fix: stop sending temperature to Anthropic models that reject it

scripts/refresh_catalog.py

@@ -38,7 +38,11 @@
 
 from dotenv import load_dotenv
 
-from duh.providers.catalog import MODEL_CATALOG, NO_TEMPERATURE_MODELS
+from duh.providers.catalog import (
+    ANTHROPIC_NO_TEMPERATURE_MODELS,
+    MODEL_CATALOG,
+    NO_TEMPERATURE_MODELS,
+)
 
 # truefoundry provider directory names (differ from our provider_id keys)
 _TF_BASE = "https://raw.githubusercontent.com/truefoundry/models/main/providers"
@@ -196,6 +200,33 @@ def probe_openai_temperature(model_id: str) -> bool | None:
     return True
 
 
+def probe_anthropic_temperature(model_id: str) -> bool | None:
+    """True if the model accepts temperature, False if it rejects, None on error."""
+    key = os.environ.get("ANTHROPIC_API_KEY")
+    if not key:
+        return None
+    resp = _post_json(
+        "https://api.anthropic.com/v1/messages",
+        {
+            "x-api-key": key,
+            "anthropic-version": "2023-06-01",
+            "Content-Type": "application/json",
+        },
+        {
+            "model": model_id,
+            "max_tokens": 16,
+            "temperature": 0.7,
+            "messages": [{"role": "user", "content": "hi"}],
+        },
+    )
+    if resp is None:
+        return None
+    err = resp.get("error") if isinstance(resp, dict) else None
+    if isinstance(err, dict):
+        return False if "temperature" in err.get("message", "") else None
+    return True
+
+
 # ── report ─────────────────────────────────────────────────────
 
 
@@ -292,6 +323,28 @@ def main() -> int:
             else:
                 print(f"  {mid}: ok ({'no-temp' if rejects else 'temp'})")
 
+        print("\n## BEHAVIOR — Anthropic temperature probe vs catalog set")
+        anthropic_ids = {m["model_id"] for m in MODEL_CATALOG.get("anthropic", [])}
+        for mid in sorted(anthropic_ids):
+            accepts = probe_anthropic_temperature(mid)
+            if accepts is None:
+                print(f"  {mid}: probe inconclusive")
+                continue
+            listed = mid in ANTHROPIC_NO_TEMPERATURE_MODELS
+            rejects = not accepts
+            if rejects and not listed:
+                print(
+                    f"  {mid}: REJECTS temp but NOT in "
+                    "ANTHROPIC_NO_TEMPERATURE_MODELS — ADD"
+                )
+            elif accepts and listed:
+                print(
+                    f"  {mid}: accepts temp but IS in "
+                    "ANTHROPIC_NO_TEMPERATURE_MODELS — REMOVE"
+                )
+            else:
+                print(f"  {mid}: ok ({'no-temp' if rejects else 'temp'})")
+
     if args.write_snapshot:
         _SNAPSHOT.write_text(json.dumps(new_snapshot, indent=2, sort_keys=True) + "\n")
         print(f"\nSnapshot written to {_SNAPSHOT}")

src/duh/providers/anthropic.py

@@ -23,7 +23,11 @@
     TokenUsage,
     ToolCallData,
 )
-from duh.providers.catalog import MODEL_CATALOG, PROVIDER_CAPS
+from duh.providers.catalog import (
+    ANTHROPIC_NO_TEMPERATURE_MODELS,
+    MODEL_CATALOG,
+    PROVIDER_CAPS,
+)
 
 if TYPE_CHECKING:
     from collections.abc import AsyncIterator
@@ -120,10 +124,12 @@ async def send(
         kwargs: dict[str, Any] = {
             "model": model_id,
             "max_tokens": max_tokens,
-            "temperature": temperature,
             "system": system,
             "messages": api_messages,
         }
+        # Newest thinking models reject temperature; older ones accept it.
+        if model_id not in ANTHROPIC_NO_TEMPERATURE_MODELS:
+            kwargs["temperature"] = temperature
         if stop_sequences:
             kwargs["stop_sequences"] = stop_sequences
         if tools:
@@ -243,10 +249,11 @@ async def stream(
         kwargs: dict[str, Any] = {
             "model": model_id,
             "max_tokens": max_tokens,
-            "temperature": temperature,
             "system": system,
             "messages": api_messages,
         }
+        if model_id not in ANTHROPIC_NO_TEMPERATURE_MODELS:
+            kwargs["temperature"] = temperature
         if stop_sequences:
             kwargs["stop_sequences"] = stop_sequences
 

src/duh/providers/catalog.py

@@ -297,6 +297,15 @@
     "gpt-5.5",
 }
 
+# ── Anthropic models that reject the temperature param ─────────
+# The newest thinking models (Opus 4.7+) deprecated `temperature`; older
+# models (Opus 4.6, Sonnet, Haiku) still accept it. Verified empirically.
+
+ANTHROPIC_NO_TEMPERATURE_MODELS: set[str] = {
+    "claude-opus-4-8",
+    "claude-opus-4-7",
+}
+
 # ── Providers that are challenger-only (not proposer-eligible) ──
 
 CHALLENGER_ONLY_PROVIDERS: set[str] = {"perplexity"}

tests/unit/test_config.py

@@ -337,10 +337,16 @@ def test_cloudflare_configured_from_env(self, tmp_path, monkeypatch):
         )
 
     def test_cloudflare_absent_without_env(self, tmp_path, monkeypatch):
-        """Missing either var -> no cloudflare provider."""
+        """Missing either var -> no cloudflare provider.
+
+        Set to empty rather than delete: load_config() runs load_dotenv(),
+        which would otherwise repopulate these from the repo .env. An empty
+        env var is treated as "existing" (not overridden) and as absent by
+        _resolve_cloudflare.
+        """
         self._isolate(tmp_path, monkeypatch)
-        monkeypatch.delenv("CLOUDFLARE_ACCOUNT_ID", raising=False)
-        monkeypatch.delenv("CLOUDFLARE_WORKERS_AI_TOKEN", raising=False)
+        monkeypatch.setenv("CLOUDFLARE_ACCOUNT_ID", "")
+        monkeypatch.setenv("CLOUDFLARE_WORKERS_AI_TOKEN", "")
 
         cfg = load_config()
         assert "cloudflare" not in cfg.providers
@@ -349,7 +355,7 @@ def test_cloudflare_requires_both(self, tmp_path, monkeypatch):
         """Only the account ID (no token) -> not configured."""
         self._isolate(tmp_path, monkeypatch)
         monkeypatch.setenv("CLOUDFLARE_ACCOUNT_ID", "acct123")
-        monkeypatch.delenv("CLOUDFLARE_WORKERS_AI_TOKEN", raising=False)
+        monkeypatch.setenv("CLOUDFLARE_WORKERS_AI_TOKEN", "")
 
         cfg = load_config()
         assert "cloudflare" not in cfg.providers

tests/unit/test_providers_anthropic.py

@@ -313,3 +313,45 @@ async def test_unhealthy_on_error(self):
         client.messages.create.side_effect = Exception("connection failed")
         provider = AnthropicProvider(client=client)
         assert await provider.health_check() is False
+
+
+# ── Temperature handling (newest models reject it) ───────────────
+
+
+class TestTemperatureHandling:
+    async def test_omits_temperature_for_no_temp_model(self):
+        """Opus 4.8 rejects temperature -> it must not be sent."""
+        client = _make_client()
+        provider = AnthropicProvider(client=client)
+        await provider.send(
+            [PromptMessage(role="user", content="hi")],
+            "claude-opus-4-8",
+            temperature=0.7,
+        )
+        call_kwargs = client.messages.stream.call_args.kwargs
+        assert "temperature" not in call_kwargs
+
+    async def test_sends_temperature_for_regular_model(self):
+        """Opus 4.6 still accepts temperature -> it is sent."""
+        client = _make_client()
+        provider = AnthropicProvider(client=client)
+        await provider.send(
+            [PromptMessage(role="user", content="hi")],
+            "claude-opus-4-6",
+            temperature=0.55,
+        )
+        call_kwargs = client.messages.stream.call_args.kwargs
+        assert call_kwargs["temperature"] == 0.55
+
+    async def test_stream_omits_temperature_for_no_temp_model(self):
+        """The streaming path also omits temperature for Opus 4.8."""
+        client = _make_client()
+        provider = AnthropicProvider(client=client)
+        async for _ in provider.stream(
+            [PromptMessage(role="user", content="hi")],
+            "claude-opus-4-8",
+            temperature=0.7,
+        ):
+            pass
+        call_kwargs = client.messages.stream.call_args.kwargs
+        assert "temperature" not in call_kwargs