phronomy 0.8.0 → 0.9.0

data/lib/phronomy/agent/base.rb

@@ -1,6 +1,5 @@
 # frozen_string_literal: true
 
-require "digest"
 require "securerandom"
 require_relative "concerns/retryable"
 require_relative "concerns/guardrailable"
@@ -255,10 +254,10 @@ module Phronomy
         # the first time +invoke+ is called. The cache persists for the lifetime
         # of the process; call {.static_knowledge_refresh!} to force a reload.
         #
-        # @param sources [Array<Phronomy::Agent::Context::Knowledge::Source::Base>]
+        # @param sources [Array<Phronomy::Agent::Context::Knowledge::Base>]
         # @example
         #   class PolicyAgent < Phronomy::Agent::Base
-        #     static_knowledge Phronomy::Agent::Context::Knowledge::Source::StaticKnowledge.new(POLICY_TEXT)
+        #     static_knowledge Phronomy::Agent::Context::Knowledge::StaticKnowledge.new(POLICY_TEXT)
         #   end
         # @api public
         def static_knowledge(*sources)
@@ -269,7 +268,7 @@ module Phronomy
         end
 
         # Returns the registered static knowledge sources.
-        # @return [Array<Phronomy::Agent::Context::Knowledge::Source::Base>]
+        # @return [Array<Phronomy::Agent::Context::Knowledge::Base>]
         # @api public
         def static_knowledge_sources
           @static_knowledge_sources || []
@@ -302,80 +301,6 @@ module Phronomy
           @static_knowledge_chunks = nil
         end
 
-        # Registers a callback that is invoked before every LLM call so the
-        # application can remove stale or irrelevant messages from the
-        # conversation history.
-        #
-        # The block receives a {Phronomy::Agent::Context::Conversation::TrimContext} and may call
-        # +ctx.remove(seqs)+ to drop messages by seq number. Changes affect
-        # only the current invocation; the underlying memory store is unchanged.
-        #
-        # @yield [ctx] Phronomy::Agent::Context::Conversation::TrimContext
-        # @example Drop the oldest message when over 80% of budget is used
-        #   on_trim do |ctx|
-        #     limit = ctx.budget&.available(used: 0) || Float::INFINITY
-        #     ctx.remove(ctx.message_elements.first[:seq]) if ctx.total_tokens > limit * 0.8
-        #   end
-        # @api public
-        def on_trim(&block)
-          @on_trim_callback = block
-        end
-
-        # @return [Proc, nil]
-        # @api private
-        def _on_trim_callback
-          @on_trim_callback
-        end
-
-        # Registers a callback that decides whether compaction should run.
-        # Evaluated before every LLM call (after on_trim). If the block returns
-        # truthy AND an +on_compact+ callback is also registered, the compact
-        # pipeline is executed.
-        #
-        # The block receives a read-only {Phronomy::Agent::Context::Conversation::TriggerContext}.
-        #
-        # @yield [ctx] Phronomy::Agent::Context::Conversation::TriggerContext
-        # @return [Boolean] truthy → run on_compact; falsy → skip
-        # @example Trigger when messages exceed 70% of token budget
-        #   on_compaction_trigger do |ctx|
-        #     limit = ctx.budget&.available(used: 0) || Float::INFINITY
-        #     ctx.total_tokens > limit * 0.7
-        #   end
-        # @api public
-        def on_compaction_trigger(&block)
-          @on_compaction_trigger_callback = block
-        end
-
-        # @return [Proc, nil]
-        # @api private
-        def _on_compaction_trigger_callback
-          @on_compaction_trigger_callback
-        end
-
-        # Registers a callback that performs the actual compaction when the
-        # +on_compaction_trigger+ callback fires. The block receives a
-        # {Phronomy::Agent::Context::Conversation::CompactionContext} and should call +ctx.compact+
-        # to specify which messages to summarise.
-        #
-        # @yield [ctx] Phronomy::Agent::Context::Conversation::CompactionContext
-        # @example Replace the first 4 messages with a short summary
-        #   on_compact do |ctx|
-        #     ctx.compact(0..3) do |elements|
-        #       texts = elements.map { |e| e[:message].content }.join(" | ")
-        #       "Earlier conversation summary: #{texts}"
-        #     end
-        #   end
-        # @api public
-        def on_compact(&block)
-          @on_compact_callback = block
-        end
-
-        # @return [Proc, nil]
-        # @api private
-        def _on_compact_callback
-          @on_compact_callback
-        end
-
         # When enabled, attaches Anthropic prompt-cache markers to the system
         # message so that the fixed instructions are served from cache on
         # subsequent turns, reducing input-token costs.
@@ -453,7 +378,7 @@ module Phronomy
 
       # Registers an anonymous handoff tool class on this agent instance.
       # Called by Runner during construction when routes are configured.
-      # @param tool_class [Class<Phronomy::Tool::Base>]
+      # @param tool_class [Class<Phronomy::Agent::Context::Capability::Base>]
       # @return [self]
       # @api private
       def _add_handoff_tool(tool_class)
@@ -482,7 +407,6 @@ module Phronomy
       # @param thread_id [String, nil] conversation thread identifier, forwarded
       #   to the compaction context when on_compact is configured.
       # @param config    [Hash] additional runtime options:
-      #   +:knowledge_sources+ (Array) — dynamic knowledge sources for this turn
       #   +:user_id+    (+String+, optional) — caller identity forwarded to the tracer
       #   +:session_id+ (+String+, optional) — session identity forwarded to the tracer
       # @param invocation_context [Phronomy::InvocationContext, nil] optional first-class context
@@ -519,60 +443,9 @@ module Phronomy
           thread_id, config = _apply_invocation_context(thread_id, config, invocation_context)
         end
         if Phronomy.configuration.event_loop
-          # Protect against blocking the EventLoop thread itself.
-          if Phronomy::EventLoop.current?
-            raise Phronomy::Error,
-              "Cannot call Agent#invoke (EventLoop mode) from within an EventLoop " \
-              "entry action. Use agent.run_as_child(input, ctx: ctx) instead."
-          end
-
-          # Build an effective config that includes the invoke_timeout scope's
-          # CancellationToken before constructing the FSM.  This ensures that
-          # every LLM, tool, and RAG call made inside _invoke_impl observes
-          # cancellation when the deadline fires.
-          timeout_sec = self.class.invoke_timeout
-          effective_config, scope = if timeout_sec
-            s = Phronomy::Concurrency::CancellationScope.new(parent_token: config[:cancellation_token])
-            s.deadline_in(timeout_sec)
-            [config.merge(cancellation_token: s.token), s]
-          else
-            [config, nil]
-          end
-
-          fsm = Agent::FSM.new(
-            agent: self,
-            input: input,
-            messages: messages,
-            thread_id: thread_id || SecureRandom.uuid,
-            config: effective_config
-          )
-          completion_queue = Phronomy::EventLoop.instance.register(fsm)
-          result = if scope
-            scope.pop_queue(completion_queue) do
-              raise Phronomy::TimeoutError,
-                "Agent #{self.class.name} invoke timed out after #{timeout_sec}s"
-            end
-          else
-            completion_queue.pop
-          end
-          raise result if result.is_a?(Exception)
-          result
+          _invoke_via_event_loop(input, messages: messages, thread_id: thread_id, config: config)
         else
-          # Guard: calling invoke from inside a scheduler task would block the task
-          # against itself when using a cooperative backend.  Use invoke_async
-          # instead to compose agents without introducing a blocking wait.
-          if Phronomy::Task.current
-            msg = "#{self.class.name}#invoke called from inside a scheduler task. " \
-              "This blocks the scheduler until the inner invocation completes, preventing " \
-              "other tasks from making progress. Use invoke_async + await instead."
-            if Phronomy.configuration.strict_runtime_guards
-              raise Phronomy::SchedulerReentrancyError, msg
-            elsif Phronomy.configuration.logger
-              Phronomy.configuration.logger.warn(msg)
-            else
-              Kernel.warn("[phronomy] WARNING: #{msg}")
-            end
-          end
+          _check_scheduler_reentrancy
           invoke_async(input, messages: messages, thread_id: thread_id, config: config).await
         end
       end
@@ -687,19 +560,11 @@ module Phronomy
         raise
       end
 
-      # Returns the {LlmContextWindow::ContextVersionCache} built during the most recent
-      # {#invoke} call on this agent instance.  The thread-local cache entry is
-      # cleaned up in the +ensure+ block of {#invoke}, but a reference is kept
-      # in +@last_context_version_cache+ so callers can inspect it after invoke
-      # returns.
-      #
-      # NOTE: Not thread-safe.  When the same Agent instance is used concurrently,
-      # +@last_context_version_cache+ reflects the most recent +invoke+ on *any*
-      # thread.  For per-invocation isolation, use a separate Agent instance per
-      # thread.
+      # @deprecated The context version cache has been removed. Returns nil.
+      #   Retained for backward compatibility with callers using safe navigation (+&.reset+).
       # @api private
       def context_version_cache
-        @last_context_version_cache
+        nil
       end
 
       private
@@ -722,29 +587,75 @@ module Phronomy
         [effective_thread_id, effective_config]
       end
 
-      # Streaming implementation for #stream.
-      def _stream_impl(input, messages: [], thread_id: nil, config: {}, &block)
-        caller_meta = {}
-        caller_meta[:user_id] = config[:user_id] if config[:user_id]
-        caller_meta[:session_id] = config[:session_id] if config[:session_id]
-        if (ic = config[:invocation_context])
-          caller_meta[:task_id] = ic.task_id if ic.task_id
-          caller_meta[:parent_task_id] = ic.parent_task_id if ic.parent_task_id
+      def _invoke_via_event_loop(input, messages:, thread_id:, config:)
+        if Phronomy::EventLoop.current?
+          raise Phronomy::Error,
+            "Cannot call Agent#invoke (EventLoop mode) from within an EventLoop " \
+            "entry action. Use agent.run_as_child(input, ctx: ctx) instead."
+        end
+
+        timeout_sec = self.class.invoke_timeout
+        effective_config, scope = if timeout_sec
+          s = Phronomy::Concurrency::CancellationScope.new(parent_token: config[:cancellation_token])
+          s.deadline_in(timeout_sec)
+          [config.merge(cancellation_token: s.token), s]
+        else
+          [config, nil]
+        end
+
+        fsm = Agent::FSM.new(
+          agent: self,
+          input: input,
+          messages: messages,
+          thread_id: thread_id || SecureRandom.uuid,
+          config: effective_config
+        )
+        completion_queue = Phronomy::EventLoop.instance.register(fsm)
+        result = if scope
+          scope.pop_queue(completion_queue) do
+            raise Phronomy::TimeoutError,
+              "Agent #{self.class.name} invoke timed out after #{timeout_sec}s"
+          end
+        else
+          completion_queue.pop
         end
+        raise result if result.is_a?(Exception)
+        result
+      end
 
-        trace("agent.invoke", input: input, **caller_meta) do |_span|
+      def _check_scheduler_reentrancy
+        return unless Phronomy::Task.current
+
+        msg = "#{self.class.name}#invoke called from inside a scheduler task. " \
+          "This blocks the scheduler until the inner invocation completes, preventing " \
+          "other tasks from making progress. Use invoke_async + await instead."
+        if Phronomy.configuration.strict_runtime_guards
+          raise Phronomy::SchedulerReentrancyError, msg
+        elsif Phronomy.configuration.logger
+          Phronomy.configuration.logger.warn(msg)
+        else
+          Kernel.warn("[phronomy] WARNING: #{msg}")
+        end
+      end
+
+      # Streaming implementation for #stream.
+      def _stream_impl(input, messages: [], thread_id: nil, config: {}, &block)
+        trace("agent.invoke", input: input, **_build_caller_meta(config)) do |_span|
           run_input_guardrails!(input)
 
           chat = build_chat
           user_message = extract_message(input)
+          context = build_context(
+            input,
+            messages: messages,
+            thread_id: thread_id,
+            config: config,
+            budget: build_token_budget,
+            instruction: build_instructions(input),
+            tools: self.class.tools + _handoff_tools
+          )
+          _apply_context_to_chat(chat, context)
 
-          # Assemble context (system prompt + history). Override #build_context to
-          # inject custom context editing logic at the Agent subclass level.
-          context = build_context(input, messages: messages, thread_id: thread_id, config: config)
-          apply_instructions(chat, context[:system]) if context[:system]
-          context[:messages].each { |msg| chat.messages << msg }
-
-          # Wire per-event callbacks to yield StreamEvents.
           current_tool_call = nil
           chat.on_tool_call do |tool_call|
             current_tool_call = tool_call
@@ -758,32 +669,9 @@ module Phronomy
             }))
           end
 
-          # Run before_completion hooks (global → class → instance) before the LLM call.
           run_before_completion_hooks!(chat, config)
 
-          # Route the LLM streaming call through the configured LLMAdapter.
-          # Chunks are pushed into a token queue by the pool worker thread and
-          # drained here (on the caller's side) so that the user block is never
-          # executed on a BlockingAdapterPool worker thread.
-          # The queue capacity is bounded by Configuration#stream_queue_max_size
-          # (nil = unbounded) to provide backpressure against a fast LLM producer.
-          adapter = Phronomy.configuration.llm_adapter
-          chunk_queue = Phronomy::Concurrency::AsyncQueue.new(max_size: Phronomy.configuration.stream_queue_max_size)
-          pending = adapter.stream_async(chat, user_message, config: config, enqueue_to: chunk_queue)
-
-          # Drain the chunk queue on this side (scheduler task / caller thread).
-          loop do
-            chunk = chunk_queue.pop
-            break if chunk.nil? # queue closed — LLM streaming complete
-            block.call(StreamEvent.new(type: :token, payload: {content: chunk.content}))
-            check_cancellation!(config, "invocation cancelled during streaming")
-          end
-
-          response = pending.await
-
-          output = response.content
-          usage = Phronomy::TokenUsage.from_tokens(response.tokens)
-
+          output, usage = _drain_stream(chat, user_message, config, &block)
           run_output_guardrails!(output)
 
           result = {output: output, messages: chat.messages, usage: usage}
@@ -797,139 +685,172 @@ module Phronomy
       # inject custom context editing logic without having to override
       # the full #invoke_once pipeline.
       #
-      # @param input     [String, Hash] the user's input for this turn
-      # @param messages  [Array<RubyLLM::Message>] raw conversation history
-      # @param thread_id [String, nil] conversation thread identifier
-      # @param config    [Hash] the invocation config (see #invoke)
-      # @return [Hash] { system: String|nil, messages: Array }
+      # The keyword arguments +budget+, +instruction+, +tools+, and +knowledge+
+      # carry pre-computed values. Override them in a subclass call to +super+
+      # to inject custom context without recomputing the defaults.
+      #
+      # @param input       [String, Hash] the user's input for this turn
+      # @param messages    [Array<RubyLLM::Message>] raw conversation history
+      # @param thread_id   [String, nil] conversation thread identifier
+      # @param config      [Hash] the invocation config (see #invoke)
+      # @param budget      [LlmContextWindow::TokenBudget, nil] pre-computed token budget
+      # @param instruction [String, nil] pre-computed system instruction
+      # @param tools       [Array<Class>] tool classes to expose
+      # @param knowledge   [Array<Hash>] knowledge chunks ({ content:, type:, source: })
+      # @return [Hash] { system: String|nil, messages: Array, tool_classes: Array }
       # @api public
-      def build_context(input, messages: [], thread_id: nil, config: {})
-        history = prepare_history(messages: messages, thread_id: thread_id, config: config)
-        budget = build_token_budget
-        system_text = build_cached_system_text(input)
-        user_message = extract_message(input)
-
+      def build_context(
+        input,
+        messages: [],
+        thread_id: nil,
+        config: {},
+        budget: build_token_budget,
+        instruction: build_instructions(input),
+        tools: self.class.tools + _handoff_tools,
+        knowledge: self.class.static_knowledge_chunks + instance_knowledge_chunks
+      )
         assembler = LlmContextWindow::Assembler.new(budget: budget)
-        assembler.add_instruction(system_text) if system_text
-        fetch_knowledge_chunks(user_message, config).each do |chunk|
-          assembler.add_knowledge(chunk[:content], type: chunk[:type], source: chunk[:source])
+        assembler.add_instruction(instruction) if instruction
+        assembler.add_capability(tools)
+        knowledge.each { |chunk| assembler.add_knowledge(chunk[:content], type: chunk[:type] || :static, trusted: true, source: chunk[:source]) }
+
+        msgs = Array(messages)
+
+        if budget && budget_exceeded?(msgs)
+          # Default strategy when the token budget is tight:
+          # 1. Compact: keep the most recent half of the messages verbatim and
+          #    replace the older half with a brief omission marker.
+          # 2. Trim: if the compacted history still exceeds the budget, call
+          #    trim_to_budget with the :safe strategy, which discards the oldest
+          #    message one at a time until the history fits.
+          # Subclasses can override build_context to apply a different strategy
+          # (e.g. LLM-based summarisation) before calling super.
+          keep = [msgs.size / 2, 2].max
+          msgs = compact_messages(msgs, keep_tail: keep) do |dropped|
+            "[#{dropped.size} earlier messages omitted]"
+          end
+          remaining = assembler.available_for_messages
+          msgs = trim_to_budget(msgs, remaining: remaining, strategy: :safe)
         end
-        assembler.add_messages(history)
-        assembler.build
+
+        assembler.add_messages(msgs)
+        @last_context = assembler.build
       end
       protected :build_context
 
-      # Fetches knowledge chunks from all registered sources concurrently.
-      #
-      # Each source is spawned as a separate task within a {Phronomy::TaskGroup};
-      # the RAG concurrency gate enforces the +max_concurrent_rag_fetches+ cap.
-      # Results are returned in registration order (spawn order) as a flat array.
+      # Keeps the last +keep+ messages from +messages+, discarding older ones.
+      # Use this inside a +build_context+ override to trim conversation history.
       #
-      # This method is available to subclasses as a building block when
-      # overriding {#build_context}. Pass a custom +query+ to implement
-      # multi-hop RAG or other retrieval strategies.
-      #
-      # @param query  [String] RAG query string (typically the current user message)
-      # @param config [Hash]   invocation config; relevant keys:
-      #   +:knowledge_sources+, +:rag_failure_policy+, +:cancellation_token+, +:rag_timeout+
-      # @return [Array<Hash>] flat list of chunk hashes with +:content+, +:type+, +:source+
-      # @api private
-      def fetch_knowledge_chunks(query, config)
-        sources = Array(config[:knowledge_sources])
-        return [] if sources.empty?
-
-        check_cancellation!(config, "invocation cancelled before RAG fetch")
-
-        # :skip (default) — ignore per-source failures so the agent can still
-        # answer with partial context. :fail surfaces the first error immediately.
-        failure_policy =
-          case config[:rag_failure_policy]
-          when :fail then :fail_fast
-          else :skip_failed
-          end
+      # @param messages [Array<RubyLLM::Message>] conversation history
+      # @param keep     [Integer] number of messages to retain (from the tail)
+      # @return [Array<RubyLLM::Message>]
+      # @api public
+      def trim_messages(messages, keep:)
+        Array(messages).last(keep)
+      end
+      protected :trim_messages
 
-        group = Phronomy::Runtime.instance.task_group(failure_policy: failure_policy)
-        bp = Phronomy.configuration.backpressure
-        rag_on_full = (bp == :raise) ? :reject : (bp || :wait)
-        rag_bp_timeout = Phronomy.configuration.backpressure_timeout
-
-        # Spawn all fetches concurrently. Results are returned in spawn order
-        # (i.e. registration order of knowledge sources) by TaskGroup#await_all.
-        sources.each do |ks|
-          group.spawn do
-            Phronomy::Runtime.instance.gate(:rag).acquire(on_full: rag_on_full, timeout: rag_bp_timeout) do
-              result, elapsed_ms = Phronomy::Runtime.measure_ms do
-                ks.fetch_async(
-                  query: query,
-                  cancellation_token: config[:cancellation_token],
-                  timeout: config[:rag_timeout]
-                ).await
-              end
-              Phronomy.configuration.logger&.debug { "RAG fetch from #{ks.class.name} completed in #{elapsed_ms}ms" }
-              result
-            end
-          end
-        end
+      # Removes the oldest messages one at a time until the count is within +limit+.
+      #
+      # @param messages [Array<RubyLLM::Message>] conversation history
+      # @param limit    [Integer] maximum number of messages to retain
+      # @return [Array<RubyLLM::Message>]
+      # @api public
+      def drop_messages_over(messages, limit:)
+        msgs = Array(messages).dup
+        msgs.shift while msgs.size > limit
+        msgs
+      end
+      protected :drop_messages_over
 
-        # await_all returns results in spawn order; nil entries indicate
-        # skipped failures when using :skip_failed.
-        group.await_all.flat_map { |chunks| Array(chunks) }
+      # Replaces all but the last +keep_tail+ messages with a single system summary.
+      # The block receives the dropped messages and must return a summary String.
+      #
+      # @param messages  [Array<RubyLLM::Message>] conversation history
+      # @param keep_tail [Integer] number of recent messages to preserve verbatim
+      # @yield  [Array<RubyLLM::Message>] the messages being summarised
+      # @yieldreturn [String] summary text
+      # @return [Array<RubyLLM::Message>]
+      # @api public
+      def compact_messages(messages, keep_tail:, &summariser)
+        msgs = Array(messages)
+        return msgs if msgs.size <= keep_tail
+        tail = msgs.last(keep_tail)
+        dropped = msgs.first(msgs.size - keep_tail)
+        summary_text = summariser.call(dropped)
+        [RubyLLM::Message.new(role: :system, content: summary_text)] + tail
       end
-      protected :fetch_knowledge_chunks
+      protected :compact_messages
 
-      # Runs the on_trim / on_compaction_trigger / on_compact pipeline on the
-      # supplied message array and returns the final Array of message objects
-      # ready to pass to the Assembler.
+      # Trims +messages+ to fit within +remaining+ tokens using the given
+      # +strategy+. Returns the trimmed message array without touching the
+      # assembler. The caller is responsible for passing the result to
+      # +assembler.add_messages+ and calling +assembler.build+.
       #
-      # Override this method in a subclass to customize how conversation
-      # history is filtered or compressed before context assembly.
+      # Supported strategies:
+      #   +:safe+ — discard the oldest message one at a time (default)
       #
-      # @param messages  [Array<RubyLLM::Message>] raw conversation history
-      # @param thread_id [String, nil] conversation thread identifier
-      # @param config    [Hash] additional invocation options
-      # @return [Array] filtered and/or compacted message objects
+      # @param messages  [Array<RubyLLM::Message>] conversation history
+      # @param remaining [Integer, nil] token allowance for messages; when +nil+
+      #   the messages are returned unchanged
+      # @param strategy  [Symbol] trim strategy (default +:safe+)
+      # @return [Array<RubyLLM::Message>]
       # @api public
-      def prepare_history(messages: [], thread_id: nil, config: {})
-        budget = build_token_budget
-        elements = build_message_elements(Array(messages))
-
-        if (trim_cb = self.class._on_trim_callback)
-          trim_ctx = Context::Conversation::TrimContext.new(message_elements: elements, budget: budget)
-          trim_cb.call(trim_ctx)
-          elements = trim_ctx.message_elements
+      def trim_to_budget(messages, remaining:, strategy: :safe)
+        return Array(messages) unless remaining
+        msgs = Array(messages)
+        loop do
+          used = msgs.sum { |m| LlmContextWindow::TokenEstimator.estimate(m.content.to_s) }
+          return msgs if used <= remaining
+          break if msgs.empty?
+          msgs = trim_messages(msgs, keep: msgs.size - 1)
         end
+        msgs
+      end
+      protected :trim_to_budget
 
-        if (trigger_cb = self.class._on_compaction_trigger_callback)
-          trigger_ctx = Context::Conversation::TriggerContext.new(message_elements: elements, budget: budget)
-          if trigger_cb.call(trigger_ctx)
-            if (compact_cb = self.class._on_compact_callback)
-              compact_ctx = Context::Conversation::CompactionContext.new(
-                message_elements: elements,
-                budget: budget,
-                thread_id: thread_id
-              )
-              compact_cb.call(compact_ctx)
-              elements = build_message_elements(compact_ctx.result_messages)
-            end
-          end
-        end
+      # Returns +true+ when the estimated token usage of +messages+ exceeds
+      # +threshold+ times the available context budget.
+      # Always returns +false+ when no token budget is available.
+      #
+      # @param messages  [Array<RubyLLM::Message>] conversation history
+      # @param threshold [Float] fraction of the available budget (default 0.8)
+      # @return [Boolean]
+      # @api public
+      def budget_exceeded?(messages, threshold: 0.8)
+        return false unless (b = build_token_budget)
+        total = Array(messages).sum { |m| LlmContextWindow::TokenEstimator.estimate(m.content.to_s) }
+        limit = b.available(used: 0)
+        total > limit * threshold
+      end
+      protected :budget_exceeded?
 
-        elements.map { |e| e[:message] }
+      # Registers a per-instance knowledge source. Knowledge chunks from all
+      # registered sources are included in every LLM call via +build_context+.
+      #
+      # @param source [#fetch] any object responding to +fetch(query:)+
+      # @return [void]
+      # @api public
+      def add_knowledge_source(source)
+        @instance_knowledge_sources ||= []
+        @instance_knowledge_sources << source
       end
-      protected :prepare_history
+      protected :add_knowledge_source
+
+      # Returns knowledge chunks fetched from all instance-level knowledge sources.
+      #
+      # @return [Array<Hash>]
+      # @api private
+      def instance_knowledge_chunks
+        return [] unless @instance_knowledge_sources
+        @instance_knowledge_sources.flat_map { |ks| ks.fetch(query: nil) }
+      end
+      protected :instance_knowledge_chunks
 
       # Performs a single (non-retried) invocation. Extracted so that #invoke can
       # wrap it in a retry loop without duplicating the LLM interaction logic.
       def invoke_once(input, messages: [], thread_id: nil, config: {})
-        caller_meta = {}
-        caller_meta[:user_id] = config[:user_id] if config[:user_id]
-        caller_meta[:session_id] = config[:session_id] if config[:session_id]
-        if (ic = config[:invocation_context])
-          caller_meta[:task_id] = ic.task_id if ic.task_id
-          caller_meta[:parent_task_id] = ic.parent_task_id if ic.parent_task_id
-        end
-
-        trace("agent.invoke", input: input, **caller_meta) do |_span|
+        trace("agent.invoke", input: input, **_build_caller_meta(config)) do |_span|
           Agent::InvocationPipeline.new(self).run(
             input,
             messages: messages,
@@ -939,6 +860,39 @@ module Phronomy
         end
       end
 
+      def _build_caller_meta(config)
+        meta = {}
+        meta[:user_id] = config[:user_id] if config[:user_id]
+        meta[:session_id] = config[:session_id] if config[:session_id]
+        if (ic = config[:invocation_context])
+          meta[:task_id] = ic.task_id if ic.task_id
+          meta[:parent_task_id] = ic.parent_task_id if ic.parent_task_id
+        end
+        meta
+      end
+
+      def _apply_context_to_chat(chat, context)
+        apply_instructions(chat, context[:system]) if context[:system]
+        (context[:tool_classes] || []).each { |tc| chat.with_tool(prepare_tool_class(tc)) }
+        context[:messages].each { |msg| chat.messages << msg }
+      end
+
+      def _drain_stream(chat, user_message, config, &block)
+        adapter = Phronomy.configuration.llm_adapter
+        chunk_queue = Phronomy::Concurrency::AsyncQueue.new(max_size: Phronomy.configuration.stream_queue_max_size)
+        pending = adapter.stream_async(chat, user_message, config: config, enqueue_to: chunk_queue)
+
+        loop do
+          chunk = chunk_queue.pop
+          break if chunk.nil?
+          block.call(StreamEvent.new(type: :token, payload: {content: chunk.content}))
+          check_cancellation!(config, "invocation cancelled during streaming")
+        end
+
+        response = pending.await
+        [response.content, Phronomy::TokenUsage.from_tokens(response.tokens)]
+      end
+
       # Builds a TokenBudget for this agent's model if possible.
       # When context_window is set at the class level, that value is used directly
       # (bypassing the RubyLLM catalogue) — useful for locally-hosted models where
@@ -965,54 +919,6 @@ module Phronomy
         nil
       end
 
-      # Converts a flat Array of message objects into the internal message_elements
-      # format used by TrimContext, TriggerContext, and CompactionContext.
-      # Each element receives a 0-based synthetic seq number.
-      #
-      # @param messages [Array] message-like objects with #role and #content
-      # @return [Array<Hash>]
-      # @api public
-      def build_message_elements(messages)
-        Array(messages).each_with_index.map do |msg, idx|
-          tokens = LlmContextWindow::TokenEstimator.estimate(msg.content.to_s)
-          {seq: idx, message: msg, tokens: tokens, role: msg.role}
-        end
-      end
-
-      # Builds (or returns a cached) system prompt text.
-      # The fingerprint is a SHA-256 digest of the instruction text concatenated
-      # with the content of every registered static knowledge source.
-      # When the fingerprint is unchanged the ContextVersionCache returns the
-      # previously assembled text without re-fetching any sources.
-      #
-      # @param input [String, Hash] the agent's current input (used for template evaluation)
-      # @return [String, nil] assembled system text, or nil when empty
-      # @api public
-      def build_cached_system_text(input)
-        instruction = build_instructions(input)
-
-        static_chunks = self.class.static_knowledge_chunks
-
-        fingerprint = Digest::SHA256.hexdigest(
-          [instruction.to_s, *static_chunks.map { |c| c[:content] }].join("\0")
-        )
-
-        cache = (@context_version_cache ||= LlmContextWindow::ContextVersionCache.new)
-        unless cache.valid?(fingerprint)
-          parts = [instruction]
-          static_chunks.each do |chunk|
-            parts << LlmContextWindow::Assembler.xml_tag(chunk[:content], type: chunk[:type], trusted: true)
-          end
-          cache.update(fingerprint: fingerprint, system_text: parts.compact.join("\n\n"))
-        end
-
-        # Persist a reference on the instance so that context_version_cache
-        # remains accessible after invoke completes.
-        @last_context_version_cache = cache
-
-        cache.system_text.empty? ? nil : cache.system_text
-      end
-
       # Returns the chat class to instantiate for this invocation.
       # When EventLoop mode is enabled ({Phronomy.configuration.event_loop}),
       # returns {ParallelToolChat} so that concurrent tool dispatch is enabled.
@@ -1039,10 +945,6 @@ module Phronomy
           RubyLLM.chat(**opts)
         end
         chat.with_temperature(t) if t
-        self.class.tools.each do |tool_class|
-          chat.with_tool(prepare_tool_class(tool_class))
-        end
-        _handoff_tools.each { |tc| chat.with_tool(tc) }
         chat
       end
 
@@ -1102,7 +1004,7 @@ module Phronomy
       # Builds the final tool class to register with the chat.
       #
       # When an already-instantiated tool object is passed (e.g. a
-      # {Phronomy::Tool::McpTool} returned by +McpTool.from_server+), it is
+      # {Phronomy::Tools::Mcp} returned by +Phronomy::Tools::Mcp.from_server+), it is
       # returned as-is.  RubyLLM's +with_tool+ accepts both classes and
       # instances, so no wrapping is needed.
       #
@@ -1110,7 +1012,7 @@ module Phronomy
       #   1. Alias override — when the Hash form of .tools maps this class to an
       #      explicit name, an anonymous subclass with that tool_name is returned.
       #   2. Scope policy   — when a scope is declared on the tool, the configured
-      #      {Phronomy::Tool::ScopePolicy} (or the default) is evaluated.
+      #      {Phronomy::Agent::Context::Capability::ScopePolicy} (or the default) is evaluated.
       #      +:reject+ wraps the tool to return a denial message without executing.
       #      +:approve+ behaves like requiring approval (same as step 3 when the
       #      tool does not already have +requires_approval+).
@@ -1120,7 +1022,7 @@ module Phronomy
       #      (tool_name, args) and, if it returns falsy, the tool returns a denial
       #      message instead of executing.
       def prepare_tool_class(tool_class)
-        # When an instantiated tool object is passed (e.g. McpTool.from_server
+        # When an instantiated tool object is passed (e.g. Phronomy::Tools::Mcp.from_server
         # returns an instance, not a class), skip class-level processing and
         # return it directly. RubyLLM#with_tool handles both forms.
         return tool_class unless tool_class.is_a?(Class)
@@ -1139,7 +1041,7 @@ module Phronomy
         # Step 2: evaluate scope policy.
         scope = resolved.scope
         if scope
-          policy = @scope_policy || Phronomy::Tool::ScopePolicy::DEFAULT
+          policy = @scope_policy || Phronomy::Agent::Context::Capability::ScopePolicy::DEFAULT
           decision = policy.call(resolved, scope, self)
           case decision
           when :reject