Agent Skills: Gold Buy Signal Skill

Gold Buy Signal Skill

Overview

Lucas trades gold short-term/micro: small positions with a stop-loss and take-profit, in either direction — not a buy-and-hold play. This skill answers "should I buy gold now?" with a plain BUY / SELL / HOLD call and a one-line reason, weighted across 8 independent sources:

• 3 scraped prediction sites (deterministic, ~20s/run, on a 30-min cron)
• TradingAgents, a separate multi-agent LLM framework (non-deterministic, ~10-15 min/run, on its own daily cron) — see the dedicated section below
• 4 market drivers (real yields via TIP ETF, US dollar index, CFTC speculator positioning, GC=F intraday momentum) — plain HTTP fetches, same 30-min cadence as the scraped sites

All the fetching/scoring logic already exists as a deterministic script. Do NOT re-fetch the 3 sites yourself with WebFetch/browser tools and re-derive a verdict by eyeballing the pages — that's slower, more expensive, and less consistent than the engine. Only fall back to manual fetching if the engine itself is broken (see Troubleshooting).

When to Use This Skill

• User runs /gold-buy-signal
• User asks "should I buy gold now/today?", "is gold a buy?", "gold looking bullish or bearish?", or similar
• User asks to check/refresh the gold signal, or to see the gold dashboard

Step 1: Run the engine

/home/lucas/.config/gold-signal/run.sh

This is host-only (not inside a DDEV container) — run it as a plain host Bash command. It prints one JSON line to stdout (also written to ~/.config/gold-signal/state/latest.json) and takes ~15-25s (3 headless Chromium page loads). If the caller is inside a container and this path isn't reachable, tell the user and stop — don't try to reimplement the fetch.

If a run is already in flight (flock), the command returns almost instantly with no output — in that case read state/latest.json instead, it'll be current within ~30s.

Step 2: Parse the JSON and answer

The payload shape:

{
  "generated_at_utc": "...",
  "overall_score": -1.0..1.0,
  "verdict": "BUY" | "SELL" | "HOLD" | "UNKNOWN",
  "agreement": "unanimous bullish" | "unanimous bearish" | "split" | "mixed/neutral",
  "bullish_count": N, "neutral_count": N, "bearish_count": N,
  "short_term_score": -1.0..1.0,   // today/tomorrow lane — THE one for micro trades
  "medium_term_score": -1.0..1.0,  // multi-day/weekly lane
  "reason": "one-line explanation, already composed",
  "sources": [ {id, label, url, ok, vote, applied_weight, headline, detail, error}, ... ]
}

sources has 8 entries: goldpricewatch, litefinance, becoin, tradingagents, plus drivers real-rates, dxy, cot, intraday-momentum — same shape, same weighted-vote participation for all. tradingagents's detail.full_reasoning carries its full debate writeup when present.

When the two lane scores disagree, say so explicitly — e.g. "overall HOLD, but note short-term is bearish while the macro/weekly picture is bullish". For Lucas's micro-trade style the short lane is the actionable one; a medium-term bullish tilt is context, not an entry signal.

Answer the user directly using verdict and reason — they're already composed, don't second-guess or rewrite the logic. Format per the user's default terse style:

[BUY/SELL/HOLD]. [reason from payload, tightened to one line].
[1 line per source: label — headline]

Example:

HOLD. Split 2 bullish / 1 neutral / 1 bearish of 4, weighted score barely positive (+0.01) — no real edge.
GoldPriceWatch: Bullish call but only 30% historically direction-accurate (down-weighted hard)
LiteFinance: analyst's base case is short (bearish) below $4,576, stop $4,612
BeCoin: Bullish technical read, 4/5 signals up
TradingAgents: Hold on GLD — overbought technicals vs. medium-term support, sitting out

If verdict is "UNKNOWN" (all sources failed to fetch/parse), say so plainly and suggest retrying — don't guess a verdict.

Always surface it when historical_direction_accuracy_pct on any source is below 50% — a source discloses this itself, and it means their own track record is worse than a coin flip on direction. It's already reflected in the vote weight, but the user should still see it named, not buried.

This is informational consolidation of 3rd-party model outputs, not financial advice — don't present it as more certain than it is. A "BUY" or "SELL" here means "the 4 sources lean that way right now," not a guarantee.

Step 3 (optional): offer the dashboard

http://127.0.0.1:8934/ (bookmarked) — served by the gold-signal-web systemd --user unit, always up. Full per-source breakdown (raw parsed fields, support/resistance levels, stop-loss levels from LiteFinance, scenario probabilities from BeCoin) plus a track record section: each source graded against its own later-reported price once its call's horizon passes (1 day for LiteFinance/BeCoin, 7 days for GoldPriceWatch) — green/red hit-dots per source, so "is this source worth trusting" is visible directly rather than taken on the source's own word. Mention it exists if the user wants more than the one-liner; every page load auto-refreshes the data in the background (see ~/.config/gold-signal/README.md).

TradingAgents — the 4th source

TradingAgents is a separate multi-agent LLM trading framework, installed at ~/workspace/proxiblue/trading-agents, running fully local via Ollama (qwen2.5:7b-instruct — see that project's .env for why not the other locally-pulled models). It analyzes GLD (SPDR Gold ETF, the closest liquid Yahoo-Finance-backed proxy TradingAgents' equity/ETF-shaped analysts can work with — not spot XAU/USD like the other 3 sources, so expect some tracking difference) and produces a 5-tier rating (Buy / Overweight / Hold / Underweight / Sell) via a full analyst-debate-risk pipeline.

It's now a genuine 4th vote in sources — engine.py's load_tradingagents_source() maps its rating to the same -1..+1 scale (Buy=+1, Overweight=+0.5, Hold=0, Underweight=-0.5, Sell=-1) and folds it into the weighted score with its own configurable weight (tradingagents.weight in config.yaml, default 0.5 — no self-disclosed accuracy to auto-discount by yet, unlike GoldPriceWatch).

It does NOT run on the 30-min cron — a run takes ~10-15 min locally, far too slow for that cadence and too much repeated GPU load to boot. It runs on its own separate daily cron instead (~/workspace/proxiblue/trading-agents/run_gold_check.sh), writing ~/workspace/proxiblue/trading-agents/state/gold_check.json, which gold-signal reads on every one of ITS runs — so the TradingAgents vote in any given gold-signal payload can be up to ~24h old (marked stale: true and excluded from the score past stale_hours, same graceful-degradation pattern as any other source going quiet).

To refresh it on request: ~/workspace/proxiblue/trading-agents/venv/bin/python ~/workspace/proxiblue/trading-agents/run_gold_check.py GLD — warn the user this takes 10-15 minutes before running it interactively. gold-signal won't see the fresh result until its own next run afterward.

Troubleshooting

• A source shows "ok": false: that site likely changed its page layout — the regex in ~/.config/gold-signal/engine.py (functions parse_goldpricewatch / parse_litefinance / parse_becoin) needs updating. Don't hand-patch the JSON; fix the parser. Use the error field and the site's current rendered text (fetch it via a headless browser tool, not curl — all 3 are JS-rendered) to find what changed.
• All sources fail / verdict: "UNKNOWN": could be a real fetch problem (network, site down) or the host load guard skipped the run (check ~/monitor/metrics-*.csv — the engine logs "skip: load1=... " when it bails early). Re-run once load settles.
• tradingagents source shows "ok": false: either no run has happened yet (state/gold_check.json missing), or its last run errored (error field carries the reason — often a market-data hiccup or the local model hallucinating a bad ticker suffix; see that project's run_gold_check.py, which already retries once). Not gold-signal's code to fix — go to ~/workspace/proxiblue/trading-agents.
• Engine won't run at all (missing venv, playwright not installed): see ~/.config/gold-signal/README.md for setup — it's a plain python3 -m venv venv && ./venv/bin/pip install -r requirements.txt plus Playwright's Chromium (already cached fleet-wide under ~/.cache/ms-playwright from pb-watch).

Success Criteria

• One BUY/SELL/HOLD verdict with a short, honest reason
• Each source's own reliability caveats (esp. low disclosed accuracy, or a stale/failed TradingAgents run) surfaced
• No re-derivation of the scoring logic by eye — the engine's verdict/reason are authoritative
• Clearly labeled as informational, not investment advice