Tool calling

Before the LLM ever runs, the copilot decides which facts a question is about. That routing is deterministic, golden-pinned, and defined in crates/copilot-core/src/tool.rs. It is the query command of the core.

The query command

{ "cmd": "query", "question": "why did BTC dump?" }

returns

{ "tool_calls": [
  { "tool": "get_fact", "symbol": "BTCUSDT", "kind": "price_move" },
  { "tool": "get_fact", "symbol": "BTCUSDT", "kind": "liquidation_cluster" }
] }

A ToolCall names the tool (always get_fact today), the symbol to read, and the fact kind. The list contains one call per (kind, symbol) that both the question routes to and exists in the current context, sorted by (kind, symbol) ascending. Build the context first (build_context); query reads it.

The routing table

The question is lowercased and matched, case-insensitively, against a fixed keyword table. Every matched keyword contributes its fact kinds; the union is requested.

Keyword (substring)	Fact kinds requested
dump, crash, sell	price_move, liquidation_cluster
drop, fall, rise, move	price_move
pump, rally, moon	price_move, oi_change
surge	price_move, volatility_spike
liquidat, cascade	liquidation_cluster
leverage	oi_change, liquidation_cluster
funding	funding_flip
open interest	oi_change
order book, orderbook, imbalance, book	orderbook_imbalance
volatil	volatility_spike

If a question matches no keyword, the router falls back to all six fact kinds — the context speaks for itself rather than returning nothing.

Determinism

Routing uses ordered set operations (a BTreeSet keyed by (kind, symbol)), so the result both deduplicates and comes out in a fixed order. The same question against the same context always yields the same tool calls, in every language. The golden corpus and the conformance tests pin the routing.

From routing to an answer

query is the deterministic first step. The CLI's ask subcommand runs it, reads the routed facts out of the context, renders them into a prompt, and hands that to the LLM adapter (LLM_ADAPTER.md). The routing decides what the model is allowed to see; the model only reasons over facts that were actually derived from the feed.

See also

Facts · LLM adapter · Grounding · Cookbook.