tokenr-ruby 0.1.4 → 0.1.5

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e749473bced8cc1dd80e3173317ce1308b2a6fd2ff24cacdcf535744474d53bb
-  data.tar.gz: b127cc4c1bee8cdba896a8a8cfd9c0b211c041f9a9a5b05dcc2e250902471ad9
+  metadata.gz: 011b6bda965d1ebb0baa5949398253f2e010b86bd8301b3166745494c254394e
+  data.tar.gz: ecd282c0a1e3ede564bef3b1ff90240b530f36c59222e3c11a56e5e8d2b7a2d1
 SHA512:
-  metadata.gz: e8198874061d6ca0378f49fa7c2f40909e938dd03d72ab03c2b4cef7fc64b9cc17eabfb70fda66ccec3675cd0bb2d745ea095f9d4264d44487d2e75656c34627
-  data.tar.gz: b0e5cc6cdbdbb3942477e6637465826451ed09e60ee9cfa4f3ff8a99e9061193fd567158e5edec211cb287fb493e4ea04ffc47a9a573252f53ddeffb7cc4e57f
+  metadata.gz: 78b3feeecae3404792b229e412c66bae6530d4bf489232affc33cfd8616bd22cb02f2c764bac9a0e01cd5ff8b52deeca8c7a3531cf30f6e4c5b7aaf953e35586
+  data.tar.gz: eac13684791536a785e1bc18f426b08a8e7914b7371f3e10849864b78c88d4724660f5cbc1bb5dd9db537e5b85f31ab083355f5bb7abf5ef0e364f9ce1e9d000

data/README.md

@@ -185,12 +185,33 @@ Tokenr.client.get_costs_by_agent(limit: 20)
 Tokenr.client.get_timeseries(interval: "day")
 ```
 
+## Prompt Caching
+
+Both OpenAI and Anthropic support prompt caching, and the SDK handles it automatically.
+
+**OpenAI** includes cached tokens inside `prompt_tokens`. The SDK reads `prompt_tokens_details["cached_tokens"]` and separates them so Tokenr can price each category at the correct rate.
+
+**Anthropic** reports cache tokens as separate fields (`cache_creation_input_tokens` and `cache_read_input_tokens`). The SDK passes these through directly.
+
+For manual tracking, you can pass cache tokens explicitly:
+
+```ruby
+Tokenr.track(
+  provider:           "anthropic",
+  model:              "claude-sonnet-4-20250514",
+  input_tokens:       500,
+  output_tokens:      200,
+  cache_read_tokens:  8000,
+  cache_write_tokens: 2000,
+)
+```
+
 ## How It Works
 
 1. `Tokenr::Integrations::OpenAI.wrap(client)` returns a thin wrapper around your existing client
 2. After each call the wrapper reads token counts from the response `usage` field
 3. Events are pushed onto an in-process queue and flushed to Tokenr in the background
-4. If tracking fails for any reason, the exception is swallowed — your app is unaffected
+4. If tracking fails for any reason, the exception is swallowed and your app is unaffected
 5. On process exit, `at_exit` flushes any remaining queued events
 
 ## Supported Providers
@@ -212,6 +233,20 @@ Tokenr.client.get_timeseries(interval: "day")
 export TOKENR_TOKEN="your-token-here"
 ```
 
+## Running Tests
+
+```bash
+# Unit and mock tests (no API keys needed)
+bundle exec rspec
+
+# Live integration tests (makes real API calls, costs fractions of a cent)
+OPENAI_API_KEY=sk-... ANTHROPIC_API_KEY=sk-ant-... bundle exec rspec spec/live_integration_spec.rb -fd
+```
+
+The live tests make a real call to each provider, then verify that the token counts in the Tokenr payload match what the provider actually returned. This includes a test that triggers Anthropic prompt caching and confirms cache tokens are extracted correctly.
+
+Note: the live tests require the `ruby-openai` and `anthropic` gems to be installed. They are not in the Gemfile since they are optional runtime dependencies.
+
 ## Security
 
 This SDK is open source so you can audit exactly what data is sent and when. The short version:

data/lib/tokenr/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Tokenr
-  VERSION = "0.1.4"
+  VERSION = "0.1.5"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: tokenr-ruby
 version: !ruby/object:Gem::Version
-  version: 0.1.4
+  version: 0.1.5
 platform: ruby
 authors:
 - Tokenr