Struct ToolSelector 

pub struct ToolSelector { /* private fields */ }

Dynamic tool selector that loads tools based on conversation context.

Core tools (~18) are always available. Contextual tool groups activate when relevant keywords appear in recent messages and stay active for the remainder of the bout (monotonic growth) to maximize prompt cache hits.

The tool executor can still run ANY tool regardless of what the LLM sees — capabilities are checked at execution time, not selection time. The model never loses capability: shell_exec is always in the core set and can do anything the contextual tools can via command-line equivalents.

Implementations

impl ToolSelector

pub fn new(capabilities: &[Capability]) -> Self

Create a new selector for a fighter with the given capabilities.

Groups whose capabilities are granted but have no representation in the core tool set are auto-activated on construction. This prevents a capability regression where, e.g., a fighter with SourceControl wouldn’t get git tools on turn 1 unless the user explicitly said “git”.

pub fn select_tools( &mut self, messages: &[Message], ) -> (Vec<ToolDefinition>, bool)

Select tools based on conversation context.

Returns (tools, tools_changed) where tools_changed is true if the tool list differs from the previous call (useful for cache invalidation decisions).

Scans the last 4 messages (current + 3 prior) for keywords. Once a group activates, it stays active for the rest of the bout.

pub fn active_group_count(&self) -> usize

Return the number of currently active contextual tool groups.