core/agent: always emit UsageEvent with api or estimate

Every model round yields usage; prefer provider usage, else estimate
from request messages and assistant content.

src/plyngent/agent/__init__.py

@@ -20,4 +20,6 @@ from .tools import ToolRegistry as ToolRegistry
 from .tools import schema_from_callable as schema_from_callable
 from .tools import tool as tool
 from .usage import TokenUsage as TokenUsage
+from .usage import chars_to_tokens as chars_to_tokens
+from .usage import estimate_token_usage as estimate_token_usage
 from .usage import token_usage_from_api as token_usage_from_api

src/plyngent/agent/events.py

@@ -41,7 +41,10 @@ class CancelledEvent(Struct, tag_field="type", tag="cancelled"):
 
 
 class UsageEvent(Struct, tag_field="type", tag="usage"):
-    """Token usage for one model completion (one tool-loop round)."""
+    """Token usage for one model completion (one tool-loop round).
+
+    ``source`` is mirrored from ``usage.source`` (``api`` / ``estimate``).
+    """
 
     usage: TokenUsage
 

src/plyngent/agent/loop.py

@@ -35,7 +35,7 @@ from .events import (
     ToolResultEvent,
     UsageEvent,
 )
-from .usage import token_usage_from_api
+from .usage import resolve_round_usage, token_usage_from_api
 
 if TYPE_CHECKING:
     from collections.abc import AsyncIterator, Awaitable, Callable, Sequence
@@ -126,9 +126,9 @@ async def _non_stream_round(
     if isinstance(assistant.content, str) and assistant.content:
         yield TextDeltaEvent(content=assistant.content)
     yield AssistantMessageEvent(message=assistant)
-    usage = token_usage_from_api(response.usage)
-    if usage is not None:
-        yield UsageEvent(usage=usage)
+    yield UsageEvent(
+        usage=resolve_round_usage(response.usage, param.messages, assistant),
+    )
 
 
 async def _stream_round(
@@ -140,6 +140,7 @@ async def _stream_round(
     Pattern: ``stream = await client.chat_completions(..., stream=True)`` then
     ``async for chunk in stream``. Tool-call deltas are merged after the stream.
     Requests ``stream_options.include_usage`` so providers may send a final usage chunk.
+    Falls back to a char-based token estimate when the provider omits usage.
     """
     stream_param = msgspec.structs.replace(
         param,
@@ -148,12 +149,12 @@ async def _stream_round(
     stream = await client.chat_completions(stream_param, stream=True)
     content_parts: list[str] = []
     tool_deltas: list[StreamToolCallDelta] = []
-    last_usage = None
+    last_api_usage: object = UNSET
 
     async for chunk in stream:
-        usage = token_usage_from_api(chunk.usage)
-        if usage is not None:
-            last_usage = usage
+        parsed = token_usage_from_api(chunk.usage)
+        if parsed is not None:
+            last_api_usage = chunk.usage
         if not chunk.choices:
             continue
         choice = chunk.choices[0]
@@ -176,8 +177,9 @@ async def _stream_round(
         tool_calls=tool_calls,
     )
     yield AssistantMessageEvent(message=assistant)
-    if last_usage is not None:
-        yield UsageEvent(usage=last_usage)
+    yield UsageEvent(
+        usage=resolve_round_usage(last_api_usage, param.messages, assistant),
+    )
 
 
 async def _assistant_round(