[Fix] Properly raise rate limit errors from Anthropic

lib/ruby_llm/error.rb

@@ -62,6 +62,13 @@ class << self
         /reduce the length of messages/i
       ].freeze
 
+      RATE_LIMIT_PATTERNS = [
+        /rate limit/i,
+        /per minute/i,
+        /per hour/i,
+        /per day/i
+      ].freeze
+
       def parse_error(provider:, response:) # rubocop:disable Metrics/PerceivedComplexity
         message = provider&.parse_error(response)
 
@@ -102,6 +109,7 @@ def parse_error(provider:, response:) # rubocop:disable Metrics/PerceivedComplex
 
       def context_length_exceeded?(message)
         return false if message.to_s.empty?
+        return false if RATE_LIMIT_PATTERNS.any? { |pattern| message.match?(pattern) }
 
         CONTEXT_LENGTH_PATTERNS.any? { |pattern| message.match?(pattern) }
       end

spec/ruby_llm/error_middleware_spec.rb

@@ -48,6 +48,18 @@
       end.to raise_error(RubyLLM::RateLimitError)
     end
 
+    it 'keeps token-per-minute rate limit 429 errors as RateLimitError, not ContextLengthExceededError' do
+      # Anthropic returns 429 with a message mentioning "input tokens per minute", which matches
+      # the /input[_\s-]?token/i context length pattern — it should be classified as a rate limit.
+      msg = "This request would exceed your organization's rate limit of 30,000 input tokens per minute"
+      response = Struct.new(:status, :body).new(429, %({"error":{"message":"#{msg}"}}))
+      provider = instance_double(RubyLLM::Provider, parse_error: msg)
+
+      expect do
+        described_class.parse_error(provider: provider, response: response)
+      end.to raise_error(RubyLLM::RateLimitError)
+    end
+
     it 'maps context-length-like 400 errors to ContextLengthExceededError' do
       msg = "This model's maximum context length is 8192 tokens."
       response = Struct.new(:status, :body).new(400, %({"error":{"message":"#{msg}"}}))