mirror of
https://github.com/stablyai/orca.git
synced 2026-11-19 04:23:57 +00:00
Claude reports deferred categories (MCP tools, system tools loaded on demand) that sit outside the used total, but summarizeContextUsage dropped the flag, so the popover listed them as window percentages next to real usage and the rows summed to well over 100%. Keep the deferred flag on each summary row and render deferred rows in their own group under a divider, with token counts instead of a share of the window and a note that they are not counted in the used total. Fixes #39878 Co-authored-by: Neil <neil@stably.ai>
83 lines
2.9 KiB
TypeScript
83 lines
2.9 KiB
TypeScript
import type { StructuredAgentContextUsage } from '../agent-session/structured-agent-session-context-usage'
|
|
|
|
/** Reduce provider context breakdowns to the rows shown by the composer ring. */
|
|
export type NativeChatContextUsageSummary = {
|
|
usedTokens: number
|
|
windowTokens: number
|
|
percentage: number
|
|
estimated: boolean
|
|
/** Largest first, as a share of the window; empty rows are dropped, and deferred rows sit outside the used total. */
|
|
rows: readonly { name: string; tokens: number; percentage: number; deferred?: true }[]
|
|
}
|
|
|
|
/** Past this the window is nearly spent, so the ring turns red to say so. */
|
|
const CRITICAL_PERCENTAGE = 90
|
|
|
|
/** What a ring draws: its fill clamped to 0-100, the red state, and the token counts as shown. */
|
|
export function contextUsageRingReadout(usage: NativeChatContextUsageSummary): {
|
|
filled: number
|
|
critical: boolean
|
|
used: string
|
|
window: string
|
|
} {
|
|
return {
|
|
filled: Math.min(Math.max(usage.percentage, 0), 100),
|
|
critical: usage.percentage >= CRITICAL_PERCENTAGE,
|
|
used: formatContextTokenCount(usage.usedTokens),
|
|
window: formatContextTokenCount(usage.windowTokens)
|
|
}
|
|
}
|
|
|
|
function rowPercentage(tokens: number, windowTokens: number): number {
|
|
return windowTokens > 0 ? Math.round((tokens / windowTokens) * 1000) / 10 : 0
|
|
}
|
|
|
|
export function summarizeContextUsage(
|
|
usage: StructuredAgentContextUsage
|
|
): NativeChatContextUsageSummary {
|
|
return {
|
|
usedTokens: usage.usedTokens,
|
|
windowTokens: usage.windowTokens,
|
|
percentage: usage.percentage,
|
|
estimated: usage.estimated,
|
|
rows: usage.categories
|
|
.filter((row) => row.tokens > 0)
|
|
.map((row) => ({
|
|
name: row.name,
|
|
tokens: row.tokens,
|
|
percentage: rowPercentage(row.tokens, usage.windowTokens),
|
|
...(row.deferred === true ? { deferred: true as const } : {})
|
|
}))
|
|
.sort((left, right) => right.tokens - left.tokens)
|
|
}
|
|
}
|
|
|
|
/** The CLI can repeat a category name, so a row's key counts which repeat it is. */
|
|
export function keyContextUsageRows(
|
|
rows: NativeChatContextUsageSummary['rows']
|
|
): { key: string; row: NativeChatContextUsageSummary['rows'][number] }[] {
|
|
const seen = new Map<string, number>()
|
|
return rows.map((row) => {
|
|
const repeat = seen.get(row.name) ?? 0
|
|
seen.set(row.name, repeat + 1)
|
|
return { key: `${repeat}:${row.name}`, row }
|
|
})
|
|
}
|
|
|
|
/** `18.6k`, `981.4k`, `1M`: capital M, since a lowercase one reads as minutes. */
|
|
export function formatContextTokenCount(tokens: number): string {
|
|
const safe = Math.max(0, tokens)
|
|
// Compared after rounding, so 999,960 reads `1M` rather than `1000k`.
|
|
if (Math.round(safe / 100) >= 10_000) {
|
|
return `${trimZero((safe / 1_000_000).toFixed(1))}M`
|
|
}
|
|
if (safe >= 1_000) {
|
|
return `${trimZero((safe / 1_000).toFixed(1))}k`
|
|
}
|
|
return String(Math.round(safe))
|
|
}
|
|
|
|
function trimZero(value: string): string {
|
|
return value.endsWith('.0') ? value.slice(0, -2) : value
|
|
}
|