Module: Legion::LLM::Inference::Executor::ContextWindow

Included in:
Legion::LLM::Inference::Executor
Defined in:
lib/legion/llm/inference/executor/context_window.rb

Overview

ContextWindow methods extracted from Executor verbatim (P4b §1.5, refactor-under-green). Functional message-list transformations: empty/thinking/tool-result trimming and context-window-aware compaction. Operates on the messages argument; reads

Instance Method Summary collapse

Instance Method Details

#compact_to_fit(messages, target_tokens) ⇒ Object 



69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 69

def compact_to_fit(messages, target_tokens)
  return messages if estimate_message_tokens(messages) <= target_tokens

  filtered = messages.reject do |msg|
    role = (msg[:role] || msg['role']).to_s
    role == 'tool' && (msg[:content] || msg['content']).to_s.length > 500
  end
  messages = filtered.map do |msg|
    role = (msg[:role] || msg['role']).to_s
    next msg unless role == 'tool'

    content = (msg[:content] || msg['content']).to_s
    content.length > 200 ? msg.merge(content: "#{content[0, 200]}\n[compacted]") : msg
  end

  return messages if estimate_message_tokens(messages) <= target_tokens

  messages = messages.last(messages.size / 2) while messages.size > 2 && estimate_message_tokens(messages) > target_tokens
  messages
end

#empty_assistant_message?(msg) ⇒ Boolean

Returns:

  • (Boolean) 


219
220
221
222
223
224
225
226
227
228
229
230
231
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 219

def empty_assistant_message?(msg)
  return false unless msg.is_a?(Hash)
  return false unless (msg[:role] || msg['role']).to_s == 'assistant'

  content = msg[:content] || msg['content']
  has_content = content.is_a?(String) ? !content.strip.empty? : !content.nil?
  return false if has_content

  tool_calls = msg[:tool_calls] || msg['tool_calls']
  return false if tool_calls.is_a?(Array) && tool_calls.any?

  true
end

#enforce_context_window(messages) ⇒ Object 



24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 24

def enforce_context_window(messages)
  context_window = resolved_context_window
  @context_accounting[:component_status][:context_window] = :observed
  return messages unless context_window&.positive?

  threshold = (context_window * Legion::Settings[:llm][:context_curation][:context_window_threshold]).to_i
  tool_budget = estimate_tool_token_budget
  available_for_messages = threshold - tool_budget
  estimated = estimate_message_tokens(messages)
  return messages if estimated <= available_for_messages

  log.warn "[llm][executor] action=context_compaction request_id=#{@request.id} " \
           "estimated_tokens=#{estimated} context_window=#{context_window} " \
           "threshold=#{threshold} tool_budget=#{tool_budget} available=#{available_for_messages}"

  preserve_after = last_user_message_index(messages)
  recent = messages[preserve_after..]
  older = messages[0...preserve_after]

  target_tokens = available_for_messages - estimate_message_tokens(recent)
  compacted = compact_to_fit(older, target_tokens)

  result = compacted + recent
  after_tokens = estimate_message_tokens(result)
  saved = [estimated - after_tokens, 0].max

  @context_accounting[:tokens][:context_window_saved_estimated_tokens] += saved
  @context_accounting[:counts][:context_window_message_count_before] = messages.size
  @context_accounting[:counts][:context_window_message_count_after] = result.size
  @context_accounting[:events] << ContextAccounting.event(
    event_type:    :context_window_enforcement,
    component:     :context_window,
    before_tokens: estimated,
    after_tokens:  after_tokens,
    before_count:  messages.size,
    after_count:   result.size,
    metadata:      { context_window: context_window, threshold: threshold }
  )

  log.info "[llm][executor] action=context_compaction_complete request_id=#{@request.id} " \
           "before=#{messages.size} after=#{result.size} " \
           "tokens_before=#{estimated} tokens_after=#{after_tokens}"
  result
end

#estimate_message_tokens(messages) ⇒ Object 



96
97
98
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 96

def estimate_message_tokens(messages)
  messages.sum { |m| ((m[:content] || m['content']).to_s.length / 4.0).ceil }
end

#estimate_tool_token_budget ⇒ Object 



100
101
102
103
104
105
106
107
108
109
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 100

def estimate_tool_token_budget
  tools = @request.tools
  return 0 if tools.nil? || tools.empty?

  tool_list = tools.is_a?(Hash) ? tools.values : Array(tools)
  tool_list.sum do |tool|
    json_repr = tool.respond_to?(:to_h) ? Legion::JSON.dump(tool.to_h) : tool.to_s
    (json_repr.length / 3.5).ceil
  end
end

#last_user_message_index(messages) ⇒ Object 



208
209
210
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 208

def last_user_message_index(messages)
  messages.rindex { |m| (m[:role] || m['role']).to_s == 'user' } || messages.size
end

#native_dispatch_messages ⇒ Object 



12
13
14
15
16
17
18
19
20
21
22
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 12

def native_dispatch_messages
  messages = apply_conversation_breakpoint(@request.messages)
  rejected = messages.count { |m| empty_assistant_message?(m) }
  if rejected.positive?
    log.warn "[llm][executor] action=strip_empty_assistants request_id=#{@request.id} removed=#{rejected}"
    messages = messages.reject { |m| empty_assistant_message?(m) }
  end
  messages = strip_thinking_from_history(messages)
  messages = trim_oversized_tool_results(messages)
  enforce_context_window(messages)
end

#reduce_messages_for_dispatch(messages) ⇒ Object

Lane-INDEPENDENT reduction applied before dispatch: empty-assistant prune + leading-thinking strip + oversized-tool-result trim. PURE — no call it to measure exactly what native_dispatch_messages will send (minus the lane-dependent enforce_context_window compaction, which is correctly excluded from routing since it depends on the chosen lane).



117
118
119
120
121
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 117

def reduce_messages_for_dispatch(messages)
  msgs = Array(messages).reject { |m| empty_assistant_message?(m) }
  msgs = strip_thinking_pure(msgs)
  trim_oversized_tool_results_pure(msgs)
end

#resolved_context_window ⇒ Object 



90
91
92
93
94
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 90

def resolved_context_window
  @resolved_offering_metadata&.dig(:limits, :context_window) ||
    @resolved_offering_metadata&.dig(:context_window) ||
    @resolved_offering_metadata&.dig('limits', 'context_window')
end

#strip_leading_thinking_block(text) ⇒ Object 



167
168
169
170
171
172
173
174
175
176
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 167

def strip_leading_thinking_block(text)
  result = text.lstrip
  THINKING_TAG_PAIRS.each do |open_tag, close_tag|
    next unless result.start_with?(open_tag)

    close_idx = result.index(close_tag, open_tag.length)
    return close_idx ? result[(close_idx + close_tag.length)..].lstrip : ''
  end
  text
end

#strip_thinking_from_history(messages) ⇒ Object 



141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 141

def strip_thinking_from_history(messages)
  before_tokens = ContextAccounting.estimate_message_tokens(messages)
  result = strip_thinking_pure(messages)
  stripped_count = messages.zip(result).count { |before, after| before != after }

  after_tokens = ContextAccounting.estimate_message_tokens(result)
  saved = [before_tokens - after_tokens, 0].max
  @context_accounting[:component_status][:thinking_strip] = :observed
  if saved.positive?
    @context_accounting[:tokens][:stripped_thinking_estimated_tokens] += saved
    @context_accounting[:counts][:stripped_thinking_message_count] += stripped_count
    @context_accounting[:events] << ContextAccounting.event(
      event_type:    :thinking_stripped,
      component:     :stripped_thinking,
      before_tokens: before_tokens,
      after_tokens:  after_tokens,
      before_count:  messages.size,
      after_count:   result.size,
      metadata:      { stripped_count: stripped_count }
    )
  end

  log.info "[llm][executor] action=strip_thinking_history request_id=#{@request.id} stripped=#{stripped_count}" if stripped_count.positive?
  result
end

#strip_thinking_pure(messages) ⇒ Object

Pure leading-thinking strip. Shared by strip_thinking_from_history (which adds accounting) and reduce_messages_for_dispatch.



125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 125

def strip_thinking_pure(messages)
  preserve_after = last_user_message_index(messages)
  messages.each_with_index.map do |msg, idx|
    next msg if idx >= preserve_after
    next msg unless (msg[:role] || msg['role']).to_s == 'assistant'

    content = msg[:content] || msg['content']
    next msg unless content.is_a?(String)

    cleaned = strip_leading_thinking_block(content)
    next msg if cleaned == content

    msg.merge(content: cleaned)
  end
end

#tool_result_message?(msg) ⇒ Boolean

Returns:

  • (Boolean) 


212
213
214
215
216
217
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 212

def tool_result_message?(msg)
  return false unless msg.is_a?(Hash)

  role = (msg[:role] || msg['role']).to_s
  role == 'tool' || msg.key?(:tool_call_id) || msg.key?('tool_call_id')
end

#trim_oversized_tool_results(messages) ⇒ Object 



198
199
200
201
202
203
204
205
206
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 198

def trim_oversized_tool_results(messages)
  result = trim_oversized_tool_results_pure(messages)
  trimmed_count = messages.zip(result).count { |before, after| before != after }
  if trimmed_count.positive?
    log.info "[llm][executor] action=trim_tool_results request_id=#{@request.id} trimmed=#{trimmed_count} " \
             "max_chars=#{Legion::Settings[:llm][:tool_result_max_dispatch_chars].to_i}"
  end
  result
end

#trim_oversized_tool_results_pure(messages) ⇒ Object

Pure oversized-tool-result trim. Shared by trim_oversized_tool_results (which adds logging) and reduce_messages_for_dispatch.



180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
 
# File 'lib/legion/llm/inference/executor/context_window.rb', line 180

def trim_oversized_tool_results_pure(messages)
  max_chars = Legion::Settings[:llm][:tool_result_max_dispatch_chars].to_i
  return messages unless max_chars.positive?

  preserve_after = last_user_message_index(messages)
  messages.each_with_index.map do |msg, idx|
    next msg if idx >= preserve_after
    next msg unless tool_result_message?(msg)

    content = msg[:content] || msg['content']
    next msg unless content.is_a?(String) && content.length > max_chars

    msg.merge(content: "#{content[0, max_chars]}\n\n[TRUNCATED: showing first #{max_chars} of #{content.length} chars. " \
                       'If you need more content, make multiple smaller targeted requests ' \
                       '(e.g. read specific line ranges, grep for specific patterns, or request smaller sections).]')
  end
end