feat(cli): add ogx launch opencode command

[nathan-weinberg]

What does this PR do?

Add a new ogx launch command group with opencode as the first subcommand. The command queries the running OGX server for available LLM models, generates an OpenCode-compatible provider configuration, and launches OpenCode with all models available via the OPENCODE_CONFIG_CONTENT environment variable.

Closes #5639

Test Plan

Automated tests added, you can try this manually via the following steps:

1. start a local OGX server (I recommend having some SaaS inference provider enabled such as gemini)
2. in a different terminal, run ogx launch opencode
3. try it out - all models available in your OGX server should be visible if you run /models in OpenCode and search for OGX

[cdoern]

I feel like ogx launch could also start an ogx server if one is not running 🤔

[nathan-weinberg]

I feel like ogx launch could also start an ogx server if one is not running 🤔

For ogx launch opencode I decided against adding any server boot logic - this is partially due to the fact that I think launching a server as part of this is a bit heavyweight and would require a bit more interactive prompting than I would like (all the server configuration would need to be determined, processes, and launched on the fly) - otherwise partially due to the fact that this supports a remote server as well

Perhaps ogx launch could be a replacement for ogx stack run?

[mattf]

nice

[mattf]

where did this magic come from?

[nathan-weinberg]

These are placeholder defaults Claude came up with - after some digging, the OGX models API doesn't currently expose context/output limits but OpenCode requires them in its config.

We could either pull these from model metadata if/when that's available, or make them CLI arguments. These are just conservative defaults that should work for most models.

[mattf]

this is a place where ogx should contribute to the ecosystem. most model providers document their context lengths, but don't make them programmatically discoverable.

models.dev is handy for discovery.

however, ogx also works w/ vllm / ollama / nim, where each deployment is bespoke and cannot be covered by models.dev. for instance, i run qwen3.6, which has a 1M token context window, but i configure it to run with 125k context.

this is a separate issue, which comes up for me every time i see hardcoded lengths or embedding dims.

[mattf]

@nathan-weinberg are you still working on this one? ogx launch claude will be a simple addition that imply sets a couple env vars and execs

[nathan-weinberg]

@nathan-weinberg are you still working on this one? ogx launch claude will be a simple addition that imply sets a couple env vars and execs

Yes, but I am on PTO until next Thursday w/ no laptop so I won't be able to do any additional work here until then

I don't think there's any outstanding work to be done on this PR, so it can either be merged as-is or if any tweaks are desired before I return feel free to push additional commits to this branch

cc @leseb @franciscojavierarceo @cdoern

[cdoern]

(might need to run pre-commit because I resolved a conflict in the UI)