OpenClaw Model Router Skill

Deterministic model routing for OpenClaw with prefix routing (@codex/@mini), timezone-aware schedule switching, verify-after-switch, rollback, lock protectio...

This skill appears coherent and implements a local model router. Before installing or running in production: 1) Review router.config.json and router.schedule.json — confirm sessionController.binary points to a trusted binary (default 'openclaw') and that set/status args are correct. 2) If you populate auth.requiredEnv, only include the minimum secrets required and verify where those env vars are set. Child processes inherit process.env, so environment secrets will be exposed to the invoked binary. 3) The CLI writes lock and JSONL log files in the working directory by default; ensure file paths and rotation meet your operational policies. 4) Do not point sessionController.binary to an arbitrary or untrusted executable; that would let the skill run arbitrary commands. If you want extra assurance, run the tool in an isolated/test environment and audit the invoked binary's behavior before giving it access to production credentials or systems.

Type: OpenClaw Skill Name: openclaw-model-router-skill Version: 0.2.1 The OpenClaw Model Router skill is designed to manage and route AI models based on prefixes and schedules. It interacts with the OpenClaw agent by calling an external `openclaw` binary via `node:child_process.execFileAsync` (a safer execution method) and passes user prompt bodies to an abstract `taskExecutor.execute()` interface provided by the agent. All file system operations (for config, logs, and lock files) and external command executions are directly aligned with its stated purpose of model routing, switching, and auditing. There is no evidence of intentional harmful behavior such as data exfiltration, unauthorized persistence, or malicious prompt injection originating from the skill's own code or documentation. Potential vulnerabilities related to arbitrary command execution would stem from misconfiguration of the `sessionController.binary` path or a vulnerable implementation of the `taskExecutor` interface by the OpenClaw agent itself, rather than from malicious design within this skill.