bug: synthesis provider reuses DEWEY_EMBEDDING_ENDPOINT instead of its own env var

[yvonnedevlinrh]

Summary

llm/config.go reuses the DEWEY_EMBEDDING_ENDPOINT environment variable to resolve the Ollama endpoint for the synthesis provider (used by dewey compile and knowledge curation). This conflates two independent concerns - embedding and text generation - and makes it impossible to point them at different Ollama instances via env vars.

Root Cause

llm/config.go reads DEWEY_EMBEDDING_ENDPOINT in two places:

• ReadSynthesisConfig() line 84 (legacy DEWEY_COMPILE_MODEL path)
• synthConfigFromEnv() line 116 (no-config-file fallback)

Both fall back to http://localhost:11434 when unset. There is no DEWEY_SYNTHESIS_ENDPOINT or equivalent - the synthesis provider silently shares whatever endpoint the embedding provider uses.

From llm/config.go:116-119:

endpoint := os.Getenv("DEWEY_EMBEDDING_ENDPOINT")
if endpoint == "" {
    endpoint = "http://localhost:11434"
}

Impact

• Users running embedding and generation on separate Ollama instances (e.g., different GPUs, different machines) cannot configure them independently via env vars.
• The shared env var name is misleading - setting DEWEY_EMBEDDING_ENDPOINT silently changes synthesis behavior with no indication.
• If Bug: Dewey does not read OLLAMA_HOST and exits when embedding model is missing #61 adds OLLAMA_HOST fallback to the embedding path, the synthesis path should get the same treatment, but it also needs its own override to avoid permanently coupling the two.

Suggested Fix

Introduce a DEWEY_SYNTHESIS_ENDPOINT env var with the fallback chain:

1. DEWEY_SYNTHESIS_ENDPOINT (app-specific override)
2. OLLAMA_HOST (ecosystem standard, once Bug: Dewey does not read OLLAMA_HOST and exits when embedding model is missing #61 lands)
3. http://localhost:11434 (default)

The config.yaml synthesis.endpoint field already exists and works correctly - this is only about the env var path.

Related

• Bug: Dewey does not read OLLAMA_HOST and exits when embedding model is missing #61 - OLLAMA_HOST not read + hard exit on missing model (addresses the embedding side)