turnkit 0.2.5 → 0.2.6

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 271ce272a71a97aa2991a580f36205e4cef8e19466e2e480b0ac6f0f0225d51f
-  data.tar.gz: b9a0503f499d3eb850e7eece6f508b6fbc206d6398263f6005520b7ef716493b
+  metadata.gz: 34429a11d156c9631705ec193c77c2ad166fb3dffc182a7b730cffd38b52f694
+  data.tar.gz: c497d2042388a33e80c037145e82a6adf1cc47286073441b7fb7f21fcd4a89b7
 SHA512:
-  metadata.gz: f8772f25a95c44b2ba3d1a17a3e89d0ba142d862e798cee6daef9c54e04deaa3d8dee77deae48b5a77f7b6051b467a14c355aabf5115b1ce89832a27c87eb1b6
-  data.tar.gz: 9b12cccaa55c8d791168eca90655e3b9db89409b69fe59f8b45d23bef71aeec296c538696af44e484da9884dfde4ace67bbfd81d4a6647783f1f7f299ef0e485
+  metadata.gz: 330444b7c8964271b8f11ec562f22c331cf6f00d470880082edc1efa263c33708e68b436ed29276c417ec173044993ecb105c05c788fa84405ac34f90f9521a2
+  data.tar.gz: 5bb9900c687ffa6c9eed0678c0d1a36bba08c79ceb0a4ab3767046e772394b4794c5f81f2b9a52142411873f8aea1bc19e148b101b00fc4fd1cb6fe89933f531

data/CHANGELOG.md

@@ -1,5 +1,9 @@
 # Changelog
 
+## 0.2.6 - 2026-06-07
+
+- Add automatic context compaction for long conversations. TurnKit now stores append-only `context_summary` messages and projects compacted history into future model calls while keeping the full transcript durable.
+
 ## 0.2.5 - 2026-06-06
 
 - Add per-agent and per-turn provider thinking configuration.

data/README.md

@@ -148,6 +148,93 @@ turn = conversation.run!
 puts turn.output_text
 ```
 
+### Context compaction
+
+TurnKit automatically compacts long conversations. Older messages are summarized for future model calls, while the original transcript remains stored durably.
+
+```ruby
+conversation = agent.conversation
+conversation.ask("Work through this long task.")
+```
+
+By default, compaction is enabled and uses the current turn model for the summary call. If a turn runs with `gpt-5`, compaction uses `gpt-5` unless you configure a separate summary model.
+
+Disable compaction globally:
+
+```ruby
+TurnKit.compaction = false
+```
+
+Use a different model for summaries:
+
+```ruby
+TurnKit.compaction = {
+  model: "gpt-4.1-mini"
+}
+```
+
+You can also configure the compaction threshold and estimated context limit:
+
+```ruby
+TurnKit.compaction = {
+  model: "gpt-4.1-mini",
+  threshold: 0.75,
+  context_limit: 128_000
+}
+```
+
+Configure compaction for one agent:
+
+```ruby
+agent = TurnKit::Agent.new(
+  name: "engineer",
+  model: "gpt-5",
+  compaction: {
+    model: "gpt-4.1-mini",
+    threshold: 0.75,
+    context_limit: 128_000
+  }
+)
+```
+
+In this example, normal turns use `gpt-5` and compaction summaries use `gpt-4.1-mini`.
+
+Override the model for one manual compaction:
+
+```ruby
+conversation.compact!(model: "gpt-4.1-mini")
+conversation.compact!(focus: "billing migration", model: "gpt-4.1-mini")
+```
+
+Disable compaction for a single turn:
+
+```ruby
+conversation.ask("Continue", compact: false)
+```
+
+Manually compact a conversation:
+
+```ruby
+conversation.compact!
+conversation.compact!(focus: "billing migration")
+```
+
+Compaction is append-only: TurnKit stores a `context_summary` message with metadata describing the message range it replaces for model projection. The original messages are not deleted, so `conversation.messages` remains the full durable transcript. Future model calls see a compacted projection that includes a reference-only summary and the recent tail.
+
+The model-visible projection uses a synthetic summary exchange followed by recent messages:
+
+```text
+user: What did we do so far?
+assistant: [CONTEXT COMPACTION — REFERENCE ONLY] ...
+user: latest request
+```
+
+For a local smoke test without calling a real provider, run:
+
+```sh
+ruby script/manual_compaction.rb
+```
+
 ### Tools
 
 Create a tool:
@@ -539,6 +626,7 @@ TurnKit.cost_limit = nil
 TurnKit.cost_rates = {}
 TurnKit.cost_calculator = nil
 TurnKit.prompt_cache = :auto
+TurnKit.compaction = true
 ```
 
 Override an agent:
@@ -567,6 +655,7 @@ agent = TurnKit::Agent.new(
 | `cost_rates` | Override prices by model. |
 | `cost_calculator` | Override cost calculation. |
 | `prompt_cache` | Use provider prompt caching. |
+| `compaction` | Enable, disable, or configure automatic context compaction. |
 
 ## Contributing
 

data/lib/turnkit/agent.rb

@@ -4,11 +4,11 @@ module TurnKit
   class Agent
     attr_reader :name, :description, :model, :instructions, :tools, :skills, :available_skills, :sub_agents
     attr_reader :client, :store, :max_iterations, :timeout, :cost_limit, :max_depth, :max_tool_executions
-    attr_reader :prompt_sections, :system_prompt, :prompt_mode, :thinking
+    attr_reader :prompt_sections, :system_prompt, :prompt_mode, :thinking, :compaction
 
     def initialize(name:, description: "", model: nil, instructions: "", tools: [], skills: [], available_skills: [], sub_agents: [],
       system_prompt: nil, prompt_sections: nil, prompt_mode: nil, client: nil, store: nil,
-      max_iterations: nil, timeout: nil, cost_limit: nil, max_depth: nil, max_tool_executions: nil, thinking: nil)
+      max_iterations: nil, timeout: nil, cost_limit: nil, max_depth: nil, max_tool_executions: nil, thinking: nil, compaction: nil)
       @name = name.to_s
       @description = description.to_s
       @model = model
@@ -28,6 +28,7 @@ module TurnKit
       @max_depth = max_depth
       @max_tool_executions = max_tool_executions
       @thinking = self.class.normalize_thinking(thinking)
+      @compaction = compaction
       raise ArgumentError, "name is required" if @name.empty?
     end
 

data/lib/turnkit/compaction.rb

@@ -0,0 +1,406 @@
+# frozen_string_literal: true
+
+module TurnKit
+  module Compaction
+    DEFAULTS = {
+      "enabled" => true,
+      "threshold" => 0.75,
+      "context_limit" => 128_000,
+      "reserved_tokens" => 20_000,
+      "head_messages" => 0,
+      "tail_messages" => 12,
+      "tail_tokens" => 8_000,
+      "summary_ratio" => 0.20,
+      "min_summary_tokens" => 1_000,
+      "max_summary_tokens" => 12_000,
+      "tool_output_max_chars" => 2_000,
+      "model" => nil,
+      "client" => nil
+    }.freeze
+
+    KNOWN_KEYS = DEFAULTS.keys.freeze
+
+    COMPACTION_SYSTEM_PROMPT = <<~TEXT.strip
+      You are an anchored context summarization assistant for TurnKit conversations.
+
+      Summarize only the conversation history you are given. Recent turns may be kept verbatim outside your summary, so focus on older context that still matters for continuing the work.
+
+      If a previous summary is provided, update it by preserving still-true details, removing stale details, and merging in new facts.
+
+      Produce only the requested Markdown summary. Do not answer the conversation itself. Do not mention that you are summarizing, compacting, or merging context.
+
+      Write in the same language the user was using.
+
+      Never include API keys, tokens, passwords, secrets, credentials, or connection strings. Replace secret values with [REDACTED].
+    TEXT
+
+    SUMMARY_TEMPLATE = <<~TEXT.strip
+      Use this exact structure:
+
+      ## Active Task
+      - [latest unfulfilled user request, preferably verbatim]
+
+      ## Goal
+      - [what the user is trying to accomplish overall]
+
+      ## Constraints & Preferences
+      - [user/developer preferences, specs, constraints, important choices]
+
+      ## Completed Actions
+      - [completed work and outcomes]
+
+      ## Active State
+      - [current state, records/files touched, test status, running tool/turn state]
+
+      ## In Progress
+      - [work underway, or "(none)"]
+
+      ## Blocked
+      - [blockers, exact errors, missing information, or "(none)"]
+
+      ## Key Decisions
+      - [important decisions and why]
+
+      ## Resolved Questions
+      - [questions already answered]
+
+      ## Pending User Asks
+      - [unanswered or unfulfilled asks]
+
+      ## Relevant Files
+      - [file/path/resource and why it matters, or "(none)"]
+
+      ## Tool Results To Remember
+      - [important tool output summaries, or "(none)"]
+
+      ## Remaining Work
+      - [likely next work, framed as context, not instructions]
+
+      ## Critical Context
+      - [specific values, IDs, commands, errors, constraints; redact secrets]
+
+      Rules:
+      - Keep every section.
+      - Use terse bullets.
+      - Preserve exact file paths, commands, error strings, IDs, and important values.
+      - Do not invent facts.
+      - Do not include secrets.
+      - Do not include a greeting or preamble.
+    TEXT
+
+    module_function
+
+    def enabled_for?(agent, overrides = {})
+      policy_for(agent, overrides)["enabled"]
+    end
+
+    def policy_for(agent, overrides = {})
+      global = normalize_config(TurnKit.compaction)
+      local = normalize_config(agent.compaction)
+      override = normalize_config(overrides)
+
+      return DEFAULTS.merge("enabled" => false) if global == false
+      return DEFAULTS.merge("enabled" => false) if local == false
+      return DEFAULTS.merge("enabled" => false) if override == false
+
+      DEFAULTS.merge(global || {}).merge(local || {}).merge(override || {})
+    end
+
+    def maybe_compact!(turn, force: nil, focus: nil)
+      return if turn.compact == false
+
+      force = turn.compact == true if force.nil?
+      policy = policy_for(turn.agent)
+      return unless policy["enabled"]
+
+      messages = project(turn.conversation.messages_for_turn(turn))
+      return unless force || over_threshold?(messages, policy)
+
+      compact!(turn.conversation, agent: turn.agent, turn: turn, focus: focus, auto: true, overrides: policy, force: true)
+    rescue StandardError => error
+      TurnKit.logger&.warn("TurnKit compaction failed: #{error.class}: #{error.message}")
+      nil
+    end
+
+    def compact!(conversation, agent:, turn: nil, focus: nil, auto: false, overrides: {}, force: true)
+      policy = policy_for(agent, overrides)
+      raise CompactionError, "compaction is disabled" unless policy["enabled"]
+
+      messages = turn ? conversation.messages_for_turn(turn) : conversation.messages
+      projected = project(messages)
+      selected = select_messages(projected, policy)
+      return nil if selected.nil? && auto
+      raise CompactionError, "not enough messages to compact" unless selected
+
+      selected_tokens = estimate_messages_tokens(selected.fetch("middle"))
+      return nil if auto && !force && !over_threshold?(projected, policy)
+
+      summary = generate_summary(
+        agent: agent,
+        policy: policy,
+        messages: selected.fetch("middle"),
+        previous_summary: selected["previous_summary"]&.text,
+        focus: focus,
+        target_tokens: summary_budget(selected_tokens, policy),
+        fallback_model: turn&.model || conversation.model || agent.effective_model,
+        conversation_id: conversation.id,
+        turn_id: turn&.id
+      )
+
+      append_summary(conversation, turn: turn, summary: summary, selected: selected, policy: policy, focus: focus, auto: auto, input_tokens: selected_tokens)
+    rescue CompactionError
+      raise
+    rescue StandardError => error
+      raise CompactionError, "#{error.class}: #{error.message}"
+    end
+
+    def project(messages)
+      rows = Array(messages).sort_by { |message| [ message.sequence.to_i, message.id ] }
+      summaries = active_summaries(rows)
+      ranges = summaries.filter_map { |summary| range_for(summary) }
+      summaries_by_id = summaries.to_h { |summary| [ summary.id, summary ] }
+      inserted = {}
+      projected = []
+
+      rows.each do |message|
+        summaries.each do |summary|
+          range = range_for(summary)
+          next unless range
+          next if inserted[summary.id]
+          next unless range.begin <= message.sequence.to_i
+
+          projected << summary
+          inserted[summary.id] = true
+        end
+
+        if message.context_summary?
+          projected << message if summaries_by_id[message.id] && !inserted[message.id] && !range_for(message)
+          inserted[message.id] = true if summaries_by_id[message.id]
+          next
+        end
+
+        next if ranges.any? { |range| range.cover?(message.sequence.to_i) }
+
+        projected << message
+      end
+
+      summaries.each do |summary|
+        next if inserted[summary.id]
+
+        projected << summary
+        inserted[summary.id] = true
+      end
+
+      projected
+    end
+
+    def estimate_messages_tokens(messages)
+      Array(messages).sum { |message| estimate_text_tokens(message.text) + 8 }
+    end
+
+    def estimate_text_tokens(text)
+      (text.to_s.length / 4.0).ceil
+    end
+
+    def summary_budget(input_tokens, policy)
+      budget = (input_tokens.to_i * policy["summary_ratio"].to_f).ceil
+      budget = [ budget, policy["min_summary_tokens"].to_i ].max
+      [ budget, policy["max_summary_tokens"].to_i ].min
+    end
+
+    def over_threshold?(messages, policy)
+      usable = [ policy["context_limit"].to_i - policy["reserved_tokens"].to_i, 1 ].max
+      estimate_messages_tokens(messages) >= (usable * policy["threshold"].to_f)
+    end
+
+    def select_messages(messages, policy)
+      rows = Array(messages)
+      return nil if rows.length <= policy["head_messages"].to_i + 1
+
+      previous_summary = rows.reverse.find(&:context_summary?)
+      candidates = rows.reject(&:context_summary?)
+      return nil if candidates.length <= policy["head_messages"].to_i + 1
+
+      head_count = policy["head_messages"].to_i
+      tail_start = tail_start_index(candidates, policy)
+      tail_start = [ tail_start, head_count ].max
+      tail_start = expand_tail_start_for_tool_pairs(candidates, tail_start)
+      middle = candidates[head_count...tail_start]
+      return nil if middle.nil? || middle.empty?
+
+      from_sequence = middle.first.sequence.to_i
+      through_sequence = middle.last.sequence.to_i
+      if previous_summary
+        from_sequence = [ from_sequence, previous_summary.sequence.to_i ].min
+        through_sequence = [ through_sequence, previous_summary.sequence.to_i ].max
+      end
+
+      {
+        "middle" => middle,
+        "previous_summary" => previous_summary,
+        "replaces_from_sequence" => from_sequence,
+        "replaces_through_sequence" => through_sequence,
+        "tail_start_sequence" => candidates[tail_start]&.sequence
+      }
+    end
+
+    def build_prompt(previous_summary:, focus:, target_tokens:)
+      parts = []
+      if previous_summary && !previous_summary.empty?
+        parts << <<~TEXT.strip
+          Update the anchored summary below using the conversation history above.
+
+          Preserve still-true details, remove stale details, and merge in new facts. Remove stale details that are no longer relevant or have been superseded.
+
+          <previous-summary>
+          #{previous_summary}
+          </previous-summary>
+        TEXT
+      else
+        parts << <<~TEXT.strip
+          Create a structured context checkpoint for the conversation history above.
+
+          This summary will replace older TurnKit messages in future model prompts while the original messages remain stored durably.
+        TEXT
+      end
+
+      if focus && !focus.to_s.strip.empty?
+        parts << <<~TEXT.strip
+          Focus topic: "#{focus}"
+
+          Preserve extra detail related to this focus topic. Summarize unrelated context more aggressively, but do not omit constraints or active blockers that affect the current task.
+        TEXT
+      end
+
+      parts << "Target length: approximately #{target_tokens} tokens."
+      parts << SUMMARY_TEMPLATE
+      parts.join("\n\n")
+    end
+
+    def normalize_config(value)
+      case value
+      when nil, true
+        nil
+      when false
+        false
+      when Hash
+        attrs = value.transform_keys(&:to_s)
+        unknown = attrs.keys - KNOWN_KEYS
+        raise ConfigError, "unknown compaction options: #{unknown.join(", ")}" if unknown.any?
+
+        attrs
+      else
+        raise ConfigError, "compaction must be true, false, nil, or a Hash"
+      end
+    end
+
+    def range_for(summary)
+      metadata = summary.compaction_metadata
+      from = metadata["replaces_from_sequence"]
+      through = metadata["replaces_through_sequence"]
+      return nil unless from && through
+
+      (from.to_i..through.to_i)
+    end
+
+    def active_summaries(messages)
+      summaries = Array(messages).select(&:context_summary?).sort_by { |summary| summary.sequence.to_i }
+      active = []
+
+      summaries.reverse_each do |summary|
+        next if active.any? { |newer| (range_for(newer)&.cover?(summary.sequence.to_i)) }
+
+        active << summary
+      end
+
+      active.reverse
+    end
+
+    def tail_start_index(messages, policy)
+      max_messages = policy["tail_messages"].to_i
+      max_tokens = policy["tail_tokens"].to_i
+      count = 0
+      tokens = 0
+      index = messages.length
+
+      (messages.length - 1).downto(0) do |i|
+        message_tokens = estimate_text_tokens(messages[i].text) + 8
+        break if count >= max_messages
+        break if count.positive? && tokens + message_tokens > max_tokens
+
+        count += 1
+        tokens += message_tokens
+        index = i
+      end
+
+      index
+    end
+
+    def expand_tail_start_for_tool_pairs(messages, tail_start)
+      index = tail_start
+      while index.positive? && messages[index]&.tool_result?
+        call_id = messages[index].metadata["tool_call_id"]
+        call_index = (index - 1).downto(0).find do |i|
+          messages[i].tool_call? && Array(messages[i].metadata["tool_calls"]).any? { |call| call["id"] == call_id || call[:id] == call_id }
+        end
+        break unless call_index
+
+        index = call_index
+      end
+      index
+    end
+
+    def generate_summary(agent:, policy:, messages:, previous_summary:, focus:, target_tokens:, fallback_model:, conversation_id:, turn_id:)
+      client = policy["client"] || agent.effective_client
+      model = policy["model"] || fallback_model
+      safe_messages = messages.map { |message| sanitize_message(message, policy) }
+      prompt = build_prompt(previous_summary: previous_summary, focus: focus, target_tokens: target_tokens)
+      result = client.chat(
+        model: model,
+        messages: MessageProjection.for(safe_messages) + [ { role: :user, content: prompt } ],
+        tools: [],
+        instructions: COMPACTION_SYSTEM_PROMPT,
+        metadata: { compaction: true, conversation_id: conversation_id, turn_id: turn_id }
+      )
+      text = result.text.to_s.strip
+      raise CompactionError, "compaction model returned an empty summary" if text.empty?
+
+      text
+    end
+
+    def sanitize_message(message, policy)
+      return message unless message.tool_result?
+
+      max = policy["tool_output_max_chars"].to_i
+      return message if max <= 0 || message.text.length <= max
+
+      attrs = message.to_h
+      text = "#{message.text[0, max]}\n\n[Tool result truncated for compaction]"
+      Message.new(attrs.merge("text" => text, "content" => [ { "type" => "text", "text" => text } ]))
+    end
+
+    def append_summary(conversation, turn:, summary:, selected:, policy:, focus:, auto:, input_tokens:)
+      model = policy["model"] || turn&.model || conversation.model || conversation.agent.effective_model
+      conversation.append_message(
+        role: "assistant",
+        kind: "context_summary",
+        text: summary,
+        turn_id: turn&.id,
+        metadata: {
+          "compaction" => {
+            "auto" => auto,
+            "focus" => focus,
+            "replaces_from_sequence" => selected.fetch("replaces_from_sequence"),
+            "replaces_through_sequence" => selected.fetch("replaces_through_sequence"),
+            "tail_start_sequence" => selected["tail_start_sequence"],
+            "summary_model" => model,
+            "input_tokens" => input_tokens,
+            "summary_tokens" => estimate_text_tokens(summary),
+            "created_for_turn_id" => turn&.id,
+            "created_at" => Clock.now.iso8601
+          }.compact
+        }
+      )
+    end
+  end
+end

data/lib/turnkit/conversation.rb

@@ -26,15 +26,16 @@ module TurnKit
       async ? turn : turn.run!
     end
 
-    def run!(trigger_message_id: nil, model: nil, budget: nil, parent_turn: nil, parent_tool_execution: nil, depth: 0, agent: self.agent, thinking: THINKING_UNSET)
-      build_turn(trigger_message_id: trigger_message_id, model: model, budget: budget, parent_turn: parent_turn, parent_tool_execution: parent_tool_execution, depth: depth, agent: agent, thinking: thinking).run!
+    def run!(trigger_message_id: nil, model: nil, budget: nil, parent_turn: nil, parent_tool_execution: nil, depth: 0, agent: self.agent, thinking: THINKING_UNSET, compact: nil)
+      build_turn(trigger_message_id: trigger_message_id, model: model, budget: budget, parent_turn: parent_turn, parent_tool_execution: parent_tool_execution, depth: depth, agent: agent, thinking: thinking, compact: compact).run!
     end
 
-    def build_turn(trigger_message_id: nil, model: nil, budget: nil, parent_turn: nil, parent_tool_execution: nil, depth: 0, agent: self.agent, thinking: THINKING_UNSET)
+    def build_turn(trigger_message_id: nil, model: nil, budget: nil, parent_turn: nil, parent_tool_execution: nil, depth: 0, agent: self.agent, thinking: THINKING_UNSET, compact: nil)
       snapshot = latest_message_sequence
       effective_thinking = thinking.equal?(THINKING_UNSET) ? agent.effective_thinking : Agent.normalize_thinking(thinking)
       options = { "trigger_message_id" => trigger_message_id }.compact
       options["thinking"] = effective_thinking
+      options["compact"] = compact unless compact.nil?
       record = store.create_turn(
         "conversation_id" => id,
         "agent_name" => agent.name,
@@ -49,6 +50,11 @@ module TurnKit
       Turn.new(agent: agent, conversation: self, record: record, store: store, budget: budget, depth: depth)
     end
 
+    def compact!(focus: nil, model: nil)
+      overrides = { "model" => model }.compact
+      TurnKit::Compaction.compact!(self, agent: agent, focus: focus, auto: false, overrides: overrides)
+    end
+
     def messages
       store.list_messages(id).map { |attrs| Message.new(attrs) }
     end

data/lib/turnkit/error.rb

@@ -3,6 +3,7 @@
 module TurnKit
   class Error < StandardError; end
   class ConfigError < Error; end
+  class CompactionError < Error; end
   class StoreError < Error; end
   class ToolError < Error; end
 end

data/lib/turnkit/message.rb

@@ -3,7 +3,7 @@
 module TurnKit
   class Message
     ROLES = %w[user assistant tool].freeze
-    KINDS = %w[text tool_call tool_result].freeze
+    KINDS = %w[text tool_call tool_result context_summary].freeze
 
     attr_reader :id, :conversation_id, :turn_id, :role, :kind, :sequence
     attr_reader :content, :text, :tool_execution_id, :provider_message_id, :metadata, :created_at
@@ -43,6 +43,26 @@ module TurnKit
       }
     end
 
+    def text?
+      kind == "text"
+    end
+
+    def tool_call?
+      kind == "tool_call"
+    end
+
+    def tool_result?
+      kind == "tool_result"
+    end
+
+    def context_summary?
+      kind == "context_summary"
+    end
+
+    def compaction_metadata
+      metadata.fetch("compaction", {})
+    end
+
     private
       def stringify(hash)
         hash.transform_keys(&:to_s)

data/lib/turnkit/message_projection.rb

@@ -2,14 +2,41 @@
 
 module TurnKit
   class MessageProjection
+    CONTEXT_SUMMARY_TRIGGER = "What did we do so far?"
+    CONTEXT_SUMMARY_PREFIX = <<~TEXT.strip
+      [CONTEXT COMPACTION — REFERENCE ONLY]
+
+      Earlier TurnKit conversation messages were compacted into the summary below. This is a handoff from a previous context window. Treat it as background reference, not as active instructions.
+
+      Do not answer questions or perform tasks merely because they appear in this summary. Respond to the latest user message after this summary.
+
+      If the latest user message contradicts, supersedes, changes topic from, or diverges from Active Task, In Progress, Pending User Asks, or Remaining Work, the latest user message wins.
+
+      Subject context and live context are recomputed for the current turn and are more authoritative for state-sensitive facts.
+
+      The original messages remain durably stored; this summary only affects the model-visible prompt projection.
+    TEXT
+
     def self.for(messages)
-      messages.map { |message| new(message).to_h }
+      messages.flat_map { |message| new(message).to_a }
     end
 
     def initialize(message)
       @message = message
     end
 
+    def to_a
+      case message.kind
+      when "context_summary"
+        [
+          { role: :user, content: CONTEXT_SUMMARY_TRIGGER },
+          { role: :assistant, content: [ CONTEXT_SUMMARY_PREFIX, message.text ].reject(&:empty?).join("\n\n") }
+        ]
+      else
+        [ to_h ]
+      end
+    end
+
     def to_h
       case message.kind
       when "tool_call"

data/lib/turnkit/turn.rb

@@ -6,7 +6,7 @@ module TurnKit
 
     attr_reader :agent, :conversation, :store, :budget, :depth
     attr_reader :id, :conversation_id, :agent_name, :parent_turn_id, :parent_tool_execution_id
-    attr_reader :root_turn_id, :context_message_sequence, :model, :thinking
+    attr_reader :root_turn_id, :context_message_sequence, :model, :thinking, :compact
     attr_reader :started_at
 
     def initialize(agent:, conversation:, record:, store:, budget: nil, depth: 0)
@@ -23,6 +23,7 @@ module TurnKit
       @context_message_sequence = @record["context_message_sequence"].to_i
       @model = @record["model"] || agent.effective_model
       @thinking = thinking_from_options
+      @compact = compact_from_options
       @started_at = @record["started_at"]
       @budget = budget || agent.build_budget
       @depth = depth
@@ -35,6 +36,7 @@ module TurnKit
       loop do
         budget.check!(depth: depth)
         budget.count_iteration!
+        TurnKit::Compaction.maybe_compact!(self)
 
         result = agent.effective_client.chat(
           model: model,
@@ -97,6 +99,7 @@ module TurnKit
     def reload
       @record = store.load_turn(id)
       @thinking = thinking_from_options
+      @compact = compact_from_options
       self
     end
 
@@ -106,7 +109,7 @@ module TurnKit
 
     private
       def llm_messages
-        MessageProjection.for(conversation.messages_for_turn(self))
+        MessageProjection.for(TurnKit::Compaction.project(conversation.messages_for_turn(self)))
       end
 
       def thinking_from_options
@@ -116,6 +119,11 @@ module TurnKit
         agent.effective_thinking
       end
 
+      def compact_from_options
+        options = (@record["options"] || {}).transform_keys(&:to_s)
+        options["compact"] if options.key?("compact")
+      end
+
       def persist_assistant_message(result)
         if result.tool_calls?
           conversation.append_message(

data/lib/turnkit/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module TurnKit
-  VERSION = "0.2.5"
+  VERSION = "0.2.6"
 end

data/lib/turnkit.rb

@@ -25,6 +25,7 @@ require_relative "turnkit/prompt_contribution"
 require_relative "turnkit/system_prompt"
 require_relative "turnkit/store"
 require_relative "turnkit/memory_store"
+require_relative "turnkit/compaction"
 require_relative "turnkit/tool"
 require_relative "turnkit/tool_call"
 require_relative "turnkit/tool_execution"
@@ -43,6 +44,7 @@ module TurnKit
     attr_accessor :default_model, :client, :store, :logger
     attr_accessor :max_iterations, :timeout, :max_depth, :max_tool_executions
     attr_accessor :cost_limit, :prompt_cache
+    attr_accessor :compaction
     attr_accessor :cost_rates, :cost_calculator
     attr_accessor :prompt_sections, :prompt_behavior, :available_skills
     attr_accessor :prompt_data_max_chars, :context_contributors
@@ -59,6 +61,7 @@ module TurnKit
   self.max_depth = 3
   self.max_tool_executions = 100
   self.prompt_cache = :auto
+  self.compaction = true
   self.cost_rates = {}
   self.prompt_sections = SystemPrompt::DEFAULT_SECTIONS.dup
   self.prompt_data_max_chars = 20_000

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: turnkit
 version: !ruby/object:Gem::Version
-  version: 0.2.5
+  version: 0.2.6
 platform: ruby
 authors:
 - Sam Couch
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2026-06-06 00:00:00.000000000 Z
+date: 2026-06-07 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: ruby_llm
@@ -42,6 +42,7 @@ files:
 - lib/turnkit/budget.rb
 - lib/turnkit/client.rb
 - lib/turnkit/clock.rb
+- lib/turnkit/compaction.rb
 - lib/turnkit/conversation.rb
 - lib/turnkit/cost.rb
 - lib/turnkit/error.rb