A tool that monitors public sentiment around technical products over time, correlating feedback with release history on a scrubbable timeline.

1. You provide: a product name, short description, and known entry-point URLs (e.g. a changelog page, Twitter handle, product site).
2. Ingestion Agent (Nimbly-powered): scrapes public releases and feedback from Twitter/X, YouTube, tech review sites, and blogs. Each piece of feedback is scored 1–10 (1 = very negative, 10 = very positive).
3. Analysis Agent: aggregates feedback into an Impression Summary, detects significant drops in sentiment, and publishes a summary post to cited.md via Senso.
4. Timeline UI: a scrubbable view showing release events and impression score over time side by side.

User Input
    │
    ▼
Ingestion Agent (Nimbly)
    ├─ Scrapes: changelogs, Twitter/X, YouTube, review sites
    ├─ Writes: releases, release_feedback (with 1–10 score)
    └─ Logs: agent_passes

Analysis Agent
    ├─ Reads: release_feedback since last pass
    ├─ Produces: impression summary (positive/negative themes)
    ├─ Detects: significant score drops
    └─ Writes: action_outputs → triggers cited.md post via Senso

Frontend (Vercel / TypeScript)
    ├─ Product setup form
    ├─ Feedback feed with scores and source links
    └─ Scrubbable timeline: releases + impression inflection points

Notes:

• release_feedback.release_id is nullable — feedback isn't always tied to a specific release.
• source_type enum: twitter, youtube, blog, review_site, other.
• agent_passes is an audit log for debugging agent effectiveness (tool calls made, what was written).
• action_outputs drives both the cited.md post and any future notification hooks.

1. User creates a product via the UI.
2. Initialize agent enriches metadata (links, description) using nimble_search and fetch_url.
3. Refresh agent discovers releases via Nimble and writes releases.
4. Feedback agent searches the web via Nimble (currently Reddit + X/Twitter only, rolling last 3 months in UTC), scores sentiment, and writes rows to release_feedback.
5. Summarize agent reads release_feedback, merges gaps with release dates, and writes feedback_summaries.
6. Each pass is logged to agent_passes.

On an existing product, use the ··· menu: Refresh releases, Fetch public feedback, and Refresh analysis to run those steps independently.

1. Analysis Agent reads new release_feedback rows since its last pass.
2. Generates an Impression Summary: key positive themes, key negative themes, overall score delta.
3. If the rolling score drops meaningfully (e.g. >2 points over a window), writes a row to action_outputs and publishes a post to cited.md via Senso.

1. Frontend queries releases and action_outputs for a product, ordered by date.
2. Renders a scrubbable timeline with two layers: release markers and impression-change markers.
3. Clicking a marker shows the source feedback items and the published summary.

• Scoring model: LLM-based 1–10 per feedback item, or a two-pass (classify → score)?
• Impression drop threshold: fixed delta or rolling average window?
• Scheduling: manual trigger only for MVP, or cron-based re-scrape?
• Auth: single-tenant for hackathon or multi-product dashboard from day one?