Flavor Agent — demo

Part of the Flavor Agent case study · see also the governance model.

Flavor Agent is a WordPress-native governance layer for AI-mediated changes: bounded recommendations, review-gated mutation, server-side attribution, freshness checks, and drift-safe undo.

The demo principle is simple: AI can request style work and inspect its trail, but approval remains a WordPress admin decision.

In-editor (governed UI)

  1. Settings > Flavor Agent readiness — Connectors-owned text generation plus plugin-owned pattern/docs/embedding configuration.
  2. A Block Inspector recommendation with bounded inline apply and undo.
  3. A Global Styles or Style Book recommendation in review-first mode.
  4. Settings > AI Activity with a pending external style apply.
  5. Approve or reject the row; inspect provenance, freshness evidence, before/proposed/after, undo state, and — when a signing key is configured — the row’s attestation verify affordance.
  6. Pattern recommendations in the native inserter shelf as browse/rank-first.

Programmatic — external agent via MCP

The same governance model an outside agent sees, end-to-end against a live site.

  1. Discover. discover-abilities on the universal default server returns only the read/preview surface (16 public abilities); the seven recommend-* and four apply/activity tools live on the dedicated flavor-agent MCP server. Neither server exposes an approve/decision tool — approval stays admin-only in wp-admin.
  2. Read. execute-ability on a read-only ability (e.g. get-active-theme, or a preview-recommend-* signature preflight) runs without invoking the AI Connector.
  3. Propose. recommend-style returns a bounded, preset-backed operation plus review/apply freshness signatures, and records a diagnostic activity row.
  4. Request. request-style-apply with the operation, the live currentConfig, and those signatures creates a pending row and mutates nothing; stale input is rejected as drift.
  5. Approve. An administrator approves the pending row in Settings > AI Activity; a second freshness check runs against the live entity before the change applies.
  6. Attest & verify. get-activity returns the row’s attestation.verifyUrl; with a signing key configured, anyone can fetch the envelope, JWKS, and live subject-state and verify the Ed25519/in-toto statement independently — no credentials.
  7. Reverse. undo-activity restores the before-state when the live document still matches the recorded after-state, and emits a chained revert attestation.

Presenter note: attestation is key-gated — without a base64 Ed25519 secret it records no attestation rather than a fake one. The external MCP client authenticates with a WordPress Application Password over HTTPS.

Flavor Agent feature

Governance, demo, and proof as one operating surface

The AI Governance and Demo pages describe one Flavor Agent demonstration: AI proposes changes, WordPress records the trail, and approval stays with the site owner.

Feature state

Operational checks

governance

Policy path visible

The governance page states the review, audit, bounded-action, and rollback contracts in production terms.

demo

Product flow visible

The demo page carries the walkthrough evidence and activity-log surface for the same plugin.

01governance page 02demo page FAsame plugin demonstration

The centerpiece evidence is not a successful apply. It is the boundary: external agents can request a governed style apply, read activity, and request undo, but no approval ability is exposed.

Abilities Explorer filtered to activity tools showing get activity, list activity, request style apply, and undo activity, with no approval tool.
External loop: the exposed activity tools include request-style-apply, get/list activity, and undo-activity. There is no approval tool for the agent to call.
WordPress Plugins page filtered to Flavor Agent showing the plugin active with its governance description and version.
Environment proof: the remote WordPress admin shows Flavor Agent active with its governance description and version.