legion-llm 0.6.18 → 0.6.20

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 83b1bf0eb338e47eb37627eb7004afb30b3fba91627577034554827edd797bcf
-  data.tar.gz: 54ba4b787dfb7ebbbe6ef10f363cb237075f5b37f74467c9408db723d9a36c48
+  metadata.gz: 45d07a2c60a8663ba1b62165b3b489d49a2aac37ee1e1ec6abff7bd5f4357d6c
+  data.tar.gz: 9ee8246c75fee6d7e690b55f4e2a91b030f6b142c91dd79acb7bf66edf4d9d05
 SHA512:
-  metadata.gz: 1a546dca02b403bc025621c1a980cf081114c383f97bd3ce3eef245a7b3314cdaacc7328cbbbfd7879cbf95dd53069e48b9af8f2401099e68966e0ee4e0b624f
-  data.tar.gz: ddb6aabeca04cdbdacfd2412ee2713dfaecc4b17911b55246ec680538874df7725eabef8768fe69c8202952d6a3f14d0a0ae27696cd64ebf2727679f7947cd61
+  metadata.gz: 92b102167bb6f346fab490787baedda2f2fa6fb528713c6b055b269f747c490d56fed5d21027616bc2c6d1f7cf4069ce14bb9b7607f7a6ad07c2c69b05ce0814
+  data.tar.gz: f1fded39722bf678936df28f3bbf3ec095265bdabc28f70eaf67e64fae5519b7c58842a8432e4b4bfdc0476c9275473f75a9c83bf0c77d6f5cc2afe1fa700aeb

data/CHANGELOG.md

@@ -1,12 +1,22 @@
 # Legion LLM Changelog
 
-## [Unreleased]
+## [0.6.20] - 2026-04-06
+
+### Added
+- Per-step pipeline timing diagnostics: `[pipeline][timing]` log line with duration per step
+- Pre-pipeline timing in inference route: `gaia_ingest`, `pre_pipeline_setup`, `executor_call` durations
+- `MAX_RUBY_LLM_TOOL_ROUNDS` (25) — caps RubyLLM's unbounded tool-use loop to prevent infinite cycling
+- `install_tool_loop_guard` applied to both streaming and non-streaming provider paths
 
 ### Fixed
 - fix Process namespace collision by using ::Process::CLOCK_MONOTONIC prefix inside Legion namespace
 
 ### Added
 - `Legion::LLM::Pipeline::ToolAdapter` - wraps Tools::Base for RubyLLM sessions
+- `Profile.derive` returns `:human` for `type: :human` and `type: :user` callers (Wire Format Phase 3)
+- `Profile.derive` returns `:service` for `type: :service` callers (Wire Format Phase 3)
+- `HUMAN_SKIP` constant (empty — humans get full pipeline)
+- `SERVICE_SKIP` constant — services skip conversational steps (context, tools, knowledge)
 
 ### Changed
 - Renamed `McpToolAdapter` to `ToolAdapter` (backwards compat alias kept)

data/lib/legion/llm/pipeline/executor.rb

@@ -45,6 +45,8 @@ module Legion
 
         ASYNC_SAFE_STEPS = %i[post_response knowledge_capture response_return].freeze
 
+        MAX_RUBY_LLM_TOOL_ROUNDS = 25
+
         ASYNC_THREAD_POOL = Concurrent::FixedThreadPool.new(4, fallback_policy: :caller_runs)
 
         def initialize(request)
@@ -132,15 +134,22 @@ module Legion
         def execute_steps
           executed = 0
           skipped = 0
+          pipeline_start = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
+          step_timings = []
           STEPS.each do |step|
             if Profile.skip?(@profile, step)
               skipped += 1
               next
             end
 
+            t0 = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
             execute_step(step) { send(:"step_#{step}") }
+            elapsed_ms = ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - t0) * 1000).round
+            step_timings << "#{step}=#{elapsed_ms}ms"
             executed += 1
           end
+          total_ms = ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - pipeline_start) * 1000).round
+          log.warn("[pipeline][timing] profile=#{@profile} total=#{total_ms}ms executed=#{executed} skipped=#{skipped} #{step_timings.join(' ')}")
           annotate_top_level_span(steps_executed: executed, steps_skipped: skipped)
         end
 
@@ -464,6 +473,7 @@ module Legion
 
         def execute_provider_request_ruby_llm
           session, message_content = build_ruby_llm_session
+          install_tool_loop_guard(session)
           @raw_response = message_content ? session.ask(message_content) : session
         end
 
@@ -645,6 +655,7 @@ module Legion
           )
 
           session, message_content = build_ruby_llm_session
+          install_tool_loop_guard(session)
           @raw_response = message_content ? session.ask(message_content, &) : session
 
           @timestamps[:provider_end] = Time.now
@@ -678,6 +689,19 @@ module Legion
           inject_registry_tools(session)
         end
 
+        def install_tool_loop_guard(session)
+          return unless session.respond_to?(:on)
+
+          tool_round = 0
+          session.on(:tool_call) do |_tool_call|
+            tool_round += 1
+            if tool_round > MAX_RUBY_LLM_TOOL_ROUNDS
+              log.warn("[pipeline] tool loop cap hit: #{tool_round} rounds, halting")
+              raise Legion::LLM::PipelineError, "tool loop exceeded #{MAX_RUBY_LLM_TOOL_ROUNDS} rounds"
+            end
+          end
+        end
+
         def apply_ruby_llm_instructions(session)
           injected_system = EnrichmentInjector.inject(
             system:      @request.system,

data/lib/legion/llm/pipeline/profile.rb

@@ -21,6 +21,14 @@ module Legion
           tool_calls context_store post_response knowledge_capture
         ].freeze
 
+        HUMAN_SKIP = %i[].freeze
+
+        SERVICE_SKIP = %i[
+          conversation_uuid context_load gaia_advisory
+          rag_context tool_discovery confidence_scoring
+          tool_calls context_store knowledge_capture
+        ].freeze
+
         module_function
 
         def derive(caller_hash)
@@ -31,7 +39,9 @@ module Legion
           identity = requested_by[:identity].to_s
 
           return :quick_reply if type == :quick_reply
-          return :external unless type == :system
+          return :human       if %i[human user].include?(type)
+          return :service     if type == :service
+          return :external    unless type == :system
 
           identity.start_with?('gaia:') ? :gaia : :system
         end
@@ -41,6 +51,8 @@ module Legion
           when :gaia        then GAIA_SKIP.include?(step)
           when :system      then SYSTEM_SKIP.include?(step)
           when :quick_reply then QUICK_REPLY_SKIP.include?(step)
+          when :human       then HUMAN_SKIP.include?(step)
+          when :service     then SERVICE_SKIP.include?(step)
           else false
           end
         end

data/lib/legion/llm/routes.rb

@@ -446,8 +446,11 @@ module Legion
           last_user = messages.select { |m| (m[:role] || m['role']).to_s == 'user' }.last
           prompt    = (last_user || {})[:content] || (last_user || {})['content'] || ''
 
+          route_t0 = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
+
           if defined?(Legion::Gaia) && Legion::Gaia.respond_to?(:started?) && Legion::Gaia.started? && prompt.to_s.length.positive?
             begin
+              gaia_t0 = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
               frame = Legion::Gaia::InputFrame.new(
                 content:      prompt,
                 channel_id:   :api,
@@ -456,6 +459,8 @@ module Legion
                 metadata:     { source_type: :human_direct, salience: 0.9 }
               )
               Legion::Gaia.ingest(frame)
+              gaia_ms = ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - gaia_t0) * 1000).round
+              log.warn("[inference][timing] gaia_ingest=#{gaia_ms}ms request_id=#{request_id}")
             rescue StandardError => e
               handle_exception(e, level: :warn, operation: 'llm.routes.gaia_ingest', request_id: request_id)
             end
@@ -501,6 +506,9 @@ module Legion
             cache:           { strategy: :default, cacheable: true }
           )
 
+          setup_ms = ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - route_t0) * 1000).round
+          log.warn("[inference][timing] pre_pipeline_setup=#{setup_ms}ms request_id=#{request_id}")
+
           executor = Legion::LLM::Pipeline::Executor.new(pipeline_request)
 
           if streaming
@@ -561,7 +569,10 @@ module Legion
             end
             # rubocop:enable Metrics/BlockLength
           else
+            exec_t0 = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
             pipeline_response = executor.call
+            exec_ms = ((::Process.clock_gettime(::Process::CLOCK_MONOTONIC) - exec_t0) * 1000).round
+            log.warn("[inference][timing] executor_call=#{exec_ms}ms request_id=#{request_id}")
             raw_msg = pipeline_response.message
             content = raw_msg.is_a?(Hash) ? (raw_msg[:content] || raw_msg['content']) : raw_msg.to_s
             routing = pipeline_response.routing || {}

data/lib/legion/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Legion
   module LLM
-    VERSION = '0.6.18'
+    VERSION = '0.6.20'
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: legion-llm
 version: !ruby/object:Gem::Version
-  version: 0.6.18
+  version: 0.6.20
 platform: ruby
 authors:
 - Esity