autosearch:recent-signal-fusion

Recent Signal Fusion — Cross-Platform Recency Bundle

Adapted from last30days-skill's SourceItem / Candidate / Cluster / Report pipeline + Scira's group-mode aggregation. Turns N channels' recent outputs into a single time-weighted cluster list.

Input

input:
  topic: str                                 # research topic
  time_window: "24h" | "7d" | "30d"          # how far back to care
  channels: list[str]                        # which channel skills to call
  min_recency_ratio: float                   # reject items older than (now - time_window)
  max_items_per_channel: int                 # e.g. 30

Pipeline Stages

1. Parallel fetch — call each channel with a time-window filter in the query. Collect raw evidence lists.

2. Normalize to SourceItems:

   source_item:
     id: str                        # hash(url + title)
     url: str
     title: str
     content_snippet: str
     platform: str                  # source_channel
     posted_at: datetime
     engagement: {likes, comments, shares} | null
     language: str
     raw_evidence_ref: dict         # original Evidence slim-dict
   

3. Recency filter — drop items older than time_window. Use extract-dates skill for items without explicit posted_at.

4. Semantic clustering — group near-duplicate SourceItems across platforms:

   cluster:
     id: str
     canonical_title: str           # best representative title
     topic_keywords: list[str]      # extracted from the cluster
     sources: list[source_item_id]
     platform_spread: int           # how many distinct platforms
     earliest_posted_at: datetime
     latest_posted_at: datetime
   

5. Rank by weighted score:

   score = (
     0.4 * recency_score(latest_posted_at, time_window)
     + 0.3 * platform_spread_score(platform_spread)   # cross-platform = stronger signal
     + 0.2 * engagement_score(sum(engagement))         # aggregated across sources
     + 0.1 * platform_reliability_score(platforms)     # quality weighting
   )