fix(minimax): route M3 defaults through split reasoning

hermes_cli/runtime_provider.py

@@ -133,6 +133,39 @@ def _detect_api_mode_for_url(base_url: str) -> Optional[str]:
     return None
 
 
+_MINIMAX_M3_OPENAI_BASE_URL = "https://api.minimax.io/v1"
+
+
+def _is_minimax_m3_model(model: Any) -> bool:
+    normalized = str(model or "").strip().lower()
+    if not normalized:
+        return False
+    return normalized.split("/")[-1] == "minimax-m3"
+
+
+def _maybe_route_default_minimax_m3_openai(
+    *,
+    provider: str,
+    base_url: str,
+    api_mode: str,
+    effective_model: Any,
+    base_url_from_default: bool,
+) -> tuple[str, str]:
+    """Route built-in MiniMax-M3 to the OpenAI-compatible split-reasoning path.
+
+    MiniMax-M3 leaks provider-specific ``<mm:think>`` markers as visible
+    content on the default Anthropic-compatible endpoint. The OpenAI-compatible
+    endpoint supports ``reasoning_split`` (emitted by the MiniMax provider
+    profile), so only the uncustomized built-in route is moved there; explicit
+    user base URL overrides keep their chosen transport.
+    """
+    if provider != "minimax" or not base_url_from_default:
+        return base_url, api_mode
+    if not _is_minimax_m3_model(effective_model):
+        return base_url, api_mode
+    return _MINIMAX_M3_OPENAI_BASE_URL, "chat_completions"
+
+
 def _resolve_plain_custom_api_mode(model_cfg: Dict[str, Any], base_url: str) -> str:
     """Resolve api_mode for legacy/plain ``provider: custom`` endpoints.
 
@@ -405,7 +438,7 @@ def _resolve_runtime_from_pool_entry(
     # longer matches the model actually being used — the bug that caused
     # opencode-zen /v1 to be stripped for chat_completions requests when
     # config.default was still a Claude model.
-    effective_model = (target_model or model_cfg.get("default") or "")
+    effective_model = target_model or model_cfg.get("default") or ""
     base_url = (getattr(entry, "runtime_base_url", None) or getattr(entry, "base_url", None) or "").rstrip("/")
     api_key = getattr(entry, "runtime_api_key", None) or getattr(entry, "access_token", "")
     api_mode = "chat_completions"
@@ -479,7 +512,16 @@ def _resolve_runtime_from_pool_entry(
         # fell back to the hardcoded default).  Env var overrides win (#6039).
         pconfig = PROVIDER_REGISTRY.get(provider)
         pool_url_is_default = pconfig and base_url.rstrip("/") == pconfig.inference_base_url.rstrip("/")
-        if configured_provider == provider and pool_url_is_default:
+        env_url = ""
+        if pconfig and pconfig.base_url_env_var:
+            env_url = (
+                _getenv(pconfig.base_url_env_var, "")
+                or os.getenv(pconfig.base_url_env_var, "")
+            ).strip().rstrip("/")
+        cfg_base_url = ""
+        if provider in {"minimax", "minimax-cn"} and pool_url_is_default and env_url:
+            base_url = env_url
+        elif configured_provider == provider and pool_url_is_default:
             cfg_base_url = str(model_cfg.get("base_url") or "").strip().rstrip("/")
             if cfg_base_url:
                 base_url = cfg_base_url
@@ -501,6 +543,13 @@ def _resolve_runtime_from_pool_entry(
             detected = _detect_api_mode_for_url(base_url)
             if detected:
                 api_mode = detected
+        base_url, api_mode = _maybe_route_default_minimax_m3_openai(
+            provider=provider,
+            base_url=base_url,
+            api_mode=api_mode,
+            effective_model=effective_model,
+            base_url_from_default=bool(pool_url_is_default and not env_url and not cfg_base_url),
+        )
 
     # OpenCode base URLs end with /v1 for OpenAI-compatible models, but the
     # Anthropic SDK prepends its own /v1/messages to the base_url.  Normalize
@@ -1358,6 +1407,7 @@ def _resolve_explicit_runtime(
     model_cfg: Dict[str, Any],
     explicit_api_key: Optional[str] = None,
     explicit_base_url: Optional[str] = None,
+    target_model: Optional[str] = None,
 ) -> Optional[Dict[str, Any]]:
     explicit_api_key = str(explicit_api_key or "").strip()
     explicit_base_url = str(explicit_base_url or "").strip().rstrip("/")
@@ -1461,7 +1511,10 @@ def _resolve_explicit_runtime(
     if pconfig and pconfig.auth_type == "api_key":
         env_url = ""
         if pconfig.base_url_env_var:
-            env_url = _getenv(pconfig.base_url_env_var, "").strip().rstrip("/")
+            env_url = (
+                _getenv(pconfig.base_url_env_var, "")
+                or os.getenv(pconfig.base_url_env_var, "")
+            ).strip().rstrip("/")
 
         base_url = explicit_base_url
         if not base_url:
@@ -1493,6 +1546,18 @@ def _resolve_explicit_runtime(
                 detected = _detect_api_mode_for_url(base_url)
                 if detected:
                     api_mode = detected
+        effective_model = target_model or model_cfg.get("default") or ""
+        base_url, api_mode = _maybe_route_default_minimax_m3_openai(
+            provider=provider,
+            base_url=base_url,
+            api_mode=api_mode,
+            effective_model=effective_model,
+            base_url_from_default=bool(
+                not explicit_base_url
+                and not env_url
+                and base_url.rstrip("/") == pconfig.inference_base_url.rstrip("/")
+            ),
+        )
 
         return {
             "provider": provider,
@@ -1661,6 +1726,7 @@ def resolve_runtime_provider(
         model_cfg=model_cfg,
         explicit_api_key=explicit_api_key,
         explicit_base_url=explicit_base_url,
+        target_model=target_model,
     )
     if explicit_runtime:
         return explicit_runtime
@@ -1997,7 +2063,15 @@ def resolve_runtime_provider(
         cfg_base_url = ""
         if cfg_provider == provider:
             cfg_base_url = (model_cfg.get("base_url") or "").strip().rstrip("/")
+        env_url = ""
+        if pconfig.base_url_env_var:
+            env_url = (
+                _getenv(pconfig.base_url_env_var, "")
+                or os.getenv(pconfig.base_url_env_var, "")
+            ).strip().rstrip("/")
         base_url = cfg_base_url or creds.get("base_url", "").rstrip("/")
+        if provider in {"minimax", "minimax-cn"} and env_url:
+            base_url = env_url
         api_mode = "chat_completions"
         if provider == "copilot":
             api_mode = _copilot_runtime_api_mode(model_cfg, creds.get("api_key", ""))
@@ -2028,6 +2102,18 @@ def resolve_runtime_provider(
                 detected = _detect_api_mode_for_url(base_url)
                 if detected:
                     api_mode = detected
+        effective_model = target_model or model_cfg.get("default") or ""
+        base_url, api_mode = _maybe_route_default_minimax_m3_openai(
+            provider=provider,
+            base_url=base_url,
+            api_mode=api_mode,
+            effective_model=effective_model,
+            base_url_from_default=bool(
+                not cfg_base_url
+                and not env_url
+                and base_url.rstrip("/") == pconfig.inference_base_url.rstrip("/")
+            ),
+        )
         # Normalize the /v1 suffix for OpenCode by API mode (see comment above).
         if provider in {"opencode-zen", "opencode-go"}:
             from hermes_cli.models import normalize_opencode_base_url

plugins/model-providers/minimax/__init__.py

@@ -1,9 +1,9 @@
 """MiniMax provider profiles (international + China).
 
-The default API-key routes use anthropic_messages because their base URLs end
-with /anthropic. Users can opt MiniMax-M3 into the OpenAI-compatible endpoint
-with base_url=https://api.minimax.io/v1; that route needs MiniMax-specific
-reasoning controls in extra_body.
+Most API-key routes use anthropic_messages because their base URLs end with
+/anthropic. The runtime resolver routes the built-in MiniMax-M3 default to
+https://api.minimax.io/v1 so this profile can request split reasoning; explicit
+user base_url overrides stay user-owned.
 """
 
 from typing import Any

tests/hermes_cli/test_runtime_provider_resolution.py

@@ -21,6 +21,15 @@ def _fake_invoke_jwt(ttl_seconds=3600):
     return f"{header}.{payload}.sig"
 
 
+class _NoCredentialPool:
+    def has_credentials(self):
+        return False
+
+
+def _disable_credential_pool(monkeypatch):
+    monkeypatch.setattr(rp, "load_pool", lambda provider: _NoCredentialPool())
+
+
 def test_resolve_runtime_provider_uses_credential_pool(monkeypatch):
     class _Entry:
         access_token = "pool-token"
@@ -45,6 +54,65 @@ def select(self):
     assert resolved["source"] == "manual"
 
 
+def test_minimax_m3_credential_pool_default_route_uses_openai_endpoint(monkeypatch):
+    class _Entry:
+        access_token = "pool-token"
+        source = "manual"
+        base_url = "https://api.minimax.io/anthropic"
+
+    class _Pool:
+        def has_credentials(self):
+            return True
+
+        def select(self):
+            return _Entry()
+
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
+    monkeypatch.setattr(
+        rp,
+        "_get_model_config",
+        lambda: {"provider": "minimax", "default": "minimax/minimax-m3"},
+    )
+    monkeypatch.setattr(rp, "load_pool", lambda provider: _Pool())
+
+    resolved = rp.resolve_runtime_provider(requested="minimax")
+
+    assert resolved["provider"] == "minimax"
+    assert resolved["api_mode"] == "chat_completions"
+    assert resolved["base_url"] == "https://api.minimax.io/v1"
+    assert resolved["api_key"] == "pool-token"
+
+
+def test_minimax_m3_credential_pool_env_base_url_preserves_user_route(monkeypatch):
+    class _Entry:
+        access_token = "pool-token"
+        source = "manual"
+        base_url = "https://api.minimax.io/anthropic"
+
+    class _Pool:
+        def has_credentials(self):
+            return True
+
+        def select(self):
+            return _Entry()
+
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
+    monkeypatch.setattr(
+        rp,
+        "_get_model_config",
+        lambda: {"provider": "minimax", "default": "minimax/minimax-m3"},
+    )
+    monkeypatch.setattr(rp, "load_pool", lambda provider: _Pool())
+    monkeypatch.setenv("MINIMAX_BASE_URL", "https://api.minimax.io/anthropic")
+
+    resolved = rp.resolve_runtime_provider(requested="minimax")
+
+    assert resolved["provider"] == "minimax"
+    assert resolved["api_mode"] == "anthropic_messages"
+    assert resolved["base_url"] == "https://api.minimax.io/anthropic"
+    assert resolved["api_key"] == "pool-token"
+
+
 def test_resolve_runtime_provider_nous_pool_uses_env_base_url_override(monkeypatch):
     entry = SimpleNamespace(
         provider="nous",
@@ -1467,6 +1535,7 @@ def test_anthropic_messages_in_valid_api_modes():
 
 def test_api_key_provider_anthropic_url_auto_detection(monkeypatch):
     """API-key providers with /anthropic base URL should auto-detect anthropic_messages mode."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {})
     monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
@@ -1481,6 +1550,7 @@ def test_api_key_provider_anthropic_url_auto_detection(monkeypatch):
 
 def test_api_key_provider_explicit_api_mode_config(monkeypatch):
     """API-key providers should respect api_mode from model config."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {"api_mode": "anthropic_messages"})
     monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
@@ -1494,6 +1564,7 @@ def test_api_key_provider_explicit_api_mode_config(monkeypatch):
 
 def test_minimax_default_url_uses_anthropic_messages(monkeypatch):
     """MiniMax with default /anthropic URL should auto-detect anthropic_messages mode."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {})
     monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
@@ -1506,8 +1577,70 @@ def test_minimax_default_url_uses_anthropic_messages(monkeypatch):
     assert resolved["base_url"] == "https://api.minimax.io/anthropic"
 
 
+def test_minimax_m3_default_route_uses_openai_reasoning_split_endpoint(monkeypatch):
+    """Built-in MiniMax-M3 should use /v1 so the profile can request reasoning_split."""
+    _disable_credential_pool(monkeypatch)
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
+    monkeypatch.setattr(
+        rp,
+        "_get_model_config",
+        lambda: {"provider": "minimax", "default": "minimax/minimax-m3"},
+    )
+    monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
+    monkeypatch.delenv("MINIMAX_BASE_URL", raising=False)
+
+    resolved = rp.resolve_runtime_provider(requested="minimax")
+
+    assert resolved["provider"] == "minimax"
+    assert resolved["api_mode"] == "chat_completions"
+    assert resolved["base_url"] == "https://api.minimax.io/v1"
+
+
+def test_minimax_m3_default_route_ignores_stale_anthropic_api_mode(monkeypatch):
+    """A stale persisted api_mode must not keep M3 on the inline-reasoning route."""
+    _disable_credential_pool(monkeypatch)
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
+    monkeypatch.setattr(
+        rp,
+        "_get_model_config",
+        lambda: {
+            "provider": "minimax",
+            "default": "MiniMax-M3",
+            "api_mode": "anthropic_messages",
+        },
+    )
+    monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
+    monkeypatch.delenv("MINIMAX_BASE_URL", raising=False)
+
+    resolved = rp.resolve_runtime_provider(requested="minimax")
+
+    assert resolved["provider"] == "minimax"
+    assert resolved["api_mode"] == "chat_completions"
+    assert resolved["base_url"] == "https://api.minimax.io/v1"
+
+
+def test_minimax_m3_env_base_url_preserves_user_route(monkeypatch):
+    """An explicit MiniMax base URL remains user-owned, even for M3."""
+    _disable_credential_pool(monkeypatch)
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
+    monkeypatch.setattr(
+        rp,
+        "_get_model_config",
+        lambda: {"provider": "minimax", "default": "MiniMax-M3"},
+    )
+    monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
+    monkeypatch.setenv("MINIMAX_BASE_URL", "https://api.minimax.io/anthropic")
+
+    resolved = rp.resolve_runtime_provider(requested="minimax")
+
+    assert resolved["provider"] == "minimax"
+    assert resolved["api_mode"] == "anthropic_messages"
+    assert resolved["base_url"] == "https://api.minimax.io/anthropic"
+
+
 def test_minimax_v1_url_uses_chat_completions(monkeypatch):
     """MiniMax with /v1 base URL should use chat_completions (user override for regions where /anthropic 404s)."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {})
     monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
@@ -1522,6 +1655,7 @@ def test_minimax_v1_url_uses_chat_completions(monkeypatch):
 
 def test_minimax_cn_v1_url_uses_chat_completions(monkeypatch):
     """MiniMax-CN with /v1 base URL should use chat_completions (user override)."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax-cn")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {})
     monkeypatch.setenv("MINIMAX_CN_API_KEY", "test-minimax-cn-key")
@@ -1536,6 +1670,7 @@ def test_minimax_cn_v1_url_uses_chat_completions(monkeypatch):
 
 def test_minimax_explicit_api_mode_respected(monkeypatch):
     """Explicit api_mode config should override MiniMax auto-detection."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {"api_mode": "chat_completions"})
     monkeypatch.setenv("MINIMAX_API_KEY", "test-minimax-key")
@@ -1549,6 +1684,7 @@ def test_minimax_explicit_api_mode_respected(monkeypatch):
 
 def test_minimax_config_base_url_overrides_hardcoded_default(monkeypatch):
     """model.base_url in config.yaml should override the hardcoded default (#6039)."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {
         "provider": "minimax",
@@ -1566,6 +1702,7 @@ def test_minimax_config_base_url_overrides_hardcoded_default(monkeypatch):
 
 def test_minimax_env_base_url_still_wins_over_config(monkeypatch):
     """MINIMAX_BASE_URL env var should take priority over config.yaml model.base_url."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {
         "provider": "minimax",
@@ -1582,6 +1719,7 @@ def test_minimax_env_base_url_still_wins_over_config(monkeypatch):
 
 def test_minimax_config_base_url_ignored_for_different_provider(monkeypatch):
     """model.base_url should NOT be used when model.provider doesn't match."""
+    _disable_credential_pool(monkeypatch)
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "minimax")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {
         "provider": "openrouter",