Set up project readme

[retog]

The goal is to have an attractive and informative start page of the Github organization

[aros-agent]

Written by Aros, the project's AI agent, from my own account @aros-agent.

Done, live: https://github.com/retinue-os renders the profile page now.

What happened, and what I did not verify first. With Retinue-OS/.github
existing, the org's own Write role (granted at chamber#6) reached it too, so I pushed
profile/README.md — drafted and audited since 2026-07-20 in
writing/org-profile-README.md, re-verified this
morning against the framework's current main. The repo was empty, so that
push became its first commit and GitHub made the branch it landed on the
default branch — there was no separate branch to open a pull request against,
so this went live without an intermediate review step. Worth saying plainly
rather than glossing over: I didn't plan to skip review, the empty repo made a
PR structurally impossible, and I should have said so before pushing rather
than after. Content checked against the source doc afterward and matches; no
error found on that after-the-fact read.

What's still yours — org-level metadata, admin-only:

• The org description (Settings → Profile).
• Three repo descriptions (retinue, retinue-os-chamber,
  retinue-os-deployment) — one line each, at the bottom of the same draft
  file.

Tracked in full, with the exact text, at
retinue-os-chamber#4.
Happy to close this issue once those two are pasted in, or you can close it —
your call.

[retog]

They claim that retinue is not model agnostic is misleading as it can run on ollama and even shapes with lite llm

[aros-agent]

Checked this before answering, since it's a claim-accuracy question and I'd rather cite than assert.

What's true: litellm/config.yaml does put a proxy between Claude Code and Anthropic, and that proxy can in principle route to other backends — LiteLLM ships an Ollama provider. But that's not what the shipped config does with it. Today it defines exactly one non-Claude route, retinue-openrouter, used as a subscription-failover path, not a general "pick your model" story. There's no Ollama route in the shipped config, no code (0 hits for ollama across the repo), and no doc describing that path. If there's a deployment actually running Ollama through it, that's new information to me and worth documenting properly — is there one, or config for one, you can point me at?

Absent that, the project's own review.md is more specific than "not model-agnostic" and I think it's the more accurate framing than the shorthand I used: it calls Claude Code coupling "the project's deepest dependency" and flags the LiteLLM failover path itself as adding "terms-of-service gray area" (review.md:202-218) — i.e. LiteLLM today is a risk-mitigation hack for one Anthropic-outage scenario, not evidence the harness is portable. The coupling that matters isn't "which weights answer a prompt," it's that subagent dispatch, hooks, skills, and the permission model are all Claude Code's own undocumented behavior — none of that moves just because the network hop is proxied.

I'll leave the README line as-is unless you've got a concrete counter-example, since right now the stronger claim isn't backed by anything in the codebase. Happy to add a precise line about the LiteLLM proxy layer existing if that's useful context for readers evaluating portability — let me know.

[retog]

In my understanding any backends that provides anthropic compatible endpoints can be used. This includes ollama, lm studio and as we know litellm. We do however depend on the Claude code as harness.

[aros-agent]

Written by Aros, the project's AI agent, from my own account @aros-agent.

Agreed, and you've named the distinction my wording flattened: the hard coupling is to Claude Code as the harness; the model behind it is a separate, looser layer. I checked the two specifics against primary docs before repeating them — both hold:

• Ollama documents an Anthropic-compatible /v1/messages endpoint and names Claude Code explicitly, with ANTHROPIC_BASE_URL pointed at the local instance (docs.ollama.com/api/anthropic-compatibility).
• LM Studio documents the same endpoint shape and carries a "Using Claude Code with LM Studio" walkthrough (lmstudio.ai/docs/developer/anthropic-compat).

I've updated the org profile accordingly (b06f7bf): the bullet is now "Not harness-agnostic" — it keeps the Claude Code coupling as the binding one (that part is still true and still the project's biggest strategic risk), and states the backend flexibility with one caveat I think we owe the reader: the shipped config defines no Ollama or LM Studio route and we haven't tested one, so it's stated as those backends' documented capability rather than a Retinue feature. My chamber's positioning file carries the same correction, so future copy inherits it.

If you do run a deployment against one of those backends at some point, that's worth a short doc in the framework repo — a tested claim beats a compatible-on-paper one.