From 4c3042ea5c6b3760851078b2da3e8a5eb7ec2680 Mon Sep 17 00:00:00 2001
From: XBetter <32125213+1998x-stack@users.noreply.github.com>
Date: Wed, 28 Oct 2026 15:52:06 +0800
Subject: [PATCH 1/6] test(oci): required CI integration cannot silently skip
 without image

---
 tests/test_docker_sandbox.py | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/tests/test_docker_sandbox.py b/tests/test_docker_sandbox.py
index 6f8eb94..068261e 100644
--- a/tests/test_docker_sandbox.py
+++ b/tests/test_docker_sandbox.py
@@ -131,6 +131,8 @@ async def test_real_docker_cgroup_network_and_workspace_boundary(tmp_path):
     """Actually execute the container when CI supplies a pulled immutable image."""
     image = os.environ.get("TINY_HARNESS_TEST_DOCKER_IMAGE")
     if not image:
+        if os.environ.get("TINY_HARNESS_REQUIRE_DOCKER_INTEGRATION") == "1":
+            pytest.fail("Required Docker integration image was not configured")
         pytest.skip("Docker integration requires TINY_HARNESS_TEST_DOCKER_IMAGE")
     workspace = tmp_path / "workspace"
     workspace.mkdir()
@@ -166,6 +168,8 @@ async def test_real_docker_timeout_removes_daemon_container(tmp_path, monkeypatc
     """A timed-out Docker CLI must not orphan its daemon-owned container."""
     image = os.environ.get("TINY_HARNESS_TEST_DOCKER_IMAGE")
     if not image:
+        if os.environ.get("TINY_HARNESS_REQUIRE_DOCKER_INTEGRATION") == "1":
+            pytest.fail("Required Docker integration image was not configured")
         pytest.skip("Docker integration requires TINY_HARNESS_TEST_DOCKER_IMAGE")
     import secrets
     import subprocess
@@ -195,6 +199,8 @@ async def test_real_docker_cancel_removes_daemon_container(tmp_path, monkeypatch
     """Client-side task cancellation must also force container teardown."""
     image = os.environ.get("TINY_HARNESS_TEST_DOCKER_IMAGE")
     if not image:
+        if os.environ.get("TINY_HARNESS_REQUIRE_DOCKER_INTEGRATION") == "1":
+            pytest.fail("Required Docker integration image was not configured")
         pytest.skip("Docker integration requires TINY_HARNESS_TEST_DOCKER_IMAGE")
     import asyncio
     import secrets

From 47d33d4382fede1ad8c3ade8daa7a541b5c3bc32 Mon Sep 17 00:00:00 2001
From: XBetter <32125213+1998x-stack@users.noreply.github.com>
Date: Wed, 28 Oct 2026 15:52:14 +0800
Subject: [PATCH 2/6] ci(oci): fail when real-container integration is
 unexpectedly skipped

---
 .github/workflows/ci.yml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index c4fac82..1eafb21 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -46,6 +46,8 @@ jobs:
   oci-sandbox-smoke:
     name: Real OCI cgroup and network isolation
     runs-on: ubuntu-latest
+    env:
+      TINY_HARNESS_REQUIRE_DOCKER_INTEGRATION: "1"
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5

From 8f0d4fa65af422b3ecdde76f339c31f53e8f18db Mon Sep 17 00:00:00 2001
From: XBetter <32125213+1998x-stack@users.noreply.github.com>
Date: Wed, 28 Oct 2026 15:52:36 +0800
Subject: [PATCH 3/6] docs(oci): document actual cgroup sandbox contract and
 production trust boundaries

---
 docs/OCI_SANDBOX.md | 108 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 108 insertions(+)
 create mode 100644 docs/OCI_SANDBOX.md

diff --git a/docs/OCI_SANDBOX.md b/docs/OCI_SANDBOX.md
new file mode 100644
index 0000000..bdd8f97
--- /dev/null
+++ b/docs/OCI_SANDBOX.md
@@ -0,0 +1,108 @@
+# OCI / Docker Shell Sandbox
+
+The built-in async Shell skill has three execution modes:
+
+| `shell_sandbox` | Behavior | Isolation claim |
+|---|---|---|
+| `off` (default) | Host subprocess, requires HITL by default | **No OS isolation** |
+| `bubblewrap` | Linux user/mount/network namespaces, per-process rlimits | Host-dependent; real namespace test may skip in CI |
+| `docker` | Linux Docker OCI container, cgroup resource caps, network none | Real Docker integration test required and passing on Ubuntu CI |
+
+The Docker backend applies **only** to the registered async
+`run_command` Shell skill. File and third-party Python tools run on the
+host and are not sandboxed by this setting.
+
+## Configure a trusted, pre-pulled immutable image
+
+Choose and audit an image suitable for the commands you intend to run.
+For a local demonstration with Alpine:
+
+```bash
+docker pull alpine:3.20
+docker image inspect --format '{{index .RepoDigests 0}}' alpine:3.20
+```
+
+Copy the resulting `repository@sha256:...` into `--sandbox-image`.
+Digest pinning ensures the selected image cannot silently change during
+execution. The agent **never pulls images** (`--pull=never`). Supply
+only images you trust; the daemon's image store and Docker access are
+trusted host components.
+
+```bash
+tiny-harness \
+  --provider deepseek \
+  --workspace /absolute/path/to/disposable-project \
+  --skills files,shell \
+  --shell-sandbox docker \
+  --sandbox-image 'alpine@sha256:<your_64_hex_digest>' \
+  --sandbox-cpus 1.0 \
+  --sandbox-memory-mb 1024 \
+  --sandbox-pids-limit 64
+```
+
+The image digest in the example is a **placeholder**, not a usable pin.
+Set `--sandbox-image` to the actual inspected `repository@sha256`
+reference. No-hitl Shell still requires the separate explicit
+`--allow-unsafe-shell` opt-in; keep HITL enabled for regular use.
+
+## Runtime behavior
+
+The builder uses a fixed Docker CLI argv and passes the model's command
+only as a final argument to `/bin/sh -c` **inside** the container:
+
+- `--network=none` blocks container external networking.
+- `--read-only` makes the image root filesystem read-only, with a private
+  bounded `/tmp` tmpfs.
+- `--mount type=bind,...,target=/workspace` exposes exactly the selected
+  host project as the writable workspace. All cwd paths must remain inside
+  it, and system root directories are rejected.
+- `--cap-drop=ALL` plus `--security-opt=no-new-privileges:true`
+  removes ambient capabilities and blocks gaining additional privileges.
+- `--user uid:gid` uses the non-root host identity (root host user is
+  rejected) to preserve project file ownership.
+- `--memory` and `--memory-swap` impose container memory/swap ceilings;
+  `--cpus` sets aggregate CPU quota; `--pids-limit` limits container
+  processes. These are **cgroup** limits across container processes,
+  unlike Bubblewrap's inherited per-process rlimits.
+- Each call receives a generated unique name, a private Docker CLI config
+  directory, and is cleaned with daemon-side `docker rm -f` even after
+  timeout or task cancellation. Failures never fall back to host Shell.
+
+The existing Shell output cap and wall-clock timeout still apply.
+The Bubblewrap-only `sandbox_cpu_seconds` and `sandbox_file_mb`
+flags do not create Docker disk-usage quotas.
+
+## Real validation
+
+GitHub CI job `Real OCI cgroup and network isolation` installs package
+dependencies, pulls the test image, resolves an immutable image digest,
+then executes the Shell skill **inside a real container**. It asserts:
+
+1. The container's cgroup memory and PID limits match the requested values.
+2. The container has no ordinary network routes.
+3. The host file outside the selected workspace is not visible.
+4. Writing into the designated workspace succeeds.
+5. Timed-out and cancelled jobs leave no running/retained named container.
+
+Unlike Bubblewrap's optional smoke, the Docker job is **mandatory** and
+fails if the required image environment is missing. Its passing state is
+evidence for this tested GitHub runner configuration, not universal proof.
+
+## Trust boundary and remaining risks
+
+**Docker daemon access is highly privileged**. Never mount
+`/var/run/docker.sock`, pass daemon credentials to models, or expose the
+Docker API to untrusted users. This implementation targets a trusted local
+daemon, not hostile multi-tenant compute, rootless Kubernetes, or remote
+Docker daemons.
+
+Other risks remain: workspace contents and executable image code must be
+trusted appropriately; volume paths could be swapped by concurrent host
+processes; Docker daemon and kernel vulnerabilities are out of scope;
+system-default seccomp profiles are daemon dependent; the shared host daemon
+is not an independent security trust domain.
+
+Production-grade multi-tenant use needs a dedicated daemon/VM per tenant,
+audited images, verified seccomp/AppArmor/SELinux profiles, identity mapping
+and tighter filesystem protections, policy gates for *every* tool handler,
+and integration tests on each supported host and kernel.

From 403f1de9103dcace3885f0836a75777773358865 Mon Sep 17 00:00:00 2001
From: XBetter <32125213+1998x-stack@users.noreply.github.com>
Date: Wed, 28 Oct 2026 15:52:59 +0800
Subject: [PATCH 4/6] docs(phase6): record real OCI cgroup evidence and
 remaining trust boundaries

---
 docs/OPTIMIZATION_STATUS.md | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/docs/OPTIMIZATION_STATUS.md b/docs/OPTIMIZATION_STATUS.md
index 269d399..fe68021 100644
--- a/docs/OPTIMIZATION_STATUS.md
+++ b/docs/OPTIMIZATION_STATUS.md
@@ -28,12 +28,13 @@
 | Phase 4 | [#22](https://github.com/1998x-stack/tiny-harness/pull/22) | Draft 2020-12 JSON Schema（嵌套/组合/数组），注册时检查 Schema，拒绝远程引用 | PR CI 通过；引入 jsonschema 运行时依赖 |
 | Phase 5 | [#24](https://github.com/1998x-stack/tiny-harness/pull/24) | Linux Bubblewrap 可选 shell namespace + workspace-only bind, 缺失时 fail-closed | Python CI 通过，实际 namespace smoke 最初被跳过 |
 | Phase 5 | [#25](https://github.com/1998x-stack/tiny-harness/pull/25) | prlimit CPU/虚拟内存/单文件大小逐进程资源约束 | Python CI 通过，非 cgroup 总配额 |
+| Phase 6 | [#27](https://github.com/1998x-stack/tiny-harness/pull/27) | OCI/Docker 可选后端，固定镜像摘要、只读根、network=none、cgroup CPU/内存/PID 限制及容器回收 | GitHub Runner 实际 Docker 隔离、cgroup、超时/取消清理测试通过 |
 
 ## 待完成事项（不应标记为已修复）
 
 ### P1：继续按独立 PR 完成
 
-1. **OS 沙箱验证及增强**：已实现可选 Linux Bubblewrap namespace 和 prlimit；尚需真实机器上的隔离测试、cgroup 总资源配额、系统级加固和 Windows/macOS 后端。默认 off 仍是不隔离的 Shell，**禁止**把 opt-in Bubblewrap 描述为完整生产沙箱。
+1. **OS 沙箱验证及增强**：已实现可选 Linux Bubblewrap 与 OCI Docker 后端；Docker 的 cgroup 限制及实际网络/工作区隔离已在 GitHub Linux runner 通过。仍需独立宿主环境的恶意负载评估、seccomp/AppArmor/SELinux 强化、daemon 隔离、竞态修复和其他平台后端。默认 off 仍是不隔离的 Shell，**不能**据 CI 单环境验证声称完整生产级沙箱。
 2. **ToolDef 参数级策略**：基础风险枚举与 no-HITL Shell 显式 opt-in 已完成；后续引入细粒度 allow/ask/deny、按命令参数审批策略，扩展其他危险工具的无 HITL 拒绝边界。
 3. **Provider 全面契约**：Anthropic 自定义 endpoint、截断判断和首次输出前的有限流式重试已完成；仍需统一非流式/流式重试、DeepSeek thinking 透传、HTTP 客户端连接复用及真实 API 冒烟。
 4. **上下文窗口**：已提供可配置窗口与启发式压缩；仍需 provider 真实 usage 回收、tokenizer 精确统计、可审计摘要和更精细的窗口分层治理。

From 89e1f35a1408031cb8f03c44f0b0546cf2010925 Mon Sep 17 00:00:00 2001
From: XBetter <32125213+1998x-stack@users.noreply.github.com>
Date: Wed, 28 Oct 2026 15:53:02 +0800
Subject: [PATCH 5/6] docs(phase6): record real OCI cgroup evidence and
 remaining trust boundaries

---
 docs/security.md | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/docs/security.md b/docs/security.md
index 238de8b..82b607c 100644
--- a/docs/security.md
+++ b/docs/security.md
@@ -10,13 +10,14 @@ tiny-harness 是可执行工具的 Agent harness，**不是安全沙箱**。模
 - `ToolDef` 未声明风险时默认 mutation；未知 risk_level 在注册阶段拒绝；safe 风险等级仍执行文件路径守卫。
 - `--no-hitl` 下内置 Shell 工具默认不可执行；使用 `--no-hitl --allow-unsafe-shell` 才能显式启用（不等于沙箱）。
 - 可选 `shell_sandbox="bubblewrap"` 为 Linux 内置 async Shell 提供 mount/network/PID namespace 约束，启动失败不自动降级；`prlimit` CPU/AS/FSIZE 限制是逐进程限制，非 cgroup 总配额。参见 [SHELL_SANDBOX.md](SHELL_SANDBOX.md)。
+- 可选 `shell_sandbox="docker"` 在 Linux 上使用摘要固定的本地 OCI 镜像、非 root UID、只读根、禁网络、最小权限和 cgroup CPU/内存/PID 总限制；真实容器 CI 验证通过，但 Docker daemon、镜像、工作区挂载仍是信任边界。参见 [OCI_SANDBOX.md](OCI_SANDBOX.md)。
 - built-in Shell skill 的 async handler 运行于指定 cwd；POSIX 上超时/取消时终止进程组；不继承 `*_API_KEY` 等未在 allow-list 的环境变量。
 - 会话快照使用原子替换与受限文件权限（POSIX `0600`）。
 
 ## 未提供的隔离能力
 
 - **默认 `shell_sandbox=off` 时 Shell 没有操作系统级文件系统隔离**：即使 cwd 在 workspace，命令仍可能访问机器上该进程有权限访问的路径。开启 bubblewrap 只是选配 Linux namespace 隔离，尚未在所有目标环境做端到端验证。
-- Shell 可能访问网络、用户凭据文件、Git 配置及 SSH agent；环境变量白名单不能覆盖全部隐式凭据路径。
+- 默认不隔离的 Shell 仍可能访问网络、用户凭据文件、Git 配置及 SSH agent；环境变量白名单不能覆盖全部隐式凭据路径。Docker 模式通过独立容器约束这条路径，但不保护其他 Python 工具。
 - 历史同步 `run_command` 辅助函数保留用于兼容，并未获得与 registered async Shell 等价的进程组取消保障。
 - Path guard 的检查和 handler 的实际打开文件不是一个不可分割的 OS 操作；工作区内容可被外部进程并发替换时仍有 TOCTOU 风险。
 - 在 Windows 上，异步 handler 的 `process.kill()` 不提供 POSIX 进程组等价的孙进程终止保障。

From c37bf5ca91b93823c38e7e68f771ef3802e0440b Mon Sep 17 00:00:00 2001
From: XBetter <32125213+1998x-stack@users.noreply.github.com>
Date: Wed, 28 Oct 2026 15:53:05 +0800
Subject: [PATCH 6/6] docs(phase6): record real OCI cgroup evidence and
 remaining trust boundaries

---
 README.md | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index eb9bb15..2f86608 100644
--- a/README.md
+++ b/README.md
@@ -313,7 +313,7 @@ ANTHROPIC_API_KEY="sk-ant-..." pytest tests/test_integration.py -v -m integratio
 
 The configured `context_window` reserves `max_tokens` for model output. When input history grows too large, the harness compacts oversized tool outputs and can discard complete older turns, preserving the latest user turn and tool-call/result pairing. This budget currently uses a heuristic estimate, not model-specific tokenizers.
 
-The built-in Shell skill requires human approval by default. In `--no-hitl` mode, Shell additionally requires `--allow-unsafe-shell`. The default `shell_sandbox=off` is **not** OS-isolated; Linux users may opt into `--shell-sandbox bubblewrap` for namespace-isolated execution with per-process resource limits. See [Linux Shell Sandbox](docs/SHELL_SANDBOX.md).
+The built-in Shell skill requires human approval by default. In `--no-hitl` mode, Shell additionally requires `--allow-unsafe-shell`. The default `shell_sandbox=off` is **not** OS-isolated. Linux can use [Bubblewrap](docs/SHELL_SANDBOX.md) with per-process limits, or opt into a [Docker OCI sandbox](docs/OCI_SANDBOX.md) with container-wide cgroup limits (validated in real GitHub CI).
 
 ## Optional Linux Shell Sandbox
 
@@ -323,6 +323,19 @@ tiny-harness --provider deepseek --skills files,shell --shell-sandbox bubblewrap
 
 Requires system-installed `bubblewrap` and `prlimit`, and working unprivileged namespaces. Sandboxed mode blocks external network by default; it never falls back to a host shell if setup fails. These per-process limits are **not** cgroup-wide quotas.
 
+## OCI Docker Sandbox (Linux)
+
+```bash
+docker pull alpine:3.20
+docker image inspect --format '{{index .RepoDigests 0}}' alpine:3.20
+tiny-harness --provider deepseek --skills files,shell \
+  --shell-sandbox docker \
+  --sandbox-image 'alpine@sha256:<your_64_hex_digest>' \
+  --sandbox-cpus 1.0 --sandbox-memory-mb 1024 --sandbox-pids-limit 64
+```
+
+Use the **actual** digest returned by `docker image inspect`; the displayed digest is only a placeholder. The Docker sandbox disables external networking, requires a pre-pulled image and uses a container read-only root with a writable workspace. A real OCI CI job verifies cgroup and filesystem/network isolation. Docker daemon privileges and third-party Python tools are still outside this boundary. See [OCI Sandbox details](docs/OCI_SANDBOX.md).
+
 ## Security
 
 Tool arguments are checked with JSON Schema Draft 2020-12 and declared filesystem paths are normalized before tool execution. **These checks are not an OS sandbox**. Read [security boundaries](docs/security.md) and the [phased optimization status](docs/OPTIMIZATION_STATUS.md) before deploying to untrusted environments.