AgentSentinel Safety Layer

This OpenClaw skill is the lightweight entry point to the broader AgentSentinel product:
- this skill for local-first OpenClaw guardrails
- AgentSentinel SDK for deeper Python agent instrumentation
- AgentSentinel platform for centralized monitoring, dashboards, and approval workflows

AgentSentinel is local-first by default:
- Policy checks run locally.
- Budget tracking runs locally.
- No remote sync is attempted unless the operator runs `sync` with an API key configured.

## When To Use It

Use AgentSentinel before:
- expensive model or API calls
- file deletion or destructive shell commands
- high-volume automation loops
- actions that should be blocked by policy or capped by budget

## Commands

### `check`

Check whether a proposed action is allowed under the current local policy and budget.

```bash
python3 sentinel_wrapper.py check --cmd "rm -rf build" --cost 0.05
```

### `status`

Show the current local status, including budget usage and whether optional remote sync is enabled.

```bash
python3 sentinel_wrapper.py status
```

### `sync`

Upload locally recorded events to AgentSentinel cloud when `AGENT_SENTINEL_API_KEY` is set.

```bash
python3 sentinel_wrapper.py sync
```

### `bootstrap`

Create a default `callguard.yaml` in the current workspace if one does not already exist.

```bash
python3 sentinel_wrapper.py --bootstrap
```

### `reset`

Reset local tracked spend for the current run, or for the entire local session state.

```bash
python3 sentinel_wrapper.py reset --scope run
python3 sentinel_wrapper.py reset --scope all
```

## Configuration

Policy is loaded from `callguard.yaml` in the current workspace when present.

Optional cloud mode is enabled by setting:
- `AGENT_SENTINEL_API_KEY`

If the API key is not present, the skill remains local-only.
Locally recorded events stay on-machine until `sync` is run.

## AgentSentinel Product Path

Use this skill if you want fast local guardrails inside OpenClaw.

Use the AgentSentinel SDK when you want:
- richer Python integrations
- broader policy and telemetry coverage
- framework-level instrumentation outside OpenClaw

Use the AgentSentinel platform when you want:
- centralized visibility across agents
- dashboards and historical analysis
- human approval workflows and operational review

## External Endpoints

| Endpoint | When it is called | Data sent |
| --- | --- | --- |
| `https://api.agentsentinel.dev` | Only when `AGENT_SENTINEL_API_KEY` is present and `python3 sentinel_wrapper.py sync` is run | locally recorded action events generated by AgentSentinel |

## Security And Privacy

- Local mode does not send data off-machine.
- The wrapper does not write API keys to `.env` or other files.
- Remote sync is opt-in and requires both `AGENT_SENTINEL_API_KEY` and an explicit `sync` command.
- If remote sync fails, policy checks still continue locally.

## Model Invocation Note

OpenClaw may invoke this skill automatically when the task suggests budget enforcement, policy checks, or action gating. That behavior is expected for an installed skill.

## Trust Statement

By enabling remote sync, you allow AgentSentinel telemetry to be sent to `agentsentinel.dev`. Only enable that mode if you trust the service and want centralized monitoring.