Send temperature through extra_body; check kwargs against SDK signatures

DEVLOG.md

@@ -4,6 +4,30 @@
 
 ---
 
+## 2026-09-29（夜）：llm-smoke 第一次實跑失敗，temperature 改走 extra_body
+
+### 本次工作 / 執行摘要
+- 事故：PR #14 合併後第一次 `make llm-smoke`，每筆呼叫都在送出前就失敗，錯誤是 `TypeError: Messages.create() got an unexpected keyword argument 'temperature'`。沒有請求到達 API，花費 US$0，輸出裡沒有 key。
+- 原因：anthropic 1.x 把 `temperature`、`top_p`、`top_k` 從 `messages.create()` 的簽名拿掉了，但 API 本身沒有移除，Haiku 4.5 仍然接受。測試用的假 client 什麼參數都收，所以測不出和真 SDK 簽名不符。
+- 修法：`request_params` 改成 `extra_body={"temperature": 0.0}`，SDK 會把它原樣併進 request JSON（claude-api skill，`python/claude-api/sdk-upgrade.md` Step 6 的建議：模型仍接受、程式又依賴這個設定時，移到 extra_body，不要刪）。identity 仍記 temperature 0，語意沒變，所以身分雜湊不變，smoke journal 裡的 failed 紀錄下次會照常重試。
+- `count_tokens` 只送 model、system、messages，沒有同樣的問題；它的參數也抽成 `count_tokens_params()`，跟 `request_params()` 一樣是唯一產生參數的地方。
+- 疤痕變腳本：新增 `test_the_arguments_we_send_fit_the_installed_sdk_signatures`，把實際送出的 kwargs `bind` 到真 SDK 的 `Messages.create` 與 `Messages.count_tokens` 簽名上，不需要網路或 key；另一條測試確認假 client 收到的參數正是這兩個函式產生的，簽名測試因此涵蓋實際送出的內容。
+
+### 核心發現 / 數據
+- (無實跑數據)。定價仍是 Haiku 4.5 輸入 US$1、輸出 US$5 per MTok。
+
+### Blockers / 遇到的問題
+- (無)
+
+### Next
+- [ ] 合併後重跑 `make llm-smoke`
+
+### Files / Budget
+- `src/tinyrouter/llm.py`、`tests/test_llm.py`、`tests/llm_fakes.py`
+- API 花費：US$0
+
+---
+
 ## 2026-09-29（晚）：PR #14 審查修正（4 medium、6 low）
 
 ### 本次工作 / 執行摘要

src/tinyrouter/llm.py

@@ -136,13 +136,29 @@ def cost_usd(
 
 
 def request_params(query: str) -> dict[str, Any]:
-    """The exact Messages API arguments for one query; the single place they are built."""
+    """The exact ``messages.create`` arguments for one query; the single place they are built.
+
+    anthropic 1.x removed ``temperature`` from ``messages.create`` (passing
+    it is a TypeError), but the API did not: Haiku 4.5 still accepts it. The
+    comparison with the old project depends on temperature 0, so it goes in
+    ``extra_body``, which the SDK merges into the request JSON unchanged
+    (claude-api skill, python/claude-api/sdk-upgrade.md, Step 6).
+    """
     return {
         "model": MODEL,
         "max_tokens": MAX_TOKENS,
-        "temperature": TEMPERATURE,
         "system": SYSTEM_PROMPT,
         "messages": [{"role": "user", "content": query}],
+        "extra_body": {"temperature": float(TEMPERATURE)},
+    }
+
+
+def count_tokens_params() -> dict[str, Any]:
+    """The ``messages.count_tokens`` arguments: the prompt with a one-character query."""
+    return {
+        "model": MODEL,
+        "system": SYSTEM_PROMPT,
+        "messages": [{"role": "user", "content": "x"}],
     }
 
 
@@ -300,12 +316,8 @@ def prompt_base_tokens(client: Any, sleep: Callable[[float], None] = time.sleep)
 
     Retried and redacted like ``classify``; LLMCallError when it gives up.
     """
-    counted, _, _ = with_retries(
-        lambda: client.messages.count_tokens(
-            model=MODEL, system=SYSTEM_PROMPT, messages=[{"role": "user", "content": "x"}]
-        ),
-        sleep,
-    )
+    params = count_tokens_params()
+    counted, _, _ = with_retries(lambda: client.messages.count_tokens(**params), sleep)
     return int(counted.input_tokens)
 
 

tests/llm_fakes.py

@@ -56,6 +56,7 @@ def __init__(
         self.base_tokens = base_tokens
         self.calls: list[dict] = []
         self.count_calls = 0
+        self.count_kwargs: list[dict] = []
         self.count_failures: list[BaseException] = []
         self.response_override = None
         self._lock = threading.Lock()
@@ -87,6 +88,7 @@ def create(self, **kwargs):
 
     def count_tokens(self, **kwargs):
         self.count_calls += 1
+        self.count_kwargs.append(kwargs)
         if self.count_failures:
             raise self.count_failures.pop(0)
         return SimpleNamespace(input_tokens=self.base_tokens)

tests/test_llm.py

@@ -1,4 +1,5 @@
 import hashlib
+import inspect
 
 import pytest
 
@@ -58,10 +59,35 @@ def test_request_matches_the_old_project_settings():
     assert params == {
         "model": "claude-haiku-4-5-20251001",
         "max_tokens": 20,
-        "temperature": 0,
         "system": SYSTEM_PROMPT,
         "messages": [{"role": "user", "content": "book a flight"}],
+        "extra_body": {"temperature": 0.0},
     }
+    assert "temperature" not in params
+    assert llm.identity()["temperature"] == 0
+
+
+def test_the_arguments_we_send_fit_the_installed_sdk_signatures():
+    """2026-09-29 smoke incident: every call raised TypeError before reaching the API.
+
+    anthropic 1.8.0 dropped ``temperature`` from ``Messages.create``; the
+    fake client accepts any keyword, so no test noticed. Binding the real
+    kwargs to the real SDK signatures catches that without a network or key.
+    """
+    from anthropic.resources.messages import Messages
+
+    create = inspect.signature(Messages.create)
+    create.bind(None, **llm.request_params("book a flight"))
+    count = inspect.signature(Messages.count_tokens)
+    count.bind(None, **llm.count_tokens_params())
+
+
+def test_the_fake_client_receives_exactly_the_built_arguments():
+    client = fake_client()
+    classify("book a flight", client, sleep=lambda _: None)
+    llm.prompt_base_tokens(client, sleep=lambda _: None)
+    assert client.messages.calls == [llm.request_params("book a flight")]
+    assert client.messages.count_kwargs == [llm.count_tokens_params()]
 
 
 @pytest.mark.parametrize(
@@ -92,7 +118,7 @@ def test_classify_returns_prediction_with_usage_and_request_id():
     assert (result.input_tokens, result.output_tokens) == (250, 4)
     assert result.request_id == "req_fake_1"
     assert result.attempts == 1
-    assert client.messages.calls[-1]["temperature"] == 0
+    assert client.messages.calls[-1]["extra_body"] == {"temperature": 0.0}
 
 
 def test_classify_retries_a_429_then_succeeds_honouring_retry_after():