Two tools, one vocabulary. screen_symbols evaluates up to three AND-ed conditions across the most-traded pairs (or a list you give it) on one timeframe, in ONE quota unit. backtest_condition takes the same conditions, one symbol and one timeframe, and reports what price did after every past occurrence — with the unconditional baseline so the edge is separated from the tape's drift.

Together they answer the two questions traders actually ask: "what fits this pattern right now?" and "has this pattern meant anything before?"

The condition vocabulary

A condition is { metric, op, value, period?, period2? }. op is lt / gt (latest value) or crossAbove / crossBelow (a one-bar event: the previous bar was on the other side).

Periods clamp to 2–200. Conditions are AND-ed; there is no OR — run two scans for an OR and merge. Unknown metrics are rejected, so stick to the table (the full recipe list is in references/condition-cookbook.md).

Screening (screen_symbols)

Arguments: interval (default 4h), conditions (≤ 3), universe (5–100 most-traded by 24h volume, default 50) OR symbols (≤ 100, explicit), limit (≤ 50 rows), sortBy volume / change24h / metric, sortOrder.

Choose the timeframe from the user's horizon: intraday → 1h (or 15m), swing → 4h, position → 1d. Widen universe to 100 when the user wants small caps or the first scan matched little.

Every result row carries the metric values under their labels (values["RSI(14)"]), plus price, change24hPct, volume24hUsd, bias, trend, rsi14, adx14, atrPct, volumeRatio, squeeze. When matched > returned, say so and offer to raise limit.

Comparison mode: symbols with no conditions returns every listed symbol as a table — this is how you answer "compare BTC, ETH and SOL" or "how do my five coins look on the daily".

Translate intent to conditions before calling, and say what you translated it to:

• "oversold" → rsi lt 30 (or 35 for a wider net); "deeply oversold" add bbPercentB lt 0
• "overbought" → rsi gt 70
• "in an uptrend" → priceVsEma gt 0 period 200 + supertrend gt 0
• "with rising volume" → volumeRatio gt 1.5
• "squeezing" → bbWidthPct lt 4 on 4h (lower on 1h) — then sort by metric ascending
• "dumped today" → changePct lt -5 period 24 on 1h, or period 6 on 4h
• "golden cross" → emaSpread crossAbove 0 period 50 period2 200 on 1d

Backtesting (backtest_condition)

Arguments: symbol, interval, conditions (same vocabulary), horizons in bars (≤ 4; defaults ≈ 4h / 1d / 3d: 1h → [4, 24, 72], 4h → [6, 18, 42], 1d → [1, 3, 7]).

The study runs over the whole stored tape — up to 1000 candles, so roughly 41 days on 1h, 166 days on 4h, 3 years on 1d. An "occurrence" is the FIRST bar of each run where the condition held (ten consecutive oversold bars are one episode). For each horizon you get samples, avgReturnPct, medianReturnPct, winRatePct, avgMaxUpPct / avgMaxDownPct (average best and worst excursion, wick-accurate), bestPct / worstPct, baselineAvgReturnPct (every bar on the same tape) and edgePct = average − baseline. Also activeNow, currentValues, lastOccurrence, and recent[] with the last five episodes and their realised returns.

How to read it honestly:

• Edge and sample size travel together. edgePct +1.8 on 27 samples is a statement; +6 on 4 samples is an anecdote, and the tool says so in warnings. Never quote the edge alone.
• Median vs average: a positive average with a negative median means a few big winners carried it. Say which.
• Excursions are the risk picture: avgMaxDownPct −5.1 on a "bullish" setup means the average episode went 5% against you before the horizon ended. That belongs next to the win rate.
• Horizons that disagree (positive at 1d, negative at 3d) usually mean a bounce that fades — report the shape, do not pick the flattering horizon.
• Recent episodes newer than a horizon are excluded from that horizon (null in recent), never counted as zero.
• Coverage limits what the study can see: 41 days of 1h data contains one market regime. A 1d study is the only one that spans cycles.

Workflows

"What's oversold right now?" → one screen_symbols. Present the table, then (optional, one more call) backtest_condition the same condition on the top match so the answer carries "and here is what that has meant on this coin before".

Screen, then dig in → a screener row already carries bias, trend, RSI, ADX and ATR% on one timeframe. When the user wants more on the best matches (levels, VWAP, the 1h/4h/1d picture), pass the top three as symbols to ONE get_technical_indicators call (technical-analysis) — never one call per coin.

"Find me setups" / "watchlist" → decide the archetype with the user in one line (mean-reversion vs trend-continuation vs breakout), run one scan per archetype (2–3 calls), and present each list with the conditions it was built from. Do not promise an outcome for any row.

"Is X actually bullish?" → one backtest_condition on the coin and timeframe in question; if the user names no coin, BTC on 1d is the most meaningful default and say why (longest tape).

"Compare A, B, C" → screen_symbols with symbols, no conditions, on the timeframe they care about. Add get_derivatives (technical-analysis) only if the question includes funding or positioning.

Presenting

Screen: state the conditions in words and in the tool's labels, the timeframe, and the universe ("50 most-traded pairs"), then one row per symbol:

| Symbol | Price | 24h % | <metric labels…> | Bias | Trend | ADX | ATR% |

Backtest: one header line (symbol, timeframe, condition, occurrences, coverage), a horizon table —

| Horizon | Samples | Avg % | Median % | Win % | Avg best % | Avg worst % | Baseline % | Edge % |

— then the last episodes and whether the condition is active now. Close with the sample-size caveat in your own words.

Conduct

• A screen is a list of things that are true about the chart right now; a backtest is a sample of what followed in the past. Neither is a recommendation. Present the evidence, note what would invalidate the setup, and leave the trade — and the sizing — to the user.
• Never turn "70% win rate on 12 samples" into a forecast; historical stats describe their sample.
• Nothing in this connector trades — no order, close, transfer or withdrawal tool exists by design. If asked to execute, say so plainly; the decision and the trade stay with the user.
• Quote only what the tools returned; when a symbol is missing from the universe say it was not scanned rather than implying it failed the filter.
• When a specific trade idea is under discussion, end with one plain sentence that crypto derivatives are high-risk and this is market information, not financial advice.