nc://fluent

Fluent: every terminal command checked against the program's own --help before it runs. A Nodecode rewrite of Bruno-coder21/fluent by Bruno-coder21.

The agent knows how to write shell commands, not how this docker, gh or curl spells its options. Fluent reads a program's own help once, keeps what it said, and checks the commands the agent runs against it: docker run --rmm comes back with did you mean --rm? instead of an error after the container was already on its way.

Get it

Tell the agent "add the fluent cell", or run nodecode add nc://fluent.

Set it up

No key, nothing to install: it reads help only for the programs you adopt. The fluent section, all of it optional, is at the end of Use it.

Adopt a program once, and from then on every command the agent runs through (sh ...), run-program, launch-program or run-shell-command is checked against it:

/fluent adopt docker      read docker's --help and its subcommands' --help (a few seconds)
/fluent status            what is read, and the mode
/fluent show docker run   the options of `docker run`, as read
/fluent check "docker run --rmm alpine"   what the checker would say, without running anything
/fluent forget docker     forget it

The model has the same check as a tool: fluent_check with a command, and adopt: true when the program has not been read yet. That is the better way to ask before writing a command you are unsure of:

(fluent:check-command "gh pr create --drafts")
;; => (--drafts is not an option of gh pr create ..., did you mean --draft?)

In block mode (the default) a flagged call does not run, and the model reads which option the program does not have and which one it probably meant. In warn mode the call runs and the note goes up on the board; off checks nothing.

Use it

{
  "fluent": {
    "mode": "block",        // block | warn | off
    "max_subcommands": 24,  // how many subcommands one adoption reads
    "adopt_at_start": []    // programs read at every start
  }
}

What it judges, and what it leaves alone

Only long options (--name), only on a command whose help was read whole, only before --, and never a word the shell expands ($VAR, globs). Short options, positional words, subcommand names and unknown subcommands are left alone, so a new docker alias is never refused. A long option that is a unique abbreviation or a known --no- negation passes.

Reading a help page runs the program itself: prog --version and prog --help, once per subcommand. Nothing else is run, nothing is sent anywhere, and the cards are plain data in this home's fluent/ folder.

Test

From this repository's root:

.github/test-cell.sh tools/nodecode-fluent

The tests run no program and dial nothing: a stub answers a small help page, and the store is a folder of its own.

Not ported

The man-page reader, the destructive-subcommand approval prompts (nodecode-guard and Nodecode's own approvals cover those), the plugin's dashboard and counters, and the post-run offer to adopt a program that rejected an option. See NOTICE.

MIT licensed.