@diogonzafe/tokenwatch 0.4.0 → 0.6.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -542,18 +542,50 @@ Budget webhook payload:
542
542
  { "text": "[tokenwatch] Budget alert: user \"user_123\" reached $1.0031 USD (threshold: $1)" }
543
543
  ```
544
544
 
545
+ ### Cost anomaly detection
546
+
547
+ Alert when a single call is anomalously expensive compared to the recent average — catches runaway agents, infinite loops, or abusive usage:
548
+
549
+ ```ts
550
+ const tracker = createTracker({
551
+ anomalyDetection: {
552
+ multiplierThreshold: 3, // alert if call cost > 3× rolling average
553
+ webhookUrl: 'https://hooks.slack.com/...',
554
+ windowHours: 24, // lookback window (default: 24h)
555
+ mode: 'once', // 'once' (default) or 'always'
556
+ },
557
+ })
558
+ ```
559
+
560
+ Checks two axes independently for each tracked call:
561
+ - **Per-model** — compares the call's cost to the average of all prior calls for that model within `windowHours`
562
+ - **Per-user** — same check scoped to a specific user (only fires when `userId` is present)
563
+
564
+ No alert fires on the first call for a given model or user (no history = no baseline). Alerts are cleared when `reset()` is called.
565
+
566
+ Anomaly webhook payload example:
567
+ ```json
568
+ { "text": "[tokenwatch] Anomaly: model \"gpt-4o\" call cost $0.5000 is 4.2x above 24h average ($0.0119)" }
569
+ ```
570
+
571
+ `mode: 'always'` fires on every anomalous call instead of latching after the first.
572
+
545
573
  ---
546
574
 
547
575
  ## CLI
548
576
 
549
577
  ```bash
550
- npx tokenwatch sync # force update cached prices from remote
551
- npx tokenwatch prices # list all models and current prices
552
- npx tokenwatch report # show usage report from ~/.tokenwatch/usage.db
553
- npx tokenwatch help # show help
578
+ npx tokenwatch sync # force update cached prices from remote
579
+ npx tokenwatch prices # list all models and current prices
580
+ npx tokenwatch report # show usage report from ~/.tokenwatch/usage.db
581
+ npx tokenwatch dashboard # open local web dashboard (default port: 4242)
582
+ npx tokenwatch dashboard --port 8080
583
+ npx tokenwatch help # show help
554
584
  ```
555
585
 
556
- `tokenwatch report` reads the local SQLite database and prints:
586
+ ### `tokenwatch report`
587
+
588
+ Reads the local SQLite database and prints:
557
589
 
558
590
  ```
559
591
  ── tokenwatch report ──────────────────────────────
@@ -574,7 +606,101 @@ npx tokenwatch help # show help
574
606
  ───────────────────────────────────────────────────
575
607
  ```
576
608
 
577
- Requires `storage: 'sqlite'` in your app and `better-sqlite3` installed.
609
+ ### `tokenwatch dashboard`
610
+
611
+ Spins up a local web server and opens a dark-themed dashboard with real-time cost data:
612
+
613
+ - **Overview cards** — total cost, tokens, calls, burn rate per hour
614
+ - **Cost over time** — line chart bucketed by time (5min / 1h / 1day depending on filter)
615
+ - **Model breakdown** — doughnut chart + table with cost share per model
616
+ - **By user / feature** — collapsible tables, hidden when empty
617
+ - **Cost forecast** — projected daily and monthly spend based on recent burn rate
618
+ - **Time filter tabs** — 1h | 24h | 7d | 30d | All; updates chart and tables in real-time via SSE
619
+
620
+ Data updates automatically every 3 seconds without refreshing the page. Requires `storage: 'sqlite'` in your app and `better-sqlite3` installed. Zero external dependencies — pure Node.js HTTP server with Chart.js loaded from CDN.
621
+
622
+ ---
623
+
624
+ ## Production Usage
625
+
626
+ ### Storage choice
627
+
628
+ | Setup | Recommended storage |
629
+ |---|---|
630
+ | Single process (monolith, lambda, single pod) | `'sqlite'` — zero config, persists across restarts |
631
+ | Multi-instance (Kubernetes, PaaS with ≥2 pods) | `PostgresStorage` / `MySQLStorage` / `MongoStorage` — shared, unified data |
632
+ | Ephemeral / testing | `'memory'` (default) — resets on restart |
633
+
634
+ ### CI / test environments
635
+
636
+ Disable network calls and staleness warnings in CI:
637
+
638
+ ```ts
639
+ const tracker = createTracker({
640
+ syncPrices: false, // skip remote price fetch — use bundled prices
641
+ warnIfStaleAfterHours: 0, // suppress staleness warning
642
+ })
643
+ ```
644
+
645
+ ### On-prem / air-gapped deployments
646
+
647
+ The daily GitHub Actions workflow updates `prices.json` and publishes a new npm package. Teams that cannot reach GitHub at runtime have two options:
648
+
649
+ 1. **Pin and vendor** — copy `prices.json` from the installed package into your repo and commit it. Pass overrides via `customPrices` for any new models.
650
+ 2. **Self-host the sync** — fork the `scripts/scrape-prices.mjs` script and run it on your own schedule, pointing to your internal registry.
651
+
652
+ Either way, set `syncPrices: false` so the library doesn't try to fetch from GitHub at runtime.
653
+
654
+ ### Anomaly detection in production
655
+
656
+ Enable `anomalyDetection` to catch runaway agents or abuse early:
657
+
658
+ ```ts
659
+ const tracker = createTracker({
660
+ storage: new PostgresStorage(pool),
661
+ anomalyDetection: {
662
+ multiplierThreshold: 3, // alert if a call costs 3x above the rolling average
663
+ webhookUrl: 'https://hooks.slack.com/...',
664
+ windowHours: 24, // baseline window (default: 24h)
665
+ },
666
+ })
667
+ ```
668
+
669
+ ---
670
+
671
+ ## OpenTelemetry Exporter
672
+
673
+ Push tracked usage as metrics to any OTel-compatible backend (Datadog, Honeycomb, Grafana, New Relic, etc.) without changing your existing instrumentation:
674
+
675
+ ```bash
676
+ npm install @opentelemetry/api
677
+ ```
678
+
679
+ ```ts
680
+ import { createTracker } from '@diogonzafe/tokenwatch'
681
+ import { OTelExporter } from '@diogonzafe/tokenwatch/exporters'
682
+
683
+ const tracker = createTracker({
684
+ exporter: new OTelExporter(), // uses the globally-registered MeterProvider
685
+ })
686
+ ```
687
+
688
+ Four metrics are emitted per call, all with `model`, `session.id`, `user.id`, and `feature` attributes (optional fields omitted when absent):
689
+
690
+ | Metric | Type | Description |
691
+ |---|---|---|
692
+ | `tokenwatch.calls` | Counter | Number of LLM API calls |
693
+ | `tokenwatch.input_tokens` | Counter | Input tokens (includes cached + cache-creation) |
694
+ | `tokenwatch.output_tokens` | Counter | Output tokens |
695
+ | `tokenwatch.cost_usd` | Histogram | Cost per call in USD |
696
+
697
+ You must configure a `MeterProvider` before creating the exporter (e.g. using the OpenTelemetry SDK). `OTelExporter` has no compile-time dependency on `@opentelemetry/api` — it resolves it at runtime and throws a helpful error if the package is not installed.
698
+
699
+ Custom meter name:
700
+
701
+ ```ts
702
+ new OTelExporter({ meterName: 'my-service' })
703
+ ```
578
704
 
579
705
  ---
580
706
 
@@ -1,4 +1,4 @@
1
- import { I as IStorage, U as UsageEntry } from './index-CJKk1hHw.cjs';
1
+ import { d as IStorage, U as UsageEntry } from './index-D9xq0RNg.cjs';
2
2
 
3
3
  /**
4
4
  * IStorage adapter for PostgreSQL using the `pg` driver.
@@ -1,4 +1,4 @@
1
- import { I as IStorage, U as UsageEntry } from './index-CJKk1hHw.js';
1
+ import { d as IStorage, U as UsageEntry } from './index-D9xq0RNg.js';
2
2
 
3
3
  /**
4
4
  * IStorage adapter for PostgreSQL using the `pg` driver.