Add Ollama Cloud support via managed sidecar container

[nezhar]

Problem

VibePod's LLM integration works by injecting ANTHROPIC_BASE_URL / CODEX_OSS_BASE_URL into agent containers, expecting an Anthropic- or OpenAI-compatible endpoint at that URL.

However, ollama.com's cloud API only exposes its native endpoint (/api/chat with Authorization: Bearer) — not the /v1/messages (Anthropic) or /v1/chat/completions (OpenAI) compatibility layers. Those are features of the local Ollama server process, not the hosted service.

This means users who want to use Ollama cloud models today must run a local Ollama instance, defeating the purpose of a cloud-backed workflow.

Proposed Solution: Ollama Sidecar Container

Add a managed ollama container to VibePod's stack (alongside proxy and datasette). The sidecar runs Ollama locally but uses :cloud-suffixed model names to offload inference to ollama.com. Agent containers hit the sidecar's compatibility endpoints as normal.

Config:

# ~/.config/vibepod/config.yaml
llm:
  enabled: true
  base_url: "http://vibepod-ollama:11434"
  api_key: "ollama"
  model: "gpt-oss:120b-cloud"

Usage:

# Sidecar starts automatically, OLLAMA_API_KEY is passed through
VP_LLM_ENABLED=true VP_LLM_MODEL=gpt-oss:120b-cloud vp run claude

This fits the existing image namespace pattern:

ollama -> ollama/ollama:latest

With an optional env override:

VP_IMAGE_OLLAMA=ollama/ollama:latest vp run claude

Why this approach

• No changes needed to vibepod-proxy or agent containers
• Both claude and codex agents work immediately (Anthropic + OpenAI compat layers are exposed on the sidecar)
• Users get cloud inference without installing Ollama on the host
• OLLAMA_API_KEY is the only credential needed
• Lightweight — the sidecar itself does no local inference for :cloud models

Alternative Considered: Proxy Translation Layer

Extend vibepod-proxy with a mitmproxy addon that intercepts /v1/messages or /v1/chat/completions calls and rewrites them to Ollama's native https://ollama.com/api/chat format. More powerful (no sidecar needed, everything cloud-native), but significantly more complex to implement correctly, especially for streaming/SSE responses.

Could be pursued as a follow-up.

Implementation Notes

• The sidecar needs OLLAMA_API_KEY injected so cloud model auth works
• Use host.docker.internal pattern already established in the LLM docs for Docker networking
• vp run ollama or auto-start as a dependency when llm.enabled: true and base_url points to the sidecar