feat(opencode): add LiteLLM provider with auto model discovery

[balcsida]

Issue for this PR

Closes #13891

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Adds a native litellm provider that auto-discovers models from a LiteLLM proxy at startup. Previously users had to manually define every model in opencode.json — now setting LITELLM_API_KEY and LITELLM_HOST is enough.

Discovery: Fetches /model/info for rich metadata (pricing, limits, capabilities). Falls back to the standard /models endpoint for older LiteLLM versions or non-LiteLLM OpenAI-compatible proxies.

What gets mapped from /model/info:

• Per-token pricing → per-million-token (including over-200k tier)
• Context/output limits
• Capabilities: reasoning, vision, PDF, audio, video, tool calling
• Temperature support via supported_openai_params
• Interleaved thinking inference for Claude/Anthropic models

Reasoning transforms: Claude models behind LiteLLM get thinking budget variants, other models get reasoningEffort. This prevents false-positive reasoning param injection for aliased models (e.g. o3-custom → Mistral).

Files changed:

• litellm.ts (new): Discovery module — fetches, parses, maps model metadata
• provider.ts: Seeds litellm provider from env vars; custom loader calls discovery and injects models (user config takes precedence)
• transform.ts: LiteLLM-specific reasoning variant logic
• llm.ts: Adds explicit providerID === "litellm" to proxy detection

Env var	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 extra headers	{}
LITELLM_TIMEOUT	Discovery timeout in ms	5000

Builds on ideas from #13896 (fallback endpoint, over-200k pricing, temperature detection) and #14277 (clean fallback chain). Supersedes my earlier #14202 which was auto-closed for not following the PR template.

How did you verify your code works?

1. Set LITELLM_API_KEY + LITELLM_HOST pointing to a running LiteLLM 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 Claude reasoning model — thinking variants work
6. Test without proxy running — graceful fallback, provider hidden
7. Test with older proxy (no /model/info) — falls back to /models with sensible defaults

Screenshots / recordings

N/A — no UI changes

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

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

Based on my search, I found two related PRs that are NOT duplicates but are closely related:

1. PR feat(opencode): add dynamic model fetching for OpenAI-compatible provider #14277: feat(opencode): add dynamic model fetching for OpenAI-compatible provider

   • Related because the current PR builds on ideas from this PR (clean fallback chain mentioned in description)

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

   • Related because the current PR explicitly builds on ideas from this PR (fallback endpoint, over-200k pricing, temperature detection mentioned in description)

These PRs appear to be related work in the same area but are separate features. PR #14468 (the current PR) is a distinct, more comprehensive implementation of the LiteLLM provider with auto-discovery, and it references these earlier related PRs rather than duplicating them.

No duplicate PRs found.

[alexyaroshuk]

lgtm, but test fails. check e2e. this test fails:
e2e/projects/projects-switch.spec.ts:49:1 › switching back to a project opens the latest workspace session

Did you try testing locally? bun test:e2e:local