mirror of
https://gitea.toothfairyai.com/ToothFairyAI/tf_code.git
synced 2026-04-03 23:53:46 +00:00
fix(core): ensure compaction is more reliable, add reserve token buffer to ensure that input window has enough room to compact (#12924)
Co-authored-by: James Lal <james@littlebearlabs.io>
This commit is contained in:
@@ -5,6 +5,7 @@ import type { JSONSchema } from "zod/v4/core"
|
||||
import type { Provider } from "./provider"
|
||||
import type { ModelsDev } from "./models"
|
||||
import { iife } from "@/util/iife"
|
||||
import { Flag } from "@/flag/flag"
|
||||
|
||||
type Modality = NonNullable<ModelsDev.Model["modalities"]>["input"][number]
|
||||
|
||||
@@ -17,6 +18,8 @@ function mimeToModality(mime: string): Modality | undefined {
|
||||
}
|
||||
|
||||
export namespace ProviderTransform {
|
||||
export const OUTPUT_TOKEN_MAX = Flag.OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX || 32_000
|
||||
|
||||
// Maps npm package to the key the AI SDK expects for providerOptions
|
||||
function sdkKey(npm: string): string | undefined {
|
||||
switch (npm) {
|
||||
@@ -723,29 +726,8 @@ export namespace ProviderTransform {
|
||||
return { [key]: options }
|
||||
}
|
||||
|
||||
export function maxOutputTokens(
|
||||
npm: string,
|
||||
options: Record<string, any>,
|
||||
modelLimit: number,
|
||||
globalLimit: number,
|
||||
): number {
|
||||
const modelCap = modelLimit || globalLimit
|
||||
const standardLimit = Math.min(modelCap, globalLimit)
|
||||
|
||||
if (npm === "@ai-sdk/anthropic" || npm === "@ai-sdk/google-vertex/anthropic") {
|
||||
const thinking = options?.["thinking"]
|
||||
const budgetTokens = typeof thinking?.["budgetTokens"] === "number" ? thinking["budgetTokens"] : 0
|
||||
const enabled = thinking?.["type"] === "enabled"
|
||||
if (enabled && budgetTokens > 0) {
|
||||
// Return text tokens so that text + thinking <= model cap, preferring 32k text when possible.
|
||||
if (budgetTokens + standardLimit <= modelCap) {
|
||||
return standardLimit
|
||||
}
|
||||
return modelCap - budgetTokens
|
||||
}
|
||||
}
|
||||
|
||||
return standardLimit
|
||||
export function maxOutputTokens(model: Provider.Model): number {
|
||||
return Math.min(model.limit.output, OUTPUT_TOKEN_MAX) || OUTPUT_TOKEN_MAX
|
||||
}
|
||||
|
||||
export function schema(model: Provider.Model, schema: JSONSchema.BaseSchema | JSONSchema7): JSONSchema7 {
|
||||
|
||||
Reference in New Issue
Block a user