fix: inject reasoning_config into MoA aggregator extra_body (#64187)

[AlexFucuson9]

Summary

When provider: moa is configured, the acting aggregator never receives the user's reasoning_effort in extra_body. The aggregator silently runs at the backend's default reasoning level.

Root Cause

build_kwargs() gates reasoning emission on agent._supports_reasoning_extra_body(), which checks the agent's own provider / base_url. On the MoA path these are the virtual moa / moa://local, which always fail the gate. The aggregator's real provider is only resolved later in moa_loop.py::create() via _slot_runtime(), but by then the reasoning decision has already been made (and was False).

Fix

Three changes:

1. Stash reasoning_config on MoAClient — Pass agent.reasoning_config through MoAClient.__init__ → MoAChatCompletions.__init__ so create() has access to the user's configured reasoning settings.

2. Re-evaluate the gate against the resolved aggregator — In create(), after resolving the aggregator slot's real runtime, call the new _moa_slot_supports_reasoning() helper to check if the resolved provider supports reasoning. If so, inject extra_body[reasoning] with the configured effort level.

3. New helper _moa_slot_supports_reasoning() — Mirrors the gate logic in AIAgent._supports_reasoning_extra_body but evaluated against the resolved slot identity (Nous Portal, GitHub Models, LM Studio, OpenRouter with reasoning-capable model prefixes).

Testing

• ruff check agent/moa_loop.py agent/agent_init.py — all checks passed
• All 17 existing MoA tests pass
• The new helper is exercised indirectly through the existing test_moa_aggregator_cost_slot tests

Fixes #64187

[alt-glitch]

This was generated by AI during triage.

Duplicate of #64220 — both fix #64187 (MoA aggregator never receives the configured reasoning_effort because the reasoning gate keys on the virtual moa/moa://local identity) by re-evaluating the gate against the resolved aggregator slot in agent/moa_loop.py. #64220 is the earlier, cleaner canonical. Flagging for the maintainer to pick one.

[teknium1]

Closing — the underlying bug is fixed by PR #64756 (merged). Your diagnosis of #64187 was correct (reasoning gates keyed to the virtual moa://local identity never fire for the aggregator). Thanks @AlexFucuson9.

Your PR predates #64631 (per-slot MoA reasoning, merged), which moved this plumbing: reasoning now flows through call_llm(reasoning_config=...) with provider-profile projection per aggregator provider. The merged fix resolves the aggregator as: slot reasoning_effort > agent.reasoning_overrides for the slot's model > global agent.reasoning_effort.

Note: #64220 was a parallel fix for the same issue submitted shortly before yours — crediting both of you.