sentiment-cachelisted

# Sentiment Cache Skill ## When to use Trigger this skill when the user says: - "Score sentiment" / "Refresh composite sentiment" - "Run the sentiment classifier" - "Show me the contrarian flags" - "Show me the sentiment cache" Do NOT trigger for: - Raw fetch (use `reddit-sentiment` or `stocktwits-sentiment` directly) - Single-shot agent sentiment reads during analysis (read the JSON file directly) - Professional news scoring (Alpha Vantage already scores those; use `us-news`) ## Why this exists The raw fetchers (`reddit-sentiment`, `stocktwits-sentiment`) collect posts and messages but don't classify them. StockTwits has *some* user-tagged Bull/Bear labels, but coverage is partial (~40-60% on liquid US names) and the labels are self-reported (gameable). This skill: 1. Reads the raw caches from both sources 2. Sends untagged message bodies through the authenticated OpenAI/Codex route for sentiment classification (bull/bear/neutral with conviction 0-1) 3. Combines user-tagged labels + LLM scores + Reddit post sentiment into a per-ticker composite 4. Applies the contrarian-filter logic (extreme bull → fade flag; extreme bear with constructive setup → buy flag) 5. Writes the canonical `sentiment.json` that the dashboard reads Per AGENTS.md §4, retail sentiment is a **contrarian filter**, not an additive bull signal. Extreme retail bullishness + extended technicals = downgrade conviction; extreme retail bearishness + Phase 1 setup = upgrade conviction. ## Provider and mod