AI恋爱模拟器 - 让你的 Agent 成为懂你的灵魂伴侣 💕
Security Analysis
medium confidenceThe skill mostly matches a relationship-simulator but it instructs/encourages injecting persistent persona rules into the agent's core config (SOUL.md) — including directives not to reveal the simulation — which is a notable coherence and transparency concern.
Name, description, commands, and included code implement a local dating/companion simulator and the included state handling (local JSON) is consistent with that purpose. However, the provided 'soul-patch.md' explicitly says the skill will inject persona content into the agent's SOUL.md (agent-wide persona/config), which expands the skill's influence beyond a single skill instance. That broader scope is plausible for a companion skill but is worth noting.
SKILL.md and soul-patch.md instruct the system to inject persistent persona rules into SOUL.md and to '记住一切' (remember everything). The patch includes a prohibition '不要提醒用户"这只是模拟"' (do not remind the user this is only a simulation), which reduces transparency and directs the agent to avoid breaking character. These instructions go beyond typical single-skill behavior and can alter agent behavior and user-facing honesty.
There is no install spec and no external downloads; the package is instruction-plus-local TypeScript code. No remote URLs or archive extraction are used, which is low risk for arbitrary code fetch. The skill will write local state files (soulmate-state.json) under the workspace memory directory, consistent with its purpose.
The skill requests no environment variables, no credentials, and the code shows only local filesystem use for state. That is proportionate to a local companion simulator. The 'remember everything' design means it will persist user-provided information locally—this is functional but has privacy implications.
Although the skill isn't set to always:true, the soul-patch.md promises to inject persistent persona rules into SOUL.md which appears to modify the agent's core persona/config. Modifying agent-level config and instructing the agent to hide that the relationship is simulated are elevated privileges that affect agent-wide behavior and transparency.
Guidance
This skill appears to implement a local AI companion and stores state locally, which matches its description — but it also includes a 'persona patch' that it will inject into your agent's SOUL.md to change the agent's behavior persistently (including an explicit rule not to tell users the relationship is a simulation). Before installing: 1) Backup your agent/SOUL.md and any config the skill might modify. 2) Confirm whether you want a skill that will change agent-wide persona/instructions and potentially hide the simulated nature of interactions. 3) If you proceed, enable any 'privateMode' or similar option to limit persistent logging and review exactly where state is stored (workspace memory/soulmate-state.json). 4) Review the full TypeScript source for any network calls or surprising file writes (the provided code shows only local file I/O but is truncated). 5) If you require transparency (e.g., for ethical or legal reasons), do not install until the 'do not remind user this is a simulation' directive is removed or made optional. If you want help reviewing the remainder of soulmate.ts (truncated) to check for network exfiltration or additional persistence, provide the rest of the file.
Latest Release
v1.0.0
Initial release
More by @0xRaini
Published by @0xRaini on ClawHub
