Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

fix: eliminate duplicate proxy logging and guard against token inflation

Tejas Chopra039399a017d5
diff --git a/headroom/proxy/handlers/anthropic.py b/headroom/proxy/handlers/anthropic.py
index 064a6683a..365a866bc 100644
--- a/headroom/proxy/handlers/anthropic.py
+++ b/headroom/proxy/handlers/anthropic.py
@@ -681,7 +681,17 @@ class AnthropicHandlerMixin:
# Flag compression failure for observability
_compression_failed = True
- tokens_saved = max(0, original_tokens - optimized_tokens)
+ # Guard: if "optimization" inflated tokens, revert to originals
+ if optimized_tokens > original_tokens:
+ logger.warning(
+ f"[{request_id}] Optimization inflated tokens "
+ f"({original_tokens} -> {optimized_tokens}), reverting to original messages"
+ )
+ optimized_messages = original_messages
+ optimized_tokens = original_tokens
+ transforms_applied = []
+
+ tokens_saved = original_tokens - optimized_tokens
optimization_latency = (time.time() - start_time) * 1000
# Hook: post_compress — let hooks observe compression results
@@ -1592,9 +1602,18 @@ class AnthropicHandlerMixin:
# Use pipeline's token counts for consistency with pipeline logs
original_tokens = result.tokens_before
optimized_tokens = result.tokens_after
+ # Guard: if "optimization" inflated tokens, revert to originals
+ if optimized_tokens > original_tokens:
+ logger.warning(
+ f"[{request_id}] Batch item optimization inflated tokens "
+ f"({original_tokens} -> {optimized_tokens}), reverting"
+ )
+ optimized_messages = messages
+ optimized_tokens = original_tokens
+
total_original_tokens += original_tokens
total_optimized_tokens += optimized_tokens
- tokens_saved = max(0, original_tokens - optimized_tokens)
+ tokens_saved = original_tokens - optimized_tokens
total_tokens_saved += tokens_saved
# CCR Tool Injection: Inject retrieval tool if compression occurred
diff --git a/headroom/proxy/handlers/batch.py b/headroom/proxy/handlers/batch.py
index 14076ec63..8ce5fe683 100644
--- a/headroom/proxy/handlers/batch.py
+++ b/headroom/proxy/handlers/batch.py
@@ -158,9 +158,18 @@ class BatchHandlerMixin:
# Use pipeline's token counts for consistency with pipeline logs
original_tokens = result.tokens_before
optimized_tokens = result.tokens_after
+ # Guard: if "optimization" inflated tokens, revert to originals
+ if optimized_tokens > original_tokens:
+ logger.warning(
+ f"[{request_id}] Batch item optimization inflated tokens "
+ f"({original_tokens} -> {optimized_tokens}), reverting"
+ )
+ optimized_messages = messages
+ optimized_tokens = original_tokens
+
total_original_tokens += original_tokens
total_optimized_tokens += optimized_tokens
- tokens_saved = max(0, original_tokens - optimized_tokens)
+ tokens_saved = original_tokens - optimized_tokens
total_tokens_saved += tokens_saved
# CCR Tool Injection: Inject retrieval tool if compression occurred
diff --git a/headroom/proxy/handlers/gemini.py b/headroom/proxy/handlers/gemini.py
index 3ded54954..d11abb6ad 100644
--- a/headroom/proxy/handlers/gemini.py
+++ b/headroom/proxy/handlers/gemini.py
@@ -265,7 +265,17 @@ class GeminiHandlerMixin:
_compression_failed = True
logger.warning(f"[{request_id}] Gemini optimization failed: {e}")
- tokens_saved = max(0, original_tokens - optimized_tokens)
+ # Guard: if "optimization" inflated tokens, revert to originals
+ if optimized_tokens > original_tokens:
+ logger.warning(
+ f"[{request_id}] Optimization inflated tokens "
+ f"({original_tokens} -> {optimized_tokens}), reverting to original messages"
+ )
+ optimized_messages = messages
+ optimized_tokens = original_tokens
+ transforms_applied = []
+
+ tokens_saved = original_tokens - optimized_tokens
optimization_latency = (time.time() - start_time) * 1000
# Query Echo: disabled — hurts prefix caching in long conversations.
diff --git a/headroom/proxy/handlers/openai.py b/headroom/proxy/handlers/openai.py
index 328ab2367..984484731 100644
--- a/headroom/proxy/handlers/openai.py
+++ b/headroom/proxy/handlers/openai.py
@@ -298,7 +298,17 @@ class OpenAIHandlerMixin:
# Flag compression failure for observability
_compression_failed = True
- tokens_saved = max(0, original_tokens - optimized_tokens)
+ # Guard: if "optimization" inflated tokens, revert to originals
+ if optimized_tokens > original_tokens:
+ logger.warning(
+ f"[{request_id}] Optimization inflated tokens "
+ f"({original_tokens} -> {optimized_tokens}), reverting to original messages"
+ )
+ optimized_messages = original_messages
+ optimized_tokens = original_tokens
+ transforms_applied = []
+
+ tokens_saved = original_tokens - optimized_tokens
optimization_latency = (time.time() - start_time) * 1000
# Hook: post_compress
@@ -791,7 +801,17 @@ class OpenAIHandlerMixin:
except Exception as e:
logger.warning(f"[{request_id}] Responses API optimization failed: {e}")
- tokens_saved = max(0, original_tokens - optimized_tokens)
+ # Guard: if "optimization" inflated tokens, revert to originals
+ if optimized_tokens > original_tokens:
+ logger.warning(
+ f"[{request_id}] Optimization inflated tokens "
+ f"({original_tokens} -> {optimized_tokens}), reverting to original messages"
+ )
+ optimized_messages = messages
+ optimized_tokens = original_tokens
+ transforms_applied = []
+
+ tokens_saved = original_tokens - optimized_tokens
optimization_latency = (time.time() - start_time) * 1000
# Convert compressed messages back to Responses API items
@@ -1056,7 +1076,13 @@ class OpenAIHandlerMixin:
if instructions and opt and opt[0].get("role") == "system":
body["instructions"] = opt[0]["content"]
opt = opt[1:]
- body["input"] = messages_to_responses_items(opt, input_data, preserved)
+ if result.tokens_after <= original_tokens:
+ body["input"] = messages_to_responses_items(opt, input_data, preserved)
+ else:
+ logger.warning(
+ f"[{request_id}] WS optimization inflated tokens "
+ f"({original_tokens} -> {result.tokens_after}), reverting"
+ )
tokens_saved = max(0, original_tokens - result.tokens_after)
first_msg_raw = json.dumps(body)
logger.info(
diff --git a/headroom/proxy/helpers.py b/headroom/proxy/helpers.py
index fea2d8273..41faa0626 100644
--- a/headroom/proxy/helpers.py
+++ b/headroom/proxy/helpers.py
@@ -78,8 +78,15 @@ def _setup_file_logging() -> None:
handler.setFormatter(
logging.Formatter("%(asctime)s - %(name)s - %(levelname)s - %(message)s")
)
- # Attach to the headroom root logger so all sub-loggers are captured
- logging.getLogger("headroom").addHandler(handler)
+ # Attach to the headroom root logger so all sub-loggers are captured.
+ # Disable propagation to root to avoid duplicate writes when
+ # wrap.py redirects stderr to the same log file.
+ headroom_logger = logging.getLogger("headroom")
+ if not any(
+ isinstance(h, RotatingFileHandler) for h in headroom_logger.handlers
+ ):
+ headroom_logger.addHandler(handler)
+ headroom_logger.propagate = False
except OSError:
# Non-fatal: can't write logs (read-only fs, permissions, etc.)
pass