[Unvalidated] Terminus 2

benchmarks/nemotron_3.5_super/eval_container_config.yaml

@@ -9,11 +9,10 @@ config_paths:
 - benchmarks/swebench/verified/opencode.yaml
 - benchmarks/swebench/multilingual/opencode.yaml
 - benchmarks/deepswe/opencode.yaml
-- benchmarks/terminal_bench_2_1/opencode.yaml
+- benchmarks/terminal_bench_2_1/terminus_2.yaml
 - benchmarks/swebench/pro/opencode.yaml
 - responses_api_models/vllm_model/configs/vllm_model.yaml
-# - nemo_gym/sandbox/providers/opensandbox/configs/opensandbox.yaml
-# SWE Bench Pro
+- benchmarks/nemotron_3.5_super/sandbox_utils.yaml
 
 ########################################
 # START Secrets
@@ -109,9 +108,9 @@ opencode_sandboxed_agent_deepswe:
   responses_api_agents:
     opencode_sandboxed_agent:
       datasets: []
-opencode_sandboxed_agent_terminal_bench_2_1:
+terminus_2_sandboxed_agent_terminal_bench_2_1:
   responses_api_agents:
-    opencode_sandboxed_agent:
+    terminus_2_sandboxed_agent:
       datasets: []
 opencode_sandboxed_agent_swebench_pro:
   responses_api_agents:

benchmarks/nemotron_3.5_super/eval_container_config_with_staged.yaml

@@ -32,9 +32,9 @@ opencode_sandboxed_agent_deepswe:
           prepare_script: benchmarks/deepswe/prepare.py
           num_repeats: 5
 
-opencode_sandboxed_agent_terminal_bench_2_1:
+terminus_2_sandboxed_agent_terminal_bench_2_1:
   responses_api_agents:
-    opencode_sandboxed_agent:
+    terminus_2_sandboxed_agent:
       datasets:
         - name: terminal_bench_2_1
           type: benchmark

benchmarks/terminal_bench_2_1/terminus_2.yaml

@@ -0,0 +1,21 @@
+config_paths:
+- responses_api_agents/terminus_2_sandboxed_agent/configs/terminus_2_sandboxed_agent.yaml
+- resources_servers/terminal_bench_2_1/configs/terminal_bench_2_1.yaml
+
+
+terminus_2_sandboxed_terminal_bench_2_1_resources_server:
+  _inherit_from: terminal_bench_2_1_resources_server
+
+terminus_2_sandboxed_agent_terminal_bench_2_1:
+  _inherit_from: terminus_2_sandboxed_agent
+  responses_api_agents:
+    terminus_2_sandboxed_agent:
+      resources_server:
+        type: resources_servers
+        name: terminus_2_sandboxed_terminal_bench_2_1_resources_server
+      datasets:
+        - name: terminal_bench_2_1
+          type: benchmark
+          jsonl_fpath: benchmarks/terminal_bench_2_1/data/benchmark.jsonl
+          prepare_script: benchmarks/terminal_bench_2_1/prepare.py
+          num_repeats: 8

responses_api_agents/terminus_2_sandboxed_agent/README.md

@@ -0,0 +1,39 @@
+# Harbor Terminus 2 Agent
+
+This agent runs Harbor's `Terminus2` control loop in the task sandbox supplied by
+the NeMo Gym resources server. It adapts the small Harbor environment interface
+that Terminus uses (`exec` and `is_dir`) to `AsyncSandbox`; task state therefore
+remains owned by the resources server.
+
+```bash
+gym env start \
+    --config responses_api_models/vllm_model/configs/vllm_model.yaml \
+    --config nemo_gym/sandbox/providers/opensandbox/configs/opensandbox.yaml \
+    --config responses_api_agents/terminus_2_sandboxed_agent/configs/terminus_2_sandboxed_agent.yaml \
+    --config resources_servers/swebench/configs/swebench.yaml
+```
+
+To run one row from a benchmark JSONL after starting the servers:
+
+```bash
+python responses_api_agents/terminus_2_sandboxed_agent/client.py \
+    +benchmark_jsonl=benchmarks/swebench/data/swebench_verified_benchmark.jsonl
+```
+
+The agent calls the configured model server exclusively through the Responses
+API. Its returned response contains every model request and response from the
+Terminus trajectory. Set `dump_trajectory: true` to also have Harbor write its
+per-turn JSON trajectory files; it is `false` by default.
+
+## Download and mount Tmux binary
+```bash
+curl -fL \
+  -o tmux-3.7c-linux-x86_64.tar.gz \
+  https://github.com/tmux/tmux-builds/releases/download/v3.7c/tmux-3.7c-linux-x86_64.tar.gz
+tar -xzf tmux-3.7c-linux-x86_64.tar.gz
+mv tmux tmux-3.7c-linux-x86_64
+
+# e.g. copy to mounted S3 bucket
+aws s3 cp tmux-3.7c-linux-x86_64 \
+  s3://tmux/tmux-3.7c-linux-x86_64
+```

responses_api_agents/terminus_2_sandboxed_agent/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0