Per-node local/edge LLM endpoints not configurable; local-ollama preset broken; JSONC parser corrupts URL strings

[RanjithRagavan]

Summary

Running ARTEMIS against local / edge OpenAI-compatible endpoints (Ollama, llama.cpp-server, vLLM, LM Studio) per agent node is currently not configurable, and the shipped local-ollama preset is broken at startup. This blocks the roadmap item On-Device Lightweight VLMs (README §Roadmap).

Details

ModelFactory already routes ollama / vllm / custom providers to ChatOpenAI (artemis/llm/router.py, ~L369–386), but the config layer cannot reach it:

1. Schema gap — LLM (artemis/config/llm.py L71–105) has no api_base, api_key, max_tokens, timeout_seconds, or is_multimodal fields, even though ModelEndpoint (router.py L87–133) already defines them.
2. Passthrough gap — _resolve_endpoint() (artemis/services/llm.py ~L1043–1089) builds ModelEndpoint from the per-node config but silently drops api_base / api_key / max_tokens. The only way to reach a local server today is the global OPENAI_BASE_URL env var, which forces every agent node onto the same endpoint. A hybrid setup (local VLM for the flash explorer tier, cloud model for the planner) is impossible.
3. Broken preset — the local-ollama preset in config/artemis.jsonc (L44–48) uses provider: "openai" with no api_base, and validate_provider() raises without OPENAI_API_KEY, so the one preset advertising local usage fails at startup.
4. Latent parser bug — strip_json_comments() (artemis/utils/file.py) uses a regex that also strips // inside quoted strings, so any artemis.jsonc containing a URL value like "api_base": "http://localhost:11434/v1" gets corrupted at load time. This blocks any config-level fix for the above.

Relation to existing work

PR #128 covers admin-UI/env credential plumbing (CUSTOM_LLM_API_KEY etc.) and validate_provider branches for custom providers — complementary to this. It does not touch _resolve_endpoint() per-node passthrough, the preset, or the JSONC parser bug.

Proposed fix

• Add the five optional endpoint fields to the LLM schema and forward them in _resolve_endpoint() (primary and fallback nodes).
• Let validate_provider() skip cloud-key requirements for ollama / vllm / custom.
• Repair the local-ollama preset and document local-endpoint variables in .env.example.
• Rewrite strip_json_comments() as a string-aware scanner.
• Add provider-surface unit tests (tests/unit/test_llm_router.py).

I have this implemented and will open a PR shortly.