# LLM Task (plugin)

Adds an **optional** agent tool `llm-task` for running **JSON-only** LLM tasks
(drafting, summarizing, classifying) with optional JSON Schema validation.

Designed to be called from workflow engines (for example, Lobster via
`openclaw.invoke --each`) without adding new OpenClaw code per workflow.

## Enable

1. Enable the plugin:

```json
{
  "plugins": {
    "entries": {
      "llm-task": { "enabled": true }
    }
  }
}
```

2. Allowlist the tool (it is registered with `optional: true`):

```json
{
  "agents": {
    "list": [
      {
        "id": "main",
        "tools": { "allow": ["llm-task"] }
      }
    ]
  }
}
```

## Config (optional)

```json
{
  "plugins": {
    "entries": {
      "llm-task": {
        "enabled": true,
        "llm": {
          "allowModelOverride": true,
          "allowedCompletionModels": ["openai/gpt-5.6-sol"],
          "allowAuthProfileOverride": true
        },
        "config": {
          "defaultProvider": "openai",
          "defaultModel": "gpt-5.6-sol",
          "defaultAuthProfileId": "main",
          "maxTokens": 800,
          "timeoutMs": 30000
        }
      }
    }
  }
}
```

The host-owned `llm` policy authorizes model/profile overrides. Its
`allowedCompletionModels` restricts every completion, including the resolved agent
default. Run `openclaw doctor --fix` once for entries created by older releases;
Doctor grants the shipped selection permissions and moves legacy
`config.allowedModels` values into `llm.allowedCompletionModels` without widening them.

## Tool API

### Parameters

- `prompt` (string, required)
- `input` (any, optional)
- `schema` (object, optional JSON Schema)
- `provider` (string, optional)
- `model` (string, optional)
- `thinking` (string, optional)
- `authProfileId` (string, optional)
- `temperature` (number, optional)
- `maxTokens` (number, optional)
- `timeoutMs` (number, optional)

### Output

Returns `details.json` containing the parsed JSON (and validates against
`schema` when provided).

## Notes

- The tool is **JSON-only** and instructs the model to output only JSON
  (no code fences, no commentary).
- No tools are exposed to the model for this run.
- Side effects should be handled outside this tool (for example, approvals in
  Lobster) before calling tools that send messages/emails.

## Bundled extension note

This extension depends on OpenClaw internal modules (the embedded agent runner).
It is intended to ship as a **bundled** OpenClaw extension (like `lobster`) and
be enabled via `plugins.entries` + tool allowlists.

It is **not** currently designed to be copied into
`~/.openclaw/extensions` as a standalone plugin directory.