Deploy Project Epsilon Space bundle

.env.app.example

@@ -1,3 +1,11 @@
 OPENROUTER_API_KEY=
+OPENROUTER_MODEL=google/gemma-4-31b-it
+OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
 OPENROUTER_SITE_URL=http://localhost:7860
 OPENROUTER_APP_NAME=Autonomous Executive Assistant Sandbox
+
+# Hackathon-compatible aliases. Set OPENAI_API_KEY to the same OpenRouter key
+# only when running inference.py under a validator that expects this name.
+OPENAI_API_KEY=
+API_BASE_URL=https://openrouter.ai/api/v1
+MODEL_NAME=google/gemma-4-31b-it

.env.training.example

@@ -1,6 +1,13 @@
 OPENROUTER_API_KEY=
 OPENROUTER_MODEL=google/gemma-4-31b-it
+OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
 OPENROUTER_SITE_URL=http://localhost:8888
 OPENROUTER_APP_NAME=Autonomous Executive Assistant Sandbox Training
 OPENROUTER_TEMPERATURE=0.1
 OPENROUTER_MAX_TOKENS=600
+
+# Hackathon-compatible aliases for inference.py. OPENAI_API_KEY should contain
+# the same OpenRouter API key when using the validator-facing script.
+OPENAI_API_KEY=
+API_BASE_URL=https://openrouter.ai/api/v1
+MODEL_NAME=google/gemma-4-31b-it

README.md

@@ -65,7 +65,7 @@ To make that possible under hackathon constraints, we replaced live services wit
 - **Environment state:** in-memory SQLite workspace simulating emails, todos, files, and action history
 - **OpenEnv contract:** Pydantic models defining observations, actions, rewards, and policy decisions
 - **Execution loop:** shared `EpisodeRunner` used by tests, scripts, notebook experiments, and the Gradio app
-- **Policies:** deterministic baseline, tabular RL checkpoint replay, and optional OpenRouter-backed live policy execution
+- **Policies:** deterministic baseline, tabular RL checkpoint replay, and optional OpenRouter-backed live policy execution through the OpenAI client compatibility layer
 - **UI layer:** Gradio control room plus visible workspace snapshots for judges
 
 ## Seeded Judge Tasks
@@ -95,6 +95,8 @@ The environment includes a stakeholder email asking for exact metrics from a loc
 - App port: `7860`
 - Entry point: `python app.py`
 - Optional secret: `OPENROUTER_API_KEY`
+- OpenAI-compatible base URL: `https://openrouter.ai/api/v1`
+- Model: `google/gemma-4-31b-it`
 - A trained RL checkpoint is bundled in `artifacts/checkpoints/` so the `rl` policy is available immediately in the demo.
 - The bundled RL artifact lives at `artifacts/checkpoints/q_policy_notebook.json`
 - The Space is deployed from the same repository used for local tests and notebook-backed experiments

docs/HF_SPACE_README.md

@@ -65,7 +65,7 @@ To make that possible under hackathon constraints, we replaced live services wit
 - **Environment state:** in-memory SQLite workspace simulating emails, todos, files, and action history
 - **OpenEnv contract:** Pydantic models defining observations, actions, rewards, and policy decisions
 - **Execution loop:** shared `EpisodeRunner` used by tests, scripts, notebook experiments, and the Gradio app
-- **Policies:** deterministic baseline, tabular RL checkpoint replay, and optional OpenRouter-backed live policy execution
+- **Policies:** deterministic baseline, tabular RL checkpoint replay, and optional OpenRouter-backed live policy execution through the OpenAI client compatibility layer
 - **UI layer:** Gradio control room plus visible workspace snapshots for judges
 
 ## Seeded Judge Tasks
@@ -95,6 +95,8 @@ The environment includes a stakeholder email asking for exact metrics from a loc
 - App port: `7860`
 - Entry point: `python app.py`
 - Optional secret: `OPENROUTER_API_KEY`
+- OpenAI-compatible base URL: `https://openrouter.ai/api/v1`
+- Model: `google/gemma-4-31b-it`
 - Bundled RL checkpoint path: `artifacts/checkpoints/q_policy_notebook.json`
 - The Space is deployed from the same repository used for local tests and notebook-backed experiments
 

inference.py

@@ -17,15 +17,25 @@ TASKS = [
 
 
 def build_openai_compatible_policy() -> OpenRouterPolicy:
-    api_key = os.environ.get("OPENAI_API_KEY", "").strip()
-    base_url = os.environ.get("API_BASE_URL", "").strip()
-    model_name = os.environ.get("MODEL_NAME", "").strip()
+    api_key = os.environ.get("OPENROUTER_API_KEY", "").strip() or os.environ.get(
+        "OPENAI_API_KEY", ""
+    ).strip()
+    base_url = (
+        os.environ.get("OPENROUTER_BASE_URL", "").strip()
+        or os.environ.get("API_BASE_URL", "").strip()
+        or "https://openrouter.ai/api/v1"
+    )
+    model_name = (
+        os.environ.get("OPENROUTER_MODEL", "").strip()
+        or os.environ.get("MODEL_NAME", "").strip()
+        or "google/gemma-4-31b-it"
+    )
     if not api_key:
-        raise RuntimeError("OPENAI_API_KEY is required.")
+        raise RuntimeError("OPENROUTER_API_KEY or OPENAI_API_KEY is required.")
     if not base_url:
-        raise RuntimeError("API_BASE_URL is required.")
+        raise RuntimeError("API_BASE_URL or OPENROUTER_BASE_URL is required.")
     if not model_name:
-        raise RuntimeError("MODEL_NAME is required.")
+        raise RuntimeError("MODEL_NAME or OPENROUTER_MODEL is required.")
     config = OpenRouterConfig(
         api_key=api_key,
         base_url=base_url,

scripts/deploy_hf_space.py

@@ -40,14 +40,14 @@ def build_parser() -> argparse.ArgumentParser:
     )
     parser.add_argument(
         "--team-name",
-        default=os.environ.get("HF_SPACE_TEAM_NAME", "Project Epsilon"),
+        default=os.environ.get("HF_SPACE_TEAM_NAME", "Team Epsilon"),
         help="Team name shown in the generated HF README.",
     )
     parser.add_argument(
         "--hf-usernames",
         default=os.environ.get(
             "HF_SPACE_TEAM_USERNAMES",
-            "HF_USERNAME_1,HF_USERNAME_2,HF_USERNAME_3",
+            "flickinshots,ShreyaKhatik,itsayushdey",
         ),
         help="Comma-separated HF usernames for the HF README placeholders.",
     )
@@ -61,6 +61,16 @@ def build_parser() -> argparse.ArgumentParser:
         default=os.environ.get("OPENROUTER_API_KEY", "").strip(),
         help="Optional secret to set on the Space during deployment.",
     )
+    parser.add_argument(
+        "--api-base-url",
+        default=os.environ.get("API_BASE_URL", "https://openrouter.ai/api/v1").strip(),
+        help="OpenAI-compatible API base URL for inference.py. Defaults to OpenRouter.",
+    )
+    parser.add_argument(
+        "--model-name",
+        default=os.environ.get("MODEL_NAME", "google/gemma-4-31b-it").strip(),
+        help="OpenRouter model id for inference.py. Defaults to Gemma 4.",
+    )
     parser.add_argument(
         "--private",
         action="store_true",
@@ -159,6 +169,9 @@ def main() -> int:
         if checkpoint_path is not None:
             messages.append(f"Bundled RL checkpoint: {checkpoint_path.relative_to(stage_dir)}")
         messages.append(maybe_set_space_secret(api, config.repo_id, "OPENROUTER_API_KEY", args.openrouter_api_key))
+        messages.append(maybe_set_space_secret(api, config.repo_id, "OPENAI_API_KEY", args.openrouter_api_key))
+        messages.append(maybe_set_space_variable(api, config.repo_id, "API_BASE_URL", args.api_base_url))
+        messages.append(maybe_set_space_variable(api, config.repo_id, "MODEL_NAME", args.model_name))
         messages.append(maybe_set_space_variable(api, config.repo_id, "OPENROUTER_APP_NAME", config.title))
         messages.append(maybe_set_space_variable(api, config.repo_id, "OPENROUTER_SITE_URL", config.app_url))
 

src/executive_assistant/deployment.py

@@ -155,7 +155,7 @@ To make that possible under hackathon constraints, we replaced live services wit
 - **Environment state:** in-memory SQLite workspace simulating emails, todos, files, and action history
 - **OpenEnv contract:** Pydantic models defining observations, actions, rewards, and policy decisions
 - **Execution loop:** shared `EpisodeRunner` used by tests, scripts, notebook experiments, and the Gradio app
-- **Policies:** deterministic baseline, tabular RL checkpoint replay, and optional OpenRouter-backed live policy execution
+- **Policies:** deterministic baseline, tabular RL checkpoint replay, and optional OpenRouter-backed live policy execution through the OpenAI client compatibility layer
 - **UI layer:** Gradio control room plus visible workspace snapshots for judges
 
 ## Seeded Judge Tasks
@@ -185,6 +185,8 @@ The environment includes a stakeholder email asking for exact metrics from a loc
 - App port: `{config.app_port}`
 - Entry point: `python app.py`
 - Optional secret: `OPENROUTER_API_KEY`
+- OpenAI-compatible base URL: `https://openrouter.ai/api/v1`
+- Model: `google/gemma-4-31b-it`
 - {checkpoint_note}
 - The bundled RL artifact lives at `artifacts/checkpoints/{config.checkpoint_name}`
 - The Space is deployed from the same repository used for local tests and notebook-backed experiments

tests/test_inference.py

@@ -4,6 +4,19 @@ from inference import build_openai_compatible_policy
 from src.executive_assistant.config import OpenRouterConfig
 
 
+def test_inference_prefers_openrouter_api_with_openai_client(monkeypatch) -> None:
+    monkeypatch.setenv("OPENROUTER_API_KEY", "openrouter-key")
+    monkeypatch.setenv("OPENROUTER_BASE_URL", "https://openrouter.ai/api/v1")
+    monkeypatch.setenv("OPENROUTER_MODEL", "google/gemma-4-31b-it")
+    monkeypatch.setenv("OPENAI_API_KEY", "wrong-openai-key")
+    monkeypatch.setenv("API_BASE_URL", "https://api.openai.com/v1")
+    monkeypatch.setenv("MODEL_NAME", "wrong-model")
+    policy = build_openai_compatible_policy()
+    assert policy.config.api_key == "openrouter-key"
+    assert policy.config.base_url == "https://openrouter.ai/api/v1"
+    assert policy.config.model_name == "google/gemma-4-31b-it"
+
+
 def test_openrouter_config_accepts_hackathon_env_names(monkeypatch) -> None:
     monkeypatch.delenv("OPENROUTER_API_KEY", raising=False)
     monkeypatch.delenv("OPENROUTER_BASE_URL", raising=False)