Refactor Agent Registry Wiring and Fix Config Options

[jmchilton]

Wire agent registry using our Lagom container instead of module-level singletons - pros and cons of course but this is the documented Galaxy best practice and I think has some real important upsides. Also fixes #21996 while I'm in there.

How to test the changes?

(Select all options that apply)

• This is a refactoring of components with existing test coverage.

License

• I agree to license these and all my past contributions to the core galaxy codebase under the MIT license.

[jmchilton]

I started working on something unrelated and realized this went astray - config.agents was dead code laying around that should have been cleaned up and we should have attached this stuff to inference_services.

From the agent:

What agent_plumbing does (3 commits)

Good structural improvements:

1. Removes global singleton agent_registry — replaced with build_default_registry() factory function, registry injected via Lagom DI in app.py
2. Makes get_agent non-optional on GalaxyAgentDependencies — removes None checks, orchestrator no longer does from galaxy.agents import agent_registry (circular import smell removed)
3. Adds per-agent enabled flag — build_default_registry(config) reads config.agents to skip disabled agents, with "router always on" guard
4. Removes HAS_AGENTS try/except guards in managers/agents.py and api/agents.py — treats pydantic-ai as a hard dependency
5. Good test coverage — 7 new tests for registry building, disabled agents, error messages

The configuration problem

Yes, the branch went the wrong direction on config. It doubled down on config.agents (the dead config dict) for the enabled flag:

• build_default_registry() reads config.agents[agent_type]["enabled"] to decide registration
• AgentService.execute_agent() also checks config.agents as a guard
• api/agents.py list endpoint checks config.agents for enabled status

But the actual runtime config that agents use for model/temperature/api_key is config.inference_services (read in base.py:_get_agent_config()). So now you have:

Concern	Read from
model, temperature, api_key, api_base_url	inference_services
enabled	agents

Two config dicts for the same logical concept. An admin would need to configure both:

inference_services:
  custom_tool:
    model: "anthropic:claude-sonnet-4-5"
agents:
  custom_tool:
    enabled: false  # different dict!

What should probably happen

The enabled flag should live in inference_services alongside model/temperature, and config.agents / _load_agent_config() should be removed entirely. Something like:

inference_services:
  default:
    model: "gpt-4o-mini"
    temperature: 0.7
  custom_tool:
    enabled: false  # same place as model config

Then build_default_registry() and _get_agent_config() both read from inference_services — one config to rule them all.

The structural DI improvements in this branch (killing the global singleton, non-optional get_agent, removing HAS_AGENTS guards) are all good and worth keeping. It's just the config source that needs redirecting.