Pattern Skill

Purpose

Turn a trader's pattern idea — either written in natural language or drawn directly on the chart — into a runnable JavaScript detection script that scans the active OHLC dataset and returns match regions. Vibe Trade loads the script into the code editor, runs it against the data, and surfaces results in the Pattern Analysis bottom-panel tab.

This skill is the go-to for any "find me X in the price history" question.

When to use this skill

Vibe Trade should dispatch to the Pattern Skill when the user:

• Describes a known technical pattern by name — "bull flag", "double

bottom", "head and shoulders", "ascending triangle", "breakout", etc.

• Has just drawn a region on the chart via the pattern selector (a

serialized SHAPE: [...] fingerprint arrives as the user's message)

• Asks for a custom indicator — "RSI smoothed with a 20 EMA", "momentum

oscillator with z-score normalization", etc.

• Pastes Pine Script code and asks for a JavaScript equivalent

Instructions

1. If the message is a fingerprint (contains SHAPE: and Sliding window):

• First analyze the shape — what pattern does it resemble, what's the

trend, the volatility, the implied direction?

• Ask the user for confirmation before generating a detection script.
• The chat router stores the fingerprint as context.pending_fingerprint

and waits for a yes/proceed before dispatching again.

1. If the message confirms a pending fingerprint (context.pending_fingerprint

is set and the user said yes), generate a shape-matching script:

• Use Pearson correlation with threshold 0.50 — NOT 0.85+. Strict

thresholds silently produce zero matches, which looks like the skill is broken.

• Fall back to returning the top 5 candidates regardless of threshold.
• Emit script_editor.load with the generated JS.
• Emit bottom_panel.activate_tab → pattern_analysis.

1. If the message is a named pattern ("bull flag", "double bottom"):

• Generate a detection script using price-structure rules, not shape

correlation.

• Use forgiving thresholds (3–5% tolerance, not 1%).
• Emit the same two tool_calls as above.

1. If the message is an indicator request ("custom RSI...", "ATR-based

stop"):

• Generate an indicator script that returns an array of values (or null

for insufficient data).

• Wrap the result as script_type: "indicator" so the frontend registers

it in the Resources dropdown.

1. If the user already has a script loaded and asks for a modification

(edit mode — context.pattern_script is non-empty):

• Return the modified script in full.
• Do NOT switch the view to Code — the user is iterating and wants the

chat feedback inline.

Inputs

| Key | Type | Meaning | |---|---|---| | message | string | Natural-language description or SHAPE fingerprint | | context.pending_fingerprint | string | Previously-analyzed fingerprint awaiting confirmation | | context.pattern_script | string | Existing script — triggers edit-mode | | context.dataset_id | string | Active dataset id (used when running the script) |

Outputs

Returns a SkillResponse with:

• reply — short plain-language explanation of the script
• script — the generated JavaScript source
• script_type — "pattern" / "indicator" / "pine_convert"
• data.parameters — parameter hints extracted from the script
• data.indicators_used — list of indicator helpers referenced
• data.default_params / data.indicator_name — indicator mode only
• tool_calls — see below

Tools used

This skill may emit any of the following tool_calls:

| Tool | When | Payload | |---|---|---| | chart.pattern_selector | User asks to "mark a pattern" or similar | true to open, false to close | | chart.highlight_matches | After a successful run with matches | PatternMatch[] | | chart.draw_markers | To annotate specific bars | Marker[] | | chart.focus_range | To zoom the chart to a match region | {startTime, endTime} | | script_editor.load | Always, when a script is generated | The JS source string | | script_editor.run | When the user asks "run it now" | — | | bottom_panel.activate_tab | After generating a script | "pattern_analysis" | | bottom_panel.set_data | To push results into a tab's store slot | {target, data} | | notify.toast | Non-blocking status ("Found 7 matches") | {level, message} |

Examples

Natural-language input > "Find bull flags after a 5% rally."

→ Returns a script that scans for a strong uptrend followed by a consolidation in a downward-sloping channel, with 3% tolerance on the flag bounds. Script loads into the editor; bottom panel switches to Pattern Analysis.

Indicator input > "Custom RSI smoothed with a 20-period EMA."

→ Returns an indicator function with default_params = {rsi_period: 14, ema_period: 20}. The frontend registers it in Resources and enables it on the chart.

Fingerprint input (from chart.pattern_selector) > Find this 285-bar pattern (scale-free): > SHAPE: [0.02, 0.25, 0.67, ...] > Sliding window: 285 bars

→ First pass: analyze the shape, ask for confirmation. → On confirmation: generate a correlation-based detector with threshold 0.50 and fallback-to-top-5 so the user always has something to look at.

Underlying implementation

This skill is wired through core/agents/processors.py::_pattern_processor, which calls core/agents/pattern_agent.py::PatternAgent.generate. The agent owns PATTERN_SYSTEM_PROMPT, INDICATOR_SYSTEM_PROMPT, and PINE_CONVERT_PROMPT.