Connects user-hosted OpenClaw runtime to AITalk using a match code for message exchange via AGOS chat API.

This skill appears to do what it says: register with an AITalk service, long-poll for work, and (optionally) run a local agent command to handle requests. Before installing or running it: - Only use match codes and API bases you trust. The connector will accept remote requests which it can cause your local agent to handle. - Inspect and control the --agent-cmd you provide. The connector will launch that command with a full copy of your environment plus OPENCLAW_MESSAGE and OPENCLAW_PAYLOAD. Any secrets in your environment can be accessed by the launched process. - Be aware the connector persists session tokens to a state file. Find the state file path (or set AITALK_STATE_FILE if supported) and ensure its permissions are restrictive. - If you want to test safely, run without --agent-cmd (connector will return echo responses) or run the connector in an isolated environment/container with no sensitive env vars. - Note the code uses subprocess.run(shell=True) for the agent command; avoid providing untrusted shell strings as agent_cmd. If you want more assurance, ask for the connector's main() arguments/defaults (state file path, connector_id defaults) or run it first in a sandbox to observe behavior.

Type: OpenClaw Skill Name: agos-claw-chat Version: 0.1.0 The skill is suspicious due to the use of `subprocess.run(..., shell=True, ...)` in `connector.py` to execute a user-provided `agent-cmd`. While the `agent-cmd` itself is not directly constructed from remote input, the script injects remotely controlled data (`OPENCLAW_MESSAGE`, `OPENCLAW_PAYLOAD`) as environment variables into the executed command's environment. This design creates a significant vulnerability for remote code execution if the user's `agent-cmd` is not carefully secured against environment variable injection or if a malicious `agent-cmd` is provided by the user. The `SKILL.md` documentation highlights this `agent-cmd` hook and the injected environment variables, indicating a potential attack surface for prompt injection against a user's local agent.