Module: Legion::LLM::API::Namespaces::OpenAI::Responses

Extended by:
Legion::Logging::Helper
Defined in:
lib/legion/llm/api/namespaces/openai/responses.rb

Class Method Summary collapse

Class Method Details

.build_output_reasoning(pipeline_response) ⇒ Object 



485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 485

def self.build_output_reasoning(pipeline_response)
  thinking_data = pipeline_response.respond_to?(:thinking) ? pipeline_response.thinking : nil
  log.info "[llm][responses] build_output_reasoning thinking_data=#{thinking_data.inspect}"
  text = extract_thinking_text(thinking_data)
  log.info "[llm][responses] build_output_reasoning extracted_text_length=#{text.length}"
  return [] if text.empty?

  # OpenAI Responses API format: type: "thinking" with thinking text field
  [
    {
      type:     'thinking',
      id:       "thnk_#{SecureRandom.hex(12)}",
      thinking: text,
      status:   'completed'
    }
  ]
end

.build_output_tool_calls(pipeline_response) ⇒ Object 



470
471
472
473
474
475
476
477
478
479
480
481
482
483
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 470

def self.build_output_tool_calls(pipeline_response)
  tools_data = pipeline_response.respond_to?(:tools) ? pipeline_response.tools : nil
  return [] unless tools_data.is_a?(Array) && !tools_data.empty?

  tools_data.filter_map do |tc|
    name  = tc.respond_to?(:name) ? tc.name : (tc[:name] || tc['name'])
    args  = tc.respond_to?(:arguments) ? tc.arguments : (tc[:arguments] || tc['arguments'] || {})
    tc_id = tc.respond_to?(:id) ? tc.id : (tc[:id] || tc['id'] || "call_#{SecureRandom.hex(8)}")
    next unless name

    { type: 'function_call', id: "fc_#{SecureRandom.hex(12)}", call_id: tc_id,
      name: name.to_s, arguments: args.is_a?(String) ? args : Legion::JSON.dump(args), status: 'completed' }
  end
end

.build_tool_declarations(tools) ⇒ Object 



249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 249

def self.build_tool_declarations(tools)
  return [] unless tools.is_a?(Array) && !tools.empty?

  tools.filter_map do |tool|
    fn = nil
    t  = tool.respond_to?(:transform_keys) ? tool.transform_keys(&:to_sym) : tool
    fn = t[:function] || t
    fn = fn.transform_keys(&:to_sym) if fn.respond_to?(:transform_keys)
    next unless fn[:name].to_s.length.positive?

    Legion::LLM::Types::ToolDefinition.build(
      name:        fn[:name].to_s,
      description: fn[:description].to_s,
      parameters:  fn[:parameters] || {},
      source:      { type: :client, executable: true }
    )
  rescue StandardError => e
    tool_name = fn.is_a?(Hash) ? fn[:name] : nil
    Legion::Logging::Helper.log.warn("[llm][api][namespaces][openai][responses] build_tool failed name=#{tool_name} error=#{e.message}")
    nil
  end
end

.build_usage(tokens) ⇒ Object 



571
572
573
574
575
576
577
578
579
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 571

def self.build_usage(tokens)
  i = extract_token(tokens, :input_tokens)
  o = extract_token(tokens, :output_tokens)
  result = { input_tokens: i, output_tokens: o, total_tokens: i + o }
  # Preserve output token breakdown (e.g. reasoning_tokens from Responses API)
  details = tokens[:output_tokens_details] || tokens['output_tokens_details']
  result[:output_tokens_details] = details if details.is_a?(Hash) && !details.empty?
  result
end

.call_executor(executor, upstream_body: nil) ⇒ Object

rubocop:enable Metrics/AbcSize



448
449
450
451
452
453
454
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 448

def self.call_executor(executor, upstream_body: nil, &)
  if executor.respond_to?(:call_responses) && executor.respond_to?(:provider_supports_responses?) && executor.provider_supports_responses?
    executor.call_responses(body: upstream_body, stream: true, &)
  else
    executor.call_stream(&)
  end
end

.call_executor_sync(executor, upstream_body: nil) ⇒ Object 



456
457
458
459
460
461
462
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 456

def self.call_executor_sync(executor, upstream_body: nil)
  if executor.respond_to?(:call_responses) && executor.respond_to?(:provider_supports_responses?) && executor.provider_supports_responses?
    executor.call_responses(body: upstream_body, stream: false)
  else
    executor.call
  end
end

.close_thinking_item(out, output_items, sequence:) ⇒ Object 



532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 532

def self.close_thinking_item(out, output_items, sequence:)
  state = current_thinking_state(output_items)
  return unless state && state[:status] == 'in_progress'

  output_index = output_items.index(state)
  state[:status] = 'completed'
  text = state[:thinking].to_s
  out << sse('response.thinking.done',
             { type: 'response.thinking.done', sequence_number: sequence.call,
               output_index: output_index, item_id: state[:id], text: text })
  out << sse('response.thinking_part.done',
             { type: 'response.thinking_part.done', sequence_number: sequence.call,
               output_index: output_index, item_id: state[:id],
               part: { type: 'thinking', thinking: text } })
  out << sse('response.output_item.done',
             { type: 'response.output_item.done', sequence_number: sequence.call,
               output_index: output_index, item: state })
end

.current_thinking_state(output_items) ⇒ Object 



551
552
553
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 551

def self.current_thinking_state(output_items)
  output_items.find { |item| item[:type] == 'thinking' && item[:status] == 'in_progress' }
end

.emit_reasoning_delta(out, _request_id, output_items, text, sequence:) ⇒ Object 



503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 503

def self.emit_reasoning_delta(out, _request_id, output_items, text, sequence:)
  return if text.empty?

  state = current_thinking_state(output_items)
  unless state
    state = {
      type:     'thinking',
      id:       "thnk_#{SecureRandom.hex(12)}",
      thinking: +'',
      status:   'in_progress'
    }
    output_items << state
    output_index = output_items.length - 1
    out << sse('response.output_item.added',
               { type: 'response.output_item.added', sequence_number: sequence.call,
                 output_index: output_index, item: state })
    out << sse('response.thinking_part.added',
               { type: 'response.thinking_part.added', sequence_number: sequence.call,
                 output_index: output_index, item_id: state[:id],
                 part: { type: 'thinking', thinking: '' } })
  end

  output_index = output_items.index(state)
  state[:thinking] << text
  out << sse('response.thinking.delta',
             { type: 'response.thinking.delta', sequence_number: sequence.call,
               output_index: output_index, item_id: state[:id], delta: text })
end

.extract_thinking_config(body) ⇒ Object

Extract thinking/reasoning config from OpenAI Responses API request. OpenAI format: { reasoning: { effort: “low|medium|high” } } Convert to Anthropic thinking config for downstream providers.



227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 227

def self.extract_thinking_config(body)
  reasoning = body[:reasoning] || body['reasoning']
  return nil unless reasoning

  effort = if reasoning.is_a?(Hash)
             reasoning[:effort] || reasoning['effort']
           else
             reasoning
           end

  # Budget must be strictly less than max_tokens (Anthropic constraint).
  # Use conservative defaults — test payloads typically use max_output_tokens: 2048.
  # Preserve the effort value so OpenAI-compatible providers can extract it
  # via openai_reasoning_effort(thinking), while Anthropic providers use budget_tokens.
  case effort.to_s
  when 'low'
    { type: 'enabled', budget_tokens: 512, effort: effort.to_s }
  when 'high', 'medium'
    { type: 'enabled', budget_tokens: 1024, effort: effort.to_s }
  end
end

.extract_thinking_text(value) ⇒ Object 



555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 555

def self.extract_thinking_text(value)
  return '' if value.nil?
  return value.to_s if value.is_a?(String)

  if value.is_a?(Hash)
    normalized = value.transform_keys { |key| key.respond_to?(:to_sym) ? key.to_sym : key }
    text = normalized[:content] || normalized[:text] || normalized[:thinking] || normalized[:reasoning]
    return text.to_s if text
  end

  return value.content.to_s if value.respond_to?(:content) && value.content
  return value.text.to_s if value.respond_to?(:text) && value.text

  value.to_s
end

.extract_token(tokens, key) ⇒ Object 



581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 581

def self.extract_token(tokens, key)
  return 0 if tokens.nil?

  if tokens.is_a?(Hash)
    v = tokens[key] || tokens[key.to_s]
    return v.to_i unless v.nil?

    alt = key == :input_tokens ? :input : :output
    v2  = tokens[alt] || tokens[alt.to_s]
    return v2.to_i unless v2.nil?

    return 0
  end

  method_name = { input_tokens: :input_tokens, output_tokens: :output_tokens,
                  input: :input_tokens, output: :output_tokens }[key]
  return tokens.public_send(method_name).to_i if method_name && tokens.respond_to?(method_name)

  0
end

.flush_pending_tool_calls(messages, pending) ⇒ Object 



211
212
213
214
215
216
217
218
219
220
221
222
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 211

def self.flush_pending_tool_calls(messages, pending)
  return if pending.empty?

  messages << {
    role:       'assistant',
    content:    '',
    tool_calls: pending.map do |tc|
      { id: tc[:id], type: 'function', function: { name: tc[:name], arguments: tc[:arguments] } }
    end
  }
  pending.clear
end

.format_response(pipeline_response, request_id:, model:) ⇒ Object 



272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 272

def self.format_response(pipeline_response, request_id:, model:)
  routing       = pipeline_response.routing || {}
  tokens        = pipeline_response.tokens || {}
  raw_msg       = pipeline_response.message
  content       = raw_msg.is_a?(Hash) ? (raw_msg[:content] || raw_msg['content']).to_s : raw_msg.to_s
  resolved_model = (routing[:model] || routing['model'] || model).to_s
  tool_calls = build_output_tool_calls(pipeline_response)
  reasoning = build_output_reasoning(pipeline_response)

  output = [*reasoning, *tool_calls, {
    type:    'message',
    id:      "msg_#{SecureRandom.hex(12)}",
    role:    'assistant',
    content: [{ type: 'output_text', text: content }],
    status:  'completed'
  }]

  # Per OpenAI Responses API spec: when tool calls are present, the response
  # must signal that client-side execution is required. Using 'completed' tells
  # the client the turn is done and it should not execute the tool calls.
  status = tool_calls.any? ? 'in_progress' : 'completed'

  result = { id: request_id, object: 'response', created_at: Time.now.to_i,
    model: resolved_model, output: output, usage: build_usage(tokens), status: status }

  if tool_calls.any?
    result[:action_required] = {
      type:           'function_calls',
      function_calls: tool_calls
    }
  end

  result
end

.native_responses_supported?(executor, _upstream_body) ⇒ Boolean

Returns:

  • (Boolean) 


464
465
466
467
468
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 464

def self.native_responses_supported?(executor, _upstream_body)
  executor.respond_to?(:call_responses) &&
    executor.respond_to?(:provider_supports_responses?) &&
    executor.provider_supports_responses?
end

.normalize_input_array(input) ⇒ Object

— Support methods —



177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 177

def self.normalize_input_array(input)
  messages = []
  pending_tool_calls = []

  input.each do |item|
    item = item.transform_keys(&:to_sym) if item.respond_to?(:transform_keys)
    case item[:type]&.to_s
    when 'function_call'
      pending_tool_calls << {
        id:        item[:call_id] || item[:id],
        name:      item[:name].to_s,
        arguments: item[:arguments].is_a?(String) ? item[:arguments] : Legion::JSON.dump(item[:arguments] || {})
      }
    when 'function_call_output'
      flush_pending_tool_calls(messages, pending_tool_calls)
      messages << { role: 'tool', tool_call_id: item[:call_id], content: item[:output].to_s }
    else
      flush_pending_tool_calls(messages, pending_tool_calls)
      role = item[:role]&.to_s
      next unless role

      # OpenAI Responses API uses "developer" as a higher-trust system role.
      # All downstream providers only understand the standard four roles.
      role = 'system' if role == 'developer'

      content = item[:content]
      content = content.to_s if content && !content.is_a?(Array)
      messages << { role: role, content: content }.compact
    end
  end
  flush_pending_tool_calls(messages, pending_tool_calls)
  messages
end

.registered(app) ⇒ Object

rubocop:disable Metrics/AbcSize



17
18
19
20
21
22
23
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
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
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
166
167
168
169
170
171
172
173
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 17

def self.registered(app) # rubocop:disable Metrics/AbcSize
  log.debug('[llm][api][namespaces][openai][responses] registering routes')

  app.post '/v1/responses' do
    require_llm!
    request_started_at = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
    body = parse_request_body

    request_id = env['HTTP_X_CLIENT_REQUEST_ID'] || "resp_#{SecureRandom.hex(16)}"

    input = body[:input]
    messages = case input
               when Array
                 Responses.normalize_input_array(input)
               when String
                 [{ role: 'user', content: input }]
               else
                 return openai_error('input is required (string or array)',
                                     type: 'invalid_request_error', status_code: 400)
               end

    messages = [{ role: 'system', content: body[:instructions].to_s }] + messages if body[:instructions]

    model       = body[:model] || Legion::Settings[:llm][:default_model] || 'default'
    streaming   = body[:stream] == true
    tool_decls  = Responses.build_tool_declarations(body[:tools])
    thinking    = Responses.extract_thinking_config(body)

    ext_provider = env['HTTP_X_LEGION_PROVIDER'] || body[:provider]
    ext_tier     = env['HTTP_X_LEGION_TIER']     || body[:tier]
    ext_instance = env['HTTP_X_LEGION_INSTANCE'] || body[:instance]

    routing = { provider: ext_provider, instance: ext_instance, model: model }.compact
    extra   = {}
    extra[:tier] = ext_tier.to_sym if ext_tier

    log.info("[llm][api][namespaces][openai][responses] action=accepted request_id=#{request_id} model=#{model} stream=#{streaming}")

    inference_request = Legion::LLM::Inference::Request.build(
      id:              request_id,
      messages:        messages,
      routing:         routing,
      tools:           tool_decls,
      caller:          build_server_caller(source: 'openai_responses', path: request.path, env: env),
      conversation_id: env['HTTP_X_LEGION_CONVERSATION_ID'] || env['HTTP_THREAD_ID'],
      stream:          streaming,
      thinking:        thinking,
      cache:           { strategy: :default, cacheable: true },
      extra:           extra.empty? ? {} : extra
    )
    executor = Legion::LLM::Inference::Executor.new(inference_request)

    if streaming
      content_type 'text/event-stream'
      headers 'Cache-Control' => 'no-cache', 'Connection' => 'keep-alive', 'X-Accel-Buffering' => 'no'
      stream do |out|
        pipeline_response = Responses.stream_response(out, executor, request_id: request_id, model: model, upstream_body: body)
        tool_calls = Responses.build_output_tool_calls(pipeline_response)
        log_api_completion_summary(
          namespace:         'namespaces][openai][responses',
          request_id:        request_id,
          pipeline_response: pipeline_response,
          stream:            true,
          started_at:        request_started_at,
          tool_calls:        tool_calls,
          stop_reason:       tool_calls.any? ? 'requires_action' : 'completed'
        )
      rescue StandardError => e
        handle_exception(e, level: :error, handled: false, operation: 'llm.api.namespaces.openai.responses.stream', request_id: request_id)
        out << "event: error\ndata: #{Legion::JSON.dump({ type: 'server_error', message: e.message })}\n\n"
      end
    else
      pipeline_response = Responses.call_executor_sync(executor, upstream_body: body)
      response_body     = Responses.format_response(pipeline_response, request_id: request_id, model: model)
      tool_calls = Responses.build_output_tool_calls(pipeline_response)
      log_api_completion_summary(
        namespace:         'namespaces][openai][responses',
        request_id:        request_id,
        pipeline_response: pipeline_response,
        stream:            false,
        started_at:        request_started_at,
        tool_calls:        tool_calls,
        stop_reason:       response_body[:status]
      )
      content_type :json
      status 200
      Legion::JSON.dump(response_body)
    end
  rescue Legion::LLM::AuthError => e
    handle_exception(e, level: :error, handled: true, operation: 'llm.api.namespaces.openai.responses.auth')
    openai_error(e.message, type: 'authentication_error', status_code: 401)
  rescue Legion::LLM::RateLimitError => e
    handle_exception(e, level: :warn, handled: true, operation: 'llm.api.namespaces.openai.responses.rate_limit')
    openai_error(e.message, type: 'rate_limit_error', code: 'rate_limit_exceeded', status_code: 429)
  rescue Legion::LLM::ProviderDown, Legion::LLM::ProviderError => e
    handle_exception(e, level: :error, handled: true, operation: 'llm.api.namespaces.openai.responses.provider')
    openai_error(e.message, type: 'server_error', status_code: 502)
  rescue StandardError => e
    handle_exception(e, level: :error, handled: false, operation: 'llm.api.namespaces.openai.responses')
    openai_error(e.message, type: 'server_error', status_code: 500)
  end
  app.get '/v1/responses/:id' do
    log.debug("[llm][api][namespaces][openai][responses] action=retrieve id=#{params[:id]}")
    openai_error("Response '#{params[:id]}' not found", type: 'invalid_request_error',
                                                        code: 'response_not_found', status_code: 404)
  end

  app.delete '/v1/responses/:id' do
    log.debug("[llm][api][namespaces][openai][responses] action=delete id=#{params[:id]}")
    content_type :json
    Legion::JSON.dump({ id: params[:id], object: 'response', deleted: true })
  end

  app.post '/v1/responses/:id/cancel' do
    log.debug("[llm][api][namespaces][openai][responses] action=cancel id=#{params[:id]}")
    openai_error("Response '#{params[:id]}' not found or already completed",
                 type: 'invalid_request_error', status_code: 404)
  end

  app.get '/v1/responses/:id/input_items' do
    log.debug("[llm][api][namespaces][openai][responses] action=input_items id=#{params[:id]}")
    content_type :json
    Legion::JSON.dump({ object: 'list', data: [], has_more: false })
  end

  app.post '/v1/responses/:id/input_tokens/count' do
    body  = parse_request_body
    input = body[:input]
    model = body[:model] || params[:id]
    messages = case input
               when Array  then Responses.normalize_input_array(input)
               when String then [{ role: 'user', content: input }]
               else []
               end
    result = Legion::LLM::TokenEstimation.estimate(messages: messages, model: model.to_s)
    content_type :json
    Legion::JSON.dump(result)
  rescue StandardError => e
    handle_exception(e, level: :error, handled: false, operation: 'llm.api.namespaces.openai.responses.count_tokens')
    openai_error(e.message, type: 'server_error', status_code: 500)
  end

  app.post '/v1/responses/:id/compact' do
    log.debug("[llm][api][namespaces][openai][responses] action=compact id=#{params[:id]}")
    openai_error("Response '#{params[:id]}' not found", type: 'invalid_request_error', status_code: 404)
  end

  # Legacy alias — preserved for clients using the pre-namespace path.
  app.post '/api/llm/inference/v1/responses' do
    log.debug('[llm][api][namespaces][openai][responses] action=legacy_alias forwarding to /v1/responses handler')
    call env.merge('PATH_INFO' => '/v1/responses')
  end

  log.debug('[llm][api][namespaces][openai][responses] routes registered')
rescue StandardError => e
  handle_exception(e, level: :error, handled: false, operation: 'llm.api.namespaces.openai.responses.register')
end

.sse(name, payload) ⇒ Object 



602
603
604
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 602

def self.sse(name, payload)
  "event: #{name}\ndata: #{Legion::JSON.dump(payload)}\n\n"
end

.stream_response(out, executor, request_id:, model:, upstream_body: nil) ⇒ Object

rubocop:disable Metrics/AbcSize



308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
 
# File 'lib/legion/llm/api/namespaces/openai/responses.rb', line 308

def self.stream_response(out, executor, request_id:, model:, upstream_body: nil)
  created_at = Time.now.to_i
  seq        = 0
  base_resp  = { id: request_id, object: 'response', created_at: created_at,
                 status: 'in_progress', model: model, output: [], usage: nil }

  out << sse('response.created',       { type: 'response.created',       sequence_number: seq += 1, response: base_resp })
  out << sse('response.in_progress',   { type: 'response.in_progress',   sequence_number: seq += 1, response: base_resp })

  msg_id = "msg_#{SecureRandom.hex(12)}"
  msg_index = 0
  message_opened = false
  output_items = []

  open_message = lambda do
    next if message_opened

    msg_index = output_items.length
    message_item = { id: msg_id, type: 'message', role: 'assistant', content: [], status: 'in_progress' }
    output_items << message_item
    out << sse('response.output_item.added',  { type: 'response.output_item.added',  sequence_number: seq += 1, output_index: msg_index,
                                                item: message_item })
    out << sse('response.content_part.added', { type: 'response.content_part.added', sequence_number: seq += 1, output_index: msg_index,
                                                content_index: 0, item_id: msg_id,
                                                part: { type: 'output_text', text: '', annotations: [] } })
    message_opened = true
  end

  full_text = +''
  full_reasoning = +''
  pending_tool_calls = {} # id => { name:, arguments:, output_index: }
  pipeline_response = call_executor(executor, upstream_body: upstream_body) do |chunk|
    thinking = chunk.respond_to?(:thinking) ? extract_thinking_text(chunk.thinking) : ''
    unless thinking.empty?
      full_reasoning << thinking
      emit_reasoning_delta(out, request_id, output_items, thinking, sequence: -> { seq += 1 })
    end

    # Handle tool call deltas from streaming responses.
    # These emit SSE events in real-time so the client sees tool calls
    # as they arrive. At the end, build_output_tool_calls provides the
    # final consolidated list (which filters out server-executed tools).
    if chunk.respond_to?(:tool_calls) && chunk.tool_calls && !chunk.tool_calls.empty?
      close_thinking_item(out, output_items, sequence: -> { seq += 1 })
      chunk.tool_calls.each do |tc_id, tc|
        tc_id_str = tc_id.to_s
        tc_name = tc.respond_to?(:name) ? tc.name.to_s : ''
        tc_args = tc.respond_to?(:arguments) ? tc.arguments.to_s : ''
        next if tc_args.empty? && tc_name.empty?

        unless pending_tool_calls[tc_id_str]
          idx = output_items.length
          out << sse('response.output_item.added',
                     { type: 'response.output_item.added', sequence_number: seq += 1,
                       output_index: idx, item: { id: tc_id_str, type: 'function_call', name: tc_name,
                         call_id: tc_id_str, arguments: '', status: 'in_progress' } })
          pending_tool_calls[tc_id_str] = { id: tc_id_str, name: tc_name, arguments: +'', output_index: idx }
          output_items << { id: tc_id_str, type: 'function_call', name: tc_name,
                            call_id: tc_id_str, arguments: pending_tool_calls[tc_id_str][:arguments], status: 'in_progress' }
        end

        pending_tc = pending_tool_calls[tc_id_str]
        out << sse('response.function_call_arguments.delta',
                   { type: 'response.function_call_arguments.delta', sequence_number: seq += 1,
                     output_index: pending_tc[:output_index], item_id: tc_id_str, delta: tc_args })
        pending_tc[:arguments] << tc_args
      end
    end

    text = chunk.respond_to?(:content) ? chunk.content.to_s : chunk.to_s
    next if text.empty?

    close_thinking_item(out, output_items, sequence: -> { seq += 1 })
    open_message.call
    full_text << text
    out << sse('response.output_text.delta', { type: 'response.output_text.delta', sequence_number: seq += 1,
                                               output_index: msg_index, content_index: 0, item_id: msg_id, delta: text })
  end

  routing        = pipeline_response.routing || {}
  tokens         = pipeline_response.tokens  || {}
  resolved_model = (routing[:model] || routing['model'] || model).to_s
  usage          = build_usage(tokens)

  if full_reasoning.empty?
    final_reasoning = extract_thinking_text(pipeline_response.respond_to?(:thinking) ? pipeline_response.thinking : nil)
    unless final_reasoning.empty?
      full_reasoning << final_reasoning
      emit_reasoning_delta(out, request_id, output_items, final_reasoning, sequence: -> { seq += 1 })
    end
  end
  close_thinking_item(out, output_items, sequence: -> { seq += 1 })

  open_message.call
  out << sse('response.output_text.done',   { type: 'response.output_text.done',   sequence_number: seq += 1,
                                              output_index: msg_index, content_index: 0, item_id: msg_id, text: full_text })
  out << sse('response.content_part.done',  { type: 'response.content_part.done',  sequence_number: seq += 1,
                                              output_index: msg_index, content_index: 0, item_id: msg_id,
                                              part: { type: 'output_text', text: full_text, annotations: [] } })

  completed_item = { id: msg_id, type: 'message', role: 'assistant', status: 'completed',
                     content: [{ type: 'output_text', text: full_text, annotations: [] }] }
  out << sse('response.output_item.done', { type: 'response.output_item.done', sequence_number: seq += 1,
                                            output_index: msg_index, item: completed_item })
  output_items[msg_index] = completed_item

  # Complete any pending streaming tool calls with their final arguments.
  pending_tool_calls.each_value do |pending|
    out << sse('response.function_call_arguments.done',
               { type: 'response.function_call_arguments.done', sequence_number: seq += 1,
                 output_index: pending[:output_index], item_id: pending[:id],
                 arguments: pending[:arguments] })
    out << sse('response.output_item.done',
               { type: 'response.output_item.done', sequence_number: seq += 1,
                 output_index: pending[:output_index],
                 item: { id: pending[:id], type: 'function_call', name: pending[:name],
                         call_id: pending[:id], arguments: pending[:arguments], status: 'completed' } })
  end

  # Determine final status based on whether there are function calls
  # that require client-side execution. Per OpenAI Responses API spec,
  # the final event must be response.done (not response.completed) when
  # function calls need client execution.
  has_tool_calls = pending_tool_calls.any?
  out << if has_tool_calls
           sse('response.done', { type: 'response.done', sequence_number: seq + 1,
             response: { id: request_id, object: 'response', created_at: created_at,
               status: 'requires_action', model: resolved_model,
               output: output_items, usage: usage,
               action_required: { type: 'function_calls', function_calls: output_items.select { |i| i[:type] == 'function_call' } } } })
         else
           sse('response.completed', { type: 'response.completed', sequence_number: seq + 1,
             response: { id: request_id, object: 'response', created_at: created_at,
               status: 'completed', model: resolved_model,
               output: output_items, usage: usage } })
         end
  pipeline_response
end