Constrained autonomous delegate for identity-proxied communication. Handles incoming messages (Email, Discord, Slack, WhatsApp) by analyzing intent and applying declarative authority policies before generating responses. Use when the user wants to delegate communication tasks while maintaining identity fidelity and enforcing strict non-overreach boundaries.

What to consider before installing: - Trust and provenance: The skill's source/homepage are unknown and author contact looks local/test. Treat it as third-party experimental code and prefer testing in an isolated environment before using with real accounts. - Local file writes: The skill reads/writes under ~/.openclaw/adg (profile, policies, memory, audit). Expect creation of training_samples.jsonl and style.yaml by reanchor_style.py. If you don't want files created at that path, run it in a sandbox or adjust the base path. - Missing pieces / oversell: The code is a scaffold — there are no channel adapters included (no actual Email/Discord/Slack/WhatsApp connectors) and the referenced policy DSL file is not present. Policy enforcement depends on files you must supply; do not assume safe defaults. - Review policies first: Before enabling autonomous replies, create and inspect the policy and contacts YAMLs that live in ~/.openclaw/adg/policies and ~/.openclaw/adg/profile. A misconfigured allowlist could enable undesired autonomous replies. - Dependencies & runtime: The scripts use PyYAML (yaml). Run in a controlled Python environment (venv) and review imports. The manifest expects a long-running Python service — ensure the runtime's isolation and permission controls are enforced by your platform. - Watermark & audit: The tool adds a subtle watermark marker (signature '⛧') and writes audit logs. If that behavior is undesirable, plan to modify or disable watermarking and confirm audit storage is acceptable. - Safe testing: Start with draft-only/default mode and non-sensitive channels/accounts. Confirm policy gates behave as you expect. If you plan to enable persistent/autonomous behavior later, re-audit the code and configuration and ensure network egress remains disabled. Because the package is largely coherent but incomplete/oversold and makes persistent changes on the filesystem, exercise caution (test in a sandbox, inspect or author the policy files, and confirm the runtime isolation) before granting it access to real messaging channels or sensitive accounts.

Type: OpenClaw Skill Name: agent-doppelganger Version: 1.0.0 The OpenClaw AgentSkills skill bundle 'agent-doppelganger' is classified as benign. Both the code (scripts/adg.py, scripts/reanchor_style.py) and documentation (SKILL.md, references/specification.md) consistently emphasize a strong safety-first design, with explicit policy gates, confidence checks, and verification steps occurring before any response generation or sending. The manifest.yaml strictly enforces permissions, denying outbound network access and shell execution, and limiting filesystem access to the skill's designated configuration and audit directories (~/.openclaw/adg/). There is no evidence of data exfiltration, malicious execution, persistence mechanisms, or prompt injection attempts against the agent itself; instead, the design actively mitigates such risks.