fix(agentexec): count Claude cache tokens as prompt tokens, as the built-in loop does

Claude's input_tokens excludes cache reads and writes; the built-in
loop's Anthropic provider folds them into PromptTokens so budgets count
total prompt size. The harness adapter now does the same, so one budget
means the same on either runner.
This commit is contained in:
sam committed 2026-09-26 09:11:57 +10:00
1 parent 4bbfd03c5c
commit 0eef8ed5ba
2 files changed
+14 -19

No files matched your search

+10 -15
View File
@@ -55,13 +55,8 @@ func (o *claudeCodeHarnessOutput) Line(line []byte, report ToolCallReporter) {
func (o *claudeCodeHarnessOutput) assistant(raw json.RawMessage, _ ToolCallReporter) {
var message struct {
ID string `json:"id"`
Usage *struct {
InputTokens int `json:"input_tokens"`
OutputTokens int `json:"output_tokens"`
CacheReadInputTokens int `json:"cache_read_input_tokens"`
CacheCreationInputTokens int `json:"cache_creation_input_tokens"`
} `json:"usage"`
ID string `json:"id"`
Usage json.RawMessage `json:"usage"`
Content []struct {
Type string `json:"type"`
ID string `json:"id"`
@@ -72,13 +67,8 @@ func (o *claudeCodeHarnessOutput) assistant(raw json.RawMessage, _ ToolCallRepor
if json.Unmarshal(raw, &message) != nil {
return
}
if message.Usage != nil && (message.ID == "" || o.markMessage(message.ID)) {
o.messageUsage = addUsage(o.messageUsage, usageFromFields(
message.Usage.InputTokens,
message.Usage.OutputTokens,
message.Usage.CacheReadInputTokens,
message.Usage.CacheCreationInputTokens,
))
if usage, ok := decodeClaudeUsage(message.Usage); ok && (message.ID == "" || o.markMessage(message.ID)) {
o.messageUsage = addUsage(o.messageUsage, usage)
}
for _, block := range message.Content {
if block.Type == "tool_use" && block.ID != "" && block.Name != "" {
@@ -147,9 +137,14 @@ func decodeClaudeUsage(raw json.RawMessage) (Usage, bool) {
return usageFromFields(fields.InputTokens, fields.OutputTokens, fields.CacheReadInputTokens, fields.CacheCreationInputTokens), true
}
// usageFromFields folds cache reads and writes into PromptTokens, as the
// built-in loop's Anthropic provider does: Claude's input_tokens excludes
// them, and budgets must count a harness run the way they count a built-in
// one.
func usageFromFields(input, output, cacheRead, cacheCreation int) Usage {
prompt := input + cacheRead + cacheCreation
return Usage{
PromptTokens: input, CompletionTokens: output, TotalTokens: input + output,
PromptTokens: prompt, CompletionTokens: output, TotalTokens: prompt + output,
CachedTokens: cacheRead, CacheCreationTokens: cacheCreation,
}
}
@@ -23,7 +23,7 @@ func TestClaudeCodeHarnessOutputFixtures(t *testing.T) {
{Tool: "Bash", Detail: "hello"},
{Tool: "Bash", Detail: errorDetail, Failed: true},
},
wantUsage: Usage{PromptTokens: 6, CompletionTokens: 248, TotalTokens: 254, CachedTokens: 78105, CacheCreationTokens: 24309},
wantUsage: Usage{PromptTokens: 102420, CompletionTokens: 248, TotalTokens: 102668, CachedTokens: 78105, CacheCreationTokens: 24309},
wantSession: "9c22c15c-80ef-42c6-a503-8d087ec47d92",
},
{
@@ -32,7 +32,7 @@ func TestClaudeCodeHarnessOutputFixtures(t *testing.T) {
{Tool: "Bash", Detail: "hello"},
{Tool: "Bash", Detail: errorDetail, Failed: true},
},
wantUsage: Usage{PromptTokens: 6, CompletionTokens: 121, TotalTokens: 127, CachedTokens: 78105, CacheCreationTokens: 24309},
wantUsage: Usage{PromptTokens: 102420, CompletionTokens: 121, TotalTokens: 102541, CachedTokens: 78105, CacheCreationTokens: 24309},
wantSession: "9c22c15c-80ef-42c6-a503-8d087ec47d92",
},
{
@@ -41,7 +41,7 @@ func TestClaudeCodeHarnessOutputFixtures(t *testing.T) {
{Tool: "Bash", Detail: "hello"},
{Tool: "Bash", Detail: errorDetail, Failed: true},
},
wantUsage: Usage{PromptTokens: 6, CompletionTokens: 248, TotalTokens: 254, CachedTokens: 78105, CacheCreationTokens: 24309},
wantUsage: Usage{PromptTokens: 102420, CompletionTokens: 248, TotalTokens: 102668, CachedTokens: 78105, CacheCreationTokens: 24309},
wantSession: "9c22c15c-80ef-42c6-a503-8d087ec47d92",
},
}
@@ -120,7 +120,7 @@ func TestClaudeCodeHarnessOutputToolResultContent(t *testing.T) {
if !reflect.DeepEqual(calls, wantCalls) {
t.Errorf("tool calls = %#v, want %#v", calls, wantCalls)
}
wantUsage := Usage{PromptTokens: 2, CompletionTokens: 3, TotalTokens: 5, CachedTokens: 5, CacheCreationTokens: 7}
wantUsage := Usage{PromptTokens: 14, CompletionTokens: 3, TotalTokens: 17, CachedTokens: 5, CacheCreationTokens: 7}
if got := output.Usage(); got != wantUsage {
t.Errorf("usage = %#v, want %#v", got, wantUsage)
}