Stock Recap — convergence-ranked picks

This skill produces a market recap and a short list of stocks the data agrees on, each with technicals. It pulls four independent legs and ranks names by how many of them point at the same ticker (equal-blend convergence):

1. Screeners — all 10 TraderDaddy Pro screeners at their default settings (Momentum Pullback, Gamma Scan, Coiled Springs, Daily Cuts, CSP Wheel, LEAPS, Leveraged, Small-Cap, Volatility Surge, Bullish Pullback).
2. Options flow — the biggest premium / smart-money unusual activity today, including the INSTITUTIONAL_ALPHA tier and repeat-strike conviction.
3. New CBOE option listings — names that just became optionable.
4. Hedge funds (13F) — TickerTrace top buys/sells, cross-fund convergence (multiple funds into the same name), and divergences.

It surfaces a dedicated under-$100 section so the list isn't all $700+ mega-caps. The visual read lives on Ghost Flow (TradingView), not here — the recap is the data layer; Ghost Flow is the picture (see "Reading the charts" below).

It opens with a regime gate + week-ahead block — the Sunday-night context a daily recap can't give you (all from the NEW dev API, api.traderdaddy.pro/api/v1, X-API-Key in repo-root .env_td_api; every leg is best-effort and degrades alone):

• 🔴/🟠/🟡/🟢 Regime banner — blends index put/call (SPY/QQQ/IWM) with SPY net gamma into a RISK-OFF → RISK-ON score. In a RISK-OFF / negative-gamma tape the convergence bar auto-tightens (names need one extra agreeing leg before they make the shortlist).
• 🗓️ The Week Ahead — the macro clock (CPI/NFP/FOMC etc., flagged by impact), a hinge-event call-out, ⚠️ earnings landmines (shortlist names that report this week — a swing entry into a print is a coin-flip, tagged in the Edge column and detail), and a one-line sector tilt (which sectors money leaned into/out of).
• 📋 Track record — a one-line forward hit-rate from score_history.mjs (below), so the list carries its own honesty check.

Mike's edge: his best picks come from Momentum Pullback names that have ALREADY started recovering — not the ones still buried in the hole. The script has a dedicated Recovery Watch section that flags pullback names where the turn is confirmed on the tape: RSI reclaimed bullish (≥50 / bullish zone), price back above the 21EMA, week no longer bleeding, A/B entry grade. It deliberately EXCLUDES deep-oversold names crossing up inside a downtrend — Ghost Flow grades those SHORT (they're bounce candidates, not longs; MIDD/CRS on 2026-07-19 were the lesson). It marks which recoveries already have bullish flow and fund buying behind them. Always call this out — and if the list is empty in a weak tape, say so plainly (nothing to chase beats a list of knives).

Four tiers of the turn (earliest → confirmed → graduated → tracking):

1. 🌅 Turning Up — the EARLIEST catch. Computes StochK(14/3) off chart-data (screeners only give today's value) and fires when K just crossed back up through 40 after dipping below — the moment the turn starts, before RSI/EMA confirm. FIG's 2026-06-30 reclaim at $18 (then +32%) is the archetype. Fresh cross within RECAP_TURN_LOOKBACK (default 2) sessions, K ≤ RECAP_TURN_KMAX (default 68, past which it's already extended). Scanned across every screener (not just pullback) so a FIG-class name is caught at the cross — it sat in csp-wheel before its 06-30 reclaim. A Source column shows which screener(s) surfaced each name. Earlier = better price but noisier; always tell Mike to confirm on Ghost Flow before sizing.
2. 🔄 Recovery Watch — the CONFIRMED tier above (RSI≥50 + above 21EMA + week not bleeding + A/B).
3. 🎓 Graduated from Watch — a name that was on the PERSISTENT 👁️ Watch list (dropped below 40 on a prior run) and just reclaimed 40. Shows was→now price, % since it went on watch, and days tracked. 🔥 = fresh cross (≤2 sessions). This is the payoff of tracking across runs: watch the drop, own the reclaim — the exact FIG catch.
4. 👁️ Watch — persistent, cross-run — names that FELL through 40. They land here and stay across runs (even after they drop off every screener; those show off ✂️) until they reclaim 40 (→ 🎓) or age out (RECAP_WATCH_MAX_DAYS, default 21). NOT actionable while here. The loop: drop → watch → cross back up → 🎓 Graduated. State lives in runs/_watchlist.json.

Scan universe: the 40-line scan runs over the full union of every screener's names (any name that landed on ≥1 screener this run, ~170 names), so a FIG- class name is caught at its cross even if it's not a pullback name yet. New watch slots are earned by pullback names only (keeps the store tight), but once on the list a name is tracked purely off StochK — so it's never lost between the drop and the reclaim, even after it leaves every screener. Set RECAP_DEBUG_WATCH=1 to log the store-only fetch each run.

How to run it

1. Run the gather script from the repo root:

   node .claude/skills/stock-recap/scripts/gather.mjs
   

   It writes a timestamped folder under .claude/skills/stock-recap/runs/<date_time>/ containing report.md (the full recap) and raw/*.json (every raw API pull, for backtesting or a deeper look). It also prints the report to stdout and a final JSON line with {reportPath, rawDir, shortlist, health}.

   • Takes ~30–90s (screeners are the slow part; they're cached 5 min server-side).
   • No setup needed: it auto-reads AGENT_API_KEY from the repo-root .env_agent_api, hits production (Railway + TickerTrace), and presents a browser User-Agent (the edge WAF 403s the bare Node UA).

   Weekly mode (Sunday night): prefix RECAP_WINDOW=week:

   RECAP_WINDOW=week node .claude/skills/stock-recap/scripts/gather.mjs
   

   This widens the flow window to the whole week and tightens the premium/score thresholds (weekly flow is noisier), and re-titles the report "Sunday Setup — Week Ahead." The regime gate + week-ahead calendar render in either mode, but they're what make the weekly run a real Sunday list. Default (no env) is the daily today recap.

2. Cron / commit-to-history — for the scheduled daily run use the wrapper:

   node .claude/skills/stock-recap/scripts/run_and_commit.mjs
   

   It runs gather.mjs, snapshots the output into the git-tracked history/ dir — history/<YYYY-MM-DD>.md (full report), history/<YYYY-MM-DD>.json (compact shortlist + health + watch-list), and history/_watchlist.json (the persistent store snapshot) — then git commits ONLY those paths (pathspec- restricted, so the rest of the working tree is never swept in) and pushes to origin (push failure is non-fatal — the commit stays saved locally). Same-day reruns overwrite the day's files; git keeps the diff history. The run dirs under runs/ stay gitignored; history/*.json is un-ignored via a .gitignore negation. Disclaw cron stock-recap daily (commit history) (job 8f0675b4, 0 17 * * 1-5 — weekday 5pm ET, post-close) invokes this.

3. Read the generated report.md (the path is in the script's final output).

4. ALWAYS eyeball the charts before presenting. Render the shortlist + tier names and actually read the PNGs — never present off the tables alone. On 2026-08-02 this reversed 5 of 8 calls in one run.

   cd .claude/skills/stock-recap
   node scripts/render_chart.mjs ZION GOOGL BMNR --out /tmp/recap_charts_<date>
   node scripts/render_chart.mjs ORCL NKE --days 300 --out /tmp/recap_charts_<date>/long
   

   Charts render as SVG in headless chromium, in the Gamma Map palette, with live TD Pro dealer levels drawn on: red walls, green support, amber PIN, purple dashed gamma flip, OI on every label. Default window is ~4 months (--days 88). Pass --no-gex to skip the options-chain fetch.

   Use --days 300 on anything you're about to call a downtrend or a breakout. 90 days is not enough to see a multi-month trendline, and that trendline IS Mike's #1 setup. The two views can disagree completely: at 90d NKE looked like dead chop under a falling 55EMA; at 300d it's a four-month base with a firm floor and flattening EMAs, coiling right under the descending line off the peak. Mike caught that and I hadn't. Conversely ORCL looked like a bounce at 90d and a worse accelerating breakdown at 300d.

   The distinction that matters — decelerating base vs active waterfall:

   • Setup: lower highs, but the decline has flattened, EMAs going horizontal, a floor tested repeatedly, price coiling under the trendline. Not triggered until it closes above the line, but it's worth watching.
   • Knife: still making lower lows (esp. undercutting the prior major low), EMAs fanned out and steeply falling, 200SMA rolling over overhead. No amount of RSI/Stoch turn makes this a long.

   Table artifacts the charts catch, every time:

   • ADX has no direction. High ADX on a falling stock = strong DOWNtrend. Never cite it as bullish without the chart (ORCL, ADX 41, in freefall).
   • A weekly % can hide a gap. RDDT's "-21% on the week" was ONE gap candle on 5x volume, the most recent bar. Fresh gap ≠ pullback.
   • Dead tickers score A+. Now filtered by DEAD_RANGE_PCT, but sanity-check the y-axis range anyway.

   Then send the PNGs to Discord grouped as keep-vs-cut so Mike sees what you saw.

5. Ghost Flow is still the confirm. The report ships no PNGs by design; Mike confirms the visual on Ghost Flow (his TradingView indicator), which encodes the whole decision — GRADE, W GATE, FLOW (CMF), SQUEEZE, %R EXHAUST, VOL PREMIUM, ADX/RSI, BOUNCE, ATR-based RISK sizing, and gamma walls on the price axis — far past what a bare candle+EMA render can show. When presenting the recap, tell Mike which shortlist names to pull up on Ghost Flow and what to look for (is the W-GATE open or WAIT, is FLOW positive, is it bouncing or rolling). The data tables here are the screen; Ghost Flow is the confirm — your own chart read in step 3 is the filter that decides what's even worth him opening.

6. Present a tight, plain-English recap to Mike — don't just dump the file. Lead with what matters:

   • The 2–4 highest-conviction names — judged on the CHART first, rank second. If the top-ranked name has a bad chart, say so and lead with the one that doesn't. Name which to confirm on Ghost Flow before sizing.
   • Anything you're cutting from the shortlist and why. The list is a starting universe, not a recommendation.
   • The Recovery Watch — Momentum Pullback names that have already reclaimed (RSI back ≥50, above the 21EMA, week not bleeding), especially ones with flow 🟢 and fund ✅ confirmation (his setup). NOT deep-oversold names in a downtrend. Point him at Ghost Flow to confirm (W-GATE open, FLOW positive, bouncing not rolling). If it's empty, say so — nothing to chase beats a list of knives.
   • The 💵 Under-$100 picks — Mike specifically wants accessible names, not just $700+ mega-caps. Always surface a few.
   • A one-line flow/fund tone read (net bullish vs bearish premium; what the funds are buying/selling; any notable cross-fund convergence).
   • ⚠️ Conflicts — names where bullish sources disagree with bearish flow or fund selling (e.g. funds buying but options heavily bearish). Flag, don't bury.
   • New CBOE listings only if there's something worth noting.

7. Point Mike at the saved report.md for the full tables/charts, and mention the raw/ JSON is there if he wants to dig in or backtest.

Scoring the track record (score_history.mjs)

The recap grades itself. After runs have aged a forward window (default 5 trading days), run:

node .claude/skills/stock-recap/scripts/score_history.mjs

For every old-enough run it looks up each shortlist name's price N trading days later (same source the run used — /api/agent/ticker/:symbol/chart-data), computes the forward return, and writes runs/<stamp>/scored.json plus a rolling runs/_scorecard.json. The next gather.mjs run reads _scorecard.json back in and prints the 📋 Track record line. It's idempotent (a scored run is cached; pass --rescore to redo) and best-effort. Buckets the results by Reversal-Watch vs base and by leg count. Tune with --horizon N.

⚠️ The convergence thesis did not survive contact with the data. At 401 picks / 30 runs the leg-count buckets are inverted: 2-leg 48% · 3-leg 26% · 4+-leg 17%, against a 47% base. More agreement predicts WORSE forward returns. Working theory: legs past two come from flow + funds, which stack correlated mega-cap noise rather than confirming a setup. As of 2026-08-02 LEG_RANK_CAP = 2 caps the rank contribution and a confirmed reversal (51%, the only bucket beating base) outranks raw agreement. Legs still gate the shortlist; they no longer buy position on it. Re-check as the sample grows — if it flips back, raise the cap. Worth running weekly (e.g. before the Sunday pull) so the track record stays fresh.

Tuning (optional env vars)

• RECAP_WINDOW=week — weekly Sunday-night mode (default today). Widens the flow window and tightens flow thresholds; re-titles to "Sunday Setup — Week Ahead."
• RECAP_SHORTLIST=20 — shortlist size (default 15).
• RECAP_MAX_PRICE=50 — cutoff for the under-$N affordable section (default 100).
• RECAP_FLOW_MIN_SCORE / RECAP_FLOW_MIN_PREMIUM — flow filter floors (weekly defaults to a higher premium floor since weekly flow is noisier).
• RECAP_TIMEOUT_MS=90000 — per-request timeout (default 60s).
• TD_API_URL / TICKERTRACE_API_URL / TD_DEV_API_URL — override base URLs (the last is the NEW dev plane backing regime + week-ahead; key in .env_td_api).

Notes / gotchas

• Convergence is directional. Only bullish-aligned legs (screener hit, bullish flow, fund buying, cross-fund buying, new-CBOE) add to a name's score. Bearish flow or fund selling against a bullish name is recorded as a conflict, not agreement, and is penalised in the ranking. This is a long-idea finder.
• Technicals enrichment: names that arrive via flow/funds only (no technical screener) are back-filled from /api/agent/ticker/:symbol (RSI, ADX, EMA stack, call/put walls, expected move). Stochastic isn't available there, so the Stoch column shows — for those.
• 13F data is TickerTrace, not TD Pro's /api/institutional/* (the agent key is scoped to /api/agent/* and can't reach it). If TickerTrace is down, the hedge-fund leg degrades gracefully and the rest still runs.
• Each run is a fresh point-in-time snapshot; nothing is overwritten. Old run folders are safe to delete (they're gitignored under .claude/).
• Charts are scripts/render_chart.mjs — hand-built SVG rendered in headless chromium (playwright), styled to match the Gamma Map product in docs/gex-chart/ and carrying live TD Pro dealer levels. Node only; no venv, no matplotlib. The earlier "no chart PNGs by design" call was wrong: the render was ugly, not useless. Reading the charts reversed 5 of 8 calls in a single run on 2026-08-02.
• Note that dealer-hud is a browser extension that draws over TradingView, so there is no renderer in it to lift. Copy the look from docs/gex-chart/ instead.