# opencode-inside

The [opencode](https://opencode.ai) coding agent running inside a capsule, on
DeepSeek V4.1 Flash through OpenRouter. Builds on
[openrouter-secrets](../openrouter-secrets): same network rule, same key handling.

## What it shows

- **A Dockerfile** instead of `runtime.image`. m64 builds it on the first serve.
  Everything is installed at build time and pinned: `FROM` by digest, `npm install
  -g opencode-ai@1.18.23`, every pip package with `==`. At run time the container
  has no network except what you grant.
- **Warming at build time** (`warm_opencode.sh`): opencode installs npm packages on
  its first start. The build does that once, where it has network, so a run
  never tries to reach the npm registry.
- **The `agent.session` kind** (`agent_gateway.py`): a small session API that m64's
  front lets through. Create a session, send it a task, read its files. See
  `AGENT.md` for every route.
- **`[runtime.app]`**: opencode's own server (web UI and API) on a second port,
  reported as `app_url`. `/human/` links to it, and `opencode attach <app_url>`
  drives the agent from your own terminal.
- **The key in the gate**: `opencode.json` points opencode at
  `http://openrouter.ai` with `apiKey: "{env:OPENROUTER_API_KEY}"`. In the
  container that variable is only a placeholder; the gate puts the real key in.

## Run it

```sh
m64 config secrets set OPENROUTER_API_KEY      # once
m64 check examples/opencode-inside
m64 serve examples/opencode-inside --allow-external-network
m64 status                                      # url and app_url
```

The first serve builds the image: a few minutes. `capsule.toml` targets
`linux-x86_64`; the Dockerfile builds the same on ARM (`linux-aarch64`).

## What to look at

- `/human/`: links to the full opencode web UI, and shows the `opencode attach`
  line for your own terminal.
- `/agent`: the session API, with copy-paste `curl` lines.
- `m64 net <id>`: the agent's connections. Only `openrouter.ai:443` is allowed;
  anything else it tries is refused and logged.
- Sessions share one container and one user: serve one capsule per person or
  project, not per session.