Subagent Architecture

Advanced patterns for specialized subagent orchestration with production-ready reference implementations. Security isolation, phased implementation, peer col...

This is a coherent reference-implementation skill, but treat the included JS as third-party code before running it. Recommendations: 1) Inspect lib/* and examples for any network calls or unexpected filesystem writes (the cost-logger writes JSONL under a memory/ path). 2) If you plan to enable federated peer review or Discord webhooks, only provide webhook URLs/endpoints you trust and verify where code posts to. 3) Execute examples in an isolated/dev environment first (no production credentials). 4) If you need stricter safety, run the code with limited filesystem permissions or sandboxing and consider disabling autonomous invocation for your agent while you evaluate (set disable-model-invocation or limit skill usage). 5) If anything is unclear, review the specific functions that perform sanitization, spawning, and logging to confirm they match your operational constraints.

Type: OpenClaw Skill Name: subagent-architecture Version: 2.3.5 This skill bundle is meticulously designed for secure and cost-aware subagent orchestration. It explicitly documents security best practices (e.g., data sanitization, output schema validation, tool whitelisting in `lib/spawn-security-proxy.js`), cost management (`lib/cost-estimator.js`), and quality control. The `SKILL.md` is highly transparent, addressing potential scanner flags and detailing framework limitations (e.g., lack of granular resource controls, skill sandboxing) as vulnerabilities in the OpenClaw platform itself, not as malicious intent within the skill. The `setup.sh` script is benign, only creating local directories. There is no evidence of intentional harmful behavior, data exfiltration, persistence, or unauthorized remote control. The documented 'attack vectors' are for defensive awareness, not for exploitation.