chat-interactive

chat-interactive — multi-turn chat console for ext-infer

A small standalone Composer app that wraps Displace\Infer\Model::chat() into an interactive Symfony Console command. Use it to:

• kick the tires on a freshly downloaded GGUF model
• demo ext-infer to teammates without writing scaffolding
• crib the multi-turn Prompt accumulation pattern for your own apps

Install

You'll need ext-infer itself installed first — either via pie install displace/ext-infer (once we publish releases; see RELEASE.md) or by running make install from the repository root. Verify it's loaded:

php -r 'echo extension_loaded("infer") ? "yes\n" : "no\n";'

Then install the Symfony Console dep:

cd examples/chat-interactive
composer install

If you're testing against a -d extension=… development build instead of a system-installed ext-infer, composer install will complain that the ext-infer PHP extension is missing. Bypass the check:

composer install --ignore-platform-req=ext-infer

Composer's platform check reads php.ini, not -d overrides, so the flag is only needed during dev. Once pie install displace/ext-infer lands the extension into the global ini, it goes away.

Run

./bin/chat /path/to/some-model.gguf

Or with options:

./bin/chat ~/models/Qwen3-0.6B-Q8_0.gguf \
    --system="You are a salty pirate. Speak only in metaphors about the sea." \
    --temperature=0.9 \
    --max-tokens=256

Add -v to see the model's internal <think> reasoning alongside its answer (Qwen3 / DeepSeek R1 / other reasoning models).

Commands

Inside the chat prompt:

Command	Effect
/reset	Clear conversation history; keep the system prompt.
/show	Dump every message in the conversation as a table.
/exit	Quit. Ctrl+D also works.

Options

Usage:
  chat [options] [--] <model>

Arguments:
  model                            Path to a GGUF model file.

Options:
  -s, --system=SYSTEM              System prompt (default: helpful assistant).
  -m, --max-tokens=MAX-TOKENS      Maximum tokens per assistant turn (default: 512).
  -t, --temperature=TEMPERATURE    Sampling temperature (default: 0.7).
  -v, --verbose                    Print the model's <think> reasoning.

What's worth lifting from this code

src/ChatCommand.php is the meat. Three patterns worth copy-pasting:

1. Reuse the system prompt across /reset

$base         = Prompt::system($system);
$conversation = $base;
// ... loop ...
if ($line === '/reset') {
    $conversation = $base;   // immutable; safe to share the original
}

Prompt is immutable, so $base never changes no matter how many ->withUser(...) calls the conversation has gone through. That's the key thing the DateTimeImmutable-style API buys you.

2. Don't feed <think> blocks back as history

$conversation = $conversation->withAssistant($response->answer());
//                                                     ^^^^^^^^^^
// answer(), not text() — otherwise the model sees its own reasoning
// on the next turn and tends to derail.

3. Handle length truncation visibly

if ($response->finishReason() === 'length') {
    $io->writeln('<comment>(truncated — bump --max-tokens)</comment>');
}

Better to surface a hint than to leave the user wondering why the answer cut off mid-word.

Extending

The command lives in one class (ChatCommand) by design — drop it into a Symfony app and it'll just work as a sub-command. To add a new command (say, embed), add a src/EmbedCommand.php, register it in bin/chat, and the Symfony Application will dispatch on argv.