The OpenAI Custom Server request targets the per-account OpenAI integration
(/app/accounts/{id}/settings/integrations/openai), not the instance-level
Captain config. The CAPTAIN_OPEN_AI_PROVIDER provider-switch added in 46fe70398
was out of the intended scope and is new behavior that did not previously exist,
so it is removed. This restores the original Captain behavior (CAPTAIN_OPEN_AI_ENDPOINT
remains usable as before, no provider gating). The correct implementation goes in
config/integration/apps.yml for the per-account openai integration.
Add an operator-level provider switch for Captain AI so it can use any
OpenAI-compatible base URL (vLLM, OpenRouter, proxies, etc.) in addition to the
default api.openai.com endpoint.
- installation_config.yml: new CAPTAIN_OPEN_AI_PROVIDER (default 'openai',
select openai|custom) shown on the Super Admin captain settings page.
- lib/llm/config.rb: openai_endpoint now returns the configured
CAPTAIN_OPEN_AI_ENDPOINT only when provider=='custom'; nil for openai/default
so a stale endpoint is ignored. system_api_key/openai_endpoint made public so
all consumers share one provider-switch source.
- Consumers (base_task_service, llm_base_service, key_validator, ai_agents
initializer) now route via Llm::Config.openai_endpoint instead of reading
CAPTAIN_OPEN_AI_ENDPOINT directly.
- installation_config.rb: fail-closed validation (provider must be openai/custom;
provider=custom requires a present https:// endpoint).
Approved by independent five-key reviews deleg_b141098a (found+fixed one logic
gap) and deleg_3088b922 (passed=true, blocking arrays empty).
RubyLLM bundles a static models.json that doesn't know about models
released after the gem was published. Self-hosted users configuring
newer models hit ModelNotFoundError.
Added a rake task that refreshes the registry from models.dev and saves
to disk. ~~Called during Docker image build so every deploy gets fresh
model data. Falls back silently to the bundled registry if models.dev is
unreachable.~~
Commit the models.json file to code so it is available across
deployments.
---------
Co-authored-by: Sojan Jose <sojan@pepalo.com>
