turnkit 0.2.2 → 0.2.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: be681d2deacaf1e3be9de2eb84eef412a686baf90a8b0c0a41280cf6a76ecc55
-  data.tar.gz: f0e6d232f50a67ce4a2cd5c46360549b7755a7b6ab100968bd9a2bf16f3cab0a
+  metadata.gz: 2c02ad5eef683595c702a33806438f414ed2da9e18c607a8b314bba4ae442404
+  data.tar.gz: 4da3877b7c20aecae1dd77e6df4497bb64a3909d28419fb1413feb37fa5fa298
 SHA512:
-  metadata.gz: dc9fbeca56bbdc7e737a56dcbb0caa87eb17186c035f052285532749e1e27546884d020c216c72f008950ad38053fc67dbd71e5cfd8d572f169029d4a78ba116
-  data.tar.gz: ae8b955e099d1d81026ff34b3bae9e4a5009122e1e8cccaa64aed0675f888c0cb12fcae503781daa6ad710e8f9f56aec2387c1ce856cf66d6850430141b28dfe
+  metadata.gz: b5de4c365826d8a4154d2ee013fe0f7289796b91b63eb34ad81693993eb55b8f8d0282f8415e7798f9eb698d2f6f4aa52b79949e1c89c0c64effe506cf26ef0b
+  data.tar.gz: b168324cf4f97485ce7854006565441fd0fe67e1f84835805d98d67f27a2a793fe2ce8bd27a6939c6ccbf3cc92023bc93c8aff5e8049fb0b2991a50548d211d6

data/CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## 0.2.3 - 2026-06-06
+
+- Add Anthropic prompt cache support for stable system prompt sections.
+- Track cache write tokens and aggregate model costs on turns.
+- Refresh README usage examples for prompt caching and usage tracking.
+
 ## 0.2.0 - 2026-06-04
 
 - Add configurable system prompt sections and custom system prompt builders.

data/README.md

@@ -26,33 +26,9 @@ Set a provider key:
 
 ```sh
 export ANTHROPIC_API_KEY=...
-# or OPENAI_API_KEY=..., GEMINI_API_KEY=..., OPENROUTER_API_KEY=...
 ```
 
-TurnKit uses RubyLLM by default. Choose the provider by choosing a RubyLLM model name:
-
-```ruby
-TurnKit.default_model = "claude-sonnet-4-5" # Anthropic
-# TurnKit.default_model = "gpt-4.1-mini"    # OpenAI
-# TurnKit.default_model = "gemini-2.5-flash" # Gemini
-```
-
-You can also override the model per agent or per run.
-
-To use a different model SDK, provide a client object that responds to `chat`:
-
-```ruby
-class MyClient < TurnKit::Client
-  def chat(model:, messages:, tools:, instructions:, temperature: nil, metadata: nil)
-    # Call your provider here.
-    TurnKit::Result.new(text: "provider response", model: model)
-  end
-end
-
-TurnKit.client = MyClient.new
-```
-
-Ask an agent:
+Create an agent:
 
 ```ruby
 require "turnkit"
@@ -68,6 +44,22 @@ puts turn.output_text
 
 ## Usage
 
+Choose a model:
+
+```ruby
+TurnKit.default_model = "claude-sonnet-4-5"
+```
+
+Use OpenAI:
+
+```sh
+export OPENAI_API_KEY=...
+```
+
+```ruby
+TurnKit.default_model = "gpt-4.1-mini"
+```
+
 Create a conversation:
 
 ```ruby
@@ -101,7 +93,7 @@ class SaveReport < TurnKit::Tool
 end
 ```
 
-Use the tool:
+Use a tool:
 
 ```ruby
 agent = TurnKit::Agent.new(
@@ -125,142 +117,99 @@ agent = TurnKit::Agent.new(
 )
 ```
 
-List available skills:
+Delegate to sub-agents:
 
 ```ruby
-research = TurnKit::Skill.from_file(
-  "skills/research.md",
-  description: "Use for source-backed research tasks."
+writer = TurnKit::Agent.new(
+  name: "writer",
+  description: "Draft concise copy."
 )
 
-agent = TurnKit::Agent.new(
-  name: "researcher",
-  instructions: "Prefer primary sources.",
-  tools: [WebSearch, ReadWebPage],
-  available_skills: [research]
+editor = TurnKit::Agent.new(
+  name: "editor",
+  sub_agents: [writer]
 )
-```
 
-Add subject context:
-
-```ruby
-article = Article.find(1)
-conversation = agent.conversation(subject: article)
+turn = editor.conversation.ask("Ask the writer for three headlines.")
+puts turn.output_text
 ```
 
-Choose prompt sections:
+Use prompt caching:
 
 ```ruby
-agent = TurnKit::Agent.new(
-  name: "writer",
-  instructions: "Write plainly.",
-  prompt_sections: %i[agent instructions tools environment]
-)
+TurnKit.prompt_cache = :auto
 ```
 
-Build a custom prompt:
+Disable prompt caching:
 
 ```ruby
-agent = TurnKit::Agent.new(
-  name: "custom",
-  instructions: "Answer in JSON.",
-  system_prompt: ->(prompt) {
-    [
-      prompt.agent_section,
-      prompt.instructions_section,
-      "Return only valid JSON."
-    ].compact.join("\n\n")
-  }
-)
+TurnKit.prompt_cache = :off
 ```
 
-Use safe prompt data blocks for pipeline-specific prompts:
+Split custom prompts:
 
 ```ruby
 agent = TurnKit::Agent.new(
-  name: "researcher",
-  system_prompt: ->(prompt) {
-    [
-      prompt.section(:agent),
-      prompt.section(:behavior),
-      prompt.untrusted_section(
-        :retrieval_context,
-        ExternalSearch.results_for("turnkit"),
-        label: "Retrieved external evidence."
-      ),
-      prompt.section(:tools),
-      prompt.section(:environment)
-    ].compact.join("\n\n")
-  }
+  name: "cached",
+  system_prompt: [
+    "Stable instructions and tool guidance.",
+    TurnKit::SystemPrompt::CACHE_BOUNDARY,
+    "Dynamic subject and live context."
+  ].join("\n")
 )
 ```
 
-Choose a prompt mode:
+Inspect usage:
 
 ```ruby
-TurnKit::Agent.new(name: "main", prompt_mode: :full)    # default sections
-TurnKit::Agent.new(name: "worker", prompt_mode: :minimal) # agent, instructions, behavior, tools, environment
-TurnKit::Agent.new(name: "raw", prompt_mode: :none)     # tiny TurnKit identity prompt
+record = TurnKit.store.load_turn(turn.id)
+record.fetch("usage")
 ```
 
-TurnKit automatically uses the minimal prompt mode for delegated sub-agent turns unless the child agent sets its own `prompt_mode`.
-
-Inject live context on each turn:
+Return usage from custom clients:
 
 ```ruby
-TurnKit.context_contributors << ->(context) {
-  TurnKit::LiveContextContribution.new(
-    name: "account",
-    content: AccountSummary.for(context.conversation.metadata["account_id"]),
-    trusted: false
-  )
-}
+class MyClient < TurnKit::Client
+  def chat(model:, messages:, tools:, instructions:, temperature: nil, metadata: nil)
+    TurnKit::Result.new(
+      text: "provider response",
+      model: model,
+      usage: TurnKit::Usage.new(
+        input_tokens: 100,
+        output_tokens: 20,
+        cached_tokens: 80,
+        cache_write_tokens: 100
+      )
+    )
+  end
+end
 ```
 
-Live context and subject context are rendered below `TurnKit::SystemPrompt::CACHE_BOUNDARY`, so provider adapters can reuse the stable prefix in the future.
-
-Add model-specific prompt guidance:
+Split instructions inside custom clients:
 
 ```ruby
-TurnKit.model_prompt_contributors[/claude/] = ->(context) {
-  TurnKit::PromptContribution.new(
-    stable_prefix: "Provider guidance for #{context.model}.",
-    section_overrides: {
-      behavior: "Be concise, tool-aware, and explicit about uncertainty."
-    }
-  )
-}
+stable, dynamic = TurnKit::SystemPrompt.split_cache_boundary(instructions)
 ```
 
-Inspect prompt shape without storing raw prompt text:
+Send `stable` with provider cache controls.
 
-```ruby
-prompt = TurnKit::SystemPrompt.new(agent: agent, turn: turn, conversation: conversation)
-prompt.report
-# => { "chars" => ..., "hash" => ..., "stable_chars" => ..., "dynamic_chars" => ... }
-```
+Send `dynamic` as normal prompt content.
 
-Delegate to sub-agents:
+Use a custom client:
 
 ```ruby
-writer = TurnKit::Agent.new(
-  name: "writer",
-  description: "Draft concise copy."
-)
-
-editor = TurnKit::Agent.new(
-  name: "editor",
-  sub_agents: [writer]
-)
-
-turn = editor.conversation.ask("Ask the writer for three headlines.")
-puts turn.output_text
+TurnKit.client = MyClient.new
 ```
 
 Install Rails persistence:
 
 ```sh
 bin/rails generate turnkit:install
+```
+
+Run migrations:
+
+```sh
 bin/rails db:migrate
 ```
 
@@ -269,7 +218,6 @@ Configure Rails:
 ```ruby
 TurnKit.store = TurnKit::ActiveRecordStore.new
 TurnKit.default_model = "claude-sonnet-4-5"
-TurnKit.timeout = 300
 ```
 
 Reconcile stale turns:
@@ -289,9 +237,10 @@ TurnKit.timeout = 300
 TurnKit.max_depth = 3
 TurnKit.max_tool_executions = 100
 TurnKit.cost_limit = nil
+TurnKit.prompt_cache = :auto
 ```
 
-Override defaults per agent:
+Override an agent:
 
 ```ruby
 agent = TurnKit::Agent.new(
@@ -303,29 +252,18 @@ agent = TurnKit::Agent.new(
 )
 ```
 
-Override the model for a single conversation or turn:
-
-```ruby
-conversation = agent.conversation(model: "claude-opus-4-1")
-turn = conversation.run!(model: "gpt-4.1-mini")
-```
-
 | Option | Description |
 | --- | --- |
-| `default_model` | Set the default RubyLLM model. The model name determines the provider. |
-| `client` | Set the model client. Defaults to `TurnKit::Adapters::RubyLLM.new`. |
+| `default_model` | Set the default RubyLLM model. |
+| `client` | Set the model client. |
 | `store` | Set the conversation store. |
 | `max_iterations` | Limit model calls per turn. |
 | `timeout` | Limit seconds per root turn. |
 | `max_depth` | Limit sub-agent nesting. |
 | `max_tool_executions` | Limit tool calls per root turn. |
 | `cost_limit` | Limit cost per root turn. |
-| `prompt_sections` | Set default system prompt sections. |
-| `prompt_behavior` | Override the default behavior section text. |
-| `prompt_data_max_chars` | Limit data-block content rendered into prompts. |
-| `context_contributors` | Add live per-turn prompt context blocks. |
-| `system_prompt_contributors` | Add global prompt prefix/suffix/section overrides. |
-| `model_prompt_contributors` | Add model-matched prompt contributions. |
+| `prompt_cache` | Use provider prompt caching. |
+| `prompt_sections` | Set default prompt sections. |
 
 ## Contributing
 

data/lib/turnkit/adapters/ruby_llm.rb

@@ -9,7 +9,7 @@ module TurnKit
         configure_from_environment
 
         chat = ::RubyLLM.chat(model: model)
-        chat.with_instructions(instructions) if instructions && !instructions.empty?
+        add_instructions(chat, instructions, model: model)
         chat.with_temperature(temperature) if temperature
         Array(tools).each { |tool| chat.with_tool(ruby_llm_tool(tool)) }
         Array(messages).each { |message| add_message(chat, message) }
@@ -55,6 +55,37 @@ module TurnKit
           )
         end
 
+        def add_instructions(chat, instructions, model:)
+          return if instructions.nil? || instructions.empty?
+
+          if prompt_cache_enabled? && anthropic_model?(model) && instructions.include?(SystemPrompt::CACHE_BOUNDARY)
+            stable, dynamic = SystemPrompt.split_cache_boundary(instructions)
+            add_system_message(chat, stable, cache: true)
+            add_system_message(chat, dynamic, cache: false)
+          else
+            chat.with_instructions(instructions)
+          end
+        end
+
+        def add_system_message(chat, content, cache: false)
+          content = content.to_s.strip
+          return if content.empty?
+
+          if cache
+            content = ::RubyLLM::Providers::Anthropic::Content.new(content, cache: true)
+          end
+
+          chat.add_message(role: :system, content: content)
+        end
+
+        def prompt_cache_enabled?
+          TurnKit.prompt_cache != :off
+        end
+
+        def anthropic_model?(model)
+          model.to_s.start_with?("claude")
+        end
+
         def ruby_llm_tool_calls(tool_calls)
           return nil if tool_calls.nil? || tool_calls.empty?
 
@@ -88,9 +119,10 @@ module TurnKit
             ToolCall.new(id: call.id, name: call.name, arguments: call.arguments)
           end
           usage = Usage.new(
-            input_tokens: response.respond_to?(:input_tokens) ? response.input_tokens : 0,
-            output_tokens: response.respond_to?(:output_tokens) ? response.output_tokens : 0,
-            cached_tokens: response.respond_to?(:cached_tokens) ? response.cached_tokens : 0
+            input_tokens: token_value(response, :input_tokens),
+            output_tokens: token_value(response, :output_tokens),
+            cached_tokens: token_value(response, :cached_tokens),
+            cache_write_tokens: token_value(response, :cache_creation_tokens)
           )
           Result.new(
             text: response.respond_to?(:content) ? response.content.to_s : response.to_s,
@@ -99,6 +131,10 @@ module TurnKit
             model: response.respond_to?(:model_id) ? response.model_id : model
           )
         end
+
+        def token_value(response, method)
+          response.respond_to?(method) ? response.public_send(method).to_i : 0
+        end
     end
   end
 end

data/lib/turnkit/turn.rb

@@ -123,9 +123,12 @@ module TurnKit
           "input_tokens" => current["input_tokens"].to_i + usage.input_tokens,
           "output_tokens" => current["output_tokens"].to_i + usage.output_tokens,
           "cached_tokens" => current["cached_tokens"].to_i + usage.cached_tokens,
+          "cache_write_tokens" => current["cache_write_tokens"].to_i + usage.cache_write_tokens,
           "total_tokens" => current["total_tokens"].to_i + usage.total_tokens
         }
-        update!(usage: totals, heartbeat_at: Clock.now)
+        attributes = { usage: totals, heartbeat_at: Clock.now }
+        attributes[:cost] = @record["cost"].to_f + usage.cost.to_f if usage.cost
+        update!(attributes)
       end
 
       def update!(attributes)

data/lib/turnkit/usage.rb

@@ -2,17 +2,18 @@
 
 module TurnKit
   class Usage
-    attr_reader :input_tokens, :output_tokens, :cached_tokens, :cost
+    attr_reader :input_tokens, :output_tokens, :cached_tokens, :cache_write_tokens, :cost
 
-    def initialize(input_tokens: 0, output_tokens: 0, cached_tokens: 0, cost: nil)
+    def initialize(input_tokens: 0, output_tokens: 0, cached_tokens: 0, cache_write_tokens: 0, cost: nil)
       @input_tokens = input_tokens.to_i
       @output_tokens = output_tokens.to_i
       @cached_tokens = cached_tokens.to_i
+      @cache_write_tokens = cache_write_tokens.to_i
       @cost = cost
     end
 
     def total_tokens
-      input_tokens + output_tokens + cached_tokens
+      input_tokens + output_tokens + cached_tokens + cache_write_tokens
     end
 
     def to_h
@@ -20,6 +21,7 @@ module TurnKit
         "input_tokens" => input_tokens,
         "output_tokens" => output_tokens,
         "cached_tokens" => cached_tokens,
+        "cache_write_tokens" => cache_write_tokens,
         "total_tokens" => total_tokens,
         "cost" => cost
       }.compact

data/lib/turnkit/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module TurnKit
-  VERSION = "0.2.2"
+  VERSION = "0.2.3"
 end

data/lib/turnkit.rb

@@ -41,7 +41,7 @@ module TurnKit
   class << self
     attr_accessor :default_model, :client, :store, :logger
     attr_accessor :max_iterations, :timeout, :max_depth, :max_tool_executions
-    attr_accessor :cost_limit
+    attr_accessor :cost_limit, :prompt_cache
     attr_accessor :prompt_sections, :prompt_behavior, :available_skills
     attr_accessor :prompt_data_max_chars, :context_contributors
     attr_accessor :system_prompt_contributors, :model_prompt_contributors
@@ -56,6 +56,7 @@ module TurnKit
   self.timeout = 300
   self.max_depth = 3
   self.max_tool_executions = 100
+  self.prompt_cache = :auto
   self.prompt_sections = SystemPrompt::DEFAULT_SECTIONS.dup
   self.prompt_data_max_chars = 20_000
   self.available_skills = []

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: turnkit
 version: !ruby/object:Gem::Version
-  version: 0.2.2
+  version: 0.2.3
 platform: ruby
 authors:
 - Sam Couch
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2026-06-05 00:00:00.000000000 Z
+date: 2026-06-06 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: ruby_llm