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
fix(proxy): let model discovery honor the custom base URL (#4330)
Vaibhav Hajare911c75ffdd63
diff --git a/headroom/providers/proxy_routes.py b/headroom/providers/proxy_routes.pyindex dd444cb0a..470f73020 100644--- a/headroom/providers/proxy_routes.py+++ b/headroom/providers/proxy_routes.py@@ -9,6 +9,7 @@ from typing import Anyfrom fastapi import FastAPI, HTTPException, Request, WebSocketfrom fastapi.responses import Response+from headroom.log_safety import redact_urlfrom headroom.providers.cloudcode import normalize_cloudcode_passthrough_pathfrom headroom.providers.codex.endpoints import codex_backend_urlfrom headroom.providers.codex.headers import drop_header@@ -528,7 +529,7 @@ def register_provider_routes(app: FastAPI, proxy: Any) -> None:custom_base = request.headers.get("x-headroom-base-url", "").strip()if custom_base:if not await is_safe_upstream_url_async(custom_base):- logger.warning("rejecting unsafe x-headroom-base-url: %r", custom_base)+ logger.warning("rejecting unsafe x-headroom-base-url: %s", redact_url(custom_base))raise HTTPException(status_code=400, detail="Rejected unsafe upstream base URL")# Only the destination is overridden: keep the provider the request's# auth selects, so Anthropic-auth requests still get model-iddiff --git a/tests/test_provider_proxy_routes.py b/tests/test_provider_proxy_routes.pyindex 855d6a564..bb109d65d 100644--- a/tests/test_provider_proxy_routes.py+++ b/tests/test_provider_proxy_routes.py@@ -6,6 +6,7 @@ from typing import Anyfrom unittest.mock import patchimport httpx+import pytestfrom fastapi.responses import JSONResponsefrom fastapi.testclient import TestClient@@ -1444,3 +1445,118 @@ def test_anthropic_messages_strips_ansi_model_id_before_upstream() -> None:assert response.status_code == 200assert fake_http_client.bodies[0]["model"] == "claude-opus-4-8"+++def test_model_metadata_routes_honor_the_custom_base_url(monkeypatch) -> None:+ # A client that names its own upstream must get its model list from that+ # upstream, not from the proxy's default OpenAI one: the chat and+ # responses handlers already honor the header (#4302).+ monkeypatch.setenv("HEADROOM_ALLOWED_BASE_URLS", "custom.example")+ calls: list[tuple[str, str, str]] = []++ async def fake_passthrough(self, request, base_url, sub_path="", provider_name=""): # type: ignore[no-untyped-def]+ calls.append((request.url.path, base_url, provider_name))+ return JSONResponse(+ {+ "path": request.url.path,+ "base_url": base_url,+ "sub_path": sub_path,+ "provider": provider_name,+ }+ )++ monkeypatch.setattr(HeadroomProxy, "handle_passthrough", fake_passthrough)++ with TestClient(_app()) as client:+ listed = client.get(+ "/v1/models",+ headers={"x-headroom-base-url": "https://custom.example"},+ )+ fetched = client.get(+ "/v1/models/gpt-5",+ headers={"x-headroom-base-url": "https://custom.example/"},+ )+ # An Anthropic-compatible gateway names its upstream the same way.+ anthropic = client.get(+ "/v1/models",+ headers={+ "x-api-key": "gateway-key",+ "x-headroom-base-url": "https://custom.example",+ },+ )++ assert listed.json()["base_url"] == "https://custom.example"+ assert fetched.json()["base_url"] == "https://custom.example"+ assert calls[0][2] == "openai"+ assert anthropic.json()["base_url"] == "https://custom.example"+ assert calls[-1][2] == "anthropic"+++def test_model_metadata_routes_reject_an_unsafe_custom_base_url(monkeypatch) -> None:+ # A client-named upstream goes through the same SSRF guard as the other+ # routes: an unsafe one is refused, and the request never reaches an+ # upstream. Falling back to the configured upstream would have sent the+ # authenticated request, and its credentials, to a host the client did+ # not name.+ monkeypatch.setenv("HEADROOM_ALLOWED_BASE_URLS", "custom.example")+ calls: list[tuple[str, str, str]] = []++ async def fake_passthrough(self, request, base_url, sub_path="", provider_name=""): # type: ignore[no-untyped-def]+ calls.append((request.url.path, base_url, provider_name))+ return JSONResponse(+ {+ "path": request.url.path,+ "base_url": base_url,+ "sub_path": sub_path,+ "provider": provider_name,+ }+ )++ monkeypatch.setattr(HeadroomProxy, "handle_passthrough", fake_passthrough)++ credential = {"authorization": "Bearer synthetic-private-gateway-key"}+ with TestClient(_app()) as client:+ listed = client.get(+ "/v1/models",+ headers={**credential, "x-headroom-base-url": "http://169.254.169.254"},+ )+ fetched = client.get(+ "/v1/models/gpt-5",+ headers={**credential, "x-headroom-base-url": "http://169.254.169.254"},+ )++ assert listed.status_code == 400+ assert listed.json()["detail"] == "Rejected unsafe upstream base URL"+ assert fetched.status_code == 400+ # No upstream call at all: the rejected request must not be redirected to+ # the configured upstream, which would disclose the credential.+ assert calls == []++[email protected]("path", ["/v1/models", "/v1/models/gpt-5"])+def test_model_metadata_rejection_redacts_url_credentials(monkeypatch, caplog, path) -> None:+ """Reject unsafe destinations without exposing URL credentials in the warning."""+ monkeypatch.setenv("HEADROOM_ALLOWED_BASE_URLS", "")+ calls = []++ async def unexpected_passthrough(*args, **kwargs):+ calls.append(True)+ raise AssertionError("Rejected metadata request reached upstream")++ monkeypatch.setattr(HeadroomProxy, "handle_passthrough", unexpected_passthrough)+ with caplog.at_level("WARNING"), TestClient(_app()) as client:+ response = client.get(+ path,+ headers={+ "authorization": "Bearer synthetic-gateway-key",+ "x-headroom-base-url": (+ "https://synthetic-user:[email protected]/v1"+ "?api_key=synthetic-query-secret"+ ),+ },+ )+ assert response.status_code == 400+ assert calls == []+ assert "rejecting unsafe x-headroom-base-url" in caplog.text+ for secret in ("synthetic-user", "synthetic-password", "synthetic-query-secret"):+ assert secret not in caplog.text