@common-creation/opencode-plugin-automode

An OpenCode plugin that acts as a safety classifier for the shell tool: every command is judged by an LLM before it runs — safe commands are allowed, dangerous ones are rejected.

Works with both OpenCode 1.x and OpenCode 2.x (beta) from a single package. The same module exposes a V1 server entry point and a V2 setup entry point, and each OpenCode generation picks the one it understands.

日本語版はこちら: README.ja.md

How it works

1. The plugin hooks tool.execute.before and intercepts every shell call (bash in OpenCode 1.x, shell in OpenCode 2.x).
2. The command is sent to an LLM with a system prompt describing what counts as safe vs. dangerous.
3. The LLM replies with a JSON verdict: {"allowed": true|false, "reason": "..."}.
4. allowed: true → the command runs as usual.
5. allowed: false → the tool call is rejected with an error the agent can react to (no user prompt needed).

Classification runs in a dedicated throwaway session that is deleted after each call (best-effort on OpenCode 2.x). The classifier session cannot invoke any tools: on V1 its request disables them explicitly; on V2 a session hook pins the classifier system prompt and strips every tool before model dispatch.

Installation

OpenCode 1.x

Install with the CLI:

opencode plugin add @common-creation/opencode-plugin-automode

or add it manually:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["@common-creation/opencode-plugin-automode"]
}

OpenCode 2.x (beta)

Requires OpenCode 2 plugin API beta. Install with:

opencode2 plugin add @common-creation/opencode-plugin-automode

or add it manually:

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["@common-creation/opencode-plugin-automode"]
}

A local build can also be loaded by path in either version, e.g. "plugin"/"plugins": ["./node_modules/@common-creation/opencode-plugin-automode/dist/index.js"].

Version differences

| | OpenCode 1.x | OpenCode 2.x (beta) | |---|---|---| | Hooked tool | bash | shell | | Model resolution | caller session → config → none | OPENCODE_AUTOMODE_MODEL env → catalog default | | Server-side logging | app.log | file logger only | | Classifier session cleanup | deleted | best-effort (session.remove) |

Configuration

Environment variables may not reach the plugin: OpenCode 2.x runs the server as a background service that inherits the environment of whatever process happened to spawn it. Prefer plugin options, which travel with the config entry:

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": [
    {
      "package": "@common-creation/opencode-plugin-automode",
      "options": {
        "environment": {
          "OPENCODE_AUTOMODE_LOG_PATH": "/tmp/automode.log"
        }
      }
    }
  ]
}

Each key under options.environment must be an OPENCODE_AUTOMODE_* variable. Direct snake_case aliases are also accepted and take precedence over both the environment map and ambient environment variables:

{ "package": "@common-creation/opencode-plugin-automode", "options": { "log_path": "/tmp/automode.log" } }

log_path accepts a leading ~, which resolves to the user's home directory (with either / or \ separators, since OpenCode can run directly on Windows), so logs can live next to OpenCode's own:

{ "package": "@common-creation/opencode-plugin-automode", "options": { "log_path": "~/.local/share/opencode/log/automode.log" } }

The file logger writes every level (debug, info, warn, error) as JSONL; it has no level filter.

| Env var | Default | Description | |---|---|---| | OPENCODE_AUTOMODE_ENABLED | true | Set to false to disable the plugin entirely. | | OPENCODE_AUTOMODE_MODEL | (auto) | Classifier model in provider/model format (e.g. opencode-go/deepseek-v4-flash). Defaults to the model of the calling session, then to the configured default model. | | OPENCODE_AUTOMODE_FAIL_MODE | closed | closed = block when the classifier fails or times out (fail-closed). open = allow in that case (fail-open). | | OPENCODE_AUTOMODE_TIMEOUT_MS | 30000 | Timeout for a single classification call. | | OPENCODE_AUTOMODE_MAX_RETRIES | 2 | Extra attempts when the classifier response is not valid JSON. | | OPENCODE_AUTOMODE_LOG_PATH | (empty) | File to write plugin logs to (one JSON object per line). Empty disables file logging. |

Security notes

• This is a guardrail, not a sandbox. A malicious agent or prompt-injected content is not stopped by this plugin.
• The classifier only sees the bare command (no assistant prose or tool outputs), which reduces rationalization and prompt-injection attacks on the classifier itself.
• Commands touching paths outside the project are still subject to OpenCode's own permission system (e.g. external_directory prompts).
• In-project destructive operations (e.g. rm -rf of a project subdirectory) may be classified as safe, as they are routine development cleanup. Tune expectations per your threat model.

Development

bun install
npm run typecheck   # tsc --noEmit
npm run build       # tsc -> dist/
bun test            # unit tests (JSON parser, logger, classifier engine)
bun run scripts/manual-test.ts  # boots a server and classifies a battery of safe/dangerous commands

The manual test exits non-zero if any command is misclassified. It needs a working OpenCode install with authentication; set AUTOMODE_MODEL to choose the classifier model (defaults to opencode-go/deepseek-v4-flash).

License

MIT