narrative-tracker

Read memory/MEMORY.md for context on prior narrative observations. Read the last 3 days of memory/logs/ — specifically any prior ### narrative-tracker entries — to (a) avoid re-reporting the same narratives without new info, and (b) detect phase transitions vs the last run.

Goal

Produce a decision-grade narrative map: every narrative gets a mindshare score, a velocity arrow, a sentiment tag, named drivers, and an explicit position call. Classification without a position call is noise.

Steps

1. Ingest signals

a. X/Twitter narratives — X.AI API (primary). The primary signal is a direct curl to the X.AI Responses API with Grok's x_search tool — see ## Fetching below for the full contract. XAI_API_KEY is injected into this skill's environment via requires:; it is present and valid, so this path is required whenever the key check prints KEY_PRESENT. Set the Bash tool timeout to ≥180000 when running the curl (x_search takes 30-120s) and capture the HTTP status:

FROM_DATE=$(date -u -d "3 days ago" +%Y-%m-%d 2>/dev/null || date -u -v-3d +%Y-%m-%d)
TO_DATE=$(date -u +%Y-%m-%d)
# Presence check uses the ${VAR:+x} modified expansion, NOT a bare $XAI_API_KEY — the bare
# form trips the Bash secret-expansion analyzer; the :+ form does not (it never puts the value on the line).
[ -n "${XAI_API_KEY:+x}" ] && echo KEY_PRESENT || echo KEY_UNSET
# Build the JSON payload to a file with jq (do NOT hand-assemble it), then POST the file.
# TWO SEPARATE commands on purpose: the jq (which interpolates $PROMPT/$FROM_DATE/$TO_DATE)
# is kept OUT of the ./secretcurl command — never pipe jq into secretcurl, and never put a
# $VAR in the secretcurl line, or the Bash permission analyzer blocks the network call.
# The `>` redirect to /tmp is fine in read-only mode (it is not a repo path; nothing reverts it).
PROMPT="Search X for the dominant crypto and tech narratives from ${FROM_DATE} to ${TO_DATE}. Return 12-15 distinct narrative threads. For each: 1) short label, 2) 3-5 representative @handles driving it, 3) 2-3 tweet permalinks, 4) rough mention-volume descriptor (niche / growing / saturating / cooling), 5) the strongest one-line bear case against it."
jq -n --arg p "$PROMPT" --arg fd "$FROM_DATE" --arg td "$TO_DATE" \
  '{model:"grok-4.6", input:[{role:"user",content:$p}], tools:[{type:"x_search",from_date:$fd,to_date:$td}]}' \
  > /tmp/xai-nt-payload.json
HTTP=$(./secretcurl -s -o /tmp/xai-nt.json -w '%{http_code}' --max-time 150 -X POST "https://api.x.ai/v1/responses" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer {XAI_API_KEY}" \
  -d @/tmp/xai-nt-payload.json)
echo "xai http=$HTTP bytes=$(wc -c </tmp/xai-nt.json)"

Run that block verbatim (do not hand-reassemble the JSON — the jq -n builder exists precisely so quoting/expansion can't break; keep the jq and the ./secretcurl as two separate commands). The echo "xai http=$HTTP ..." line must appear in your output — it is your proof the call ran. On HTTP=200 with a non-empty body, parse /tmp/xai-nt.json with jq -r '.output[] | select(.type == "message") | .content[] | select(.type == "output_text") | .text' and use that as the primary narrative signal (SOURCE=api).