analytics-collector

name	analytics-collector
description	Collect X and LinkedIn post metrics for the Genfeed content loop and emit normalized Metric JSON for feedback scoring.
license	MIT
metadata	{"author":"genfeedai","version":"1.0.0"}

Analytics Collector

You measure what happened after publishing. Given the ids of posts that went out, you pull their engagement from X and LinkedIn and normalize each to one Metric. You are the analytic in trend -> remix -> produce -> post -> analytic -> repeat.

This is where the loop closes. The orchestrator pipes each Metric you emit into gf record-metric <id>; the connector recomputes that item's feedbackScore; and gf feedback <term> turns those scores into a multiplier that lifts previously-successful themes to the top of the next trend scan. You produce the raw measurement that makes the factory learn.

Pure worker: reads one token from the environment, persists nothing, writes no manifest itself.

---

Run It

bun run scripts/collect.ts --platform x --ids 1799111,1799222
bun run scripts/collect.ts --platform linkedin --ids "urn:li:share:6844785523593134080"

A ContentItem (or an array of {platform, postId}) on stdin is an alternative to --ids — every derivative with a postId is collected:

bun run ../genfeed-connector/gf.ts get <id> \
  | bun run scripts/collect.ts --platform x

Requires Bun 1.1+. Zero dependencies — only Node built-ins and global fetch.

---

Platforms

X

• Endpoint: GET https://api.x.com/2/tweets?ids=...&tweet.fields=public_metrics,organic_metrics,non_public_metrics.
• public_metrics works with an app-only or user token; organic_metrics / non_public_metrics (impressions, link clicks) need a user-context token for the author. If the richer fields are denied, it automatically retries with public_metrics alone.
• Mapping: impressions <- impression_count, likes <- like_count, comments <- reply_count, shares <- retweet_count + quote_count, clicks <- url_link_clicks.

LinkedIn

• Endpoint: GET https://api.linkedin.com/rest/socialActions/<urn> with LinkedIn-Version + X-Restli-Protocol-Version: 2.0.0.
• Mapping: likes <- likesSummary.totalLikes, comments <- commentsSummary.aggregatedTotalComments.
• Impressions/clicks for member posts require organization analytics scopes and are left undefined here.

Per-id failures warn to stderr and are skipped; the run returns whatever it could collect.

---

Output

{
  "count": 1,
  "metrics": [
    {
      "platform": "x",
      "postId": "1799111",
      "impressions": 10400,
      "likes": 612,
      "comments": 41,
      "shares": 88,
      "clicks": 230,
      "capturedAt": "2026-06-08T12:00:00.000Z"
    }
  ]
}

Each metric matches the Metric shape in genfeed-connector/lib/schema.ts.

---

How It Closes The Loop

export X_BEARER_TOKEN="$(bun run ../genfeed-connector/gf.ts token x)"

# 1. measure
bun run scripts/collect.ts --platform x --ids "$POST_ID" > metrics.json

# 2. record each metric — this recomputes feedbackScore on the item
jq -c '.metrics[]' metrics.json | while read -r m; do
  bun run ../genfeed-connector/gf.ts record-metric "$ITEM_ID" <<<"$m"
done

# 3. next cycle: trend-scout's raw scores are multiplied by (1 + feedback),
#    so themes that performed before rise to the top
bun run ../genfeed-connector/gf.ts feedback "ai agents"   # -> { "term": "ai agents", "multiplier": 0.74 }

collect measures, the connector remembers, trend-scout re-ranks. That three-skill cycle is the locked loop.

---

Security

• Tokens (X_BEARER_TOKEN, LINKEDIN_ACCESS_TOKEN) are read from the environment, used in memory, and never written anywhere. Resolve them through the connector.
• All calls are read-only GETs against the platform analytics endpoints; this worker never posts, edits, or deletes.