llm.rb 11.1.0 → 11.2.0

data/lib/llm/providers/bedrock/signature.rb

@@ -8,7 +8,7 @@ class LLM::Bedrock
   # Signs HTTP requests and headers with AWS Signature V4.
   #
   # Returns the signed headers as a Hash through #to_h, ready to merge
-  # into a Net::HTTPRequest or other HTTP client. Everything else is
+  # into an {LLM::Transport::Request} or other HTTP client. Everything else is
   # private.
   #
   # Uses only Ruby's stdlib (openssl, digest) with no external deps.
@@ -89,8 +89,8 @@ class LLM::Bedrock
     end
 
     ##
-    # @param [Net::HTTPRequest] req
-    # @return [Net::HTTPRequest]
+    # @param [LLM::Transport::Request] req
+    # @return [LLM::Transport::Request]
     def sign!(req)
       to_h.each { |k, v| req[k] = v }
       req

data/lib/llm/providers/bedrock.rb

@@ -217,7 +217,7 @@ module LLM
       body = LLM.json.dump(payload)
       path = stream ? "/model/#{model_id}/converse-stream" \
                     : "/model/#{model_id}/converse"
-      req = Net::HTTP::Post.new(path, headers)
+      req = LLM::Transport::Request.post(path, headers)
       transport.set_body_stream(req, StringIO.new(body))
       [req, messages, body]
     end

data/lib/llm/providers/google/files.rb

@@ -45,7 +45,7 @@ class LLM::Google
     # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params.merge!(key: key))
-      req = Net::HTTP::Get.new("/v1beta/files?#{query}", headers)
+      req = LLM::Transport::Request.get("/v1beta/files?#{query}", headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :files)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -64,7 +64,7 @@ class LLM::Google
     # @return [LLM::Response]
     def create(file:, **params)
       file = LLM.File(file)
-      req = Net::HTTP::Post.new(request_upload_url(file:), {})
+      req = LLM::Transport::Request.post(request_upload_url(file:), {})
       req["content-length"] = file.bytesize
       req["X-Goog-Upload-Offset"] = 0
       req["X-Goog-Upload-Command"] = "upload, finalize"
@@ -91,7 +91,7 @@ class LLM::Google
     def get(file:, **params)
       file_id = file.respond_to?(:name) ? file.name : file.to_s
       query = URI.encode_www_form(params.merge!(key: key))
-      req = Net::HTTP::Get.new("/v1beta/#{file_id}?#{query}", headers)
+      req = LLM::Transport::Request.get("/v1beta/#{file_id}?#{query}", headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :file)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -111,7 +111,7 @@ class LLM::Google
     def delete(file:, **params)
       file_id = file.respond_to?(:name) ? file.name : file.to_s
       query = URI.encode_www_form(params.merge!(key: key))
-      req = Net::HTTP::Delete.new("/v1beta/#{file_id}?#{query}", headers)
+      req = LLM::Transport::Request.delete("/v1beta/#{file_id}?#{query}", headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -128,7 +128,7 @@ class LLM::Google
     private
 
     def request_upload_url(file:)
-      req = Net::HTTP::Post.new("/upload/v1beta/files?key=#{key}", headers)
+      req = LLM::Transport::Request.post("/upload/v1beta/files?key=#{key}", headers)
       req["X-Goog-Upload-Protocol"] = "resumable"
       req["X-Goog-Upload-Command"] = "start"
       req["X-Goog-Upload-Header-Content-Length"] = file.bytesize

data/lib/llm/providers/google/images.rb

@@ -40,7 +40,7 @@ class LLM::Google
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response]
     def create(prompt:, n: 1, image_size: nil, aspect_ratio: nil, person_generation: nil, model: "imagen-4.0-generate-001", **params)
-      req  = Net::HTTP::Post.new("/v1beta/models/#{model}:predict?key=#{key}", headers)
+      req  = LLM::Transport::Request.post("/v1beta/models/#{model}:predict?key=#{key}", headers)
       body = LLM.json.dump({
         parameters: {
           sampleCount: n,

data/lib/llm/providers/google/models.rb

@@ -39,7 +39,7 @@ class LLM::Google
     # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params.merge!(key: key))
-      req = Net::HTTP::Get.new("/v1beta/models?#{query}", headers)
+      req = LLM::Transport::Request.get("/v1beta/models?#{query}", headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :models)
       tracer.on_request_finish(operation: "request", res:, span:)

data/lib/llm/providers/google.rb

@@ -56,7 +56,7 @@ module LLM
     def embed(input, model: "gemini-embedding-001", **params)
       model = model.respond_to?(:id) ? model.id : model
       path = ["/v1beta/models/#{model}", "embedContent?key=#{@key}"].join(":")
-      req = Net::HTTP::Post.new(path, headers)
+      req = LLM::Transport::Request.post(path, headers)
       req.body = LLM.json.dump({content: {parts: [{text: input}]}})
       res, span, tracer = execute(request: req, operation: "embeddings", model:)
       res = ResponseAdapter.adapt(res, type: :embedding)
@@ -205,7 +205,7 @@ module LLM
       action = stream ? "streamGenerateContent?key=#{@key}&alt=sse" : "generateContent?key=#{@key}"
       model.respond_to?(:id) ? model.id : model
       path = ["/v1beta/models/#{model}", action].join(":")
-      req  = Net::HTTP::Post.new(path, headers)
+      req  = LLM::Transport::Request.post(path, headers)
       messages = build_complete_messages(prompt, params, role)
       body = LLM.json.dump({contents: adapt(messages)}.merge!(params))
       transport.set_body_stream(req, StringIO.new(body))

data/lib/llm/providers/ollama/models.rb

@@ -40,7 +40,7 @@ class LLM::Ollama
     # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new("/api/tags?#{query}", headers)
+      req = LLM::Transport::Request.get("/api/tags?#{query}", headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :models)
       tracer.on_request_finish(operation: "request", res:, span:)

data/lib/llm/providers/ollama.rb

@@ -48,7 +48,7 @@ module LLM
     # @return [LLM::Response]
     def embed(input, model: default_model, **params)
       params   = {model:}.merge!(params)
-      req      = Net::HTTP::Post.new("/v1/embeddings", headers)
+      req      = LLM::Transport::Request.post("/v1/embeddings", headers)
       req.body = LLM.json.dump({input:}.merge!(params))
       res, span, tracer = execute(request: req, operation: "embeddings", model:)
       res = ResponseAdapter.adapt(res, type: :embedding)
@@ -129,7 +129,7 @@ module LLM
     def build_complete_request(prompt, params, role)
       messages = build_complete_messages(prompt, params, role)
       body = LLM.json.dump({messages: [adapt(messages)].flatten}.merge!(params))
-      req = Net::HTTP::Post.new("/api/chat", headers)
+      req = LLM::Transport::Request.post("/api/chat", headers)
       transport.set_body_stream(req, StringIO.new(body))
       req
     end

data/lib/llm/providers/openai/audio.rb

@@ -32,7 +32,7 @@ class LLM::OpenAI
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response]
     def create_speech(input:, voice: "alloy", model: "gpt-4o-mini-tts", response_format: "mp3", **params)
-      req = Net::HTTP::Post.new(path("/audio/speech"), headers)
+      req = LLM::Transport::Request.post(path("/audio/speech"), headers)
       req.body = LLM.json.dump({input:, voice:, model:, response_format:}.merge!(params))
       io = StringIO.new("".b)
       res, span, tracer = execute(request: req, operation: "request") { _1.read_body { |chunk| io << chunk } }
@@ -55,7 +55,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     def create_transcription(file:, model: "whisper-1", **params)
       multi = LLM::Multipart.new(params.merge!(file: LLM.File(file), model:))
-      req = Net::HTTP::Post.new(path("/audio/transcriptions"), headers)
+      req = LLM::Transport::Request.post(path("/audio/transcriptions"), headers)
       req["content-type"] = multi.content_type
       transport.set_body_stream(req, multi.body)
       res, span, tracer = execute(request: req, operation: "request")
@@ -79,7 +79,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     def create_translation(file:, model: "whisper-1", **params)
       multi = LLM::Multipart.new(params.merge!(file: LLM.File(file), model:))
-      req = Net::HTTP::Post.new(path("/audio/translations"), headers)
+      req = LLM::Transport::Request.post(path("/audio/translations"), headers)
       req["content-type"] = multi.content_type
       transport.set_body_stream(req, multi.body)
       res, span, tracer = execute(request: req, operation: "request")

data/lib/llm/providers/openai/files.rb

@@ -40,7 +40,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new(path("/files?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/files?#{query}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :enumerable)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -60,7 +60,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     def create(file:, purpose: "assistants", **params)
       multi = LLM::Multipart.new(params.merge!(file: LLM.File(file), purpose:))
-      req = Net::HTTP::Post.new(path("/files"), headers)
+      req = LLM::Transport::Request.post(path("/files"), headers)
       req["content-type"] = multi.content_type
       transport.set_body_stream(req, multi.body)
       res, span, tracer = execute(request: req, operation: "request")
@@ -83,7 +83,7 @@ class LLM::OpenAI
     def get(file:, **params)
       file_id = file.respond_to?(:id) ? file.id : file
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new(path("/files/#{file_id}?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/files/#{file_id}?#{query}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :file)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -105,7 +105,7 @@ class LLM::OpenAI
     def download(file:, **params)
       query = URI.encode_www_form(params)
       file_id = file.respond_to?(:id) ? file.id : file
-      req = Net::HTTP::Get.new(path("/files/#{file_id}/content?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/files/#{file_id}/content?#{query}"), headers)
       io = StringIO.new("".b)
       res, span, tracer = execute(request: req, operation: "request") { |res| res.read_body { |chunk| io << chunk } }
       res = LLM::Response.new(res).tap { _1.define_singleton_method(:file) { io } }
@@ -125,7 +125,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     def delete(file:)
       file_id = file.respond_to?(:id) ? file.id : file
-      req = Net::HTTP::Delete.new(path("/files/#{file_id}"), headers)
+      req = LLM::Transport::Request.delete(path("/files/#{file_id}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
       tracer.on_request_finish(operation: "request", res:, span:)

data/lib/llm/providers/openai/images.rb

@@ -50,7 +50,7 @@ class LLM::OpenAI
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response]
     def create(prompt:, model: "dall-e-3", response_format: "b64_json", **params)
-      req = Net::HTTP::Post.new(path("/images/generations"), headers)
+      req = LLM::Transport::Request.post(path("/images/generations"), headers)
       req.body = LLM.json.dump({prompt:, n: 1, model:, response_format:}.merge!(params))
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :image)
@@ -76,7 +76,7 @@ class LLM::OpenAI
     def create_variation(image:, model: "dall-e-2", response_format: "b64_json", **params)
       image = LLM.File(image)
       multi = LLM::Multipart.new(params.merge!(image:, model:, response_format:))
-      req = Net::HTTP::Post.new(path("/images/variations"), headers)
+      req = LLM::Transport::Request.post(path("/images/variations"), headers)
       req["content-type"] = multi.content_type
       transport.set_body_stream(req, multi.body)
       res, span, tracer = execute(request: req, operation: "request")
@@ -102,7 +102,7 @@ class LLM::OpenAI
     def edit(image:, prompt:, model: "dall-e-2", response_format: "b64_json", **params)
       image = LLM.File(image)
       multi = LLM::Multipart.new(params.merge!(image:, prompt:, model:, response_format:))
-      req = Net::HTTP::Post.new(path("/images/edits"), headers)
+      req = LLM::Transport::Request.post(path("/images/edits"), headers)
       req["content-type"] = multi.content_type
       transport.set_body_stream(req, multi.body)
       res, span, tracer = execute(request: req, operation: "request")

data/lib/llm/providers/openai/models.rb

@@ -39,7 +39,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new(path("/models?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/models?#{query}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :models)
       tracer.on_request_finish(operation: "request", res:, span:)

data/lib/llm/providers/openai/moderations.rb

@@ -47,7 +47,7 @@ class LLM::OpenAI
     # @param [String, LLM::Model] model The model to use
     # @return [LLM::Response]
     def create(input:, model: "omni-moderation-latest", **params)
-      req = Net::HTTP::Post.new(path("/moderations"), headers)
+      req = LLM::Transport::Request.post(path("/moderations"), headers)
       input = RequestAdapter::Moderation.new(input).adapt
       req.body = LLM.json.dump({input:, model:}.merge!(params))
       res, span, tracer = execute(request: req, operation: "request")

data/lib/llm/providers/openai/responses.rb

@@ -40,7 +40,7 @@ class LLM::OpenAI
       params = [params, adapt_schema(params), adapt_tools(tools)].inject({}, &:merge!).compact
       role, stream = params.delete(:role), params.delete(:stream)
       params[:stream] = true if @provider.streamable?(stream) || stream == true
-      req = Net::HTTP::Post.new(path("/responses"), headers)
+      req = LLM::Transport::Request.post(path("/responses"), headers)
       messages = build_complete_messages(prompt, params, role)
       @provider.tracer.set_request_metadata(user_input: extract_user_input(messages, fallback: prompt))
       body = LLM.json.dump({input: [adapt(messages, mode: :response)].flatten}.merge!(params))
@@ -61,7 +61,7 @@ class LLM::OpenAI
     def get(response, **params)
       response_id = response.respond_to?(:id) ? response.id : response
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new(path("/responses/#{response_id}?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/responses/#{response_id}?#{query}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :responds)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -76,7 +76,7 @@ class LLM::OpenAI
     # @return [LLM::Object] Response body
     def delete(response)
       response_id = response.respond_to?(:id) ? response.id : response
-      req = Net::HTTP::Delete.new(path("/responses/#{response_id}"), headers)
+      req = LLM::Transport::Request.delete(path("/responses/#{response_id}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
       tracer.on_request_finish(operation: "request", res:, span:)

data/lib/llm/providers/openai/vector_stores.rb

@@ -31,7 +31,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new(path("/vector_stores?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/vector_stores?#{query}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :enumerable)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -47,7 +47,7 @@ class LLM::OpenAI
     # @return [LLM::Response]
     # @see https://platform.openai.com/docs/api-reference/vector_stores/create OpenAI docs
     def create(name:, file_ids: nil, **params)
-      req = Net::HTTP::Post.new(path("/vector_stores"), headers)
+      req = LLM::Transport::Request.post(path("/vector_stores"), headers)
       req.body = LLM.json.dump(params.merge({name:, file_ids:}).compact)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
@@ -72,7 +72,7 @@ class LLM::OpenAI
     # @see https://platform.openai.com/docs/api-reference/vector_stores/retrieve OpenAI docs
     def get(vector:)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
-      req = Net::HTTP::Get.new(path("/vector_stores/#{vector_id}"), headers)
+      req = LLM::Transport::Request.get(path("/vector_stores/#{vector_id}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -89,7 +89,7 @@ class LLM::OpenAI
     # @see https://platform.openai.com/docs/api-reference/vector_stores/modify OpenAI docs
     def modify(vector:, name: nil, **params)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
-      req = Net::HTTP::Post.new(path("/vector_stores/#{vector_id}"), headers)
+      req = LLM::Transport::Request.post(path("/vector_stores/#{vector_id}"), headers)
       req.body = LLM.json.dump(params.merge({name:}).compact)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
@@ -105,7 +105,7 @@ class LLM::OpenAI
     # @see https://platform.openai.com/docs/api-reference/vector_stores/delete OpenAI docs
     def delete(vector:)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
-      req = Net::HTTP::Delete.new(path("/vector_stores/#{vector_id}"), headers)
+      req = LLM::Transport::Request.delete(path("/vector_stores/#{vector_id}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -122,7 +122,7 @@ class LLM::OpenAI
     # @see https://platform.openai.com/docs/api-reference/vector_stores/search OpenAI docs
     def search(vector:, query:, **params)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
-      req = Net::HTTP::Post.new(path("/vector_stores/#{vector_id}/search"), headers)
+      req = LLM::Transport::Request.post(path("/vector_stores/#{vector_id}/search"), headers)
       req.body = LLM.json.dump(params.merge({query:}).compact)
       res, span, tracer = execute(request: req, operation: "retrieval")
       res = ResponseAdapter.adapt(res, type: :enumerable)
@@ -140,7 +140,7 @@ class LLM::OpenAI
     def all_files(vector:, **params)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new(path("/vector_stores/#{vector_id}/files?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/vector_stores/#{vector_id}/files?#{query}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :enumerable)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -159,7 +159,7 @@ class LLM::OpenAI
     def add_file(vector:, file:, attributes: nil, **params)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
       file_id = file.respond_to?(:id) ? file.id : file
-      req = Net::HTTP::Post.new(path("/vector_stores/#{vector_id}/files"), headers)
+      req = LLM::Transport::Request.post(path("/vector_stores/#{vector_id}/files"), headers)
       req.body = LLM.json.dump(params.merge({file_id:, attributes:}).compact)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
@@ -190,7 +190,7 @@ class LLM::OpenAI
     def update_file(vector:, file:, attributes:, **params)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
       file_id = file.respond_to?(:id) ? file.id : file
-      req = Net::HTTP::Post.new(path("/vector_stores/#{vector_id}/files/#{file_id}"), headers)
+      req = LLM::Transport::Request.post(path("/vector_stores/#{vector_id}/files/#{file_id}"), headers)
       req.body = LLM.json.dump(params.merge({attributes:}).compact)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
@@ -209,7 +209,7 @@ class LLM::OpenAI
       vector_id = vector.respond_to?(:id) ? vector.id : vector
       file_id = file.respond_to?(:id) ? file.id : file
       query = URI.encode_www_form(params)
-      req = Net::HTTP::Get.new(path("/vector_stores/#{vector_id}/files/#{file_id}?#{query}"), headers)
+      req = LLM::Transport::Request.get(path("/vector_stores/#{vector_id}/files/#{file_id}?#{query}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
       tracer.on_request_finish(operation: "request", res:, span:)
@@ -226,7 +226,7 @@ class LLM::OpenAI
     def delete_file(vector:, file:)
       vector_id = vector.respond_to?(:id) ? vector.id : vector
       file_id = file.respond_to?(:id) ? file.id : file
-      req = Net::HTTP::Delete.new(path("/vector_stores/#{vector_id}/files/#{file_id}"), headers)
+      req = LLM::Transport::Request.delete(path("/vector_stores/#{vector_id}/files/#{file_id}"), headers)
       res, span, tracer = execute(request: req, operation: "request")
       res = LLM::Response.new(res)
       tracer.on_request_finish(operation: "request", res:, span:)

data/lib/llm/providers/openai.rb

@@ -52,7 +52,7 @@ module LLM
     # @raise (see LLM::Provider#request)
     # @return (see LLM::Provider#embed)
     def embed(input, model: "text-embedding-3-small", **params)
-      req = Net::HTTP::Post.new(path("/embeddings"), headers)
+      req = LLM::Transport::Request.post(path("/embeddings"), headers)
       req.body = LLM.json.dump({input:, model:}.merge!(params))
       res, span, tracer = execute(request: req, operation: "embeddings", model:)
       res = ResponseAdapter.adapt(res, type: :embedding)
@@ -222,7 +222,7 @@ module LLM
     def build_complete_request(prompt, params, role)
       messages = build_complete_messages(prompt, params, role)
       body = LLM.json.dump({messages: adapt(messages, mode: :complete).flatten}.merge!(params))
-      req = Net::HTTP::Post.new(completions_path, headers)
+      req = LLM::Transport::Request.post(completions_path, headers)
       transport.set_body_stream(req, StringIO.new(body))
       [req, messages]
     end

data/lib/llm/skill.rb

@@ -159,7 +159,7 @@ module LLM
       params[:concurrency] = concurrency if concurrency
       agent = Class.new(LLM::Agent) do
         instructions(instructions)
-        tools(inherit_tools ? ctx.params[:tools] : tools)
+        tools(inherit_tools ? [*ctx.params[:tools]].reject(&:skill?) : tools)
         tracer(tracer)
       end.new(ctx.llm, params)
       agent.messages.concat(messages_for(ctx))

data/lib/llm/tool.rb

@@ -202,6 +202,13 @@ class LLM::Tool
     false
   end
 
+  ##
+  # Returns true if the tool is a skill
+  # @return [Boolean]
+  def self.skill?
+    false
+  end
+
   ##
   # Returns a function bound to this tool instance.
   # @return [LLM::Function]
@@ -216,6 +223,20 @@ class LLM::Tool
     self.class.mcp?
   end
 
+  ##
+  # Returns true if the tool is an A2A tool
+  # @return [Boolean]
+  def a2a?
+    self.class.a2a?
+  end
+
+  ##
+  # Returns true if the tool is a skill
+  # @return [Boolean]
+  def skill?
+    self.class.skill?
+  end
+
   ##
   # Called when an in-flight tool run is interrupted.
   # Tools can override this to implement cooperative cleanup.