"j %SrSSKJr SSKJr SSKJr SSKJr SSK J r J r Sr \S "S S S 9r S \S'\SSj5rSSjrSSjrSSjrg )uRecord run usage, attributing it to the agent run that produced it. An agent run span reports the usage of the requests *that run* made — not its nested runs', which report their own. Summing the agent-run spans in a trace then gives the conversation's total, matching the sum of the `chat` spans underneath them, which is the point of keeping agent-run usage in its own attribute namespace: a backend can add these up without counting anything twice. `RunUsage` alone can't say who produced what. It is accumulated into in place, and the multi-agent delegation pattern hands the *same* object to concurrent delegates (`usage=ctx.usage`), so neither the object's contents nor an end-minus-start delta on it distinguishes this run's requests from a sibling's — concurrent delegates absorb each other. Which run is producing is a property of the call stack, so that is what this follows. [`accumulate`] [] makes a run's `RunUsage` the one credited for as long as its span is open, and restores the enclosing run's on the way out; the `record_*` functions credit only that innermost run. Asyncio copies the context when a task is created, so concurrent delegates each start from the parent's and replace it with their own, never seeing each other's. This module owns the *only* in-place mutation of a run's usage. Incrementing a `RunUsage` field directly leaves its tokens off the run's span, so `tests/test_usage_attribution.py` fails on a bare `requests += `, `tool_calls += `, or `.incr(` that isn't marked `# usage-attribution: ok` with a reason. ) annotations) Generator)contextmanager) ContextVar) RequestUsageRunUsage) accumulaterecord_requestrecord_tool_call record_usagezRunUsage | Nonezpydantic_ai.usage_attributionN)defaultzContextVar[RunUsage | None]_activec## [RU5nSv [RU5 g![RU5 f=f7f)zCredit `run_usage` with what is recorded in this context, until the block exits. A nested run replaces it for the length of its own span, so what the nested run records is its own; resetting on the way out hands crediting back to the enclosing run. N)rsetreset) run_usagetokens [/home/mande/repo/quber/.venv/lib/python3.13/site-packages/pydantic_ai/_usage_attribution.pyr r &s5 KK "E  e esA 3A A  A cU=RS- sl[R5=nbU=RS- slgg)zRCount one model request against this run's usage and against the run that made it.rN)requestsrgetusagers rr r 4s7 NNaN[[]" /a0cU=RS- sl[R5=nbU=RS- slgg)zYCount one successful tool call against this run's usage and against the run that made it.rN) tool_callsrrrs rr r ;s: [[]" /!0rcxURU5 [R5=nbURU5 gg)uAdd recorded usage to this run's usage and to the run that produced it. `recorded` is one response's `RequestUsage`, or the `RunUsage` delta a durable operation accumulated across the boundary — which carries its own requests and tool calls. N)incrrr)rrecordedrs rr r Bs2  JJx[[]" /x 0r)rr returnzGenerator[None])rr r!None)rr r zRunUsage | RequestUsager!r")__doc__ __future__rcollections.abcr contextlibr contextvarsrrrr __all__r__annotations__r r r r rrr+s]0#%%") N'12C'DEdnr's $s   "!r