAS,YAAY,KAAoB,OAAyB,OAA8C,OAAsB;AACpI,OAAM;AACN,OAAM;AACN,aAAY,OAAO,MAAM,UAAU,MAAM,QAAQ;AACjD,OAAM,mBAAmB;AACzB,OAAM,YAAY,iBAAiB,QAAQ,MAAM,UAAU,OAAO,MAAM;CAExE,MAAM,WAAW,wBAAwB,MAAM,UAAU,MAAM,QAAQ;AACvE,OAAM,WAAW,KAAK;EACpB,OAAO,SAAS;EAChB,aAAa;EACb,cAAc;EACf,CAAC;AAEF,YAAW,KAAK,MAAM;;AAGxB,SAAS,gBAAgB,KAAoB,SAAsD;CACjG,MAAM,QAAQ,uBAAuB,QAAQ;AAC7C,OAAM,OAAO;AACb,QAAO,yBAAyB,KAAK,MAAM;;AAG7C,SAAS,yBAAyB,KAAoB,OAAoD;AACxG,QAAO;EACL,sBAAsB;EACtB,cAAc,OAAO,EAAE,YAAY,YAAY;AAC7C,OAAI;IACF,MAAM,SAAS,MAAM,YAAY;AAEjC,UAAM;AACN,UAAM;AACN,aAAS,MAAM,OAAO,OAAO,MAAM;AACnC,gBAAY,OAAO,MAAM,UAAU,MAAM,SAAS,OAAO,UAAU,QAAQ;AAC3E,UAAM,mBAAmB,OAAO,aAAa;AAE7C,QAAI,OAAO,UAAU,GACnB,OAAM,iBAAiB,OAAO,SAAS;IAGzC,MAAM,gBAA0B,EAAE;AAClC,SAAK,MAAM,QAAQ,OAAO,QACxB,KAAI,KAAK,SAAS,aAAa;AAC7B,WAAM,aAAa,KAAK,KAAK,SAAS;AACtC,mBAAc,KAAK,KAAK,SAAS;AACjC,SAAI,MAAM,YAAY;MACpB,MAAM,MAAM,OAAO,KAAK,UAAU,WAAW,cAAc,KAAK,MAAM,GAAG,KAAK;AAC9E,YAAM,kBAAkB,KAAK;OAC3B,MAAM,KAAK;OACX,OAAO,iBAAiB,KAAK,KAAK,UAAU,MAAM,kBAAkB;OACrE,CAAC;;;IAKR,MAAM,gBAAgB,wBAAwB,MAAM,UAAU,OAAO,UAAU,WAAW,MAAM,QAAQ;AACxG,UAAM,WAAW,KAAK;KACpB,OAAO,cAAc;KACrB,aAAa,OAAO,MAAM,YAAY,SAAS;KAC/C,cAAc,OAAO,MAAM,aAAa,SAAS;KACjD,GAAI,cAAc,SAAS,IAAI,EAAE,WAAW,eAAe,GAAG,EAAE;KACjE,CAAC;AAEF,eAAW,KAAK,MAAM;AACtB,WAAO;YACA,OAAO;AACd,gBAAY,KAAK,OAAO,OAAO,MAAM;AACrC,UAAM;;;EAIV,YAAY,OAAO,EAAE,UAAU,YAAY;GACzC,MAAM,cAAc,KAAK,KAAK;GAC9B,IAAI;GAEJ,IAAI;GACJ,IAAI;GACJ,IAAI;GACJ,IAAI;GACJ,MAAM,kBAA4B,EAAE;GACpC,MAAM,yCAAyB,IAAI,KAAiD;GACpF,IAAI;GAEJ,IAAI;AACJ,OAAI;AACF,qBAAiB,MAAM,UAAU;YAC1B,OAAO;AACd,gBAAY,KAAK,OAAO,OAAO,MAAM;AACrC,UAAM;;GAGR,MAAM,EAAE,QAAQ,GAAG,SAAS;GAE5B,MAAM,kBAAkB,IAAI,gBAG1B;IACA,UAAU,OAAO,YAAY;AAC3B,SAAI,CAAC,kBAAkB,MAAM,SAAS,aACpC,kBAAiB,KAAK,KAAK;AAG7B,SAAI,MAAM,SAAS,oBAAoB;AACrC,sBAAgB,KAAK,MAAM,SAAS;AACpC,UAAI,MAAM,WACR,wBAAuB,IAAI,MAAM,IAAI;OAAE,MAAM,MAAM;OAAU,QAAQ,EAAE;OAAE,CAAC;;AAI9E,SAAI,MAAM,SAAS,sBAAsB,MAAM,YAAY;MACzD,MAAM,SAAS,uBAAuB,IAAI,MAAM,GAAG;AACnD,UAAI,OACF,QAAO,OAAO,KAAK,MAAM,MAAM;;AAInC,SAAI,MAAM,SAAS,oBAAoB,MAAM,YAAY;MACvD,MAAM,SAAS,uBAAuB,IAAI,MAAM,GAAG;AACnD,UAAI,QAAQ;OACV,MAAM,MAAM,cAAc,OAAO,OAAO,KAAK,GAAG,CAAC;AACjD,aAAM,kBAAkB,KAAK;QAC3B,MAAM,OAAO;QACb,OAAO,iBAAiB,KAAK,OAAO,MAAM,MAAM,kBAAkB;QACnE,CAAC;AACF,8BAAuB,OAAO,MAAM,GAAG;;;AAI3C,SAAI,MAAM,SAAS,UAAU;AAC3B,oBAAc;OACZ,aAAa,MAAM,MAAM,YAAY,SAAS;OAC9C,cAAc,MAAM,MAAM,aAAa,SAAS;OAChD,iBAAiB,MAAM,MAAM,YAAY,aAAa;OACtD,kBAAkB,MAAM,MAAM,YAAY,cAAc;OACxD,iBAAiB,MAAM,MAAM,aAAa,aAAa;OACxD;AACD,2BAAqB,MAAM,aAAa;;AAG1C,SAAI,MAAM,SAAS,qBAAqB;AACtC,UAAI,MAAM,QAAS,iBAAgB,MAAM;AACzC,UAAI,MAAM,GAAI,oBAAmB,MAAM;;AAGzC,SAAI,MAAM,SAAS,QACjB,eAAc,MAAM,iBAAiB,QAAQ,MAAM,MAAM,UAAU,OAAO,MAAM,MAAM;AAGxF,gBAAW,QAAQ,MAAM;;IAG3B,QAAQ;AACN,WAAM;AACN,WAAM;AAEN,SAAI,aAAa;AACf,YAAM,MAAM,eAAe,YAAY;AACvC,YAAM,MAAM,gBAAgB,YAAY;AACxC,YAAM,MAAM,mBAAmB,YAAY;AAC3C,YAAM,MAAM,oBAAoB,YAAY;AAC5C,YAAM,MAAM,mBAAmB,YAAY;;AAG7C,iBAAY,OAAO,MAAM,UAAU,MAAM,SAAS,cAAc;AAChE,WAAM,mBAAmB;AAEzB,WAAM,aAAa,KAAK,GAAG,gBAAgB;AAE3C,SAAI,iBACF,OAAM,iBAAiB;AAGzB,SAAI,eACF,OAAM,qBAAqB,iBAAiB;AAE9C,WAAM,iBAAiB,KAAK,KAAK,GAAG;AAEpC,SAAI,YAAa,OAAM,YAAY;KAEnC,MAAM,gBAAgB,wBAAwB,MAAM,UAAU,iBAAiB,MAAM,QAAQ;AAC7F,WAAM,WAAW,KAAK;MACpB,OAAO,cAAc;MACrB,aAAa,aAAa,eAAe;MACzC,cAAc,aAAa,gBAAgB;MAC3C,GAAI,gBAAgB,SAAS,IAAI,EAAE,WAAW,CAAC,GAAG,gBAAgB,EAAE,GAAG,EAAE;MAC1E,CAAC;AAEF,gBAAW,KAAK,MAAM;;IAEzB,CAAC;AAEF,UAAO;IACL,QAAQ,OAAO,YAAY,gBAAgB;IAC3C,GAAG;IACJ;;EAEJ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8CH,SAAgB,uBACd,SACA,SACsB;CACtB,IAAI;CACJ,IAAI;AAEJ,KAAI,YAAY,WAAW,QAAQ,QAAQ;AACzC,UAAQ,QAAQ;AAChB,QAAM,MAAM;QACP;AACL,QAAM;AACN,UAAQ,uBAAuB,QAAQ;AACvC,QAAM,OAAO;;CAGf,MAAM,iBAAiD;EAErD,QAAQ,QAAsB;AAC5B,SAAM,sBAAsB,KAAK,KAAK;;EAGxC,iBAAiB,OAA8B;GAC7C,MAAM,YAA6B;IACjC,MAAO,MAAM,SAAkC;IAC/C,YAAY,MAAM;IAClB,SAAS,MAAM;IAChB;AACD,OAAI,CAAC,MAAM,WAAW,MAAM,MAC1B,WAAU,QAAQ,MAAM,iBAAiB,QAAQ,MAAM,MAAM,UAAU,OAAO,MAAM,MAAM;AAE5F,SAAM,eAAe,KAAK,UAAU;;EAGtC,SAAS,QAAuB;AAC9B,OAAI,MAAM,oBACR,OAAM,kBAAkB,KAAK,KAAK,GAAG,MAAM;AAE7C,cAAW,KAAK,MAAM;;;AAK1B,QAAO,yBAAyB,IAAI,kBAAkB,CAAC"}
+{"version":3,"file":"index.mjs","names":[],"sources":["../../src/ai/index.ts"],"sourcesContent":["import { gateway, wrapLanguageModel, bindTelemetryIntegration } from 'ai'\nimport type { GatewayModelId, TelemetryIntegration, OnStartEvent, OnToolCallFinishEvent, OnFinishEvent } from 'ai'\nimport type { LanguageModelV3, LanguageModelV3Middleware, LanguageModelV3StreamPart } from '@ai-sdk/provider'\nimport type { RequestLogger } from '../types'\n\n/**\n * Fine-grained control over tool call input capture.\n */\nexport interface ToolInputsOptions {\n  /**\n   * Max character length for the stringified input JSON.\n   * Inputs exceeding this limit are truncated with a `…` suffix.\n   */\n  maxLength?: number\n  /**\n   * Custom transform applied to each captured input before storing.\n   * Receives the parsed input and tool name; return value is stored.\n   * Runs before `maxLength` truncation.\n   */\n  transform?: (input: unknown, toolName: string) => unknown\n}\n\n/**\n * Pricing entry for a single model: cost per 1 million tokens in dollars.\n */\nexport interface ModelCost {\n  input: number\n  output: number\n}\n\n/**\n * Options for `createAILogger` and `createAIMiddleware`.\n */\nexport interface AILoggerOptions {\n  /**\n   * When enabled, `toolCalls` contains `{ name, input }` objects instead of plain tool name strings.\n   * Opt-in because inputs can be large and may contain sensitive data.\n   *\n   * - `true` — capture all inputs as-is\n   * - `{ maxLength, transform }` — capture with truncation or custom transform\n   * @default false\n   */\n  toolInputs?: boolean | ToolInputsOptions\n  /**\n   * Pricing map for estimating request cost.\n   * Keys are model IDs (e.g. `'claude-sonnet-4.6'`, `'gpt-4o'`), values are\n   * `{ input, output }` in dollars per 1M tokens.\n   *\n   * When provided, the wide event includes `ai.estimatedCost` (in dollars).\n   *\n   * @example\n   * ```ts\n   * const ai = createAILogger(log, {\n   *   cost: {\n   *     'claude-sonnet-4.6': { input: 3, output: 15 },\n   *     'gpt-4o': { input: 2.5, output: 10 },\n   *   },\n   * })\n   * ```\n   */\n  cost?: Record<string, ModelCost>\n}\n\n/**\n * Per-step token usage breakdown for multi-step agent runs.\n */\nexport interface AIStepUsage {\n  model: string\n  inputTokens: number\n  outputTokens: number\n  toolCalls?: string[]\n}\n\n/**\n * Tool execution detail captured via `TelemetryIntegration`.\n */\nexport interface AIToolExecution {\n  name: string\n  durationMs: number\n  success: boolean\n  error?: string\n}\n\n/**\n * Embedding metadata captured via `captureEmbed`.\n */\nexport interface AIEmbeddingData {\n  model?: string\n  tokens: number\n  dimensions?: number\n  count?: number\n}\n\n/**\n * Shape of the `ai` field written to the wide event, and the public\n * snapshot returned by `AILogger.getMetadata()`.\n *\n * `model` and `provider` are populated after the first model call.\n * They may be undefined when only `captureEmbed` has been called or\n * before any AI activity has happened.\n */\nexport interface AIEventData {\n  calls: number\n  model?: string\n  models?: string[]\n  provider?: string\n  inputTokens: number\n  outputTokens: number\n  totalTokens: number\n  cacheReadTokens?: number\n  cacheWriteTokens?: number\n  reasoningTokens?: number\n  finishReason?: string\n  toolCalls?: string[] | Array<{ name: string, input: unknown }>\n  responseId?: string\n  steps?: number\n  stepsUsage?: AIStepUsage[]\n  msToFirstChunk?: number\n  msToFinish?: number\n  tokensPerSecond?: number\n  error?: string\n  tools?: AIToolExecution[]\n  totalDurationMs?: number\n  embedding?: AIEmbeddingData\n  estimatedCost?: number\n}\n\n/**\n * Public alias for the metadata snapshot returned by `AILogger.getMetadata()`.\n *\n * Mirrors the shape of the `ai` field on the wide event, so the same object\n * can be persisted, surfaced to end-users, or compared across runs.\n */\nexport type AIMetadata = AIEventData\n\n/**\n * Callback fired on every metadata update (per step, per embed, on error,\n * and on integration completion). Receives a structured snapshot.\n */\nexport type AIMetadataListener = (metadata: AIMetadata) => void\n\nexport interface AILogger {\n  /**\n   * Wrap a language model with evlog middleware.\n   * All `generateText` and `streamText` calls using the wrapped model\n   * are captured automatically into the wide event.\n   *\n   * Accepts a `LanguageModelV3` object or a model string (e.g. `'anthropic/claude-sonnet-4.6'`).\n   * Strings are resolved via the AI SDK gateway.\n   *\n   * Also works with pre-wrapped models (e.g. from supermemory, guardrails):\n   * `ai.wrap(withSupermemory(base, orgId))` composes correctly.\n   *\n   * @example\n   * ```ts\n   * const ai = createAILogger(log)\n   * const model = ai.wrap('anthropic/claude-sonnet-4.6')\n   *\n   * // Also accepts a model object\n   * const model = ai.wrap(anthropic('claude-sonnet-4.6'))\n   * ```\n   */\n  wrap: (model: LanguageModelV3 | GatewayModelId) => LanguageModelV3\n\n  /**\n   * Manually capture token usage from an `embed()` or `embedMany()` result.\n   * Embedding models use a different type than language models, so they\n   * cannot be wrapped with middleware.\n   *\n   * @example\n   * ```ts\n   * const { embedding, usage } = await embed({ model: embeddingModel, value: query })\n   * ai.captureEmbed({ usage })\n   *\n   * // With model info (v2)\n   * ai.captureEmbed({ usage, model: 'text-embedding-3-small', dimensions: 1536 })\n   *\n   * // After embedMany\n   * ai.captureEmbed({ usage, count: texts.length })\n   * ```\n   */\n  captureEmbed: (result: {\n    usage: { tokens: number }\n    model?: string\n    dimensions?: number\n    count?: number\n  }) => void\n\n  /**\n   * Get a snapshot of the current AI execution metadata.\n   *\n   * Returns the same structured object that is written to the `ai` field\n   * of the wide event. Safe to call at any time — including inside the\n   * AI SDK's `onFinish` callback, after `await generateText()`, or while a\n   * stream is in progress.\n   *\n   * The returned snapshot is a fresh copy: mutating it does not affect\n   * subsequent calls or the underlying state.\n   *\n   * @example Persist execution history after a run\n   * ```ts\n   * const ai = createAILogger(log, { cost: { ... } })\n   *\n   * await generateText({ model: ai.wrap('anthropic/claude-sonnet-4.6'), prompt })\n   *\n   * const metadata = ai.getMetadata()\n   * await db.insert('ai_runs', metadata)\n   * ```\n   *\n   * @example Surface usage to end-users in a streaming response\n   * ```ts\n   * const result = streamText({\n   *   model: ai.wrap('anthropic/claude-sonnet-4.6'),\n   *   messages,\n   *   onFinish: () => {\n   *     const { totalTokens, estimatedCost } = ai.getMetadata()\n   *     trackUsage(userId, { totalTokens, estimatedCost })\n   *   },\n   * })\n   * ```\n   */\n  getMetadata: () => AIMetadata\n\n  /**\n   * Get the current estimated cost in dollars.\n   *\n   * Returns `undefined` if no `cost` map was provided to `createAILogger`,\n   * or if the model is not in the pricing map.\n   *\n   * Convenience for `getMetadata().estimatedCost`.\n   *\n   * @example\n   * ```ts\n   * const ai = createAILogger(log, {\n   *   cost: { 'claude-sonnet-4.6': { input: 3, output: 15 } },\n   * })\n   *\n   * await generateText({ model: ai.wrap('anthropic/claude-sonnet-4.6'), prompt })\n   *\n   * console.log(`Cost: $${ai.getEstimatedCost()?.toFixed(4)}`)\n   * ```\n   */\n  getEstimatedCost: () => number | undefined\n\n  /**\n   * Subscribe to metadata updates.\n   *\n   * The callback fires every time the underlying state flushes — once per\n   * step (in multi-step agent runs), once per `captureEmbed` call, on\n   * model errors, and once on `createEvlogIntegration`'s `onFinish`.\n   *\n   * Each invocation receives a fresh snapshot (same shape as `getMetadata`).\n   * Returns an unsubscribe function.\n   *\n   * @example Stream incremental usage updates to the client\n   * ```ts\n   * const ai = createAILogger(log)\n   *\n   * ai.onUpdate((metadata) => {\n   *   pushToClient({ type: 'ai-progress', metadata })\n   * })\n   *\n   * const result = streamText({ model: ai.wrap('...'), messages })\n   * ```\n   *\n   * @example Cleanup\n   * ```ts\n   * const off = ai.onUpdate((metadata) => { ... })\n   * // later\n   * off()\n   * ```\n   */\n  onUpdate: (callback: AIMetadataListener) => () => void\n\n  /**\n   * Internal accumulator state exposed for `createEvlogIntegration` to share.\n   * @internal\n   */\n  _state: AccumulatorState\n}\n\ninterface UsageAccumulator {\n  inputTokens: number\n  outputTokens: number\n  cacheReadTokens: number\n  cacheWriteTokens: number\n  reasoningTokens: number\n}\n\nfunction addUsage(\n  acc: UsageAccumulator,\n  usage: {\n    inputTokens: { total: number | undefined, cacheRead?: number | undefined, cacheWrite?: number | undefined }\n    outputTokens: { total: number | undefined, reasoning?: number | undefined }\n  },\n): void {\n  acc.inputTokens += usage.inputTokens.total ?? 0\n  acc.outputTokens += usage.outputTokens.total ?? 0\n  acc.cacheReadTokens += usage.inputTokens.cacheRead ?? 0\n  acc.cacheWriteTokens += usage.inputTokens.cacheWrite ?? 0\n  acc.reasoningTokens += usage.outputTokens.reasoning ?? 0\n}\n\n/**\n * When using `gateway('google/gemini-3-flash')`, the model object has\n * `provider: 'gateway'` and `modelId: 'google/gemini-3-flash'`.\n * This extracts the real provider and model name from the modelId.\n */\nfunction resolveProviderAndModel(provider: string, modelId: string): { provider: string, model: string } {\n  if (provider !== 'gateway' || !modelId.includes('/')) {\n    return { provider, model: modelId }\n  }\n  const slashIndex = modelId.indexOf('/')\n  return {\n    provider: modelId.slice(0, slashIndex),\n    model: modelId.slice(slashIndex + 1),\n  }\n}\n\n/**\n * Create the evlog AI middleware that captures AI SDK data into a wide event.\n *\n * Use this when you need explicit middleware composition with other wrappers\n * (e.g. supermemory, guardrails). For most cases, use `createAILogger` instead.\n *\n * Note: `captureEmbed` is not available with the raw middleware — use\n * `createAILogger` if you need embedding capture.\n *\n * @example Nuxt API route with supermemory\n * ```ts\n * import { createAIMiddleware } from 'evlog/ai'\n * import { wrapLanguageModel } from 'ai'\n *\n * export default defineEventHandler(async (event) => {\n *   const log = useLogger(event)\n *\n *   const model = wrapLanguageModel({\n *     model: withSupermemory(base, orgId),\n *     middleware: [createAIMiddleware(log, { toolInputs: true })],\n *   })\n * })\n * ```\n */\nexport function createAIMiddleware(log: RequestLogger, options?: AILoggerOptions): LanguageModelV3Middleware {\n  return buildMiddleware(log, options)\n}\n\n/**\n * Create an AI logger that captures AI SDK data into the wide event.\n *\n * Uses model middleware (`wrapLanguageModel`) to transparently intercept\n * all LLM calls. `onFinish` and `onStepFinish` remain free for user code.\n *\n * @example\n * ```ts\n * import { createAILogger } from 'evlog/ai'\n *\n * const log = useLogger(event)\n * const ai = createAILogger(log)\n * const model = ai.wrap('anthropic/claude-sonnet-4.6')\n *\n * const result = streamText({\n *   model,\n *   messages,\n *   onFinish: ({ text }) => saveConversation(text),\n * })\n * ```\n *\n * @example Capture tool call inputs\n * ```ts\n * const ai = createAILogger(log, { toolInputs: true })\n * ```\n */\nexport function createAILogger(log: RequestLogger, options?: AILoggerOptions): AILogger {\n  const state = createAccumulatorState(options)\n  state._log = log\n  const middleware = buildMiddlewareFromState(log, state)\n\n  return {\n    wrap: (model: LanguageModelV3 | GatewayModelId) => {\n      const resolved = typeof model === 'string' ? gateway(model) : model\n      return wrapLanguageModel({ model: resolved, middleware })\n    },\n\n    captureEmbed: (result: {\n      usage: { tokens: number }\n      model?: string\n      dimensions?: number\n      count?: number\n    }) => {\n      state.calls++\n      state.usage.inputTokens += result.usage.tokens\n      state.embedding = {\n        tokens: (state.embedding?.tokens ?? 0) + result.usage.tokens,\n        ...(result.model ? { model: result.model } : state.embedding?.model ? { model: state.embedding.model } : {}),\n        ...(result.dimensions ? { dimensions: result.dimensions } : state.embedding?.dimensions ? { dimensions: state.embedding.dimensions } : {}),\n        ...(result.count ? { count: (state.embedding?.count ?? 0) + result.count } : state.embedding?.count ? { count: state.embedding.count } : {}),\n      }\n      flushState(log, state)\n    },\n\n    getMetadata: () => buildMetadata(state),\n\n    getEstimatedCost: () => computeEstimatedCost(state),\n\n    onUpdate: (callback: AIMetadataListener) => {\n      state.subscribers.add(callback)\n      return () => {\n        state.subscribers.delete(callback)\n      }\n    },\n\n    _state: state,\n  }\n}\n\ninterface AccumulatorState {\n  calls: number\n  steps: number\n  usage: UsageAccumulator\n  models: string[]\n  lastProvider: string | undefined\n  allToolCalls: string[]\n  allToolCallInputs: Array<{ name: string, input: unknown }>\n  stepsUsage: AIStepUsage[]\n  lastFinishReason: string | undefined\n  lastMsToFirstChunk: number | undefined\n  lastMsToFinish: number | undefined\n  lastError: string | undefined\n  lastResponseId: string | undefined\n  toolInputs: boolean\n  toolInputsOptions: ToolInputsOptions | undefined\n  toolExecutions: AIToolExecution[]\n  generationStartTime: number | undefined\n  totalDurationMs: number | undefined\n  embedding: AIEmbeddingData | undefined\n  costMap: Record<string, ModelCost> | undefined\n  subscribers: Set<AIMetadataListener>\n  /** @internal Logger reference for integration flush */\n  _log?: RequestLogger\n}\n\nfunction resolveToolInputs(raw?: boolean | ToolInputsOptions): { enabled: boolean, options: ToolInputsOptions | undefined } {\n  if (!raw) return { enabled: false, options: undefined }\n  if (raw === true) return { enabled: true, options: undefined }\n  return { enabled: true, options: raw }\n}\n\nfunction processToolInput(input: unknown, toolName: string, options: ToolInputsOptions | undefined): unknown {\n  let value = input\n  if (options?.transform) {\n    value = options.transform(value, toolName)\n  }\n  if (options?.maxLength) {\n    const str = typeof value === 'string' ? value : JSON.stringify(value)\n    if (str.length > options.maxLength) {\n      return `${str.slice(0, options.maxLength)}…`\n    }\n  }\n  return value\n}\n\nfunction createAccumulatorState(options?: AILoggerOptions): AccumulatorState {\n  const { enabled, options: captureOpts } = resolveToolInputs(options?.toolInputs)\n  return {\n    calls: 0,\n    steps: 0,\n    usage: {\n      inputTokens: 0,\n      outputTokens: 0,\n      cacheReadTokens: 0,\n      cacheWriteTokens: 0,\n      reasoningTokens: 0,\n    },\n    models: [],\n    lastProvider: undefined,\n    allToolCalls: [],\n    allToolCallInputs: [],\n    stepsUsage: [],\n    lastFinishReason: undefined,\n    lastMsToFirstChunk: undefined,\n    lastMsToFinish: undefined,\n    lastError: undefined,\n    lastResponseId: undefined,\n    toolInputs: enabled,\n    toolInputsOptions: captureOpts,\n    toolExecutions: [],\n    generationStartTime: undefined,\n    totalDurationMs: undefined,\n    embedding: undefined,\n    costMap: options?.cost,\n    subscribers: new Set(),\n  }\n}\n\nfunction computeEstimatedCost(state: AccumulatorState): number | undefined {\n  if (!state.costMap) return undefined\n  const lastModel = state.models[state.models.length - 1]\n  if (!lastModel) return undefined\n  const pricing = state.costMap[lastModel]\n  if (!pricing) return undefined\n  const inputCost = (state.usage.inputTokens / 1_000_000) * pricing.input\n  const outputCost = (state.usage.outputTokens / 1_000_000) * pricing.output\n  const total = inputCost + outputCost\n  return total > 0 ? Math.round(total * 1_000_000) / 1_000_000 : undefined\n}\n\nfunction buildMetadata(state: AccumulatorState): AIMetadata {\n  const uniqueModels = [...new Set(state.models)]\n  const lastModel = state.models[state.models.length - 1]\n\n  const data: AIMetadata = {\n    calls: state.calls,\n    inputTokens: state.usage.inputTokens,\n    outputTokens: state.usage.outputTokens,\n    totalTokens: state.usage.inputTokens + state.usage.outputTokens,\n  }\n\n  if (lastModel) data.model = lastModel\n  if (state.lastProvider) data.provider = state.lastProvider\n  if (uniqueModels.length > 1) data.models = uniqueModels\n  if (state.usage.cacheReadTokens > 0) data.cacheReadTokens = state.usage.cacheReadTokens\n  if (state.usage.cacheWriteTokens > 0) data.cacheWriteTokens = state.usage.cacheWriteTokens\n  if (state.usage.reasoningTokens > 0) data.reasoningTokens = state.usage.reasoningTokens\n  if (state.lastFinishReason) data.finishReason = state.lastFinishReason\n  if (state.toolInputs && state.allToolCallInputs.length > 0) {\n    data.toolCalls = state.allToolCallInputs.map(t => ({ ...t }))\n  } else if (state.allToolCalls.length > 0) {\n    data.toolCalls = [...state.allToolCalls]\n  }\n  if (state.lastResponseId) data.responseId = state.lastResponseId\n  if (state.steps > 1) {\n    data.steps = state.steps\n    data.stepsUsage = state.stepsUsage.map(s => ({ ...s, ...(s.toolCalls ? { toolCalls: [...s.toolCalls] } : {}) }))\n  }\n  if (state.lastMsToFirstChunk !== undefined) data.msToFirstChunk = state.lastMsToFirstChunk\n  if (state.lastMsToFinish !== undefined) {\n    data.msToFinish = state.lastMsToFinish\n    if (state.usage.outputTokens > 0 && state.lastMsToFinish > 0) {\n      data.tokensPerSecond = Math.round((state.usage.outputTokens / state.lastMsToFinish) * 1000)\n    }\n  }\n  if (state.lastError) data.error = state.lastError\n  if (state.toolExecutions.length > 0) data.tools = state.toolExecutions.map(t => ({ ...t }))\n  if (state.totalDurationMs !== undefined) data.totalDurationMs = state.totalDurationMs\n  if (state.embedding) data.embedding = { ...state.embedding }\n  const cost = computeEstimatedCost(state)\n  if (cost !== undefined) data.estimatedCost = cost\n\n  return data\n}\n\nfunction notifySubscribers(state: AccumulatorState, metadata: AIMetadata): void {\n  if (state.subscribers.size === 0) return\n  for (const subscriber of state.subscribers) {\n    try {\n      subscriber(metadata)\n    } catch {\n      // Subscribers must not break the AI flow.\n    }\n  }\n}\n\nfunction flushState(log: RequestLogger, state: AccumulatorState): void {\n  const data = buildMetadata(state)\n  log.set({ ai: data } as Record<string, unknown>)\n  notifySubscribers(state, data)\n}\n\nfunction recordModel(state: AccumulatorState, provider: string, modelId: string, responseModelId?: string): void {\n  const resolved = resolveProviderAndModel(provider, responseModelId ?? modelId)\n  state.models.push(resolved.model)\n  state.lastProvider = resolved.provider\n}\n\nfunction safeParseJSON(input: string): unknown {\n  try {\n    return JSON.parse(input)\n  } catch {\n    return input\n  }\n}\n\nfunction recordError(log: RequestLogger, state: AccumulatorState, model: { provider: string, modelId: string }, error: unknown): void {\n  state.calls++\n  state.steps++\n  recordModel(state, model.provider, model.modelId)\n  state.lastFinishReason = 'error'\n  state.lastError = error instanceof Error ? error.message : String(error)\n\n  const resolved = resolveProviderAndModel(model.provider, model.modelId)\n  state.stepsUsage.push({\n    model: resolved.model,\n    inputTokens: 0,\n    outputTokens: 0,\n  })\n\n  flushState(log, state)\n}\n\nfunction buildMiddleware(log: RequestLogger, options?: AILoggerOptions): LanguageModelV3Middleware {\n  const state = createAccumulatorState(options)\n  state._log = log\n  return buildMiddlewareFromState(log, state)\n}\n\nfunction buildMiddlewareFromState(log: RequestLogger, state: AccumulatorState): LanguageModelV3Middleware {\n  return {\n    specificationVersion: 'v3',\n    wrapGenerate: async ({ doGenerate, model }) => {\n      try {\n        const result = await doGenerate()\n\n        state.calls++\n        state.steps++\n        addUsage(state.usage, result.usage)\n        recordModel(state, model.provider, model.modelId, result.response?.modelId)\n        state.lastFinishReason = result.finishReason.unified\n\n        if (result.response?.id) {\n          state.lastResponseId = result.response.id\n        }\n\n        const stepToolCalls: string[] = []\n        for (const item of result.content) {\n          if (item.type === 'tool-call') {\n            state.allToolCalls.push(item.toolName)\n            stepToolCalls.push(item.toolName)\n            if (state.toolInputs) {\n              const raw = typeof item.input === 'string' ? safeParseJSON(item.input) : item.input\n              state.allToolCallInputs.push({\n                name: item.toolName,\n                input: processToolInput(raw, item.toolName, state.toolInputsOptions),\n              })\n            }\n          }\n        }\n\n        const resolvedModel = resolveProviderAndModel(model.provider, result.response?.modelId ?? model.modelId)\n        state.stepsUsage.push({\n          model: resolvedModel.model,\n          inputTokens: result.usage.inputTokens.total ?? 0,\n          outputTokens: result.usage.outputTokens.total ?? 0,\n          ...(stepToolCalls.length > 0 ? { toolCalls: stepToolCalls } : {}),\n        })\n\n        flushState(log, state)\n        return result\n      } catch (error) {\n        recordError(log, state, model, error)\n        throw error\n      }\n    },\n\n    wrapStream: async ({ doStream, model }) => {\n      const streamStart = Date.now()\n      let firstChunkTime: number | undefined\n\n      let streamUsage: UsageAccumulator | undefined\n      let streamFinishReason: string | undefined\n      let streamModelId: string | undefined\n      let streamResponseId: string | undefined\n      const streamToolCalls: string[] = []\n      const streamToolInputBuffers = new Map<string, { name: string, chunks: string[] }>()\n      let streamError: string | undefined\n\n      let doStreamResult: Awaited<ReturnType<typeof doStream>>\n      try {\n        doStreamResult = await doStream()\n      } catch (error) {\n        recordError(log, state, model, error)\n        throw error\n      }\n\n      const { stream, ...rest } = doStreamResult\n\n      const transformStream = new TransformStream<\n        LanguageModelV3StreamPart,\n        LanguageModelV3StreamPart\n      >({\n        transform(chunk, controller) {\n          if (!firstChunkTime && chunk.type === 'text-delta') {\n            firstChunkTime = Date.now()\n          }\n\n          if (chunk.type === 'tool-input-start') {\n            streamToolCalls.push(chunk.toolName)\n            if (state.toolInputs) {\n              streamToolInputBuffers.set(chunk.id, { name: chunk.toolName, chunks: [] })\n            }\n          }\n\n          if (chunk.type === 'tool-input-delta' && state.toolInputs) {\n            const buffer = streamToolInputBuffers.get(chunk.id)\n            if (buffer) {\n              buffer.chunks.push(chunk.delta)\n            }\n          }\n\n          if (chunk.type === 'tool-input-end' && state.toolInputs) {\n            const buffer = streamToolInputBuffers.get(chunk.id)\n            if (buffer) {\n              const raw = safeParseJSON(buffer.chunks.join(''))\n              state.allToolCallInputs.push({\n                name: buffer.name,\n                input: processToolInput(raw, buffer.name, state.toolInputsOptions),\n              })\n              streamToolInputBuffers.delete(chunk.id)\n            }\n          }\n\n          if (chunk.type === 'finish') {\n            streamUsage = {\n              inputTokens: chunk.usage.inputTokens.total ?? 0,\n              outputTokens: chunk.usage.outputTokens.total ?? 0,\n              cacheReadTokens: chunk.usage.inputTokens.cacheRead ?? 0,\n              cacheWriteTokens: chunk.usage.inputTokens.cacheWrite ?? 0,\n              reasoningTokens: chunk.usage.outputTokens.reasoning ?? 0,\n            }\n            streamFinishReason = chunk.finishReason.unified\n          }\n\n          if (chunk.type === 'response-metadata') {\n            if (chunk.modelId) streamModelId = chunk.modelId\n            if (chunk.id) streamResponseId = chunk.id\n          }\n\n          if (chunk.type === 'error') {\n            streamError = chunk.error instanceof Error ? chunk.error.message : String(chunk.error)\n          }\n\n          controller.enqueue(chunk)\n        },\n\n        flush() {\n          state.calls++\n          state.steps++\n\n          if (streamUsage) {\n            state.usage.inputTokens += streamUsage.inputTokens\n            state.usage.outputTokens += streamUsage.outputTokens\n            state.usage.cacheReadTokens += streamUsage.cacheReadTokens\n            state.usage.cacheWriteTokens += streamUsage.cacheWriteTokens\n            state.usage.reasoningTokens += streamUsage.reasoningTokens\n          }\n\n          recordModel(state, model.provider, model.modelId, streamModelId)\n          state.lastFinishReason = streamFinishReason\n\n          state.allToolCalls.push(...streamToolCalls)\n\n          if (streamResponseId) {\n            state.lastResponseId = streamResponseId\n          }\n\n          if (firstChunkTime) {\n            state.lastMsToFirstChunk = firstChunkTime - streamStart\n          }\n          state.lastMsToFinish = Date.now() - streamStart\n\n          if (streamError) state.lastError = streamError\n\n          const resolvedModel = resolveProviderAndModel(model.provider, streamModelId ?? model.modelId)\n          state.stepsUsage.push({\n            model: resolvedModel.model,\n            inputTokens: streamUsage?.inputTokens ?? 0,\n            outputTokens: streamUsage?.outputTokens ?? 0,\n            ...(streamToolCalls.length > 0 ? { toolCalls: [...streamToolCalls] } : {}),\n          })\n\n          flushState(log, state)\n        },\n      })\n\n      return {\n        stream: stream.pipeThrough(transformStream),\n        ...rest,\n      }\n    },\n  }\n}\n\n/**\n * Create an AI SDK `TelemetryIntegration` that captures tool execution\n * timing, errors, and total generation wall time into the wide event.\n *\n * Complements the middleware-based `createAILogger`: the middleware captures\n * token usage and streaming metrics at the model level, while the integration\n * captures application-level lifecycle events (tool execution, total duration).\n *\n * When passed an `AILogger`, shares its accumulator so both paths write to\n * the same `ai.*` field. Can also be used standalone with a `RequestLogger`.\n *\n * @example Combined with middleware (recommended)\n * ```ts\n * import { createAILogger, createEvlogIntegration } from 'evlog/ai'\n *\n * const log = useLogger(event)\n * const ai = createAILogger(log)\n *\n * const result = await generateText({\n *   model: ai.wrap('anthropic/claude-sonnet-4.6'),\n *   tools: { getWeather },\n *   experimental_telemetry: {\n *     isEnabled: true,\n *     integrations: [createEvlogIntegration(ai)],\n *   },\n * })\n * ```\n *\n * @example Standalone (no middleware wrapping)\n * ```ts\n * import { createEvlogIntegration } from 'evlog/ai'\n *\n * const integration = createEvlogIntegration(log)\n *\n * const result = await generateText({\n *   model: openai('gpt-4o'),\n *   experimental_telemetry: {\n *     isEnabled: true,\n *     integrations: [integration],\n *   },\n * })\n * ```\n */\nexport function createEvlogIntegration(\n  logOrAi: RequestLogger | AILogger,\n  options?: AILoggerOptions,\n): TelemetryIntegration {\n  let log: RequestLogger\n  let state: AccumulatorState\n\n  if ('_state' in logOrAi && logOrAi._state) {\n    state = logOrAi._state\n    log = state._log!\n  } else {\n    log = logOrAi as RequestLogger\n    state = createAccumulatorState(options)\n    state._log = log\n  }\n\n  class EvlogIntegration implements TelemetryIntegration {\n\n    onStart(_event: OnStartEvent) {\n      state.generationStartTime = Date.now()\n    }\n\n    onToolCallFinish(event: OnToolCallFinishEvent) {\n      const execution: AIToolExecution = {\n        name: (event.toolCall as { toolName: string }).toolName,\n        durationMs: event.durationMs,\n        success: event.success,\n      }\n      if (!event.success && event.error) {\n        execution.error = event.error instanceof Error ? event.error.message : String(event.error)\n      }\n      state.toolExecutions.push(execution)\n    }\n\n    onFinish(_event: OnFinishEvent) {\n      if (state.generationStartTime) {\n        state.totalDurationMs = Date.now() - state.generationStartTime\n      }\n      flushState(log, state)\n    }\n  \n  }\n\n  return bindTelemetryIntegration(new EvlogIntegration())\n}\n"],"mappings":";;AAiSA,SAAS,SACP,KACA,OAIM;AACN,KAAI,eAAe,MAAM,YAAY,SAAS;AAC9C,KAAI,gBAAgB,MAAM,aAAa,SAAS;AAChD,KAAI,mBAAmB,MAAM,YAAY,aAAa;AACtD,KAAI,oBAAoB,MAAM,YAAY,cAAc;AACxD,KAAI,mBAAmB,MAAM,aAAa,aAAa;;;;;;;AAQzD,SAAS,wBAAwB,UAAkB,SAAsD;AACvG,KAAI,aAAa,aAAa,CAAC,QAAQ,SAAS,IAAI,CAClD,QAAO;EAAE;EAAU,OAAO;EAAS;CAErC,MAAM,aAAa,QAAQ,QAAQ,IAAI;AACvC,QAAO;EACL,UAAU,QAAQ,MAAM,GAAG,WAAW;EACtC,OAAO,QAAQ,MAAM,aAAa,EAAE;EACrC;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BH,SAAgB,mBAAmB,KAAoB,SAAsD;AAC3G,QAAO,gBAAgB,KAAK,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BtC,SAAgB,eAAe,KAAoB,SAAqC;CACtF,MAAM,QAAQ,uBAAuB,QAAQ;AAC7C,OAAM,OAAO;CACb,MAAM,aAAa,yBAAyB,KAAK,MAAM;AAEvD,QAAO;EACL,OAAO,UAA4C;AAEjD,UAAO,kBAAkB;IAAE,OADV,OAAO,UAAU,WAAW,QAAQ,MAAM,GAAG;IAClB;IAAY,CAAC;;EAG3D,eAAe,WAKT;AACJ,SAAM;AACN,SAAM,MAAM,eAAe,OAAO,MAAM;AACxC,SAAM,YAAY;IAChB,SAAS,MAAM,WAAW,UAAU,KAAK,OAAO,MAAM;IACtD,GAAI,OAAO,QAAQ,EAAE,OAAO,OAAO,OAAO,GAAG,MAAM,WAAW,QAAQ,EAAE,OAAO,MAAM,UAAU,OAAO,GAAG,EAAE;IAC3G,GAAI,OAAO,aAAa,EAAE,YAAY,OAAO,YAAY,GAAG,MAAM,WAAW,aAAa,EAAE,YAAY,MAAM,UAAU,YAAY,GAAG,EAAE;IACzI,GAAI,OAAO,QAAQ,EAAE,QAAQ,MAAM,WAAW,SAAS,KAAK,OAAO,OAAO,GAAG,MAAM,WAAW,QAAQ,EAAE,OAAO,MAAM,UAAU,OAAO,GAAG,EAAE;IAC5I;AACD,cAAW,KAAK,MAAM;;EAGxB,mBAAmB,cAAc,MAAM;EAEvC,wBAAwB,qBAAqB,MAAM;EAEnD,WAAW,aAAiC;AAC1C,SAAM,YAAY,IAAI,SAAS;AAC/B,gBAAa;AACX,UAAM,YAAY,OAAO,SAAS;;;EAItC,QAAQ;EACT;;AA6BH,SAAS,kBAAkB,KAAiG;AAC1H,KAAI,CAAC,IAAK,QAAO;EAAE,SAAS;EAAO,SAAS,KAAA;EAAW;AACvD,KAAI,QAAQ,KAAM,QAAO;EAAE,SAAS;EAAM,SAAS,KAAA;EAAW;AAC9D,QAAO;EAAE,SAAS;EAAM,SAAS;EAAK;;AAGxC,SAAS,iBAAiB,OAAgB,UAAkB,SAAiD;CAC3G,IAAI,QAAQ;AACZ,KAAI,SAAS,UACX,SAAQ,QAAQ,UAAU,OAAO,SAAS;AAE5C,KAAI,SAAS,WAAW;EACtB,MAAM,MAAM,OAAO,UAAU,WAAW,QAAQ,KAAK,UAAU,MAAM;AACrE,MAAI,IAAI,SAAS,QAAQ,UACvB,QAAO,GAAG,IAAI,MAAM,GAAG,QAAQ,UAAU,CAAC;;AAG9C,QAAO;;AAGT,SAAS,uBAAuB,SAA6C;CAC3E,MAAM,EAAE,SAAS,SAAS,gBAAgB,kBAAkB,SAAS,WAAW;AAChF,QAAO;EACL,OAAO;EACP,OAAO;EACP,OAAO;GACL,aAAa;GACb,cAAc;GACd,iBAAiB;GACjB,kBAAkB;GAClB,iBAAiB;GAClB;EACD,QAAQ,EAAE;EACV,cAAc,KAAA;EACd,cAAc,EAAE;EAChB,mBAAmB,EAAE;EACrB,YAAY,EAAE;EACd,kBAAkB,KAAA;EAClB,oBAAoB,KAAA;EACpB,gBAAgB,KAAA;EAChB,WAAW,KAAA;EACX,gBAAgB,KAAA;EAChB,YAAY;EACZ,mBAAmB;EACnB,gBAAgB,EAAE;EAClB,qBAAqB,KAAA;EACrB,iBAAiB,KAAA;EACjB,WAAW,KAAA;EACX,SAAS,SAAS;EAClB,6BAAa,IAAI,KAAK;EACvB;;AAGH,SAAS,qBAAqB,OAA6C;AACzE,KAAI,CAAC,MAAM,QAAS,QAAO,KAAA;CAC3B,MAAM,YAAY,MAAM,OAAO,MAAM,OAAO,SAAS;AACrD,KAAI,CAAC,UAAW,QAAO,KAAA;CACvB,MAAM,UAAU,MAAM,QAAQ;AAC9B,KAAI,CAAC,QAAS,QAAO,KAAA;CAGrB,MAAM,QAFa,MAAM,MAAM,cAAc,MAAa,QAAQ,QAC9C,MAAM,MAAM,eAAe,MAAa,QAAQ;AAEpE,QAAO,QAAQ,IAAI,KAAK,MAAM,QAAQ,IAAU,GAAG,MAAY,KAAA;;AAGjE,SAAS,cAAc,OAAqC;CAC1D,MAAM,eAAe,CAAC,GAAG,IAAI,IAAI,MAAM,OAAO,CAAC;CAC/C,MAAM,YAAY,MAAM,OAAO,MAAM,OAAO,SAAS;CAErD,MAAM,OAAmB;EACvB,OAAO,MAAM;EACb,aAAa,MAAM,MAAM;EACzB,cAAc,MAAM,MAAM;EAC1B,aAAa,MAAM,MAAM,cAAc,MAAM,MAAM;EACpD;AAED,KAAI,UAAW,MAAK,QAAQ;AAC5B,KAAI,MAAM,aAAc,MAAK,WAAW,MAAM;AAC9C,KAAI,aAAa,SAAS,EAAG,MAAK,SAAS;AAC3C,KAAI,MAAM,MAAM,kBAAkB,EAAG,MAAK,kBAAkB,MAAM,MAAM;AACxE,KAAI,MAAM,MAAM,mBAAmB,EAAG,MAAK,mBAAmB,MAAM,MAAM;AAC1E,KAAI,MAAM,MAAM,kBAAkB,EAAG,MAAK,kBAAkB,MAAM,MAAM;AACxE,KAAI,MAAM,iBAAkB,MAAK,eAAe,MAAM;AACtD,KAAI,MAAM,cAAc,MAAM,kBAAkB,SAAS,EACvD,MAAK,YAAY,MAAM,kBAAkB,KAAI,OAAM,EAAE,GAAG,GAAG,EAAE;UACpD,MAAM,aAAa,SAAS,EACrC,MAAK,YAAY,CAAC,GAAG,MAAM,aAAa;AAE1C,KAAI,MAAM,eAAgB,MAAK,aAAa,MAAM;AAClD,KAAI,MAAM,QAAQ,GAAG;AACnB,OAAK,QAAQ,MAAM;AACnB,OAAK,aAAa,MAAM,WAAW,KAAI,OAAM;GAAE,GAAG;GAAG,GAAI,EAAE,YAAY,EAAE,WAAW,CAAC,GAAG,EAAE,UAAU,EAAE,GAAG,EAAE;GAAG,EAAE;;AAElH,KAAI,MAAM,uBAAuB,KAAA,EAAW,MAAK,iBAAiB,MAAM;AACxE,KAAI,MAAM,mBAAmB,KAAA,GAAW;AACtC,OAAK,aAAa,MAAM;AACxB,MAAI,MAAM,MAAM,eAAe,KAAK,MAAM,iBAAiB,EACzD,MAAK,kBAAkB,KAAK,MAAO,MAAM,MAAM,eAAe,MAAM,iBAAkB,IAAK;;AAG/F,KAAI,MAAM,UAAW,MAAK,QAAQ,MAAM;AACxC,KAAI,MAAM,eAAe,SAAS,EAAG,MAAK,QAAQ,MAAM,eAAe,KAAI,OAAM,EAAE,GAAG,GAAG,EAAE;AAC3F,KAAI,MAAM,oBAAoB,KAAA,EAAW,MAAK,kBAAkB,MAAM;AACtE,KAAI,MAAM,UAAW,MAAK,YAAY,EAAE,GAAG,MAAM,WAAW;CAC5D,MAAM,OAAO,qBAAqB,MAAM;AACxC,KAAI,SAAS,KAAA,EAAW,MAAK,gBAAgB;AAE7C,QAAO;;AAGT,SAAS,kBAAkB,OAAyB,UAA4B;AAC9E,KAAI,MAAM,YAAY,SAAS,EAAG;AAClC,MAAK,MAAM,cAAc,MAAM,YAC7B,KAAI;AACF,aAAW,SAAS;SACd;;AAMZ,SAAS,WAAW,KAAoB,OAA+B;CACrE,MAAM,OAAO,cAAc,MAAM;AACjC,KAAI,IAAI,EAAE,IAAI,MAAM,CAA4B;AAChD,mBAAkB,OAAO,KAAK;;AAGhC,SAAS,YAAY,OAAyB,UAAkB,SAAiB,iBAAgC;CAC/G,MAAM,WAAW,wBAAwB,UAAU,mBAAmB,QAAQ;AAC9E,OAAM,OAAO,KAAK,SAAS,MAAM;AACjC,OAAM,eAAe,SAAS;;AAGhC,SAAS,cAAc,OAAwB;AAC7C,KAAI;AACF,SAAO,KAAK,MAAM,MAAM;SAClB;AACN,SAAO;;;AAIX,SAAS,YAAY,KAAoB,OAAyB,OAA8C,OAAsB;AACpI,OAAM;AACN,OAAM;AACN,aAAY,OAAO,MAAM,UAAU,MAAM,QAAQ;AACjD,OAAM,mBAAmB;AACzB,OAAM,YAAY,iBAAiB,QAAQ,MAAM,UAAU,OAAO,MAAM;CAExE,MAAM,WAAW,wBAAwB,MAAM,UAAU,MAAM,QAAQ;AACvE,OAAM,WAAW,KAAK;EACpB,OAAO,SAAS;EAChB,aAAa;EACb,cAAc;EACf,CAAC;AAEF,YAAW,KAAK,MAAM;;AAGxB,SAAS,gBAAgB,KAAoB,SAAsD;CACjG,MAAM,QAAQ,uBAAuB,QAAQ;AAC7C,OAAM,OAAO;AACb,QAAO,yBAAyB,KAAK,MAAM;;AAG7C,SAAS,yBAAyB,KAAoB,OAAoD;AACxG,QAAO;EACL,sBAAsB;EACtB,cAAc,OAAO,EAAE,YAAY,YAAY;AAC7C,OAAI;IACF,MAAM,SAAS,MAAM,YAAY;AAEjC,UAAM;AACN,UAAM;AACN,aAAS,MAAM,OAAO,OAAO,MAAM;AACnC,gBAAY,OAAO,MAAM,UAAU,MAAM,SAAS,OAAO,UAAU,QAAQ;AAC3E,UAAM,mBAAmB,OAAO,aAAa;AAE7C,QAAI,OAAO,UAAU,GACnB,OAAM,iBAAiB,OAAO,SAAS;IAGzC,MAAM,gBAA0B,EAAE;AAClC,SAAK,MAAM,QAAQ,OAAO,QACxB,KAAI,KAAK,SAAS,aAAa;AAC7B,WAAM,aAAa,KAAK,KAAK,SAAS;AACtC,mBAAc,KAAK,KAAK,SAAS;AACjC,SAAI,MAAM,YAAY;MACpB,MAAM,MAAM,OAAO,KAAK,UAAU,WAAW,cAAc,KAAK,MAAM,GAAG,KAAK;AAC9E,YAAM,kBAAkB,KAAK;OAC3B,MAAM,KAAK;OACX,OAAO,iBAAiB,KAAK,KAAK,UAAU,MAAM,kBAAkB;OACrE,CAAC;;;IAKR,MAAM,gBAAgB,wBAAwB,MAAM,UAAU,OAAO,UAAU,WAAW,MAAM,QAAQ;AACxG,UAAM,WAAW,KAAK;KACpB,OAAO,cAAc;KACrB,aAAa,OAAO,MAAM,YAAY,SAAS;KAC/C,cAAc,OAAO,MAAM,aAAa,SAAS;KACjD,GAAI,cAAc,SAAS,IAAI,EAAE,WAAW,eAAe,GAAG,EAAE;KACjE,CAAC;AAEF,eAAW,KAAK,MAAM;AACtB,WAAO;YACA,OAAO;AACd,gBAAY,KAAK,OAAO,OAAO,MAAM;AACrC,UAAM;;;EAIV,YAAY,OAAO,EAAE,UAAU,YAAY;GACzC,MAAM,cAAc,KAAK,KAAK;GAC9B,IAAI;GAEJ,IAAI;GACJ,IAAI;GACJ,IAAI;GACJ,IAAI;GACJ,MAAM,kBAA4B,EAAE;GACpC,MAAM,yCAAyB,IAAI,KAAiD;GACpF,IAAI;GAEJ,IAAI;AACJ,OAAI;AACF,qBAAiB,MAAM,UAAU;YAC1B,OAAO;AACd,gBAAY,KAAK,OAAO,OAAO,MAAM;AACrC,UAAM;;GAGR,MAAM,EAAE,QAAQ,GAAG,SAAS;GAE5B,MAAM,kBAAkB,IAAI,gBAG1B;IACA,UAAU,OAAO,YAAY;AAC3B,SAAI,CAAC,kBAAkB,MAAM,SAAS,aACpC,kBAAiB,KAAK,KAAK;AAG7B,SAAI,MAAM,SAAS,oBAAoB;AACrC,sBAAgB,KAAK,MAAM,SAAS;AACpC,UAAI,MAAM,WACR,wBAAuB,IAAI,MAAM,IAAI;OAAE,MAAM,MAAM;OAAU,QAAQ,EAAE;OAAE,CAAC;;AAI9E,SAAI,MAAM,SAAS,sBAAsB,MAAM,YAAY;MACzD,MAAM,SAAS,uBAAuB,IAAI,MAAM,GAAG;AACnD,UAAI,OACF,QAAO,OAAO,KAAK,MAAM,MAAM;;AAInC,SAAI,MAAM,SAAS,oBAAoB,MAAM,YAAY;MACvD,MAAM,SAAS,uBAAuB,IAAI,MAAM,GAAG;AACnD,UAAI,QAAQ;OACV,MAAM,MAAM,cAAc,OAAO,OAAO,KAAK,GAAG,CAAC;AACjD,aAAM,kBAAkB,KAAK;QAC3B,MAAM,OAAO;QACb,OAAO,iBAAiB,KAAK,OAAO,MAAM,MAAM,kBAAkB;QACnE,CAAC;AACF,8BAAuB,OAAO,MAAM,GAAG;;;AAI3C,SAAI,MAAM,SAAS,UAAU;AAC3B,oBAAc;OACZ,aAAa,MAAM,MAAM,YAAY,SAAS;OAC9C,cAAc,MAAM,MAAM,aAAa,SAAS;OAChD,iBAAiB,MAAM,MAAM,YAAY,aAAa;OACtD,kBAAkB,MAAM,MAAM,YAAY,cAAc;OACxD,iBAAiB,MAAM,MAAM,aAAa,aAAa;OACxD;AACD,2BAAqB,MAAM,aAAa;;AAG1C,SAAI,MAAM,SAAS,qBAAqB;AACtC,UAAI,MAAM,QAAS,iBAAgB,MAAM;AACzC,UAAI,MAAM,GAAI,oBAAmB,MAAM;;AAGzC,SAAI,MAAM,SAAS,QACjB,eAAc,MAAM,iBAAiB,QAAQ,MAAM,MAAM,UAAU,OAAO,MAAM,MAAM;AAGxF,gBAAW,QAAQ,MAAM;;IAG3B,QAAQ;AACN,WAAM;AACN,WAAM;AAEN,SAAI,aAAa;AACf,YAAM,MAAM,eAAe,YAAY;AACvC,YAAM,MAAM,gBAAgB,YAAY;AACxC,YAAM,MAAM,mBAAmB,YAAY;AAC3C,YAAM,MAAM,oBAAoB,YAAY;AAC5C,YAAM,MAAM,mBAAmB,YAAY;;AAG7C,iBAAY,OAAO,MAAM,UAAU,MAAM,SAAS,cAAc;AAChE,WAAM,mBAAmB;AAEzB,WAAM,aAAa,KAAK,GAAG,gBAAgB;AAE3C,SAAI,iBACF,OAAM,iBAAiB;AAGzB,SAAI,eACF,OAAM,qBAAqB,iBAAiB;AAE9C,WAAM,iBAAiB,KAAK,KAAK,GAAG;AAEpC,SAAI,YAAa,OAAM,YAAY;KAEnC,MAAM,gBAAgB,wBAAwB,MAAM,UAAU,iBAAiB,MAAM,QAAQ;AAC7F,WAAM,WAAW,KAAK;MACpB,OAAO,cAAc;MACrB,aAAa,aAAa,eAAe;MACzC,cAAc,aAAa,gBAAgB;MAC3C,GAAI,gBAAgB,SAAS,IAAI,EAAE,WAAW,CAAC,GAAG,gBAAgB,EAAE,GAAG,EAAE;MAC1E,CAAC;AAEF,gBAAW,KAAK,MAAM;;IAEzB,CAAC;AAEF,UAAO;IACL,QAAQ,OAAO,YAAY,gBAAgB;IAC3C,GAAG;IACJ;;EAEJ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8CH,SAAgB,uBACd,SACA,SACsB;CACtB,IAAI;CACJ,IAAI;AAEJ,KAAI,YAAY,WAAW,QAAQ,QAAQ;AACzC,UAAQ,QAAQ;AAChB,QAAM,MAAM;QACP;AACL,QAAM;AACN,UAAQ,uBAAuB,QAAQ;AACvC,QAAM,OAAO;;CAGf,MAAM,iBAAiD;EAErD,QAAQ,QAAsB;AAC5B,SAAM,sBAAsB,KAAK,KAAK;;EAGxC,iBAAiB,OAA8B;GAC7C,MAAM,YAA6B;IACjC,MAAO,MAAM,SAAkC;IAC/C,YAAY,MAAM;IAClB,SAAS,MAAM;IAChB;AACD,OAAI,CAAC,MAAM,WAAW,MAAM,MAC1B,WAAU,QAAQ,MAAM,iBAAiB,QAAQ,MAAM,MAAM,UAAU,OAAO,MAAM,MAAM;AAE5F,SAAM,eAAe,KAAK,UAAU;;EAGtC,SAAS,QAAuB;AAC9B,OAAI,MAAM,oBACR,OAAM,kBAAkB,KAAK,KAAK,GAAG,MAAM;AAE7C,cAAW,KAAK,MAAM;;;AAK1B,QAAO,yBAAyB,IAAI,kBAAkB,CAAC"}