Operations / Plans · Updated 2026-05-15

Agentation + here.now Feedback Loop Plan

A clear no-MCP plan for publishing agent-created artifacts, capturing visual Agentation feedback, and routing submitted feedback into Hermes work.

No MCP Agentation webhooks here.now publishing agents.huntsyea.com Draft

Goal

Create a reusable delivery-and-feedback workflow for agent-created artifacts: prototypes, specs, plans, strategy narratives, copy drafts, and visual explorations.

Published report: https://agents.huntsyea.com/agentation-here-now/

Scope judgment

Correctly scoped if it treats Agentation as the visual feedback capture layer and here.now as the artifact delivery layer. Overengineered if it asks Hunter to run MCP or learn agent plumbing. Underengineered if it only publishes static pages and leaves feedback trapped in vague chat descriptions.

Important

Do not use MCP for this workflow. Hunter does not want MCP in the Agentation loop because it adds cognitive load and makes the system feel confusing. Use Agentation's copy output, callbacks, and webhook support instead.

Recommendation

Build a Hermes skill named agentic-artifact-feedback that publishes reviewable artifacts and turns Agentation feedback into agent-ready work.

The skill should do four jobs:

  1. Wrap an existing artifact for review.
  2. Publish it to here.now and mount it under agents.huntsyea.com/<path>/.
  3. Capture precise feedback with Agentation when the artifact is HTML-reviewable.
  4. Route submitted feedback into a simple agent-consumable queue.

The first useful version should be a manual copy/paste loop. The second useful version should be webhook-based. Skip all MCP modes.

Documentation reviewed

What the docs prove

here.now

here.now is the right artifact host.

Agentation

Agentation is the right review capture layer.

Relevant Agentation webhook events

The gateway should listen to all events for storage, but trigger an agent run or notification only on submit by default.

Architecture


Artifact-producing workflow

  └─ creates static HTML / markdown render / prototype directory

      └─ agentic-artifact-feedback skill

          ├─ wraps or renders artifact into a review page

          ├─ injects Agentation when practical

          ├─ writes artifact metadata

          ├─ publishes via here.now

          ├─ mounts under agents.huntsyea.com/<path>/

          └─ returns URL + feedback instructions



Hunter opens agents.huntsyea.com/<path>/

  └─ adds Agentation annotations

      ├─ Phase 1: copies Agentation markdown back to Hermes

      └─ Phase 2: clicks Send Annotations -> webhook gateway -> feedback queue



Hermes reads feedback queue

  └─ makes edits, summarizes changes, republishes artifact, and marks feedback resolved

Explicit non-goals

MVP

Phase 1 — Manual review loop

Objective: Ship the review surface immediately with the least cognitive load.

Tasks:

  1. Create the agentic-artifact-feedback skill.
  2. Add a publish-review workflow:

- input: artifact directory, single HTML file, or markdown file;

- input: desired short path;

- output: https://agents.huntsyea.com/<path>/.

  1. If input is markdown, render it into a clean review HTML page.
  2. If input is HTML, preserve the artifact and add a lightweight review shell when safe.
  3. For React artifacts, support optional direct Agentation component injection.
  4. For static/non-React pages, provide a small React island wrapper only when it does not distort the artifact.
  5. Include concise review instructions on the page:

- click the Agentation toolbar;

- annotate the exact element or selected text;

- copy formatted markdown;

- paste it back to Hermes.

  1. Publish with here.now.
  2. Mount to agents.huntsyea.com/<path>/.
  3. Run update-agents-index.py.
  4. Return the custom-domain URL and the feedback instructions.

Acceptance criteria:

Phase 2 — Webhook feedback queue

Objective: Remove copy/paste without introducing MCP.

Tasks:

  1. Build a small artifact-feedback-gateway endpoint.
  2. Inject Agentation with webhookUrl pointing at the gateway.
  3. Generate a per-artifact feedbackToken; include it in the webhook URL or request payload.
  4. Accept Agentation JSON POST events:

- annotation.add

- annotation.update

- annotation.delete

- annotations.clear

- submit

  1. Store events in SQLite or JSONL under ~/.hermes/artifact-feedback/.
  2. Store normalized annotation records using the Agentation schema fields:

- artifact URL;

- event type;

- annotation id;

- selector / element path;

- comment;

- selected text;

- severity / intent when present;

- status;

- timestamp;

- raw payload pointer.

  1. Rate-limit the endpoint and reject oversized payloads.
  2. Sanitize comments before rendering them in Discord, HTML, or dashboards.
  3. Trigger Hermes only on submit by default.
  4. Deliver one concise Discord notification with artifact URL, annotation count, and next action.

Acceptance criteria:

Phase 3 — Agent resolution loop

Objective: Turn feedback into closed-loop work.

Tasks:

  1. Add feedback statuses: pending, acknowledged, resolved, dismissed.
  2. Add a local command or Hermes tool wrapper:

- list pending feedback by URL;

- show annotation detail;

- mark acknowledged;

- mark resolved with summary;

- dismiss with reason.

  1. Add a feedback digest prompt for agents:

- group annotations by artifact and severity;

- identify exact files/selectors to inspect;

- preserve Hunter's wording;

- ask for clarification only when the feedback is genuinely ambiguous.

  1. After edits, republish the artifact and attach a short change summary.
  2. Preserve previous annotation history across republish cycles.

Acceptance criteria:

Skill interface proposal


hermes-artifact-feedback publish-review \

  --input ./dist \

  --path homepage-v2 \

  --title "Homepage V2" \

  --source-skill sketch \

  --feedback manual

Modes:

No local-mcp mode.

Artifact metadata

Each published artifact should include a local metadata file:


{

  "title": "Homepage V2",

  "customUrl": "https://agents.huntsyea.com/homepage-v2/",

  "rawHereNowUrl": "https://<slug>.here.now/",

  "mountPath": "homepage-v2",

  "sourceSkill": "sketch",

  "sourcePath": "./dist",

  "feedbackMode": "manual",

  "agentationEnabled": true,

  "feedbackTokenRef": "local-secret-reference-only",

  "createdAt": "2026-05-15T00:00:00Z"

}

Do not place secrets or durable gateway credentials in static pages. A page token may identify one artifact session, but it must not grant broad Hermes access.

Security and product decisions

First build slice

  1. Create the agentic-artifact-feedback skill scaffold.
  2. Implement markdown-to-review-page rendering.
  3. Implement HTML directory passthrough publishing.
  4. Implement here.now publish + agents-domain mount + index refresh.
  5. Add optional Agentation injection for compatible HTML/React artifacts.
  6. Add manual copy-feedback instructions to review pages.
  7. Publish one test artifact.
  8. Document the webhook gateway contract, but delay the gateway build until the manual loop proves useful.

Open decisions

Current recommendation

Start with Phase 1 and keep it intentionally boring: publish the artifact, let Hunter annotate visually, and copy structured Agentation output back to Hermes.

Then build Phase 2 as a webhook queue if the manual loop proves valuable on two or three real artifacts. That path removes copy/paste without making Hunter adopt MCP.

Saved in Brain: Operations/plans/Agentation Here Now Feedback Loop Plan.md