Proper CLI: one-shot `ask`, `chat`, override flags, `config show`

[GeigerJ2]

The console script today (aiida-agents -> cli.py:main) is just a bare REPL with no subcommands. We should to turn it into a thin Click CLI. Click because verdi is Click and it's the idiom we and our users already live in, so the patterns from aiida-core carry over.

The real motivation isn't the REPL, which we already have, it's one-shot mode: aiida-agents ask "what are the last 5 processes?". That's scriptable, testable, CI-able, and it's how I'd actually dogfood the agent without sitting in an interactive loop. aiida-agents chat keeps today's interactive REPL.

I'd also add --provider / --model / --profile as override flags sitting on top of the typed settings, so precedence reads flag > env > file > default. --profile especially: the agent runs against a profile, so people should be able to point it at one inline instead of doing verdi profile setdefault gymnastics first.

And aiida-agents config show: read-only, prints the effective settings and where each value came from. It's our verdi config list analogue, and it earns its place precisely because once config comes from several layers people start asking "why is it using that model and not mine?". Showing provenance answers that in one command. I'd hold off on config set / writing files until someone actually needs persistence beyond .env (YAGNI for now).

Depends on the typed-settings work, #6.

[GeigerJ2]

Just as a reminder, also aiida-agents rag init to build the text corpus for the RAG and do the embeddings (to be decided, possibly later via CI, and shipped with the wheels). But, at least have the CLI entry point.

[GeigerJ2]

If an ollama model is not available, but selected through the configuration, the user should be prompted if it should be downloaded, if approved, ollama pull should execute, and the package should continue as desired, using the just downloaded model.

[GeigerJ2]

One more for the CLI pile: a aiida-agents check (or warmup/doctor, whatever reads best) that verifies the configured model is actually usable and warms it as a side effect.

The point is that it reads the resolved settings (provider + AIIDA_AGENTS_MODEL) rather than me repeating the model name on the side, so it does the right thing per provider:

• ollama: fire a 1-token generation to load the model into memory, so the first real query isn't a cold load (which on the big local models can be large. Same invocation confirms the model is pulled; if it 404s, tell me ollama pull <model> instead of failing mid-session (ties into the auto-pull idea).
• cloud (openai-compatible / anthropic): there's no "warm", but the same tiny request checks the API key + endpoint reachability, which is exactly what I want to know before starting a session rather than on the first question.

It pairs naturally with config show and depends on the same Click structure, so it should land with that work, not before.

Also, we should add a spinner for when there is background computation ongoing, so it does not seem like things are stuck.

[GeigerJ2]

Ideally, also multi-line inputs, and vim keybindings. I'd like the migration from claude code to be seamless.

[GeigerJ2]

See also: #14