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.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
Fixing the build AND the perf logs
Tejas Chopra76e2cb1928ac
diff --git a/headroom/perf/analyzer.py b/headroom/perf/analyzer.pyindex cd65c53c8..862acff27 100644--- a/headroom/perf/analyzer.py+++ b/headroom/perf/analyzer.py@@ -39,8 +39,18 @@ _TOIN_RE = re.compile(def _parse_kv(kv_str: str) -> dict[str, str]:- """Parse key=value pairs from a PERF log line."""+ """Parse key=value pairs from a PERF log line.++ The ``transforms=`` field is always last and its value may contain spaces+ (e.g. ``transforms=router:excluded:tool*32 read_lifecycle:stale*17``).+ Everything after ``transforms=`` is captured as a single value.+ """result: dict[str, str] = {}+ # Handle transforms= specially since its value contains spaces+ if "transforms=" in kv_str:+ before, transforms_val = kv_str.split("transforms=", 1)+ result["transforms"] = transforms_val.strip()+ kv_str = beforefor part in kv_str.split():if "=" in part:k, v = part.split("=", 1)@@ -143,7 +153,21 @@ def parse_log_files(last_n_hours: float = 168.0) -> PerfReport:if m:kv = _parse_kv(m.group("kv"))transforms_str = kv.get("transforms", "none")- transforms = transforms_str.split(",") if transforms_str != "none" else []+ # Handle both old comma-separated and new space-separated *N format+ if transforms_str == "none":+ transforms: list[str] = []+ elif "*" in transforms_str or " " in transforms_str:+ # New format: "router:excluded:tool*32 read_lifecycle:stale*17"+ transforms = []+ for part in transforms_str.split():+ if "*" in part:+ name, _ = part.rsplit("*", 1)+ else:+ name = part+ transforms.append(name)+ else:+ # Old comma-separated format+ transforms = transforms_str.split(",")report.perf_records.append(PerfRecord(timestamp=m.group("ts"),diff --git a/headroom/proxy/server.py b/headroom/proxy/server.pyindex 241ca2b36..13415e543 100644--- a/headroom/proxy/server.py+++ b/headroom/proxy/server.py@@ -175,6 +175,22 @@ def _setup_file_logging() -> None:_setup_file_logging()++def _summarize_transforms(transforms: list[str]) -> str:+ """Collapse repeated transforms into counted summary.++ e.g. ['router:excluded:tool', 'router:excluded:tool', 'read_lifecycle:stale']+ → 'router:excluded:tool*2 read_lifecycle:stale'+ """+ if not transforms:+ return "none"+ counts: dict[str, int] = {}+ for t in transforms:+ counts[t] = counts.get(t, 0) + 1+ parts = [f"{k}*{v}" if v > 1 else k for k, v in counts.items()]+ return " ".join(parts)++# Maximum request body size (100MB - increased to support image-heavy requests)MAX_REQUEST_BODY_SIZE = 100 * 1024 * 1024@@ -2286,7 +2302,7 @@ class HeadroomProxy:f"tok_saved={tokens_saved} "f"cache_read={cr} cache_write={cw} cache_hit_pct={chp} "f"opt_ms={optimization_latency:.0f} "- f"transforms={','.join(transforms_applied) if transforms_applied else 'none'}"+ f"transforms={_summarize_transforms(transforms_applied)}")# Remove compression headers since httpx already decompressed the response@@ -3909,7 +3925,7 @@ class HeadroomProxy:f"cache_read={cache_read_tokens} cache_write={cache_write_tokens} "f"cache_hit_pct={cache_hit_pct} "f"opt_ms={optimization_latency:.0f} "- f"transforms={','.join(transforms_applied) if transforms_applied else 'none'}"+ f"transforms={_summarize_transforms(transforms_applied)}")# Normalize input tokens based on provider semantics:@@ -4098,7 +4114,7 @@ class HeadroomProxy:f"tok_saved={tokens_saved} "f"cache_read=0 cache_write=0 cache_hit_pct=0 "f"opt_ms={optimization_latency:.0f} "- f"transforms={','.join(transforms_applied) if transforms_applied else 'none'}"+ f"transforms={_summarize_transforms(transforms_applied)}")return StreamingResponse(