anthropic 1.79.0 → 1.81.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 74e6cbcdcbbf6d31ad5a08947e0b061fb8be75a367c6be4aac63cf7a715279fc
-  data.tar.gz: 9dcb1523d20c8af4dce1dadd9d6387f6dca92c0c2c8647eed98caa4a1e915197
+  metadata.gz: 154fbf0eeb1ef7997d8bc33f44e588fbe3fb3327aed6be57fba6d6459703d580
+  data.tar.gz: 974d03138f614db3ecc816de5c3d25e4862951faf84b0679ee8058758a89789e
 SHA512:
-  metadata.gz: 4d5532e5feec100f1edd36c35e00bc9b61192624fcd785710373fcb439fdc300cc0b499142df492bdc92098ebfe4a29ec36216fb0e8636f0c0e272c42865cdd0
-  data.tar.gz: 734afbe0a5d61ca96d6872c8ac7ee0ec771ac76b07349843b826e14cad1afa8a46619ed6ee416b3ba48405c4f3453800a174f15fe75f7e78a8f774b3384f979b
+  metadata.gz: e8701990520600baf05a9a240c836ba0245272b08a4df77a52d13da4d6c29cfaf4c884671ec3d2e8cec60a067ab5a859feff5630dbe1d3b796372bf691402408
+  data.tar.gz: bdad1379435404bc3ad2ab1534024c9a7baaa3bfbce296080e51fec7a31c9f1a95650f86c3e42ff94eb9b1dcd30c3f412bca8718148cbbd50b05b33c05624392

data/CHANGELOG.md

@@ -1,5 +1,56 @@
 # Changelog
 
+## [1.81.0](https://github.com/anthropics/anthropic-sdk-ruby/compare/v1.80.0...v1.81.0) (2026-10-20)
+
+### Features
+
+* **api:** add data_refreshed_at and has_more to analytics report pages
+* **api:** add workspace rate-limit updates to the Admin API
+* **api:** make User Profiles generally available
+* **client:** yield server-sent events with unrecognized event names
+* **google-cloud:** use the v1 endpoints and add the organization endpoint
+
+
+### Bug Fixes
+
+* **client:** `messages.stream` and `beta.messages.stream` now take the same parameters as `stream_raw`; their Sorbet and RBS signatures take one params hash
+* **types:** type the keyword parameters of messages.stream and beta.messages.stream
+
+
+### Chores
+
+* **ci:** publish gems from a fixed Ubuntu 24.04 runner
+* **ci:** set up Ruby 3.2 with a newer setup-ruby in the build and dependency version jobs
+* **ci:** skip the changelog check on draft pull requests
+* **docs:** add credential expiry descriptions to the Managed Agents API reference
+* **docs:** add Messages API request descriptions
+* **docs:** correct plugin manifest_version and reach descriptions
+* **docs:** improve Managed Agents memory store descriptions
+* **docs:** improve Managed Agents vault and credential descriptions
+* **tests:** cover per-request timeouts and later union variants
+* **tests:** make the tracing tests independent of test order
+* **tests:** send events that the stream yields in the tracing test
+
+## [1.80.0](https://github.com/anthropics/anthropic-sdk-ruby/compare/v1.79.0...v1.80.0) (2026-10-15)
+
+### Features
+
+* **api:** add between_tools to thinking types in model capabilities
+* **api:** add the speed parameter and usage.speed to Messages
+* **api:** return a Console organization's usage-tier spend cap as a read-only spend limit
+
+
+### Chores
+
+* **ci:** speed up breaking change detection
+* **docs:** describe headers on Managed Agents and telemetry endpoints
+* **tests:** cover trace context propagation and failing telemetry
+
+
+### Documentation
+
+* **api:** note that some models need a beta header to use speed
+
 ## [1.79.0](https://github.com/anthropics/anthropic-sdk-ruby/compare/v1.78.0...v1.79.0) (2026-10-14)
 
 ### Features
@@ -8,7 +59,9 @@
 * **api:** add organization_type_not_supported error to spend limit lists
 * **api:** add telemetry destination test endpoint
 * **api:** add telemetry destinations to the Admin API
-* **client:** add OpenTelemetry tracing for API calls, streaming and tool runs, on by default: once the application registers a tracer provider each call emits a span, and each request carries the trace context in the headers of the registered propagator, such as `traceparent`. `open_telemetry: false` or `ANTHROPIC_OPEN_TELEMETRY=false` turns both off, and `open_telemetry: {propagation: false}` or `ANTHROPIC_OPEN_TELEMETRY_PROPAGATION=false` only the headers, where a variable applies only to a client that is given no `open_telemetry` option. Adds the dependency `opentelemetry-api` (`>= 1.2.4`, `< 2`)
+* **client:** add native support for OpenTelemetry tracing
+
+  If your app registers a global tracer provider, the SDK will create spans with it after upgrading, and they'll be exported with the rest of your traces. Requests will also get trace context headers from the global propagator. To disable, pass `open_telemetry: false` to the client or set `ANTHROPIC_OPEN_TELEMETRY=false`. Learn more in the [documentation](https://platform.claude.com/docs/en/manage-claude/opentelemetry/sdk).
 
 
 ### Bug Fixes

data/README.md

@@ -15,7 +15,7 @@ Add to your application's Gemfile:
 <!-- x-release-please-start-version -->
 
 ```ruby
-gem "anthropic", "~> 1.79.0"
+gem "anthropic", "~> 1.81.0"
 ```
 
 <!-- x-release-please-end -->

data/lib/anthropic/client.rb

@@ -82,6 +82,9 @@ module Anthropic
     # @return [Anthropic::Resources::Skills]
     attr_reader :skills
 
+    # @return [Anthropic::Resources::UserProfiles]
+    attr_reader :user_profiles
+
     # @return [Anthropic::Resources::Organization]
     attr_reader :organization
 
@@ -342,6 +345,7 @@ module Anthropic
       @models = Anthropic::Resources::Models.new(client: self)
       @files = Anthropic::Resources::Files.new(client: self)
       @skills = Anthropic::Resources::Skills.new(client: self)
+      @user_profiles = Anthropic::Resources::UserProfiles.new(client: self)
       @organization = Anthropic::Resources::Organization.new(client: self)
       @beta = Anthropic::Resources::Beta.new(client: self)
     end

data/lib/anthropic/helpers/aws/client.rb

@@ -91,9 +91,9 @@ module Anthropic
 
           # In API-key mode the gateway key rides the base client's `api_key`
           # slot; {#resolve_default_credentials?} keeps first-party credential
-          # auto-discovery out of the picture when it is absent. The gateway
-          # serves the full first-party API, so every resource the parent wires
-          # up stays available.
+          # auto-discovery out of the picture when it is absent. Every resource
+          # the parent wires up stays available; {#build_request} rejects the
+          # one route the gateway does not serve.
           super(
             api_key: effective_api_key,
             base_url: resolved_base_url,
@@ -105,6 +105,23 @@ module Anthropic
           )
         end
 
+        # @api private
+        #
+        # Fails fast on user profile action URLs, which the gateway does not serve.
+        #
+        # @param req [Hash{Symbol=>Object}]
+        # @param opts [Hash{Symbol=>Object}]
+        # @return [Hash{Symbol=>Object}]
+        private def build_request(req, opts)
+          # Id-parameterized routes pass `path` as an Array whose first element
+          # is the format string (e.g. `["v1/user_profiles/%1$s/action_urls", id]`).
+          if Array(req[:path]).first.to_s.match?(%r{\Av1/user_profiles/[^/]+/action_urls})
+            raise NotImplementedError.new("User profile action URLs are not supported on the AWS gateway.")
+          end
+
+          super
+        end
+
         # @api private
         #
         # @return [String]

data/lib/anthropic/helpers/bedrock/client.rb

@@ -169,6 +169,10 @@ module Anthropic
           raise NotImplementedError, "Skills are not supported on Bedrock."
         end
 
+        def user_profiles
+          raise NotImplementedError, "User profiles are not supported on Bedrock."
+        end
+
         # @api private
         #
         # @param req [Hash{Symbol=>Object}] .
@@ -318,6 +322,8 @@ module Anthropic
             message = "Please instead use https://docs.anthropic.com/en/api/claude-on-amazon-bedrock#list-available-models to list available models on Bedrock."
             # rubocop:enable Layout/LineLength
             raise NotImplementedError.new(message)
+          in %r{^v1/user_profiles}
+            raise NotImplementedError.new("User profiles are not supported on Bedrock.")
           else
           end
         end

data/lib/anthropic/helpers/bedrock/mantle_client.rb

@@ -131,6 +131,13 @@ module Anthropic
                 "Skills are not supported on Bedrock Mantle. Only Messages (/v1/messages) is supported."
         end
 
+        # User profiles are not supported on Bedrock Mantle. Only Messages (/v1/messages) is supported.
+        def user_profiles
+          raise NotImplementedError,
+                "User profiles are not supported on Bedrock Mantle. " \
+                "Only Messages (/v1/messages) is supported."
+        end
+
         # @api private
         #
         # @return [Hash{String=>String}]
@@ -146,10 +153,36 @@ module Anthropic
         # @api private
         #
         # @return [#call]
-        private def provider_middleware = method(:aws_auth_provider)
+        private def provider_middleware
+          ->(req, nxt) { aws_auth_provider(adapt_request(req), nxt) }
+        end
+
+        # @api private
+        #
+        # Mantle drops custom request headers, so Messages and token counting send the user profile ID
+        # in the body.
+        #
+        # @param req [Anthropic::APIRequest]
+        # @return [Anthropic::APIRequest]
+        private def adapt_request(req)
+          path = req.url.path.to_s
+          return req unless req.method == :post && path.end_with?("/v1/messages", "/v1/messages/count_tokens")
+
+          body = req.body
+          profile_headers, headers = req.headers.partition do |name, _|
+            name.to_s.casecmp?("anthropic-user-profile-id")
+          end
+          user_profile_id = profile_headers.last&.last
+          return req if user_profile_id.to_s.empty? || !body.is_a?(Hash)
+
+          unless body.keys.any? { _1.to_s == "user_profile_id" }
+            body = {**body, user_profile_id: user_profile_id}
+          end
+          req.with(headers: headers.to_h, body: body)
+        end
 
         # Restricted Beta service that only exposes messages. Other beta resources
-        # (models, files, skills) are not supported on Bedrock Mantle.
+        # (models, files, skills, user profiles) are not supported on Bedrock Mantle.
         class MantleBeta
           # @return [Anthropic::Resources::Beta::Messages]
           attr_reader :messages
@@ -173,6 +206,12 @@ module Anthropic
             raise NotImplementedError,
                   "Beta skills are not supported on Bedrock Mantle. Only Messages (/v1/messages) is supported."
           end
+
+          def user_profiles
+            raise NotImplementedError,
+                  "Beta user profiles are not supported on Bedrock Mantle. " \
+                  "Only Messages (/v1/messages) is supported."
+          end
         end
       end
     end

data/lib/anthropic/helpers/google_cloud/client.rb

@@ -10,17 +10,26 @@ module Anthropic
       # `:rawPredict` publisher-model API (publisher model IDs, messages
       # only). This client speaks the full first-party Anthropic API: requests
       # pass through the gateway unchanged — standard `/v1/*` paths, standard
-      # model names, the complete API surface. The deprecated text Completions
-      # endpoint is intentionally not exposed.
+      # model names. The deprecated text Completions endpoint is intentionally
+      # not exposed, and the gateway does not serve the User Profiles API.
       class Client < Anthropic::Client
         # OAuth2 scope required to call the gateway.
         SCOPE = "https://www.googleapis.com/auth/cloud-platform"
 
+        LOCATION_URL_TEMPLATE =
+          "https://claude.googleapis.com/v1/projects/%<project>s/locations/%<location>s"
+
         # Gateway base URL template; override via `base_url` or
         # `ANTHROPIC_GOOGLE_CLOUD_BASE_URL`.
-        URL_TEMPLATE =
-          "https://claude.googleapis.com/v1alpha" \
-          "/projects/%<project>s/locations/%<location>s/workspaces/%<workspace_id>s/invoke"
+        URL_TEMPLATE = "#{LOCATION_URL_TEMPLATE}/workspaces/%<workspace_id>s/invoke".freeze
+
+        # Gateway base URL template for organization-level routes, which are
+        # served from a base with no workspace in it.
+        ORGANIZATION_URL_TEMPLATE = "#{LOCATION_URL_TEMPLATE}/claudeOrganization/invoke".freeze
+
+        ORGANIZATION_PATH_PREFIX = "/v1/organizations/"
+
+        private_constant :LOCATION_URL_TEMPLATE, :ORGANIZATION_PATH_PREFIX
 
         # @return [String, nil]
         attr_reader :project
@@ -53,12 +62,14 @@ module Anthropic
         #   explicit `nil` falls back to the same defaults.
         #
         # @param workspace_id [String, nil] Defaults to
-        #   `ENV["ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID"]`. Required unless
-        #   `skip_auth` is true and an explicit `base_url` is given.
+        #   `ENV["ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID"]`. Required when the base
+        #   URL is derived, except for organization-level routes
+        #   (`/v1/organizations/`), which need no workspace ID.
         #
-        # @param base_url [String, nil] Override the gateway base URL. Defaults to
-        #   `ENV["ANTHROPIC_GOOGLE_CLOUD_BASE_URL"]`, otherwise derived from
-        #   `project`, `location`, and `workspace_id` via {URL_TEMPLATE}.
+        # @param base_url [String, nil] Override the gateway base URL for every
+        #   route. Defaults to `ENV["ANTHROPIC_GOOGLE_CLOUD_BASE_URL"]`, otherwise
+        #   derived from `project`, `location`, and `workspace_id` via
+        #   {URL_TEMPLATE}.
         #
         # @param google_credentials [#apply, nil] A `googleauth`-compatible
         #   credentials object. Mutually exclusive with `skip_auth`.
@@ -67,10 +78,9 @@ module Anthropic
         #   token string. Takes precedence over `google_credentials`. Mutually
         #   exclusive with `skip_auth`.
         #
-        # @param skip_auth [Boolean] When true, no token is attached and
-        #   `workspace_id` is only required when the base URL must be derived.
-        #   For callers fronting with their own authenticated proxy. Mutually
-        #   exclusive with `token_provider` and `google_credentials`.
+        # @param skip_auth [Boolean] When true, no token is attached. For callers
+        #   fronting with their own authenticated proxy. Mutually exclusive with
+        #   `token_provider` and `google_credentials`.
         #
         # @param max_retries [Integer] Max number of retries to attempt after a
         #   failed retryable request.
@@ -132,23 +142,23 @@ module Anthropic
           # so nil can't reach the derived URL (matches the other SDKs).
           @location = location || ENV["ANTHROPIC_GOOGLE_CLOUD_LOCATION"] || "global"
 
-          if workspace_id.nil? && !skip_auth
-            raise ArgumentError.new(
-              "No workspace ID was given. Set the `workspace_id` argument or the " \
-              "`ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID` environment variable."
-            )
-          end
           @workspace_id = workspace_id
 
           base_url = nil if base_url.to_s.empty?
-          base_url ||= derive_base_url
+          unless base_url
+            @organization_base_url = URI.parse(derive_base_url(ORGANIZATION_URL_TEMPLATE))
+            # Without a workspace ID there is no workspace base to derive, so
+            # `base_url` is the organization base.
+            base_url = @workspace_id.nil? ? @organization_base_url.to_s : derive_base_url(URL_TEMPLATE)
+          end
 
           # No first-party credential is ever passed up, and
           # {#resolve_default_credentials?} keeps the parent from discovering
           # one; `base_url` is always resolved by this point, so the parent's
           # `ANTHROPIC_BASE_URL` fallback never applies. The parent wires up
-          # every resource; {#completions} withholds the one the gateway does
-          # not serve.
+          # every resource; {#completions} and {#user_profiles} withhold the
+          # ones the gateway does not serve, and {#build_request} refuses the
+          # user profile routes however the resource is reached.
           super(
             base_url: base_url,
             timeout: timeout,
@@ -176,6 +186,83 @@ module Anthropic
           )
         end
 
+        # User profiles are not supported on the Google Cloud gateway.
+        #
+        # @raise [NotImplementedError]
+        def user_profiles
+          raise NotImplementedError.new("User profiles are not supported on the Google Cloud gateway.")
+        end
+
+        # @api private
+        #
+        # @param req [Hash{Symbol=>Object}] .
+        #
+        #   @option req [Symbol] :method
+        #
+        #   @option req [String, Array<String>] :path
+        #
+        #   @option req [Hash{String=>Array<String>, String, nil}, nil] :query
+        #
+        #   @option req [Hash{String=>String, Integer, Array<String, Integer, nil>, nil}, nil] :headers
+        #
+        #   @option req [Object, nil] :body
+        #
+        #   @option req [Symbol, Integer, Array<Symbol, Integer>, Proc, nil] :unwrap
+        #
+        #   @option req [Class<Anthropic::Internal::Type::BasePage>, nil] :page
+        #
+        #   @option req [Class<Anthropic::Internal::Type::BaseStream>, nil] :stream
+        #
+        #   @option req [Anthropic::Internal::Type::Converter, Class, nil] :model
+        #
+        # @param opts [Hash{Symbol=>Object}] .
+        #
+        #   @option opts [String, nil] :idempotency_key
+        #
+        #   @option opts [Hash{String=>Array<String>, String, nil}, nil] :extra_query
+        #
+        #   @option opts [Hash{String=>String, nil}, nil] :extra_headers
+        #
+        #   @option opts [Object, nil] :extra_body
+        #
+        #   @option opts [Integer, nil] :max_retries
+        #
+        #   @option opts [Float, nil] :timeout
+        #
+        # @raise [ArgumentError]
+        # @raise [NotImplementedError]
+        # @return [Hash{Symbol=>Object}]
+        private def build_request(req, opts)
+          # Id-parameterized routes pass `path` as an Array whose first element
+          # is the format string (e.g. `["v1/user_profiles/%1$s", id]`).
+          if Array(req[:path]).first.to_s.start_with?("v1/user_profiles")
+            raise NotImplementedError.new("User profiles are not supported on the Google Cloud gateway.")
+          end
+
+          request = super
+          return request if @organization_base_url.nil?
+
+          url = request.fetch(:url)
+          under_base_url =
+            url.path.start_with?("#{base_url.path}/") &&
+            Anthropic::Internal::Util.uri_origin(url) == Anthropic::Internal::Util.uri_origin(base_url)
+          return request unless under_base_url
+
+          path = url.path.delete_prefix(base_url.path)
+          if path.start_with?(ORGANIZATION_PATH_PREFIX)
+            url = url.dup
+            url.path = "#{@organization_base_url.path}#{path}"
+            {**request, url: url}
+          elsif @workspace_id.nil?
+            raise ArgumentError.new(
+              "No workspace ID was given. Set the `workspace_id` argument or the " \
+              "`ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID` environment variable."
+            )
+          else
+            request
+          end
+        end
+
         # @api private
         #
         # Hard guarantee that no first-party `X-Api-Key` / `Authorization`
@@ -230,9 +317,10 @@ module Anthropic
 
         # @api private
         #
+        # @param template [String]
         # @return [String]
         # @raise [ArgumentError]
-        private def derive_base_url
+        private def derive_base_url(template)
           if @project.nil?
             raise ArgumentError.new(
               "No project found. Set the `project` argument, set the " \
@@ -240,13 +328,7 @@ module Anthropic
               "application default credentials with a project."
             )
           end
-          if @workspace_id.nil?
-            raise ArgumentError.new(
-              "No workspace ID was given. Set the `workspace_id` argument or the " \
-              "`ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID` environment variable."
-            )
-          end
-          format(URL_TEMPLATE, project: @project, location: @location, workspace_id: @workspace_id)
+          format(template, project: @project, location: @location, workspace_id: @workspace_id)
         end
 
         # @api private

data/lib/anthropic/helpers/refusal_fallback.rb

@@ -38,10 +38,11 @@ module Anthropic
   # — and the fallback's events are spliced onto the still-open stream, so the
   # client sees one continuous message: a `fallback` content block at each
   # model boundary, monotonic block indices, and per-hop `usage.iterations` on
-  # the final `message_delta`. Only `model` is honored from each entry on this
-  # path: the credit token is redeemable only against the refused request's
-  # body, so the other per-entry overrides would be rejected. A
-  # `between_tools` thinking config is always sent as `disabled` on this path.
+  # the final `message_delta`. Only `model` and `speed` are honored from each
+  # entry on this path. The other per-entry overrides are ignored, because the
+  # credit token is tied to the refused request's body. A `speed` value replaces
+  # the request's, and `speed: nil` omits it. A `between_tools` thinking config
+  # is always sent as `disabled` on this path.
   #
   # Credit tokens are sent in the object form with `mode: :best_effort`, so a
   # failing redemption never rejects the retry — it proceeds at normal price
@@ -49,6 +50,10 @@ module Anthropic
   # beta this requires is sent by default on every request the middleware
   # handles; the `betas:` option controls this.
   #
+  # A `speed: "fast"` request that falls back to an entry with no `speed`, and
+  # is rejected with a 429 or 529, is retried once on that entry without
+  # `speed`. Set `speed: "fast"` on the entry to keep it fast-only.
+  #
   # To keep later requests on the model that accepted, pass a
   # {BetaFallbackState} via the `fallback_state` request option; requests
   # sharing that state start directly at the pinned fallback.
@@ -72,7 +77,7 @@ module Anthropic
     #   …) patch the original request body for that hop — a value overrides
     #   the param, an explicit `nil` unsets it, an absent key keeps the
     #   original value (an absent `thinking:` sends `between_tools` as
-    #   `disabled`).
+    #   `disabled`). On the streaming path only `speed:` patches it.
     # @param betas [Array<String>] betas added to the `anthropic-beta` header
     #   of every `/v1/messages` request this middleware handles. Defaults to
     #   `["fallback-credit-2026-07-01"]`; pass `[]` to send none.
@@ -129,24 +134,24 @@ module Anthropic
       end
 
       streaming = req.streaming?
+      entry = @fallbacks.fetch(start_index) unless start_index == -1
       initial_body =
-        if start_index == -1
+        if entry.nil?
           body
         elsif streaming
-          # Only `model` is honored on the streaming path so the credit token
-          # — redeemable only against the refused request's body — stays valid
-          # for the spliced hops.
-          degrade_between_tools(body.merge(model: @fallbacks.fetch(start_index).fetch(:model)))
+          # Only `model` and `speed` are honored on the streaming path: the credit
+          # token is tied to the refused request's body.
+          degrade_between_tools(patch(body, entry.slice(:speed)).merge(model: entry.fetch(:model)))
         else
-          apply_fallback(body, @fallbacks.fetch(start_index))
+          apply_fallback(body, entry)
         end
-      res = nxt.call(req.with(body: initial_body))
+      res = entry ? send_hop(req, nxt, entry, initial_body).first : nxt.call(req.with(body: initial_body))
       return res unless res.status < 300
 
       if streaming
         first_hop = start_index + 1
         return res unless first_hop < @fallbacks.length
-        return splice_fallback_stream(req, res, nxt, body, first_hop, pin)
+        return splice_fallback_stream(req, res, nxt, initial_body, first_hop, pin)
       end
 
       handle_non_streaming(req, res, nxt, body, start_index, pin)
@@ -276,6 +281,25 @@ module Anthropic
       {token: token, mode: Anthropic::Models::Beta::BetaFallbackCreditTokenParam::Mode::BEST_EFFORT}
     end
 
+    # Send a request built from `entry`. One that is fast only because the
+    # original request was, and is rejected with a 429 or 529, is resent once
+    # without `speed`, and that response stands in for the first.
+    #
+    # @param req [Anthropic::APIRequest]
+    # @param nxt [#call]
+    # @param entry [Hash{Symbol=>Object}]
+    # @param hop_body [Hash{Symbol=>Object}]
+    # @return [Array(Anthropic::APIResponse, Boolean)] the response, and whether
+    #   it answers the resend.
+    def send_hop(req, nxt, entry, hop_body)
+      res = nxt.call(req.with(body: hop_body))
+      return [res, false] unless [429, 529].include?(res.status)
+      return [res, false] if entry.key?(:speed) || hop_body[:speed].to_s != "fast"
+
+      Anthropic::Internal::Util.close_fused!(res.body)
+      [nxt.call(req.with(body: hop_body.except(:speed))), true]
+    end
+
     # --- non-streaming ------------------------------------------------------
 
     # @param req [Anthropic::APIRequest]
@@ -303,7 +327,7 @@ module Anthropic
         hop_body = hop_body.merge(fallback_credit_token: credit_token_param(token)) if token
 
         begin
-          hop_res = nxt.call(req.with(body: hop_body))
+          hop_res = send_hop(req, nxt, entry, hop_body).first
         rescue StandardError
           hop_res = nil
         end
@@ -359,7 +383,7 @@ module Anthropic
     # @param req [Anthropic::APIRequest]
     # @param res [Anthropic::APIResponse]
     # @param nxt [#call]
-    # @param body [Hash{Symbol=>Object}] the stripped request body
+    # @param body [Hash{Symbol=>Object}] the body built for stream A
     # @param first_hop [Integer]
     # @param pin [Proc]
     # @return [Anthropic::APIResponse]
@@ -399,10 +423,12 @@ module Anthropic
       iterations = [iteration_usage("message", a[:model] || "", a.dig(:refused, :usage))]
 
       (first_hop...@fallbacks.length).each do |hop|
-        model = @fallbacks.fetch(hop).fetch(:model).to_s
+        entry = @fallbacks.fetch(hop)
+        model = entry.fetch(:model).to_s
         has_next = hop + 1 < @fallbacks.length
 
-        res_b, sent, failed = issue_hop(req, nxt, degrade_between_tools(body), model, token, base, partial)
+        hop_body = degrade_between_tools(patch(body, entry.slice(:speed)))
+        res_b, sent, failed = issue_hop(req, nxt, hop_body, entry, token, base, partial)
 
         if failed
           next if has_next
@@ -465,15 +491,19 @@ module Anthropic
       end
     end
 
-    # Issue one fallback hop with a single 400-on-prefill retry.
+    # Issue one fallback hop with a single 400-on-prefill retry, which stays
+    # without `speed` once the hop has been resent without it.
     #
     # @return [Array(Anthropic::APIResponse, Array<Hash>, false), Array(nil, Array<Hash>, true)]
     #   the response, the continuation actually sent (`base + partial`, or
     #   `base` if the prefill was dropped on retry), and whether the hop failed.
-    def issue_hop(req, nxt, body, model, token, base, partial)
+    def issue_hop(req, nxt, body, entry, token, base, partial)
       continuation = base + partial
       2.times do |attempt|
-        hop_body = body.merge(model: model, fallback_credit_token: credit_token_param(token))
+        hop_body = body.merge(
+          model: entry.fetch(:model).to_s,
+          fallback_credit_token: credit_token_param(token)
+        )
         unless continuation.empty?
           hop_body = hop_body.merge(
             messages: body.fetch(:messages) + [{role: :assistant, content: continuation}]
@@ -481,10 +511,11 @@ module Anthropic
         end
 
         begin
-          res = nxt.call(req.with(body: hop_body))
+          res, resent = send_hop(req, nxt, entry, hop_body)
         rescue StandardError
           return [nil, continuation, true]
         end
+        body = body.except(:speed) if resent
         return [res, continuation, false] if res.status < 300
 
         if attempt.zero? && res.status == 400 && !partial.empty?

data/lib/anthropic/helpers/vertex/client.rb

@@ -141,6 +141,10 @@ module Anthropic
           raise NotImplementedError, "Skills are not supported on Vertex."
         end
 
+        def user_profiles
+          raise NotImplementedError, "Managing user profiles is not supported on Vertex."
+        end
+
         # @api private
         #
         # @param req [Hash{Symbol=>Object}] .
@@ -185,6 +189,9 @@ module Anthropic
           if path.start_with?("v1/messages/batches")
             raise NotImplementedError.new("The Batch API is not supported in the Vertex client yet")
           end
+          if path.start_with?("v1/user_profiles")
+            raise NotImplementedError.new("Managing user profiles is not supported on Vertex.")
+          end
 
           super
         end