ClawHub

Feedback Controller

v1.0.0

Feedback Controller — Closed-Loop Agent Skill for Correcting Execution Drift. Use it when the user needs a disciplined protocol and fixed output contract for...

⭐ 0· 45·0 current·0 all-time
byCubic AI@clarkchenkai
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
✓
Purpose & Capability
Name, description, and provided materials (protocol, patterns, escalation rules) match the stated purpose. No unrelated env vars, binaries, or install steps are requested.
ℹ
Instruction Scope
SKILL.md confines the agent to a five‑step correction protocol and a fixed six‑part output contract. It does grant the agent discretion to 'retry, switch tools, narrow scope, rewrite, or escalate' and to inspect produced outputs and execution traces — reasonable for a correction controller but broad in practice. If you need to limit automatic tool switching or retries, enforce those constraints at the agent/tool-permission level.
✓
Install Mechanism
Instruction-only skill with no install spec and no code to write to disk; lowest-risk install footprint.
✓
Credentials
No environment variables, credentials, or config paths are required or referenced. The skill does not request secrets or external service tokens.
ℹ
Persistence & Privilege
Registry flags show normal autonomous invocation allowed (disable-model-invocation: false). agents/openai.yaml sets allow_implicit_invocation: true, which may let the platform call this skill automatically in some contexts — not inherently dangerous, but if you want tighter control, consider disabling implicit invocation or restricting when the agent can apply correction actions.
Assessment
This skill is internally consistent and does what it says: a disciplined protocol for diagnosing and correcting drift. It does not ask for credentials or install code. The main operational concern is scope: the protocol explicitly allows retries and tool switching, so review and constrain the agent's tool permissions and implicit-invocation settings if you want to prevent automated corrective actions in high‑risk workflows. If you require human approval for escalations or changes with material consequences, enforce that in the agent policy or by disabling implicit invocation for this skill.

Like a lobster shell, security has layers — review code before you run it.

latestvk974kst9zxb6xmf6bfb2p6wbbx84hn3t

License

MIT-0
Free to use, modify, and redistribute. No attribution required.
Termshttps://spdx.org/licenses/MIT-0.html

Comments