llm_logs 0.3.1 → 0.4.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: c20c411ed2868a4bbb60216620534d48385b4559c1ef57f8be6182dba9bf1b84
-  data.tar.gz: 451f2cc00575d72296057be68ef287d156cd0d83fc7ed085a863a071ab528660
+  metadata.gz: 71e4dcee0febd77db39e2966ffb5ed3e1e8f5e66dd5b9ff6d12812e44a58db8f
+  data.tar.gz: 37409caad632015780f85f625a97bb0ee45e37e4c0f38a25bde8b8e50ce88af8
 SHA512:
-  metadata.gz: ec33e8232e3d068134f10171d4cec80041f1239703996c28d64b8889ad1813559eadc7bf8d65311518e10d640f8f44218e1677ae581ed27ca24880c610bd728a
-  data.tar.gz: b9180ecd77a042dd187237562689d72773af3f5c822d231ada48bfadcabc2fe9e82c86ba537ed65b2c3b25be9aaac96943cfc1f212f9e39e63c112e6fedb5aa5
+  metadata.gz: 4039c5a3f448495da3a0c6b72cd56082fa37a1e13bf803ecf863a3720f77a764d91104f22b796382f9526e8a97b29458ee5adfc4b868701c7ba7224649ffccc7
+  data.tar.gz: ccfb679fc19b675a27de5f2aa4b52c20597aba015ff9462e2d0f11b9890b33b3ecd9d62b06360ce20573cc92ec0c905e98d4777273cbcc667fade3db5df9a3c3

data/README.md

@@ -105,6 +105,31 @@ params = prompt.build(
 # => { model: "claude-sonnet-4", messages: [...], temperature: 0.3, max_tokens: 2048 }
 ```
 
+### Reasoning Effort
+
+Reasoning models take an effort tier rather than a temperature. Set it like any
+other model param, from the admin form or a prompt `.md`:
+
+```yaml
+model: gpt-5.6-terra
+model_params:
+  reasoning_effort: low
+```
+
+Values are validated against `LlmLogs.reasoning_effort_options`, which defaults to
+`%w[none low medium high xhigh max]` — the union of what current providers accept.
+Narrow it if your app targets a single provider (Anthropic has no `none` tier):
+
+```ruby
+LlmLogs.setup do |config|
+  config.reasoning_effort_options = %w[low medium high]
+end
+```
+
+Leaving the field blank stores nothing, so the model's own default applies.
+Applying the value to a request is the caller's job — the gem stores and validates
+it, it does not build your chat object.
+
 ### Versioning
 
 Every save creates a new version. Previous versions are never modified.

data/app/controllers/llm_logs/prompts_controller.rb

@@ -26,14 +26,14 @@ module LlmLogs
     def create
       @prompt = Prompt.new(prompt_params)
 
-      if @prompt.save
-        if version_params[:messages].present?
-          @prompt.update_content!(**version_params)
-        end
-        redirect_to prompt_path(@prompt), notice: "Prompt created."
-      else
-        render :new, status: :unprocessable_entity
+      ActiveRecord::Base.transaction do
+        @prompt.save!
+        @prompt.update_content!(**version_params) if version_params[:messages].present?
       end
+      redirect_to prompt_path(@prompt), notice: "Prompt created."
+    rescue ActiveRecord::RecordInvalid => e
+      surface_version_errors(e)
+      render :new, status: :unprocessable_entity
     end
 
     def edit
@@ -44,14 +44,15 @@ module LlmLogs
     def update
       @prompt = Prompt.find(params[:id])
 
-      if @prompt.update(prompt_params)
-        if version_params[:messages].present?
-          @prompt.update_content!(**version_params)
-        end
-        redirect_to prompt_path(@prompt), notice: "Prompt updated."
-      else
-        render :edit, status: :unprocessable_entity
+      ActiveRecord::Base.transaction do
+        @prompt.update!(prompt_params)
+        @prompt.update_content!(**version_params) if version_params[:messages].present?
       end
+      redirect_to prompt_path(@prompt), notice: "Prompt updated."
+    rescue ActiveRecord::RecordInvalid => e
+      @current_version = @prompt.current_version
+      surface_version_errors(e)
+      render :edit, status: :unprocessable_entity
     end
 
     def destroy
@@ -70,6 +71,12 @@ module LlmLogs
       raw.except(:tags_input)
     end
 
+    def surface_version_errors(error)
+      return if error.record == @prompt
+
+      error.record.errors.full_messages.each { |message| @prompt.errors.add(:base, message) }
+    end
+
     def version_params
       raw = params.require(:prompt).permit(:model, :changelog, model_params: {})
       messages = parse_messages

data/app/models/llm_logs/batch/adapters/bedrock.rb

@@ -96,10 +96,17 @@ module LlmLogs
             "messages" => [{"role" => "user", "content" => payload["input"]}]
           }
           body["system"] = payload["instructions"] if payload["instructions"]
+          body["temperature"] = payload["temperature"] if payload["temperature"]
+          body.merge!(adaptive_thinking(payload["reasoning_effort"])) if payload["reasoning_effort"].present?
           body.merge!(structured_output(payload["schema"])) if payload["schema"]
           body
         end
 
+        # Anthropic takes an effort tier as adaptive thinking rather than a bare param.
+        def adaptive_thinking(effort)
+          {"thinking" => {"type" => "adaptive"}, "output_config" => {"effort" => effort}}
+        end
+
         # Anthropic structured output via a single forced tool (verify encoding against QA).
         def structured_output(schema)
           spec = schema["schema"] || schema

data/app/models/llm_logs/batch/adapters/openai_responses.rb

@@ -15,6 +15,7 @@ module LlmLogs
               id: request.custom_id,
               instructions: payload["instructions"],
               temperature: payload["temperature"],
+              **reasoning_extra(payload["reasoning_effort"]),
               **schema_extra(payload["schema"])
             )
           end
@@ -36,6 +37,14 @@ module LlmLogs
 
         private
 
+        # The Responses API nests effort under `reasoning:`; `reasoning_effort` is the
+        # Chat Completions spelling and is silently ignored here.
+        def reasoning_extra(effort)
+          return {} if effort.blank?
+
+          {reasoning: {effort: effort}}
+        end
+
         # No memoization: LlmLogs.batch_adapters holds one shared instance of this adapter
         # for the life of the process, so caching the resumed handle here would go stale
         # for a long-running PollJob worker (in-progress batches would never re-fetch a

data/app/models/llm_logs/batch.rb

@@ -18,7 +18,7 @@ module LlmLogs
     scope :recent, -> { order(created_at: :desc) }
     scope :unreconciled, -> { where.not(status: %i[reconciled failed expired]) }
 
-    def self.enqueue(purpose:, model:, input:, instructions:, schema:, routing:, temperature: nil)
+    def self.enqueue(purpose:, model:, input:, instructions:, schema:, routing:, temperature: nil, reasoning_effort: nil)
       BatchRequest.create!(
         purpose: purpose,
         model: model,
@@ -29,7 +29,8 @@ module LlmLogs
           "input" => input,
           "instructions" => instructions,
           "schema" => schema,
-          "temperature" => temperature
+          "temperature" => temperature,
+          "reasoning_effort" => reasoning_effort
         }.compact
       )
     end

data/app/models/llm_logs/prompt_version.rb

@@ -7,6 +7,13 @@ module LlmLogs
 
     validates :version_number, presence: true, uniqueness: { scope: :prompt_id }
     validates :messages, presence: true
+    validate :reasoning_effort_is_supported
+
+    def reasoning_effort
+      return nil unless model_params.is_a?(Hash)
+
+      model_params["reasoning_effort"] || model_params[:reasoning_effort]
+    end
 
     def variables
       messages.flat_map { |msg| msg["content"].to_s.scan(/\{\{[#^]?([^\/}]+)\}\}/) }.flatten.uniq.sort
@@ -27,5 +34,18 @@ module LlmLogs
       params.merge!(model_params.symbolize_keys) if model_params.present?
       params
     end
+
+    private
+
+    # Providers reject an unknown effort tier at request time; catching it here means
+    # a typo in the admin form or a prompt .md fails loudly at write time instead.
+    def reasoning_effort_is_supported
+      return if reasoning_effort.blank?
+
+      allowed = LlmLogs.reasoning_effort_options
+      return if allowed.include?(reasoning_effort.to_s)
+
+      errors.add(:model_params, "reasoning_effort #{reasoning_effort.inspect} is not supported (expected one of: #{allowed.join(", ")})")
+    end
   end
 end

data/app/views/llm_logs/prompts/_form.html.erb

@@ -1,4 +1,12 @@
 <%= form_with(model: prompt, url: prompt.persisted? ? prompt_path(prompt) : prompts_path, class: "space-y-6") do |f| %>
+  <% if prompt.errors[:base].any? %>
+    <div class="rounded-md bg-red-50 p-3 ring-1 ring-red-200">
+      <% prompt.errors[:base].each do |error| %>
+        <p class="text-sm text-red-700"><%= error %></p>
+      <% end %>
+    </div>
+  <% end %>
+
   <div class="bg-white rounded-lg shadow-sm ring-1 ring-gray-900/5 p-6 space-y-4">
     <h2 class="text-lg font-medium text-gray-900">Prompt Details</h2>
 
@@ -83,6 +91,17 @@
           value="<%= current_version&.model_params&.dig('max_tokens') %>"
           class="w-full rounded-md border-gray-300 shadow-sm text-sm px-3 py-2 border">
       </div>
+      <div>
+        <label class="block text-sm font-medium text-gray-700 mb-1">Reasoning Effort</label>
+        <select name="prompt[model_params][reasoning_effort]"
+          class="w-full rounded-md border-gray-300 shadow-sm text-sm px-3 py-2 border">
+          <%= options_for_select(
+                [["Model default", ""]] + LlmLogs.reasoning_effort_options.map { |o| [o, o] },
+                current_version&.model_params&.dig("reasoning_effort").to_s
+              ) %>
+        </select>
+        <p class="mt-1 text-xs text-gray-500">Only applies to reasoning models. Leave as default to send nothing.</p>
+      </div>
     </div>
   </div>
 

data/app/views/llm_logs/prompts/show.html.erb

@@ -44,22 +44,6 @@
         </div>
       </div>
 
-      <% if @current_version.model_params.present? && @current_version.model_params.any? %>
-        <div class="bg-white rounded-lg shadow-sm ring-1 ring-gray-900/5">
-          <div class="px-4 py-3 border-b border-gray-200">
-            <h2 class="text-sm font-medium text-gray-900">Model Parameters</h2>
-          </div>
-          <div class="p-4">
-            <dl class="grid grid-cols-2 gap-2 text-sm">
-              <% @current_version.model_params.each do |key, value| %>
-                <dt class="text-gray-500"><%= key %></dt>
-                <dd class="text-gray-900 font-mono"><%= value %></dd>
-              <% end %>
-            </dl>
-          </div>
-        </div>
-      <% end %>
-
       <% if @current_version.default_variables.present? && @current_version.default_variables.any? %>
         <div class="bg-white rounded-lg shadow-sm ring-1 ring-gray-900/5">
           <div class="px-4 py-3 border-b border-gray-200">
@@ -101,6 +85,22 @@
       </div>
     </div>
 
+    <% if @current_version&.model_params.present? %>
+      <div class="mt-4 bg-white rounded-lg shadow-sm ring-1 ring-gray-900/5">
+        <div class="px-4 py-3 border-b border-gray-200">
+          <h2 class="text-sm font-medium text-gray-900">Model Parameters</h2>
+        </div>
+        <div class="p-4">
+          <dl class="grid grid-cols-2 gap-2 text-sm">
+            <% @current_version.model_params.each do |key, value| %>
+              <dt class="text-gray-500"><%= key %></dt>
+              <dd class="text-gray-900 font-mono text-right"><%= value %></dd>
+            <% end %>
+          </dl>
+        </div>
+      </div>
+    <% end %>
+
     <details class="mt-4 bg-white rounded-lg shadow-sm ring-1 ring-gray-900/5">
       <summary class="cursor-pointer px-4 py-3 text-sm font-medium text-gray-900">SDK Usage</summary>
       <div class="px-4 pb-4">

data/lib/llm_logs/configuration.rb

@@ -3,7 +3,7 @@ module LlmLogs
     BedrockBatch = Struct.new(:role_arn, :s3_bucket, :s3_prefix, :min_records, :model_matcher, :region, keyword_init: true)
 
     attr_accessor :enabled, :auto_instrument, :retention_days, :prompts_source_path, :prompt_subfolders,
-                  :batch_enabled, :batch_provider, :page_size, :bedrock_batch
+                  :batch_enabled, :batch_provider, :page_size, :bedrock_batch, :reasoning_effort_options
 
     def initialize
       @enabled             = true
@@ -15,6 +15,10 @@ module LlmLogs
       @batch_provider      = :openai_responses
       @page_size           = 50
       @bedrock_batch       = nil
+      # Effort tiers offered in the prompt form. Union of what current providers
+      # accept -- OpenAI takes all six, Anthropic has no "none". Narrow this in an
+      # initializer if your app targets one provider.
+      @reasoning_effort_options = %w[none low medium high xhigh max]
     end
   end
 

data/lib/llm_logs/version.rb

@@ -1,3 +1,3 @@
 module LlmLogs
-  VERSION = "0.3.1"
+  VERSION = "0.4.0"
 end

data/lib/llm_logs.rb

@@ -46,6 +46,10 @@ module LlmLogs
     configuration.page_size = page_size
   end
 
+  def self.reasoning_effort_options
+    configuration.reasoning_effort_options
+  end
+
   def self.batch_enabled?
     configuration.batch_enabled
   end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: llm_logs
 version: !ruby/object:Gem::Version
-  version: 0.3.1
+  version: 0.4.0
 platform: ruby
 authors:
 - Anton