fix(codex): record duration_ms after streaming in tutorials [greptile]

The sync and async-base harness_codex tutorials computed duration_ms before the
stream was consumed, so TurnUsage.duration_ms captured only subprocess spawn
overhead, not inference time — monitoring on that field would show wrong data.
Make CodexTurn.duration_ms / cost_usd public mutable attributes (the true
wall-clock is only known after the stream ends), and set turn.duration_ms after
process.wait() in both tutorials so it measures the full turn.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: declan-scale

examples/tutorials/00_sync/harness_codex/project/acp.py

@@ -136,11 +136,9 @@ async def handle_message_send(
         await process.stdin.drain()
         process.stdin.close()
 
-        duration_ms = int(time.monotonic() * 1000) - start_ms
         turn = CodexTurn(
             events=_process_stdout(process),
             model=MODEL,
-            duration_ms=duration_ms,
         )
 
         emitter = UnifiedEmitter(
@@ -154,6 +152,10 @@ async def handle_message_send(
 
         await process.wait()
 
+        # Record the real wall-clock duration AFTER streaming completes; setting
+        # it before the stream ran would capture only subprocess spawn overhead.
+        turn.duration_ms = int(time.monotonic() * 1000) - start_ms
+
         if turn_span:
             usage = turn.usage()
             turn_span.output = {

examples/tutorials/10_async/00_base/harness_codex/project/acp.py

@@ -172,11 +172,9 @@ async def handle_task_event_send(params: SendEventParams):
         await process.stdin.drain()
         process.stdin.close()
 
-        duration_ms = int(time.monotonic() * 1000) - start_ms
         turn = CodexTurn(
             events=_process_stdout(process),
             model=MODEL,
-            duration_ms=duration_ms,
         )
 
         emitter = UnifiedEmitter(
@@ -189,6 +187,10 @@ async def handle_task_event_send(params: SendEventParams):
 
         await process.wait()
 
+        # Record the real wall-clock duration AFTER streaming completes; setting
+        # it before the stream ran would capture only subprocess spawn overhead.
+        turn.duration_ms = int(time.monotonic() * 1000) - start_ms
+
         # Persist the new thread ID so subsequent turns resume the same session.
         usage = turn.usage()
         if usage.model:

src/agentex/lib/adk/_modules/_codex_turn.py

@@ -152,8 +152,11 @@ def __init__(
     ) -> None:
         self._raw_events = events
         self._model = model
-        self._duration_ms = duration_ms
-        self._cost_usd = cost_usd
+        # Public + mutable: the true wall-clock duration (and cost) is usually
+        # only known after the stream is consumed, so callers may set these
+        # after construction and before calling usage().
+        self.duration_ms = duration_ms
+        self.cost_usd = cost_usd
 
         # Populated by the on_result callback once the stream is exhausted.
         self._result: dict[str, Any] | None = None
@@ -196,6 +199,6 @@ def usage(self) -> TurnUsage:
             model=self._model,
             tool_call_count=self._result.get("tool_call_count", 0),
             reasoning_count=self._result.get("reasoning_count", 0),
-            duration_ms=self._duration_ms,
-            cost_usd=self._cost_usd,
+            duration_ms=self.duration_ms,
+            cost_usd=self.cost_usd,
         )