[management, proxy] Add prompt-cache token and cost accounting to agent network usage (#6900)

Co-authored-by: braginini <bangvalo@gmail.com>
This commit is contained in:
Maycon Santos
2026-07-26 21:42:41 +02:00
committed by GitHub
co-authored by braginini
parent 1e5b0a5c89
commit 4f6247b5c3
29 changed files with 1582 additions and 170 deletions
@@ -69,15 +69,18 @@ func applyBedrockInvokeChunk(payload []byte, usage *llm.Usage, completion *strin
}
// converseStreamEvent captures the Converse stream frames carrying completion
// text (contentBlockDelta) and the final token usage (metadata).
// text (contentBlockDelta) and the final token usage (metadata). Cache buckets
// are additive to inputTokens (AWS write bucket: cacheWriteInputTokens).
type converseStreamEvent struct {
Delta *struct {
Text string `json:"text"`
} `json:"delta"`
Usage *struct {
InputTokens int64 `json:"inputTokens"`
OutputTokens int64 `json:"outputTokens"`
TotalTokens int64 `json:"totalTokens"`
InputTokens int64 `json:"inputTokens"`
OutputTokens int64 `json:"outputTokens"`
TotalTokens int64 `json:"totalTokens"`
CacheReadTokens int64 `json:"cacheReadInputTokens"`
CacheWriteTokens int64 `json:"cacheWriteInputTokens"`
} `json:"usage"`
}
@@ -105,6 +108,12 @@ func applyConverseStreamEvent(eventType string, payload []byte, usage *llm.Usage
if ev.Usage.TotalTokens > 0 {
usage.TotalTokens = ev.Usage.TotalTokens
}
if ev.Usage.CacheReadTokens > 0 {
usage.CachedInputTokens = ev.Usage.CacheReadTokens
}
if ev.Usage.CacheWriteTokens > 0 {
usage.CacheCreationTokens = ev.Usage.CacheWriteTokens
}
}
}
}
@@ -66,6 +66,24 @@ func TestAccumulateBedrockStream_Converse(t *testing.T) {
require.Equal(t, "pong", completion, "converse text deltas concatenated")
}
// The converse-stream metadata frame's camelCase cache fields must reach the billed cache buckets.
func TestAccumulateBedrockStream_ConverseCacheBuckets(t *testing.T) {
var body bytes.Buffer
body.Write(bedrockFrame(t, "contentBlockDelta", mustJSON(t, map[string]any{"delta": map[string]any{"text": "pong"}})))
body.Write(bedrockFrame(t, "metadata", mustJSON(t, map[string]any{"usage": map[string]any{
"inputTokens": 11, "outputTokens": 3, "totalTokens": 30,
"cacheReadInputTokens": 7, "cacheWriteInputTokens": 9,
}})))
usage, completion := accumulateBedrockStream(body.Bytes())
require.Equal(t, int64(11), usage.InputTokens, "input tokens from metadata frame")
require.Equal(t, int64(3), usage.OutputTokens, "output tokens from metadata frame")
require.Equal(t, int64(7), usage.CachedInputTokens, "cache-read tokens from metadata frame")
require.Equal(t, int64(9), usage.CacheCreationTokens, "cache-write tokens from metadata frame")
require.Equal(t, int64(30), usage.TotalTokens, "provider-reported total wins")
require.Equal(t, "pong", completion)
}
func TestAccumulateBedrockStream_Truncated(t *testing.T) {
// A body cut mid-frame must not panic; partial usage is returned.
full := bedrockFrame(t, "metadata", mustJSON(t, map[string]any{"usage": map[string]any{"inputTokens": 11, "outputTokens": 3}}))