llm_meta_client 1.6.0 → 1.7.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: a1afc833ffdebd20e85c17d3d86ba6f89d8da0ac893bf0e19e33b0071e5be7de
-  data.tar.gz: be3d6e7dd8809c5b084e9c5cc714d6a5bf969dc70ae097a4d79cea079e63f7a8
+  metadata.gz: f5b2078839265ff4e964a4437b2d4865c4fcb9d520d3adee38160b04716dceed
+  data.tar.gz: f70e94937945bfc7bb2d51e65d3c229055e434ea130e3181760c4c12c9281e41
 SHA512:
-  metadata.gz: 8844927644808f81a7eabda2d5d19e7b65674545ed8d65fa3d336b6b7b4261b90e1c5f160eccdd439fd52507b32793305419104382746b9ea6a137d6d01e48eb
-  data.tar.gz: 1ae00410b164fc95a0c9cc06dc59fa4e0730331f95d1c787792ca9dbd8db829bd1d8f0db5d2f1f23c5c2c6dfee89f55d8dce93d200746b8963795c7a1cd377c6
+  metadata.gz: 4ccd30289df04520d3c8be1343bdba1e183204b898b0c763d8a7bc5a70b8b49627bf4c57487786faa32be90c0cd2d38617aee0567214d93fa21f850ddd08ec61
+  data.tar.gz: d16fbce1d748fe0cbe516c0af6b6d8dab825eaec3b362d901591f578cba460a9ed59c07805e054282557d81045a235387c072123e1a17691a1b450ea8c2e00fb

data/CHANGELOG.md

@@ -5,6 +5,12 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.7.0] - 2026-07-06
+
+### Added
+
+- `ServerQuery#stream` and `#call` accept a new `messages:` kwarg — an array of role-tagged prior turns (`[{role: "user"|"assistant"|"system", content: "..."}]`) forwarded to `llm_meta_server` as `messages` in the JSON body. When present, the current-turn text stays in `prompt:` (so image/document attachments still bind to it) and the server pre-seeds `LLM::Session`'s message buffer with the history. Fixes a class of "sticky context" bugs where models re-executed the prior turn's task instead of the new instruction because the whole conversation was being packed into a single user message.
+
 ## [1.6.0] - 2026-07-02
 
 ### Added

data/lib/llm_meta_client/server_query.rb

@@ -12,11 +12,20 @@ module LlmMetaClient
     # Returns the final assistant content. If tool calls fired, the returned
     # string mirrors the synchronous #call format (response + markdown
     # "Tool calls" section appended) so persistence stays consistent.
-    def stream(id_token, api_key_uuid, model_id, context, user_content, tool_ids: [], generation_settings: {}, image_context: nil, image: nil, images: nil, document: nil)
+    def stream(id_token, api_key_uuid, model_id, context, user_content, tool_ids: [], generation_settings: {}, image_context: nil, image: nil, images: nil, document: nil, messages: nil)
       if image_context.present?
         prompt_text = user_content.is_a?(Hash) ? (user_content[:prompt] || user_content["prompt"]).to_s : user_content.to_s
         debug_log "Streaming image request to LLM: \n===>\n#{prompt_text}\n(with #{image_context.size} prior turn(s))\n===>"
         body = { prompt: prompt_text, image_context: image_context }
+      elsif messages.present?
+        # New (v1.7+) role-tagged wire format. Prior turns travel as a
+        # proper `messages:` array; the current user turn's text is sent as
+        # `prompt`. Callers who use this form should NOT include the
+        # current turn in `messages` — it stays in `prompt` so attachments
+        # (image/document) can attach to it via the existing pipeline.
+        prompt_text = user_content.is_a?(Hash) ? (user_content[:prompt] || user_content["prompt"]).to_s : user_content.to_s
+        debug_log "Streaming role-tagged request to LLM: \n===>\n#{prompt_text}\n(with #{messages.size} prior message(s))\n===>"
+        body = { prompt: prompt_text, messages: messages }
       else
         context_and_user_content = "Context:#{context}, User Prompt: #{user_content}"
         debug_log "Streaming request to LLM: \n===>\n#{context_and_user_content}\n===>"
@@ -62,12 +71,18 @@ module LlmMetaClient
       collected_tool_calls.any? ? combine_with_tool_calls(assembled, collected_tool_calls) : assembled
     end
 
-    def call(id_token, api_key_uuid, model_id, context, user_content, tool_ids: [], generation_settings: {}, document: nil)
+    def call(id_token, api_key_uuid, model_id, context, user_content, tool_ids: [], generation_settings: {}, document: nil, messages: nil)
       debug_log "Context: #{context}"
-      context_and_user_content = "Context:#{context}, User Prompt: #{user_content}"
-      debug_log "Request to LLM: \n===>\n#{context_and_user_content}\n===>"
+      if messages.present?
+        prompt_text = user_content.is_a?(Hash) ? (user_content[:prompt] || user_content["prompt"]).to_s : user_content.to_s
+        debug_log "Role-tagged request to LLM: \n===>\n#{prompt_text}\n(with #{messages.size} prior message(s))\n===>"
+        payload = prompt_text
+      else
+        payload = "Context:#{context}, User Prompt: #{user_content}"
+        debug_log "Request to LLM: \n===>\n#{payload}\n===>"
+      end
 
-      response = request(api_key_uuid, id_token, model_id, context_and_user_content, tool_ids, generation_settings, document: document)
+      response = request(api_key_uuid, id_token, model_id, payload, tool_ids, generation_settings, document: document, messages: messages)
 
       unless response.success?
         raise Exceptions::ServerError, build_error_message(response.code.to_i, response.parsed_response)
@@ -116,7 +131,7 @@ module LlmMetaClient
       lines.join("\n")
     end
 
-    def request(api_key_uuid, id_token, model_id, user_content, tool_ids, generation_settings, document: nil)
+    def request(api_key_uuid, id_token, model_id, user_content, tool_ids, generation_settings, document: nil, messages: nil)
       headers = { "Content-Type" => "application/json" }
       headers["Authorization"] = "Bearer #{id_token}" if id_token.present?
 
@@ -124,6 +139,7 @@ module LlmMetaClient
       body[:tool_ids] = tool_ids if tool_ids.present?
       body[:generation_settings] = generation_settings if generation_settings.present?
       body[:document] = document if document.present?
+      body[:messages] = messages if messages.present?
 
       HTTParty.post(
         url(api_key_uuid, model_id),

data/lib/llm_meta_client/version.rb

@@ -1,3 +1,3 @@
 module LlmMetaClient
-  VERSION = "1.6.0"
+  VERSION = "1.7.0"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: llm_meta_client
 version: !ruby/object:Gem::Version
-  version: 1.6.0
+  version: 1.7.0
 platform: ruby
 authors:
 - dhq_boiler