feat: sandbox_agent swe resources server

environments/swe/README.md

@@ -0,0 +1,22 @@
+# swe
+
+Any agent harness for swe tasks in a sandbox.
+Run by `sandbox_agent`, scored by `anyswe` resources server. 
+
+Prepare data:
+
+```bash
+python environments/swe/prepare.py --input-jsonl <swebench_verified.jsonl> --limit 20
+```
+
+Start env:
+```bash
+ng_run "+config_paths=[environments/swe/config.yaml]"
+```
+
+Generate rollouts:
+```bash
+ng_collect_rollouts ... 
+```
+
+TODO more content

environments/swe/config.yaml

@@ -0,0 +1,97 @@
+anyswe:
+  resources_servers:
+    anyswe:
+      entrypoint: app.py
+      domain: agent
+      verified: false
+      description: Grades SWE task patches in fresh sandboxes with the official SWE-bench evaluation scripts.
+      value: Software engineering agent capabilities.
+      max_pass_to_pass: 20
+      eval_timeout: 1800
+      sandbox_provider:
+        opensandbox:
+          connection:
+            domain: ${oc.env:OPENSANDBOX_DOMAIN}
+            api_key: ${oc.env:OPENSANDBOX_API_KEY}
+            protocol: http
+            use_server_proxy: true
+            request_timeout_s: 1800
+sandbox_opencode_swe:
+  responses_api_agents:
+    sandbox_agent:
+      entrypoint: app.py
+      domain: agent
+      verified: false
+      description: opencode CLI harness solving SWE-bench tasks inside per-instance OpenSandbox images.
+      value: Software engineering agent capabilities with in-sandbox grading.
+      resources_server:
+        type: resources_servers
+        name: anyswe
+      model_server:
+        type: responses_api_models
+        name: policy_model
+      concurrency: 64
+      agent_module: responses_api_agents.opencode_agent.app
+      agent_class: OpenCodeAgent
+      agent_config_class: OpenCodeAgentConfig
+      agent_config:
+        resources_server:
+          type: resources_servers
+          name: anyswe
+        concurrency: 1
+        command: opencode
+        model: nvinf/policy_model
+        thinking: true
+        timeout: 3600
+        opencode_version: 1.17.8
+        repo_dir: /testbed
+        extra_args:
+          - --dangerously-skip-permissions
+        system_prompt: |
+          You are a software engineering agent. The repository is checked out at /testbed with the
+          bug already reproduced there. Investigate the issue described by the user, edit the code
+          under /testbed to fix it, and do not modify tests. Work only with your bash tool.
+        opencode_config:
+          permission:
+            bash: allow
+            edit: allow
+            webfetch: allow
+          provider:
+            nvinf:
+              npm: "@ai-sdk/openai-compatible"
+              options:
+                baseURL: __SANDBOX_MODEL_URL__/v1
+                apiKey: gym  # pragma: allowlist secret
+              models:
+                policy_model:
+                  interleaved:
+                    field: reasoning
+                  limit:
+                    context: 262144
+                    output: 262144
+      sandbox_provider:
+        opensandbox:
+          connection:
+            domain: ${oc.env:OPENSANDBOX_DOMAIN}
+            api_key: ${oc.env:OPENSANDBOX_API_KEY}
+            protocol: http
+            use_server_proxy: true
+            request_timeout_s: 1800
+      sandbox_image: ubuntu:24.04
+      sandbox_spec:
+        ttl_s: 7200
+        resources:
+          cpu: 2
+          memory_mib: 8192
+          disk_gib: 30
+      setup_commands:
+        - "command -v curl > /dev/null 2>&1 || (apt-get -qq update && apt-get -y -qq install curl ca-certificates > /dev/null 2>&1) || (yum -y -q install curl ca-certificates > /dev/null 2>&1) || true"
+        - "mkdir -p /deps && curl -fsSL https://github.com/astral-sh/python-build-standalone/releases/download/20241219/cpython-3.12.8+20241219-x86_64-unknown-linux-gnu-install_only.tar.gz | tar xz -C /deps --strip-components=1"
+        - "/deps/bin/python3 -m pip -q install nemo-gym"
+        - "curl -fsSL https://opencode.ai/install | VERSION=1.17.8 bash; command -v opencode >/dev/null 2>&1 || ln -sf $HOME/.opencode/bin/opencode /usr/local/bin/opencode"
+      sandbox_python: /deps/bin/python3
+      rollout_timeout: 4200
+      datasets:
+      - name: example
+        type: example
+        jsonl_fpath: environments/swe/data/example.jsonl

environments/swe/prepare.py

@@ -0,0 +1,114 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert SWE-bench rows into sandbox_agent rows graded by anyswe. TODO: r2e etc"""
+
+import argparse
+import json
+from pathlib import Path
+
+
+def _as_list(v) -> list[str]:
+    if isinstance(v, str):
+        try:
+            return list(json.loads(v))
+        except (json.JSONDecodeError, TypeError):
+            return [v] if v else []
+    return list(v or [])
+
+
+def _instance_image(container_formatter, instance_id: str) -> str:
+    fmt = container_formatter[0] if isinstance(container_formatter, list) else container_formatter
+    fmt = fmt or "swebench/sweb.eval.x86_64.{instance_id}"
+    if fmt.startswith("docker://"):
+        fmt = fmt[len("docker://") :]
+    tag = instance_id.replace("__", "_1776_").lower()
+    image = fmt.format(instance_id=tag)
+    if ":" not in image.rsplit("/", 1)[-1]:
+        image += ":latest"
+    return image
+
+
+def _benchmark(info: dict, override: str) -> str:
+    if override:
+        return override
+    name = str(info.get("dataset_name") or "")
+    if "R2E-Gym" in name:
+        return "r2e-gym"
+    if "Multilingual" in name:
+        return "swe-bench-multilingual"
+    return "swe-bench"
+
+
+def convert(row: dict, benchmark: str = "") -> dict:
+    info = row.get("problem_info") or row.get("verifier_metadata")
+    if not isinstance(info, dict):
+        meta = (row.get("responses_create_params") or {}).get("metadata")
+        info = meta if isinstance(meta, dict) else row
+    inst = (
+        json.loads(info["instance_dict"])
+        if isinstance(info.get("instance_dict"), str)
+        else dict(info["instance_dict"])
+    )
+    instance_id = info["instance_id"]
+    fail_to_pass = _as_list(inst.get("FAIL_TO_PASS") or inst.get("fail_to_pass"))
+    pass_to_pass = _as_list(inst.get("PASS_TO_PASS") or inst.get("pass_to_pass"))
+    if not fail_to_pass:
+        raise ValueError(f"no test directives for {instance_id}")
+    return {
+        "responses_create_params": {
+            "input": [{"role": "user", "content": inst.get("problem_statement") or info.get("problem_statement", "")}],
+            "metadata": {
+                "docker_image": _instance_image(info.get("container_formatter"), instance_id),
+                "patch_workdir": "/testbed",
+            },
+        },
+        "verifier_metadata": {
+            "instance_id": instance_id,
+            "benchmark": _benchmark(info, benchmark),
+            "test_patch": inst.get("test_patch", ""),
+            "fail_to_pass": fail_to_pass,
+            "pass_to_pass": pass_to_pass,
+            "instance_dict": inst,
+        },
+    }
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input-jsonl", required=True)
+    parser.add_argument("--output-jsonl", default="environments/swe/data/example.jsonl")
+    parser.add_argument("--limit", type=int, default=0)
+    parser.add_argument("--benchmark", default="", choices=["", "swe-bench", "swe-bench-multilingual", "r2e-gym"])
+    args = parser.parse_args()
+
+    rows = []
+    for line in Path(args.input_jsonl).read_text().splitlines():
+        if not line.strip():
+            continue
+        try:
+            rows.append(convert(json.loads(line), args.benchmark))
+        except Exception as e:
+            print(f"skip: {e}")
+        if args.limit and len(rows) >= args.limit:
+            break
+
+    out = Path(args.output_jsonl)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text("".join(json.dumps(r) + "\n" for r in rows))
+    print(f"wrote {len(rows)} rows to {out}")
+
+
+if __name__ == "__main__":
+    main()

resources_servers/anyswe/README.md

@@ -0,0 +1,15 @@
+# anyswe
+
+Grades SWE task patches.
+
+The agent (e.g. `sandbox_agent` with `patch_workdir: /testbed`) captures the rollout's
+`git diff` into response metadata as `model_patch`. This verifier rebuilds the task's
+instance container in a fresh sandbox, applies the patch, and runs the official SWE-bench
+evaluation scripts (`swebench` package, `make_test_spec` + log parsers) via
+`verify_task`. Reward is 1.0 when the instance is resolved.
+
+Dataset rows need `verifier_metadata` with `instance_id`, `test_patch`, `fail_to_pass`,
+`pass_to_pass`, and `instance_dict` (the raw SWE-bench instance row), plus
+`responses_create_params.metadata.docker_image` naming the instance image.
+
+See `environments/swe` for a full wiring with the sandboxed opencode harness.