diff --git a/examples-guards/areal/examples_manifest.yaml b/examples-guards/areal/examples_manifest.yaml
index 895f3c480..5492c0b0b 100644
--- a/examples-guards/areal/examples_manifest.yaml
+++ b/examples-guards/areal/examples_manifest.yaml
@@ -2,11 +2,8 @@ version: 1
 scan:
   root: examples
   include_extensions: ['.py', '.sh']
-  # files-only 模式（docs/examples-guard-engine.md §2.7）：scan 只保留
-  # root + include_extensions 两个键；实现模块、上游自测等一律进
-  # unsupported 段（带分组注释与理由）。
-  # 对账口径：release 树（当前 v2.1.0）114 个 .py/.sh =
-  # supported(18) + unsupported(96)。
+  # 当前 v2.1.1，扫描共 161 个
+  # 支持用例（4）+ 可看护暂不支持用例（34）+ 不可看护对象（123）
 
 # 看护 release 线（workflow target_ref 留空，由引擎按 latest release
 # 解析）。supported 配方用专用镜像 ghcr.io/hwvanici/areal_npu:v1.0.5-a2
@@ -19,134 +16,35 @@ scan:
 # - scheduler.type=null 的 yaml（math 系）必须补 scheduler.type=local。
 # - norm：reward=group + adv=batch 实测可过；reward/adv 为 batch 级的需加
 #   mean_level=null（HCCL 不支持 float64 allreduce）。
-# - setup_timeout=600 兜底数据集在线冷下载（HF_ENDPOINT=huggingface.co）。
+# - setup_timeout=600 兜底数据集在线冷下载（HF_ENDPOINT=hf-mirror.com）。
 # - rollout.setup_timeout 默认仅 300s，vLLM-Ascend 多模态 server 预热（mm
 #   processor warmup + tokenizer 深拷贝）在慢机上会超，RL 条目统一
 #   +rollout.setup_timeout=1800 兜底。
 supported:
-  # CLEVR 计数 GRPO（复用 areal-vlm-grpo profile；重数据：10.4GB 下载 +
-  # 7 万图单进程预处理，故 setup_timeout 3600 / timeout 180）
-  # 已知问题：offload+disk 的存盘聚合走 CPU gloo，会超上游 120s 硬编码等待
-  # （remote_inf_engine.py）；1.5B 条目已去 offload 规避，3B 待内存验证。
-  - path: examples/vlm/clevr_count_70k_grpo.py
-    profile: areal-vlm-grpo
+  # SWE 轨迹 SFT（纯本地，无 rollout/沙箱；数据用 fixture 的 pair-format JSONL，
+  # 路径含 swe token 自动路由进 SWE 管线；配置复用 gsm8k_sft.yaml + swe 覆盖）
+  - path: examples/swe/train_sft.py
+    profile: areal-swe-sft
     runner: linux-aarch64-a2-2
     npu_devices: '0,1'
     image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
     overlay_args:
-      - --config examples/vlm/clevr_count_70k_grpo.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - +rollout.setup_timeout=1800
-      - actor.backend=fsdp:d2p1t1
-      - +actor.fsdp.offload_params=true
-      - +actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.35
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=3600
-      - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-      - vllm.max_model_len=8576
-    timeout_minutes: 300
-  # tau2-bench 客服多轮 RL（保留真实 UserSimulator，solo_mode=false）：用户侧由
-  # run_example 在另一张卡拉本地 vLLM（同一份 Qwen3-1.7B 权重、多 served-model-name
-  # 别名吸收 econfig.user_llm 的 litellm 命名），econfig.user_llm_base_url 指过去；
-  # agent 侧照常走 AReaL proxy。max_steps/new_tokens/n_samples/batch_size 压到
-  # 冒烟量级（单集远低于入口硬编码的 600s episode 超时）。
-  - path: examples/tau2/train.py
-    profile: areal-tau2-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/tau2/config_1.7b_airline.yaml
-      - cluster.n_gpus_per_node=1
-      - rollout.backend=vllm:d1p1t1
-      - actor.backend=fsdp:d1p1t1
-      - +actor.fsdp.offload_params=true
-      - +actor.weight_update_mode=disk
-      - actor.enable_tree_training=false
-      # mb 容量必须 ≥ 单条最长轨迹（tau2 airline 轨迹 ~4978 token，1024 时
-      # ffd_allocate 直接 RuntimeError）；24576+pad_to_maximum=true 实测 OOM
-      # （4978 垫到 24576 的 5 倍 padding）——8192 + 不垫的组合。
-      - actor.pad_to_maximum=false
-      - actor.mb_spec.max_tokens_per_mb=8192
-      - ref.mb_spec.max_tokens_per_mb=8192
-      - actor.path=${AREAL_MODEL_PATH}
-      - econfig.max_steps=2
-      - econfig.user_llm_base_url=http://127.0.0.1:18888/v1/
-      - econfig.user_llm_args.max_completion_tokens=64
-      - gconfig.n_samples=1
-      - gconfig.max_new_tokens=64
-      - gconfig.max_tokens=8192
-      - train_dataset.batch_size=1
-      - valid_dataset.batch_size=1
-      - stats_logger.wandb.mode=disabled
-      - vllm.gpu_memory_utilization=0.3
-      - vllm.max_model_len=8192
-      - +rollout.setup_timeout=1800
-      - total_train_steps=1
-      - total_train_epochs=1
-    timeout_minutes: 150
-  # Qwen2.5-VL-3B Geometry3K GRPO（双卡）
-  - path: examples/vlm/geometry3k_grpo.py
-    profile: areal-vlm-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/vlm_npu/qwen2_5_vl_3b_geometry3k_grpo.yaml
+      - --config examples/math/gsm8k_sft.yaml
       - scheduler.type=local
       - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      # mm 预热会超默认 300s 健康等待（mm server 预热见头部通用配方）
-      - +rollout.setup_timeout=1800
       - actor.backend=fsdp:d2p1t1
-      - +actor.fsdp.offload_params=true
-      - +actor.weight_update_mode=disk
       - actor.path=${AREAL_MODEL_PATH}
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
+      - experiment_name=swe-sft-smoke
       - train_dataset.batch_size=2
+      - train_dataset.path=${FIXTURE_DIR}/swe_sft_pairs.jsonl
       - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=600
-      - +valid_dataset.setup_timeout=600
+      - valid_dataset.path=${FIXTURE_DIR}/swe_sft_pairs.jsonl
+      - +swe.pre_split=true
+      - +swe.dump_samples=0
       - +total_train_steps=1
-      - vllm.max_model_len=8576
-    timeout_minutes: 150
-  # GSM8K GRPO（用上游 NPU 专用 yaml）
-  - path: examples/math/gsm8k_rl.py
-    profile: areal-math-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/math/gsm8k_grpo_npu.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - actor.backend=fsdp:d2p1t1
-      - ++actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
       - +train_dataset.setup_timeout=600
       - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-    timeout_minutes: 150
+    timeout_minutes: 90
   # GSM8K SFT（纯 FSDP，无 rollout）
   - path: examples/math/gsm8k_sft.py
     profile: areal-math-sft
@@ -165,265 +63,7 @@ supported:
       - +valid_dataset.setup_timeout=600
       - +total_train_steps=1
     timeout_minutes: 150
-  # BoBa GRPO（本地 jsonl，setup 预下载到 AREAL_BOBA_DATA）
-  - path: examples/math/boba_grpo.py
-    profile: areal-math-boba
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/math/boba_grpo.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - actor.backend=fsdp:d2p1t1
-      - ++actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
-      - actor.ppo_n_minibatches=1
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - rollout.enable_rollout_tracing=false
-      - perf_tracer.enabled=false
-      - train_dataset.path=${AREAL_BOBA_DATA}
-      - train_dataset.batch_size=2
-      - +total_train_steps=1
-      - ++actor.scheduling_spec.0.env_vars.TASK_QUEUE_ENABLE=1
-    timeout_minutes: 150
-  # Countdown GRPO（数据由 setup 现场生成）
-  - path: examples/countdown/train.py
-    profile: areal-countdown-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/countdown/train_config.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - +rollout.setup_timeout=1800
-      - actor.backend=fsdp:d2p1t1
-      - +actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - +total_train_steps=1
-    timeout_minutes: 150
-  # GSM8K 多轮自我纠错 GRPO（复用 areal-math-grpo profile）
-  - path: examples/multi_turn_math/gsm8k_rl_mt.py
-    profile: areal-math-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/multi_turn_math/gsm8k_grpo_mt.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - actor.backend=fsdp:d2p1t1
-      - +actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=600
-      - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-    timeout_minutes: 150
-  # TIR 工具集成推理 GRPO（带内工具调用 + 本地 python/calculator；torl 数据
-  # 仅 GitHub 来源、job 容器不可达，parquet 以 fixture 入库）
-  - path: examples/tir/train_tir.py
-    profile: areal-tir-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/tir/tir_math_config.yaml
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - +rollout.setup_timeout=1800
-      - actor.backend=fsdp:d2p1t1
-      - +actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=128
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=600
-      - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-      - vllm.max_model_len=4096
-    timeout_minutes: 150
-  # Scaffolding RLVR（旧版 SPMD yaml：+scheduler/+backend，~ 删 legacy 键，
-  # python -m 启动）；worker 写死 model="default" 而 vLLM 按模型路径名服务会
-  # 404，故 setup 建 TARGET_ROOT/default 软链 + vllm.model=default。
-  - path: examples/scaffolding/gsm8k_rlvr_scaffolding.py
-    profile: areal-scaffold-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/scaffolding/gsm8k_rlvr_scaffolding.yaml
-      - +scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - +rollout.backend=vllm:d1p1t2
-      - +rollout.setup_timeout=1800
-      - +actor.backend=fsdp:d2p1t1
-      - +ref.backend=fsdp:d2p1t1
-      - +actor.fsdp.offload_params=true
-      - ~actor.behav_imp_weight_cap
-      - actor.path=${AREAL_MODEL_PATH}
-      - vllm.model=default
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=600
-      - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-    timeout_minutes: 150
-  # OpenAI Agents SDK 多 agent handoff GRPO（LLM 走本地 vLLM）
-  - path: examples/openai_agents/train_agents.py
-    profile: areal-agents-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/openai_agents/config.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - +rollout.setup_timeout=1800
-      - actor.backend=fsdp:d2p1t1
-      - +actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.3
-      - +gconfig.max_new_tokens=64
-      - gconfig.max_tokens=256
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=600
-      - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-    timeout_minutes: 150
   
-  # camel-ai ChatAgent 多轮 GRPO（GSM8K）。入口脚本内定义了 Config，必须由
-  # run_example.sh 以「import 成模块再调 main()」启动；camel-ai 由 setup 另装。
-  - path: examples/camel/train.py
-    profile: areal-camel-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/camel/config.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - actor.backend=fsdp:d2p1t1
-      - +actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - vllm.gpu_memory_utilization=0.3
-      - gconfig.max_new_tokens=64
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=600
-      - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-    timeout_minutes: 150
-  # Agent Workflow（OpenAI SDK agent，GSM8K）。agent 用 AReaL 注入的本地 proxy
-  # endpoint（非外部 LLM、无需 key）；复用 areal-agents-grpo。
-  - path: examples/agent_workflow/train.py
-    profile: areal-agents-grpo
-    runner: linux-aarch64-a2-2
-    npu_devices: '0,1'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/agent_workflow/config.yaml
-      - scheduler.type=local
-      - cluster.n_gpus_per_node=2
-      - rollout.backend=vllm:d1p1t2
-      - +rollout.setup_timeout=1800
-      - actor.backend=fsdp:d2p1t1
-      - ++actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - actor.mb_spec.max_tokens_per_mb=1024
-      - ref.mb_spec.max_tokens_per_mb=1024
-      - vllm.gpu_memory_utilization=0.3
-      - ++gconfig.max_new_tokens=64
-      - gconfig.max_tokens=256
-      - gconfig.n_samples=1
-      - train_dataset.batch_size=2
-      - valid_dataset.batch_size=2
-      - +train_dataset.setup_timeout=600
-      - +valid_dataset.setup_timeout=600
-      - +total_train_steps=1
-    timeout_minutes: 150
-  # Hermes 在线 RL（唯一覆盖 online 模式 + Agent Service + per-session key）；
-  # _version 回退 v1（默认 v2 强制 SGLang-only 的 awex，vLLM 404）。
-  - path: examples/hermes/train.py
-    profile: areal-hermes-grpo
-    runner: linux-aarch64-a2-1
-    npu_devices: '0'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/hermes/config.yaml
-      - rollout.backend=vllm:d1p1t1
-      - rollout._version=v1
-      - actor.backend=fsdp:d1p1t1
-      - actor._version=v1
-      - ++actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - rollout.admin_api_key=sk-areal-ci
-      - +rollout.agent.admin_api_key=sk-areal-ci
-      - vllm.gpu_memory_utilization=0.25
-      - vllm.max_model_len=16384
-      - gconfig.max_new_tokens=256
-      - +rollout.setup_timeout=1800
-      - total_train_steps=1
-      - total_train_epochs=1
-    timeout_minutes: 150
-  # OpenClaw 在线 RL：任意 OpenAI 兼容 agent runtime（chat-completions 协议，
-  # 无 runtime 依赖）；驱动由 run_example.sh 直接 HTTP。
-  - path: examples/openclaw/train.py
-    profile: areal-openclaw-grpo
-    runner: linux-aarch64-a2-1
-    npu_devices: '0'
-    image: ghcr.nju.edu.cn/hwvanici/areal_npu:v1.0.5-a2
-    overlay_args:
-      - --config examples/openclaw/config.yaml
-      - rollout.backend=vllm:d1p1t1
-      - actor.backend=fsdp:d1p1t1
-      - ++actor.weight_update_mode=disk
-      - actor.path=${AREAL_MODEL_PATH}
-      - train_dataset.batch_size=1
-      - actor.reward_norm=null
-      - actor.adv_norm=null
-      - vllm.gpu_memory_utilization=0.2
-      - vllm.max_model_len=16384
-      - gconfig.max_new_tokens=256
-      - +rollout.setup_timeout=1800
-      - total_train_steps=1
-      - total_train_epochs=1
-    timeout_minutes: 150
   # 在线推理服务（rollout-only）：online_rollout.py 起 gateway 后等外部流量；
   # run_example.sh 用直接 HTTP 造流量（start_session→chat/completions→set_reward
   # →refresh），收满 batch_size 轨迹即打印 Rollout complete 退出。
@@ -471,28 +111,43 @@ supported:
       - ++rollout.setup_timeout=1800
     timeout_minutes: 150
 unsupported:
-  # ===== 上游源码 bug（改上游源码才能通，无 overlay 可绕）=======================
-  # 以下 5 条都是上游脚本 / 引擎自身的 bug，实测必崩：
-  # 1) clevr SFT：loader 读错 processor 键名——clevr_count_70k.py:110 用旧键
-  #    token_type_ids 取 mm_token_type_ids（transformers 5.x 下恒 None），SFT
-  #    collate 阶段 torch.tensor(None) 崩。（geometry3k_grpo 读同一错键无碍：GRPO
-  #    不 collate 数据集样本。）
-  # 2) hhrlhf RW：引擎在打包前就读 x["cu_seqlens"]（rw_engine.py:52→21），而 collate
-  #    只产 input_ids/attention_mask（rw_trainer.py:61-73）→ KeyError。
-  # 3) hhrlhf DPO：loss 聚合硬编码 fp64（functional.py:829-835），Ascend 不支持。
-  # 4) virl39k GRPO：reward_fn 以函数对象传入，过 RPC 时 JSON 序列化失败（正常条目
-  #    传"点路径"字符串）；同 bug 的 vlm_npu/geometry3k_grpo.py 见下方工具段。
-  # 5) agent_service：入口硬编码 LocalScheduler(experiment_name=..., trial_name=...,
-  #    gpu_devices=[])，不传 fileroot 也不传 log_dir，local.py:158 断言必炸；CLI 仅
-  #    --admin-api-key、LocalScheduler 无环境变量兜底，无 overlay 出口（README 的
-  #    Quick Start 与 Programmatic 片段同一写法，同样必炸）。sidecar 方案本身已实测
-  #    可行（本地 online_rollout 网关正常就绪），配方沉淀在 run_example.sh 的
-  #    agent-service 分支与 setup §1g，上游修复后可直接复用。
+  # ===== 上游源码 bug（实测必崩，改上游源码才能通，无 overlay 可绕）=============
+  # loader 读错 processor 键名：clevr_count_70k.py:111 用旧键 token_type_ids 取
+  # mm_token_type_ids（transformers 5.x 下恒 None），SFT collate 阶段
+  # torch.tensor(None) 崩（geometry3k_grpo 读同一错键无碍：GRPO 不 collate 数据集样本）
   - examples/vlm/clevr_count_70k_sft.py
+  # 引擎在打包前就读 x["cu_seqlens"]（rw_engine.py:22/50/126），而 collate 只产
+  # input_ids/attention_mask（rw_trainer.py:55）→ KeyError
   - examples/alignment/hhrlhf_rw.py
+  # loss 聚合硬编码 fp64（functional.py:840-846，dpo_pair_logratios；docstring 载明
+  # 有意设计，需上游参数化），Ascend 不支持
   - examples/alignment/hhrlhf_dpo.py
+  # reward_fn 以函数对象传入，过 RPC 时 JSON 序列化失败（正常条目传"点路径"字符串）
   - examples/vlm_npu/virl39k_grpo.py
+  # 入口硬编码 LocalScheduler(experiment_name=..., trial_name=..., gpu_devices=[])，
+  # 不传 fileroot 也不传 log_dir，local.py:170 断言必炸；CLI 仅 --admin-api-key、
+  # LocalScheduler 无环境变量兜底，无 overlay 出口（README 的 Quick Start 与
+  # Programmatic 片段同一写法，同样必炸）。sidecar 方案本身已实测可行（本地
+  # online_rollout 网关正常就绪），配方沉淀在 run_example.sh 的 agent-service 分支
+  # 与 setup §1g，上游修复后可直接复用
   - examples/agent_service/run_agent_service.py
+  # --- 上游 v2.1.1 回归：stats_tracker 改 fp64，NPU HCCL 不支持 at::kDouble ---
+  # 11a1907e(#1736) 把统计累加/reduce 改成 fp64；训练能跑完、最后 export_stats 崩
+  # （2026-11-05 实测）。无配置/环境变量出口；钉回 v2.1.0 或等上游修复后迁回
+  - examples/vlm/clevr_count_70k_grpo.py
+  - examples/tau2/train.py
+  - examples/vlm/geometry3k_grpo.py
+  - examples/math/gsm8k_rl.py
+  - examples/math/boba_grpo.py
+  - examples/countdown/train.py
+  - examples/multi_turn_math/gsm8k_rl_mt.py
+  - examples/tir/train_tir.py
+  - examples/scaffolding/gsm8k_rlvr_scaffolding.py
+  - examples/openai_agents/train_agents.py
+  - examples/camel/train.py
+  - examples/agent_workflow/train.py
+  - examples/hermes/train.py
+  - examples/openclaw/train.py
 
   # ===== 上游 demo 设计缺陷（时序竞态，双向不稳定；2026-10-16 实测定位）=========
   # hitl demo（zeroclaw 第三方 CLI 兼容性验证，价值仍在）：admin key 被
@@ -528,10 +183,31 @@ unsupported:
   # 沙箱必须 AEnvironment 集群（k8s+Helm），CI 无集群；自建路线实测锁死（容器无 
   # CAP_SYS_ADMIN、mount 被拒，起不了 DinD/k8s）。回归：单容器自托管或注入 AENV_SYSTEM_URL
   - examples/swe/train_swe_rl.py
+  # Arena RL：需 Arena 在线 Stream 服务与 slurm/sbatch 集群提交
+  - examples/swe/qwen38_flash_next/train_rl.py
+  - examples/swe/rl/qwen38_27b/qwen27_train_rl.py
+  # SGLang rollout 硬依赖（脚本断言），NPU 不支持
+  - examples/math/gsm8k_score_centering.py
+  # SGLang rollout + Megatron/AWEX 栈（8 卡），NPU 不支持
+  - examples/cpu_staged_offload/dapo-math_rl_cpu_staged.py
+  # SGLang rollout + Megatron/AWEX 栈（8 卡），NPU 不支持
+  - examples/mopd/gsm8k_qwen3_14b_to_0_6b.py
+  # SGLang rollout + Megatron + R3 router replay（SGLang-only 特性），NPU 不支持
+  - examples/r3/gsm8k_multiturn.py
+  # 同 r3 配置族（sglang rollout + megatron actor + Moonlight，8 卡），NPU 不支持
+  - examples/r3/gsm8k_rlvr.py
+  # 多模态 agent 轨迹仅支持 SGLang rollout（client.py 运行时校验 vLLM 拒绝），NPU 不支持
+  - examples/vlm/geometry3k_agent_grpo.py
+  # Daytona 云沙箱评分示例，需 DAYTONA_API_KEY（见 README），CI 无外部服务
+  - examples/sandbox_daytona/reward_example.py
+
+  # ===== 可看护但重复（与已看护用例同链路，不单列省 CI）=====
+  # 副本现带 virl39k 同款 RPC bug（见上），修好后仍与正主重复
+  - examples/vlm_npu/geometry3k_grpo.py  # examples/vlm/geometry3k_grpo.py
+
 
-  # ===== 工具套件（非独立入口；被上面的入口 import 的实现 / 测试 / 包标识 / helper / CLI）=====
-  # scan 到但都不是可独立运行的示例，只被入口间接调用：
-  # -- 上游自测（scaffolding/tests、tir/tests）与 pytest 基础设施
+  # ===== 不可看护的用例=====
+  # 上游自测与 pytest 基础设施（scaffolding/tests、tir/tests），非可守卫 example。
   - examples/scaffolding/tests/test_controllers.py
   - examples/scaffolding/tests/test_scaffolding_llm_integration.py
   - examples/scaffolding/tests/test_self_contained.py
@@ -544,7 +220,7 @@ unsupported:
   - examples/tir/tests/test_tir_workflow_async.py
   - examples/tir/tests/test_tool_manager_async.py
   - examples/tir/tests/test_tool_registry_lazy.py
-  # -- 包标识文件（__init__.py）
+  # 包标识文件（__init__.py）。
   - examples/hermes/__init__.py
   - examples/scaffolding/__init__.py
   - examples/scaffolding/core/__init__.py
@@ -553,7 +229,13 @@ unsupported:
   - examples/terminal_bench/agent/__init__.py
   - examples/terminal_bench/workflow/__init__.py
   - examples/tir/tools/__init__.py
-  # -- scaffolding 框架实现（run_example.sh 以 python -m 跑 scaffolding/*.py 时 import）
+  - examples/__init__.py
+  - examples/cpu_staged_offload/__init__.py
+  - examples/swe/__init__.py
+  # helper / 实现模块：被同目录 example import 或经 config 引用，非入口（scaffolding / tir /
+  # terminal_bench / search_agent 的框架实现，各 example 的 config / utils / reward，
+  # data_engineering / prm / tau2 的配套模块，swe（含 Arena 工作流与 worker）/ hermes /
+  # openclaw / agent_service 的支撑模块）。
   - examples/scaffolding/_compat.py
   - examples/scaffolding/controllers.py
   - examples/scaffolding/core/controller.py
@@ -570,7 +252,6 @@ unsupported:
   - examples/scaffolding/task.py
   - examples/scaffolding/worker.py
   - examples/scaffolding/workflow.py
-  # -- tir 工具链实现（train_tir.py 调用）
   - examples/tir/prompts.py
   - examples/tir/tir_workflow.py
   - examples/tir/tool_manager.py
@@ -578,64 +259,82 @@ unsupported:
   - examples/tir/tools/calculator_tool.py
   - examples/tir/tools/daytona_python_tool.py
   - examples/tir/tools/python_tool.py
-  # -- terminal_bench 实现（train.py 调用）
   - examples/terminal_bench/agent/camel_terminal_agent.py
   - examples/terminal_bench/agent/chat_agent_trace.py
   - examples/terminal_bench/agent/prompts.py
   - examples/terminal_bench/agent_rl_config.py
   - examples/terminal_bench/workflow/camel_rlvr_workflow.py
   - examples/terminal_bench/workflow/pre_build_tasks_utils.py
-  # -- search_agent/tongyi_deepresearch 实现（entry 调用）
   - examples/search_agent/tongyi_deepresearch/config.py
   - examples/search_agent/tongyi_deepresearch/prompt.py
   - examples/search_agent/tongyi_deepresearch/react_agent.py
   - examples/search_agent/tongyi_deepresearch/tool_search.py
   - examples/search_agent/tongyi_deepresearch/tool_visit.py
-  # -- 各 example 的单文件 config / utils / reward helper
   - examples/agent_workflow/configs.py
   - examples/countdown/reward_score.py
   - examples/multi_turn_math/config.py
   - examples/openai_agents/config.py
   - examples/tau2/utils.py
-  # -- profile rank helper（run_qwen3_30b_a3b_sft_profile.sh 内嵌调用）
-  - examples/profile/profile_rank_utils.py
-  # -- SWE env/train config dataclass（train_swe_rl.py import）
   - examples/swe/utils.py
-  # -- SWE rollout 过滤函数（按组内 reward 方差过滤全对/全错样本）
   - examples/swe/filter_function.py
-  # -- 其它被入口 import 的实现 / 环境 / demo
+  - examples/swe/agent.py
+  - examples/swe/prefix_matchers.py
+  - examples/swe/preprocessors.py
   - examples/agent_service/agent.py
   - examples/docs/debug/cmp_rollout.py
-  - examples/sandbox_daytona/reward_example.py
   - examples/tau2/agent.py
-  # -- openclaw 在线链路的驱动与工具（run_example 已改为直接 HTTP，这些脚本不再调用）
   - examples/openclaw/_fmt.py
   - examples/openclaw/demo_lifecycle.py
   - examples/openclaw/set_reward.py
   - examples/openclaw/start_session.py
-  # -- Hermes 的 AgentRunnable 实现（Agent Service 进程内每 session 加载）
   - examples/hermes/hermes.py
-  # -- SWE 的 agent workflow 实现（train_swe_rl.py 调用）
-  - examples/swe/agent.py
-  # -- SWE InteractionCache 前缀匹配器
-  - examples/swe/prefix_matchers.py
-  # -- SWE 消息预处理器
-  - examples/swe/preprocessors.py
-  # -- hermes 会话工具（纯 HTTP；run_example 已改为直接 HTTP，hermes_loop 不再调用）
   - examples/hermes/hermes_loop.py
   - examples/hermes/set_reward.py
   - examples/hermes/start_session.py
-  # -- SGLang 工具链补丁（环境修补工具，非入口）
-  - examples/swe/apply_sgl_radix_cache_patch.py
-  # -- profile 性能剖析三件套（CUDA 专属工具链，NPU 无对应产物）
+  - examples/data_engineering/tir/data_calibration.py
+  - examples/data_engineering/tir/data_recipe.py
+  - examples/prm/scorers.py
+  - examples/tau2/message_compat.py
+  - examples/swe/arena_agent.py
+  - examples/swe/arena_client.py
+  - examples/swe/arena_config.py
+  - examples/swe/arena_dataset_snapshot.py
+  - examples/swe/arena_reward_breakdown.py
+  - examples/swe/arena_reward_mapping.py
+  - examples/swe/arena_types.py
+  - examples/swe/config.py
+  - examples/swe/reward_transforms.py
+  - examples/swe/qwen38_flash_next/actor_worker.py
+  - examples/swe/qwen38_flash_next/gdn_cp_compat.py
+  - examples/swe/qwen38_flash_next/grad_norm_guard.py
+  - examples/swe/qwen38_flash_next/ple_chunked.py
+  - examples/swe/qwen38_flash_next/proxy.py
+  - examples/swe/qwen38_flash_next/template_defaults.py
+  - examples/swe/rl/qwen38_27b/configure_sglang_logging.py
+  - examples/swe/rl/qwen38_27b/qwen27_actor_worker.py
+  # 工具 / CLI：数据生成、评估器、profile 剖析工具链（CUDA 专属）、SGLang 补丁与 Arena 提交/校验工具，非入口。
+  - examples/countdown/countdown.py
+  - examples/math/gsm8k_eval.py
+  - examples/profile/profile_rank_utils.py
   - examples/profile/postprocess_profile.py
   - examples/profile/run_qwen3_30b_a3b_sft_profile.sh
   - examples/profile/train_sft_profile.py
-  # -- 特殊 CLI（countdown 数据生成 / gsm8k_eval 评估器，非入口）
-  - examples/countdown/countdown.py
-  - examples/math/gsm8k_eval.py
-  # -- vlm_npu 包装层 / 坏副本（geometry3k_grpo.py 同 virl39k 的 RPC bug；.sh 为包装）
-  - examples/vlm_npu/geometry3k_grpo.py
+  - examples/swe/apply_sgl_radix_cache_patch.py
+  - examples/swe/arena_preflight.py
+  - examples/swe/arena_stream_config_projection.py
+  - examples/swe/patch_sglang_qwen3_reasoning.py
+  - examples/swe/rl/qwen38_27b/source_snapshot.py
+  - examples/swe/rl/qwen38_27b/validate_recover_preflight.py
+  # .sh 脚本（vlm_npu 包装、swe slurm/sbatch 提交）。
   - examples/vlm_npu/qwen2_5_vl_3b_geometry3k_grpo.sh
   - examples/vlm_npu/qwen2_5_vl_3b_virl39k_grpo_multinode.sh
   - examples/vlm_npu/qwen3_vl_2b_geometry3k_grpo.sh
+  - examples/swe/qwen38_flash_next/rl_controller.sh
+  - examples/swe/qwen38_flash_next/sbatch_sft_qwen38_flash_next.sh
+  - examples/swe/qwen38_flash_next/sft_worker.sh
+  - examples/swe/qwen38_flash_next/submit_rl.sh
+  - examples/swe/qwen38_flash_next/verify_bridge.sh
+  - examples/swe/rl/qwen38_27b/runtime.sh
+  - examples/swe/rl/qwen38_27b/submit.sh
+  - examples/swe/submit_arena.sh
+  - examples/swe/submit_arena_flash_v3.sh
diff --git a/examples-guards/areal/fixtures/swe_sft_pairs.jsonl b/examples-guards/areal/fixtures/swe_sft_pairs.jsonl
new file mode 100644
index 000000000..37fb85e3e
--- /dev/null
+++ b/examples-guards/areal/fixtures/swe_sft_pairs.jsonl
@@ -0,0 +1,4 @@
+{"messages": [{"role": "user", "content": "utils.py 的 parse_date 对 '2024-02-30' 抛 ValueError，请修复为把非法日期收敛到当月最后一天。"}, {"role": "assistant", "content": "问题在构造时用静态 31 天上限校验日期。把日期收敛到当月实际天数即可：\n\n```python\nimport calendar\n\ndef parse_date(s):\n    year, month, day = map(int, s.split('-'))\n    day = min(day, calendar.monthrange(year, month)[1])\n    return datetime.date(year, month, day)\n```"}]}
+{"messages": [{"role": "user", "content": "tests/test_cache.py 挂了：clear() 之后 LRU 索引里还有残留 key，怎么修？"}, {"role": "assistant", "content": "clear() 只清了数据字典，没有清有序索引，导致查询命中幽灵条目。两个结构都要重置：\n\n```python\ndef clear(self):\n    self._data.clear()\n    self._index.clear()\n```"}]}
+{"messages": [{"role": "user", "content": "api/client.py 的重试逻辑在超时后没有退避，直接打爆了下游。"}, {"role": "assistant", "content": "重试之间需要指数退避。最小改法：\n\n```python\nfor attempt in range(self.max_retries):\n    try:\n        return self.request(url)\n    except TimeoutError:\n        time.sleep(2 ** attempt)\n```"}, {"role": "user", "content": "这样最后一次失败也会 sleep，能不能避免？"}, {"role": "assistant", "content": "把 sleep 放在非最后一次重试之前即可：\n\n```python\nfor attempt in range(self.max_retries):\n    try:\n        return self.request(url)\n    except TimeoutError:\n        if attempt < self.max_retries - 1:\n            time.sleep(2 ** attempt)\n```"}]}
+{"messages": [{"role": "user", "content": "models/user.py 的 to_dict() 递归序列化遇到 datetime 就崩。"}, {"role": "assistant", "content": "给序列化加一个类型分派，datetime 转 ISO 字符串：\n\n```python\ndef to_dict(obj):\n    if isinstance(obj, datetime.datetime):\n        return obj.isoformat()\n    if hasattr(obj, '__dict__'):\n        return {k: to_dict(v) for k, v in obj.__dict__.items()}\n    return obj\n```"}, {"role": "user", "content": "测试里嵌套的 list 也会崩，怎么处理？"}, {"role": "assistant", "content": "补上 list 分支，逐项递归：\n\n```python\ndef to_dict(obj):\n    if isinstance(obj, datetime.datetime):\n        return obj.isoformat()\n    if isinstance(obj, list):\n        return [to_dict(item) for item in obj]\n    if hasattr(obj, '__dict__'):\n        return {k: to_dict(v) for k, v in obj.__dict__.items()}\n    return obj\n```"}]}
diff --git a/examples-guards/areal/scripts/run_example.sh b/examples-guards/areal/scripts/run_example.sh
index 803c73ee4..2c6d3ba87 100644
--- a/examples-guards/areal/scripts/run_example.sh
+++ b/examples-guards/areal/scripts/run_example.sh
@@ -49,9 +49,6 @@ cd "$TARGET_ROOT"
 # (e.g. boba_grpo.py's reward_fn "examples.math.boba_grpo.boba_reward_fn");
 # the example dir covers sibling modules imported by the script itself.
 export PYTHONPATH="$TARGET_ROOT:$(dirname "$LAUNCH_PATH")${PYTHONPATH:+:$PYTHONPATH}"
-# examples/scaffolding/*.py use package-relative imports (from ._compat
-# import ...), so they must run as modules, not as plain scripts.
-#
 # Forked runtime logs (data-worker / data-router / data-gateway under
 # <fileroot>/logs/) are NOT part of this job's stdout, so a crash there is
 # otherwise invisible; dump them when the example fails.
@@ -208,288 +205,6 @@ if [[ "$LAUNCH_PATH" == *.py ]]; then
 fi
 
 case "$EXAMPLE_REL" in
-  examples/scaffolding/*.py)
-    MODULE="${EXAMPLE_REL%.py}"
-    MODULE="${MODULE//\//.}"
-    "$PYTHON" -m "$MODULE" "${EXTRA_ARGS[@]}" || run_rc=$?
-    ;;
-  examples/hermes/train.py)
-    # Online RL: train.py blocks waiting for an externally-driven session
-    # lifecycle, so a plain foreground run would hang until timeout.
-    # Orchestrate one full loop instead: background trainer (starts the
-    # vLLM rollout engine + proxy gateway) -> Agent Service (HermesAgent)
-    # -> start_session over HTTP (mint session key; see step 3 for why the
-    # CLI is bypassed) -> one piped conversation via hermes_loop (EOF exits
-    # cleanly) -> set_reward -> start_session refresh (ends the session,
-    # exports the trajectory) -> the trainer's _OnlineAgent unblocks, trains
-    # its single step and exits.
-    HERMES_ADMIN_KEY="areal-agent-ci"
-    TRAIN_LOG="$PWD/hermes_train.log"
-    AGENT_LOG="$PWD/hermes_agent.log"
-
-    hermes_cleanup() {
-      areal agent stop >/dev/null 2>&1 || true
-      if [[ -n "${TRAINER_PID:-}" ]] && kill -0 "$TRAINER_PID" 2>/dev/null; then
-        kill "$TRAINER_PID" 2>/dev/null || true
-      fi
-    }
-    trap hermes_cleanup EXIT
-
-    # 1) Trainer in the background; its log is the source of truth for
-    #    the proxy/inference gateway address (rl_trainer logs
-    #    "Proxy gateway available at http://host:port" once online mode
-    #    is up - engine load takes minutes).
-    "$PYTHON" "$LAUNCH_PATH" "${EXTRA_ARGS[@]}" >"$TRAIN_LOG" 2>&1 &
-    TRAINER_PID=$!
-
-    GW=$(wait_gateway "$TRAIN_LOG" "$TRAINER_PID" trainer) || exit 1
-    echo "inference gateway: $GW"
-
-    # 2) Agent Service with the Hermes agent (run blocks until the stack
-    #    is healthy, then exits leaving the service up).
-    #    hermes-agent enables 30+ toolsets by default; their definitions ride
-    #    EVERY /v1/responses request and overflow vllm's context, making the
-    #    agent death-loop its own compressor. HermesAgent reads
-    #    HERMES_DISABLED_TOOLSETS in __init__ (hermes.py:118) - i.e. in the
-    #    Agent Service worker process - so it must be EXPORTED here, BEFORE
-    #    `areal agent run`, for the workers to inherit it. Keep only "clarify".
-    export HERMES_DISABLED_TOOLSETS="web,search,x_search,vision,video,image_gen,video_gen,computer_use,terminal,skills,browser,cronjob,file,tts,todo,memory,context_engine,session_search,project,code_execution,delegation,homeassistant,kanban,discord,discord_admin,yuanbao,feishu_doc,feishu_drive,spotify,debugging,safe,coding"
-    if ! areal agent run --agent examples.hermes.hermes.HermesAgent \
-        --admin-api-key "$HERMES_ADMIN_KEY" --force >"$AGENT_LOG" 2>&1; then
-      echo "areal agent run failed; tail of $AGENT_LOG:"
-      tail -n 120 "$AGENT_LOG"
-      exit 1
-    fi
-    AGENT_GW=$(grep -oE 'gateway=[A-Za-z0-9._:/-]+' "$AGENT_LOG" | tail -1 | cut -d= -f2 || true)
-    if [[ -z "$AGENT_GW" ]]; then
-      AGENT_GW=$(areal agent status 2>/dev/null | grep -oE 'https?://[A-Za-z0-9._:-]+' | tail -1 || true)
-    fi
-    if [[ -z "$AGENT_GW" ]]; then
-      echo "agent gateway not found; tail of $AGENT_LOG:"
-      tail -n 120 "$AGENT_LOG"
-      exit 1
-    fi
-    echo "agent gateway: $AGENT_GW"
-
-    # 3) Mint the session key on the inference gateway. Talk HTTP directly
-    #    instead of examples/hermes/start_session.py: that CLI hardcodes the
-    #    V2 gateway contract (HTTP 201 + {"sessions":[{"session_api_key":..}]})
-    #    while the v1 proxy gateway returns 200 + flat {"api_key","session_id"}
-    #    (proxy_gateway.py NEW path returns the worker response as-is) - the
-    #    CLI dies on a perfectly successful session, the retry loop then
-    #    spawns zombie sessions until all 3 workers' capacity is exhausted
-    #    and everything 429s. 429 ("no capacity") is retried: workers only
-    #    register in the readiness queue after the trainer is fully up.
-    SESS_KEY=""
-    for attempt in $(seq 1 24); do
-      BODY=$(http_post "$GW" rl/start_session sk-areal-ci '{"task_id":"ci-hermes-smoke"}' 180)
-      SESS_KEY=$(printf '%s' "$BODY" | json_pick '["api_key"]')
-      [[ -n "$SESS_KEY" ]] && break
-      echo "[session attempt $attempt] no session yet (capacity/registering)"
-      sleep 5
-    done
-    if [[ -z "$SESS_KEY" ]]; then
-      echo "start_session never returned a session key"
-      tail -n 40 "$TRAIN_LOG"
-      exit 1
-    fi
-    echo "session: $SESS_KEY"
-
-    # 4) Drive ONE agent turn directly over HTTP, bypassing hermes_loop.py:
-    #    that CLI hardcodes its httpx client timeout to 120s (:47), which the
-    #    ~11.5K-token Hermes round-trip exceeds on this NPU - it dies with
-    #    "Read timed out" while the server is still working. Sending the same
-    #    /v1/responses POST ourselves with a 600s client timeout avoids the
-    #    hardcoded limit (no source patch). inf-* route the agent's LLM calls
-    #    through the inference gateway under the session key (self-evolution
-    #    capture); model id "default" is InferenceDataProxy's served name.
-    agent_turn() {
-      "$PYTHON" - "$AGENT_GW" "$HERMES_ADMIN_KEY" "$1" "$GW" "$SESS_KEY" <<'PY'
-import json
-import sys
-import urllib.error
-import urllib.request
-
-base, key, content, inf_base, sess_key = sys.argv[1:6]
-payload = {
-    "input": [{"type": "message", "content": content}],
-    "model": "hermes-agent",
-    "user": "hermes-ci",
-    "inf_base_url": inf_base,
-    "inf_model": "default",
-    "session_api_key": sess_key,
-}
-req = urllib.request.Request(
-    f"{base.rstrip('/')}/v1/responses",
-    data=json.dumps(payload).encode(),
-    headers={"Content-Type": "application/json", "Authorization": f"Bearer {key}"},
-    method="POST",
-)
-try:
-    with urllib.request.urlopen(req, timeout=600) as resp:
-        data = json.loads(resp.read().decode())
-except urllib.error.HTTPError as e:
-    print(f"HTTP {e.code}: {e.read().decode()[:300]}", file=sys.stderr)
-    sys.exit(1)
-except OSError as e:
-    print(f"{type(e).__name__}: {e}", file=sys.stderr)
-    sys.exit(1)
-texts = []
-for item in data.get("output", []):
-    if item.get("type") == "message":
-        for block in item.get("content", []):
-            if block.get("type") == "output_text":
-                texts.append(block["text"])
-print("".join(texts) if texts else json.dumps(data)[:300])
-PY
-    }
-
-    if ! AGENT_REPLY=$(agent_turn "Say hello and introduce yourself in one sentence."); then
-      echo "agent turn failed; tails:"
-      tail -n 60 "$AGENT_LOG"
-      tail -n 40 "$TRAIN_LOG"
-      exit 1
-    fi
-    echo "agent reply: $AGENT_REPLY"
-
-    # 5) Reward the session, then refresh it: the refresh call (start_session
-    #    with the existing api_key) ends the old session, waits for the
-    #    training pipeline to cycle and starts a new one with the same key -
-    #    the trajectory export unblocks the trainer's _OnlineAgent. Both go
-    #    through http_post (v1 contract: 200 on success). Refresh can 429 while
-    #    the pipeline cycles (refresh_timeout on the gateway side), so retry.
-    REWARD_RESP=$(http_post "$GW" rl/set_reward "$SESS_KEY" '{"reward":1.0}' 180)
-    if [[ -z "$REWARD_RESP" ]]; then
-      echo "set_reward failed; tail of $TRAIN_LOG:"
-      tail -n 40 "$TRAIN_LOG"
-      exit 1
-    fi
-    echo "reward applied: $REWARD_RESP"
-
-    REFRESHED=""
-    for attempt in $(seq 1 12); do
-      BODY=$(http_post "$GW" rl/start_session sk-areal-ci \
-        "{\"task_id\":\"ci-hermes-smoke\",\"api_key\":\"$SESS_KEY\"}" 180)
-      NEW_KEY=$(printf '%s' "$BODY" | json_pick '["api_key"]')
-      if [[ -n "$NEW_KEY" ]]; then
-        REFRESHED=1
-        echo "session refreshed (key: $NEW_KEY)"
-        break
-      fi
-      echo "[refresh attempt $attempt] pipeline still cycling"
-      sleep 10
-    done
-    if [[ -z "$REFRESHED" ]]; then
-      echo "session refresh never succeeded; the trainer may be stuck waiting"
-      tail -n 40 "$TRAIN_LOG"
-      exit 1
-    fi
-
-    # 6) Trainer trains its single step and exits. Bounded wait: a rejected
-    #    trajectory (empty/error-only session) leaves the trainer waiting
-    #    for the NEXT session forever - instead of holding the runner until
-    #    the job timeout, fail after 15 min with the trainer log dumped.
-    wait_bounded "$TRAINER_PID" 900 "$TRAIN_LOG"
-    if (( WAIT_TIMED_OUT )); then
-      run_rc=1
-      echo "trainer still running after 900s - likely waiting for another session (rejected trajectory?); tail of $TRAIN_LOG:"
-      tail -n 120 "$TRAIN_LOG"
-    elif wait "$TRAINER_PID"; then
-      echo "hermes online-rl loop completed"
-    else
-      run_rc=$?
-      echo "trainer exited rc=$run_rc; tail of $TRAIN_LOG:"
-      tail -n 120 "$TRAIN_LOG"
-    fi
-    ;;
-  examples/openclaw/train.py)
-    # Online RL via the proxy gateway (same architecture as hermes): the agent
-    # runtime is any OpenAI-compatible client. We drive one episode DIRECTLY
-    # over HTTP instead of the upstream examples/openclaw/demo_lifecycle.py,
-    # whose /chat/completions client timeout is hardcoded to 30s (:264) - the
-    # NPU generation exceeds that, so it dies with "Read timed out" before
-    # set_reward. Doing the same calls ourselves with a long client timeout
-    # avoids the hardcoded limit (no source patch).
-    OC_ADMIN_KEY="sk-test123456"
-    TRAIN_LOG="$PWD/openclaw_train.log"
-
-    openclaw_cleanup() {
-      if [[ -n "${TRAINER_PID:-}" ]] && kill -0 "$TRAINER_PID" 2>/dev/null; then
-        kill "$TRAINER_PID" 2>/dev/null || true
-      fi
-    }
-    trap openclaw_cleanup EXIT
-
-    # 1) Trainer in the background; its log carries the gateway address.
-    "$PYTHON" "$LAUNCH_PATH" "${EXTRA_ARGS[@]}" >"$TRAIN_LOG" 2>&1 &
-    TRAINER_PID=$!
-
-    GW=$(wait_gateway "$TRAIN_LOG" "$TRAINER_PID" trainer) || exit 1
-    echo "inference gateway: $GW"
-
-    # 2) start_session (admin) -> per-session key. 429 ("no capacity") is
-    #    retried: workers only register once the trainer is fully up.
-    SESS_KEY=""
-    for attempt in $(seq 1 24); do
-      BODY=$(http_post "$GW" rl/start_session "$OC_ADMIN_KEY" '{"task_id":"ci-openclaw-smoke"}' 60)
-      SESS_KEY=$(printf '%s' "$BODY" | json_pick '["api_key"]')
-      [[ -n "$SESS_KEY" ]] && break
-      echo "[session attempt $attempt] no session yet"
-      sleep 5
-    done
-    [[ -n "$SESS_KEY" ]] || { echo "start_session never returned a key"; tail -n 40 "$TRAIN_LOG"; exit 1; }
-    echo "session: $SESS_KEY"
-
-    # 3) chat/completions under the session key - REAL generation (tokens +
-    #    logprobs recorded for RL). Long client timeout: the NPU generates
-    #    slowly (~1.5 tok/s), far past demo_lifecycle's hardcoded 30s.
-    CHAT_BODY='{"model":"default","messages":[{"role":"system","content":"You are a helpful math tutor. Show your work step by step."},{"role":"user","content":"Solve step by step: What is 12 * 15 + 3?"}],"temperature":0.7,"max_tokens":256}'
-    CHAT=$(http_post "$GW" chat/completions "$SESS_KEY" "$CHAT_BODY" 600)
-    [[ -n "$CHAT" ]] || { echo "chat/completions failed"; tail -n 40 "$TRAIN_LOG"; exit 1; }
-    echo "chat completion: $(printf '%s' "$CHAT" | cut -c1-200)"
-
-    # 4) set_reward (session key), then refresh: start_session with the key
-    #    ends the session and exports the trajectory -> the trainer's
-    #    _OnlineAgent unblocks and trains its single step.
-    REWARD=$(http_post "$GW" rl/set_reward "$SESS_KEY" '{"reward":1.0}' 60)
-    [[ -n "$REWARD" ]] || { echo "set_reward failed"; tail -n 40 "$TRAIN_LOG"; exit 1; }
-    echo "reward applied: $REWARD"
-
-    REFRESHED=""
-    for attempt in $(seq 1 12); do
-      BODY=$(http_post "$GW" rl/start_session "$OC_ADMIN_KEY" \
-        "{\"task_id\":\"ci-openclaw-smoke\",\"api_key\":\"$SESS_KEY\"}" 130)
-      NEW_KEY=$(printf '%s' "$BODY" | json_pick '["api_key"]')
-      [[ -n "$NEW_KEY" ]] && { REFRESHED=1; echo "session refreshed (key: $NEW_KEY)"; break; }
-      echo "[refresh attempt $attempt] pipeline still cycling"
-      sleep 10
-    done
-    [[ -n "$REFRESHED" ]] || { echo "refresh never succeeded; trainer may be stuck"; tail -n 40 "$TRAIN_LOG"; exit 1; }
-
-    # 5) Trainer trains its single step and exits (bounded).
-    wait_bounded "$TRAINER_PID" 1200 "$TRAIN_LOG"
-    if (( WAIT_TIMED_OUT )); then
-      run_rc=1
-      echo "trainer wait timed out; tail of $TRAIN_LOG:"
-      tail -n 120 "$TRAIN_LOG"
-    elif wait "$TRAINER_PID"; then
-      echo "openclaw online-rl loop completed"
-    else
-      run_rc=$?
-      echo "trainer exited rc=$run_rc; tail of $TRAIN_LOG:"
-      tail -n 120 "$TRAIN_LOG"
-    fi
-    ;;
-  examples/camel/train.py)
-    # train.py defines its config dataclass (AgentRLConfig) at module scope.
-    # Whether run as a plain script or via `python -m`, the class's __module__
-    # is "__main__" (verified), so the engine worker cannot import it and
-    # deserialize_value degrades the config to a plain dict -> config.seed
-    # AttributeError. Importing the file as a module instead gives the class
-    # the path "examples.camel.train.AgentRLConfig" (importable in the worker,
-    # repo root is on PYTHONPATH) and we call main() explicitly.
-    "$PYTHON" -c "import sys; from examples.camel.train import main; main(sys.argv[1:])" "${EXTRA_ARGS[@]}" || run_rc=$?
-    ;;
   examples/experimental/inference_service/online_rollout.py)
     # Rollout-only online example: online_rollout.py starts the gateway then
     # BLOCKS in rollout_batch(data=None) waiting for EXTERNAL traffic. We
@@ -548,195 +263,6 @@ PY
       tail -n 120 "$ROLLOUT_LOG"
     fi
     ;;
-  examples/experimental/inference_service/human_in_the_loop_demo.py)
-    # The demo brings its own driver (launches online_rollout.py itself,
-    # temporarily points ~/.zeroclaw/config.toml at the local gateway, runs
-    # 4 HITL rounds via the zeroclaw CLI, then checks the rollout log). Its
-    # exit code is LOOSE though: zeroclaw failures are caught (empty answer,
-    # reward 0) and a missing "Rollout complete" only prints a warning
-    # before "Demo finished" + exit 0. Make the pass criterion strict:
-    # exit 0 AND "Rollout complete" in the demo output (on success it echoes
-    # the rollout log tail incl. "Databatch detected").
-    HITL_LOG="$PWD/hitl_demo.log"
-    # Diagnose the zeroclaw build BEFORE the demo swallows its stderr
-    # (check_output discards the captured output on failure): run the demo's
-    # exact invocation shape against the placeholder config
-    # (custom:http://localhost:1). rc=2 + usage text = the installed CLI
-    # lacks the demo's interface - zeroclaw.com's installer ships the
-    # OFFICIAL build, while the demo targets the dhh1995 fork. A connection
-    # error instead means the syntax is fine.
-    # Dump the official CLI surface first: version (for later pinning) and
-    # whether `agent -m` / --session-state-file exist at all.
-    { zeroclaw --version; echo "--------"; zeroclaw --help; echo "--------"; zeroclaw agent --help; } \
-      >"$PWD/zc_help.log" 2>&1 || true
-    echo "--- zeroclaw version / --help / agent --help (head -n 60) ---"
-    head -n 60 "$PWD/zc_help.log" 2>/dev/null || true
-    timeout 30 zeroclaw agent -m "probe" \
-      --session-state-file "$PWD/zc_probe.json" >"$PWD/zc_probe.log" 2>&1 \
-      || echo "zeroclaw probe rc=$? (usage error = wrong build; connection error = syntax OK)"
-    cat "$PWD/zc_probe.log" 2>/dev/null || true
-    demo_rc=0
-    "$PYTHON" "$LAUNCH_PATH" "${EXTRA_ARGS[@]}" >"$HITL_LOG" 2>&1 || demo_rc=$?
-    if (( demo_rc == 0 )) && grep -q 'Rollout complete' "$HITL_LOG"; then
-      echo "hitl demo completed"
-      grep 'Rollout complete' "$HITL_LOG" | tail -1
-    else
-      run_rc=1
-      echo "hitl demo failed (rc=$demo_rc, or no 'Rollout complete' - hollow run?); tail of $HITL_LOG:"
-      tail -n 80 "$HITL_LOG"
-    fi
-    ;;
-  examples/tau2/train.py)
-    # Tau2 keeps its REAL user simulator (solo_mode=false): host the user
-    # LLM locally on the runner's 2nd card with the SAME weights (zero extra
-    # download) and point econfig.user_llm_base_url at it (the overlay sets
-    # the URL). The trainer is pinned to card 0 via npu_devices, so the two
-    # never collide. Multiple --served-model-name aliases absorb whatever
-    # model string the tau2 client sends (econfig.user_llm uses litellm
-    # "openai/..." naming) - vLLM matches the model field strictly.
-    tau2_cleanup() {
-      if [[ -n "${USER_SIM_PID:-}" ]] && kill -0 "$USER_SIM_PID" 2>/dev/null; then
-        kill "$USER_SIM_PID" 2>/dev/null || true
-      fi
-    }
-    trap tau2_cleanup EXIT
-
-    # 1) User-simulator vLLM on physical device 1 via the shared helper
-    #    (sets USER_SIM_PID/USER_SIM_LOG; the EXIT trap above kills it).
-    start_user_sim "$PWD/tau2_user_sim.log" || exit 1
-
-    # 2) Trainer in the foreground; its exit code is authoritative (the
-    #    user-sim server is torn down by the EXIT trap).
-    tau2_rc=0
-    "$PYTHON" "$LAUNCH_PATH" "${EXTRA_ARGS[@]}" || tau2_rc=$?
-    if (( tau2_rc == 0 )); then
-      echo "tau2 training completed"
-    else
-      run_rc=$tau2_rc
-      echo "tau2 trainer exited rc=$tau2_rc; user-sim tail:"
-      tail -n 40 "$USER_SIM_LOG"
-    fi
-    ;;
-  examples/agent_service/run_agent_service.py)
-    # The Agent Service hosts NO model itself: ClaudeAgent shells out via
-    # claude-agent-sdk, whose ClaudeSDKClient spawns the Claude Code CLI -
-    # the API traffic goes wherever ANTHROPIC_BASE_URL points. Point it at a
-    # SIDECAR online_rollout stack: its v2 gateway serves the Anthropic
-    # /v1/messages route and the data proxy translates Anthropic->OpenAI onto
-    # the local vLLM, so no real Anthropic key is needed. The service's own
-    # stdin loop is interactive: hold it open with a FIFO writer, drive ONE
-    # /v1/responses request (the shape documented in the example README),
-    # then close the FIFO -> EOF -> clean exit 0 = pass. The sidecar's
-    # rollout_batch never completes (no RL sessions are driven); it is just
-    # the LLM backend and is killed at the end.
-    AS_ADMIN_KEY="areal-agent-admin"
-    AS_SIDECAR_KEY="sk-test123456"
-    AS_SVC_LOG="$PWD/agent_service.log"
-    AS_ROLLOUT_LOG="$PWD/agent_service_rollout.log"
-    AS_FIFO="$PWD/agent_service_stdin.fifo"
-
-    agent_service_cleanup() {
-      exec 3>&- 2>/dev/null || true
-      if [[ -n "${AS_SVC_PID:-}" ]] && kill -0 "$AS_SVC_PID" 2>/dev/null; then
-        kill "$AS_SVC_PID" 2>/dev/null || true
-      fi
-      if [[ -n "${AS_ROLLOUT_PID:-}" ]] && kill -0 "$AS_ROLLOUT_PID" 2>/dev/null; then
-        kill "$AS_ROLLOUT_PID" 2>/dev/null || true
-      fi
-    }
-    trap agent_service_cleanup EXIT
-
-    # 1) Sidecar: online_rollout.py (vLLM + gateway serving /v1/messages,
-    #    registered model name "default").
-    "$PYTHON" examples/experimental/inference_service/online_rollout.py \
-      --config examples/experimental/inference_service/online_rollout.yaml \
-      rollout.backend=vllm:d1p1t1 \
-      actor.backend=fsdp:d1p1t1 \
-      actor.path="$AREAL_MODEL_PATH" \
-      train_dataset.batch_size=1 \
-      vllm.gpu_memory_utilization=0.25 \
-      vllm.max_model_len=16384 \
-      gconfig.max_new_tokens=256 \
-      rollout.setup_timeout=1800 \
-      >"$AS_ROLLOUT_LOG" 2>&1 &
-    AS_ROLLOUT_PID=$!
-
-    AS_INF_GW=$(wait_gateway "$AS_ROLLOUT_LOG" "$AS_ROLLOUT_PID" sidecar) || exit 1
-    echo "sidecar inference gateway: $AS_INF_GW"
-
-    # 2) The Agent Service, stdin held open by a FIFO writer on fd 3 (a plain
-    #    background run would hit EOF immediately and exit). Claude Code CLI
-    #    env: traffic -> the sidecar gateway; nonessential phones-home
-    #    disabled (those hosts are unreachable from the CN runners and would
-    #    stall the CLI); generous API timeout for the slow NPU generation.
-    mkfifo "$AS_FIFO"
-    ANTHROPIC_BASE_URL="$AS_INF_GW" \
-    ANTHROPIC_API_KEY="$AS_SIDECAR_KEY" \
-    CLAUDE_MODEL=default \
-    CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
-    DISABLE_TELEMETRY=1 \
-    API_TIMEOUT_MS=600000 \
-      "$PYTHON" "$LAUNCH_PATH" "${EXTRA_ARGS[@]}" <"$AS_FIFO" >"$AS_SVC_LOG" 2>&1 &
-    AS_SVC_PID=$!
-    exec 3>"$AS_FIFO"
-
-    # 3) Wait for the service gateway ("Gateway: http://..."), then /health.
-    AS_AGENT_GW=""
-    for i in $(seq 1 60); do
-      AS_AGENT_GW=$(grep 'Gateway:' "$AS_SVC_LOG" 2>/dev/null | tail -1 \
-        | grep -oE 'https?://[A-Za-z0-9._:-]+' | tail -1 || true)
-      [[ -n "$AS_AGENT_GW" ]] && break
-      kill -0 "$AS_SVC_PID" 2>/dev/null || break
-      sleep 5
-    done
-    if [[ -n "$AS_AGENT_GW" ]]; then
-      for i in $(seq 1 24); do
-        http_healthy "${AS_AGENT_GW}/health" && break
-        kill -0 "$AS_SVC_PID" 2>/dev/null || break
-        sleep 5
-      done
-    fi
-    if [[ -z "$AS_AGENT_GW" ]] || ! http_healthy "${AS_AGENT_GW}/health"; then
-      echo "agent service gateway not healthy; tail of $AS_SVC_LOG:"
-      tail -n 120 "$AS_SVC_LOG"
-      exit 1
-    fi
-    echo "agent service gateway: $AS_AGENT_GW"
-
-    # 4) ONE conversation through the service (the README's documented
-    #    request shape). The worker spawns the Claude Code CLI, whose API
-    #    calls land on the sidecar's /v1/messages -> local vLLM.
-    AS_RESP=$(http_post "$AS_AGENT_GW" v1/responses "$AS_ADMIN_KEY" \
-      '{"input":[{"type":"message","content":"Say hello in one sentence."}],"model":"claude-agent","user":"ci-agent-service"}' 600)
-    if [[ -z "$AS_RESP" ]] || ! printf '%s' "$AS_RESP" | grep -q 'completed'; then
-      run_rc=1
-      echo "agent turn failed; response head + service tail:"
-      printf '%s\n' "$AS_RESP" | cut -c1-400
-      tail -n 80 "$AS_SVC_LOG"
-    else
-      echo "agent reply: $(printf '%s' "$AS_RESP" | grep -oE '"text": *"[^"]{0,120}' | head -1)"
-    fi
-
-    # 5) Close the FIFO writer -> the interactive loop sees EOF -> cleanup +
-    #    exit 0. The service's exit code is the pass criterion.
-    exec 3>&-
-    wait_bounded "$AS_SVC_PID" 120 "$AS_SVC_LOG"
-    if (( WAIT_TIMED_OUT == 0 )); then
-      as_rc=0
-      wait "$AS_SVC_PID" || as_rc=$?
-      if (( as_rc == 0 )) && (( run_rc == 0 )); then
-        echo "agent_service completed"
-      else
-        run_rc=${as_rc:-1}
-        echo "agent service exited rc=$as_rc; tail of $AS_SVC_LOG:"
-        tail -n 120 "$AS_SVC_LOG"
-      fi
-    else
-      run_rc=1
-      echo "agent service did not exit after EOF; tail of $AS_SVC_LOG:"
-      tail -n 120 "$AS_SVC_LOG"
-    fi
-    ;;
   examples/experimental/inference_service/tau2_rollout.py)
     # Rollout-only tau2: the SAME real agent + user simulator as the
     # training entry (Tau2AgentWorkflow, airline domain), but driven through
diff --git a/examples-guards/areal/scripts/setup_example.sh b/examples-guards/areal/scripts/setup_example.sh
index d380a9a89..273df94b7 100644
--- a/examples-guards/areal/scripts/setup_example.sh
+++ b/examples-guards/areal/scripts/setup_example.sh
@@ -16,67 +16,20 @@ fi
 
 PROFILE="$1"
 
-SUPPORTED_PROFILES="areal-vlm-grpo areal-vlm-mt-grpo areal-vlm-sft areal-clevr-sft areal-tir-grpo areal-scaffold-grpo areal-agents-grpo areal-math-grpo areal-math-sft areal-math-aime areal-math-boba areal-countdown-grpo areal-hermes-grpo areal-openclaw-grpo areal-camel-grpo areal-online-rollout areal-hitl-demo areal-tau2-grpo areal-agent-service areal-align"
-AIME_PREP=0
-BOBA_PREP=0
-COUNTDOWN_PREP=0
-HHRLHF_PREP=0
-TORL_PREP=0
-SCAFFOLD_PREP=0
-CLEVR_SUBSET_PREP=0
-HERMES_PREP=0
-OPENCLAW_PREP=0
-CAMEL_PREP=0
-HITL_PREP=0
+SUPPORTED_PROFILES="areal-math-sft areal-swe-sft areal-online-rollout areal-tau2-grpo"
 TAU2_PREP=0
-AGENT_SERVICE_PREP=0
 case "$PROFILE" in
-  areal-vlm-grpo) MODEL_ID="Qwen/Qwen2.5-VL-3B-Instruct" ;;
-  areal-vlm-mt-grpo) MODEL_ID="Qwen/Qwen3-VL-2B-Instruct" ;;
-  areal-vlm-sft) MODEL_ID="Qwen/Qwen3-VL-2B-Instruct" ;;
-  # clevr_count_70k_sft.py: see section 10 (local subset; the full 70k train
-  # split blows the data worker when the SFT loader materializes pixel_values).
-  areal-clevr-sft) MODEL_ID="Qwen/Qwen2.5-VL-3B-Instruct"; CLEVR_SUBSET_PREP=1 ;;
-  # tir/train_tir.py: see section 8 (fixture staging; the loader's own
-  # GitHub download is unreachable from the CI job containers).
-  areal-tir-grpo) MODEL_ID="Qwen/Qwen2.5-Math-1.5B"; TORL_PREP=1 ;;
-  # gsm8k_rlvr_scaffolding.py: 1.5B model + section 9 (model alias symlink).
-  areal-scaffold-grpo) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct"; SCAFFOLD_PREP=1 ;;
-  areal-agents-grpo) MODEL_ID="Qwen/Qwen2-1.5B-Instruct" ;;
-  # gsm8k_rl.py and gsm8k_eval.py share the same model.
-  areal-math-grpo) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct" ;;
   areal-math-sft) MODEL_ID="Qwen/Qwen3-1.7B" ;;
-  areal-math-aime) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct"; AIME_PREP=1 ;;
-  areal-math-boba) MODEL_ID="deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"; BOBA_PREP=1 ;;
-  areal-countdown-grpo) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct"; COUNTDOWN_PREP=1 ;;
-  # hermes/train.py: online RL; agent runtime package installed in
-  # section 1c, orchestration in run_example.sh's hermes branch.
-  areal-hermes-grpo) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct"; HERMES_PREP=1 ;;
-  # openclaw/train.py: online RL through the OpenAI-compatible proxy gateway;
-  # no agent runtime package to install (demo_lifecycle.py is pure HTTP).
-  # No source patch (no-patch policy) - demo_lifecycle.py runs as shipped.
-  # Reuses hermes' model.
-  areal-openclaw-grpo) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct"; OPENCLAW_PREP=1 ;;
-  # camel/train.py: camel-ai ChatAgent GRPO. Needs the undeclared camel-ai
-  # package (section 1d) and a `python -m` launch (run_example.sh camel branch).
-  areal-camel-grpo) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct"; CAMEL_PREP=1 ;;
+  # swe/train_sft.py: dataset path must contain an "swe" token to route
+  # into the SWE pipeline (fixture swe_sft_pairs.jsonl via FIXTURE_DIR).
+  areal-swe-sft) MODEL_ID="Qwen/Qwen3-1.7B" ;;
   # online_rollout.py: rollout-only online demo; no extra runtime to install
   # (drive is run_example.sh's online_rollout branch).
   areal-online-rollout) MODEL_ID="Qwen/Qwen3-1.7B" ;;
-  # human_in_the_loop_demo.py: zeroclaw-driven HITL demo; section 1e installs
-  # the CLI and plants its home config skeleton.
-  areal-hitl-demo) MODEL_ID="Qwen/Qwen3-0.6B"; HITL_PREP=1 ;;
-  # tau2/train.py: multi-turn customer-service RL (real user simulator; the
-  # user LLM is hosted locally by run_example.sh's tau2 branch). Section 1f
-  # installs the tau2-bench fork + data.
+  # tau2_rollout.py: rollout-only tau2 (agent + user simulator); section 1f
+  # installs the tau2-bench fork + data, run_example.sh's tau2_rollout branch
+  # hosts the user-sim vLLM on card 1.
   areal-tau2-grpo) MODEL_ID="Qwen/Qwen3-1.7B"; TAU2_PREP=1 ;;
-  # agent_service/run_agent_service.py: Claude Agent SDK micro-service; the
-  # Claude Code CLI it spawns is installed in section 1g; run_example.sh's
-  # agent-service branch points its API traffic at a local sidecar gateway.
-  areal-agent-service) MODEL_ID="Qwen/Qwen3-0.6B"; AGENT_SERVICE_PREP=1 ;;
-  # hhrlhf_dpo.py: DPO pipeline is model-agnostic; upstream yaml's 7B swapped
-  # for 1.5B (7B fp32 init OOMs the 32GB cards in the pool, 15GB per run).
-  areal-align) MODEL_ID="Qwen/Qwen2.5-1.5B-Instruct"; HHRLHF_PREP=1 ;;
   *)
     echo "unknown profile: ${PROFILE} (supported: ${SUPPORTED_PROFILES})" >&2
     exit 1
@@ -89,8 +42,8 @@ if [[ -z "${TARGET_ROOT:-}" || -z "${GITHUB_WORKSPACE:-}" || -z "${GITHUB_ENV:-}
 fi
 
 # -------------------------------------------------------
-# Shared helpers: mirror pip installs, shallow clones, tarball
-# fetches - one place for the mirror URLs and CI timeout budget.
+# Shared helpers: mirror pip installs and shallow clones - one place
+# for the mirror URLs.
 # -------------------------------------------------------
 pip_tuna() {
   uv pip install --system --index-url https://pypi.tuna.tsinghua.edu.cn/simple "$@"
@@ -101,17 +54,18 @@ pip_aliyun() {
 }
 
 shallow_clone() {
-  git clone --depth 1 --branch "$2" "$1" "$3"
-}
-
-# Tarball fetch with a long budget and one resume retry: the zeroclaw
-# release comes from github.com, whose throughput from the CN runners is
-# marginal - 15.5MB once needed >600s and got cut at 91%. -C - resumes
-# the partial file instead of restarting from byte 0.
-fetch_tarball() {
-  curl -fsSL --connect-timeout 30 --max-time 1800 -C - "$1" -o "$2" \
-    || curl -fsSL --connect-timeout 30 --max-time 1800 -C - "$1" -o "$2" \
-    || { echo "failed to download $1" >&2; exit 2; }
+  local url=$1 branch=$2 dest=$3 attempt
+  # CI 出口经 gh-proxy/squid，长传输偶发 curl 18/early EOF 截断；
+  # 失败时目录可能残留，重试前清理，最多 3 次。
+  for attempt in 1 2 3; do
+    rm -rf "$dest"
+    if git clone --depth 1 --branch "$branch" "$url" "$dest"; then
+      return 0
+    fi
+    echo "shallow_clone failed (attempt ${attempt}/3): ${url}@${branch}" >&2
+    sleep 15
+  done
+  return 1
 }
 
 # -------------------------------------------------------
@@ -161,83 +115,6 @@ VLLM_TARGET_DEVICE=empty uv pip install --system --no-deps --no-build-isolation
   -e "$VLLM_SRC"
 python -c "import torch, torch_npu, vllm, vllm_ascend; from vllm.entrypoints.openai.utils import validate_json_request; print(f'vllm {vllm.__version__} + vllm-ascend 0.22.1rc1 layout OK; torch {torch.__version__}, torch_npu {torch_npu.__version__}')"
 
-# -------------------------------------------------------
-# 1c. hermes profile: the agent runtime package. The model is the
-# standard Qwen2.5-1.5B-Instruct (downloaded by section 3); hermes-agent
-# itself is model-agnostic - it drives whatever OpenAI-compatible endpoint
-# the rollout engine serves. The Agent Service imports
-# examples.hermes.hermes.HermesAgent from the target checkout.
-# Index: tsinghua (not aliyun) - aliyun's PEP 658 .metadata sidecar for
-# hermes-agent's python-dotenv dep 404s intermittently; tuna mirrors the
-# metadata reliably (same mirror as transformers-quick-start.yml).
-# -------------------------------------------------------
-if [[ "$HERMES_PREP" == 1 ]]; then
-  pip_tuna hermes-agent==0.19.0
-fi
-
-# -------------------------------------------------------
-# 1d. camel profile: the camel-ai agent runtime.
-# examples/camel/train.py imports camel.agents.ChatAgent, and
-# areal.experimental.camel.openai_model imports camel.logger / camel.messages /
-# camel.models.base_model / camel.types / camel.utils; camel-ai is NOT an
-# AReaL-declared dependency, so it is installed here. camel-ai goes in with
-# --no-deps so the job's NPU torch/vllm stack is never touched (camel is
-# API-based); its lighter runtime deps come in separately (none pull torch).
-# -------------------------------------------------------
-if [[ "$CAMEL_PREP" == 1 ]]; then
-  pip_tuna --no-deps camel-ai
-  pip_tuna tiktoken docstring-parser colorama python-dotenv jsonschema
-fi
-
-# -------------------------------------------------------
-# 1e. hitl-demo profile: the zeroclaw chat client + its home
-# config skeleton. human_in_the_loop_demo.py drives its own traffic through
-# the zeroclaw CLI (`zeroclaw agent -m ... --session-state-file ...`), so the
-# binary must be on PATH. The demo's FIRST action on the config is
-# shutil.copy2(~/.zeroclaw/config.toml, .demo_bak) -> FileNotFoundError if the
-# file is missing, and its regex patch silently skips a missing
-# default_provider line - so plant a minimal skeleton in the runner home dir
-# (the demo backs it up and restores it on exit; AReaL source is untouched).
-# If zeroclaw.com turns out unreachable from the CN runners, switch to the
-# fork's GitHub Releases aarch64-linux tarball.
-# -------------------------------------------------------
-if [[ "$HITL_PREP" == 1 ]]; then
-  # PINNED to v0.6.0: the last release line whose CLI matches the demo's
-  # hardcoded invocation `zeroclaw agent -m ... --session-state-file ...`.
-  # v0.6.0's agent subcommand has NO --agent flag and its Providers list
-  # documents `custom:<URL>` - the inline OpenAI-compatible provider the
-  # demo's config patcher writes. v0.7+ introduced the agent-alias system
-  # and made --agent REQUIRED ("Required - there is no default agent"),
-  # so the current official build (zeroclaw.com / releases/latest, v0.8.x)
-  # rejects the demo's invocation with a usage error (observed rc=2).
-  # zeroclaw.com's installer has no version flag, so fetch the pinned
-  # release tarball directly - github release downloads are reachable from
-  # the CN runners (zeroclaw.com's own installer fetches from there).
-  # /usr/local/bin is on the default PATH, so no PATH plumbing is needed.
-  ZC_VERSION=v0.6.0
-  ZC_ASSET="zeroclaw-aarch64-unknown-linux-gnu.tar.gz"
-  ZC_URL="https://github.com/zeroclaw-labs/zeroclaw/releases/download/${ZC_VERSION}/${ZC_ASSET}"
-  ZC_TMP="$(mktemp -d)"
-  # a 404 here usually means the aarch64 asset is missing in $ZC_VERSION
-  fetch_tarball "$ZC_URL" "$ZC_TMP/$ZC_ASSET"
-  tar -xzf "$ZC_TMP/$ZC_ASSET" -C "$ZC_TMP"
-  [[ -x "$ZC_TMP/zeroclaw" ]] || { echo "$ZC_ASSET did not contain a zeroclaw binary" >&2; exit 2; }
-  install -m 0755 "$ZC_TMP/zeroclaw" /usr/local/bin/zeroclaw
-  rm -rf "$ZC_TMP"
-  zeroclaw --version || { echo "zeroclaw failed to run" >&2; exit 2; }
-  # Same flat keys the demo's patcher regexes target (plus workspace_dir,
-  # mirroring what v0.6.0's own installer writes); the demo rewrites
-  # default_provider/api_key to the live gateway on start.
-  mkdir -p "$HOME/.zeroclaw"
-  cat > "$HOME/.zeroclaw/config.toml" <<'TOML'
-default_provider = "custom:http://localhost:1"
-default_model = "default"
-api_key = "placeholder"
-workspace_dir = "/tmp/zeroclaw-workspace"
-TOML
-  echo "planted minimal ~/.zeroclaw/config.toml skeleton"
-fi
-
 # -------------------------------------------------------
 # 1f. tau2 profile: the tau2-bench fork (async orchestrator
 # + user simulator) and its task data. TAU2_DATA_DIR must reach the trainer
@@ -254,42 +131,6 @@ if [[ "$TAU2_PREP" == 1 ]]; then
   echo "tau2-bench fork installed; TAU2_DATA_DIR=$TAU2_DIR/data"
 fi
 
-# -------------------------------------------------------
-# 1g. agent-service profile: the Claude Agent SDK stack.
-# examples/agent_service's ClaudeAgent shells out through claude-agent-sdk,
-# whose ClaudeSDKClient spawns the Claude Code CLI (an npm package) - so
-# besides the pip package we need Node >= 18 and the CLI itself. The NPU
-# image ships no node, so install a pinned Node from the npmmirror CDN
-# (CN-reachable, unlike nodejs.org) and the CLI from the npmmirror
-# registry. The API traffic is pointed at the LOCAL gateway by
-# run_example.sh's agent-service branch (ANTHROPIC_BASE_URL) - no real
-# Anthropic key is ever used.
-# -------------------------------------------------------
-if [[ "$AGENT_SERVICE_PREP" == 1 ]]; then
-  pip_tuna claude-agent-sdk
-  AS_NEED_NODE=1
-  if command -v node >/dev/null 2>&1; then
-    AS_NODE_MAJOR="$(node --version | sed 's/^v\([0-9]*\).*/\1/')"
-    if (( AS_NODE_MAJOR >= 18 )); then AS_NEED_NODE=0; fi
-  fi
-  if [[ "$AS_NEED_NODE" == 1 ]]; then
-    AS_NODE_VERSION=v22.14.0
-    AS_NODE_DIR="/usr/local/node-${AS_NODE_VERSION}"
-    fetch_tarball \
-      "https://registry.npmmirror.com/-/binary/node/${AS_NODE_VERSION}/node-${AS_NODE_VERSION}-linux-arm64.tar.gz" \
-      /tmp/as-node.tar.gz
-    mkdir -p "$AS_NODE_DIR"
-    tar -xzf /tmp/as-node.tar.gz -C "$AS_NODE_DIR" --strip-components=1
-    export PATH="$AS_NODE_DIR/bin:$PATH"
-    echo "PATH=$PATH" >> "$GITHUB_ENV"
-  fi
-  node --version || { echo "node unavailable for the claude CLI" >&2; exit 2; }
-  npm install -g @anthropic-ai/claude-code \
-    --registry=https://registry.npmmirror.com \
-    || { echo "claude code CLI install failed" >&2; exit 2; }
-  claude --version || { echo "claude CLI failed to run" >&2; exit 2; }
-fi
-
 # -------------------------------------------------------
 # 2. Runtime Environment setup.
 # -------------------------------------------------------
@@ -337,9 +178,8 @@ export AREAL_MODEL_ID="$MODEL_ID"
 
 # The resolved path is captured into the AREAL_MODEL_PATH shell variable (tee
 # keeps the diagnostics streaming into the job log) and written to GITHUB_ENV
-# for the run step. Later sections of this script - countdown's tokenizer
-# (section 6) and the scaffolding model alias (section 9) - consume it, so it
-# must be the real on-disk download dir.
+# for the run step (overlay args reference AREAL_MODEL_PATH), so it must be
+# the real on-disk download dir.
 AREAL_MODEL_PATH=$(python3 <<'PY' | tee /dev/stderr | sed -n 's/^AREAL_MODEL_PATH=//p' | tail -n 1
 import os
 import subprocess
@@ -391,215 +231,3 @@ PY
   echo "model resolution printed no AREAL_MODEL_PATH line" >&2
   exit 2
 }
-
-# -------------------------------------------------------
-# 4. AIME dataset + derived config (areal-math-aime).
-# -------------------------------------------------------
-# areal/dataset/aime.py loads <path>/aime_train.parquet and
-# <path>/aime_test.parquet from a LOCAL directory (upstream ships no aime
-# config; the loader takes no HF repo id). Convert the public MathArena
-# parquets (same flow as examples/distillation/mopd/README.md), then derive
-# the GRPO config from gsm8k_grpo_npu.yaml: point the datasets at the local
-# dir and set rollout.agent to null (aime_rl.py runs the standard
-# RLVRWorkflow, not an agent workflow). The generated config is exposed to
-# overlay args via AREAL_AIME_CONFIG.
-if [[ "$AIME_PREP" == 1 ]]; then
-python3 <<'PY'
-import os
-
-import pandas as pd
-import yaml
-from huggingface_hub import hf_hub_download
-
-workspace = os.environ["GITHUB_WORKSPACE"]
-target_root = os.environ["TARGET_ROOT"]
-aime_dir = os.path.join(workspace, "areal_data", "aime")
-os.makedirs(aime_dir, exist_ok=True)
-
-# hf_hub_download honors HF_ENDPOINT set in section 3.
-for repo, dst in [
-    ("MathArena/aime_2025", "aime_train.parquet"),
-    ("MathArena/aime_2026", "aime_test.parquet"),
-]:
-    src = hf_hub_download(
-        repo_id=repo, repo_type="dataset",
-        filename="data/train-00000-of-00001.parquet",
-    )
-    df = pd.read_parquet(src).rename(columns={"problem": "question"})
-    df = df[["question", "answer"]]
-    df.to_parquet(os.path.join(aime_dir, dst))
-    print(f"prepared {dst}: {len(df)} rows", flush=True)
-
-base = os.path.join(target_root, "examples", "math", "gsm8k_grpo_npu.yaml")
-with open(base, encoding="utf-8") as fh:
-    cfg = yaml.safe_load(fh)
-cfg["experiment_name"] = "aime-grpo"
-cfg["train_dataset"]["path"] = aime_dir
-cfg["valid_dataset"]["path"] = aime_dir
-config_path = os.path.join(aime_dir, "aime_grpo_npu.yaml")
-with open(config_path, "w", encoding="utf-8") as fh:
-    yaml.safe_dump(cfg, fh, sort_keys=False, allow_unicode=True)
-
-with open(os.environ["GITHUB_ENV"], "a") as fh:
-    fh.write(f"AREAL_AIME_CONFIG={config_path}\n")
-print(f"generated {config_path}", flush=True)
-PY
-fi
-
-# -------------------------------------------------------
-# 5. BoBa dataset (areal-math-boba).
-# -------------------------------------------------------
-# boba_grpo.py loads train_dataset.path as a LOCAL jsonl file
-# (load_dataset("json", data_files=path)); the upstream yaml ships the HF
-# repo id, which would fail with FileNotFoundError. Pre-download the public
-# AReaL-boba-106k.jsonl and expose its local path via AREAL_BOBA_DATA.
-if [[ "$BOBA_PREP" == 1 ]]; then
-python3 <<'PY'
-import os
-import shutil
-
-from huggingface_hub import hf_hub_download
-
-src = hf_hub_download(
-    repo_id="inclusionAI/AReaL-boba-Data", repo_type="dataset",
-    filename="AReaL-boba-106k.jsonl",
-)
-dst = os.path.join(
-    os.environ["GITHUB_WORKSPACE"], "areal_data", "boba", "AReaL-boba-106k.jsonl"
-)
-os.makedirs(os.path.dirname(dst), exist_ok=True)
-shutil.copyfile(src, dst)
-with open(os.environ["GITHUB_ENV"], "a") as fh:
-    fh.write(f"AREAL_BOBA_DATA={dst}\n")
-print(f"downloaded boba dataset to {dst}", flush=True)
-PY
-fi
-
-# -------------------------------------------------------
-# 6. Countdown dataset (areal-countdown-grpo).
-# -------------------------------------------------------
-# countdown.py writes ./data/countdown/qwen/*.jsonl relative to CWD; run it
-# from TARGET_ROOT so the generated files land where train_config.yaml's
-# default train_dataset.path=data/countdown/qwen/train_e.jsonl expects them
-# (run_example.sh also cds to TARGET_ROOT). The default generation size is
-# 500k samples; 32/8 is plenty for the 1-step smoke. The tokenizer comes
-# from the model predownloaded in section 3, so generation is offline.
-if [[ "$COUNTDOWN_PREP" == 1 ]]; then
-  MODEL_DIR="$AREAL_MODEL_PATH"
-  mkdir -p "$TARGET_ROOT/data/countdown/qwen"
-  (cd "$TARGET_ROOT" && python3 examples/countdown/countdown.py \
-    --num_samples 32 --eval_size 8 --tokenizer_path "$MODEL_DIR")
-fi
-
-# -------------------------------------------------------
-# 7. HH-RLHF dataset (areal-align: hhrlhf_dpo.py / hhrlhf_rw.py).
-# -------------------------------------------------------
-# Anthropic/hh-rlhf is public (NOT gated), but its jsonl.gz files live in
-# per-subset subdirectories (harmless-base/, helpful-base/, ...), so the
-# upstream loader's load_dataset(path, split=...) without a config name
-# fails with a multi-config error. Work around it boba-style: download the
-# harmless-base subset online (via HF_ENDPOINT) into a flat local
-# dir; load_dataset(<dir>, split="train"/"test") then infers splits from
-# the file names. Exposed to overlay args via AREAL_HHRLHF_DATA.
-# NOTE: the dir is deliberately named "hh-rlhf" 鈥?areal's dataset dispatch
-# matches the literal substring "hh-rlhf" in the path; a dir named
-# "hhrlhf" falls through to the load_from_disk fallback and errors out.
-if [[ "$HHRLHF_PREP" == 1 ]]; then
-python3 <<'PY'
-import os
-import shutil
-
-from huggingface_hub import hf_hub_download
-
-dst_dir = os.path.join(os.environ["GITHUB_WORKSPACE"], "areal_data", "hh-rlhf")
-os.makedirs(dst_dir, exist_ok=True)
-for filename in ("train.jsonl.gz", "test.jsonl.gz"):
-    src = hf_hub_download(
-        repo_id="Anthropic/hh-rlhf", repo_type="dataset",
-        filename=f"harmless-base/{filename}",
-    )
-    shutil.copyfile(src, os.path.join(dst_dir, filename))
-with open(os.environ["GITHUB_ENV"], "a") as fh:
-    fh.write(f"AREAL_HHRLHF_DATA={dst_dir}\n")
-print(f"prepared hh-rlhf harmless-base at {dst_dir}", flush=True)
-PY
-fi
-
-# -------------------------------------------------------
-# 8. ToRL dataset (areal-tir-grpo): fixture staging.
-# -------------------------------------------------------
-# The torl_data loader downloads its parquets from github.com at load time
-# (areal/dataset/torl_data.py), but the CI job containers cannot reach
-# github.com (mainland network, no proxy; the runner agent pulls code on the
-# host side). No HF/ModelScope mirror of ToRL exists (ModelScope 404;
-# upstream ships GitHub only), so the two tiny parquets are vendored as
-# fixtures - the documented last resort. prepare_torl_data() skips its
-# download once /tmp/areal/torl_data/_SUCCESS exists, so staging the files
-# plus the flag here makes the loader fully offline. The data service worker
-# runs in the same container, so /tmp is shared with the run step.
-# Fixture pin (GAIR-NLP/ToRL @ main): train.parquet blob 57bd7d5d (6035244
-# bytes), test.parquet blob e8ca20fd (51798 bytes).
-if [[ "$TORL_PREP" == 1 ]]; then
-  SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-  TORL_FIXTURE="$SCRIPT_DIR/../fixtures"
-  for f in train.parquet test.parquet; do
-    [[ -s "$TORL_FIXTURE/$f" ]] || {
-      echo "missing fixture: $TORL_FIXTURE/$f (see section 8 comment for source)" >&2
-      exit 2
-    }
-  done
-  mkdir -p /tmp/areal/torl_data
-  cp "$TORL_FIXTURE/train.parquet" "$TORL_FIXTURE/test.parquet" /tmp/areal/torl_data/
-  : > /tmp/areal/torl_data/_SUCCESS
-  echo "staged torl fixtures at /tmp/areal/torl_data (loader download skipped)"
-fi
-
-# -------------------------------------------------------
-# 9. Scaffolding model alias (areal-scaffold-grpo).
-# -------------------------------------------------------
-# examples/scaffolding hardcodes model="default" in its OpenAI-compatible
-# worker (workflow.py:102, a TRT-LLM convention), while AReaL's vLLM server
-# serves under the model path - requests 404 and every rollout comes back
-# empty, which then trips the "total loss weight must be positive" assert.
-# vLLMConfig has no served_model_name passthrough (strict dataclass, unknown
-# keys are rejected at the structured merge), so serve the model under the
-# literal name "default" instead: vLLM uses the raw --model string as the
-# served model name, so a relative "default" path resolving to the model dir
-# makes the name match. The server subprocess inherits the run CWD
-# (TARGET_ROOT), hence the symlink location.
-if [[ "$SCAFFOLD_PREP" == 1 ]]; then
-  MODEL_DIR="$AREAL_MODEL_PATH"
-  [[ -d "$MODEL_DIR" ]] || { echo "model dir missing: $MODEL_DIR" >&2; exit 2; }
-  ln -sfn "$MODEL_DIR" "$TARGET_ROOT/default"
-  echo "linked $TARGET_ROOT/default -> $MODEL_DIR (scaffolding model alias)"
-fi
-
-# -------------------------------------------------------
-# 10. CLEVR subset (areal-clevr-sft).
-# -------------------------------------------------------
-# The clevr SFT loader materializes per-sample pixel_values across the whole
-# split (clevr_count_70k.py:104-127), unlike the RL loader which only keeps
-# JPEG bytes, so the 70k train split blows the data worker (host OOM / tens
-# of GB of arrow cache). The example hardcodes split="train" and
-# get_custom_dataset ignores train_dataset.split when the script passes a
-# non-None split (dataset/__init__.py:322,328), and the loader takes no extra
-# kwargs - so subset via path instead: train_dataset.path points at a dir
-# whose name contains "clevr_count_70k" (dispatch substring) holding
-# train.parquet (load_dataset infers the split from the file name). The
-# absolute slice train[:200] downloads only the first shard (~0.5GB of the
-# 10.4GB), and the parquet keeps the HF feature metadata so images stay PIL.
-if [[ "$CLEVR_SUBSET_PREP" == 1 ]]; then
-python3 <<'PY'
-import os
-
-from datasets import load_dataset
-
-dst = os.path.join(os.environ["GITHUB_WORKSPACE"], "areal_data", "clevr_count_70k")
-os.makedirs(dst, exist_ok=True)
-ds = load_dataset("BUAADreamer/clevr_count_70k", split="train[:200]")
-ds.to_parquet(os.path.join(dst, "train.parquet"))
-with open(os.environ["GITHUB_ENV"], "a") as fh:
-    fh.write(f"AREAL_CLEVR_SUBSET_DIR={dst}\n")
-print(f"prepared clevr subset ({len(ds)} rows) at {dst}", flush=True)
-PY
-fi