Bump to v0.5.19: cache stats fixes, Codex WS fix, compression-vs-cache tracking

v0.5.19 includes:
- Fix Codex WebSocket HTTP 500: ChatGPT auth routing to chatgpt.com,
correct beta header, HTTP POST streaming fallback
- Token-level cache hit rate (not binary per-request)
- Compression-vs-cache tracking: tokens_saved vs cache_bust_tokens per session
- Pass uncached_input_tokens for OpenAI and Gemini handlers
- Fix async test failures in test_proxy_scalability

headroom/__init__.py

@@ -153,7 +153,7 @@ from .transforms import (
     TransformPipeline,
 )
 
-__version__ = "0.5.18"
+__version__ = "0.5.19"
 
 __all__ = [
     # Main client

headroom/proxy/handlers/gemini.py

@@ -341,12 +341,15 @@ class GeminiHandlerMixin:
                         f"[{request_id}] Failed to extract cached tokens from Gemini response: {e}"
                     )
 
+                uncached_input_tokens = max(0, total_input_tokens - cache_read_tokens)
+
                 if self.cost_tracker:
                     self.cost_tracker.record_tokens(
                         model,
                         tokens_saved,
                         optimized_tokens,
                         cache_read_tokens=cache_read_tokens,
+                        uncached_tokens=uncached_input_tokens,
                     )
 
                 await self.metrics.record_request(
@@ -359,6 +362,7 @@ class GeminiHandlerMixin:
                     overhead_ms=optimization_latency,
                     waste_signals=waste_signals_dict,
                     cache_read_tokens=cache_read_tokens,
+                    uncached_input_tokens=uncached_input_tokens,
                 )
 
                 if tokens_saved > 0:

headroom/proxy/handlers/openai.py

@@ -512,12 +512,16 @@ class OpenAIHandlerMixin:
                     messages=optimized_messages,
                 )
 
+                # OpenAI has no write penalty — uncached = total - cached
+                uncached_input_tokens = max(0, total_input_tokens - cache_read_tokens)
+
                 if self.cost_tracker:
                     self.cost_tracker.record_tokens(
                         model,
                         tokens_saved,
                         optimized_tokens,
                         cache_read_tokens=cache_read_tokens,
+                        uncached_tokens=uncached_input_tokens,
                     )
 
                 # Cache
@@ -537,6 +541,7 @@ class OpenAIHandlerMixin:
                     pipeline_timing=pipeline_timing,
                     waste_signals=waste_signals_dict,
                     cache_read_tokens=cache_read_tokens,
+                    uncached_input_tokens=uncached_input_tokens,
                 )
 
                 if tokens_saved > 0:

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "headroom-ai"
-version = "0.5.18"
+version = "0.5.19"
 description = "The Context Optimization Layer for LLM Applications - Cut costs by 50-90%"
 readme = "README.md"
 license = "Apache-2.0"