chore(release): add AUTHOR_MAP entry for 2001Y (#54065 salvage)

agent/moa_loop.py

@@ -69,6 +69,17 @@ def _slot_label(slot: dict[str, str]) -> str:
     return f"{slot.get('provider', '').strip()}:{slot.get('model', '').strip()}"
 
 
+def _reference_system_prompt(slot: dict[str, str]) -> str:
+    role_prompt = str(slot.get("role_prompt") or "").strip()
+    if not role_prompt:
+        return _REFERENCE_SYSTEM_PROMPT
+    return (
+        f"{_REFERENCE_SYSTEM_PROMPT}\n\n"
+        "Reference-specific role instruction for this model:\n"
+        f"{role_prompt}"
+    )
+
+
 def _slot_runtime(slot: dict[str, str]) -> dict[str, Any]:
     """Resolve a reference/aggregator slot to real runtime call kwargs.
 
@@ -145,7 +156,7 @@ def _run_reference(
         # it is analyzing state for an aggregator, not acting on the task. The
         # trimmed view (_reference_messages) already strips the agent's own
         # system prompt, so this is the only system message the reference sees.
-        messages = [{"role": "system", "content": _REFERENCE_SYSTEM_PROMPT}, *ref_messages]
+        messages = [{"role": "system", "content": _reference_system_prompt(slot)}, *ref_messages]
         response = call_llm(
             task="moa_reference",
             messages=messages,
@@ -506,7 +517,14 @@ class MoAChatCompletions:
                 f"{m.get('role')}:{m.get('content')}" for m in ref_messages
             ).encode("utf-8", "replace")
         ).hexdigest()
-        _cache_key = (self.preset_name, _sig, tuple(_slot_label(s) for s in reference_models))
+        _cache_key = (
+            self.preset_name,
+            _sig,
+            tuple(
+                (_slot_label(s), str(s.get("role_prompt") or "").strip())
+                for s in reference_models
+            ),
+        )
         _refs_from_cache = _cache_key == self._ref_cache_key and bool(self._ref_cache_outputs)
 
         if _refs_from_cache:

hermes_cli/moa_cmd.py

@@ -57,7 +57,14 @@ def _pick_slot(current: dict[str, str] | None = None) -> dict[str, str]:
     current_model = (current or {}).get("model", "")
     model_default = models.index(current_model) if current_model in models else 0
     model = models[_prompt_choice(f"Select model for {provider.get('slug')}", models, model_default)]
-    return {"provider": str(provider.get("slug") or ""), "model": str(model)}
+    slot = {"provider": str(provider.get("slug") or ""), "model": str(model)}
+    # The interactive picker edits only provider/model today. Preserve an
+    # existing per-reference role prompt so a quick model re-save does not erase
+    # richer config.yaml/dashboard-authored MoA roles.
+    role_prompt = str((current or {}).get("role_prompt") or "").strip()
+    if role_prompt:
+        slot["role_prompt"] = role_prompt
+    return slot
 
 
 def _print_config(config: dict[str, Any]) -> None:
@@ -72,6 +79,9 @@ def _print_config(config: dict[str, Any]) -> None:
         print("  Reference models:")
         for idx, slot in enumerate(preset["reference_models"], start=1):
             print(f"    {idx}. {slot['provider']}:{slot['model']}")
+            role_prompt = str(slot.get("role_prompt") or "").strip()
+            if role_prompt:
+                print(f"       role: {role_prompt}")
         agg = preset["aggregator"]
         print(f"  Aggregator: {agg['provider']}:{agg['model']}")
 

hermes_cli/moa_config.py

@@ -42,7 +42,7 @@ def _coerce_int(value: Any, default: int) -> int:
             return default
 
 
-def _clean_slot(slot: Any) -> dict[str, str] | None:
+def _clean_slot(slot: Any, *, preserve_role_prompt: bool = False) -> dict[str, str] | None:
     if not isinstance(slot, dict):
         return None
     provider = str(slot.get("provider") or "").strip()
@@ -56,7 +56,12 @@ def _clean_slot(slot: Any) -> dict[str, str] | None:
     # an invalid slot is dropped, falling back to the preset's defaults.
     if provider.lower() == "moa":
         return None
-    return {"provider": provider, "model": model}
+    clean = {"provider": provider, "model": model}
+    if preserve_role_prompt:
+        role_prompt = str(slot.get("role_prompt") or "").strip()
+        if role_prompt:
+            clean["role_prompt"] = role_prompt
+    return clean
 
 
 def _default_preset() -> dict[str, Any]:
@@ -80,7 +85,7 @@ def _normalize_preset(raw: Any) -> dict[str, Any]:
         # defaults instead of crashing the iteration, mirroring the tolerance
         # for the scalar fields below (reference_temperature / max_tokens).
         raw_refs = [raw_refs] if isinstance(raw_refs, dict) else []
-    refs = [_clean_slot(item) for item in raw_refs]
+    refs = [_clean_slot(item, preserve_role_prompt=True) for item in raw_refs]
     refs = [item for item in refs if item is not None]
     if not refs:
         refs = deepcopy(DEFAULT_MOA_REFERENCE_MODELS)

hermes_cli/web_server.py

@@ -868,6 +868,10 @@ class ModelAssignment(BaseModel):
 class MoaModelSlot(BaseModel):
     provider: str = ""
     model: str = ""
+    # Optional per-reference guidance. Aggregator slots accept the field in the
+    # API payload for round-trip compatibility, but normalize_moa_config drops
+    # it for aggregators because only reference calls consume role prompts.
+    role_prompt: str = ""
 
 
 class MoaPresetPayload(BaseModel):

scripts/release.py

@@ -45,6 +45,7 @@ ACP_REGISTRY_MANIFEST = REPO_ROOT / "acp_registry" / "agent.json"
 
 # Auto-extracted from noreply emails + manual overrides
 AUTHOR_MAP = {
+    "15167896+2001Y@users.noreply.github.com": "2001Y",  # PR #54065 salvage (per-reference role_prompt for MoA reference models)
     "186512915+lEWFkRAD@users.noreply.github.com": "lEWFkRAD",  # PR #53848 salvage (stream the MoA aggregator response to the user)
     "193368749+jimmyjohansson84@users.noreply.github.com": "jimmyjohansson84",  # PR #27123 salvage (Kanban unknown-skill warn-instead-of-crash; #27136)
     "gxalong@gmail.com": "Jeffgithub0029",  # PR #28558 salvage (chunk Telegram text *after* MarkdownV2/HTML formatting so escaping inflation can't push a send over the 4096 UTF-16 limit; #28557)

tests/hermes_cli/test_moa_cmd.py

@@ -0,0 +1,59 @@
+from types import SimpleNamespace
+
+
+def test_cmd_moa_configure_preserves_reference_role_prompt(monkeypatch, tmp_path, capsys):
+    """Interactive configure edits provider/model but must not erase role prompts."""
+    home = tmp_path / ".hermes"
+    home.mkdir()
+    (home / "config.yaml").write_text(
+        """
+moa:
+  default_preset: review
+  presets:
+    review:
+      reference_models:
+        - provider: openrouter
+          model: x-ai/grok-4.3
+          role_prompt: Find real-world edge cases.
+      aggregator:
+        provider: openrouter
+        model: openai/gpt-5.5
+""".strip(),
+        encoding="utf-8",
+    )
+    monkeypatch.setenv("HERMES_HOME", str(home))
+
+    from hermes_cli import moa_cmd
+    from hermes_cli.config import load_config
+
+    monkeypatch.setattr(
+        moa_cmd,
+        "_model_options",
+        lambda: [
+            {
+                "slug": "openrouter",
+                "name": "OpenRouter",
+                "models": ["x-ai/grok-4.3", "openai/gpt-5.5"],
+            }
+        ],
+    )
+
+    def fake_prompt(title, rows, default=0):
+        if title == "Add another reference model?":
+            return 1
+        return default
+
+    monkeypatch.setattr(moa_cmd, "_prompt_choice", fake_prompt)
+
+    moa_cmd.cmd_moa(SimpleNamespace(moa_command="configure", name="review"))
+
+    cfg = load_config()
+    refs = cfg["moa"]["presets"]["review"]["reference_models"]
+    assert refs == [
+        {
+            "provider": "openrouter",
+            "model": "x-ai/grok-4.3",
+            "role_prompt": "Find real-world edge cases.",
+        }
+    ]
+    assert "role: Find real-world edge cases." in capsys.readouterr().out

tests/hermes_cli/test_moa_config.py

@@ -42,6 +42,55 @@ def test_normalize_moa_config_preserves_named_presets():
     assert cfg["reference_models"] == [{"provider": "openai-codex", "model": "gpt-5.5"}]
 
 
+def test_normalize_moa_config_preserves_reference_role_prompt():
+    """Reference slots may carry per-model role guidance for advisory calls."""
+    cfg = normalize_moa_config(
+        {
+            "presets": {
+                "review": {
+                    "reference_models": [
+                        {
+                            "provider": "openrouter",
+                            "model": "x-ai/grok-4.3",
+                            "role_prompt": "Find real-world edge cases.",
+                        }
+                    ]
+                }
+            }
+        }
+    )
+
+    assert cfg["presets"]["review"]["reference_models"] == [
+        {
+            "provider": "openrouter",
+            "model": "x-ai/grok-4.3",
+            "role_prompt": "Find real-world edge cases.",
+        }
+    ]
+
+
+def test_normalize_moa_config_drops_aggregator_role_prompt():
+    """role_prompt is reference-only; aggregator slots remain provider/model pairs."""
+    cfg = normalize_moa_config(
+        {
+            "presets": {
+                "review": {
+                    "aggregator": {
+                        "provider": "openrouter",
+                        "model": "openai/gpt-5.5",
+                        "role_prompt": "ignored",
+                    }
+                }
+            }
+        }
+    )
+
+    assert cfg["presets"]["review"]["aggregator"] == {
+        "provider": "openrouter",
+        "model": "openai/gpt-5.5",
+    }
+
+
 def test_legacy_flat_config_becomes_default_preset():
     cfg = normalize_moa_config(
         {

tests/hermes_cli/test_web_server.py

@@ -492,6 +492,40 @@ class TestWebServerEndpoints:
         assert cfg["moa"]["reference_models"] == payload["reference_models"]
         assert cfg["moa"]["aggregator"] == payload["aggregator"]
 
+    def test_put_moa_models_preserves_reference_role_prompts(self):
+        from hermes_cli.config import load_config
+
+        payload = {
+            "presets": {
+                "review": {
+                    "reference_models": [
+                        {
+                            "provider": "openrouter",
+                            "model": "x-ai/grok-4.3",
+                            "role_prompt": "Find real-world edge cases.",
+                        }
+                    ],
+                    "aggregator": {
+                        "provider": "openrouter",
+                        "model": "openai/gpt-5.5",
+                        "role_prompt": "ignored on aggregators",
+                    },
+                }
+            },
+            "default_preset": "review",
+        }
+
+        resp = self.client.put("/api/model/moa", json=payload)
+        assert resp.status_code == 200
+        assert resp.json()["presets"]["review"]["reference_models"][0]["role_prompt"] == "Find real-world edge cases."
+        roundtrip = self.client.get("/api/model/moa").json()
+        resp2 = self.client.put("/api/model/moa", json=roundtrip)
+        assert resp2.status_code == 200
+        cfg = load_config()
+        preset = cfg["moa"]["presets"]["review"]
+        assert preset["reference_models"][0]["role_prompt"] == "Find real-world edge cases."
+        assert "role_prompt" not in preset["aggregator"]
+
     # ── GET /api/media (remote image display) ───────────────────────────
 
     def test_get_media_serves_image_in_root(self):

tests/run_agent/test_moa_loop_mode.py

@@ -401,6 +401,34 @@ def test_run_reference_prepends_advisory_system_prompt(monkeypatch):
     assert msgs[-1]["role"] == "user"
 
 
+def test_run_reference_appends_slot_role_prompt_to_system_prompt(monkeypatch):
+    """A reference slot's role_prompt reaches that reference, not just aggregator."""
+    from agent.moa_loop import _REFERENCE_SYSTEM_PROMPT, _run_reference
+
+    captured = {}
+
+    def fake_call_llm(**kwargs):
+        captured.update(kwargs)
+        return _response("edge-case advice")
+
+    monkeypatch.setattr("agent.moa_loop.call_llm", fake_call_llm)
+
+    label, text = _run_reference(
+        {
+            "provider": "openrouter",
+            "model": "x-ai/grok-4.3",
+            "role_prompt": "Focus on real-world minority knockout requirements.",
+        },
+        [{"role": "user", "content": "evaluate this plan"}],
+    )
+
+    assert text == "edge-case advice"
+    system = captured["messages"][0]
+    assert system["role"] == "system"
+    assert _REFERENCE_SYSTEM_PROMPT in system["content"]
+    assert "real-world minority knockout requirements" in system["content"]
+
+
 def test_moa_facade_references_get_trimmed_messages(monkeypatch, tmp_path):
     home = tmp_path / ".hermes"
     home.mkdir()
@@ -466,6 +494,99 @@ moa:
     assert agg_call["tools"] is not None
 
 
+def test_moa_facade_applies_distinct_reference_role_prompts(monkeypatch, tmp_path):
+    """YAML role_prompt values reach the matching reference calls via facade."""
+    home = tmp_path / ".hermes"
+    home.mkdir()
+    (home / "config.yaml").write_text(
+        """
+moa:
+  default_preset: review
+  presets:
+    review:
+      reference_models:
+        - provider: openrouter
+          model: x-ai/grok-4.3
+          role_prompt: Edge {case} critic.
+        - provider: openrouter
+          model: z-ai/glm-5.2
+          role_prompt: Design presentation critic.
+      aggregator:
+        provider: openrouter
+        model: openai/gpt-5.5
+""".strip(),
+        encoding="utf-8",
+    )
+    monkeypatch.setenv("HERMES_HOME", str(home))
+    reference_system_prompts = {}
+
+    def fake_call_llm(**kwargs):
+        if kwargs["task"] == "moa_reference":
+            reference_system_prompts[kwargs["model"]] = kwargs["messages"][0]["content"]
+            return _response(f"advice from {kwargs['model']}")
+        return _response("aggregator acted")
+
+    monkeypatch.setattr("agent.moa_loop.call_llm", fake_call_llm)
+
+    from agent.moa_loop import MoAChatCompletions
+
+    facade = MoAChatCompletions("review")
+    facade.create(messages=[{"role": "user", "content": "q"}], tools=[])
+
+    assert "Edge {case} critic." in reference_system_prompts["x-ai/grok-4.3"]
+    assert "Design presentation critic." in reference_system_prompts["z-ai/glm-5.2"]
+    assert "Design presentation critic." not in reference_system_prompts["x-ai/grok-4.3"]
+
+
+def test_moa_reference_cache_key_includes_role_prompt(monkeypatch, tmp_path):
+    """Changing only role_prompt must invalidate the per-turn reference cache."""
+    home = tmp_path / ".hermes"
+    home.mkdir()
+    config_path = home / "config.yaml"
+
+    def write_config(role_prompt: str) -> None:
+        config_path.write_text(
+            f"""
+moa:
+  default_preset: review
+  presets:
+    review:
+      reference_models:
+        - provider: openrouter
+          model: x-ai/grok-4.3
+          role_prompt: {role_prompt}
+      aggregator:
+        provider: openrouter
+        model: openai/gpt-5.5
+""".strip(),
+            encoding="utf-8",
+        )
+
+    write_config("first critic")
+    monkeypatch.setenv("HERMES_HOME", str(home))
+    ref_prompts = []
+
+    def fake_call_llm(**kwargs):
+        if kwargs["task"] == "moa_reference":
+            ref_prompts.append(kwargs["messages"][0]["content"])
+            return _response("advice")
+        return _response("acted")
+
+    monkeypatch.setattr("agent.moa_loop.call_llm", fake_call_llm)
+
+    from agent.moa_loop import MoAChatCompletions
+
+    facade = MoAChatCompletions("review")
+    messages = [{"role": "user", "content": "same turn"}]
+    facade.create(messages=messages, tools=[])
+    write_config("second critic")
+    facade.create(messages=messages, tools=[])
+
+    assert len(ref_prompts) == 2
+    assert "first critic" in ref_prompts[0]
+    assert "second critic" in ref_prompts[1]
+
+
 def test_moa_disabled_preset_skips_references(monkeypatch, tmp_path):
     home = tmp_path / ".hermes"
     home.mkdir()

website/docs/user-guide/features/mixture-of-agents.md

@@ -80,8 +80,14 @@ moa:
       reference_models:
         - provider: openai-codex
           model: gpt-5.5
+          role_prompt: >-
+            Check the plan as a rigorous implementation critic. Surface hidden
+            assumptions, missed tests, and tool-use risks.
         - provider: openrouter
           model: deepseek/deepseek-v4-pro
+          role_prompt: >-
+            Check feasibility, edge cases, and whether the extra perspective is
+            worth its cost.
       aggregator:
         provider: openrouter
         model: anthropic/claude-opus-4.8
@@ -97,6 +103,13 @@ Default preset:
 - reference: `openrouter:deepseek/deepseek-v4-pro`
 - aggregator / acting model: `openrouter:anthropic/claude-opus-4.8`
 
+`role_prompt` is optional and reference-only. When present, Hermes appends it
+to that reference model's private advisory system prompt, so each reference can
+specialize (for example: logic critic, real-world edge-case critic, design
+critic, or implementation verifier). It is trusted operator configuration:
+don't fill it from untrusted user content, and keep it short because it is sent
+on every reference call.
+
 ## Terminal preset management
 
 ```bash