Harden GenPlan container runtime

src/robocode/utils/apptainer_sandbox.py

@@ -77,6 +77,7 @@
     _filtered_repo_mounts,
     _find_repo_root,
     _get_claude_oauth_token,
+    _kindergarden_asset_volumes,
     _mcp_prestart_wrapper,
     container_python,
 )
@@ -235,6 +236,7 @@ def _build_apptainer_cmd(
     agent_cmd: list[str],
     extra_binds: list[str] | None = None,
     ss_pybullet_abs: str | None = None,
+    kindergarden_asset_binds: list[str] | None = None,
 ) -> list[str]:
     """Assemble the full ``apptainer exec`` command line.
 
@@ -289,6 +291,8 @@ def _build_apptainer_cmd(
         ]
     if ss_pybullet_abs is not None:
         cmd += ["--bind", f"{ss_pybullet_abs}:/robocode/third-party/ss-pybullet:ro"]
+    for bind in kindergarden_asset_binds or []:
+        cmd += ["--bind", bind]
     for bind in extra_binds or []:
         cmd += ["--bind", bind]
     cmd += [
@@ -403,6 +407,9 @@ async def run_agent_in_apptainer_sandbox(
             ss_pybullet_abs=(
                 str(ss_pybullet.resolve()) if ss_pybullet is not None else None
             ),
+            kindergarden_asset_binds=(
+                [] if config.blackbox else _kindergarden_asset_volumes()
+            ),
             kinder_baselines_abs=(
                 str(filtered_kinder_baselines.resolve())
                 if filtered_kinder_baselines is not None
@@ -520,6 +527,7 @@ def run_genplan_in_apptainer(
         bilevel_env: list[str] = []
         bilevel_bind: list[str] = []
         ss_pybullet_bind: list[str] = []
+        asset_binds = _kindergarden_asset_volumes()
         if ss_pybullet is not None:
             ss_pybullet_bind = [
                 "--bind",
@@ -545,6 +553,7 @@ def run_genplan_in_apptainer(
             f"{filtered_src.resolve()}:/robocode/src",
             "--bind",
             f"{filtered_kindergarden.resolve()}:/robocode/third-party/kindergarden",
+            *(tok for bind in asset_binds for tok in ("--bind", bind)),
             *ss_pybullet_bind,
             *bilevel_bind,
             str(sif_path),

src/robocode/utils/docker_sandbox.py

@@ -46,7 +46,7 @@
 import tempfile
 import time
 import uuid
-from collections.abc import Iterator
+from collections.abc import Callable, Iterator
 from contextlib import ExitStack, contextmanager, nullcontext
 from dataclasses import dataclass
 from pathlib import Path
@@ -176,6 +176,36 @@ def _find_repo_root() -> Path:
     )
 
 
+_LARGE_KINDERGARDEN_ASSET_DIRS = (
+    Path("src/kinder/envs/dynamic3d/models/assets"),
+    Path("src/kinder/envs/kinematic3d/assets"),
+)
+
+
+def _ignore_kindergarden_copy_patterns(
+    kindergarden: Path, blackbox: bool
+) -> Callable[[str, list[str]], set[str]]:
+    """Skip non-source payloads when staging Kindergarden into /tmp."""
+    root = kindergarden.resolve()
+    skip_names = {"tests", "docs", "demos"}
+    if blackbox:
+        skip_names.add("envs")
+
+    def ignore(dirpath: str, names: list[str]) -> set[str]:
+        ignored = {name for name in names if name in skip_names}
+        try:
+            rel_dir = Path(dirpath).resolve().relative_to(root)
+        except ValueError:
+            rel_dir = Path()
+        for name in names:
+            child = rel_dir / name
+            if child in _LARGE_KINDERGARDEN_ASSET_DIRS:
+                ignored.add(name)
+        return ignored
+
+    return ignore
+
+
 def _copy_kindergarden_without_tests(
     kindergarden: Path, dest: Path, *, blackbox: bool = False
 ) -> None:
@@ -184,14 +214,36 @@ def _copy_kindergarden_without_tests(
     ``demos/`` holds recorded solutions. Blackbox also skips ``kinder/envs/``;
     the installable skeleton stays so ``uv sync --frozen`` succeeds.
     """
-    skip = ("tests", "docs", "demos") + (("envs",) if blackbox else ())
+    # Large meshes/textures are source-adjacent data, not source code. Copying
+    # them into every per-run staging directory exhausts /tmp when several
+    # replicates start, so callers bind-mount the originals read-only.
     shutil.copytree(
         kindergarden,
         dest,
-        ignore=shutil.ignore_patterns(*skip),
+        ignore=_ignore_kindergarden_copy_patterns(kindergarden, blackbox),
     )
 
 
+def _kindergarden_asset_mounts() -> list[tuple[Path, str]]:
+    """Return existing bulky Kindergarden asset dirs and their container paths."""
+    kindergarden = _find_repo_root() / "third-party" / "kindergarden"
+    container_root = "/robocode/third-party/kindergarden"
+    mounts: list[tuple[Path, str]] = []
+    for rel_path in _LARGE_KINDERGARDEN_ASSET_DIRS:
+        host_path = kindergarden / rel_path
+        if host_path.is_dir():
+            mounts.append((host_path, f"{container_root}/{rel_path.as_posix()}:ro"))
+    return mounts
+
+
+def _kindergarden_asset_volumes() -> list[str]:
+    """Return Docker/Apptainer volume specs for bulky Kindergarden assets."""
+    return [
+        f"{host.resolve()}:{container}"
+        for host, container in _kindergarden_asset_mounts()
+    ]
+
+
 # The two kinder-baselines subpackages robocode depends on (editable path deps in
 # uv.lock). Only these are needed for the sandbox's `uv sync --frozen`; the rest
 # of the monorepo is not copied/mounted.
@@ -328,6 +380,7 @@ def _docker_run_prefix(
     env_args: list[str] | None = None,
     map_host_gateway: bool = False,
     ss_pybullet: Path | None = None,
+    kindergarden_asset_volumes: list[str] | None = None,
     extra_volumes: list[str] | None = None,
     env_server_port: int | None = None,
 ) -> list[str]:
@@ -382,6 +435,8 @@ def _docker_run_prefix(
             "-v",
             f"{filtered_kindergarden.resolve()}:/robocode/third-party/kindergarden",
         ]
+    for volume in kindergarden_asset_volumes or []:
+        cmd += ["-v", volume]
     if filtered_kinder_baselines is not None:
         cmd += [
             "-v",
@@ -449,6 +504,7 @@ def run_genplan_in_docker(
             auth_args,
             firewall_domains,
             ss_pybullet=ss_pybullet,
+            kindergarden_asset_volumes=_kindergarden_asset_volumes(),
         ) + [DOCKER_PYTHON, "-m", "robocode.approaches.genplan_driver"]
         logger.info("Starting genplan Docker container %s", container_name)
         try:
@@ -736,6 +792,9 @@ async def run_agent_in_docker_sandbox(
             auth_args,
             firewall_domains,
             ss_pybullet=ss_pybullet,
+            kindergarden_asset_volumes=(
+                [] if config.blackbox else _kindergarden_asset_volumes()
+            ),
             extra_volumes=session_volumes + tel_volumes,
             env_args=[
                 "-e",

tests/utils/test_apptainer_sandbox.py

@@ -182,7 +182,6 @@ def test_build_cmd_basic_shape(tmp_path: Path) -> None:
     assert "--cleanenv" in cmd
     pwd_idx = cmd.index("--pwd")
     assert cmd[pwd_idx + 1] == "/sandbox"
-
     # Env vars are passed as `--env KEY=val` pairs.
     assert "CLAUDE_CODE_MAX_OUTPUT_TOKENS=8192" in cmd
     assert "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=70" in cmd
@@ -207,6 +206,30 @@ def test_build_cmd_basic_shape(tmp_path: Path) -> None:
     assert cmd[-3:] == ["claude", "--print", "hello"]
 
 
+def test_build_cmd_mounts_kindergarden_assets_read_only(tmp_path: Path) -> None:
+    """Dynamic3D assets are shared rather than copied into per-run staging."""
+    bind = (
+        "/host/assets:/robocode/third-party/kindergarden/src/kinder/envs/"
+        "dynamic3d/models/assets:ro"
+    )
+    config = ApptainerSandboxConfig(
+        sandbox_dir=tmp_path / "sandbox",
+        sif_path=tmp_path / "robocode-sandbox.sif",
+    )
+    cmd = _build_apptainer_cmd(
+        config,
+        sandbox_abs="/host/sandbox",
+        src_abs="/host/src",
+        kindergarden_abs="/host/kindergarden",
+        kinder_baselines_abs=None,
+        auth_args=[],
+        firewall_domains=[],
+        agent_cmd=["claude"],
+        kindergarden_asset_binds=[bind],
+    )
+    assert bind in cmd
+
+
 def test_build_cmd_bilevel_conditional(tmp_path: Path) -> None:
     """The kinder-baselines bind and --extra bilevel sync appear only when requested."""
 

tests/utils/test_docker_sandbox.py

@@ -42,6 +42,7 @@
     DOCKER_PYTHON,
     DockerSandboxConfig,
     _build_docker_auth_args,
+    _copy_kindergarden_without_tests,
     _copy_src,
     _docker_run_prefix,
     _filtered_repo_mounts,
@@ -191,6 +192,27 @@ def test_filtered_repo_mounts_default_excludes_demos() -> None:
         assert not (kindergarden / "demos").exists()
 
 
+def test_kindergarden_copy_excludes_large_assets(
+    tmp_path: Path,
+) -> None:
+    """Large read-only scene assets are not duplicated in each /tmp staging tree."""
+    source = tmp_path / "source"
+    dynamic_assets = source / "src/kinder/envs/dynamic3d/models/assets"
+    kinematic_assets = source / "src/kinder/envs/kinematic3d/assets"
+    dynamic_assets.mkdir(parents=True)
+    kinematic_assets.mkdir(parents=True)
+    (dynamic_assets / "large-mesh.bin").write_bytes(b"mesh")
+    (kinematic_assets / "large-texture.png").write_bytes(b"texture")
+    (source / "pyproject.toml").touch()
+
+    destination = tmp_path / "destination"
+    _copy_kindergarden_without_tests(source, destination)
+
+    assert (destination / "pyproject.toml").exists()
+    assert not (destination / "src/kinder/envs/dynamic3d/models/assets").exists()
+    assert not (destination / "src/kinder/envs/kinematic3d/assets").exists()
+
+
 def test_filtered_repo_mounts_default_keeps_package_skeleton() -> None:
     """The demos exclusion must not take the installable skeleton with it."""
     with _filtered_repo_mounts() as (_, kindergarden, _tmp, _ss):
@@ -310,6 +332,27 @@ def test_docker_run_prefix_adds_extra_volumes(tmp_path: Path) -> None:
     assert cmd[cmd.index(volume) - 1] == "-v"
 
 
+def test_docker_run_prefix_mounts_kindergarden_assets_read_only(tmp_path: Path) -> None:
+    """Dynamic3D assets are shared rather than copied into per-run staging."""
+    volume = (
+        f"{tmp_path.resolve()}/assets:"
+        "/robocode/third-party/kindergarden/src/kinder/envs/dynamic3d/"
+        "models/assets:ro"
+    )
+    cmd = _docker_run_prefix(
+        "c",
+        "img",
+        tmp_path,
+        tmp_path / "src",
+        tmp_path / "kg",
+        None,
+        [],
+        [],
+        kindergarden_asset_volumes=[volume],
+    )
+    assert volume in cmd
+
+
 def test_docker_run_prefix_mounts_sandbox_not_parent_run_dir(tmp_path: Path) -> None:
     """Only the agent child directory, not its parent Hydra run, is mounted."""
     run_dir = tmp_path / "run"