mirror of
https://github.com/tiennm99/goclaw.git
synced 2026-10-12 14:15:00 +00:00
thotam
503909d3b6
SetLastPromptTokens received TotalUsage.PromptTokens — the sum of prompt tokens across every think→act→observe iteration of a run — instead of the final iteration's own usage. A tool-heavy run inflated the stored value by the iteration count, so the sessions list showed e.g. 901K "context used" against a 258K window, and maybeSummarize over-triggered compaction once the inflation exceeded the 40% overhead clamp (observed: 83 compactions on a 4-message session). Track the final iteration's usage separately (ThinkState.LastUsage) and persist its actual context size: prompt tokens + cached segments (Anthropic reports input_tokens excluding cache read/creation) + output tokens (the final reply joins history, so it occupies the next request's prompt). The run-cumulative total still feeds AccumulateTokens for billing.
56 lines
2.2 KiB
Go
56 lines
2.2 KiB
Go
package providers
|
|
|
|
// CallUsage attributes one LLM call (or a tool's internal LLM call) to its
|
|
// provider/model with the tokens it consumed and its cost. Usage is embedded
|
|
// so its token fields serialize flat (prompt_tokens, cache_read_input_tokens, …)
|
|
// alongside type/name/provider/model — this same value is both the in-memory
|
|
// accumulation record and the webhook `calls[]` element.
|
|
type CallUsage struct {
|
|
Type string `json:"type"` // "llm_call" | "tool_call"
|
|
Name string `json:"name"` // e.g. "9router/cx/gpt-5.6 #3" or "read_image"
|
|
Provider string `json:"provider"`
|
|
Model string `json:"model"`
|
|
Usage // embedded, NO json tag → promotes token fields to flat CallUsage JSON
|
|
CostUSD float64 `json:"cost_usd,omitempty"`
|
|
}
|
|
|
|
// SumCallUsage folds per-call usage into one aggregate: summed tokens/cache,
|
|
// OR-ed PromptTokensIncludeCachedSegments. Zero value for an empty slice.
|
|
func SumCallUsage(calls []CallUsage) Usage {
|
|
var total Usage
|
|
for _, c := range calls {
|
|
total.PromptTokens += c.PromptTokens
|
|
total.CompletionTokens += c.CompletionTokens
|
|
total.TotalTokens += c.TotalTokens
|
|
total.CacheReadTokens += c.CacheReadTokens
|
|
total.CacheCreationTokens += c.CacheCreationTokens
|
|
total.ThinkingTokens += c.ThinkingTokens
|
|
if c.PromptTokensIncludeCachedSegments {
|
|
total.PromptTokensIncludeCachedSegments = true
|
|
}
|
|
}
|
|
return total
|
|
}
|
|
|
|
// ContextTokens returns the full prompt-side size of a single LLM call —
|
|
// the actual context the model saw — normalizing provider cache accounting.
|
|
// Anthropic-style usage reports input_tokens EXCLUDING cached segments, so
|
|
// cache read/creation tokens must be added back; OpenAI-style usage
|
|
// (PromptTokensIncludeCachedSegments=true) already includes them.
|
|
// Only meaningful for per-call usage, not run-cumulative totals.
|
|
func (u Usage) ContextTokens() int {
|
|
if u.PromptTokensIncludeCachedSegments {
|
|
return u.PromptTokens
|
|
}
|
|
return u.PromptTokens + u.CacheReadTokens + u.CacheCreationTokens
|
|
}
|
|
|
|
// SumCallCost totals the per-call USD cost (0 when pricing was unavailable).
|
|
func SumCallCost(calls []CallUsage) float64 {
|
|
var total float64
|
|
for _, c := range calls {
|
|
total += c.CostUSD
|
|
}
|
|
return total
|
|
}
|