Command Routing & Dispatch

Related: Agent Orchestration, Command Queue, Command Suggest, Simple Text Heuristics.

Everything in _AI is a “typed command”: a natural-language phrase tokenized into a word vector and dispatched to a subsystem. This page traces how a phrase reaches the code that handles it.

Input sources

Typed commands are read from the console (getline(cin, ...) in the main loop); speech commands arrive through the audio pipeline. Both produce a vector<string> of tokens.

The router

AugmentedActions routes by the leading token(s). An early phase matches multi-word heads (routeEarlyByHead); otherwise a head keyword selects a subsystem handler:

head == "computer" | "macro" | "action"  -> tryComputer(command)
head == "driving"                          -> tryDriving(command)
head == "net"                              -> networking features
head == "rl" ...                           -> RL surface (net / life / curriculum / ...)

Module handlers

Each subsystem exposes HandleSpeechCommand(command), which walks the token vector by index: skip the subsystem head, read a verb, and branch. For example computer rl lands in _Computers::HandleSpeechCommand, which forwards the rl subtree to _ComputerRL::HandleSpeechCommand.

Worked trace: computer rl explore

1. Tokens: [computer, rl, explore].
2. tryComputer → _Computers::HandleSpeechCommand.
3. command[1] == "rl" → _ComputerRL::HandleSpeechCommand.
4. verb explore → enable neural policy, set curiosity/epsilon, dry-run off, mode = Explore, start the desktop loop.

Intent classifier and NLU fallback

When a phrase does not match a known verb, the CommandIntentClassifier and the RLLanguageBridge (SuggestCommand) can map free-form language onto a typed command — e.g. an English instruction rewritten to computer rl ... and re-dispatched.

Adding a command

1. Add a verb branch in the subsystem's HandleSpeechCommand.
2. Register a catalog/help entry so it appears in suggestions and help output.
3. (Optional) add an intent-classifier example and a Lua global.