| name | fetch-signals |
| description | Query Google Trends, Reddit, YouTube, news and Wikipedia pageviews to retrieve raw volume and velocity statistics for a topic. Use when the Trend Intelligence agent needs fresh trend/discussion/search signals for a query or topic. Always fetch source text via Jina Reader and respect the search_cache 30-day no-repeat rule. |
| metadata | {"agent":"trend-intelligence","source":"Project Atlas Agent Skills Manifest §2","layer":"L1-trend"} |
fetch-signals
Harvest real-time demand signals from external sources for a given query/topic.
When to use
- Daily trend-discovery sweep, or on a real signal change (news event, pageview spike).
- A topic needs current volume/velocity before scoring.
Sources (spec §5.5 mapping)
- Brave Search (live) via the
brave-search MCP server — call it directly for web volume.
High-volume deterministic polling instead routes through the NestJS SearchService, which adds the
search_cache 30-day gate + Redis rate limiter (SYS-SEARCH).
- News (live) via
SearchService's currents source (Currents API, primary) — confirmed working,
1,000 req/day free tier, daily-reset cap. freenews (FreeNewsApi.io) is registered as a secondary/
fallback source but was unresponsive in live testing (server accepts the connection, never replies);
treat it as best-effort, not primary.
- Google SERP (live) via the
brightdata MCP server's search_engine tool (engine: "google") —
structured organic results (title, URL, description) for demand-signal research. Complements Brave.
- YouTube video data (live) via
brightdata MCP's web_data_youtube_videos tool — structured
metadata (views, likes, description, channel) for a known video URL. Use after search_engine finds
candidate video URLs to enrich competitor analysis.
- Reddit discussion signal (live) via
brightdata MCP's web_data_reddit_posts tool — structured
post data (title, score, comments, subreddit) for a known Reddit post URL. Replaces the Reddit API
approach (which requires a commercial contract); Bright Data handles compliance. Surface Reddit URLs
first via search_engine (site:reddit.com <topic>), then read each post with web_data_reddit_posts.
- Google Trends + GDELT — still not implemented, no client exists. Wikipedia Pageviews also pending.
- Fetch & clean page text with Jina Reader (avoids headless browser on the VPS, spec §11.3).
Function signature (manifest contract)
def fetch_signals(source: str, query: str) -> list:
"""Return raw metric signals mapped by datetime."""
Inputs / Outputs
- Input:
source — must be one of SearchService's real SearchSource values:
brave, exa, currents, freenews (NOT the google_trends|reddit|youtube|news|wikipedia
labels this doc used to list -- those aren't real source identifiers at the implementation
level; currents/freenews are what to pass for a news signal), query.
- Output: list of
{datetime, metric, value} raw signals.
Cost discipline (spec §6)
- Enforce
search_cache 30-day no-repeat per query; only re-query on a real signal change.
- Batch multiple candidate queries per call where possible.
Implementation
Call the temporal-bridge MCP tool start_workflow("fetchSignalsWorkflow", "trend-signals", [{"source": source, "query": query}]), then poll get_workflow_result(workflow_id). Runs the already-implemented fetchSignals Temporal activity (server/src/modules/temporal/activities/trend-signals.activities.ts), which goes through the cached, rate-limited SearchService (the search_cache 30-day gate + Redis limiter apply automatically).
Backend dependency
search_cache is implemented (DatabaseModule); youtube_topics/_trends tables (Railway) — stubbed until wired.- API keys: Brave and Currents are live (
BRAVE_API_KEY, CURRENTS_API_KEY, via SearchService, SYS-SEARCH).
FREENEWS_API_KEY is also set but the provider itself is unresponsive (see Sources above).
YouTube, Reddit, Google Trends, Wikipedia still pending -- no client exists for any of them. Jina key is available (JINA_API).
Model
deepseek-direct/deepseek-chat (cheap, high-volume). Output must be schema-only JSON (spec §11.3).
