docs: add OpenCode integration example

docs/guide/integrations/index.md

@@ -47,4 +47,5 @@ lang: en-US
 
 | Title | Author | Date | Tags |
 | --- | --- | --- | --- |
+| [OpenCode Integration Guide](./opencode.md) | Sam-Hui-dot | 2026-07-05 | integration, opencode, coding-agent, agent |
 | [Pi Agent Integration Guide](./pi-agent.md) | chaojixinren | 2026-07-01 | integration, pi-agent, coding-agent, agent |

docs/guide/integrations/opencode.md

@@ -0,0 +1,227 @@
+---
+title: OpenCode Integration Guide
+author: Sam-Hui-dot
+date: 2026-07-05
+tags:
+  - integration
+  - opencode
+  - coding-agent
+  - agent
+lang: en-US
+---
+
+# OpenCode Integration Guide
+
+[中文文档](../../zh/guide/integrations/opencode.md)
+
+Run [OpenCode](https://opencode.ai/), an open-source terminal coding agent,
+inside CubeSandbox MicroVMs. This guide pairs with the runnable
+[`examples/opencode-integration`](https://github.com/TencentCloud/CubeSandbox/tree/master/examples/opencode-integration)
+project.
+
+## Integration Target and Version
+
+| Component | Version |
+|---|---|
+| OpenCode CLI | `opencode-ai` pinned by `OPENCODE_VERSION` build arg |
+| Node.js | 22, installed through NodeSource |
+| CubeSandbox base image | `ghcr.io/tencentcloud/cubesandbox-base:2026.16` |
+| Host SDKs | `e2b>=2.4.1`, `cubesandbox>=0.3.0` |
+| CubeSandbox platform | `>=0.3.0` for pause/resume; CubeEgress required for header injection |
+
+OpenCode's documented non-interactive mode is `opencode run [message..]`. The
+example uses:
+
+```bash
+opencode run --model <provider/model> --dir /workspace --auto --format json "<prompt>"
+```
+
+## Why Run OpenCode Inside CubeSandbox
+
+OpenCode can edit files, execute commands, install packages, and call LLM APIs.
+Running it in CubeSandbox gives each coding task a disposable MicroVM boundary:
+
+| Concern | CubeSandbox pattern |
+|---|---|
+| Isolation | One MicroVM per task or session |
+| Reproducibility | Boot from a pinned template image |
+| Fast reuse | Pause/resume preserves `/workspace` and OpenCode state |
+| Network control | Default-deny egress with explicit LLM host allow rules |
+| Secret handling | CubeEgress can inject API keys on the wire |
+| Reviewability | Host scripts verify generated files and test results |
+
+## Prerequisites
+
+- A running CubeSandbox deployment; CubeAPI reachable at `http://<node>:3000`.
+- `cubemastercli` on `$PATH`.
+- Docker plus a registry reachable from Cube nodes.
+- Python 3.10+ for host-side driver scripts.
+- An LLM provider API key supported by OpenCode.
+
+## Integration Steps
+
+### 1. Build the template image
+
+```bash
+cd examples/opencode-integration
+docker build --platform linux/amd64 \
+  -t <registry>/opencode-cube:latest .
+docker push <registry>/opencode-cube:latest
+```
+
+The image installs Node.js, OpenCode, `git`, `ripgrep`, Python, and small
+inspection tools on top of the CubeSandbox base image.
+
+### 2. Register the Cube template
+
+```bash
+cubemastercli tpl create-from-image \
+  --image <registry>/opencode-cube:latest \
+  --writable-layer-size 4G \
+  --expose-port 49983 \
+  --probe 49983 \
+  --probe-path /health
+```
+
+The inherited CubeSandbox base entrypoint keeps envd on port `49983` and serves
+the standard `/health` endpoint.
+
+### 3. Configure the host driver
+
+```bash
+cp .env.example .env
+pip install -r requirements.txt
+```
+
+Set:
+
+```bash
+E2B_API_URL=http://<node>:3000
+E2B_API_KEY=e2b_000000
+CUBE_TEMPLATE_ID=<template-id>
+OPENCODE_MODEL=openai/gpt-4.1-mini
+OPENAI_API_KEY=<provider-key>
+```
+
+For custom endpoints, set `OPENCODE_BASE_URL`; the scripts write an
+`opencode.json` in the sandbox workspace with `provider.<id>.options.baseURL`.
+
+When validating against the repository's `dev-env/` VM, set
+`CUBE_DEV_SIDECAR=1` so the host-side SDK routes sandbox traffic through
+`examples/e2b-dev-sidecar`.
+
+`network_policy.py` uses the native `cubesandbox` SDK. When running that script
+against `dev-env/`, also set:
+
+```bash
+CUBE_API_URL=http://127.0.0.1:13000
+CUBE_PROXY_NODE_IP=127.0.0.1
+CUBE_PROXY_PORT_HTTP=11080
+```
+
+### 4. Run the one-shot coding task
+
+```bash
+python run_opencode.py --dry-run
+python run_opencode.py
+```
+
+The script seeds a small Python project, runs OpenCode, then verifies:
+
+- `calculator.add` was implemented.
+- `python3 -m unittest discover -v` passes.
+- `result.md` contains the marker `OPENCODE_CUBE_OK`.
+
+### 5. Demonstrate session persistence
+
+```bash
+python resume_opencode.py
+```
+
+The first turn writes `plan.md`, pauses the sandbox, reconnects to the same
+sandbox, verifies `/workspace` and OpenCode's state directory survived, then
+continues the task and checks `OPENCODE_RESUME_OK`.
+
+### 6. Restrict egress and inject credentials
+
+```bash
+python network_policy.py
+```
+
+The strict path uses the native `cubesandbox` SDK:
+
+- `allow_internet_access=False` makes egress default-deny.
+- A CubeEgress rule allows only the configured LLM host.
+- `Inject` attaches the real provider credential as an HTTP header.
+- The sandbox process receives only a placeholder key.
+
+Use `--skip-agent` to verify policy behavior without spending LLM tokens.
+
+## Key Code Snippets
+
+Headless OpenCode command construction:
+
+```python
+opencode_command(
+    prompt,
+    workspace="/workspace",
+    model="openai/gpt-4.1-mini",
+    title="cube-opencode-demo",
+)
+```
+
+Restricted egress rule:
+
+```python
+Rule(
+    name="allow_openai_llm",
+    match=Match(scheme="https", sni="api.openai.com", host="api.openai.com"),
+    action=Action(
+        allow=True,
+        audit="metadata",
+        inject=[Inject(
+            header="Authorization",
+            secret=secret,
+            format="Bearer ${SECRET}",
+        )],
+    ),
+)
+```
+
+## Caveats
+
+- OpenCode still needs a real LLM provider key for live runs.
+- `--auto` lets OpenCode perform edits and commands without interactive approval;
+  use it only inside the isolated sandbox boundary.
+- Direct env injection in `run_opencode.py` is convenient for local validation,
+  but shared clusters should prefer `network_policy.py`.
+- Custom provider support depends on OpenCode's provider configuration and the
+  target endpoint's compatibility with the selected model.
+
+## Validation
+
+From `examples/opencode-integration`:
+
+```bash
+python3 -m py_compile env_utils.py _opencode_common.py run_opencode.py resume_opencode.py network_policy.py
+python3 -m unittest discover . -p 'test_*.py'
+bash -n build-template.sh
+docker build --platform linux/amd64 -t opencode-cube:verify .
+docker run --rm opencode-cube:verify opencode --version
+```
+
+Live CubeSandbox validation:
+
+```bash
+python run_opencode.py
+python resume_opencode.py
+python network_policy.py --skip-agent
+python network_policy.py
+```
+
+## References
+
+- [OpenCode CLI documentation](https://opencode.ai/docs/cli/)
+- [OpenCode provider documentation](https://opencode.ai/docs/providers/)
+- [CubeSandbox OpenCode example](https://github.com/TencentCloud/CubeSandbox/tree/master/examples/opencode-integration)
+- [CubeSandbox Security Proxy](../security-proxy.md)

docs/zh/guide/integrations/index.md

@@ -47,4 +47,5 @@ lang: zh-CN
 
 | 标题 | 作者 | 日期 | 标签 |
 | --- | --- | --- | --- |
+| [OpenCode 集成指南](./opencode.md) | Sam-Hui-dot | 2026-07-05 | integration, opencode, coding-agent, agent |
 | [Pi Agent 集成指南](./pi-agent.md) | chaojixinren | 2026-07-01 | integration, pi-agent, coding-agent, agent |