fix(ai): address review feedback on default effort and simplify lowering

This commit is contained in:
Aiden Cline
2026-10-31 17:37:41 -05:00
Unverified
parent e96fde7476
commit 8160b0f888
14 changed files with 199 additions and 101 deletions
+22 -26
View File
@@ -1,43 +1,39 @@
// Changing the top-level reasoning effort invalidates the provider prompt cache. Protocols with a native
// per-message update freeze the top-level effort to the first recorded assistant reply and emit markers
// at the start of each later segment whose effort changed.
import type { LLMRequest } from "./schema/messages.js"
import type { LLMRequest, Message } from "./schema/messages.js"
export const resolveEffortUpdates = (
request: LLMRequest,
requested: string | undefined,
defaultEffort: string,
): { readonly effort: string | undefined; readonly markers: ReadonlyMap<number, string> } => {
validEffort?: (effort: string) => boolean,
) => {
const markers = new Map<number, string>()
if (!(request.model.route.supportsEffortUpdates?.(request) ?? false)) return { effort: requested, markers }
const first = request.messages.find((message) => message.role === "assistant" && message.effort !== undefined)
if (first === undefined) return { effort: requested, markers }
const baseline = first.effort ?? null
const desired = requested ?? null
const messages = request.messages
const effortOf = (message: Message) =>
typeof message.effort === "string" && validEffort && !validEffort(message.effort) ? undefined : message.effort
const first = messages.find((message) => message.role === "assistant" && effortOf(message) !== undefined)
const baseline = first ? effortOf(first) : undefined
if (baseline === undefined || !request.model.route.supportsEffortUpdates?.(request))
return { effort: requested, markers }
const desired = requested ?? null
let active = baseline
let segmentStart = 0
let prefilled = false
let start = 0
for (let i = 0; i < messages.length; i++) {
const message = messages[i]!
if (message.role !== "assistant") continue
let nextStart = i + 1
while (nextStart < messages.length && messages[nextStart]!.role === "tool") nextStart++
const isTrailingPrefill = message.effort === undefined && nextStart === messages.length
const target = isTrailingPrefill ? desired : message.effort
if (target !== undefined && segmentStart > 0 && target !== active) {
markers.set(segmentStart, target ?? defaultEffort)
if (messages[i]!.role !== "assistant") continue
let next = i + 1
while (messages[next]?.role === "tool") next++
// A trailing unstamped assistant is a prefill for the reply being requested.
const recorded = effortOf(messages[i]!)
const target = recorded !== undefined ? recorded : next === messages.length ? desired : undefined
if (target !== undefined && target !== active) {
markers.set(start, target ?? defaultEffort)
active = target
}
if (isTrailingPrefill) prefilled = true
segmentStart = nextStart
i = nextStart - 1
start = next
i = next - 1
}
if (!prefilled && segmentStart > 0 && desired !== active) {
markers.set(segmentStart, desired ?? defaultEffort)
}
if (desired !== active) markers.set(start, desired ?? defaultEffort)
return { effort: baseline ?? undefined, markers }
}
@@ -39,6 +39,8 @@ export const DEFAULT_BASE_URL = "https://api.anthropic.com/v1"
export const PATH = "/messages"
export const DEFAULT_MAX_TOKENS = 32_000
const MIN_THINKING_BUDGET = 1_024
const ANTHROPIC_EFFORTS = new Set(["low", "medium", "high", "xhigh", "max"])
const isAnthropicEffort = (effort: string) => ANTHROPIC_EFFORTS.has(effort)
// Claude Opus 5.5 and Haiku 5.5 default to `medium`; other Claude models default to `high`.
const defaultEffort = (id: string) => {
@@ -929,7 +931,7 @@ const lowerMessages = Effect.fnUntraced(function* (
const effort = effortMarkers.get(index)
if (effort === undefined) return
const last = messages.at(-1)
if (last?.role === "system" && last.content.length === 0 && last.output_config !== undefined) messages.pop()
if (last?.role === "system" && last.output_config !== undefined) messages.pop()
// Accepted at any position, so the text-update placement rules do not apply.
messages.push({ role: "system", content: [], output_config: { effort } })
}
@@ -1081,6 +1083,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
request,
options.effort ?? outputConfig?.effort ?? undefined,
defaultEffort(request.model.id),
isAnthropicEffort,
)
const generation = request.generation
// Allocate the 4-breakpoint budget in invalidation order: tools → system →
@@ -1425,7 +1428,10 @@ const onContentBlockDelta = Effect.fnUntraced(function* (
return [state, NO_EVENTS] satisfies StepResult
})
const onContentBlockStop = Effect.fnUntraced(function* (state: ParserState, event: AnthropicEvent) {
const onContentBlockStop = Effect.fnUntraced(function* (
state: ParserState,
event: AnthropicEvent,
) {
if (event.index === undefined) return [state, NO_EVENTS] satisfies StepResult
if (event.index in state.compactions) {
const { [event.index]: content, ...compactions } = state.compactions
+8 -4
View File
@@ -638,8 +638,6 @@ const lowerToolResultOutput = Effect.fnUntraced(function* (
return yield* Effect.forEach(content, (item) => lowerToolResultContentItem(item, request, adapter))
})
export const DEFAULT_EFFORT = "medium"
const lowerMessages = Effect.fnUntraced(function* (
request: LLMRequest,
adapter: ProviderAdapter,
@@ -1208,7 +1206,10 @@ const onReasoningSummaryPartDone = (state: ParserState, event: Event): StepResul
]
}
const onFunctionCallArgumentsDelta = Effect.fnUntraced(function* (state: ParserState, event: Event) {
const onFunctionCallArgumentsDelta = Effect.fnUntraced(function* (
state: ParserState,
event: Event,
) {
if (event.item_id === undefined) return [state, NO_EVENTS] satisfies StepResult
const tool = state.tools[event.item_id]
if (!tool) return [state, NO_EVENTS] satisfies StepResult
@@ -1236,7 +1237,10 @@ const onFunctionCallArgumentsDelta = Effect.fnUntraced(function* (state: ParserS
return [{ ...state, lifecycle, tools: result.tools }, events] satisfies StepResult
})
const onOutputItemDone = Effect.fnUntraced(function* (state: ParserState, item: NormalizedEvent["item"]) {
const onOutputItemDone = Effect.fnUntraced(function* (
state: ParserState,
item: NormalizedEvent["item"],
) {
if (!item) return [state, NO_EVENTS] satisfies StepResult
if (item.type === "compaction") {
+7 -12
View File
@@ -164,10 +164,11 @@ const adapter = {
// GPT-6 and later default to `configuration_update` support, except in `reasoning.mode: "pro"`
// or alongside automatic `context_management` compaction or `truncation: "auto"`.
const supportsEffortUpdates = (request: LLMRequest) => {
if (request.providerOptions?.contextManagement !== undefined) return false
if (request.providerOptions?.truncation === "auto") return false
const body = request.http?.body
if (request.providerOptions?.contextManagement !== undefined || body?.context_management !== undefined) return false
if (request.providerOptions?.truncation === "auto" || body?.truncation === "auto") return false
if (request.providerOptions?.reasoningMode === "pro") return false
if (Schema.is(Schema.Struct({ mode: Schema.Literal("pro") }))(request.http?.body?.reasoning)) return false
if (Schema.is(Schema.Struct({ mode: Schema.Literal("pro") }))(body?.reasoning)) return false
const override = request.model.compatibility?.supportsEffortUpdates
if (override !== undefined) return override
const match = /(?:^|\/)gpt-(\d+)(?:\.\d+)?(?:-|$)/i.exec(request.model.id)
@@ -184,10 +185,7 @@ const nativeImageTool = (tool: ToolDefinition) => {
return Schema.is(OpenAIResponsesImageGenerationTool)(native) ? native : undefined
}
const lowerTool = Effect.fnUntraced(function* (
tool: ToolDefinition,
custom: ReadonlyMap<string, OpenAIFreeform.Custom>,
) {
const lowerTool = Effect.fnUntraced(function* (tool: ToolDefinition, custom: ReadonlyMap<string, OpenAIFreeform.Custom>) {
const native = nativeImageToolInput(tool)
if (native !== undefined) {
if (Schema.is(OpenAIResponsesImageGenerationTool)(native)) return native
@@ -200,10 +198,7 @@ const lowerTool = Effect.fnUntraced(function* (
// Native namespaces hold only function tools, so deeper levels flatten into
// the leaf names the same way non-native protocols flatten the whole tree.
const lowerToolEntry = Effect.fnUntraced(function* (
tool: ToolEntry,
custom: ReadonlyMap<string, OpenAIFreeform.Custom>,
) {
const lowerToolEntry = Effect.fnUntraced(function* (tool: ToolEntry, custom: ReadonlyMap<string, OpenAIFreeform.Custom>) {
if (tool.type === "tool") return yield* lowerTool(tool, custom)
// OpenAI requires a namespace description; fall back to a generic one so a
// missing description never blocks the request.
@@ -243,7 +238,7 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
)(request.providerOptions?.contextManagement)
const options = OpenResponsesOptions.resolve(request)
const reasoning = Schema.decodeUnknownSync(ReasoningOptions)(request.providerOptions ?? {})
const updates = resolveEffortUpdates(request, options.reasoningEffort, OpenResponses.DEFAULT_EFFORT)
const updates = resolveEffortUpdates(request, options.reasoningEffort, "medium")
const generation = OpenResponses.lowerGeneration(request, { ...options, reasoningEffort: updates.effort })
const custom = OpenAIFreeform.tools(request)
return {
+1 -2
View File
@@ -260,8 +260,7 @@ export namespace Message {
export const user = (content: ContentInput) => make({ role: "user", content })
export const assistant = (content: ContentInput, options?: { readonly effort?: ReasoningEffort | null }) =>
make({ role: "assistant", content, ...options })
export const assistant = (content: ContentInput) => make({ role: "assistant", content })
/**
* Add an operator-authored instruction at this chronological point in the
+57 -24
View File
@@ -23,8 +23,10 @@ const openai = (id: string, compatibility?: { readonly supportsEffortUpdates?: b
const opus5 = anthropic("claude-opus-5")
const astra = openai("gpt-6-astra")
const reply = (content: Message.ContentInput, effort?: string | null) =>
Message.make({ role: "assistant", content, effort })
const conversation = [Message.user("Before."), Message.assistant("First.", { effort: "high" }), Message.user("After.")]
const conversation = [Message.user("Before."), reply("First.", "high"), Message.user("After.")]
const systemMessages = (body: AnthropicMessages.AnthropicMessagesBody) =>
body.messages.filter((message) => message.role === "system")
@@ -59,7 +61,7 @@ describe("cache policy", () => {
model: opus5,
messages: [
Message.user("first"),
Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })], { effort: "high" }),
reply([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })], "high"),
Message.tool({ id: "call_1", name: "lookup", result: "ok" }),
],
providerOptions: { effort: "low" },
@@ -112,9 +114,9 @@ describe("Anthropic Messages effort updates", () => {
model: opus5,
messages: [
Message.user("One."),
Message.assistant("First.", { effort: "high" }),
reply("First.", "high"),
Message.user("Two."),
Message.assistant("Second.", { effort: "low" }),
reply("Second.", "low"),
Message.user("Forked follow-up."),
],
providerOptions: { effort: "max" },
@@ -149,9 +151,9 @@ describe("Anthropic Messages effort updates", () => {
model: anthropic(id),
messages: [
Message.user("One."),
Message.assistant("First.", { effort: null }),
reply("First.", null),
Message.user("Two."),
Message.assistant("Second.", { effort: "low" }),
reply("Second.", "low"),
Message.user("Three."),
],
providerOptions: { output_config: { format } },
@@ -175,7 +177,7 @@ describe("Anthropic Messages effort updates", () => {
model: opus5,
messages: [
Message.user("Before."),
Message.assistant("First.", { effort: "high" }),
reply("First.", "high"),
Message.user("After."),
Message.assistant("Summary:"),
],
@@ -223,7 +225,7 @@ describe("Anthropic Messages effort updates", () => {
model: opus5,
messages: [
Message.user("Weather?"),
Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })], { effort: "high" }),
reply([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })], "high"),
Message.tool({ id: "call_1", name: "lookup", result: { temp: 72 } }),
],
providerOptions: { effort: "low" },
@@ -245,12 +247,7 @@ describe("Anthropic Messages effort updates", () => {
const prepared = yield* compileRequest(
ai.request({
model: opus5,
messages: [
Message.user("Fix it."),
Message.assistant("Done.", { effort: "high" }),
Message.system("Update."),
Message.user("Next."),
],
messages: [Message.user("Fix it."), reply("Done.", "high"), Message.system("Update."), Message.user("Next.")],
providerOptions: { effort: "low" },
cache: "none",
}),
@@ -333,6 +330,28 @@ describe("Anthropic Messages effort updates", () => {
}),
)
it.effect("ignores non-Anthropic `none` effort from an earlier disabled-thinking turn", () =>
Effect.gen(function* () {
const prepared = yield* compileRequest(
ai.request({
model: opus5,
messages: [
Message.user("One."),
reply("First.", "high"),
Message.user("Two."),
reply("No thinking.", "none"),
Message.user("Three."),
],
providerOptions: { effort: "low" },
cache: "none",
}),
)
expect(prepared.body.output_config).toEqual({ effort: "high" })
expect(systemMessages(prepared.body)).toEqual([{ role: "system", content: [], output_config: { effort: "low" } }])
}),
)
it.effect("omits markers for Opus 5.0 on Bedrock Mantle Messages while lowering Opus 5.5", () =>
Effect.gen(function* () {
const mantle = AmazonBedrockMantle.configure({ apiKey: "test", region: "us-east-1" })
@@ -406,12 +425,7 @@ describe("OpenAI Responses effort updates", () => {
const prepared = yield* compileRequest(
ai.request({
model: astra,
messages: [
Message.user("Before."),
Message.assistant("First.", { effort: "medium" }),
Message.assistant([], { effort: "low" }),
Message.user("After."),
],
messages: [Message.user("Before."), reply("First.", "medium"), reply([], "low"), Message.user("After.")],
providerOptions: { reasoningEffort: "xhigh" },
}),
)
@@ -433,9 +447,9 @@ describe("OpenAI Responses effort updates", () => {
model: astra,
messages: [
Message.user("One."),
Message.assistant("First.", { effort: "high" }),
reply("First.", "high"),
Message.user("Two."),
Message.assistant("Second.", { effort: "low" }),
reply("Second.", "low"),
Message.user("Forked follow-up."),
],
providerOptions: { reasoningEffort: "xhigh" },
@@ -457,9 +471,9 @@ describe("OpenAI Responses effort updates", () => {
model: astra,
messages: [
Message.user("One."),
Message.assistant("First.", { effort: null }),
reply("First.", null),
Message.user("Two."),
Message.assistant("Second.", { effort: "low" }),
reply("Second.", "low"),
Message.user("Three."),
],
}),
@@ -490,12 +504,31 @@ describe("OpenAI Responses effort updates", () => {
}),
)
const withBodyTruncation = yield* compileRequest(
ai.request({
model: astra,
messages: conversation,
providerOptions: { reasoningEffort: "low" },
http: { body: { truncation: "auto" } },
}),
)
const withBodyCompaction = yield* compileRequest(
ai.request({
model: astra,
messages: conversation,
providerOptions: { reasoningEffort: "low" },
http: { body: { context_management: [{ type: "compaction" }] } },
}),
)
expect(updates(withCompaction.body)).toEqual([])
expect(withCompaction.body.reasoning).toEqual({ effort: "low" })
expect(withCompaction.body.context_management).toEqual([{ type: "compaction" }])
expect(updates(withTruncation.body)).toEqual([])
expect(withTruncation.body.reasoning).toEqual({ effort: "low" })
expect(withTruncation.body.truncation).toBe("auto")
expect(updates(withBodyTruncation.body)).toEqual([])
expect(updates(withBodyCompaction.body)).toEqual([])
}),
)
@@ -418,7 +418,7 @@ describe("Amazon Bedrock Mantle Messages recorded", () => {
model: mantleMessages("anthropic.claude-opus-5"),
messages: [
Message.user("What is 12 + 30?"),
Message.assistant("42", { effort: "high" }),
Message.make({ role: "assistant", content: "42", effort: "high" }),
Message.user("Add 8 to that result. Reply with only the integer."),
],
providerOptions: {
+24 -4
View File
@@ -13,7 +13,7 @@ import {
type ProviderPackage,
} from "@opencode/ai/effect"
import { Auth, mergeHttpOptions } from "@opencode/ai/route"
import { Context, Effect, Layer, Schema, Struct } from "effect"
import { Context, Effect, Layer, Option, Schema, Struct } from "effect"
import { AISDK } from "./aisdk.js"
import { Credential } from "./credential.js"
import { Integration } from "./integration.js"
@@ -136,6 +136,23 @@ export interface Resolved {
readonly transport?: Provider.Transport
/** Milliseconds without streamed data before a WebSocket exchange fails; `false` disables the limit. */
readonly chunkTimeout?: number | false
/** Effort configured on the provider or model before any variant overlay. */
readonly defaultEffort?: string
}
const ConfiguredEffort = Schema.Struct({
effort: Schema.optional(Schema.String),
reasoningEffort: Schema.optional(Schema.String),
output_config: Schema.optional(Schema.Struct({ effort: Schema.optional(Schema.String) })),
outputConfig: Schema.optional(Schema.Struct({ effort: Schema.optional(Schema.String) })),
})
const decodeConfiguredEffort = Schema.decodeUnknownOption(ConfiguredEffort)
export const configuredEffort = (options: Readonly<Record<string, unknown>> | undefined) => {
if (!options) return undefined
const parsed = Option.getOrUndefined(decodeConfiguredEffort(options))
return parsed?.effort ?? parsed?.reasoningEffort ?? parsed?.output_config?.effort ?? parsed?.outputConfig?.effort
}
export type MediaError =
@@ -365,7 +382,8 @@ function prepareProviderURL(
const prepared = baseURL.replace(
/\$\{([^}]+)\}/g,
(placeholder, name: string) =>
(name === "AWS_REGION" && typeof settings.region === "string" ? settings.region : process.env[name]) ?? placeholder,
(name === "AWS_REGION" && typeof settings.region === "string" ? settings.region : process.env[name]) ??
placeholder,
)
const failure = unresolvedProviderVariables(model, prepared)
return failure ? Effect.fail(failure) : Effect.succeed(prepared)
@@ -458,14 +476,15 @@ export const layer = Layer.effect(
)
const credential = connection ? yield* integrations.connection.resolve(connection) : undefined
const selectedVariant = yield* withVariant(selected, variant)
const baseSettings = Provider.mergeOverlay(provider?.settings, Provider.modelSettings(selected.settings))
const runtimeInfo: RuntimeInfo = {
...selectedVariant,
settings: Provider.mergeOverlay(provider?.settings, Provider.modelSettings(selectedVariant.settings)),
}
return { provider, credential, runtimeInfo }
return { provider, credential, runtimeInfo, defaultEffort: configuredEffort(baseSettings) }
})
const load = Effect.fn("ModelResolver.resolveModel")(function* (selected: Info, variant?: VariantID) {
const { provider, credential, runtimeInfo } = yield* prepare(selected, variant)
const { provider, credential, runtimeInfo, defaultEffort } = yield* prepare(selected, variant)
const model = yield* fromCatalogModel(runtimeInfo, credential, dependencies)
const runtime =
provider?.activation === "enabled" &&
@@ -488,6 +507,7 @@ export const layer = Layer.effect(
cache: runtimeInfo.settings?.cache,
transport: provider?.settings?.transport,
chunkTimeout: Provider.timeout(provider?.settings?.chunkTimeout),
defaultEffort,
}
})
const loadMedia = Effect.fnUntraced(function* <Runtime>(
+8 -2
View File
@@ -29,6 +29,7 @@ import { HttpClientRequest, HttpClientResponse } from "effect/http"
import { makeLocationNode } from "@opencode/util/effect/app-node"
import { App } from "../app.js"
import { Instructions } from "../instructions/index.js"
import { ModelResolver } from "../model-resolver.js"
import { Permission } from "../permission.js"
import { PluginHooks } from "../plugin/hooks.js"
import { QuestionTool } from "../tool/plugin/question.js"
@@ -119,7 +120,7 @@ export const baseTranscript = (input: {
metadata: { experimental: { instructions: input.initial } },
},
].filter((part) => part.text.length > 0),
messages: toLLMMessages(input.messages, input.model.ref, providerMetadataKey),
messages: toLLMMessages(input.messages, input.model.ref, providerMetadataKey, input.model.defaultEffort),
}
}
@@ -287,6 +288,11 @@ export const layer = Layer.effect(
const entries = Object.entries(shaped.options)
const generation = Object.fromEntries(entries.filter(([k]) => GENERATION_KEYS.has(k))) as GenerationOptionsFields
const providerOptions = Object.fromEntries(entries.filter(([k]) => !GENERATION_KEYS.has(k)))
// Hook-supplied effort options are not recorded on assistant rows, so fall back to top-level effort.
const messages =
ModelResolver.configuredEffort(providerOptions) === undefined
? shaped.messages
: shaped.messages.map((m) => (m.effort === undefined ? m : new Message({ ...m, effort: undefined })))
const affinity = SessionAffinity.get(session)
const base = ai.request({
model: model.model,
@@ -307,7 +313,7 @@ export const layer = Layer.effect(
promptCacheKey: /^ses_[0-9a-f]{64}$/.test(affinity) ? affinity.slice(4) : affinity,
cache: model.cache?.ttl === undefined ? undefined : { ttl: model.cache.ttl },
system: shaped.system,
messages: boundImages(unsupportedParts(shaped.messages, model.capabilities, model.model.provider)),
messages: boundImages(unsupportedParts(messages, model.capabilities, model.model.provider)),
tools: Array.from(hooked, ([name, t]) => ({ ...t, name })),
toolChoice: input.toolChoice,
generation: Object.keys(generation).length === 0 ? undefined : generation,
@@ -66,6 +66,10 @@ export const resolved = (
limit: options.limit,
compaction: options.compaction,
transport: options.transport,
defaultEffort:
options.variant === undefined || options.variant === "default"
? ModelResolver.configuredEffort(model.defaults?.providerOptions)
: undefined,
})
const layer = Layer.effect(
@@ -152,7 +152,12 @@ const toolResult = (tool: SessionMessage.AssistantTool, providerMetadata: Provid
}
}
const assistant = (message: SessionMessage.Assistant, model: Model.Ref, providerMetadataKey: string) => {
const assistant = (
message: SessionMessage.Assistant,
model: Model.Ref,
providerMetadataKey: string,
defaultEffort?: string,
) => {
const sameProvider = message.model.providerID === model.providerID
const sameModel = sameProvider && message.model.id === model.id
const reuseProviderMetadata = sameModel && message.error === undefined
@@ -216,7 +221,9 @@ const assistant = (message: SessionMessage.Assistant, model: Model.Ref, provider
.map(Message.tool)
if (meaningful.length === 0) return results
const effort =
sameModel && variantEffort(model.variant) !== undefined ? variantEffort(message.model.variant) : undefined
sameModel && variantEffort(model.variant, defaultEffort) !== undefined
? variantEffort(message.model.variant, defaultEffort)
: undefined
return [
Message.make({ id: message.id, role: "assistant", content: meaningful, effort, metadata: message.metadata }),
...results,
@@ -225,12 +232,17 @@ const assistant = (message: SessionMessage.Assistant, model: Model.Ref, provider
const EFFORT_VARIANTS = new Set<string>(ReasoningEfforts)
const variantEffort = (variant: Model.VariantID | undefined) => {
if (variant === undefined || variant === "default") return null
const variantEffort = (variant: Model.VariantID | undefined, defaultEffort?: string) => {
if (variant === undefined || variant === "default") return defaultEffort ?? null
return EFFORT_VARIANTS.has(variant) ? variant : undefined
}
function toLLMMessage(message: SessionMessage.Info, model: Model.Ref, providerMetadataKey: string): Message[] {
function toLLMMessage(
message: SessionMessage.Info,
model: Model.Ref,
providerMetadataKey: string,
defaultEffort?: string,
): Message[] {
switch (message.type) {
case "agent-switched":
case "idle":
@@ -281,7 +293,7 @@ function toLLMMessage(message: SessionMessage.Info, model: Model.Ref, providerMe
}),
]
case "assistant":
return assistant(message, model, providerMetadataKey)
return assistant(message, model, providerMetadataKey, defaultEffort)
case "compaction": {
if (message.status !== "completed") return []
// History selection only keeps native windows the target model can replay.
@@ -311,4 +323,5 @@ export const toLLMMessages = (
messages: readonly SessionMessage.Info[],
model: Model.Ref,
providerMetadataKey: string = model.providerID,
) => messages.flatMap((message) => toLLMMessage(message, model, providerMetadataKey))
defaultEffort?: string,
) => messages.flatMap((message) => toLLMMessage(message, model, providerMetadataKey, defaultEffort))
@@ -1109,8 +1109,16 @@ describe("ChatGPTPlugin", () => {
Effect.flatMap((resolved) => compileRequest(ai.request({ model: resolved.model, messages }))),
Effect.map((compiled) => compiled.body),
)
const switched = [Message.user("a"), Message.assistant("b", { effort: "high" }), Message.user("c")]
const switchedBack = [...switched, Message.assistant("d", { effort: "max" }), Message.user("e")]
const switched = [
Message.user("a"),
Message.make({ role: "assistant", content: "b", effort: "high" }),
Message.user("c"),
]
const switchedBack = [
...switched,
Message.make({ role: "assistant", content: "d", effort: "max" }),
Message.user("e"),
]
const user = (text: string) => ({ role: "user", content: [{ type: "input_text", text }] })
const assistant = (text: string) => ({ role: "assistant", content: [{ type: "output_text", text }] })
const update = (effort: string) => ({ type: "configuration_update", reasoning: { effort } })
@@ -254,9 +254,11 @@ Earlier work
])
})
test("maps the default variant and no variant to null", () => {
test("maps the default variant and no variant to the configured default effort or null", () => {
expect(toLLMMessages([reply(ref())], ref("low"))[0]?.effort).toBeNull()
expect(toLLMMessages([reply(ref("default"))], ref())[0]?.effort).toBeNull()
expect(toLLMMessages([reply(ref())], ref("low"), "provider", "high")[0]?.effort).toBe("high")
expect(toLLMMessages([reply(ref("default"))], ref(), "provider", "medium")[0]?.effort).toBe("medium")
})
test("leaves effort unset across model switches, non-effort variants, and model-switched markers", () => {
+26 -14
View File
@@ -2319,7 +2319,7 @@ describe("SessionRunnerLLM", () => {
})
s.currentModel = LanguageModel.update(s.currentModel, { defaults: { providerOptions: { effort: "low" } } })
yield* s.llm.push(TestLLM.text("Later answer", "text-effort-low"))
yield* s.runPrompt("Second")
const second = yield* s.runPrompt("Second")
expect(messageRoles(s.requests[1])).toEqual(["user", "assistant", "user"])
expect(s.requests[1]?.messages[1]?.effort).toBe("high")
@@ -2335,18 +2335,17 @@ describe("SessionRunnerLLM", () => {
],
})
// Switch to `max`, run a third turn, then fork before that third turn:
// `forked` inherits `session.model` (`max`) while its copied history ends at `low`.
// Switch to `max`, run Turn 3, then fork and revert to `second`:
// both `low` and `max` switch rows are removed while `session.model` remains `max`.
yield* s.bus.publish(SessionEvent.ModelSelected, {
sessionID,
model: { ...model, variant: Model.VariantID.make("max") },
})
s.currentModel = LanguageModel.update(s.currentModel, { defaults: { providerOptions: { effort: "max" } } })
const third = yield* s.session.prompt({ sessionID, text: "Third", resume: false })
yield* s.llm.push(TestLLM.text("Third answer", "text-effort-max-1"))
yield* s.resume
yield* s.runPrompt("Third")
const forked = yield* s.session.fork({ sessionID, before: third.id })
const forked = yield* s.session.fork({ sessionID, before: second.id })
yield* s.llm.push(TestLLM.text("Forked answer", "text-effort-max-2"))
yield* s.session.prompt({ sessionID: forked.id, text: "Forked", resume: false })
yield* s.session.resume(forked.id)
@@ -2355,9 +2354,6 @@ describe("SessionRunnerLLM", () => {
expect(forkedCompiled.body).toMatchObject({
output_config: { effort: "high" },
messages: [
{ role: "user" },
{ role: "assistant" },
{ role: "system", content: [], output_config: { effort: "low" } },
{ role: "user" },
{ role: "assistant" },
{ role: "system", content: [], output_config: { effort: "max" } },
@@ -2365,8 +2361,8 @@ describe("SessionRunnerLLM", () => {
],
})
// Committed revert on the original session keeps the same frozen baseline and earlier marker.
yield* s.bus.publish(SessionEvent.RevertEvent.Committed, { sessionID, to: third.id })
// Committed revert to `second` keeps the frozen `high` baseline as well.
yield* s.bus.publish(SessionEvent.RevertEvent.Committed, { sessionID, to: second.id })
yield* s.llm.push(TestLLM.text("Reverted answer", "text-effort-max-3"))
yield* s.runPrompt("After revert")
@@ -2374,9 +2370,6 @@ describe("SessionRunnerLLM", () => {
expect(revertedCompiled.body).toMatchObject({
output_config: { effort: "high" },
messages: [
{ role: "user" },
{ role: "assistant" },
{ role: "system", content: [], output_config: { effort: "low" } },
{ role: "user" },
{ role: "assistant" },
{ role: "system", content: [], output_config: { effort: "max" } },
@@ -2385,6 +2378,25 @@ describe("SessionRunnerLLM", () => {
})
})
scenario("keeps configured default effort stable when no variant is selected", function* (s) {
s.currentModel = LanguageModel.make({
id: "claude-opus-5-5",
provider: "anthropic",
route: AnthropicMessages.route,
defaults: { providerOptions: { effort: "high" } },
})
yield* s.llm.push(TestLLM.text("First answer", "text-default-effort-1"))
yield* s.runPrompt("First")
yield* s.llm.push(TestLLM.text("Second answer", "text-default-effort-2"))
yield* s.runPrompt("Second")
const compiled = yield* compileRequest(s.requests[1]!)
expect(compiled.body).toMatchObject({
output_config: { effort: "high" },
messages: [{ role: "user" }, { role: "assistant" }, { role: "user" }],
})
})
scenario("preserves instruction values while a source is temporarily unavailable", function* (s) {
yield* s.runPrompt("First")
yield* s.bus.publish(SessionEvent.ModelSelected, {