active_harness 0.2.41 → 0.2.42

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e575886969a297ab3688f08a326e0290b1f7c433322c7a05aab8f183178e441e
-  data.tar.gz: 8b73cfb142335117fd7bde7c6da6ebca88f7218906450a1897b2cd42ea9b37d5
+  metadata.gz: 64ee242555ae4c7c33965a04607e3f72c76eb83046f62aaadebf98585866b1a8
+  data.tar.gz: 0ec32b43d65efce4ed82c3a6ab997eb42f03720cd219baf63375fd9d4abb4316
 SHA512:
-  metadata.gz: 51060b54a300332514f5993d83b5666416da70dc51493d06259106909c17a96262d28991290cbcc51b4018f324e82d06f274cd292d81ad65ad9b99a33a82060d
-  data.tar.gz: f4041988b036d9b252c0081e3e664febabaf72cbc69995629f21b6b6711c7a4283a1d9e30fb388f0bbde0a127c940e54cd3a4b325978798a4d090e93ecf639dc
+  metadata.gz: 0cb81649fd8e9a5c8ee2f2836dbda95cf94e0ce2ad5a9f60ed5325bb5666baf0c11009578b449c0b8ba45551dd670d71b78329be71a196423d08137f660d86b3
+  data.tar.gz: 2b1c6257ca5e3fd16337e57a7ab5c018de1534aa98616d6cee1be19b6a767cbd7a0ec03f1bbe578400d183a1d61ae566c6eccc91587cbe276954874633a0d5e3

data/lib/active_harness/agent/providers.rb

@@ -44,6 +44,7 @@ module ActiveHarness
     }.freeze
 
     TRANSCRIPTION_PROVIDERS = {
+      openai:      -> { Providers::Audio::OpenAI.new },
       openrouter:  -> { Providers::Audio::OpenRouter.new }
     }.freeze
 

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

@@ -0,0 +1,127 @@
+require "uri"
+require "securerandom"
+
+module ActiveHarness
+  module Providers
+    module Audio
+      class OpenAI < Base
+        ENDPOINT = "https://api.openai.com/v1/audio/transcriptions"
+
+        # OpenAI's transcription endpoint only accepts these formats — notably
+        # no flac/ogg/aac, unlike OpenRouter's version of this endpoint.
+        CONTENT_TYPES = {
+          "mp3"  => "audio/mpeg",
+          "mp4"  => "audio/mp4",
+          "mpeg" => "audio/mpeg",
+          "mpga" => "audio/mpeg",
+          "m4a"  => "audio/mp4",
+          "wav"  => "audio/wav",
+          "webm" => "audio/webm"
+        }.freeze
+
+        # @param model        [String]  e.g. "whisper-1", "gpt-4o-transcribe", "gpt-4o-mini-transcribe"
+        # @param audio_data   [String]  raw binary audio bytes
+        # @param audio_format [String]  one of CONTENT_TYPES.keys
+        # @param language     [String]  ISO-639-1 code, e.g. "en" (optional — auto-detected if omitted)
+        #
+        # Synchronous — this endpoint has no job/polling API. Unlike OpenRouter's
+        # transcription endpoint, OpenAI's is multipart/form-data only (no
+        # base64/JSON request mode).
+        def call(model:, audio_data:, audio_format:, language: nil, **_)
+          content_type = CONTENT_TYPES[audio_format]
+          unless content_type
+            raise Errors::InvalidRequestError,
+              "openai transcription does not support .#{audio_format} — use one of: #{CONTENT_TYPES.keys.join(', ')}"
+          end
+
+          boundary = SecureRandom.hex(16)
+          fields   = { "model" => model }
+          fields["language"] = language if language
+
+          body = build_multipart_body(boundary, fields, audio_data, "audio.#{audio_format}", content_type)
+          headers = {
+            "Content-Type"  => "multipart/form-data; boundary=#{boundary}",
+            "Authorization" => "Bearer #{api_key}"
+          }
+
+          raw  = HTTP.post(URI(ENDPOINT), headers: headers, body: body, timeout: 90)
+          data = parse!(raw)
+          handle_error!(data)
+
+          text = data["text"]
+          raise Errors::ProviderError, "No transcription text in response: #{data.keys}" if text.nil?
+
+          { content: text, provider: :openai, model: model, usage: extract_transcription_usage(data) }
+        end
+
+        private
+
+        def build_multipart_body(boundary, fields, file_data, filename, content_type)
+          body = +""
+          fields.each do |name, value|
+            body << "--#{boundary}\r\n"
+            body << "Content-Disposition: form-data; name=\"#{name}\"\r\n\r\n"
+            body << "#{value}\r\n"
+          end
+
+          body << "--#{boundary}\r\n"
+          body << "Content-Disposition: form-data; name=\"file\"; filename=\"#{filename}\"\r\n"
+          body << "Content-Type: #{content_type}\r\n\r\n"
+          body << file_data
+          body << "\r\n--#{boundary}--\r\n"
+          body
+        end
+
+        # whisper-1 reports usage as { type: "duration", seconds: N } — no token
+        # counts, so there's nothing to map to input_tokens/output_tokens. Newer
+        # models (gpt-4o-transcribe, gpt-4o-mini-transcribe) report
+        # { type: "tokens", input_tokens:, output_tokens:, total_tokens:, ... }.
+        # Neither reports a direct dollar cost like OpenRouter does — cost is left
+        # to the normal per-token Pricing lookup, which returns nil for
+        # duration-billed models since it has no token counts to work with.
+        def extract_transcription_usage(data)
+          u = data["usage"]
+          return nil unless u && u["type"] == "tokens"
+
+          {
+            input_tokens:  u["input_tokens"].to_i,
+            output_tokens: u["output_tokens"].to_i,
+            total_tokens:  u["total_tokens"].to_i
+          }
+        end
+
+        def api_key
+          key = config.openai_api_key.to_s
+          raise Errors::InvalidApiKeyError, "openai_api_key is not configured" if key.empty?
+          key
+        end
+
+        def handle_error!(data)
+          return unless data["error"]
+
+          msg      = data.dig("error", "message").to_s
+          code     = data.dig("error", "code").to_s
+          type     = data.dig("error", "type").to_s
+          metadata = data["error"].reject { |k, _| %w[message code type].include?(k) }
+          metadata = nil if metadata.empty?
+
+          case code
+          when "invalid_api_key", "unauthorized"
+            raise Errors::InvalidApiKeyError.new(msg,  error_code: code, metadata: metadata)
+          when "rate_limit_exceeded"
+            raise Errors::RateLimitError.new(msg,      error_code: code, metadata: metadata)
+          when "content_filter"
+            raise Errors::SafetyBlockedError.new(msg,  error_code: code, metadata: metadata)
+          else
+            case type
+            when "server_error"
+              raise Errors::ServerError.new(msg,       error_code: code, metadata: metadata)
+            else
+              raise Errors::InvalidRequestError.new(msg, error_code: code, metadata: metadata)
+            end
+          end
+        end
+      end
+    end
+  end
+end

data/lib/active_harness.rb

@@ -23,6 +23,7 @@ require_relative "active_harness/providers/vertexai"
 require_relative "active_harness/providers/custom"
 require_relative "active_harness/providers/images/openai"
 require_relative "active_harness/providers/images/openrouter"
+require_relative "active_harness/providers/audio/openai"
 require_relative "active_harness/providers/audio/openrouter"
 require "active_harness_pricing"
 require_relative "active_harness/memory"
@@ -33,7 +34,7 @@ require_relative "active_harness/pipeline"
 require_relative "active_harness/railtie" if defined?(Rails::Railtie)
 
 module ActiveHarness
-  VERSION = "0.2.41"
+  VERSION = "0.2.42"
 
   class << self
     # Configure ActiveHarness.

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: active_harness
 version: !ruby/object:Gem::Version
-  version: 0.2.41
+  version: 0.2.42
 platform: ruby
 authors:
 - the-teacher
 autorequire: 
 bindir: bin
 cert_chain: []
-date: 2026-07-19 00:00:00.000000000 Z
+date: 2026-07-20 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: concurrent-ruby
@@ -73,6 +73,7 @@ files:
 - lib/active_harness/pipeline/step.rb
 - lib/active_harness/providers/PROVIDER_CONTRACT.md
 - lib/active_harness/providers/anthropic.rb
+- lib/active_harness/providers/audio/openai.rb
 - lib/active_harness/providers/audio/openrouter.rb
 - lib/active_harness/providers/azure.rb
 - lib/active_harness/providers/base.rb