fix(moa): apply prompt-caching decoration to the aggregator's one-shot synthesis call

agent/moa_loop.py

@@ -173,18 +173,19 @@ def _slot_runtime(slot: dict[str, str]) -> dict[str, Any]:
     return out
 
 
-def _maybe_apply_advisor_cache_control(
+def _maybe_apply_moa_cache_control(
     messages: list[dict[str, Any]],
     runtime: dict[str, Any],
 ) -> list[dict[str, Any]]:
-    """Decorate an advisor request with cache_control when its route honors it.
+    """Decorate an advisor or aggregator request with cache_control when its
+    route honors it.
 
     Reuses the SAME policy function as the main agent loop
-    (``anthropic_prompt_cache_policy``) resolved against the advisor slot's
-    own provider/base_url/api_mode/model, and the SAME breakpoint layout
-    (``apply_anthropic_cache_control``, system_and_3). This keeps advisor
-    calls decorated exactly like an acting agent on that provider would be —
-    no MoA-specific caching logic to drift.
+    (``anthropic_prompt_cache_policy``) resolved against the slot's own
+    provider/base_url/api_mode/model, and the SAME breakpoint layout
+    (``apply_anthropic_cache_control``, system_and_3). This keeps advisor and
+    aggregator calls decorated exactly like an acting agent on that provider
+    would be — no MoA-specific caching logic to drift.
 
     Returns the messages unchanged on any resolution error or when the
     policy says the route doesn't honor markers.
@@ -196,8 +197,8 @@ def _maybe_apply_advisor_cache_control(
         from agent.prompt_caching import apply_anthropic_cache_control
 
         # The policy function reads agent.* only as fallbacks for kwargs we
-        # don't pass; provide a stub so an advisor slot is judged purely on
-        # its own resolved runtime.
+        # don't pass; provide a stub so the slot is judged purely on its own
+        # resolved runtime.
         stub = SimpleNamespace(provider="", base_url="", api_mode="", model="")
         should_cache, native_layout = anthropic_prompt_cache_policy(
             stub,
@@ -212,7 +213,7 @@ def _maybe_apply_advisor_cache_control(
             messages, native_anthropic=native_layout
         )
     except Exception as exc:  # pragma: no cover - decoration must never break a call
-        logger.debug("advisor cache_control decoration skipped: %s", exc)
+        logger.debug("MoA cache_control decoration skipped: %s", exc)
         return messages
 
 
@@ -268,7 +269,7 @@ def _run_reference(
         # caching is opt-in per request. OpenAI-family advisors are untouched
         # (their caching is automatic; markers are ignored harmlessly, but we
         # only decorate when the policy says the route honors them).
-        messages = _maybe_apply_advisor_cache_control(messages, runtime)
+        messages = _maybe_apply_moa_cache_control(messages, runtime)
         response = call_llm(
             task="moa_reference",
             messages=messages,
@@ -617,13 +618,27 @@ def aggregate_moa_context(
     )
 
     agg_label = _slot_label(aggregator)
+    agg_runtime = _slot_runtime(aggregator)
     try:
+        # Same cache_control decoration as _run_reference's advisor calls
+        # (see _maybe_apply_moa_cache_control) — this synthesis call is a
+        # third, independent MoA call path that 22c5048d9 did not cover (it
+        # only restored caching for the acting-aggregator turn in the
+        # persistent `provider: moa` model and for advisor fan-out). Without
+        # it, the one-shot `/moa <prompt>` command's synthesis call re-bills
+        # its full input (system-less prompt containing every joined
+        # reference output) on every invocation with zero cache_control
+        # breakpoints, even when the resolved aggregator slot is a
+        # cache-honoring route (e.g. Claude on OpenRouter/native Anthropic).
+        agg_messages = _maybe_apply_moa_cache_control(
+            [{"role": "user", "content": synth_prompt}], agg_runtime
+        )
         response = call_llm(
             task="moa_aggregator",
-            messages=[{"role": "user", "content": synth_prompt}],
+            messages=agg_messages,
             temperature=aggregator_temperature,
             max_tokens=max_tokens,
-            **_slot_runtime(aggregator),
+            **agg_runtime,
         )
         synthesis = _extract_text(response)
     except Exception as exc:

tests/agent/test_moa_aggregator_cache_control.py

@@ -0,0 +1,114 @@
+"""Regression test: the MoA aggregator's one-shot synthesis call
+(``aggregate_moa_context``, used by the ``/moa <prompt>`` command) must get
+the same Anthropic-style prompt-caching decoration as the acting-aggregator
+turn (``MoAChatCompletions.create``) and the advisor fan-out
+(``_run_reference``).
+
+22c5048d9 ("fix(moa): restore prompt caching for the aggregator and
+advisors") fixed the other two MoA call paths but never touched
+``aggregate_moa_context`` — a third, independent call path with its own
+``call_llm(task="moa_aggregator", ...)`` invocation. Without this fix, every
+``/moa <prompt>`` one-shot call re-bills its full input (system-less prompt
+containing all joined reference outputs) with zero cache_control breakpoints,
+even when the resolved aggregator slot is a cache-honoring route.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+
+
+def _response(content="synthesized guidance"):
+    message = SimpleNamespace(content=content, tool_calls=[])
+    choice = SimpleNamespace(message=message, finish_reason="stop")
+    return SimpleNamespace(choices=[choice], usage=None, model="fake")
+
+
+@pytest.fixture
+def captured_calls(monkeypatch):
+    calls = []
+
+    def fake_call_llm(**kwargs):
+        calls.append(kwargs)
+        return _response()
+
+    monkeypatch.setattr("agent.moa_loop.call_llm", fake_call_llm)
+    monkeypatch.setattr(
+        "agent.moa_loop._run_references_parallel",
+        lambda *a, **k: [("advisor-a", "advice from a", None)],
+    )
+    return calls
+
+
+def _aggregator_kwargs(calls):
+    return next(c for c in calls if c.get("task") == "moa_aggregator")
+
+
+def test_aggregator_synthesis_gets_cache_control_on_native_anthropic_route(
+    captured_calls, monkeypatch
+):
+    """A cache-honoring aggregator slot (native Anthropic) must get
+    cache_control breakpoints on its synthesis call."""
+    from agent import moa_loop
+
+    monkeypatch.setattr(
+        moa_loop,
+        "_slot_runtime",
+        lambda slot: {
+            "provider": "anthropic",
+            "model": "claude-opus-4.8",
+            "base_url": "",
+            "api_mode": "anthropic_messages",
+        },
+    )
+
+    moa_loop.aggregate_moa_context(
+        user_prompt="what should I do next?",
+        api_messages=[{"role": "user", "content": "help me plan"}],
+        reference_models=[{"provider": "openrouter", "model": "openai/gpt-5.5"}],
+        aggregator={"provider": "anthropic", "model": "claude-opus-4.8"},
+    )
+
+    agg_kwargs = _aggregator_kwargs(captured_calls)
+    synth_message = agg_kwargs["messages"][0]
+    assert synth_message["role"] == "user"
+    content = synth_message["content"]
+    # Native Anthropic layout places cache_control on inner content blocks,
+    # so a cached message's content is a list of blocks rather than a bare
+    # string once decorated.
+    assert isinstance(content, list), "expected native cache_control block layout"
+    assert any(
+        isinstance(block, dict) and "cache_control" in block for block in content
+    ), "aggregator synthesis message must carry a cache_control breakpoint"
+
+
+def test_aggregator_synthesis_untouched_on_non_caching_route(
+    captured_calls, monkeypatch
+):
+    """A non-cache-honoring aggregator slot (plain OpenAI) must not be
+    decorated — proves the guard doesn't over-fire."""
+    from agent import moa_loop
+
+    monkeypatch.setattr(
+        moa_loop,
+        "_slot_runtime",
+        lambda slot: {
+            "provider": "openai",
+            "model": "gpt-5.5",
+            "base_url": "",
+            "api_mode": "chat_completions",
+        },
+    )
+
+    moa_loop.aggregate_moa_context(
+        user_prompt="what should I do next?",
+        api_messages=[{"role": "user", "content": "help me plan"}],
+        reference_models=[{"provider": "openrouter", "model": "openai/gpt-5.5"}],
+        aggregator={"provider": "openai", "model": "gpt-5.5"},
+    )
+
+    agg_kwargs = _aggregator_kwargs(captured_calls)
+    synth_message = agg_kwargs["messages"][0]
+    assert isinstance(synth_message["content"], str), "must stay undecorated (plain string content)"