fix(moa): preserve provider identity for anthropic + bedrock slots

agent/moa_loop.py

@@ -99,7 +99,22 @@ def _slot_runtime(slot: dict[str, str]) -> dict[str, Any]:
         # provider-backed targets whose provider branch adds auth refresh,
         # request metadata, or request-shape adapters. Keep those providers
         # identified by name.
-        if resolved_provider in {"nous", "openai-codex", "xai-oauth"}:
+        # ``bedrock`` belongs here too: its provider branch builds an
+        # AWS-SigV4-signed client (or IAM-role-signed) against the
+        # bedrock-runtime endpoint. resolve_runtime_provider returns that
+        # endpoint's base_url plus a PLACEHOLDER api_key ("aws-sdk") — there is
+        # no real bearer token. Forwarding base_url+api_key makes call_llm treat
+        # it as a plain OpenAI-compatible endpoint and POST with an unsigned
+        # fake bearer, which Bedrock answers with an empty/malformed
+        # ChatCompletion (choices=None). Keeping it identified by name routes it
+        # through the real signed bedrock branch.
+        #
+        # ``anthropic`` likewise: subscription OAuth setup-tokens (sk-ant-oat*)
+        # require Bearer auth plus the ``anthropic-beta: oauth-*`` header, which
+        # only the anthropic provider branch adds. Forwarding base_url+api_key
+        # sends the OAuth token as ``x-api-key``, which Anthropic rejects with a
+        # bare 429.
+        if resolved_provider in {"nous", "anthropic", "openai-codex", "xai-oauth", "bedrock"}:
             return out
         # Pass the resolved endpoint through so call_llm builds the request for
         # the provider's actual API surface instead of auto-detecting. base_url

scripts/release.py

@@ -45,6 +45,8 @@
 
 # Auto-extracted from noreply emails + manual overrides
 AUTHOR_MAP = {
+    "chufengfan@jackroooc-2.local": "jackroofan",  # PR #54609 salvage (add anthropic to MoA _slot_runtime name-preserve set; OAuth sk-ant-oat* needs Bearer + anthropic-beta header)
+    "igor.izotov@gmail.com": "iizotov",  # PR #54912 salvage (add bedrock to MoA _slot_runtime name-preserve set; SigV4-signed client, placeholder aws-sdk api_key)
     "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/run_agent/test_moa_loop_mode.py

@@ -200,7 +200,7 @@ def fake_resolve(*, requested, target_model=None):
     assert rt == {"provider": "openai-codex", "model": "gpt-5.5"}
 
 
-@pytest.mark.parametrize("provider", ["anthropic", "minimax-oauth", "qwen-oauth"])
+@pytest.mark.parametrize("provider", ["minimax-oauth", "qwen-oauth"])
 def test_moa_provider_backed_slot_survives_aux_resolution(monkeypatch, provider):
     """MoA can pass resolved endpoints for provider-backed slots without
     call_llm flattening them to generic custom endpoints.
@@ -211,6 +211,11 @@ def test_moa_provider_backed_slot_survives_aux_resolution(monkeypatch, provider)
     via ``_resolve_task_provider_model`` (which takes everything except
     ``api_mode``, handled separately). The provider identity must survive that
     resolution rather than being flattened to ``custom``.
+
+    NOTE: providers in the ``_slot_runtime`` name-preservation set (anthropic,
+    bedrock, nous, openai-codex, xai-oauth) are intentionally NOT forwarded —
+    they're covered by their own dedicated tests. This case covers the
+    forward-the-resolved-endpoint path for providers that are NOT in the set.
     """
     from agent import moa_loop
     from agent.auxiliary_client import _resolve_task_provider_model
@@ -678,3 +683,45 @@ def fake_call_llm(**kwargs):
 
     # 2 references × 2 distinct turns = 4 reference runs.
     assert len(ref_runs) == 4
+
+
+def test_slot_runtime_anthropic_oauth_routes_through_provider_branch(monkeypatch):
+    """Native anthropic slots must NOT forward base_url/api_key.
+
+    anthropic OAuth setup-tokens (sk-ant-oat*) require Bearer auth + the
+    ``anthropic-beta: oauth-*`` header, which only the provider branch of
+    call_llm adds. If _slot_runtime forwarded base_url/api_key, call_llm would
+    treat the slot as a plain custom endpoint and send the token as x-api-key,
+    which Anthropic rejects with a bare 429. So a whitelisted provider
+    (anthropic) returns only provider/model, while a non-whitelisted provider
+    (openrouter) forwards the resolved base_url/api_key.
+    """
+    from agent import moa_loop
+
+    def fake_resolve(*, requested, target_model=None):
+        return {
+            "provider": requested,
+            "base_url": "https://resolved.example/v1",
+            "api_key": "resolved-key",
+        }
+
+    monkeypatch.setattr(
+        "hermes_cli.runtime_provider.resolve_runtime_provider", fake_resolve
+    )
+
+    # Whitelisted: anthropic must skip base_url/api_key forwarding.
+    anthropic_rt = moa_loop._slot_runtime(
+        {"provider": "anthropic", "model": "claude-opus-4-8"}
+    )
+    assert anthropic_rt == {"provider": "anthropic", "model": "claude-opus-4-8"}
+    assert "base_url" not in anthropic_rt
+    assert "api_key" not in anthropic_rt
+
+    # Non-whitelisted: openrouter still forwards the resolved endpoint.
+    other_rt = moa_loop._slot_runtime(
+        {"provider": "openrouter", "model": "some-model"}
+    )
+    assert other_rt["provider"] == "openrouter"
+    assert other_rt["model"] == "some-model"
+    assert other_rt["base_url"] == "https://resolved.example/v1"
+    assert other_rt["api_key"] == "resolved-key"