feat: #1994 Add a list of per-request usage data to Usage

src/agents/usage.py

@@ -4,6 +4,30 @@
 from pydantic.dataclasses import dataclass
 
 
+@dataclass
+class IndividualRequestUsage:
+    """Usage details for a single API request.
+
+    This is useful for cost calculation when different pricing rates apply based on
+    per-request token counts (e.g., Anthropic's 200K token threshold pricing).
+    """
+
+    input_tokens: int
+    """Input tokens for this individual request."""
+
+    output_tokens: int
+    """Output tokens for this individual request."""
+
+    total_tokens: int
+    """Total tokens (input + output) for this individual request."""
+
+    input_tokens_details: InputTokensDetails
+    """Details about the input tokens for this individual request."""
+
+    output_tokens_details: OutputTokensDetails
+    """Details about the output tokens for this individual request."""
+
+
 @dataclass
 class Usage:
     requests: int = 0
@@ -27,7 +51,34 @@ class Usage:
     total_tokens: int = 0
     """Total tokens sent and received, across all requests."""
 
+    individual_requests: list[IndividualRequestUsage] = field(default_factory=list)
+    """List of individual request usage details for accurate per-request cost calculation.
+
+    This field preserves the token counts for each individual API request made during a run.
+    This is particularly useful for providers like Anthropic that have different pricing
+    tiers based on per-request token counts (e.g., different rates for requests with more
+    or fewer than 200K tokens).
+
+    Each call to `add()` automatically creates an entry in this list if the added usage
+    represents a new request (i.e., has non-zero tokens).
+
+    Example:
+        For a run that makes 3 API calls with 100K, 150K, and 80K input tokens each,
+        the aggregated `input_tokens` would be 330K, but `individual_requests` would
+        preserve the [100K, 150K, 80K] breakdown needed for accurate cost calculation.
+    """
+
     def add(self, other: "Usage") -> None:
+        """Add another Usage object to this one, aggregating all fields.
+
+        This method automatically preserves individual request details for accurate
+        cost calculation. When adding a Usage object that represents a single request
+        (requests=1), it creates an IndividualRequestUsage entry to preserve the
+        per-request token breakdown.
+
+        Args:
+            other: The Usage object to add to this one.
+        """
         self.requests += other.requests if other.requests else 0
         self.input_tokens += other.input_tokens if other.input_tokens else 0
         self.output_tokens += other.output_tokens if other.output_tokens else 0
@@ -41,3 +92,18 @@ def add(self, other: "Usage") -> None:
             reasoning_tokens=self.output_tokens_details.reasoning_tokens
             + other.output_tokens_details.reasoning_tokens
         )
+
+        # Automatically preserve individual request details for accurate cost calculation.
+        # If the other Usage represents a single request with tokens, record it.
+        if other.requests == 1 and other.total_tokens > 0:
+            individual_usage = IndividualRequestUsage(
+                input_tokens=other.input_tokens,
+                output_tokens=other.output_tokens,
+                total_tokens=other.total_tokens,
+                input_tokens_details=other.input_tokens_details,
+                output_tokens_details=other.output_tokens_details,
+            )
+            self.individual_requests.append(individual_usage)
+        elif other.individual_requests:
+            # If the other Usage already has individual request breakdowns, merge them.
+            self.individual_requests.extend(other.individual_requests)

tests/test_usage.py

@@ -1,6 +1,6 @@
 from openai.types.responses.response_usage import InputTokensDetails, OutputTokensDetails
 
-from agents.usage import Usage
+from agents.usage import IndividualRequestUsage, Usage
 
 
 def test_usage_add_aggregates_all_fields():
@@ -50,3 +50,220 @@ def test_usage_add_aggregates_with_none_values():
     assert u1.total_tokens == 15
     assert u1.input_tokens_details.cached_tokens == 4
     assert u1.output_tokens_details.reasoning_tokens == 6
+
+
+def test_individual_request_usage_creation():
+    """Test that IndividualRequestUsage is created correctly."""
+    individual = IndividualRequestUsage(
+        input_tokens=100,
+        output_tokens=200,
+        total_tokens=300,
+        input_tokens_details=InputTokensDetails(cached_tokens=10),
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=20),
+    )
+
+    assert individual.input_tokens == 100
+    assert individual.output_tokens == 200
+    assert individual.total_tokens == 300
+    assert individual.input_tokens_details.cached_tokens == 10
+    assert individual.output_tokens_details.reasoning_tokens == 20
+
+
+def test_usage_add_preserves_single_request():
+    """Test that adding a single request Usage creates an IndividualRequestUsage entry."""
+    u1 = Usage()
+    u2 = Usage(
+        requests=1,
+        input_tokens=100,
+        input_tokens_details=InputTokensDetails(cached_tokens=10),
+        output_tokens=200,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=20),
+        total_tokens=300,
+    )
+
+    u1.add(u2)
+
+    # Should preserve the individual request details
+    assert len(u1.individual_requests) == 1
+    individual = u1.individual_requests[0]
+    assert individual.input_tokens == 100
+    assert individual.output_tokens == 200
+    assert individual.total_tokens == 300
+    assert individual.input_tokens_details.cached_tokens == 10
+    assert individual.output_tokens_details.reasoning_tokens == 20
+
+
+def test_usage_add_ignores_zero_token_requests():
+    """Test that zero-token requests don't create IndividualRequestUsage entries."""
+    u1 = Usage()
+    u2 = Usage(
+        requests=1,
+        input_tokens=0,
+        input_tokens_details=InputTokensDetails(cached_tokens=0),
+        output_tokens=0,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=0),
+        total_tokens=0,
+    )
+
+    u1.add(u2)
+
+    # Should not create an individual request entry for zero tokens
+    assert len(u1.individual_requests) == 0
+
+
+def test_usage_add_ignores_multi_request_usage():
+    """Test that multi-request Usage objects don't create IndividualRequestUsage entries."""
+    u1 = Usage()
+    u2 = Usage(
+        requests=3,  # Multiple requests
+        input_tokens=100,
+        input_tokens_details=InputTokensDetails(cached_tokens=10),
+        output_tokens=200,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=20),
+        total_tokens=300,
+    )
+
+    u1.add(u2)
+
+    # Should not create an individual request entry for multi-request usage
+    assert len(u1.individual_requests) == 0
+
+
+def test_usage_add_merges_existing_individual_requests():
+    """Test that existing individual_requests are merged when adding Usage objects."""
+    # Create first usage with individual requests
+    u1 = Usage()
+    u2 = Usage(
+        requests=1,
+        input_tokens=100,
+        input_tokens_details=InputTokensDetails(cached_tokens=10),
+        output_tokens=200,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=20),
+        total_tokens=300,
+    )
+    u1.add(u2)
+
+    # Create second usage with individual requests
+    u3 = Usage(
+        requests=1,
+        input_tokens=50,
+        input_tokens_details=InputTokensDetails(cached_tokens=5),
+        output_tokens=75,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=10),
+        total_tokens=125,
+    )
+
+    u1.add(u3)
+
+    # Should have both individual requests
+    assert len(u1.individual_requests) == 2
+
+    # First request
+    first = u1.individual_requests[0]
+    assert first.input_tokens == 100
+    assert first.output_tokens == 200
+    assert first.total_tokens == 300
+
+    # Second request
+    second = u1.individual_requests[1]
+    assert second.input_tokens == 50
+    assert second.output_tokens == 75
+    assert second.total_tokens == 125
+
+
+def test_usage_add_with_pre_existing_individual_requests():
+    """Test adding Usage objects that already have individual_requests."""
+    u1 = Usage()
+
+    # Create a usage with individual requests
+    u2 = Usage(
+        requests=1,
+        input_tokens=100,
+        input_tokens_details=InputTokensDetails(cached_tokens=10),
+        output_tokens=200,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=20),
+        total_tokens=300,
+    )
+    u1.add(u2)
+
+    # Create another usage with individual requests
+    u3 = Usage(
+        requests=1,
+        input_tokens=50,
+        input_tokens_details=InputTokensDetails(cached_tokens=5),
+        output_tokens=75,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=10),
+        total_tokens=125,
+    )
+
+    # Add u3 to u1
+    u1.add(u3)
+
+    # Should have both individual requests
+    assert len(u1.individual_requests) == 2
+    assert u1.individual_requests[0].input_tokens == 100
+    assert u1.individual_requests[1].input_tokens == 50
+
+
+def test_usage_individual_requests_default_empty():
+    """Test that individual_requests defaults to an empty list."""
+    u = Usage()
+    assert u.individual_requests == []
+
+
+def test_anthropic_cost_calculation_scenario():
+    """Test a realistic scenario for Anthropic cost calculation with 200K token thresholds."""
+    # Simulate 3 API calls: 100K, 150K, and 80K input tokens each
+    # None exceed 200K, so they should all use the lower pricing tier
+
+    usage = Usage()
+
+    # First request: 100K input tokens
+    req1 = Usage(
+        requests=1,
+        input_tokens=100_000,
+        input_tokens_details=InputTokensDetails(cached_tokens=0),
+        output_tokens=50_000,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=0),
+        total_tokens=150_000,
+    )
+    usage.add(req1)
+
+    # Second request: 150K input tokens
+    req2 = Usage(
+        requests=1,
+        input_tokens=150_000,
+        input_tokens_details=InputTokensDetails(cached_tokens=0),
+        output_tokens=75_000,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=0),
+        total_tokens=225_000,
+    )
+    usage.add(req2)
+
+    # Third request: 80K input tokens
+    req3 = Usage(
+        requests=1,
+        input_tokens=80_000,
+        input_tokens_details=InputTokensDetails(cached_tokens=0),
+        output_tokens=40_000,
+        output_tokens_details=OutputTokensDetails(reasoning_tokens=0),
+        total_tokens=120_000,
+    )
+    usage.add(req3)
+
+    # Verify aggregated totals
+    assert usage.requests == 3
+    assert usage.input_tokens == 330_000  # 100K + 150K + 80K
+    assert usage.output_tokens == 165_000  # 50K + 75K + 40K
+    assert usage.total_tokens == 495_000  # 150K + 225K + 120K
+
+    # Verify individual request preservation
+    assert len(usage.individual_requests) == 3
+    assert usage.individual_requests[0].input_tokens == 100_000
+    assert usage.individual_requests[1].input_tokens == 150_000
+    assert usage.individual_requests[2].input_tokens == 80_000
+
+    # All individual requests are under 200K threshold
+    for req in usage.individual_requests:
+        assert req.input_tokens < 200_000
+        assert req.output_tokens < 200_000