pikuri-core 0.0.5 → 0.0.7

data/lib/pikuri/agent.rb

@@ -18,10 +18,16 @@ module Pikuri
   # Two seams are visible:
   #
   # 1. **Listeners** ({ListenerList} + {Listener::Base} subclasses)
-  #    — pure consumers of the event stream. The +Agent+ is the
-  #    sole emitter; listeners never write back. New rendering or
-  #    capture targets (a web sink, a structured log) are added
-  #    here without touching {Agent}.
+  #    — pure consumers of the event stream; they never write back.
+  #    The +Agent+ emits every loop-narration {Event} variant;
+  #    extensions emit their own domain events through the
+  #    {ExtensionContext} capability facade handed to
+  #    {Extension#bind}. There is no public path from an +Agent+
+  #    reference to emission — no +listeners+ reader, no +chat+
+  #    reader, no emit method — so holding an agent grants read
+  #    access to its configuration and nothing more. New rendering
+  #    or capture targets (a web sink, a structured log) are added
+  #    as listeners without touching {Agent}.
   # 2. **Controls** ({Control::StepLimit}, {Control::Cancellable},
   #    {Control::Interloper}) — host-facing signal holders. The
   #    +Agent+ reads from them at well-defined boundaries:
@@ -32,24 +38,32 @@ module Pikuri
   #    {Control::Interloper#drain!} on every +after_tool_result+.
   #
   # The two roles are named separately so "what fires when" is a
-  # single grep for +@listeners.emit+ in this file.
+  # single grep for +@listeners.emit+ in this file (loop narration)
+  # plus the capability calls in {ExtensionContext} (domain events).
   #
-  # == Step-exhaustion rescue
+  # == Step-exhaustion policy
   #
   # If the +step_limit:+ {Control::StepLimit} trips during
-  # +Chat#ask+, {#run_loop} catches the +Exceeded+ exception,
-  # emits an {Event::FallbackNotice} to the listener stream, and
-  # hands off to {Synthesizer.run} on a fresh +RubyLLM::Chat+.
-  # The synth reuses the parent's listener stream via
-  # {ListenerList#for_sub_agent} (Terminal padded, TokenLog
-  # zeroed, recorder shared by reference) with a +name:+ derived
-  # from the parent's. The synth shares the parent's
-  # +cancellable+ so a user cancel during synthesis still works,
-  # and gets a fresh +step_limit+ at +max: 1+ (defensive — the
-  # synth has no tools and shouldn't trip it). The synth's
-  # answer becomes the value reported by
-  # {#last_assistant_content}, so callers (notably the +agent+ tool
-  # from +pikuri-subagents+) still get a usable reply.
+  # completion, {#run_loop} catches the +Exceeded+ exception and
+  # applies the budget's {Control::StepLimit#on_exhausted} policy
+  # (see that class header for how hosts pick):
+  #
+  # - +:raise+ (the default) — re-raise to the host, same shape as
+  #   the cancellation rescue below. The chat history survives and
+  #   {Control::StepLimit#reset!} fires at the next turn boundary,
+  #   so a REPL user can simply say "continue".
+  # - +:synthesize+ — emit an {Event::FallbackNotice} and run the
+  #   {Synthesizer} prompt on a nested tools-free +Agent+ (the
+  #   same construction shape the +agent+ tool from
+  #   +pikuri-subagents+ uses for sub-agents): parent's listener
+  #   stream derived via {ListenerList#for_sub_agent} (Terminal
+  #   padded, TokenLog zeroed, recorder shared by reference),
+  #   parent's +cancellable+ shared so a user cancel during
+  #   synthesis still works, a defensive +step_limit+ at +max: 1+
+  #   (the synth has no tools and shouldn't tick it). The synth's
+  #   answer becomes the value reported by
+  #   {#last_assistant_content}, so callers (notably the +agent+
+  #   tool from +pikuri-subagents+) still get a usable reply.
   #
   # == Cancellation rescue
   #
@@ -68,253 +82,13 @@ module Pikuri
     LOGGER = Pikuri.logger_for('Agent')
     private_constant :LOGGER
 
-    # Wire one +RubyLLM::Chat+ for pikuri's event stream and
-    # controls. Used by both {#initialize} (on the main chat) and
-    # {Synthesizer.run} (on the synth chat) so the two share one
-    # source of truth for "which callback emits which event."
-    #
-    # Handles the three message-level registered callbacks
-    # (+after_message+, +before_tool_call+, +after_tool_result+);
-    # the per-chunk streaming callback is separate because
-    # ruby_llm takes it as a block to +Chat#ask+ rather than a
-    # registered hook — see {.streaming_block}.
-    #
-    # @param chat [RubyLLM::Chat] the chat instance to wire
-    # @param listeners [ListenerList] the listener stream events
-    #   flow into
-    # @param step_limit [Control::StepLimit, nil] when set,
-    #   {Control::StepLimit#tick!} is poked on every
-    #   +before_tool_call+ (and raises {Control::StepLimit::Exceeded}
-    #   when over budget)
-    # @param cancellable [Control::Cancellable, nil] when set,
-    #   {Control::Cancellable#check!} is poked on every
-    #   +before_tool_call+ (and raises
-    #   {Control::Cancellable::Cancelled} when the flag is up)
-    # @param interloper [Control::Interloper, nil] when set, the
-    #   queue is drained on every +after_tool_result+, each item
-    #   appended as a +role: :user+ message and emitted as
-    #   {Event::UserTurn} with +mid_loop: true+
-    # @param on_user_message [Proc, nil] when set, called with each
-    #   drained interloper +content+ String *after* it is appended
-    #   to the chat — the per-turn {Extension#on_user_message}
-    #   dispatch (prefetch + recording). Threaded through here rather
-    #   than fired inline so {Synthesizer.run}, which reuses this
-    #   wiring without an interloper or memory, simply passes +nil+.
-    #   Only consulted when +interloper+ is also set.
-    # @return [void]
-    def self.wire_chat(chat, listeners:, step_limit: nil, cancellable: nil, interloper: nil,
-                       on_user_message: nil)
-      chat.after_message do |msg|
-        emit_after_message(msg, listeners)
-      end
-      chat.before_tool_call do |tc|
-        listeners.emit(Event::ToolCall.new(name: tc.name, arguments: tc.arguments))
-        step_limit&.tick!
-        cancellable&.check!
-      end
-      chat.after_tool_result do |result|
-        listeners.emit(Event::ToolResult.new(content: result))
-        drain_interloper(interloper, chat, listeners, on_user_message) if interloper
-      end
-    end
-
-    # Build the per-chunk streaming block passed to +Chat#ask+.
-    # Each invocation of the returned proc converts one
-    # +RubyLLM::Chunk+ into zero, one, or two delta events
-    # ({Event::ThinkingDelta} / {Event::AssistantDelta}) on
-    # +listeners+. Tool-call chunks are intentionally ignored —
-    # partial JSON has no useful rendering; the assembled
-    # +tool_calls+ surface through {Event::ToolCall} once the
-    # message completes.
-    #
-    # Lives parallel to {.wire_chat} (instead of being folded into
-    # it) because +Chat#ask+ takes the streaming block as an
-    # argument rather than a registered callback, so both
-    # {#run_loop} and {Synthesizer.run} pass it inline at the call
-    # site with +&Agent.streaming_block(listeners: ..., cancellable: ...)+.
-    #
-    # == Cancellation polling
-    #
-    # When +cancellable+ is non-nil, {Control::Cancellable#check!}
-    # fires *before* each chunk's emit. The +before_tool_call+
-    # wiring in {.wire_chat} only fires when the model requests a
-    # tool, which leaves a no-tool turn (e.g. a plain greeting)
-    # with zero cancellation points — Ctrl+C trips the flag but
-    # nothing reads it. Polling on every streamed chunk closes
-    # that gap: an in-flight Cancellation+check! raises on the
-    # next chunk delivered after the flag flips, the exception
-    # propagates out through ruby_llm's streaming path
-    # (+Chat#ask+ doesn't rescue), and {#run_loop} catches it,
-    # emits {Event::Cancelled}, and re-raises. The pre-emit
-    # ordering is deliberate: a chunk that arrives after a cancel
-    # request shouldn't render — the user has said stop.
-    #
-    # @param listeners [ListenerList] the listener stream chunk
-    #   events flow into
-    # @param cancellable [Control::Cancellable, nil] when non-nil,
-    #   polled on every chunk so a flag flipped mid-stream raises
-    #   {Control::Cancellable::Cancelled} on the very next chunk
-    # @return [Proc] a +-> (chunk) { ... }+ proc suitable for
-    #   passing to +Chat#ask+ with +&+
-    def self.streaming_block(listeners:, cancellable: nil)
-      ->(chunk) {
-        cancellable&.check!
-        emit_chunk(chunk, listeners)
-      }
-    end
-
-    # Normalize a +RubyLLM::Chat+ +after_message+ payload into
-    # zero, one, or two {Event} variants (+Thinking+ and/or
-    # +Assistant+) plus one {Event::Tokens} for the usage block.
-    # Empty thinking / empty content are filtered here so
-    # listeners never see vacuous events. Non-assistant roles
-    # (e.g. tool-role messages echoed back through
-    # +after_message+) are skipped entirely.
-    #
-    # +msg+ is a +RubyLLM::Message+. Beyond +role+, +content+,
-    # +thinking+, and the +*_tokens+ accessors used here, it also
-    # carries +msg.tool_calls+ on assistant turns that requested
-    # one and +msg.raw+ for the unparsed provider payload.
-    #
-    # @param msg [RubyLLM::Message]
-    # @param listeners [ListenerList]
-    # @return [void]
-    def self.emit_after_message(msg, listeners)
-      return unless msg.role == :assistant
-
-      text = msg.thinking&.text
-      listeners.emit(Event::Thinking.new(content: text)) if text && !text.empty?
-
-      content = msg.content
-      listeners.emit(Event::Assistant.new(content: content)) if content.is_a?(String) && !content.empty?
-
-      listeners.emit(Event::Tokens.new(
-                       input: msg.input_tokens,
-                       output: msg.output_tokens,
-                       cached: msg.cached_tokens,
-                       cache_creation: msg.cache_creation_tokens,
-                       thinking: msg.thinking_tokens,
-                       model_id: msg.model_id
-                     ))
-    end
-    private_class_method :emit_after_message
-
-    # Normalize a +RubyLLM::Chunk+ from a streaming +Chat#ask+
-    # into zero, one, or two delta events
-    # ({Event::ThinkingDelta} / {Event::AssistantDelta}). Empty
-    # +thinking.text+ and empty +content+ are filtered here so
-    # listeners never see vacuous fragments. Tool-call deltas are
-    # intentionally skipped — see {.streaming_block}.
-    #
-    # +chunk+ is a +RubyLLM::Chunk+ (subclass of +RubyLLM::Message+),
-    # so the same +.thinking+ / +.content+ accessors used in
-    # {.emit_after_message} apply.
-    #
-    # @param chunk [RubyLLM::Chunk]
-    # @param listeners [ListenerList]
-    # @return [void]
-    def self.emit_chunk(chunk, listeners)
-      thinking = chunk.thinking&.text
-      listeners.emit(Event::ThinkingDelta.new(content: thinking)) if thinking && !thinking.empty?
-
-      content = chunk.content
-      listeners.emit(Event::AssistantDelta.new(content: content)) if content.is_a?(String) && !content.empty?
-    end
-    private_class_method :emit_chunk
-
-    # Drain the interloper queue: for each pending item, append a
-    # +role: :user+ message to the chat history so the next
-    # round-trip sees it, emit an {Event::UserTurn} with
-    # +mid_loop: true+ to the listener stream so renderers see the
-    # injection, then run the per-turn {Extension#on_user_message}
-    # dispatch (so mid-loop injections are prefetched + recorded
-    # exactly like initial turns).
-    #
-    # The dispatch runs *after* the +:user+ append so any
-    # +<memory-context>+ it injects lands as a +:system+ message
-    # right behind the user turn it annotates — the same
-    # append-at-the-tail ordering {#run_loop} produces for initial
-    # turns.
-    #
-    # @param interloper [Control::Interloper]
-    # @param chat [RubyLLM::Chat]
-    # @param listeners [ListenerList]
-    # @param on_user_message [Proc, nil] per-content dispatch; +nil+
-    #   skips it (e.g. an interloper with no memory extension wired)
-    # @return [void]
-    def self.drain_interloper(interloper, chat, listeners, on_user_message = nil)
-      interloper.drain!.each do |content|
-        chat.add_message(role: :user, content: content)
-        listeners.emit(Event::UserTurn.new(content: content, mid_loop: true))
-        on_user_message&.call(content)
-      end
-    end
-    private_class_method :drain_interloper
-
-    # One-shot inference. Builds a fresh +RubyLLM::Chat+ with no
-    # tools, no MCP, no listeners, no step budget, asks +prompt+ as
-    # the single user turn, and returns the assistant's reply as a
-    # plain String. Lives parallel to {#initialize} / {#run_loop}
-    # because the use case (e.g. summarizing an MCP server's tool
-    # set into a short description block before any agent turn
-    # runs) is genuinely one-shot — there is no loop, no tool
-    # iteration, no listener stream.
-    #
-    # +prompt+ is sent as the user message. For a one-shot call
-    # there is no behavioral difference between the system slot
-    # and the user slot, so we use one parameter; pack any
-    # "instructions + data" framing into +prompt+ directly.
-    #
-    # == Cancellation
-    #
-    # {Control::Cancellable#check!} fires once before the call and
-    # once after, so a flag flipped right around the request
-    # raises {Control::Cancellable::Cancelled} promptly. The
-    # in-flight HTTP call itself is *not* interrupted — same
-    # "gentle cancel" semantic the main loop offers (see
-    # {Control::Cancellable}'s class header). For 30s synthesis
-    # passes at boot this is still a useful escape hatch: the next
-    # check raises and the call returns.
-    #
-    # == Failure
-    #
-    # Errors from the provider (HTTP failure, malformed response,
-    # +RubyLLM+ raising) propagate to the caller verbatim — there
-    # is no recovery layer here. Callers that want "fail soft on
-    # synthesis errors" (e.g. {Mcp::Servers}) rescue at their level
-    # and fall back to a default; this method stays loud.
-    #
-    # @param transport [ChatTransport] same model-resolution
-    #   triple {#initialize} uses; if +model+ is +nil+, falls
-    #   back to +RubyLLM.config.default_model+
-    # @param prompt [String] the prompt sent as the single user
-    #   turn; must be non-blank
-    # @param cancellable [Control::Cancellable, nil] when set,
-    #   checked before the call so a flag flipped right
-    #   around the request raises {Control::Cancellable::Cancelled}
-    # @return [String] the assistant's reply content
-    # @raise [ArgumentError] when +prompt+ is +nil+, empty, or
-    #   whitespace-only
-    # @raise [Control::Cancellable::Cancelled] when the
-    #   +cancellable+ flag was tripped at the pre-call check
-    def self.think(transport:, prompt:, cancellable: nil)
-      raise ArgumentError, "prompt must not be blank, got #{prompt.inspect}" \
-        if prompt.nil? || prompt.to_s.strip.empty?
-
-      transport = transport.with(model: RubyLLM.config.default_model) unless transport.model
-      cancellable&.check!
-      chat = RubyLLM.chat(**transport.to_h)
-      chat.ask(prompt)
-      last = chat.messages.reverse.find { |m| m.role == :assistant }
-      last&.content.to_s
-    end
-
-    # @param transport [ChatTransport] the model-resolution triple
-    #   (+model+ / +provider+ / +assume_model_exists+) forwarded
-    #   to +RubyLLM.chat+. Bundled into one value object so every
-    #   construction site — this constructor and the synthesizer
-    #   rescue below — can forward all three with one assignment
-    #   instead of three kwargs (where dropping one would silently
+    # @param transport [ChatTransport] the model-resolution bundle
+    #   (+model+ / +provider+ / +assume_model_exists+ and, for a model
+    #   on a non-global server, +api_base+ / +api_key+) the chat is
+    #   built from. Bundled into one value object so every construction
+    #   site — this constructor, the synthesizer rescue below, a
+    #   mid-conversation switch — can forward it with one assignment
+    #   instead of loose kwargs (where dropping one would silently
     #   route the chat elsewhere or raise
     #   +RubyLLM::ModelNotFoundError+). If +transport.model+ is
     #   +nil+, it's filled in from +RubyLLM.config.default_model+.
@@ -326,7 +100,10 @@ module Pikuri
     # @param step_limit [Control::StepLimit, nil] step budget
     #   control. When set, {Control::StepLimit#tick!} fires on
     #   every +before_tool_call+ and {Control::StepLimit#reset!}
-    #   at the start of each turn. +nil+ means "no step budget"
+    #   at the start of each turn; the budget's
+    #   {Control::StepLimit#on_exhausted} policy decides what
+    #   {#run_loop} does when it trips (see "Step-exhaustion
+    #   policy" in the class header). +nil+ means "no step budget"
     #   (the agent can loop indefinitely).
     # @param cancellable [Control::Cancellable, nil] cancellation
     #   control. When set, {Control::Cancellable#check!} fires on
@@ -339,18 +116,6 @@ module Pikuri
     #   every +after_tool_result+ and each item becomes a
     #   {Event::UserTurn} with +mid_loop: true+. +nil+ means
     #   "no mid-loop injection" (the bundled CLIs default).
-    # @param context_window [Integer, nil] explicit override for
-    #   the model's context-window cap. When set, it wins over
-    #   ruby_llm's reported value and the llama.cpp probe — see
-    #   {ContextWindowDetector} for precedence. Resolved cap is
-    #   emitted as an {Event::ContextCap} immediately after
-    #   construction.
-    # @param llama_probe_url [String, nil] llama.cpp +/props+ URL
-    #   used as the third detection source. Only consulted when
-    #   neither +context_window+ nor ruby_llm's reported value is
-    #   set. Typically derived by +bin/pikuri-chat+ from its
-    #   configured +openai_api_base+; leave +nil+ when the
-    #   configured server is anything other than llama.cpp.
     # @param id [String] unique identifier for this agent. Empty
     #   for the main agent; sub-agents get persona-rooted ids
     #   like +"researcher 0"+, +"researcher 1"+, +"file_miner 0"+, ...
@@ -362,8 +127,8 @@ module Pikuri
     #   the codebase for the persona-name load (the value the LLM
     #   picks in the +agent+ tool's +name:+ argument).
     # @param streaming [Boolean] opt into chunk-level streaming.
-    #   When +true+, {#run_loop} passes the block returned by
-    #   {.streaming_block} to +Chat#ask+, and ruby_llm requests
+    #   When +true+, {#run_loop} passes a per-chunk block to
+    #   +Chat#complete+, and ruby_llm requests
     #   SSE responses from the provider — chunks are normalized
     #   into {Event::ThinkingDelta} / {Event::AssistantDelta} on
     #   the listener stream as they arrive. When +false+ (the
@@ -386,7 +151,7 @@ module Pikuri
     # @return [Agent]
     def initialize(transport:, system_prompt:,
                    step_limit: nil, cancellable: nil, interloper: nil,
-                   context_window: nil, llama_probe_url: nil, id: '',
+                   id: '',
                    streaming: false,
                    &block)
       @transport = transport.model ? transport : transport.with(model: RubyLLM.config.default_model)
@@ -403,8 +168,6 @@ module Pikuri
       # Stashed for {#run_configure}, which runs the failure-prone
       # build phase below out of a separate method.
       @block = block
-      @context_window = context_window
-      @llama_probe_url = llama_probe_url
 
       # Register *before* the build phase so a mid-construction raise
       # is still recoverable: extensions arm their cleanup via
@@ -427,9 +190,6 @@ module Pikuri
       end
     end
 
-    # @return [RubyLLM::Chat] underlying chat; the extension seam
-    attr_reader :chat
-
     # @return [ChatTransport] the resolved transport bundle this
     #   agent was constructed with — same model id / provider /
     #   assume-model-exists flag passed to every +RubyLLM.chat+
@@ -473,10 +233,6 @@ module Pikuri
     #   each persona owns its own system prompt verbatim.
     attr_reader :system_prompt
 
-    # @return [ListenerList] the listener list attached to this
-    #   agent's chat
-    attr_reader :listeners
-
     # @return [Control::StepLimit, nil] the step-budget control
     #   this agent was constructed with, or +nil+ when none.
     attr_reader :step_limit
@@ -521,13 +277,15 @@ module Pikuri
     #   extensions).
     attr_reader :extensions
 
-    # @return [Integer, nil] context-window cap resolved by
-    #   {ContextWindowDetector} at construction time. +nil+ when
-    #   no source produced a value (custom local model with no
-    #   override and no reachable llama.cpp +/props+). Read by
-    #   extensions that spawn their own ruby_llm calls (notably
-    #   the +agent+ tool from +pikuri-subagents+, so spawned
-    #   sub-agents inherit the same cap without re-probing).
+    # @return [Integer, nil] resolved context-window cap — the
+    #   {ChatTransport#context_window} if one was given, else what
+    #   {ContextWindowDetector} probed. +nil+ when neither produced
+    #   a value (a non-llama server with no explicit cap). Re-resolved
+    #   on every model switch (see {#run_loop}'s +transport:+). Read by
+    #   extensions that spawn their own ruby_llm calls (notably the
+    #   +agent+ tool from +pikuri-subagents+, which hands a sub-agent
+    #   +parent.transport.with(context_window: this)+ so the resolved
+    #   cap rides along without a re-probe).
     attr_reader :context_window_cap
 
     # Final assistant message content for the most recent
@@ -552,11 +310,15 @@ module Pikuri
     # and any other observable output is the listeners'
     # responsibility.
     #
-    # If the +step_limit+ control trips during +ask+, the rescue
-    # branch emits an {Event::FallbackNotice} and runs
-    # {Synthesizer.run} on a fresh +RubyLLM::Chat+. The synth's
-    # answer is captured for {#last_assistant_content}; the
-    # exception does not bubble out.
+    # If the +step_limit+ control trips during completion, the
+    # rescue branch applies its {Control::StepLimit#on_exhausted}
+    # policy: +:raise+ re-raises the +Exceeded+ exception to the
+    # host (chat history intact — the next turn's +reset!+
+    # refreshes the budget, so "continue" just works);
+    # +:synthesize+ emits an {Event::FallbackNotice} and runs the
+    # {Synthesizer} prompt on a nested tools-free agent, capturing
+    # its answer for {#last_assistant_content}. See
+    # "Step-exhaustion policy" in the class header.
     #
     # If the +cancellable+ control trips during +ask+, the rescue
     # branch emits an {Event::Cancelled} and re-raises the
@@ -566,19 +328,45 @@ module Pikuri
     # Subsequent calls keep building on the same chat history, so
     # the model sees full multi-turn context.
     #
+    # == Switching models mid-conversation
+    #
+    # Passing a +transport:+ that differs from the current one
+    # switches the underlying chat to that model — via
+    # +Chat#with_model+, so the history and the registered
+    # callbacks survive — re-resolves the context-window cap, and
+    # emits an {Event::ModelSwitched} followed by a fresh
+    # {Event::ContextCap}. The switch is deliberately confined to
+    # the top of this method (a private +apply_transport!+) rather
+    # than exposed as a standalone setter: the chat is
+    # single-thread-confined, so doing it here serializes the swap
+    # with the turn on the loop's own thread — a background thread
+    # mutating +with_model+'s connection mid-completion would tear
+    # an in-flight stream. A +nil+ +transport:+ (the default) keeps
+    # the current model. The conversation is *not* re-baselined: a
+    # switch is the same conversation under a new model, so the
+    # message count and running context size carry over (the next
+    # turn's token report self-corrects to the new model's count).
+    #
     # @param user_message [String] the user's request for this
     #   turn; must not be +nil+, empty, or whitespace-only
+    # @param transport [ChatTransport, nil] when non-+nil+ and
+    #   structurally different from the current transport, switch to
+    #   it before running the turn (see above); +nil+ keeps the
+    #   current model
     # @raise [ArgumentError] if +user_message+ is +nil+, empty,
     #   or contains only whitespace — an empty turn would poison
     #   the chat history and burn a step budget on nothing
     # @raise [Control::Cancellable::Cancelled] if the registered
     #   {Control::Cancellable} was triggered during the turn;
     #   the listener stream sees an {Event::Cancelled} first
+    # @raise [Control::StepLimit::Exceeded] if the step budget
+    #   tripped and its policy is +:raise+ (the default)
     # @return [nil]
-    def run_loop(user_message:)
+    def run_loop(user_message:, transport: nil)
       raise ArgumentError, "user_message must not be blank, got #{user_message.inspect}" \
         if user_message.nil? || user_message.to_s.strip.empty?
 
+      apply_transport!(transport) if transport
       @synth_answer = nil
       @step_limit&.reset!
       @cancellable&.reset!
@@ -594,7 +382,7 @@ module Pikuri
       @listeners.emit(Event::UserTurn.new(content: user_message, mid_loop: false))
       dispatch_ext_on_user_message(user_message)
       if @streaming
-        @chat.complete(&self.class.streaming_block(listeners: @listeners, cancellable: @cancellable))
+        @chat.complete(&streaming_block)
       else
         @chat.complete
       end
@@ -603,42 +391,16 @@ module Pikuri
       @listeners.emit(Event::Cancelled.new)
       raise
     rescue Control::StepLimit::Exceeded => e
-      @listeners.emit(Event::FallbackNotice.new(
-                        reason: "agent exhausted #{e.max_steps} steps; synthesizing answer from gathered evidence"
-                      ))
+      raise unless @step_limit&.on_exhausted == :synthesize
 
-      # Synth runs under this agent's identity but on a fresh
-      # chat with a different system prompt, so it gets a
-      # distinct +_synthesizer+ suffix on the id — same +_+
-      # separator the sub-agent generator uses, so main becomes
-      # +"synthesizer"+ and a sub-agent +"researcher 0"+ becomes
-      # +"researcher 0_synthesizer"+. Any +TokenLog+ in the list
-      # tags the synth's prompt under that bracket so it's
-      # obvious from the log which turns were the rescue rather
-      # than the original loop.
-      synth_id = @id.empty? ? 'synthesizer' : "#{@id}_synthesizer"
-      synth_chat = RubyLLM.chat(**@transport.to_h)
-      # Defensive step limit on the synth: the synth has no
-      # tools so it should never trip +before_tool_call+, but
-      # guarding the budget anyway means a buggy provider that
-      # somehow returns a tool call doesn't loop forever.
-      synth_step_limit = @step_limit && Control::StepLimit.new(max: 1)
-      @synth_answer = Synthesizer.run(
-        chat: synth_chat,
-        parent_messages: @chat.messages,
-        user_message: user_message,
-        listeners: @listeners.for_sub_agent(id: synth_id),
-        step_limit: synth_step_limit,
-        cancellable: @cancellable,
-        streaming: @streaming
-      )
+      @synth_answer = Synthesizer.run_synthesizer(@extension_context, @chat.messages, user_message)
       nil
     end
 
     # Release agent-owned resources. Fires every handler registered
     # via {Configurator#on_close} (during the +Agent.new+ block) and
-    # {#on_close} (during {Extension#bind} or any post-construction
-    # call), in LIFO order — matches Ruby +ensure+-block semantics
+    # {ExtensionContext#on_close} (during {Extension#bind} or any
+    # later hook), in LIFO order — matches Ruby +ensure+-block semantics
     # so handlers registered later (which may depend on handlers
     # registered earlier) tear down first. Each handler runs inside
     # its own +rescue+; an exception is logged via
@@ -661,46 +423,6 @@ module Pikuri
       end
     end
 
-    # Register a handler called by {#close}. Symmetric to
-    # {Configurator#on_close} — same LIFO + per-handler-rescue +
-    # idempotent semantics — but available post-construction, so
-    # an {Extension}'s +bind(agent)+ can install per-agent cleanup
-    # that's keyed to this specific agent rather than the parent.
-    #
-    # @yield called with no arguments at close time
-    # @return [void]
-    def on_close(&blk)
-      raise ArgumentError, 'on_close requires a block' unless block_given?
-
-      @on_close_handlers << blk
-      nil
-    end
-
-    # Register a raw +RubyLLM::Tool+ subclass on this agent's
-    # underlying chat, bypassing the {Pikuri::Tool} strict-validation
-    # seam. Sole intended caller: {Mcp::Servers::Connect}, which uses
-    # this to lazy-add MCP-exposed tools after the LLM invokes
-    # +mcp_connect+ in a turn.
-    #
-    # The +internal_+ prefix is the warning: native pikuri tools
-    # should go through {Pikuri::Tool} so they get
-    # {Tool::Parameters} validation and the LLM-actionable
-    # +"Error: ..."+ contract. MCP tools deliberately don't — see
-    # IDEAS.md §"v1 implementation shape" / "MCP tools bypass
-    # +Pikuri::Tool+ entirely."
-    #
-    # The added tool does NOT enter +@tools+, only +@chat+'s tool
-    # list. Sub-agents (the +agent+ tool from +pikuri-subagents+)
-    # therefore cannot snapshot it — which is the whole point:
-    # activation is strictly per-agent, see IDEAS.md §"Per-agent
-    # activation, no propagation".
-    #
-    # @param ruby_llm_tool [Class] subclass of +RubyLLM::Tool+
-    # @return [void]
-    def internal_add_tool(ruby_llm_tool)
-      @chat.with_tool(ruby_llm_tool)
-    end
-
     # Short, single-line config dump suitable for a startup
     # banner or a debug print.
     #
@@ -748,42 +470,164 @@ module Pikuri
       end
       @extensions = configurator.extensions.dup
 
-      @chat = RubyLLM.chat(**@transport.to_h)
+      @chat = build_chat(@transport)
       @chat.with_instructions(@system_prompt)
       @tools.each { |t| @chat.with_tool(t.to_ruby_llm_tool) }
 
-      @context_window_cap = ContextWindowDetector.new(
-        override: @context_window,
-        ruby_llm_reported: @chat.model.context_window,
-        llama_probe_url: @llama_probe_url,
-        model_id: @chat.model.id
-      ).detect
+      # Wire @chat for pikuri's event stream and controls — the
+      # three message-level registered callbacks (+after_message+,
+      # +before_tool_call+, +after_tool_result+). The per-chunk
+      # streaming callback is separate because ruby_llm takes it as
+      # a block to +Chat#complete+ rather than a registered hook —
+      # see {#streaming_block}. Together with the +@listeners.emit+
+      # calls in {#run_loop} / {#dispatch_ext_on_user_message} this
+      # is the complete "which callback emits which event" map.
+      @chat.after_message do |msg|
+        emit_after_message(msg)
+      end
+      @chat.before_tool_call do |tc|
+        @listeners.emit(Event::ToolCall.new(name: tc.name, arguments: tc.arguments))
+        @step_limit&.tick!
+        @cancellable&.check!
+      end
+      @chat.after_tool_result do |result|
+        @listeners.emit(Event::ToolResult.new(content: result))
+        drain_interloper if @interloper
+      end
 
-      self.class.wire_chat(
-        @chat,
-        listeners: @listeners,
-        step_limit: @step_limit,
-        cancellable: @cancellable,
-        interloper: @interloper,
-        on_user_message: method(:dispatch_ext_on_user_message)
+      # Context-window cap: lets every listener that cares (notably
+      # TokenLog) pick the value off the stream before any Tokens
+      # event arrives. Re-fires on each model switch (see
+      # {#apply_transport!}).
+      detect_and_emit_context_cap!
+
+      # The runtime capability facade — constructed once, after the
+      # chat and listener list are final, and handed to every
+      # extension's #bind / #on_user_message. The ONLY object that
+      # grants emission / raw-tool-registration / close-handler
+      # capabilities; the Agent itself exposes no public path to
+      # them. See {ExtensionContext}.
+      @extension_context = ExtensionContext.new(
+        agent: self, chat: @chat, listeners: @listeners,
+        on_close_sink: @on_close_handlers
       )
 
-      # One-shot context-window cap: lets every listener that
-      # cares (notably TokenLog) pick the value off the stream
-      # before any Tokens event arrives.
+      # Bind sweep — each extension gets its chance to install
+      # per-agent state (dynamic tools via
+      # {ExtensionContext#add_raw_tool}, per-agent close hooks via
+      # {ExtensionContext#on_close}, domain-event wiring via
+      # {ExtensionContext#emit_event}, etc.) now that the chat is
+      # fully wired. See IDEAS.md §"Extension protocol design" for
+      # what #configure vs #bind are each for.
+      @extensions.each { |ext| ext.bind(@extension_context) }
+    end
+
+    # Resolve the context-window cap and announce it on the listener
+    # stream as an {Event::ContextCap}. The transport's explicit
+    # +context_window+ wins verbatim; otherwise {ContextWindowDetector}
+    # probes the server (yielding +nil+ for a non-llama one). Because the
+    # cap rides {ChatTransport}, a model switch resolves the *new*
+    # transport's cap — explicit caps don't bleed across models. Shared
+    # by {#run_configure} (construction) and {#apply_transport!} (each
+    # model switch) so "how the cap is resolved and emitted" lives in
+    # one place.
+    #
+    # @return [void]
+    def detect_and_emit_context_cap!
+      # Probe the server this transport actually targets — after a
+      # cross-server switch the chat's connection points at
+      # +@transport.api_base+, but the process-global config the
+      # detector defaults to still names the *old* server, so derive the
+      # base from the transport (falling back to the global base for a
+      # transport that rides it).
+      @context_window_cap = @transport.context_window ||
+                            ContextWindowDetector.detect(
+                              @transport,
+                              openai_base: @transport.api_base || RubyLLM.config.openai_api_base
+                            )
       @listeners.emit(Event::ContextCap.new(cap: @context_window_cap))
+    end
 
-      # Bind sweep — each extension gets its chance to install
-      # per-agent state (dynamic tools via #internal_add_tool,
-      # per-agent close hooks via #on_close, etc.) now that the
-      # chat is fully wired. See IDEAS.md §"Extension protocol
-      # design" for what #configure vs #bind are each for.
-      @extensions.each { |ext| ext.bind(self) }
+    # Switch the underlying chat to +transport+ when it differs from
+    # the current one. Called only from the top of {#run_loop} — see
+    # that method's "Switching models mid-conversation" section for why
+    # the swap is confined to the loop's own thread.
+    #
+    # Mirrors {#initialize}'s +nil+-model fill before the structural
+    # comparison, so a +transport+ that defers its model to the default
+    # doesn't read as "different" and switch spuriously.
+    # +Chat#with_model+ swaps only the model / provider / connection,
+    # leaving +@chat+'s message history and registered callbacks
+    # intact, so the conversation continues seamlessly under the new
+    # model. Emits {Event::ModelSwitched} (the narration: old → new
+    # transport, unformatted, for the chrome to present) then re-resolves
+    # and re-emits the cap via {#detect_and_emit_context_cap!}.
+    #
+    # == Cross-server switches
+    #
+    # +with_model+ alone re-resolves against the chat's *existing*
+    # connection, so it can only move between models on one server. When
+    # either the new or the old transport overrides the connection
+    # (+ChatTransport#connection_overrides?+ — a different +api_base+ /
+    # +api_key+), the swap first installs a fresh +RubyLLM::Context+ via
+    # +Chat#with_context+ (which re-points the connection and re-resolves
+    # the model in place, again preserving history + callbacks) before
+    # +with_model+ lands the new model id. The "old overrode" half of
+    # the guard handles switching *back* to a global-config model: the
+    # rebuilt context dups the process-global config, resetting the
+    # connection the previous override installed.
+    #
+    # @param transport [ChatTransport] the model to switch to
+    # @return [void]
+    def apply_transport!(transport)
+      filled = transport.model ? transport : transport.with(model: RubyLLM.config.default_model)
+      return if filled == @transport
+
+      old = @transport
+      @chat.with_context(build_context(filled)) if filled.connection_overrides? || old.connection_overrides?
+      @chat.with_model(filled.model, provider: filled.provider, assume_exists: filled.assume_model_exists)
+      @transport = filled
+      @listeners.emit(Event::ModelSwitched.new(from: old, to: filled))
+      detect_and_emit_context_cap!
+    end
+
+    # Build the chat for +transport+: through a dedicated
+    # +RubyLLM::Context+ when it overrides the connection, else through
+    # the process-global +RubyLLM.chat+ (which the construction-time
+    # path has always used).
+    #
+    # @param transport [ChatTransport]
+    # @return [RubyLLM::Chat]
+    def build_chat(transport)
+      if transport.connection_overrides?
+        build_context(transport).chat(**transport.chat_kwargs)
+      else
+        RubyLLM.chat(**transport.chat_kwargs)
+      end
+    end
+
+    # A +RubyLLM::Context+ carrying +transport+'s connection overrides
+    # mapped onto the provider's ruby_llm config slots
+    # (+#{provider}_api_base+ / +#{provider}_api_key+). +RubyLLM.context+
+    # dups the process-global config, so an absent override inherits the
+    # global value; a transport with no overrides yields a plain dup
+    # (used by {#apply_transport!} to reset a prior override). The
+    # +ChatTransport+ guarantees a non-+nil+ +provider+ whenever an
+    # override is set, so the slot name is always resolvable.
+    #
+    # @param transport [ChatTransport]
+    # @return [RubyLLM::Context]
+    def build_context(transport)
+      slug = transport.provider
+      RubyLLM.context do |c|
+        c.public_send("#{slug}_api_base=", transport.api_base) unless transport.api_base.nil?
+        c.public_send("#{slug}_api_key=", transport.api_key) unless transport.api_key.nil?
+      end
     end
 
     # Fire the per-turn {Extension#on_user_message} hook on every
     # extension that defines it, appending any returned
-    # +<memory-context>+ block to the chat as a +role: :system+
+    # String text block to the chat as a +role: :system+
     # message right after the user turn it annotates (callers append
     # the +:user+ message first; this runs last). The system role is
     # load-bearing — it tags the block as recalled reference (not new
@@ -797,9 +641,8 @@ module Pikuri
     #
     # Private and the single place the chat log grows by a memory
     # block — keeps "what mutates the log, when" one grep in this
-    # file. Fired from {#run_loop} (initial turn) and, via the
-    # +on_user_message:+ proc threaded into {.wire_chat}, from
-    # {.drain_interloper} (mid-loop interlopers). Called on every
+    # file. Fired from {#run_loop} (initial turn) and from
+    # {#drain_interloper} (mid-loop interlopers). Called on every
     # extension unconditionally — same as {Extension#configure} /
     # {Extension#bind}: the hook is part of the protocol and the
     # {Extension} module supplies a no-op default, so any extension
@@ -811,7 +654,7 @@ module Pikuri
     # @return [void]
     def dispatch_ext_on_user_message(content)
       @extensions.each do |ext|
-        message = ext.on_user_message(self, content)
+        message = ext.on_user_message(@extension_context, content)
         next unless message.is_a?(String) && !message.strip.empty?
 
         block = message.strip
@@ -820,5 +663,117 @@ module Pikuri
       end
       nil
     end
+
+    # Build the per-chunk streaming block passed to +Chat#complete+.
+    # Each invocation of the returned proc converts one
+    # +RubyLLM::Chunk+ into zero, one, or two delta events
+    # ({Event::ThinkingDelta} / {Event::AssistantDelta}) on
+    # +@listeners+. Tool-call chunks are intentionally ignored —
+    # partial JSON has no useful rendering; the assembled
+    # +tool_calls+ surface through {Event::ToolCall} once the
+    # message completes.
+    #
+    # == Cancellation polling
+    #
+    # When +@cancellable+ is non-nil, {Control::Cancellable#check!}
+    # fires *before* each chunk's emit. The +before_tool_call+
+    # wiring in {#run_configure} only fires when the model requests a
+    # tool, which leaves a no-tool turn (e.g. a plain greeting)
+    # with zero cancellation points — Ctrl+C trips the flag but
+    # nothing reads it. Polling on every streamed chunk closes
+    # that gap: an in-flight Cancellation+check! raises on the
+    # next chunk delivered after the flag flips, the exception
+    # propagates out through ruby_llm's streaming path
+    # (+Chat#complete+ doesn't rescue), and {#run_loop} catches it,
+    # emits {Event::Cancelled}, and re-raises. The pre-emit
+    # ordering is deliberate: a chunk that arrives after a cancel
+    # request shouldn't render — the user has said stop.
+    #
+    # @return [Proc] a +-> (chunk) { ... }+ proc suitable for
+    #   passing to +Chat#complete+ with +&+
+    def streaming_block
+      ->(chunk) {
+        @cancellable&.check!
+        emit_chunk(chunk)
+      }
+    end
+
+    # Normalize a +RubyLLM::Chat+ +after_message+ payload into
+    # zero, one, or two {Event} variants (+Thinking+ and/or
+    # +Assistant+) plus one {Event::Tokens} for the usage block.
+    # Empty thinking / empty content are filtered here so
+    # listeners never see vacuous events. Non-assistant roles
+    # (e.g. tool-role messages echoed back through
+    # +after_message+) are skipped entirely.
+    #
+    # +msg+ is a +RubyLLM::Message+. Beyond +role+, +content+,
+    # +thinking+, and the +*_tokens+ accessors used here, it also
+    # carries +msg.tool_calls+ on assistant turns that requested
+    # one and +msg.raw+ for the unparsed provider payload.
+    #
+    # @param msg [RubyLLM::Message]
+    # @return [void]
+    def emit_after_message(msg)
+      return unless msg.role == :assistant
+
+      text = msg.thinking&.text
+      @listeners.emit(Event::Thinking.new(content: text)) if text && !text.empty?
+
+      content = msg.content
+      @listeners.emit(Event::Assistant.new(content: content)) if content.is_a?(String) && !content.empty?
+
+      @listeners.emit(Event::Tokens.new(
+                        input: msg.input_tokens,
+                        output: msg.output_tokens,
+                        cached: msg.cached_tokens,
+                        cache_creation: msg.cache_creation_tokens,
+                        thinking: msg.thinking_tokens,
+                        model_id: msg.model_id
+                      ))
+    end
+
+    # Normalize a +RubyLLM::Chunk+ from a streaming completion
+    # into zero, one, or two delta events
+    # ({Event::ThinkingDelta} / {Event::AssistantDelta}). Empty
+    # +thinking.text+ and empty +content+ are filtered here so
+    # listeners never see vacuous fragments. Tool-call deltas are
+    # intentionally skipped — see {#streaming_block}.
+    #
+    # +chunk+ is a +RubyLLM::Chunk+ (subclass of +RubyLLM::Message+),
+    # so the same +.thinking+ / +.content+ accessors used in
+    # {#emit_after_message} apply.
+    #
+    # @param chunk [RubyLLM::Chunk]
+    # @return [void]
+    def emit_chunk(chunk)
+      thinking = chunk.thinking&.text
+      @listeners.emit(Event::ThinkingDelta.new(content: thinking)) if thinking && !thinking.empty?
+
+      content = chunk.content
+      @listeners.emit(Event::AssistantDelta.new(content: content)) if content.is_a?(String) && !content.empty?
+    end
+
+    # Drain the interloper queue: for each pending item, append a
+    # +role: :user+ message to the chat history so the next
+    # round-trip sees it, emit an {Event::UserTurn} with
+    # +mid_loop: true+ to the listener stream so renderers see the
+    # injection, then run the per-turn {Extension#on_user_message}
+    # dispatch (so mid-loop injections are prefetched + recorded
+    # exactly like initial turns).
+    #
+    # The dispatch runs *after* the +:user+ append so any
+    # +<memory-context>+ it injects lands as a +:system+ message
+    # right behind the user turn it annotates — the same
+    # append-at-the-tail ordering {#run_loop} produces for initial
+    # turns.
+    #
+    # @return [void]
+    def drain_interloper
+      @interloper.drain!.each do |content|
+        @chat.add_message(role: :user, content: content)
+        @listeners.emit(Event::UserTurn.new(content: content, mid_loop: true))
+        dispatch_ext_on_user_message(content)
+      end
+    end
   end
 end