feat: add LiteLLM provider with auto model discovery

[balcsida]

Closes #13891

What

Adds native LiteLLM provider support with automatic model discovery from the proxy's /model/info endpoint. Previously, users had to manually define every model in opencode.json when using a LiteLLM proxy. Now, setting LITELLM_API_KEY and LITELLM_HOST is enough — models are discovered automatically at startup.

Changes

• New litellm.ts: Discovery module that fetches /model/info, maps pricing from per-token to per-million-token, extracts capabilities (reasoning, vision, PDF, audio, video, tool calling), filters wildcards, and infers interleaved thinking for Claude models
• provider.ts: Seeds litellm provider when env vars are detected; adds custom loader that resolves host/key, calls discovery, and injects models (user config takes precedence)
• transform.ts: LiteLLM-specific reasoning variants — Claude/Anthropic models get thinking budget variants, others get reasoningEffort. Prevents false-positive reasoning param injection for aliased models
• llm.ts: Adds explicit providerID === "litellm" to isLiteLLMProxy check

Why not models.dev?

LiteLLM is a self-hosted proxy where available models are user-configured and vary per deployment. Unlike static providers (OpenRouter, Helicone), the model list can only be known at runtime. A models.dev entry can't represent this — runtime discovery is required.

Env vars

Variable	Description	Default
LITELLM_API_KEY	API key for the proxy	—
LITELLM_HOST	Base URL of the proxy	http://localhost:4000
LITELLM_BASE_URL	Alias for LITELLM_HOST	—
LITELLM_CUSTOM_HEADERS	JSON string of custom headers	{}
LITELLM_TIMEOUT	Discovery timeout in ms	5000

Verification

1. Set LITELLM_API_KEY + LITELLM_HOST pointing to a running proxy
2. Run opencode — LiteLLM models appear in model picker
3. Select a model and send a message — streaming works
4. Test aliased reasoning model (e.g. o3-custom → Mistral) — reasoning params NOT injected
5. Test actual reasoning model — reasoning variants work
6. Test without proxy running — graceful fallback, provider hidden

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Duplicate PR Found

PR #13896: "feat(opencode): add auto loading models for litellm providers"
#13896

This appears to be a duplicate or closely related PR addressing the same feature — automatic model loading for LiteLLM providers. Both PRs are implementing first-class LiteLLM support with auto-discovery. You should check if #13896 is still open/merged and coordinate with that work to avoid duplication.

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.