Log every model call's time and tokens

Each call through BudgetedClient emits llm.call: seconds, input,
output and cached tokens, and the answering model when a fallback
answered; a failed call logs how long it ran before it failed.
Usage now carries the server's cached-prompt count on both routes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J5Utyfzw7gF7UhjRDEtjTk

nisse-core/src/nisse_core/models/context.py

@@ -31,6 +31,7 @@ dangerous direction.
 """
 
 import json
+import time
 from collections.abc import Callable, Sequence
 from dataclasses import dataclass, replace
 from typing import Any
@@ -297,6 +298,31 @@ class BudgetedClient:
     def last_rate_limit_wait_s(self) -> float:
         return float(getattr(self._inner, "last_rate_limit_wait_s", 0.0) or 0.0)
 
+    def _log_call(self, model: str, started: float, error: str | None = None) -> None:
+        """One event per model call: how long it took and what it used, so a
+        slow turn shows whether reading or writing took the time."""
+        if self._oplog is None:
+            return
+        fields: dict[str, Any] = {
+            "model": model,
+            "seconds": round(time.monotonic() - started, 1),
+        }
+        if error is not None:
+            self._oplog.emit("llm.call", cause="failed", error=error, **fields)
+            return
+        usage = self.last_usage
+        answered = self.last_model
+        if answered and answered != model:
+            fields["answered_by"] = answered
+        self._oplog.emit(
+            "llm.call",
+            cause="answered",
+            input_tokens=int(usage.get("input", 0)),
+            output_tokens=int(usage.get("output", 0)),
+            cached_tokens=int(usage.get("cached", 0)),
+            **fields,
+        )
+
     def budget_for(self, model: str) -> Budget:
         window, model_max_output = self._limits_for(model)
         ceiling = self._max_output_tokens
@@ -360,17 +386,23 @@ class BudgetedClient:
                 )
         if require_tool_call is None:
             require_tool_call = bool(tools) and self._enforces_tool_choice(model)
-        produced = self._inner.complete(
-            fitted,
-            tools,
-            model,
-            system=system,
-            max_tokens=budget.max_output,
-            # Every caller that offers tools needs one called: the executor
-            # finishes with finish, the planner with submit_plan or
-            # ask_human, a delegation with hand_back.
-            require_tool_call=require_tool_call,
-        )
+        started = time.monotonic()
+        try:
+            produced = self._inner.complete(
+                fitted,
+                tools,
+                model,
+                system=system,
+                max_tokens=budget.max_output,
+                # Every caller that offers tools needs one called: the executor
+                # finishes with finish, the planner with submit_plan or
+                # ask_human, a delegation with hand_back.
+                require_tool_call=require_tool_call,
+            )
+        except Exception as e:
+            self._log_call(model, started, error=type(e).__name__)
+            raise
+        self._log_call(model, started)
         reported = int(self.last_usage.get("input", 0))
         if reported > 0:
             estimated = estimate_tokens(fitted, system, tools)

nisse-core/src/nisse_core/models/llm.py

@@ -244,6 +244,10 @@ class LiteLLMClient:
         self.last_usage = {
             "input": int(usage.get("prompt_tokens") or 0),
             "output": int(usage.get("completion_tokens") or 0),
+            # Prompt tokens the server reused from its cache, when it says.
+            "cached": int(
+                (usage.get("prompt_tokens_details") or {}).get("cached_tokens") or 0
+            ),
         }
         out: list[CanonicalEvent] = []
         if message.get("content"):
@@ -371,6 +375,7 @@ class AnthropicPassthroughClient:
         self.last_usage = {
             "input": int(usage.get("input_tokens") or 0),
             "output": int(usage.get("output_tokens") or 0),
+            "cached": int(usage.get("cache_read_input_tokens") or 0),
         }
         out: list[CanonicalEvent] = []
         for block in content: