Bill inference to an org + fix per-project .env loading

README.md

@@ -33,9 +33,10 @@ Create a `.env` file in the project root (or export these in your shell):
 ```bash
 HF_TOKEN=<your-hugging-face-token> # HF Router inference + Hub actions
 GITHUB_TOKEN=<github-personal-access-token>
+HF_BILL_TO=<org-name> # optional: bill Inference Providers usage to an org
 ```
 
-All API-based model calls go through Hugging Face [Inference Providers](https://huggingface.co/docs/inference-providers/en/index), so your `HF_TOKEN` must be allowed to make Inference Provider calls. If no `HF_TOKEN` is set, the CLI will prompt you to paste one on first launch unless you start on a local model. To get a `GITHUB_TOKEN` follow the tutorial [here](https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/managing-your-personal-access-tokens#creating-a-fine-grained-personal-access-token). See the [local models section below](#local-models) for instructions on using agents that run on your hardware.
+All API-based model calls go through Hugging Face [Inference Providers](https://huggingface.co/docs/inference-providers/en/index), so your `HF_TOKEN` must be allowed to make Inference Provider calls. Set `HF_BILL_TO` to an org you belong to to charge inference to that org's credits instead of your personal monthly allowance — this sends an `X-HF-Bill-To` header on every router call (main agent, research sub-agent, and compaction alike). If no `HF_TOKEN` is set, the CLI will prompt you to paste one on first launch unless you start on a local model. To get a `GITHUB_TOKEN` follow the tutorial [here](https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/managing-your-personal-access-tokens#creating-a-fine-grained-personal-access-token). See the [local models section below](#local-models) for instructions on using agents that run on your hardware.
 
 ### Usage
 

agent/config.py

@@ -4,7 +4,7 @@
 from pathlib import Path
 from typing import Any, Literal, Union
 
-from dotenv import load_dotenv
+from dotenv import find_dotenv, load_dotenv
 from fastmcp.mcp_config import (
     RemoteMCPServer,
     StdioMCPServer,
@@ -207,10 +207,12 @@ def load_config(
     Use ${VAR_NAME} in your JSON for any secret.
     Automatically loads from .env file.
     """
-    # Load .env from project root first (so it works from any directory),
-    # then CWD .env can override if present
+    # Repo .env first, then let the launch directory's .env fill in any gaps.
+    # find_dotenv needs usecwd=True to look at the launch CWD, not this file.
     load_dotenv(_PROJECT_ROOT / ".env")
-    load_dotenv(override=False)
+    cwd_dotenv = find_dotenv(usecwd=True)
+    if cwd_dotenv:
+        load_dotenv(cwd_dotenv, override=False)
 
     raw_config = _load_json_config(Path(config_path))
     if include_user_defaults:

agent/core/llm_params.py

@@ -33,6 +33,18 @@ def _resolve_hf_router_token(session_hf_token: str | None = None) -> str | None:
 # an accepted-looking value, so this stays intentionally small and generic.
 _HF_EFFORTS = {"low", "medium", "high"}
 
+# When ``HF_BILL_TO`` names an org you belong to, the router charges that org's
+# credits instead of your personal allowance — same header huggingface_hub's
+# ``bill_to=`` sets, applied here because we go through LiteLLM, not InferenceClient.
+HF_BILL_TO_ENV = "HF_BILL_TO"
+HF_BILL_TO_HEADER = "X-HF-Bill-To"
+
+
+def _resolve_hf_bill_to() -> str | None:
+    """Org to bill HF Inference Providers usage to, or None if unset."""
+    value = os.environ.get(HF_BILL_TO_ENV)
+    return value.strip() or None if value else None
+
 
 def _hf_router_effort_level(reasoning_effort: str) -> str:
     level = "low" if reasoning_effort == "minimal" else reasoning_effort
@@ -120,6 +132,10 @@ def _resolve_llm_params(
       1. session.hf_token — the user's own token (CLI / OAuth / cache file).
       2. huggingface_hub cache — ``HF_TOKEN`` / ``HUGGING_FACE_HUB_TOKEN`` /
          local ``hf auth login`` cache.
+
+    When ``HF_BILL_TO`` is set, an ``X-HF-Bill-To`` header is attached to
+    HF-router calls so Inference Providers usage is charged to that org's
+    credits instead of the token owner's personal monthly allowance.
     """
     normalized_model = strip_huggingface_model_prefix(model_name) or model_name
 
@@ -136,6 +152,9 @@ def _resolve_llm_params(
         "api_base": HF_ROUTER_BASE_URL,
         "api_key": api_key,
     }
+    bill_to = _resolve_hf_bill_to()
+    if bill_to:
+        params["extra_headers"] = {HF_BILL_TO_HEADER: bill_to}
     if reasoning_effort:
         hf_level = _hf_router_effort_level(reasoning_effort)
         if hf_level not in _HF_EFFORTS:

tests/unit/test_llm_params.py

@@ -2,13 +2,21 @@
 
 from agent.core.hf_tokens import resolve_hf_request_token
 from agent.core.llm_params import (
+    HF_BILL_TO_ENV,
+    HF_BILL_TO_HEADER,
     UnsupportedEffortError,
     _resolve_hf_router_token,
     _resolve_llm_params,
 )
 from agent.core.model_ids import HF_ROUTER_BASE_URL
 
 
+@pytest.fixture(autouse=True)
+def _clear_bill_to(monkeypatch):
+    """Keep the developer's own HF_BILL_TO out of the default-behavior tests."""
+    monkeypatch.delenv(HF_BILL_TO_ENV, raising=False)
+
+
 def test_hf_router_params_for_default_model_uses_session_token():
     params = _resolve_llm_params(
         "anthropic/claude-opus-4.8:fal-ai",
@@ -63,13 +71,39 @@ def test_router_params_fall_back_to_hf_cache_when_session_token_missing(monkeypa
     assert "extra_headers" not in params
 
 
-def test_router_params_never_set_bill_to_headers():
+def test_router_params_omit_bill_to_header_when_env_unset():
     params = _resolve_llm_params("moonshotai/Kimi-K2.7-Code", "session-token")
 
     assert params["api_key"] == "session-token"
     assert "extra_headers" not in params
 
 
+def test_router_params_add_bill_to_header_when_env_set(monkeypatch):
+    monkeypatch.setenv(HF_BILL_TO_ENV, "my-org")
+
+    params = _resolve_llm_params("moonshotai/Kimi-K2.7-Code", "session-token")
+
+    assert params["extra_headers"] == {HF_BILL_TO_HEADER: "my-org"}
+
+
+def test_router_params_strip_whitespace_and_ignore_blank_bill_to(monkeypatch):
+    monkeypatch.setenv(HF_BILL_TO_ENV, "  my-org  ")
+    assert _resolve_llm_params("moonshotai/Kimi-K2.7-Code")["extra_headers"] == {
+        HF_BILL_TO_HEADER: "my-org"
+    }
+
+    monkeypatch.setenv(HF_BILL_TO_ENV, "   ")
+    assert "extra_headers" not in _resolve_llm_params("moonshotai/Kimi-K2.7-Code")
+
+
+def test_local_model_params_never_get_bill_to_header(monkeypatch):
+    monkeypatch.setenv(HF_BILL_TO_ENV, "my-org")
+
+    params = _resolve_llm_params("ollama/llama3.1:8b")
+
+    assert "extra_headers" not in params
+
+
 def test_huggingface_prefix_is_stripped_for_router_calls():
     params = _resolve_llm_params("huggingface/openai/gpt-5.5:fal-ai")