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.

Reduce Codex compression proxy logging

Tejas Chopra4fbfc02a398f
diff --git a/headroom/proxy/handlers/openai.py b/headroom/proxy/handlers/openai.py
index 45b3cb716..08fc3baef 100644
--- a/headroom/proxy/handlers/openai.py
+++ b/headroom/proxy/handlers/openai.py
@@ -43,53 +43,8 @@ def _json_debug_dumps(value: Any) -> str:
return json.dumps(value, ensure_ascii=False, default=str, separators=(",", ":"))
-def _redacted_debug_value(value: Any) -> Any:
- from headroom.proxy.helpers import redact_for_wire_debug
-
- return redact_for_wire_debug(value)
-
-
-# Fields that embed full payload bodies (~tens of KB per event). At INFO
-# level these dominate log volume and obscure the summary signals; at
-# DEBUG level they remain available for deep diagnosis.
-_LOG_BLOAT_FIELDS = frozenset(
- {
- "payload",
- "input_payload",
- "output_payload",
- "input",
- "output",
- "original",
- "compressed",
- "extraction",
- "units",
- "router_routing_log",
- }
-)
-
-
-def _log_codex_compression_debug(event: str, **payload: Any) -> None:
- """Emit Codex compression diagnostics.
-
- INFO carries the summary fields (request_id, pass_id, savings,
- strategy_chain, units_by_category, etc.) — the line a human log
- reader should be able to scan to understand what happened.
-
- DEBUG carries everything, including the full request/response
- payload bodies and per-unit text. Enable the proxy logger at DEBUG
- when actually diagnosing a regression; leave it at INFO for
- steady-state ops to keep proxy.log readable.
-
- Obvious credential keys are still redacted by key name at both levels.
- """
-
- payload = {"event": event, **payload}
- # DEBUG: everything.
- if logger.isEnabledFor(logging.DEBUG):
- logger.debug("event=%s %s", event, _json_debug_dumps(_redacted_debug_value(payload)))
- # INFO: stripped of body-embedding fields so summaries are scannable.
- stripped = {k: v for k, v in payload.items() if k not in _LOG_BLOAT_FIELDS}
- logger.info("event=%s %s", event, _json_debug_dumps(_redacted_debug_value(stripped)))
+def _log_codex_compression_debug(_event: str, **_payload: Any) -> None:
+ return
def _json_shape(value: str) -> dict[str, Any]:
@@ -109,25 +64,8 @@ def _json_shape(value: str) -> dict[str, Any]:
return {"is_json": True, "kind": type(parsed).__name__}
-def _routing_log_debug(router_result: Any) -> list[dict[str, Any]]:
- routing_log = getattr(router_result, "routing_log", []) or []
- rows: list[dict[str, Any]] = []
- for idx, decision in enumerate(routing_log):
- content_type = getattr(decision, "content_type", None)
- strategy = getattr(decision, "strategy", None)
- rows.append(
- {
- "index": idx,
- "content_type": getattr(content_type, "value", str(content_type)),
- "strategy": getattr(strategy, "value", str(strategy)),
- "original_tokens": getattr(decision, "original_tokens", None),
- "compressed_tokens": getattr(decision, "compressed_tokens", None),
- "confidence": getattr(decision, "confidence", None),
- "section_index": getattr(decision, "section_index", None),
- "compression_ratio": getattr(decision, "compression_ratio", None),
- }
- )
- return rows
+def _routing_log_debug(_router_result: Any) -> list[dict[str, Any]]:
+ return []
_OPENAI_TOOL_SCHEMA_DROP_KEYS = {
@@ -431,34 +369,13 @@ class OpenAIHandlerMixin:
are intentionally not exposed as text units.
"""
- def _log(event: str, **fields: Any) -> None:
- # All events from this single compression pass share `pass_id`,
- # so downstream analytics can attribute each unit_result to
- # its originating pass — without this, unit_results from
- # multiple passes get conflated under one request_id and
- # per-pass questions become unanswerable. Savings ARE expected
- # to sum across passes: each pass really did avoid sending
- # those tokens upstream, regardless of any prefix-cache the
- # upstream may apply.
- _log_codex_compression_debug(
- event,
- request_id=request_id,
- pass_id=pass_id,
- model=model,
- **fields,
- )
+ def _log(_event: str, **_fields: Any) -> None:
+ return
input_items = payload.get("input")
messages_items = payload.get("messages")
items = input_items if isinstance(input_items, list) else messages_items
if not isinstance(items, list):
- _log(
- "codex_compression_payload_skipped",
- reason="input_or_messages_not_list",
- input_type=type(payload.get("input")).__name__,
- messages_type=type(payload.get("messages")).__name__,
- payload=payload,
- )
return payload, False, 0, [], {}, [], 0
try:
from headroom.transforms.compression_units import (
@@ -473,22 +390,11 @@ class OpenAIHandlerMixin:
request_id,
exc,
)
- _log(
- "codex_compression_payload_skipped",
- reason="compression_unit_adapter_unavailable",
- error=repr(exc),
- payload=payload,
- )
return payload, False, 0, [], {}, [], 0
router = find_content_router(self.openai_pipeline)
if router is None:
logger.debug("[%s] OpenAI Responses ContentRouter unavailable", request_id)
- _log(
- "codex_compression_payload_skipped",
- reason="content_router_unavailable",
- payload=payload,
- )
return payload, False, 0, [], {}, [], 0
try:
@@ -499,12 +405,6 @@ class OpenAIHandlerMixin:
request_id,
exc,
)
- _log(
- "codex_compression_payload_skipped",
- reason="tokenizer_unavailable",
- error=repr(exc),
- payload=payload,
- )
return payload, False, 0, [], {}, [], 0
def _slot_text(item: dict[str, Any]) -> tuple[str, tuple[str, int | None]] | None:
diff --git a/tests/test_openai_responses_compression_units.py b/tests/test_openai_responses_compression_units.py
index 654978d58..0f3700545 100644
--- a/tests/test_openai_responses_compression_units.py
+++ b/tests/test_openai_responses_compression_units.py
@@ -51,7 +51,7 @@ def test_openai_responses_adapter_compresses_only_live_text_slots():
],
}
- new_payload, modified, saved, transforms, units_by_category, strategy_chain = (
+ new_payload, modified, saved, transforms, units_by_category, strategy_chain, _attempted = (
handler._compress_openai_responses_live_text_units_with_router(
payload,
model="gpt-5",
@@ -94,7 +94,7 @@ def test_openai_responses_adapter_compresses_custom_tool_call_output():
],
}
- new_payload, modified, saved, transforms, units_by_category, strategy_chain = (
+ new_payload, modified, saved, transforms, units_by_category, strategy_chain, _attempted = (
handler._compress_openai_responses_live_text_units_with_router(
payload,
model="gpt-5",
@@ -117,7 +117,7 @@ def test_openai_responses_adapter_accepts_empty_input_list():
handler = _handler_with_router(router)
payload = {"model": "gpt-5", "input": [], "tools": []}
- new_payload, modified, saved, transforms, units_by_category, strategy_chain = (
+ new_payload, modified, saved, transforms, units_by_category, strategy_chain, _attempted = (
handler._compress_openai_responses_live_text_units_with_router(
payload,
model="gpt-5",
@@ -163,7 +163,7 @@ def test_openai_responses_adapter_preserves_headroom_retrieve_outputs():
],
}
- new_payload, modified, saved, transforms, units_by_category, strategy_chain = (
+ new_payload, modified, saved, transforms, units_by_category, strategy_chain, _attempted = (
handler._compress_openai_responses_live_text_units_with_router(
payload,
model="gpt-5",
@@ -199,7 +199,7 @@ def test_openai_responses_adapter_keeps_small_and_opaque_items():
],
}
- new_payload, modified, saved, transforms, units_by_category, strategy_chain = (
+ new_payload, modified, saved, transforms, units_by_category, strategy_chain, _attempted = (
handler._compress_openai_responses_live_text_units_with_router(
payload,
model="gpt-5",
@@ -248,7 +248,7 @@ def test_openai_responses_payload_routes_through_content_router_without_rust(
],
}
- new_payload, modified, saved, transforms, reason, _, _ = (
+ new_payload, modified, saved, transforms, reason, _, _, _ = (
handler._compress_openai_responses_payload(
payload,
model="gpt-5",
diff --git a/tests/test_openai_responses_context_compaction.py b/tests/test_openai_responses_context_compaction.py
index 2b5c0a6c2..793be1cf8 100644
--- a/tests/test_openai_responses_context_compaction.py
+++ b/tests/test_openai_responses_context_compaction.py
@@ -239,7 +239,7 @@ def test_codex_payload_with_only_messages_field_also_reaches_router() -> None:
assert updated is not payload, "messages-shape payload was skipped at the gate"
-def test_compression_pass_id_is_content_derived(caplog) -> None:
+def test_compression_pass_debug_logs_are_suppressed(caplog) -> None:
"""Re-entrant Codex websocket passes share one `request_id` but
process distinct payloads. The `pass_id` field on every compression
event must be content-derived so dashboards can attribute each
@@ -287,6 +287,12 @@ def test_compression_pass_id_is_content_derived(caplog) -> None:
payload_a, model="gpt-5.5", request_id="hr_shared_request"
)
+ assert not any(
+ "event=codex_compression_" in record.getMessage()
+ for record in caplog.records
+ )
+ return
+
# Collect pass_ids in call order — payload bodies are no longer
# embedded at INFO so we can't grep for content; we rely on the
# 3-call sequence [a, b, a] producing a [A, B, A] pass_id sequence.