From 2962ba2b7bb3bf518036aa31ec8e67a2825eb6e3 Mon Sep 17 00:00:00 2001
From: srojk34 <286497132+srojk34@users.noreply.github.com>
Date: Wed, 1 Jul 2026 14:01:34 +0300
Subject: [PATCH] fix(auxiliary): treat explicit model:auto sentinel, not just
 cfg_model

'auto' is a sentinel meaning "inherit from main runtime / auto-detect",
not a literal model id -- already handled for cfg_model (config-derived)
in _resolve_task_provider_model, but not for the explicit `model` kwarg.

MoA reference/aggregator slots (agent/moa_loop.py's _slot_runtime) forward
a preset's `model:` field as this explicit argument rather than through
auxiliary.<task> config, so a MoA preset configured with `model: auto`
(a natural thing to try given the existing auxiliary.*.model: auto
convention) reached this function as the explicit `model` arg and took
the `model or cfg_model` branch, bypassing the cfg_model-only sentinel
check entirely -- sending the literal string "auto" to the wire as a
model id.

Normalize both the explicit `model` and `cfg_model` the same way, fixing
this at the single chokepoint every caller (MoA included) already goes
through, rather than patching moa_loop.py separately.
---
 agent/auxiliary_client.py            |  9 +++++
 tests/agent/test_auxiliary_client.py | 49 ++++++++++++++++++++++++++++
 2 files changed, 58 insertions(+)

diff --git a/agent/auxiliary_client.py b/agent/auxiliary_client.py
index 7cb0849ea220..c678bc2a3796 100644
--- a/agent/auxiliary_client.py
+++ b/agent/auxiliary_client.py
@@ -6369,6 +6369,15 @@ def _resolve_task_provider_model(
     # which downstream consumers like ContextCompressor accept as the task output.
     # The provider-side 'auto' is handled in _resolve_auto() via main_runtime
     # fallback, so dropping cfg_model to None here lets that path do its job.
+    #
+    # The explicit `model` kwarg needs the identical normalization: MoA slots
+    # (agent/moa_loop.py's _slot_runtime) forward a preset's `model:` field as
+    # this explicit argument rather than through auxiliary.<task> config, so a
+    # user-configured `model: auto` on a MoA reference/aggregator slot reaches
+    # this function here, not as cfg_model. Only normalizing cfg_model let that
+    # literal "auto" slip through via `model or cfg_model` below.
+    if model and model.lower() == "auto":
+        model = None
     if cfg_model and cfg_model.lower() == "auto":
         cfg_model = None
 
diff --git a/tests/agent/test_auxiliary_client.py b/tests/agent/test_auxiliary_client.py
index cdc34955e857..9f61ff9668db 100644
--- a/tests/agent/test_auxiliary_client.py
+++ b/tests/agent/test_auxiliary_client.py
@@ -394,6 +394,55 @@ class TestResolveTaskProviderModel:
         assert resolved_provider == "anthropic"
         assert model == "claude-haiku-4-5-20251001"
 
+    def test_explicit_model_auto_sentinel_is_normalized(self):
+        """MoA slots (agent/moa_loop.py's _slot_runtime) forward a preset's
+        `model:` field as the explicit `model` kwarg here, not through
+        auxiliary.<task> config. Only cfg_model was normalized before, so a
+        MoA reference/aggregator slot configured with `model: auto` sent the
+        literal string "auto" to the wire as a model id."""
+        resolved_provider, model, base_url, api_key, api_mode = _resolve_task_provider_model(
+            provider="anthropic",
+            model="auto",
+        )
+
+        assert resolved_provider == "anthropic"
+        assert model is None
+
+    def test_explicit_model_auto_sentinel_case_insensitive(self):
+        resolved_provider, model, base_url, api_key, api_mode = _resolve_task_provider_model(
+            provider="anthropic",
+            model="AUTO",
+        )
+
+        assert model is None
+
+    def test_explicit_model_auto_falls_back_to_cfg_model(self, monkeypatch):
+        """When the explicit model is the "auto" sentinel, it must not shadow
+        a real configured task model — the `model or cfg_model` fallback
+        chain should still reach cfg_model exactly as if model had been
+        omitted entirely."""
+        monkeypatch.setattr(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            lambda _task: {"provider": "openai", "model": "gpt-real-model"},
+        )
+
+        resolved_provider, model, base_url, api_key, api_mode = _resolve_task_provider_model(
+            task="moa_reference",
+            model="auto",
+        )
+
+        assert model == "gpt-real-model"
+
+    def test_non_auto_model_is_unaffected(self):
+        """Regression guard: a real model name must not be touched by the
+        sentinel normalization."""
+        resolved_provider, model, base_url, api_key, api_mode = _resolve_task_provider_model(
+            provider="openai",
+            model="gpt-4o-mini",
+        )
+
+        assert model == "gpt-4o-mini"
+
 
 class TestMoaAggregatorSharedResolution:
     """The shared MoA→aggregator helper and the layers that consume it.