pikuri 0.0.1 → 0.0.3

data/lib/pikuri/agent/context_window_detector.rb

@@ -1,101 +0,0 @@
-# frozen_string_literal: true
-
-require 'faraday'
-require 'json'
-
-module Pikuri
-  class Agent
-    # Resolves the model's context-window cap from three sources, in order:
-    # an explicit override, the value ruby_llm reports for the model, or a
-    # llama.cpp +/props+ probe. Returns +nil+ if none of those produce a
-    # value.
-    #
-    # Used by {Agent#initialize} at construction time to feed
-    # {Listener::TokenLog} a cap it can render alongside the running
-    # context size (so the +ctx=12.2k/32.0k+ line tells the operator how
-    # close the conversation is to the limit).
-    #
-    # == Precedence
-    #
-    # 1. +override+ — the +Agent.new(context_window:)+ kwarg. Wins over
-    #    everything; an explicit value is the operator's statement of
-    #    truth.
-    # 2. +ruby_llm_reported+ — +RubyLLM::Model::Info#context_window+ from
-    #    {Agent#chat}'s resolved model. Populated for models in ruby_llm's
-    #    bundled registry (OpenAI, Anthropic, Gemini, …); +nil+ for custom
-    #    local model ids that fall through to +Model::Info.default+.
-    # 3. +llama_probe_url+ — HTTP GET against llama.cpp's non-standard
-    #    +/props+ endpoint. The server exposes the launched +n_ctx+ at
-    #    +default_generation_settings.n_ctx+ there. Probed only when the
-    #    first two are +nil+. Provider-specific to llama.cpp; the caller
-    #    (typically +bin/pikuri-chat+) derives the right URL from its configured
-    #    base.
-    #
-    # == Failure handling
-    #
-    # The probe is best-effort. HTTP error, timeout, non-JSON body, or a
-    # missing/invalid +n_ctx+ field all return +nil+ and log one +warn+
-    # line via +Pikuri.logger_for('ContextWindowDetector')+. This is the
-    # CLAUDE.md "secondary to the loop" carve-out — a wedged or
-    # non-llama.cpp server should not abort agent construction over a
-    # cosmetic readout.
-    class ContextWindowDetector
-      LOGGER = Pikuri.logger_for('ContextWindowDetector')
-
-      # Probe timeouts in seconds. Short on purpose: this runs synchronously
-      # during +Agent.new+ and a wedged server should not stall startup
-      # noticeably.
-      OPEN_TIMEOUT = 2
-      READ_TIMEOUT = 2
-
-      # @param override [Integer, nil] explicit cap from the caller; wins if
-      #   non-+nil+
-      # @param ruby_llm_reported [Integer, nil] value off
-      #   +RubyLLM::Chat#model.context_window+
-      # @param llama_probe_url [String, nil] full URL to llama.cpp +/props+;
-      #   +nil+ or empty string skips the probe
-      def initialize(override:, ruby_llm_reported:, llama_probe_url:)
-        @override = override
-        @ruby_llm_reported = ruby_llm_reported
-        @llama_probe_url = llama_probe_url
-      end
-
-      # @return [Integer, nil] resolved cap, or +nil+ if no source produced
-      #   one
-      def detect
-        return @override if @override
-        return @ruby_llm_reported if @ruby_llm_reported
-        return nil if @llama_probe_url.nil? || @llama_probe_url.empty?
-
-        probe_llama_cpp
-      end
-
-      private
-
-      def probe_llama_cpp
-        response = Faraday.new(
-          request: { open_timeout: OPEN_TIMEOUT, timeout: READ_TIMEOUT }
-        ).get(@llama_probe_url) do |req|
-          req.headers['Accept'] = 'application/json'
-        end
-
-        return warn_and_nil("HTTP #{response.status} from #{@llama_probe_url}") unless response.status == 200
-
-        data = JSON.parse(response.body)
-        n_ctx = data.dig('default_generation_settings', 'n_ctx')
-        return n_ctx if n_ctx.is_a?(Integer) && n_ctx.positive?
-
-        warn_and_nil(
-          "no positive integer at default_generation_settings.n_ctx in #{@llama_probe_url} response"
-        )
-      rescue Faraday::Error, JSON::ParserError => e
-        warn_and_nil("#{e.class.name.split('::').last}: #{e.message}")
-      end
-
-      def warn_and_nil(reason)
-        LOGGER.warn("llama.cpp /props probe failed: #{reason}")
-        nil
-      end
-    end
-  end
-end

data/lib/pikuri/agent/listener/in_memory_message_list.rb

@@ -1,33 +0,0 @@
-# frozen_string_literal: true
-
-module Pikuri
-  class Agent
-    module Listener
-      # Recording listener that appends every {Message} the agent emits
-      # to an in-memory list. Used by specs to assert on emissions
-      # without parsing stdout, and as the rough shape a future
-      # structured consumer (web sink, telemetry pipe) would take.
-      class InMemoryMessageList < MessageListener
-        # @return [Array<Agent::Message>] every message the listener has
-        #   seen, in order; never nil
-        attr_reader :events
-
-        def initialize
-          super
-          @events = []
-        end
-
-        # @param message [Agent::Message]
-        # @return [void]
-        def on_message(message)
-          @events << message
-        end
-
-        # @return [String] short label for {Agent#to_s}
-        def to_s
-          'InMemoryMessageList'
-        end
-      end
-    end
-  end
-end

data/lib/pikuri/agent/listener/message_listener.rb

@@ -1,93 +0,0 @@
-# frozen_string_literal: true
-
-module Pikuri
-  class Agent
-    module Listener
-      # Abstract base for listeners that consume the +Agent+'s normalized
-      # event stream as {Message} value objects. Concrete subclasses
-      # override the single {#on_message} hook and pattern-match on the
-      # variant.
-      #
-      # +#attach(chat)+ wires the three +RubyLLM::Chat+ callbacks once,
-      # builds the appropriate {Message} variant from each payload, and
-      # forwards it into {#on_message}. Empty +thinking+ and +assistant+
-      # content is filtered here so subclasses never receive vacuous
-      # events.
-      #
-      # Provider-reported token usage is a separate channel: every
-      # assistant +after_message+ event (including pure tool-call turns
-      # with empty content) emits an {Agent::Tokens} through {#on_tokens}
-      # so listeners that track context-window growth see every round
-      # trip. {Message::User} is not a +Chat+ event — {Agent#run_loop}
-      # constructs it directly and calls {#on_message}.
-      class MessageListener
-        # Wire the receiver into +chat+'s callback API. After this
-        # returns, +chat+'s loop will build {Message} variants from each
-        # event and forward them into {#on_message} on this instance.
-        #
-        # @param chat [RubyLLM::Chat]
-        # @return [void]
-        def attach(chat)
-          chat.after_message     { |msg| dispatch_chat_message(msg) }
-          chat.before_tool_call  { |tc|  on_message(Message::ToolCall.new(name: tc.name, arguments: tc.arguments)) }
-          chat.after_tool_result { |r|   on_message(Message::ToolResult.new(content: r)) }
-        end
-
-        # Single entry point for every event in the normalized stream.
-        # Concrete subclasses override this and dispatch on the variant
-        # (typically with a +case msg in Message::X(...)+ pattern). The
-        # default implementation is a no-op so listeners that only care
-        # about a subset can pattern-match and let everything else fall
-        # through.
-        #
-        # @param message [Message::User, Message::Thinking,
-        #   Message::Assistant, Message::ToolCall, Message::ToolResult,
-        #   Message::FallbackNotice]
-        # @return [void]
-        def on_message(message); end
-
-        # Hook for provider-reported token usage. Fires once per
-        # assistant turn (including pure tool-call turns with empty
-        # content) from {#dispatch_chat_message}. Default is a no-op so
-        # listeners that don't care — {Terminal}, {InMemoryMessageList} —
-        # ignore it; {TokenLog} overrides this to log and accumulate.
-        #
-        # @param tokens [Agent::Tokens]
-        # @return [void]
-        def on_tokens(tokens); end
-
-        private
-
-        # Normalizes a +RubyLLM::Chat+ +after_message+ payload into the
-        # appropriate {Message} variants for {#on_message}, plus an
-        # {Agent::Tokens} for {#on_tokens}.
-        #
-        # +msg+ is a +RubyLLM::Message+. Beyond +role+, +content+,
-        # +thinking+, and +tokens+ 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 dispatch_chat_message(msg)
-          return unless msg.role == :assistant
-
-          text = msg.thinking&.text
-          on_message(Message::Thinking.new(content: text)) if text && !text.empty?
-
-          content = msg.content
-          on_message(Message::Assistant.new(content: content)) if content.is_a?(String) && !content.empty?
-
-          on_tokens(Agent::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
-      end
-    end
-  end
-end

data/lib/pikuri/agent/listener/step_limit.rb

@@ -1,97 +0,0 @@
-# frozen_string_literal: true
-
-module Pikuri
-  class Agent
-    module Listener
-      # Standalone listener that caps the number of tool calls per
-      # +Agent#run_loop+ invocation. Not a +MessageListener+ — its job is
-      # control flow (raising on overrun), not surfacing events.
-      #
-      # ruby_llm has no built-in step budget, so this listener counts
-      # +before_tool_call+ events on the underlying +RubyLLM::Chat+ and
-      # raises {Exceeded} once the cap is crossed. +Agent#run_loop+
-      # forwards that out of +Chat#ask+; the step-exhaustion synthesizer
-      # rescues it to salvage the run.
-      class StepLimit
-        # Raised by the +before_tool_call+ callback once tool-call count
-        # exceeds +max+. Carries the budget that was tripped so rescue
-        # clauses can include it in user-facing messages.
-        class Exceeded < StandardError
-          # @return [Integer]
-          attr_reader :max_steps
-
-          # @param max_steps [Integer]
-          def initialize(max_steps)
-            @max_steps = max_steps
-            super("Agent loop exceeded #{max_steps} steps")
-          end
-        end
-
-        # @param max [Integer] hard cap on tool-call rounds; must be
-        #   positive
-        # @raise [ArgumentError] if +max+ is zero or negative
-        def initialize(max:)
-          raise ArgumentError, "max must be positive, got #{max}" if max <= 0
-
-          @max = max
-          @step = 0
-        end
-
-        # Wire the receiver into +chat+'s +before_tool_call+ callback. The
-        # callback increments the counter on every tool call and raises
-        # {Exceeded} once it crosses the configured cap.
-        #
-        # @param chat [RubyLLM::Chat]
-        # @return [void]
-        def attach(chat)
-          chat.before_tool_call do |_|
-            @step += 1
-            raise Exceeded, @max if @step > @max
-          end
-        end
-
-        # Listener protocol hook. Resets the step counter on
-        # {Agent::Message::User} — the User message marks the start of a
-        # new turn by definition, so it doubles as the lifecycle signal.
-        # Every other variant is a no-op.
-        #
-        # @param message [Agent::Message]
-        # @return [void]
-        def on_message(message)
-          case message
-          in Agent::Message::User
-            @step = 0
-          else
-            # no-op
-          end
-        end
-
-        # Sub-agent variant: a fresh +StepLimit+ at the caller-supplied
-        # +max_steps:+, or — when the key is absent from the forwarded
-        # params — at the receiver's own cap. The mutable counter and
-        # the +before_tool_call+ registration are per-chat, so the
-        # parent's instance cannot govern a sub-agent's chat; every
-        # sub-agent needs its own.
-        #
-        # The +max_steps+ default to +@max+ makes the method safe to
-        # call with no arguments, which keeps
-        # {Agent::ListenerList#for_sub_agent} usable as a bare
-        # +.for_sub_agent+ — a caller that doesn't care about the cap
-        # gets a sensible inheritance instead of an exception.
-        #
-        # @param max_steps [Integer] positive step cap for the sub-agent;
-        #   defaults to the receiver's current cap
-        # @return [StepLimit]
-        # @raise [ArgumentError] if +max_steps+ is non-positive
-        def for_sub_agent(max_steps: @max, **)
-          self.class.new(max: max_steps)
-        end
-
-        # @return [String] short config dump for {Agent#to_s}
-        def to_s
-          "StepLimit(max=#{@max})"
-        end
-      end
-    end
-  end
-end

data/lib/pikuri/agent/listener/terminal.rb

@@ -1,137 +0,0 @@
-# frozen_string_literal: true
-
-require 'rainbow'
-require 'tty-markdown'
-
-module Pikuri
-  class Agent
-    module Listener
-      # Terminal renderer for the normalized event stream: dim grey
-      # reasoning, Markdown-rendered assistant content, cyan tool-call
-      # and tool-result lines. {Message::User} is intentionally silent —
-      # the user typed it, so re-rendering it adds nothing on the CLI.
-      # This is the default listener attached by {Agent#initialize}.
-      #
-      # Optionally prepends a fixed number of leading spaces to every
-      # rendered line via the +padding:+ kwarg. Sub-agents get a fresh
-      # padded instance through this listener's {#for_sub_agent} hook
-      # (dispatched by {Agent::ListenerList#for_sub_agent}) so their
-      # output is visually indented under the parent's stream.
-      class Terminal < MessageListener
-        LOGGER = Pikuri.logger_for('Terminal')
-
-        # Cap, in characters, applied to tool-result content rendered to
-        # the terminal. Anything longer is truncated with a marker that
-        # reports the original byte size so the reader can tell a 200-
-        # char return apart from a 50KB HTML dump.
-        MAX_TOOL_RESULT_CHARS = 200
-
-        # Padding applied to a sub-agent's rendered stream. Absolute, not
-        # additive — sub-agent recursion is blocked (the sub-agent's tool
-        # set excludes +sub_agent+ itself), so the depth never exceeds 1
-        # in practice and a fixed indent reads cleanly.
-        SUB_AGENT_PADDING = 2
-
-        # @param padding [Integer] non-negative number of leading spaces
-        #   prepended to every rendered line. Defaults to 0; sub-agents
-        #   get a fresh instance with {SUB_AGENT_PADDING} via
-        #   {#for_sub_agent}.
-        def initialize(padding: 0)
-          super()
-          @padding = padding
-        end
-
-        # @return [Integer] current padding width; exposed so callers can
-        #   introspect it (and so tests can assert it).
-        attr_reader :padding
-
-        # Sub-agent variant: a fresh +Terminal+ at {SUB_AGENT_PADDING} so
-        # sub-agent output is visually indented under the parent's
-        # stream. Called by {Agent::ListenerList#for_sub_agent}; ignores
-        # any params it's handed (Terminal has no caller-provided knobs).
-        #
-        # @return [Terminal]
-        def for_sub_agent(**)
-          self.class.new(padding: SUB_AGENT_PADDING)
-        end
-
-        # @param message [Agent::Message] one of the {Agent::Message}
-        #   variants; see {Agent::Message} for the full list
-        # @return [void]
-        def on_message(message)
-          case message
-          in Message::Thinking(content:)
-            puts indent(Rainbow(content).color(85, 85, 85))
-          in Message::Assistant(content:)
-            puts indent(render_markdown(content))
-          in Message::ToolCall(name:, arguments:)
-            args = arguments.map { |k, v| "#{k}=#{v.inspect}" }.join(', ')
-            puts indent(Rainbow("→ #{name}(#{args})").cyan)
-          in Message::ToolResult(content:)
-            puts indent(Rainbow("= #{truncate_tool_result(content)}").cyan)
-          in Message::FallbackNotice(reason:)
-            puts indent(Rainbow("! #{reason}").yellow)
-          in Message::User
-            # No-op: the user just typed this; echoing it back is noise.
-          end
-        end
-
-        # @return [String] short label for {Agent#to_s}; includes the
-        #   padding suffix only when non-zero so the default form stays
-        #   +"Terminal"+
-        def to_s
-          @padding.zero? ? 'Terminal' : "Terminal(padding=#{@padding})"
-        end
-
-        private
-
-        # Prepend +@padding+ spaces to every line of +text+. Splits on
-        # +each_line+ rather than a +gsub+ trick so a trailing newline
-        # in the input doesn't produce a stray padded blank line at the
-        # end — +puts+ adds the final newline if missing, same as
-        # before.
-        #
-        # @param text [String]
-        # @return [String]
-        def indent(text)
-          return text if @padding.zero?
-
-          prefix = ' ' * @padding
-          text.to_s.each_line.map { |line| prefix + line }.join
-        end
-
-        # Render assistant Markdown for the terminal, degrading to the
-        # raw string when the renderer raises. tty-markdown / strings
-        # have known bugs around ANSI inside tables (e.g.
-        # +Strings::Wrap.insert_ansi+ raising +IndexError+); we'd rather
-        # show ugly Markdown than abort an in-flight conversation.
-        #
-        # @param content [String] assistant Markdown
-        # @return [String] rendered ANSI text, or +content+ unchanged on
-        #   render failure
-        def render_markdown(content)
-          TTY::Markdown.parse(content)
-        rescue StandardError => e
-          LOGGER.warn("TTY::Markdown render failed (#{e.class}: #{e.message}); falling back to raw text")
-          content
-        end
-
-        # Flatten whitespace and cap to {MAX_TOOL_RESULT_CHARS}. The cap
-        # keeps multi-screen dumps (rendered HTML, PDF text) from
-        # drowning the terminal stream; the byte-count suffix on a
-        # truncated result distinguishes "tool returned exactly this"
-        # from "tool returned much more, you're seeing a slice."
-        #
-        # @param content [String] tool observation
-        # @return [String] single-line display form, possibly truncated
-        def truncate_tool_result(content)
-          original_bytes = content.to_s.bytesize
-          flattened = content.to_s.gsub(/\s+/, ' ').strip
-          return flattened if flattened.length <= MAX_TOOL_RESULT_CHARS
-
-          "#{flattened[0, MAX_TOOL_RESULT_CHARS]}… (#{original_bytes} bytes total)"
-        end
-      end
-    end
-  end
-end

data/lib/pikuri/agent/listener/token_log.rb

@@ -1,166 +0,0 @@
-# frozen_string_literal: true
-
-module Pikuri
-  class Agent
-    module Listener
-      # Logs the conversation's context-window consumption per assistant
-      # turn via +Pikuri.logger_for('Tokens')+. Overrides
-      # {MessageListener#on_tokens} and emits one +INFO+ line per turn.
-      #
-      # Existence rationale: catch context-window growth before the
-      # provider raises +RubyLLM::ContextLengthExceededError+. +ctx+ is
-      # the headline number — tokens consumed by the conversation
-      # *through* this turn: this turn's prompt plus its reply, both of
-      # which the model will re-process on the next turn. The +Δ+ field
-      # shows the climb between turns; +↑+ / +↓+ are the per-turn input
-      # / output sizes that drove it (matching Claude Code's convention:
-      # +↑+ is what's sent to the model, +↓+ is what comes back).
-      #
-      # == State and scope
-      #
-      # The latest snapshot lives on the instance, so one +TokenLog+ is
-      # per-conversation. Sub-agents (with their own +RubyLLM::Chat+) get
-      # a fresh instance via {#for_sub_agent} (dispatched by
-      # {ListenerList#for_sub_agent}) so their counts log against their
-      # own chat. The synthesizer rescue keeps using the parent's
-      # instance because synth turns extend the same conversation.
-      #
-      # == Log line shape
-      #
-      # Without a cap (no detection source — see {Agent::ContextWindowDetector}):
-      #
-      #   msg #1: ctx=6.8k  Δ+6.8k  ↑6.8k  ↓0.0k
-      #   msg #2: ctx=8.4k  Δ+1.6k  ↑1.0k  ↓0.6k
-      #
-      # With a cap of e.g. 32k (set by {Agent#initialize} after detection):
-      #
-      #   msg #1: ctx=6.8k/32.0k  Δ+6.8k  ↑6.8k  ↓0.0k
-      #
-      # When the owning {Agent} has a non-empty {Agent#name} (i.e. a
-      # sub-agent), the line is prefixed with +[name] +:
-      #
-      #   [sub_agent 0] msg #1: ctx=4.2k  Δ+4.2k  ↑4.2k  ↓0.0k
-      #
-      # +ctx+ is the snapshot (+input + cached + cache_creation + output+;
-      # see {Agent::Tokens}), optionally suffixed with +/<cap>+ when
-      # {#context_window_cap} is set so the operator can see how close
-      # the conversation is to the limit. Including +output+ makes this
-      # turn's reply visible immediately — leaving it out would hide a
-      # long reply in the +ctx=+ headline until the next turn pulled it
-      # in as cached prompt. +Δ+ is signed: +Δ+X+ for growth,
-      # +Δ-X+ for shrinkage (legitimate if ruby_llm ever prunes between
-      # turns). On the first message the baseline is implicitly zero, so
-      # +Δ+ equals +ctx+. Sizes are scaled by 1024 and shown with one
-      # decimal + a +k+ suffix.
-      class TokenLog < MessageListener
-        LOGGER = Pikuri.logger_for('Tokens')
-
-        # Tokens consumed by the conversation through the most recent
-        # turn — this turn's prompt plus its reply, which together make
-        # up the bulk of what the next turn will re-process. Equals
-        # +input + cached + cache_creation + output+ from the latest
-        # {#on_tokens}. Zero until the first {#on_tokens} fires.
-        #
-        # @return [Integer]
-        attr_reader :context_window_size
-
-        # Model's context-window cap, or +nil+ if no source could supply
-        # one (see {Agent::ContextWindowDetector}). When set, the +ctx=+
-        # field renders as +ctx=<used>/<cap>+ instead of just +ctx=<used>+.
-        # {Agent#initialize} pushes the detected value here via
-        # {ListenerList#context_window_cap=} once it has run the detector,
-        # so an instance constructed bare (e.g. by callers, in tests, or
-        # by {#for_sub_agent}) defaults to +nil+.
-        #
-        # @return [Integer, nil]
-        attr_accessor :context_window_cap
-
-        # @return [String] owning agent's identifier. Empty by default
-        #   (main agent); set by {#for_sub_agent} from the sub-agent's
-        #   generated name so the log lines can be prefixed with
-        #   +[<name>] +. Read-only — for a sub-agent's listener you get a
-        #   fresh instance via {#for_sub_agent}.
-        attr_reader :name
-
-        # @param name [String] agent identifier prepended to each log
-        #   line as +[<name>] + when non-empty. Defaults to +""+ for the
-        #   main agent.
-        def initialize(name: '')
-          super()
-          @name = name
-          @msg = 0
-          @context_window_size = 0
-          @context_window_cap = nil
-        end
-
-        # Sub-agent variant: a fresh +TokenLog+ with a zeroed snapshot so
-        # the sub-agent's context-window readings track its own
-        # +RubyLLM::Chat+ rather than continuing the parent's. Picks the
-        # sub-agent's +name:+ out of the forwarded params so its log
-        # lines carry the +[<name>] + prefix; defaults to +""+ when
-        # absent. The cap is left +nil+ here; the sub-agent's
-        # {Agent#initialize} pushes the resolved cap onto every
-        # +TokenLog+ in its list via {ListenerList#context_window_cap=}
-        # immediately after construction.
-        #
-        # @param name [String] sub-agent's identifier
-        # @return [TokenLog]
-        def for_sub_agent(name: '', **)
-          self.class.new(name: name)
-        end
-
-        # Update the context-window snapshot, then log one +INFO+ line.
-        #
-        # @param tokens [Agent::Tokens]
-        # @return [void]
-        def on_tokens(tokens)
-          @msg += 1
-          input_now          = tokens.input.to_i
-          cached_now         = tokens.cached.to_i
-          cache_creation_now = tokens.cache_creation.to_i
-
-          prev_ctx = @context_window_size
-          @context_window_size = input_now + cached_now + cache_creation_now + tokens.output.to_i
-          delta = @context_window_size - prev_ctx
-
-          LOGGER.info(format_line(input_now, tokens.output.to_i, delta))
-        end
-
-        # @return [String] short label for {Agent#to_s}, including the
-        #   current context-window size (with cap suffix when set)
-        def to_s
-          "TokenLog(ctx=#{format_ctx})"
-        end
-
-        private
-
-        def format_line(input, output, delta)
-          sign = delta.negative? ? '-' : '+'
-          prefix = @name.empty? ? '' : "[#{@name}] "
-          "#{prefix}msg ##{@msg}: ctx=#{format_ctx}  Δ#{sign}#{format_k(delta.abs)}  ↑#{format_k(input)}  ↓#{format_k(output)}"
-        end
-
-        # +<used>+ when no cap is set, +<used>/<cap>+ when one is. Shared
-        # between {#format_line} and {#to_s} so the headline reads the same
-        # in the log stream and in {Agent#to_s} banners.
-        def format_ctx
-          base = format_k(@context_window_size)
-          return base if @context_window_cap.nil?
-
-          "#{base}/#{format_k(@context_window_cap)}"
-        end
-
-        # Format a token count as a 1024-scaled +k+-suffixed string.
-        # +nil+ → +0.0k+; +0+ → +0.0k+; 12_453 → +12.2k+. Uniform
-        # format keeps lines easy to scan at the cost of looking odd
-        # for very small per-turn outputs (+↓0.0k+ on tool-call acks).
-        #
-        # @param n [Integer, nil]
-        # @return [String]
-        def format_k(n)
-          format('%.1fk', n.to_i / 1024.0)
-        end
-      end
-    end
-  end
-end