Fix dashboard logging URL to use proxy path, force Docker rebuild

- log_to_dashboard now posts to /api/env/training/log (through proxy)
instead of /training/log (direct, which 404s on Next.js)
- ARG CACHEBUST in Dockerfile to invalidate build cache

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Dockerfile

@@ -19,7 +19,7 @@ COPY lib/ ./lib/
 COPY public/ ./public/
 COPY package.json package-lock.json tsconfig.json next.config.ts postcss.config.mjs eslint.config.mjs .eslintrc.json next-env.d.ts metadata.json ./
 ENV NEXT_TELEMETRY_DISABLED=1
-# Build timestamp to bust cache: 2026-03-08
+ARG CACHEBUST=3
 RUN npm run build
 
 FROM python:3.11-slim AS runner

training/train_grpo.ipynb

@@ -98,7 +98,7 @@
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": "PRINTER = 0\n\ndef extract_fold_json(response):\n    \"\"\"Extract FOLD JSON from LLM response text.\"\"\"\n    m = re.search(r\"```(?:json)?\\s*(\\{.*?\\})\\s*```\", response, re.DOTALL)\n    if m:\n        try: return json.loads(m.group(1))\n        except: pass\n    m = re.search(r'\\{[^{}]*\"vertices_coords\"[^{}]*\\}', response, re.DOTALL)\n    if m:\n        try: return json.loads(m.group(0))\n        except: pass\n    try:\n        d = json.loads(response.strip())\n        if isinstance(d, dict) and \"vertices_coords\" in d: return d\n    except: pass\n    return None\n\n\ndef log_to_dashboard(task_name, reward, shape_similarity, is_valid, error=None, fold_data=None, final_positions=None, target_positions=None):\n    \"\"\"Send training step data to the frontend dashboard via /training/log.\"\"\"\n    try:\n        requests.post(\n            f\"{OPENENV_URL.replace('/api/env', '')}/training/log\",\n            json={\n                \"task_name\": task_name,\n                \"reward\": reward,\n                \"shape_similarity\": shape_similarity,\n                \"is_valid\": is_valid,\n                \"error\": error,\n                \"fold_data\": fold_data,\n                \"final_positions\": final_positions or [],\n                \"target_positions\": target_positions or [],\n            },\n            timeout=5,\n        )\n    except:\n        pass  # Don't let dashboard logging break training\n\n\ndef valid_fold_reward(completions, **kwargs):\n    \"\"\"Reward 1 (local): +1.0 valid FOLD structure, -0.5 bad structure, -2.0 unparseable.\"\"\"\n    REQUIRED = {\"vertices_coords\", \"edges_vertices\", \"edges_assignment\"}\n    scores = []\n    for c in completions:\n        fold = extract_fold_json(c[0][\"content\"])\n        if fold is None:\n            scores.append(-2.0)\n            continue\n        # Basic structural checks\n        if not REQUIRED.issubset(fold.keys()):\n            scores.append(-0.5); continue\n        verts = fold[\"vertices_coords\"]\n        edges = fold[\"edges_vertices\"]\n        asgn = fold[\"edges_assignment\"]\n        if len(verts) < 3 or len(edges) < 3 or len(edges) != len(asgn):\n            scores.append(-0.5); continue\n        if not any(a in (\"M\",\"V\") for a in asgn):\n            scores.append(-0.5); continue\n        if not any(a == \"B\" for a in asgn):\n            scores.append(-0.5); continue\n        scores.append(1.0)\n    return scores\n\n\ndef openenv_reward(completions, task_name, **kwargs):\n    \"\"\"Reward 2 (OpenEnv API): Submit fold to environment, get simulation reward.\n\n    Calls POST /reset and POST /step on the HF Space OpenEnv environment.\n    The environment runs the fold simulation and computes shape similarity.\n    Also logs each step to the frontend training dashboard.\n    \"\"\"\n    global PRINTER\n    # task_name comes as a list from the dataset\n    tn = task_name[0] if isinstance(task_name, list) else task_name\n\n    scores = []\n    for c in completions:\n        resp = c[0][\"content\"]\n\n        # Periodic logging\n        if PRINTER % 10 == 0:\n            print(f\"\\n--- [{tn}] Sample {PRINTER} ---\\n{resp[:300]}\")\n        PRINTER += 1\n\n        # Parse the FOLD JSON from the LLM response\n        fold = extract_fold_json(resp)\n        if fold is None:\n            scores.append(-2.0)\n            log_to_dashboard(tn, -2.0, 0.0, False, error=\"No JSON parsed\")\n            continue\n\n        try:\n            # Reset environment for this task\n            env.reset(task_name=tn)\n\n            # Submit the fold to OpenEnv — environment simulates and scores it\n            result = env.step(fold)\n\n            # Get reward from the environment\n            reward = result.get(\"reward\", None)\n            obs = result.get(\"observation\", {})\n            sim = obs.get(\"shape_similarity\", 0.0)\n            is_valid = not bool(obs.get(\"error\"))\n\n            if reward is not None:\n                scores.append(float(reward))\n            else:\n                if obs.get(\"error\"):\n                    scores.append(-2.0)\n                else:\n                    reward = float(sim) * 20.0\n                    scores.append(reward)\n\n            # Log to frontend dashboard\n            log_to_dashboard(\n                task_name=tn,\n                reward=float(reward) if reward is not None else scores[-1],\n                shape_similarity=float(sim),\n                is_valid=is_valid,\n                error=obs.get(\"error\"),\n                fold_data=fold,\n                final_positions=obs.get(\"final_positions\", []),\n                target_positions=obs.get(\"target_positions\", []),\n            )\n\n        except requests.exceptions.RequestException as e:\n            print(f\"OpenEnv API error: {e}\")\n            scores.append(-1.0)\n            log_to_dashboard(tn, -1.0, 0.0, False, error=str(e))\n        except Exception as e:\n            print(f\"Reward error: {e}\")\n            scores.append(-1.0)\n            log_to_dashboard(tn, -1.0, 0.0, False, error=str(e))\n\n    return scores\n\n\nprint(\"Reward functions ready (valid_fold=local, openenv_reward=API + dashboard logging).\")"
+   "source": "PRINTER = 0\n\ndef extract_fold_json(response):\n    \"\"\"Extract FOLD JSON from LLM response text.\"\"\"\n    m = re.search(r\"```(?:json)?\\s*(\\{.*?\\})\\s*```\", response, re.DOTALL)\n    if m:\n        try: return json.loads(m.group(1))\n        except: pass\n    m = re.search(r'\\{[^{}]*\"vertices_coords\"[^{}]*\\}', response, re.DOTALL)\n    if m:\n        try: return json.loads(m.group(0))\n        except: pass\n    try:\n        d = json.loads(response.strip())\n        if isinstance(d, dict) and \"vertices_coords\" in d: return d\n    except: pass\n    return None\n\n\ndef log_to_dashboard(task_name, reward, shape_similarity, is_valid, error=None, fold_data=None, final_positions=None, target_positions=None):\n    \"\"\"Send training step data to the frontend dashboard via the proxy.\"\"\"\n    try:\n        # Goes through Next.js proxy: /api/env/training/log → localhost:8000/training/log\n        requests.post(\n            f\"{OPENENV_URL}/training/log\",\n            json={\n                \"task_name\": task_name,\n                \"reward\": reward,\n                \"shape_similarity\": shape_similarity,\n                \"is_valid\": is_valid,\n                \"error\": error,\n                \"fold_data\": fold_data,\n                \"final_positions\": final_positions or [],\n                \"target_positions\": target_positions or [],\n            },\n            timeout=5,\n        )\n    except:\n        pass  # Don't let dashboard logging break training\n\n\ndef valid_fold_reward(completions, **kwargs):\n    \"\"\"Reward 1 (local): +1.0 valid FOLD structure, -0.5 bad structure, -2.0 unparseable.\"\"\"\n    REQUIRED = {\"vertices_coords\", \"edges_vertices\", \"edges_assignment\"}\n    scores = []\n    for c in completions:\n        fold = extract_fold_json(c[0][\"content\"])\n        if fold is None:\n            scores.append(-2.0)\n            continue\n        # Basic structural checks\n        if not REQUIRED.issubset(fold.keys()):\n            scores.append(-0.5); continue\n        verts = fold[\"vertices_coords\"]\n        edges = fold[\"edges_vertices\"]\n        asgn = fold[\"edges_assignment\"]\n        if len(verts) < 3 or len(edges) < 3 or len(edges) != len(asgn):\n            scores.append(-0.5); continue\n        if not any(a in (\"M\",\"V\") for a in asgn):\n            scores.append(-0.5); continue\n        if not any(a == \"B\" for a in asgn):\n            scores.append(-0.5); continue\n        scores.append(1.0)\n    return scores\n\n\ndef openenv_reward(completions, task_name, **kwargs):\n    \"\"\"Reward 2 (OpenEnv API): Submit fold to environment, get simulation reward.\n\n    Calls POST /reset and POST /step on the HF Space OpenEnv environment.\n    The environment runs the fold simulation and computes shape similarity.\n    Also logs each step to the frontend training dashboard.\n    \"\"\"\n    global PRINTER\n    # task_name comes as a list from the dataset\n    tn = task_name[0] if isinstance(task_name, list) else task_name\n\n    scores = []\n    for c in completions:\n        resp = c[0][\"content\"]\n\n        # Periodic logging\n        if PRINTER % 10 == 0:\n            print(f\"\\n--- [{tn}] Sample {PRINTER} ---\\n{resp[:300]}\")\n        PRINTER += 1\n\n        # Parse the FOLD JSON from the LLM response\n        fold = extract_fold_json(resp)\n        if fold is None:\n            scores.append(-2.0)\n            log_to_dashboard(tn, -2.0, 0.0, False, error=\"No JSON parsed\")\n            continue\n\n        try:\n            # Reset environment for this task\n            env.reset(task_name=tn)\n\n            # Submit the fold to OpenEnv — environment simulates and scores it\n            result = env.step(fold)\n\n            # Get reward from the environment\n            reward = result.get(\"reward\", None)\n            obs = result.get(\"observation\", {})\n            sim = obs.get(\"shape_similarity\", 0.0)\n            is_valid = not bool(obs.get(\"error\"))\n\n            if reward is not None:\n                scores.append(float(reward))\n            else:\n                if obs.get(\"error\"):\n                    scores.append(-2.0)\n                else:\n                    reward = float(sim) * 20.0\n                    scores.append(reward)\n\n            # Log to frontend dashboard\n            log_to_dashboard(\n                task_name=tn,\n                reward=float(reward) if reward is not None else scores[-1],\n                shape_similarity=float(sim),\n                is_valid=is_valid,\n                error=obs.get(\"error\"),\n                fold_data=fold,\n                final_positions=obs.get(\"final_positions\", []),\n                target_positions=obs.get(\"target_positions\", []),\n            )\n\n        except requests.exceptions.RequestException as e:\n            print(f\"OpenEnv API error: {e}\")\n            scores.append(-1.0)\n            log_to_dashboard(tn, -1.0, 0.0, False, error=str(e))\n        except Exception as e:\n            print(f\"Reward error: {e}\")\n            scores.append(-1.0)\n            log_to_dashboard(tn, -1.0, 0.0, False, error=str(e))\n\n    return scores\n\n\nprint(\"Reward functions ready (valid_fold=local, openenv_reward=API + dashboard logging).\")"
   },
   {
    "cell_type": "markdown",