pwn 0.5.757 → 0.5.759

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: a9bcba2a6a58d7e3f7990a1dbbf0cc73a1234efe20bfd7dd0eae786ecc4c4c32
-  data.tar.gz: 252972a638225c5ddd3693839eb4fe7db2311529e3ceb2a32f69f0cc3056fd27
+  metadata.gz: 1ef7228011e10bd0573cab148f192e8ef963608a002d4e3d833fcd2e380fd973
+  data.tar.gz: 8f238e16e7d0d521deb4b4abdad45fed2d15fb82d92bcc8448b0877bc23b8783
 SHA512:
-  metadata.gz: e581b4c0dd6a4b1e05beccfade96bc635f48bf47c63168209b4900e6fb68a9d8e0b2394b4f6458a5aa96a9373248a0f96adb9ae0dcdeea2210f10f908bae630c
-  data.tar.gz: c87a11936f1cf59e2d6373c2d941ccf33f5aa4edac9e91fa09e1c9b5b871bcf2def227c10a6659131bc4279e05c7b6d85e28eeb23ff701d3ebadb0a3359ba129
+  metadata.gz: d6eb621a41250c5332e00f941c8b6b7444f796fcb30c009d774d7a0814f3c268809152275f2c63f7f7076450c75bbe662c0248ec6b21abb9730b44b097fdb139
+  data.tar.gz: 8ea184fdfa2018bc59f00aef8585d190907236a3d71df302c21db3eb025813111cd5a7bb289513c5a7f51a816f1c7e094f5b6c0bf420ff50d9a7a27a11ffbc65

data/Gemfile

@@ -68,7 +68,7 @@ gem 'ostruct', '0.6.3'
 gem 'packetfu', '2.0.0'
 gem 'packetgen', '4.1.1'
 gem 'pdf-reader', '2.16.0'
-gem 'pg', '1.6.3'
+gem 'pg', '1.7.0'
 gem 'pry', '0.16.0'
 gem 'pry-doc', '1.7.0'
 gem 'rake', '13.4.2'

data/lib/pwn/ai/agent/loop.rb

@@ -1923,12 +1923,16 @@ module PWN
 
           content = msg[:content]
           tool_calls = Array(msg[:tool_calls])
+          promoted = false
           # Local/thinking models (Ollama Qwen3, DeepSeek-R1, etc.) sometimes
           # return only :thinking with empty :content and no tool_calls. Promote
           # thinking so the agent does not print a blank final answer.
           if content.to_s.strip.empty? && tool_calls.empty?
             thinking = msg[:thinking].to_s
-            content = thinking unless thinking.strip.empty?
+            unless thinking.strip.empty?
+              content = thinking
+              promoted = true
+            end
           end
 
           out = {
@@ -1953,6 +1957,7 @@ module PWN
           out[:_native_content] = msg[:_native_content] if msg[:_native_content]
           thinking = msg[:thinking] || msg[:reasoning_content]
           out[:thinking] = thinking if thinking.to_s.strip != ''
+          out[:_thinking_promoted] = true if promoted
           # Local/abliterated models sometimes emit shell(...) as plain content
           # with empty tool_calls. Coerce registered call-shaped text into
           # structured tool_calls so Loop dispatches instead of FINAL-answering.
@@ -2297,6 +2302,17 @@ module PWN
           nil
         end
 
+        private_class_method def self.emit_thinking!(opts = {})
+          text = opts[:thinking].to_s.strip
+          Thread.current[:pwn_last_thinking] = text
+          Thread.current[:pwn_thinking_promoted] = opts[:promoted] ? true : false
+          return if text.empty?
+
+          opts[:on_tool]&.call('thinking', text, '')
+        rescue StandardError
+          nil
+        end
+
         # Pre-dispatch: one high-level brief for the whole upcoming tool
         # collection (pwn-ai → task is one-to-many with pwn-ai → <tool>).
         # Prefer opts[:tools] = [{name:, args:}, ...]; falls back to single name.
@@ -2304,18 +2320,13 @@ module PWN
           state = opts[:state]
           return nil unless state && defined?(TaskSummarizer) && TaskSummarizer.enabled?
 
-          thinking = opts[:thinking].to_s.strip
-          line = if thinking.empty?
-                   TaskSummarizer.about_to(
-                     name: opts[:name],
-                     args: opts[:args],
-                     state: state,
-                     request: opts[:request],
-                     tools: opts[:tools]
-                   )
-                 else
-                   thinking
-                 end
+          line = TaskSummarizer.about_to(
+            name: opts[:name],
+            args: opts[:args],
+            state: state,
+            request: opts[:request],
+            tools: opts[:tools]
+          )
           # about_to returns nil when the brief is a duplicate of the last one
           emit_task_summary(line: line, on_tool: opts[:on_tool]) if line
           line
@@ -3409,6 +3420,11 @@ module PWN
 
             calls = Array(msg[:tool_calls])
             text  = msg[:content].to_s
+            emit_thinking!(
+              thinking: msg[:thinking] || msg[:reasoning_content],
+              promoted: msg[:_thinking_promoted],
+              on_tool: on_tool
+            )
 
             # Belt-and-suspenders: plain-text shell(...) / tool forms from local
             # models under weak TEMPLATE {{ .Prompt }} become real tool_calls.
@@ -3525,7 +3541,6 @@ module PWN
                 }
               end,
               request: request,
-              thinking: msg[:thinking] || msg[:reasoning_content],
               on_tool: on_tool
             )
 

data/lib/pwn/ai/ollama.rb

@@ -436,6 +436,9 @@ module PWN
       # - function.arguments must be a Hash/Array object, not a JSON string
       #   (string args → HTTP 400 "can't find closing '}' symbol")
       # - assistant content nil + tool_calls → "" (Open WebUI form validation)
+      # - system/developer rows after the first message are merged into one
+      #   leading system message. Qwen-style templates raise HTTP 500
+      #   "System message must be at the beginning" otherwise.
       private_class_method def self.parse_tool_arguments_object(opts = {})
         raw = opts[:arguments]
         case raw
@@ -458,7 +461,7 @@ module PWN
       end
 
       private_class_method def self.normalize_messages_for_ollama(opts = {})
-        Array(opts[:messages]).filter_map do |m|
+        rows = Array(opts[:messages]).filter_map do |m|
           next unless m.is_a?(Hash)
 
           role = (m[:role] || m['role']).to_s
@@ -508,6 +511,71 @@ module PWN
 
           out
         end
+        lead_system(messages: rows, fold_into_user: opts[:fold_into_user])
+      end
+
+      # One system role is only safe when the server does not already prepend one.
+      # Models whose template raises "System message must be at the beginning"
+      # already carry a Modelfile SYSTEM prompt. Fold client system text into
+      # the first user message instead of sending a second system role.
+      private_class_method def self.lead_system(opts = {})
+        parts = []
+        rest = []
+        Array(opts[:messages]).each do |row|
+          next unless row.is_a?(Hash)
+
+          role = (row[:role] || row['role']).to_s
+          if %w[system developer].include?(role)
+            text = (row.key?(:content) ? row[:content] : row['content']).to_s
+            parts << text unless text.strip.empty?
+          else
+            rest << row
+          end
+        end
+        return rest if parts.empty?
+
+        text = parts.join("\n\n")
+        return [{ role: 'system', content: text }] + rest unless opts[:fold_into_user]
+
+        user = rest.find { |row| (row[:role] || row['role']).to_s == 'user' }
+        if user
+          body = (user.key?(:content) ? user[:content] : user['content']).to_s
+          user[:content] = "#{text}\n\n#{body}"
+          user.delete('content')
+          rest
+        else
+          [{ role: 'user', content: text }] + rest
+        end
+      end
+
+      private_class_method def self.strict_system_template?(opts = {})
+        name = opts[:model].to_s.strip
+        return false if name.empty?
+
+        @strict_system_cache ||= {}
+        return @strict_system_cache[name] if @strict_system_cache.key?(name)
+
+        @strict_system_cache[name] = template_rejects_later_system?(model: name)
+      rescue StandardError
+        false
+      end
+
+      private_class_method def self.template_rejects_later_system?(opts = {})
+        name = opts[:model].to_s
+        names = [name]
+        names << "#{name}:latest" unless name.include?(':')
+        names.each do |candidate|
+          raw = ollama_rest_call(
+            http_method: :post, rest_call: 'api/show',
+            http_body: { model: candidate }, timeout: 8, quiet: true, non_interactive: true
+          )
+          template = PWN::AI::ModelCatalog.parse_row(raw: raw)
+          text = template.is_a?(Hash) ? (template[:template] || template['template']).to_s : ''
+          return true if text.include?('System message must be at the beginning')
+        end
+        false
+      rescue StandardError
+        false
       end
 
       # Supported Method Parameters::
@@ -526,13 +594,16 @@ module PWN
       # NDJSON chunks back into a single response.
 
       public_class_method def self.chat_with_tools(opts = {})
-        engine   = PWN::Env[:ai][:ollama]
-        messages = normalize_messages_for_ollama(messages: opts[:messages])
-        raise 'ERROR: messages array is required' if messages.nil? || messages.empty?
-
+        engine = PWN::Env[:ai][:ollama]
         model = opts[:model] ||= engine[:model]
         raise 'ERROR: Model is required.  Call #get_models method for details' if model.nil?
 
+        messages = normalize_messages_for_ollama(
+          messages: opts[:messages],
+          fold_into_user: strict_system_template?(model: model)
+        )
+        raise 'ERROR: messages array is required' if messages.nil? || messages.empty?
+
         temp = opts[:temp].to_f
         temp = engine[:temp].to_f.nonzero? || 1 if temp.zero?
 
@@ -656,6 +727,10 @@ module PWN
         end
 
         http_body[:messages].push(user_role)
+        http_body[:messages] = lead_system(
+          messages: http_body[:messages],
+          fold_into_user: strict_system_template?(model: model)
+        )
 
         timeout = opts[:timeout]
         spinner = opts[:spinner]

data/lib/pwn/ai/open_web_ui.rb

@@ -468,6 +468,9 @@ module PWN
       # - function.arguments must be a Hash/Array object, not a JSON string
       #   (string args → HTTP 400 "can't find closing '}' symbol")
       # - assistant content nil + tool_calls → "" (Open WebUI form validation)
+      # - system/developer rows after the first message are merged into one
+      #   leading system message. Qwen-style templates raise HTTP 500
+      #   "System message must be at the beginning" otherwise.
       private_class_method def self.parse_tool_arguments_object(opts = {})
         raw = opts[:arguments]
         case raw
@@ -490,7 +493,7 @@ module PWN
       end
 
       private_class_method def self.normalize_messages_for_ollama(opts = {})
-        Array(opts[:messages]).filter_map do |m|
+        rows = Array(opts[:messages]).filter_map do |m|
           next unless m.is_a?(Hash)
 
           role = (m[:role] || m['role']).to_s
@@ -540,6 +543,71 @@ module PWN
 
           out
         end
+        lead_system(messages: rows, fold_into_user: opts[:fold_into_user])
+      end
+
+      # One system role is only safe when the server does not already prepend one.
+      # mini-mythos-mission-ready has a Modelfile SYSTEM prompt and a template that
+      # raises if any later message has role system. Fold those rows into the
+      # first user message instead of sending a second system role.
+      private_class_method def self.lead_system(opts = {})
+        parts = []
+        rest = []
+        Array(opts[:messages]).each do |row|
+          next unless row.is_a?(Hash)
+
+          role = (row[:role] || row['role']).to_s
+          if %w[system developer].include?(role)
+            text = (row.key?(:content) ? row[:content] : row['content']).to_s
+            parts << text unless text.strip.empty?
+          else
+            rest << row
+          end
+        end
+        return rest if parts.empty?
+
+        text = parts.join("\n\n")
+        return [{ role: 'system', content: text }] + rest unless opts[:fold_into_user]
+
+        user = rest.find { |row| (row[:role] || row['role']).to_s == 'user' }
+        if user
+          body = (user.key?(:content) ? user[:content] : user['content']).to_s
+          user[:content] = "#{text}\n\n#{body}"
+          user.delete('content')
+          rest
+        else
+          [{ role: 'user', content: text }] + rest
+        end
+      end
+
+      private_class_method def self.strict_system_template?(opts = {})
+        name = opts[:model].to_s.strip
+        return false if name.empty?
+
+        @strict_system_cache ||= {}
+        return @strict_system_cache[name] if @strict_system_cache.key?(name)
+
+        @strict_system_cache[name] = template_rejects_later_system?(model: name)
+      rescue StandardError
+        false
+      end
+
+      private_class_method def self.template_rejects_later_system?(opts = {})
+        name = opts[:model].to_s
+        names = [name]
+        names << "#{name}:latest" unless name.include?(':')
+        names.each do |candidate|
+          raw = openwebui_rest_call(
+            http_method: :post, rest_call: 'ollama/api/show',
+            http_body: { model: candidate }, timeout: 8, quiet: true, non_interactive: true
+          )
+          template = PWN::AI::ModelCatalog.parse_row(raw: raw)
+          text = template.is_a?(Hash) ? (template[:template] || template['template']).to_s : ''
+          return true if text.include?('System message must be at the beginning')
+        end
+        false
+      rescue StandardError
+        false
       end
 
       # Supported Method Parameters::
@@ -559,13 +627,16 @@ module PWN
       # Bare POST /api/chat is not an API route on stock Open WebUI (405).
 
       public_class_method def self.chat_with_tools(opts = {})
-        engine   = PWN::Env[:ai][:openwebui]
-        messages = normalize_messages_for_ollama(messages: opts[:messages])
-        raise 'ERROR: messages array is required' if messages.nil? || messages.empty?
-
+        engine = PWN::Env[:ai][:openwebui]
         model = opts[:model] ||= engine[:model]
         raise 'ERROR: Model is required.  Call #get_models method for details' unless real_config_value?(value: model)
 
+        messages = normalize_messages_for_ollama(
+          messages: opts[:messages],
+          fold_into_user: strict_system_template?(model: model)
+        )
+        raise 'ERROR: messages array is required' if messages.nil? || messages.empty?
+
         temp = opts[:temp].to_f
         temp = engine[:temp].to_f.nonzero? || 1 if temp.zero?
 
@@ -682,6 +753,10 @@ module PWN
         end
 
         http_body[:messages].push(user_role)
+        http_body[:messages] = lead_system(
+          messages: http_body[:messages],
+          fold_into_user: strict_system_template?(model: model)
+        )
 
         timeout = opts[:timeout]
         spinner = opts[:spinner]

data/lib/pwn/plugins/repl/ai_console.rb

@@ -266,7 +266,7 @@ module PWN
 
         # Own the terminal, event pump and the lifetime of exactly one request.
         class Console # rubocop:disable Metrics/ClassLength -- cohesive single-owner terminal state machine
-          COLORS = { operator: :operator, task: :task, tool: :tool, result: :result, assistant: :assistant, notice: :notice, warning: :warning }.freeze
+          COLORS = { operator: :operator, task: :task, thinking: :notice, tool: :tool, result: :result, assistant: :assistant, notice: :notice, warning: :warning }.freeze
           PALETTE = %w[cyan green yellow red white blue magenta black].freeze
 
           def initialize(pry:, input:, curses:, getch:)
@@ -480,6 +480,8 @@ module PWN
               on_tool = lambda do |name, args, result|
                 if name.to_s == 'task'
                   @events << [:task, args.to_s]
+                elsif name.to_s == 'thinking'
+                  @events << [:thinking, args.to_s]
                 else
                   @events << [:tool, "#{name}\n#{args.is_a?(String) ? args : args.inspect}"]
                   @events << [:result, result.to_s]
@@ -503,7 +505,8 @@ module PWN
                   enabled_toolsets: PWN::Env.dig(:ai, :agent, :toolsets),
                   on_tool: on_tool, steering: control, debug: @pry.config.pwn_ai_debug, debug_tee: @output
                 )
-                @events << [:assistant, final.to_s]
+                promoted = Thread.current[:pwn_thinking_promoted] && final.to_s.strip == Thread.current[:pwn_last_thinking].to_s.strip
+                @events << [:assistant, final.to_s] unless promoted
               end
             rescue Control::Stopped => e
               @events << [:warning, e.message]
@@ -1652,6 +1655,7 @@ module PWN
           puts "USAGE:
             # Ctrl+L clears only the session pane, including while busy.
             # Ctrl+T toggles the active pane between SESSION and MISSION CONTROL. Ctrl+S opens swarm.
+            # Thinking from a model that returns it is a THINKING row in the session pane, above the answer.
             # Up/Down select in menus, recall ~/.pwn/pwn_history in the mission pane, and scroll the session pane when it is active.
             # Search: Enter accepts without sending; Esc restores the draft.
             # Launch the single-owner curses console; non-terminals return unavailable.

data/lib/pwn/plugins/repl/ai_swarm.rb

@@ -279,6 +279,10 @@ module PWN
 
           def tool_callback(job, opts)
             lambda do |name, args, result|
+              if name.to_s == 'thinking'
+                opts[:on_event]&.call(:thinking, args.to_s)
+                return
+              end
               update(job, job[:state], last_tool: name)
               opts[:on_event]&.call(:tool, "#{job[:agent]} [#{job[:id]}] #{name}\n#{args.inspect}")
               opts[:on_event]&.call(:result, result.to_s)

data/lib/pwn/plugins/repl.rb

@@ -561,6 +561,19 @@ module PWN
                     mirror.call("#{header}\n#{body_out}")
                     next
                   end
+                  if name.to_s == 'thinking'
+                    body = args.is_a?(String) ? args.to_s : args.inspect
+                    timestamp = Time.now.strftime('%Y-%m-%d %H:%M:%S%z')
+                    header = "[ #{timestamp} → pwn-ai → thinking ]"
+                    puts "\e[36m#{header}\e[0m"
+                    body_out = +''
+                    body.to_s.each_line do |ln|
+                      puts "\e[36m  #{ln.rstrip}\e[0m"
+                      body_out << "  #{ln.rstrip}\n"
+                    end
+                    mirror.call("#{header}\n#{body_out}")
+                    next
+                  end
 
                   argv = args.is_a?(String) ? args.to_s : args.inspect
                   timestamp = Time.now.strftime('%Y-%m-%d %H:%M:%S%z')
@@ -592,9 +605,12 @@ module PWN
                   debug_tee: $stdout
                 )
                 $stdout.flush
-                puts "\n\e[32m#{final}\e[0m\n\n"
-                $stdout.flush
-                PWN::Plugins::Log.mirror_tui!(msg: "\n#{final}\n\n") if pi.config.pwn_ai_debug && defined?(PWN::Plugins::Log)
+                promoted = Thread.current[:pwn_thinking_promoted] && final.to_s.strip == Thread.current[:pwn_last_thinking].to_s.strip
+                unless promoted
+                  puts "\n\e[32m#{final}\e[0m\n\n"
+                  $stdout.flush
+                  PWN::Plugins::Log.mirror_tui!(msg: "\n#{final}\n\n") if pi.config.pwn_ai_debug && defined?(PWN::Plugins::Log)
+                end
                 if pi.config.pwn_ai_debug && sess_id && PWN.const_defined?(:Sessions)
                   PWN::Plugins::Log.progress(
                     msg: "session=#{sess_id}",

data/lib/pwn/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module PWN
-  VERSION = '0.5.757'
+  VERSION = '0.5.759'
 end

data/spec/lib/pwn/ai/agent/loop_spec.rb

@@ -2123,36 +2123,33 @@ describe PWN::AI::Agent::Loop do # rubocop:disable Metrics/BlockLength
     expect(src).to include('byteslice(0, 2_048)')
   end
 
-  describe 'thinking on the task TUI line' do
-    it 'prefers model thinking over the TaskSummarizer brief' do
+  describe 'thinking on the session pane' do
+    it 'keeps the task brief and emits thinking separately' do
       allow(PWN::AI::Agent::TaskSummarizer).to receive(:enabled?).and_return(true)
       allow(PWN::AI::Agent::TaskSummarizer).to receive(:about_to).and_return('task 1/3: scan')
       seen = []
       described_class.send(
-        :task_summary_about_to!,
-        state: { plan: ['scan'] },
-        tools: [{ name: 'shell' }],
-        request: 'scan the host',
+        :emit_thinking!,
         thinking: 'Need to fingerprint the service before scanning.',
         on_tool: ->(name, args, _res) { seen << [name, args] }
       )
-      expect(seen).to eq([['task', 'Need to fingerprint the service before scanning.']])
-      expect(PWN::AI::Agent::TaskSummarizer).not_to have_received(:about_to)
-    end
-
-    it 'falls back to the TaskSummarizer brief when thinking is blank' do
-      allow(PWN::AI::Agent::TaskSummarizer).to receive(:enabled?).and_return(true)
-      allow(PWN::AI::Agent::TaskSummarizer).to receive(:about_to).and_return('task 1/3: scan')
-      seen = []
       described_class.send(
         :task_summary_about_to!,
         state: { plan: ['scan'] },
         tools: [{ name: 'shell' }],
         request: 'scan the host',
-        thinking: '  ',
         on_tool: ->(name, args, _res) { seen << [name, args] }
       )
-      expect(seen).to eq([['task', 'task 1/3: scan']])
+      expect(seen).to eq([
+                           ['thinking', 'Need to fingerprint the service before scanning.'],
+                           ['task', 'task 1/3: scan']
+                         ])
+    end
+
+    it 'does not emit a blank thinking row' do
+      seen = []
+      described_class.send(:emit_thinking!, thinking: '  ', on_tool: ->(name, args, _res) { seen << [name, args] })
+      expect(seen).to eq([])
     end
   end
 

data/spec/lib/pwn/ai/ollama_spec.rb

@@ -12,10 +12,11 @@ describe PWN::AI::Ollama do
       allow(described_class).to receive(:sleep)
       allow(described_class).to receive(:rand).and_return(0.5)
       allow(PWN::AI::HttpRetry).to receive(:report_event)
+      described_class.instance_variable_set(:@strict_system_cache, nil)
       expect do
         described_class.chat_with_tools(messages: [{ role: 'user', content: 'hello' }], quiet: quiet)
       end.to raise_error(RuntimeError, /Ollama.*429.*rate limit exceeded/)
-      expect(RestClient::Request).to have_received(:execute).exactly(5).times
+      expect(RestClient::Request).to have_received(:execute).exactly(6).times
       expect(described_class).to have_received(:sleep).with(2.5).exactly(4).times
     end
   end

data/spec/lib/pwn/ai/ollama_system_message_spec.rb

@@ -0,0 +1,52 @@
+# frozen_string_literal: true
+
+require 'spec_helper'
+
+describe 'Ollama and Open WebUI system-message placement' do
+  it 'puts a later Ollama system message first so the chat template does not 500' do
+    engine = { model: 'mini-mythos-mission-ready', num_ctx: 2048, keep_alive: '1m', temp: 0.1, tool_temp: 0.1 }
+    allow(PWN::Env).to receive(:[]).and_call_original
+    allow(PWN::Env).to receive(:[]).with(:ai).and_return({ ollama: engine, active: 'ollama' })
+    captured = nil
+    allow(PWN::AI::Ollama).to receive(:ollama_rest_call) do |**kwargs|
+      captured = kwargs[:http_body]
+      '{"message":{"role":"assistant","content":"ok"}}'
+    end
+    PWN::AI::Ollama.chat_with_tools(
+      model: 'mini-mythos-mission-ready',
+      messages: [
+        { role: 'user', content: 'hello' },
+        { role: 'system', content: 'must lead' },
+        { role: 'developer', content: 'also lead' }
+      ]
+    )
+    expect(captured[:messages].map { |row| row[:role] }).to eq(%w[system user])
+    expect(captured[:messages][0][:content]).to include('must lead', 'also lead')
+  end
+
+  it 'folds Open WebUI system text into the user message when the template rejects a second system role' do
+    engine = { model: 'mini-mythos-mission-ready:latest', num_ctx: 2048, keep_alive: '1m', temp: 0.1, tool_temp: 0.1 }
+    allow(PWN::Env).to receive(:[]).and_call_original
+    allow(PWN::Env).to receive(:[]).with(:ai).and_return({ openwebui: engine, active: 'openwebui' })
+    PWN::AI::OpenWebUI.instance_variable_set(:@strict_system_cache, nil)
+    captured = nil
+    allow(PWN::AI::OpenWebUI).to receive(:openwebui_rest_call) do |**kwargs|
+      if kwargs[:rest_call].to_s.end_with?('show')
+        { template: "raise_exception('System message must be at the beginning.')" }.to_json
+      else
+        captured = kwargs[:http_body]
+        '{"message":{"role":"assistant","content":"ok"}}'
+      end
+    end
+    PWN::AI::OpenWebUI.chat_with_tools(
+      model: 'mini-mythos-mission-ready:latest',
+      messages: [
+        { role: 'system', content: 'operator rules' },
+        { role: 'user', content: 'ping' },
+        { role: 'system', content: 'ENGAGEMENT MEMORY' }
+      ]
+    )
+    expect(captured[:messages].map { |row| row[:role] }).to eq(%w[user])
+    expect(captured[:messages][0][:content]).to include('operator rules', 'ENGAGEMENT MEMORY', 'ping')
+  end
+end

data/spec/lib/pwn/ai/open_web_ui_spec.rb

@@ -12,10 +12,11 @@ describe PWN::AI::OpenWebUI do
       allow(described_class).to receive(:sleep)
       allow(described_class).to receive(:rand).and_return(0.5)
       allow(PWN::AI::HttpRetry).to receive(:report_event)
+      described_class.instance_variable_set(:@strict_system_cache, nil)
       expect do
         described_class.chat_with_tools(messages: [{ role: 'user', content: 'hello' }], quiet: quiet)
       end.to raise_error(RuntimeError, /Open WebUI.*429.*rate limit exceeded/)
-      expect(RestClient::Request).to have_received(:execute).exactly(5).times
+      expect(RestClient::Request).to have_received(:execute).exactly(6).times
       expect(described_class).to have_received(:sleep).with(2.5).exactly(4).times
     end
   end
@@ -83,6 +84,33 @@ describe PWN::AI::OpenWebUI do
     expect(captured_chat[:stream]).to be true
   end
 
+  it 'merges later system and developer messages into one leading system message' do
+    engine = { model: 'mini-mythos-mission-ready', num_ctx: 2048, keep_alive: '1m', temp: 0.1, tool_temp: 0.1 }
+    allow(PWN::Env).to receive(:[]).and_call_original
+    allow(PWN::Env).to receive(:[]).with(:ai).and_return({ openwebui: engine, active: 'openwebui' })
+    captured = nil
+    allow(described_class).to receive(:openwebui_rest_call) do |**kwargs|
+      captured = kwargs[:http_body]
+      '{"message":{"role":"assistant","content":"ok"}}'
+    end
+    described_class.chat_with_tools(
+      messages: [
+        { role: 'system', content: 'operator rules' },
+        { role: 'user', content: 'hello' },
+        { role: 'system', content: 'ENGAGEMENT MEMORY\nnotes' },
+        { role: 'developer', content: 'mid-turn reminder' },
+        { role: 'assistant', content: '', tool_calls: [{ id: 'c1', type: 'function', function: { name: 'shell', arguments: '{"cmd":"id"}' } }] }
+      ],
+      model: 'mini-mythos-mission-ready'
+    )
+    roles = captured[:messages].map { |row| row[:role] }
+    expect(roles).to eq(%w[system user assistant])
+    expect(captured[:messages][0][:content]).to include('operator rules')
+    expect(captured[:messages][0][:content]).to include('ENGAGEMENT MEMORY')
+    expect(captured[:messages][0][:content]).to include('mid-turn reminder')
+    expect(captured[:messages][2][:tool_calls].first[:function][:arguments]).to eq({ cmd: 'id' })
+  end
+
   it 'assembles native /api/chat NDJSON stream into a single message' do
     body = [
       { model: 'm', message: { role: 'assistant', content: 'Hel' }, done: false }.to_json,

data/spec/lib/pwn/plugins/repl/ai_console_spec.rb

@@ -389,6 +389,19 @@ describe 'pwn-ai curses launch' do # rubocop:disable Metrics/BlockLength -- publ
     expect(colors).to eq(operator_label: 4, operator_body: 5, assistant: 5, task: 2, tool: 1, result: 3)
   end
 
+  it 'shows model thinking in the session pane even when compact' do
+    pry = Pry.new
+    console = PWN::Plugins::REPL::AIConsole::Console.new(pry: pry, input: StringIO.new, curses: nil, getch: nil)
+    console.instance_variable_set(:@verbose, false)
+    console.add(:thinking, 'The lemon is yellow because of carotenoids.')
+    console.add(:assistant, 'Yellow.')
+    text = console.timeline_rows(60).map { |_color, row| row }.join("\n")
+    expect(text).to include('THINKING')
+    expect(text).to include('The lemon is yellow because of carotenoids.')
+    expect(text).to include('ASSISTANT')
+    expect(text).to include('Yellow.')
+  end
+
   it 'uses template colors for every unconfigured role without mutating defaults or overrides' do
     defaults = PWN::Config.env_template.dig(:ai, :tui, :theme)
     [nil, {}, 'invalid'].each do |configured|

data/third_party/pwn_rdoc.jsonl

@@ -403,6 +403,7 @@
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.dispatch_fail_n Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.dispatch_fail_n`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.duration_unsatisfied? Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.duration_unsatisfied?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.emit_task_summary Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.emit_task_summary`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.emit_thinking! Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.emit_thinking!`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.engine_transient? Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.engine_transient?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.escalate Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.escalate`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Agent::Loop.evidence_blob Usage"},{"role":"assistant","content":"`PWN::AI::Agent::Loop.evidence_blob`: "}]}
@@ -1201,10 +1202,13 @@
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.get_model Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.get_model`: Supported Method Parameters\n\nmodel = PWN::AI::Ollama.get_model(name: ‘required - tag name’)\n\nLocal tags have size and family, not token prices. get_model reads /api/show for that tag and returns it. Missing pricing stays missing.\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.get_models Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.get_models`: Supported Method Parameters\n\nresponse = PWN::AI::Ollama.get_models\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.help Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.help`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::Ollama.lead_system Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.lead_system`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.normalize_messages_for_ollama Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.normalize_messages_for_ollama`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.ollama_rest_call Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.ollama_rest_call`: Supported Method Parameters\n\nollama_rest_call(\n\nhttp_method: 'optional HTTP method (defaults to GET)',\nrest_call: 'required rest call path relative to base_uri (e.g. api/chat)',\nparams: 'optional params passed in the URI or HTTP Headers',\nhttp_body: 'optional HTTP body sent in HTTP methods that support it e.g. POST',\ntimeout: 'optional timeout in seconds (defaults to 900)',\nspinner: 'optional - display spinner (defaults to false)'\n\n)\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.parse_tool_arguments_object Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.parse_tool_arguments_object`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.real_config_value? Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.real_config_value?`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::Ollama.strict_system_template? Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.strict_system_template?`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::Ollama.template_rejects_later_system? Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.template_rejects_later_system?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::Ollama.visible_from_thinking Usage"},{"role":"assistant","content":"`PWN::AI::Ollama.visible_from_thinking`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenAI.api_endpoint Usage"},{"role":"assistant","content":"`PWN::AI::OpenAI.api_endpoint`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenAI.authors Usage"},{"role":"assistant","content":"`PWN::AI::OpenAI.authors`: Author(s)\n\n0day Inc. <support@0dayinc.com>\n"}]}
@@ -1252,10 +1256,13 @@
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.get_model Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.get_model`: Supported Method Parameters\n\nmodel = PWN::AI::OpenWebUI.get_model(name: ‘required - model id’)\n\nOpen WebUI mixes OpenAI-compat rows and proxied Ollama tags. A direct model route is tried first; otherwise the already-fetched catalog row is returned. Local tags without a price stay unpriced.\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.get_models Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.get_models`: Supported Method Parameters\n\nresponse = PWN::AI::OpenWebUI.get_models\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.help Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.help`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.lead_system Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.lead_system`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.normalize_messages_for_ollama Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.normalize_messages_for_ollama`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.openwebui_rest_call Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.openwebui_rest_call`: Supported Method Parameters\n\nopenwebui_rest_call(\n\nhttp_method: 'optional HTTP method (defaults to GET)',\nrest_call: 'required rest call path relative to base_uri',\nparams: 'optional params passed in the URI or HTTP Headers',\nhttp_body: 'optional HTTP body sent in HTTP methods that support it e.g. POST',\ntimeout: 'optional timeout in seconds (defaults to 900)',\nspinner: 'optional - display spinner (defaults to false)'\n\n)\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.parse_tool_arguments_object Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.parse_tool_arguments_object`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.real_config_value? Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.real_config_value?`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.strict_system_template? Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.strict_system_template?`: "}]}
+{"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.template_rejects_later_system? Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.template_rejects_later_system?`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::OpenWebUI.visible_from_thinking Usage"},{"role":"assistant","content":"`PWN::AI::OpenWebUI.visible_from_thinking`: "}]}
 {"messages":[{"role":"user","content":"PWN::AI::RedTeam.authors Usage"},{"role":"assistant","content":"`PWN::AI::RedTeam.authors`: Author(s)\n\n0day Inc. <support@0dayinc.com>\n"}]}
 {"messages":[{"role":"user","content":"PWN::AI::RedTeam.help Usage"},{"role":"assistant","content":"`PWN::AI::RedTeam.help`: "}]}

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: pwn
 version: !ruby/object:Gem::Version
-  version: 0.5.757
+  version: 0.5.759
 platform: ruby
 authors:
 - 0day Inc.
@@ -771,14 +771,14 @@ dependencies:
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.6.3
+        version: 1.7.0
   type: :development
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - '='
       - !ruby/object:Gem::Version
-        version: 1.6.3
+        version: 1.7.0
 - !ruby/object:Gem::Dependency
   name: pry
   requirement: !ruby/object:Gem::Requirement
@@ -6754,6 +6754,7 @@ files:
 - spec/lib/pwn/ai/model_catalog_spec.rb
 - spec/lib/pwn/ai/model_lookup_spec.rb
 - spec/lib/pwn/ai/ollama_spec.rb
+- spec/lib/pwn/ai/ollama_system_message_spec.rb
 - spec/lib/pwn/ai/open_ai_oauth_transport_spec.rb
 - spec/lib/pwn/ai/open_ai_spec.rb
 - spec/lib/pwn/ai/open_web_ui_spec.rb