Skip to content

fix(executor): honor explicit-zero Claude stream usage fields #5832

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
rome-xi wants to merge 1 commit into
base: dev
Choose a base branch
from
+254 −35
Open

fix(executor): honor explicit-zero Claude stream usage fields #5832

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 7 additions & 15 deletions internal/runtime/executor/helps/plugin_executor_usage.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,8 +40,8 @@ func ObservePluginExecutorStreamUsage(protocol string, payload []byte, buffer *S
switch strings.ToLower(strings.TrimSpace(protocol)) {
case "claude":
IterateStreamLines(payload, func(line []byte) {
if detail, ok := parseClaudeStreamLine(line); ok {
ObserveMergedStreamUsage(buffer, detail)
if update, ok := parseClaudeStreamLine(line); ok {
observeClaudeUsageUpdate(buffer, update)
}
})
case "gemini":
Expand Down Expand Up @@ -111,20 +111,12 @@ func parseClaudePayloadUsage(payload []byte) usage.Detail {
return ParseClaudeUsage([]byte(`{"usage":` + usageNode.Raw + `}`))
}

func parseClaudeStreamLine(line []byte) (usage.Detail, bool) {
payload := ExtractStreamJSONPayload(line)
if len(payload) == 0 || !gjson.ValidBytes(payload) {
return usage.Detail{}, false
}
usageNode := gjson.GetBytes(payload, "usage")
if !usageNode.Exists() {
usageNode = gjson.GetBytes(payload, "message.usage")
}
if !usageNode.Exists() {
return usage.Detail{}, false
func parseClaudeStreamLine(line []byte) (claudeUsageUpdate, bool) {
usageNode, ok := claudeStreamUsageNode(line)
if !ok {
return claudeUsageUpdate{}, false
}
detail := ParseClaudeUsage([]byte(`{"usage":` + usageNode.Raw + `}`))
return detail, true
return parseClaudeUsageUpdate(usageNode), true
}

// ObserveMergedStreamUsage updates buffer with merged usage details.
Expand Down
123 changes: 103 additions & 20 deletions internal/runtime/executor/helps/usage_helpers.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -710,13 +710,28 @@ func (b *StreamUsageBuffer) ObserveOpenAIStream(line []byte) {
}

// ObserveClaudeStream records and merges usage from a Claude SSE line.
// Present fields, including explicit zeros, overwrite prior values; omitted or
// JSON-null fields keep the previously observed value.
func (b *StreamUsageBuffer) ObserveClaudeStream(line []byte) {
if b == nil {
return
}
if detail, ok := ParseClaudeStreamUsage(line); ok {
ObserveMergedStreamUsage(b, detail)
usageNode, ok := claudeStreamUsageNode(line)
if !ok {
return
}
observeClaudeUsageUpdate(b, parseClaudeUsageUpdate(usageNode))
}

func observeClaudeUsageUpdate(b *StreamUsageBuffer, update claudeUsageUpdate) {
if b == nil {
return
}
if existing, ok := b.Detail(); ok {
b.Observe(mergeClaudeStreamUsage(existing, update), true)
return
}
b.Observe(update.detail, true)
}

// Publish emits the latest observed usage detail, if any.
Expand Down Expand Up @@ -905,36 +920,111 @@ func ParseClaudeUsage(data []byte) usage.Detail {
}

func ParseClaudeStreamUsage(line []byte) (usage.Detail, bool) {
usageNode, ok := claudeStreamUsageNode(line)
if !ok {
return usage.Detail{}, false
}
return parseClaudeUsageNode(usageNode), true
}

func claudeStreamUsageNode(line []byte) (gjson.Result, bool) {
payload := jsonPayload(line)
if len(payload) == 0 || !gjson.ValidBytes(payload) {
return usage.Detail{}, false
return gjson.Result{}, false
}
usageNode := gjson.GetBytes(payload, "usage")
if !usageNode.Exists() {
usageNode = gjson.GetBytes(payload, "message.usage")
}
if !usageNode.Exists() {
return usage.Detail{}, false
return gjson.Result{}, false
}
return parseClaudeUsageNode(usageNode), true
return usageNode, true
}

func parseClaudeUsageNode(usageNode gjson.Result) usage.Detail {
cacheReadTokens := usageNode.Get("cache_read_input_tokens").Int()
cacheCreationTokens := usageNode.Get("cache_creation_input_tokens").Int()
rawOutputTokens := usageNode.Get("output_tokens").Int()
return parseClaudeUsageUpdate(usageNode).detail
}

type claudeUsageUpdate struct {
detail usage.Detail
hasInput bool
hasOutput bool
hasCacheRead bool
hasCacheCreation bool
hasReasoning bool
}

func claudeUsageFieldPresent(node gjson.Result) bool {
return node.Exists() && node.Type != gjson.Null
}

func parseClaudeUsageUpdate(usageNode gjson.Result) claudeUsageUpdate {
inputNode := usageNode.Get("input_tokens")
outputNode := usageNode.Get("output_tokens")
cacheReadNode := usageNode.Get("cache_read_input_tokens")
cacheCreationNode := usageNode.Get("cache_creation_input_tokens")
// Anthropic reports thinking as a subset of output_tokens. Prefer the official
// nested field, then fall back to legacy aliases used by some gateways.
reasoningNode := firstExistingUsageNode(
reasoningNode := firstPresentClaudeUsageNode(
usageNode,
"output_tokens_details.thinking_tokens",
"output_tokens_details.reasoning_tokens",
"thinking_tokens",
)
reasoningTokens := reasoningNode.Int()
if reasoningTokens < 0 {
reasoningTokens = 0
detail := usage.Detail{
InputTokens: inputNode.Int(),
OutputTokens: outputNode.Int(),
ReasoningTokens: reasoningNode.Int(),
CacheReadTokens: cacheReadNode.Int(),
CacheCreationTokens: cacheCreationNode.Int(),
}
return claudeUsageUpdate{
detail: finalizeClaudeUsageDetail(detail),
hasInput: claudeUsageFieldPresent(inputNode),
hasOutput: claudeUsageFieldPresent(outputNode),
hasCacheRead: claudeUsageFieldPresent(cacheReadNode),
hasCacheCreation: claudeUsageFieldPresent(cacheCreationNode),
hasReasoning: claudeUsageFieldPresent(reasoningNode),
}
}

func firstPresentClaudeUsageNode(root gjson.Result, paths ...string) gjson.Result {
for _, path := range paths {
node := root.Get(path)
if claudeUsageFieldPresent(node) {
return node
}
}
return gjson.Result{}
}

func mergeClaudeStreamUsage(existing usage.Detail, update claudeUsageUpdate) usage.Detail {
merged := existing
if update.hasInput {
merged.InputTokens = update.detail.InputTokens
}
if update.hasOutput {
merged.OutputTokens = update.detail.OutputTokens
}
if update.hasCacheRead {
merged.CacheReadTokens = update.detail.CacheReadTokens
}
if update.hasCacheCreation {
merged.CacheCreationTokens = update.detail.CacheCreationTokens
}
if update.hasReasoning {
merged.ReasoningTokens = update.detail.ReasoningTokens
}
return finalizeClaudeUsageDetail(merged)
}

func finalizeClaudeUsageDetail(detail usage.Detail) usage.Detail {
if detail.ReasoningTokens < 0 {
detail.ReasoningTokens = 0
}
rawOutputTokens := detail.OutputTokens
reasoningTokens := detail.ReasoningTokens
nonReasoningOutput := rawOutputTokens
if reasoningTokens > 0 && reasoningTokens <= rawOutputTokens {
nonReasoningOutput = rawOutputTokens - reasoningTokens
Expand All @@ -944,14 +1034,7 @@ func parseClaudeUsageNode(usageNode gjson.Result) usage.Detail {
// is inconsistent.
nonReasoningOutput = 0
}
detail := usage.Detail{
InputTokens: usageNode.Get("input_tokens").Int(),
OutputTokens: rawOutputTokens,
ReasoningTokens: reasoningTokens,
CachedTokens: cacheReadTokens,
CacheReadTokens: cacheReadTokens,
CacheCreationTokens: cacheCreationTokens,
}
detail.CachedTokens = detail.CacheReadTokens
if detail.CachedTokens == 0 {
detail.CachedTokens = detail.CacheCreationTokens
}
Expand Down
144 changes: 144 additions & 0 deletions internal/runtime/executor/helps/usage_helpers_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1015,6 +1015,150 @@ func TestStreamUsageBufferObserveClaudeStream_MergesStartAndDelta(t *testing.T)
}
}

func TestStreamUsageBufferObserveClaudeStream_ExplicitZeroOverwritesPrior(t *testing.T) {
var buffer StreamUsageBuffer

lineStart := []byte(`data: {"type":"message_start","message":{"id":"msg_123","model":"claude-opus-5","usage":{"input_tokens":2095,"cache_creation_input_tokens":7185,"cache_read_input_tokens":100,"output_tokens":1}}}`)
buffer.ObserveClaudeStream(lineStart)

// Full cache hit: final usage reports explicit zeros for uncached input and
// cache creation, with cache_read covering the prompt.
lineDelta := []byte(`data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"input_tokens":0,"cache_creation_input_tokens":0,"cache_read_input_tokens":93780,"output_tokens":15}}`)
buffer.ObserveClaudeStream(lineDelta)

detail, ok := buffer.Detail()
if !ok {
t.Fatal("expected buffer to contain usage detail")
}
if detail.InputTokens != 0 {
t.Errorf("InputTokens = %d, want 0", detail.InputTokens)
}
if detail.OutputTokens != 15 {
t.Errorf("OutputTokens = %d, want 15", detail.OutputTokens)
}
if detail.CacheReadTokens != 93780 {
t.Errorf("CacheReadTokens = %d, want 93780", detail.CacheReadTokens)
}
if detail.CacheCreationTokens != 0 {
t.Errorf("CacheCreationTokens = %d, want 0", detail.CacheCreationTokens)
}
if detail.CachedTokens != 93780 {
t.Errorf("CachedTokens = %d, want 93780", detail.CachedTokens)
}
wantTotal := int64(0 + 15 + 93780 + 0)
if detail.TotalTokens != wantTotal {
t.Errorf("TotalTokens = %d, want %d", detail.TotalTokens, wantTotal)
}
}

func TestStreamUsageBufferObserveClaudeStream_NullInputPreservesStart(t *testing.T) {
var buffer StreamUsageBuffer

lineStart := []byte(`data: {"type":"message_start","message":{"id":"msg_123","model":"claude-opus-5","usage":{"input_tokens":2095,"cache_creation_input_tokens":7185,"cache_read_input_tokens":355598,"output_tokens":1}}}`)
buffer.ObserveClaudeStream(lineStart)

lineDelta := []byte(`data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"input_tokens":null,"output_tokens":15}}`)
buffer.ObserveClaudeStream(lineDelta)

detail, ok := buffer.Detail()
if !ok {
t.Fatal("expected buffer to contain usage detail")
}
if detail.InputTokens != 2095 {
t.Errorf("InputTokens = %d, want 2095", detail.InputTokens)
}
if detail.OutputTokens != 15 {
t.Errorf("OutputTokens = %d, want 15", detail.OutputTokens)
}
if detail.CacheReadTokens != 355598 {
t.Errorf("CacheReadTokens = %d, want 355598", detail.CacheReadTokens)
}
if detail.CacheCreationTokens != 7185 {
t.Errorf("CacheCreationTokens = %d, want 7185", detail.CacheCreationTokens)
}
}

func TestStreamUsageBufferObserveClaudeStream_OmittedLaterUsagePreservesPrior(t *testing.T) {
var buffer StreamUsageBuffer

lineStart := []byte(`data: {"type":"message_start","message":{"id":"msg_123","model":"claude-opus-5","usage":{"input_tokens":2095,"cache_creation_input_tokens":7185,"cache_read_input_tokens":355598,"output_tokens":1}}}`)
buffer.ObserveClaudeStream(lineStart)

buffer.ObserveClaudeStream([]byte(`data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"hi"}}`))

detail, ok := buffer.Detail()
if !ok {
t.Fatal("expected buffer to contain usage detail")
}
if detail.InputTokens != 2095 {
t.Errorf("InputTokens = %d, want 2095", detail.InputTokens)
}
if detail.OutputTokens != 1 {
t.Errorf("OutputTokens = %d, want 1", detail.OutputTokens)
}
if detail.CacheReadTokens != 355598 {
t.Errorf("CacheReadTokens = %d, want 355598", detail.CacheReadTokens)
}
if detail.CacheCreationTokens != 7185 {
t.Errorf("CacheCreationTokens = %d, want 7185", detail.CacheCreationTokens)
}
}

func TestStreamUsageBufferObserveClaudeStream_LatePositiveInputReplaces(t *testing.T) {
var buffer StreamUsageBuffer

lineStart := []byte(`data: {"type":"message_start","message":{"id":"msg_123","model":"claude-opus-5","usage":{"input_tokens":2095,"cache_creation_input_tokens":7185,"cache_read_input_tokens":355598,"output_tokens":1}}}`)
buffer.ObserveClaudeStream(lineStart)

lineDelta := []byte(`data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"input_tokens":42,"output_tokens":15}}`)
buffer.ObserveClaudeStream(lineDelta)

detail, ok := buffer.Detail()
if !ok {
t.Fatal("expected buffer to contain usage detail")
}
if detail.InputTokens != 42 {
t.Errorf("InputTokens = %d, want 42", detail.InputTokens)
}
if detail.OutputTokens != 15 {
t.Errorf("OutputTokens = %d, want 15", detail.OutputTokens)
}
if detail.CacheReadTokens != 355598 {
t.Errorf("CacheReadTokens = %d, want 355598", detail.CacheReadTokens)
}
if detail.CacheCreationTokens != 7185 {
t.Errorf("CacheCreationTokens = %d, want 7185", detail.CacheCreationTokens)
}
wantTotal := int64(42 + 15 + 355598 + 7185)
if detail.TotalTokens != wantTotal {
t.Errorf("TotalTokens = %d, want %d", detail.TotalTokens, wantTotal)
}
}

func TestObservePluginExecutorStreamUsage_ClaudeExplicitZeroOverwritesPrior(t *testing.T) {
var buffer StreamUsageBuffer
payload := []byte("data: {\"type\":\"message_start\",\"message\":{\"id\":\"msg_123\",\"model\":\"claude-opus-5\",\"usage\":{\"input_tokens\":2095,\"cache_creation_input_tokens\":7185,\"cache_read_input_tokens\":100,\"output_tokens\":1}}}\n" +
"data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\"},\"usage\":{\"input_tokens\":0,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":93780,\"output_tokens\":15}}\n")
ObservePluginExecutorStreamUsage("claude", payload, &buffer)

detail, ok := buffer.Detail()
if !ok {
t.Fatal("expected buffer to contain usage detail")
}
if detail.InputTokens != 0 {
t.Errorf("InputTokens = %d, want 0", detail.InputTokens)
}
if detail.OutputTokens != 15 {
t.Errorf("OutputTokens = %d, want 15", detail.OutputTokens)
}
if detail.CacheReadTokens != 93780 {
t.Errorf("CacheReadTokens = %d, want 93780", detail.CacheReadTokens)
}
if detail.CacheCreationTokens != 0 {
t.Errorf("CacheCreationTokens = %d, want 0", detail.CacheCreationTokens)
}
}

func TestStreamUsageBufferObserveClaudeStream_FailurePreservesUsage(t *testing.T) {
var buffer StreamUsageBuffer

Expand Down
Loading