Feat/ollama provider

[Aneesh-382005]

Summary

The codebase already acknowledged this gap in a comment in rust/crates/api/src/providers/mod.rs:

// When OPENAI_BASE_URL is set, the user explicitly configured an
// OpenAI-compatible endpoint. Prefer it over the Anthropic fallback
// even when the model name has no recognized prefix — this is the
// common case for local providers (Ollama, LM Studio, vLLM, etc.)
// where model names like "qwen2.5-coder:7b" don't match any prefix.

The routing logic beneath it was correct. But two layers upstream were
still blocking local model names before routing ever ran.

Before this PR, the intended workaround was:

export OPENAI_BASE_URL="http://127.0.0.1:11434/v1"
export OPENAI_API_KEY="anything"
claw --model "openai/qwen2.5-coder:7b" prompt "say hello"

This required three env vars, a fake API key, and a provider/ prefix
that Ollama doesn't understand, and it still broke due to a logic bug in
wire_model_for_base_url where an inner matches! could never fire,
causing the full openai/qwen2.5-coder:7b slug to reach Ollama's API
instead of the bare model name.

After this PR:

export OLLAMA_HOST="http://127.0.0.1:11434"
claw --model "qwen2.5-coder:7b" prompt "say hello"

One env var. No prefix. No fake key. Matches Ollama's own convention.

What was blocking it

Two separate layers rejected local model names before routing logic ran:

1. validate_model_syntax() in rusty-claude-cli/src/main.rs enforced
   provider/model format on every model name unconditionally.

2. detect_provider_kind() in providers/mod.rs matched the qwen
   prefix and routed to DashScope regardless of what endpoint was configured.

What this PR changes

• rusty-claude-cli/src/main.rs - if OLLAMA_HOST is set, skip the
  provider/model format requirement. Empty strings still rejected.
• providers/mod.rs - if OLLAMA_HOST is set, skip prefix matching
  and route directly to the OpenAI-compatible backend.
• providers/openai_compat.rs - adds OLLAMA_CONFIG constant and
  from_ollama_env() constructor that reads OLLAMA_HOST, appends /v1,
  and sets a dummy api_key since Ollama requires no auth.
• client.rs - in the OpenAi dispatch arm, when OLLAMA_HOST is set,
  use from_ollama_env() instead of the standard constructor.
• docs/local-openai-compatible-providers.md and USAGE.md updated to
  document OLLAMA_HOST as the preferred Ollama path.

Follow-up work not in this PR

• ProviderKind::Ollama as a dedicated variant so /doctor shows
  "Ollama" as the provider name rather than "OpenAI"
• Same pattern for LM Studio and llama.cpp
• Tool-call graceful degradation for models that don't support
  function calling

Anti-slop triage

• Classification: actionable-fix
• Evidence: claw --model "qwen2.5-coder:7b" prompt "x" returns
  invalid model syntax on unpatched main. With this patch and
  OLLAMA_HOST=http://127.0.0.1:11434, qwen2.5-coder:7b responds
  correctly. Regression test added in alias_resolution_tests.
• Non-destructive review result: merge candidate

Verification

• Targeted tests/docs checks ran, or the gap is explicitly recorded.
• git diff --check passes.
• No live secrets, tokens, private logs, or unrelated generated churn are included.

Resolution gate

• If this PR resolves an issue, the issue number and fix evidence are linked.
• If this PR should not merge, the rejection/defer rationale is evidence-backed and does not rely on vibes.
• I did not merge/close remote PRs or issues from an automation lane without owner approval.

[Yeachan-Heo]

Thanks for the contribution! The Ollama provider approach is solid — using OLLAMA_HOST to route and bypass colon-containing model names makes sense.

However, the diff in crates/api/src/providers/mod.rs introduces a broken merge: the added if let Some(metadata) = metadata_for_model(model) line creates a dangling unclosed block right before the existing let resolved_model = ... code. This won't compile.

Could you rebase this on the latest main and fix the conflict in detect_provider_kind? The OLLAMA_HOST early-return should come before the existing let resolved_model line, and the original if let Some(metadata) = metadata_for_model(&resolved_model) should remain intact after it.

Once the diff is clean I'm happy to merge.

[Yeachan-Heo]

Superseded by commit be8112f which adds native Ollama support via env var. This PR's approach was sound but had a duplicate bug in . The merged implementation uses as the preferred single env var, bypasses model syntax validation for Ollama names, and requires no API key.

[Aneesh-382005]

Thanks for reviewing and cleaning this up. I'll follow up with generic local provider support for all sorts of frameworks.
LM Studio, llama.cpp, and vLLM, so bare model names work across all compatible local servers, not just Ollama

[1716775457damn]

Nice work on the Ollama integration. The approach of using OLLAMA_HOST to bypass colon-based model name parsing is clean. For the follow-up on generic local provider support across LM Studio/llama.cpp/vLLM, would it make sense to also consider a config-based provider registry so users can add custom endpoints without code changes? That would make the system extensible beyond the built-in providers.

[Aneesh-382005]

Nice work on the Ollama integration. The approach of using OLLAMA_HOST to bypass colon-based model name parsing is clean. For the follow-up on generic local provider support across LM Studio/llama.cpp/vLLM, would it make sense to also consider a config-based provider registry so users can add custom endpoints without code changes? That would make the system extensible beyond the built-in providers.

That's the direction I am thinking as well. A config based registry makes more sense than handling different Local LLM frameworks as edge cases. I'm planning to work on this.