feat(core): configure session warming per provider and model (#60025)

This commit is contained in:
Aiden Cline
2026-10-31 21:05:18 -05:00
committed by GitHub
Unverified
parent 9a611d36b9
commit 296074aab2
4 changed files with 68 additions and 11 deletions
+3 -1
View File
@@ -25,6 +25,8 @@ export type ProviderCompaction = { type: "summary" } | { type: "native" }
export type ProviderCacheTTL = `${number}${"ms" | "s" | "m" | "h" | "d"}`
export type ProviderCacheWarming = boolean | { interval?: "auto" | (string & {}); duration?: string }
export type ProviderTransport = "http" | "websocket"
export type ProviderEncoding = "zstd" | "gzip" | "none"
@@ -393,7 +395,7 @@ export type LocationPublicInfo = {
export type ModelRef = { id: string; providerID: ProviderID; variant?: string }
export type ProviderCache = { ttl?: ProviderCacheTTL }
export type ProviderCache = { ttl?: ProviderCacheTTL; warming?: ProviderCacheWarming }
export type ModelVariantSettings = { compaction?: ProviderCompaction; encoding?: ProviderEncoding } & {
[x: string]: any
+21 -10
View File
@@ -4,10 +4,13 @@ import { define } from "@opencode/plugin/effect/plugin"
import type { SessionHooks } from "@opencode/plugin/effect/session"
import type { Provider } from "@opencode/schema/provider"
import type { Session } from "@opencode/schema/session"
import { Clock, Duration, Effect, Scope, Stream } from "effect"
import { ConfigWarming } from "@opencode/schema/config/warming"
import { Clock, Duration, Effect, Option, Schema, Scope, Stream } from "effect"
import { Config } from "../config.js"
import { resolveModel } from "./optimize.js"
const decodeWarming = Schema.decodeUnknownOption(ConfigWarming.Warming)
const defaults = {
prompt: "This is a keep-alive request. Do not perform any work or use tools. Reply with exactly: OK",
interval: Duration.minutes(4),
@@ -18,16 +21,24 @@ export const Plugin = define({
id: "opencode.warming",
effect: Effect.fn(function* (ctx) {
const config = yield* Config.Service
const modelTTL = (event: SessionHooks["context"]) =>
resolveModel(ctx, event).pipe(
Effect.map((model) => model.settings?.cache?.ttl),
Effect.orElseSucceed(() => undefined),
)
// A model's or provider's `cache.warming` overrides the top-level `warming`.
const loadSettings = Effect.fn("WarmingPlugin.loadSettings")(function* (event: SessionHooks["context"]) {
const warming = Config.latest(yield* config.entries(), "warming")
if (!warming) return
const settings = warming === true ? defaults : { ...defaults, ...warming }
const interval = settings.interval === "auto" ? autoInterval(yield* modelTTL(event)) : settings.interval
// Warming must never fail the request it hooks, so an unreadable catalog falls back to the top-level setting.
const cache = (yield* resolveModel(ctx, event).pipe(Effect.orElseSucceed(() => undefined)))?.settings?.cache
const local = cache?.warming === undefined ? undefined : Option.getOrUndefined(decodeWarming(cache.warming))
if (cache?.warming !== undefined && local === undefined)
yield* Effect.logWarning("invalid cache.warming setting; using the top-level warming setting", {
model: event.model,
})
const global = Config.latest(yield* config.entries(), "warming")
if (!(local ?? global)) return
// The keep-alive prompt is only set at the top level; a model's setting changes timing.
const settings = {
...defaults,
...(typeof global === "object" ? global : {}),
...(typeof local === "object" ? local : {}),
}
const interval = settings.interval === "auto" ? autoInterval(cache?.ttl) : settings.interval
// "auto" doesn't warm a model without a cache.ttl.
if (interval === undefined) return
const { duration } = settings
+30
View File
@@ -21507,6 +21507,9 @@
"properties": {
"ttl": {
"$ref": "#/components/schemas/Provider.CacheTTL"
},
"warming": {
"$ref": "#/components/schemas/Provider.CacheWarming"
}
},
"additionalProperties": false
@@ -21515,6 +21518,33 @@
"type": "string",
"pattern": "^[+-]?(?:\\d+(?:\\.\\d+)?|\\.\\d+)(?:[Ee][+-]?\\d+)?(?:ms|s|m|h|d)$"
},
"Provider.CacheWarming": {
"anyOf": [
{
"type": "boolean"
},
{
"type": "object",
"properties": {
"interval": {
"anyOf": [
{
"type": "string",
"enum": ["auto"]
},
{
"type": "string"
}
]
},
"duration": {
"type": "string"
}
},
"additionalProperties": false
}
]
},
"Provider.Compaction": {
"anyOf": [
{
+14
View File
@@ -53,9 +53,23 @@ export const CacheTTL = Schema.TemplateLiteral([Schema.Number, Schema.Literals([
.annotate({ identifier: "Provider.CacheTTL" })
export type CacheTTL = typeof CacheTTL.Type
/** Keep-alive requests for this provider's or model's sessions; overrides the top-level `warming`. */
export const CacheWarming = Schema.Union([
Schema.Boolean,
Schema.Struct({
// Settings hold JSON, so durations stay strings here; the warming plugin parses them.
interval: Schema.Union([Schema.Literal("auto"), Schema.String])
.pipe(optional)
.annotate({ description: 'Duration such as "4 minutes", or "auto" for 90% of the model\'s cache.ttl' }),
duration: Schema.String.pipe(optional).annotate({ description: 'Duration such as "30 minutes"' }),
}),
]).annotate({ identifier: "Provider.CacheWarming" })
export type CacheWarming = typeof CacheWarming.Type
/** Prompt-cache settings. `ttl` is sent to the provider as given; each provider accepts its own set. */
export const Cache = Schema.Struct({
ttl: CacheTTL.pipe(optional),
warming: CacheWarming.pipe(optional),
}).annotate({ identifier: "Provider.Cache" })
export type Cache = typeof Cache.Type