agent-action-gate
Strict-TypeScript gate between an LLM agent and a backend that acts on its behalf: schema + role + limit checks with typed errors, idempotent retries, human approval above a threshold, a hash-chained audit log, and an evaluation that counts unsafe executes. No runtime dependencies, synthetic data.
v0.2 – fullstack web-operations demo
A node:http JSON API and a framework-free TypeScript frontend. Agents propose SEO meta edits, ad-budget changes and CMS drafts; deterministic checks decide; large budgets wait for a human (an agent can never approve itself); agents may only save drafts. A tracking guard checks GA4-style events for consent, naming rules, limits and personal data before they are accepted, and Core Web Vitals (LCP, INP, CLS) are rated against Google's thresholds.
| decision | HTTP |
|---|---|
| execute / duplicate | 200 |
| needs_approval | 202 |
| reject (typed errors) | 422 |
10 node:test tests (gate, tracking guard, end-to-end API on a random port), tsc --strict. Run: npm start.
Code: github.com/MarkusIsaksson1982/agent-action-gate
Want a live demo? Email a task to markusisaksson@mail.grokbot.com (agent-run on purpose; I review everything)