fix: avoid litellm provider-name collision for custom OpenAI-compatib…

[pikann]

Summary

Agents configured with a freeform/self-hosted LLM provider name crashed on every real invocation with:

litellm.APIConnectionError: sequence item 0: expected str instance, list found

Reported in #221, but the upstream-litellm default_pt() theory in that issue doesn't hold up — I reproduced the call path locally and the openai/<model> passthrough handles multi-block message content fine.

Root cause

build_llm() decided whether to route a provider through the OpenAI-compatible passthrough by checking provider not in litellm.provider_list. That list is litellm's internal provider registry, which reserves generic-sounding names — custom, vllm, huggingface, petals, etc. — for its own non-chat, non-OpenAI-compatible handlers.

A user setting up a self-hosted endpoint naturally types "custom" into Paca's freeform "Custom…" provider field. Since "custom" is already a registered litellm provider, build_llm() skipped the OpenAI passthrough and built model_str = "custom/<model>". litellm's native custom handler then does:

prompt = " ".join([message["content"] for message in messages])

which throws exactly the reported TypeError once any message has multi-block content (the normal shape once there's a system prompt + skills).

Fix

Route based on Paca's own provider catalog (data/llm_models.json, the same list backing the frontend provider dropdown) instead of litellm's internal provider_list. Any provider name outside that catalog — regardless of whether it happens to collide with a litellm-reserved name — now always gets the OpenAI-compatible passthrough.

• Extracted the catalog loader into src/llm_catalog.py, shared by routes/llm.py and agent/builder.py so both agree on what counts as "known."
• Added regression tests for build_llm() (previously untested), including the exact "custom" collision from Custom-provider agent LLM calls fail in build_llm() — litellm default_pt() crash on multi-block content (related to #214) #221.

Fixes #221

Test plan

• pytest — 85 passed, 3 skipped (unrelated e2e)
• ruff check clean
• Reproduced the original crash locally with litellm.completion(model="custom/...", ...) and confirmed the fix routes it through openai/... instead