feat(ai): add Anthropic provider with chat parity (1/5)

Gemfile

@@ -101,6 +101,7 @@ gem "after_commit_everywhere", "~> 1.0"
 
 # AI
 gem "ruby-openai"
+gem "anthropic", "~> 1.0"
 gem "langfuse-ruby", "~> 0.1.4", require: "langfuse"
 
 group :development, :test do

Gemfile.lock

@@ -87,6 +87,10 @@ GEM
       activerecord (>= 4.2)
       activesupport
     android_key_attestation (0.3.0)
+    anthropic (1.43.0)
+      cgi
+      connection_pool
+      standardwebhooks
     ast (2.4.3)
     attr_required (1.0.2)
     aws-eventstream (1.4.0)
@@ -759,6 +763,7 @@ GEM
       faraday (>= 1.0.1, < 3.0)
       faraday-multipart (~> 1.0, >= 1.0.4)
     stackprof (0.2.27)
+    standardwebhooks (1.1.0)
     stimulus-rails (1.3.4)
       railties (>= 6.0.0)
     stringio (3.1.7)
@@ -859,6 +864,7 @@ DEPENDENCIES
   aasm
   activerecord-import
   after_commit_everywhere (~> 1.0)
+  anthropic (~> 1.0)
   aws-sdk-s3 (~> 1.208.0)
   bcrypt (~> 3.1)
   benchmark-ips

app/models/assistant/builtin.rb

@@ -60,7 +60,7 @@ def respond_to(message, assistant_message: nil)
       if assistant_message.content.blank?
         assistant_message.destroy
       else
-        # Demote partially-streamed turns to `failed` so `Responder#conversation_history` excludes them.
+        # Demote partially-streamed turns to `failed` so the responder's history builders (`#openai_messages_payload`, `#chat_message_records`) exclude them.
         assistant_message.update_columns(status: "failed")
       end
     end

app/models/assistant/responder.rb

@@ -79,7 +79,8 @@ def get_llm_response(streamer:, function_results: [], previous_response_id: nil)
         instructions: instructions,
         functions: function_tool_caller.function_definitions,
         function_results: function_results,
-        messages: conversation_history,
+        messages: openai_messages_payload,
+        conversation_history: chat_message_records,
         streamer: streamer,
         previous_response_id: previous_response_id,
         session_id: chat_session_id,
@@ -116,15 +117,37 @@ def chat
       @chat ||= message.chat
     end
 
-    def conversation_history
-      messages = []
-      return messages unless chat&.messages
+    # Memoized fetch — both `chat_message_records` and `openai_messages_payload`
+    # derive their shape from this one in-memory array so a single chat turn
+    # fires one history query instead of two.
+    def complete_chat_messages
+      return @complete_chat_messages if defined?(@complete_chat_messages)
+
+      @complete_chat_messages =
+        if chat&.messages
+          chat.messages
+              .where(type: [ "UserMessage", "AssistantMessage" ], status: "complete")
+              .includes(:tool_calls)
+              .ordered
+              .to_a
+        else
+          []
+        end
+    end
 
-      chat.messages
-          .where(type: [ "UserMessage", "AssistantMessage" ], status: "complete")
-          .includes(:tool_calls)
-          .ordered
-          .each do |chat_message|
+    # Raw Message records preceding the current turn — providers that build
+    # their own native message shape (Anthropic) consume this directly so they
+    # do not have to round-trip through the OpenAI-shaped payload below.
+    def chat_message_records
+      complete_chat_messages.reject { |m| m.id == message.id }
+    end
+
+    # Builds the OpenAI-shaped messages payload (role: "user" | "assistant" |
+    # "tool"; tool_call_id pairing) consumed by Provider::Openai's generic
+    # chat path. Anthropic uses chat_message_records instead.
+    def openai_messages_payload
+      messages = []
+      complete_chat_messages.each do |chat_message|
         if chat_message.tool_calls.any?
           messages << {
             role: chat_message.role,

app/models/chat.rb

@@ -51,10 +51,24 @@ def generate_title(prompt)
       prompt.first(80)
     end
 
-    # Returns the default AI model to use for chats
-    # Priority: AI Config > Setting
+    # Returns the default AI model to use for chats.
+    # Resolved from the configured llm_provider so installs that swap providers
+    # don't have to manually update every chat default. Falls through to a
+    # provider that actually has credentials configured, otherwise the chosen
+    # provider's classes would later raise "no LLM provider supports model …"
+    # even when the other provider is configured.
     def default_model
-      Provider::Openai.effective_model.presence || Setting.openai_model
+      prefers_anthropic = Setting.llm_provider == "anthropic"
+
+      if prefers_anthropic && Provider::Anthropic.configured?
+        Provider::Anthropic.effective_model.presence || Setting.anthropic_model
+      elsif Provider::Openai.configured?
+        Provider::Openai.effective_model.presence || Setting.openai_model
+      elsif Provider::Anthropic.configured?
+        Provider::Anthropic.effective_model.presence || Setting.anthropic_model
+      else
+        Provider::Openai.effective_model.presence || Setting.openai_model
+      end
     end
   end