feat: enable persistent model checkpoint discovery and upload API for HF storage

.env.example

@@ -44,3 +44,8 @@ LLM_CALL_DELAY=12.0
 # For Hugging Face persistent storage, set OPENENV_DATA_DIR=/data/openenv_rl
 STORAGE_ENABLED=true
 OPENENV_DATA_DIR=
+
+# Optional CSV list of extra model directories to scan for RL checkpoints.
+# Example:
+# OPENENV_MODEL_SEARCH_DIRS=/data/openenv_rl/results/best_model/phase1,/data/openenv_rl/results/best_model/phase2
+OPENENV_MODEL_SEARCH_DIRS=

app/main.py

@@ -41,7 +41,7 @@ import time
 from typing import Any, Literal
 from uuid import uuid4
 
-from fastapi import APIRouter, Body, FastAPI, HTTPException, Query, status
+from fastapi import APIRouter, Body, FastAPI, File, HTTPException, Query, UploadFile, status
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.routing import APIRoute
 from fastapi.responses import FileResponse, RedirectResponse, StreamingResponse
@@ -427,12 +427,55 @@ def _log_line_text(value: Any) -> str:
 
 
 def _phase_model_dirs() -> list[Path]:
-    base = REPO_ROOT / "results" / "best_model"
-    return [
-        base / "phase1",
-        base / "phase2",
+    """
+    Discover model directories from multiple roots.
+
+    Priority:
+    1) Explicit OPENENV_MODEL_SEARCH_DIRS (CSV of absolute/relative paths)
+    2) Persistent storage root OPENENV_DATA_DIR (HF bucket mount recommended)
+    3) Repo-local results/best_model
+    """
+    configured_dirs = (os.getenv("OPENENV_MODEL_SEARCH_DIRS") or "").strip()
+    configured: list[Path] = []
+    if configured_dirs:
+        for raw in configured_dirs.split(","):
+            s = raw.strip()
+            if not s:
+                continue
+            p = Path(s)
+            if not p.is_absolute():
+                p = (REPO_ROOT / p).resolve()
+            configured.append(p)
+
+    data_root_raw = (os.getenv("OPENENV_DATA_DIR") or "/data/openenv_rl").strip()
+    data_root = Path(data_root_raw)
+    if not data_root.is_absolute():
+        data_root = (REPO_ROOT / data_root).resolve()
+
+    repo_base = REPO_ROOT / "results" / "best_model"
+    data_base = data_root / "results" / "best_model"
+
+    candidates = [
+        *configured,
+        data_base / "phase1",
+        data_base / "phase2",
+        data_root / "best_model" / "phase1",
+        data_root / "best_model" / "phase2",
+        repo_base / "phase1",
+        repo_base / "phase2",
     ]
 
+    # Preserve order, remove duplicates.
+    deduped: list[Path] = []
+    seen: set[str] = set()
+    for p in candidates:
+        key = str(p.resolve()) if p.exists() else str(p)
+        if key in seen:
+            continue
+        seen.add(key)
+        deduped.append(p)
+    return deduped
+
 
 def _discover_phase12_zip_models() -> list[Path]:
     discovered: list[Path] = []
@@ -446,6 +489,14 @@ def _discover_phase12_zip_models() -> list[Path]:
     return unique
 
 
+def _model_storage_base_dir() -> Path:
+    data_root_raw = (os.getenv("OPENENV_DATA_DIR") or "/data/openenv_rl").strip()
+    data_root = Path(data_root_raw)
+    if not data_root.is_absolute():
+        data_root = (REPO_ROOT / data_root).resolve()
+    return data_root / "results" / "best_model"
+
+
 def _phase_from_model_path(path: Path) -> int:
     parent = path.parent.name.lower()
     if parent == "phase1":
@@ -1751,7 +1802,7 @@ def api_workflow_components() -> WorkflowComponentsResponse:
     repo_root   = REPO_ROOT
     baseline_f  = repo_root / "baseline_openai.py"
     inference_f = repo_root / "inference.py"
-    phase2_model = repo_root / "results" / "best_model" / "phase2_final.zip"
+    phase2_model = next((p for p in _discover_phase12_zip_models() if _phase_from_model_path(p) == 2), None)
     components = [
         WorkflowComponentStatus(
             component="baseline_openai.py",
@@ -1770,8 +1821,12 @@ def api_workflow_components() -> WorkflowComponentsResponse:
         WorkflowComponentStatus(
             component="phase2_final.zip",
             description="Trained Phase 2 PPO checkpoint used for local RL evaluation/execution.",
-            available=phase2_model.exists(),
-            command=r".\.venv\3.11\Scripts\python.exe -m rl.evaluate --model results/best_model/phase2_final.zip --episodes 3 --model-type maskable",
+            available=phase2_model is not None,
+            command=(
+                f".\\.venv\\3.11\\Scripts\\python.exe -m rl.evaluate --model {phase2_model} --episodes 3 --model-type maskable"
+                if phase2_model is not None
+                else r".\.venv\3.11\Scripts\python.exe -m rl.evaluate --model results/best_model/phase2_final.zip --episodes 3 --model-type maskable"
+            ),
         ),
         WorkflowComponentStatus(
             component="openenv-api",
@@ -1939,6 +1994,45 @@ def api_rl_models() -> RLModelsResponse:
     return RLModelsResponse(models=models)
 
 
+@api.post("/rl_models/upload", summary="Upload RL checkpoint zip to persistent storage")
+async def api_rl_model_upload(
+    phase: int = Query(..., ge=1, le=2, description="Model phase bucket (1 or 2)"),
+    file: UploadFile = File(..., description="Checkpoint zip file"),
+) -> dict[str, Any]:
+    name = (file.filename or "").strip()
+    if not name:
+        raise HTTPException(status_code=status.HTTP_422_UNPROCESSABLE_CONTENT, detail="Missing filename.")
+    if not name.lower().endswith(".zip"):
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="Only .zip checkpoint files are accepted.",
+        )
+
+    safe_name = Path(name).name
+    base_dir = _model_storage_base_dir()
+    target_dir = base_dir / f"phase{phase}"
+    target_dir.mkdir(parents=True, exist_ok=True)
+    target_path = target_dir / safe_name
+
+    total = 0
+    with target_path.open("wb") as out:
+        while True:
+            chunk = await file.read(1024 * 1024)
+            if not chunk:
+                break
+            out.write(chunk)
+            total += len(chunk)
+    await file.close()
+
+    return {
+        "saved": True,
+        "phase": phase,
+        "filename": safe_name,
+        "size_bytes": total,
+        "path": str(target_path),
+    }
+
+
 @api.get(
     "/rl/models",
     response_model=list[ModelInfo],

requirements.txt

@@ -12,3 +12,4 @@ anyio>=4.0,<5.0
 PyYAML>=6.0,<7.0
 sse-starlette>=2.1,<3.0
 numpy>=1.26,<3.0
+python-multipart>=0.0.9,<1.0