Compare commits

...
13 changed files with 264 additions and 20 deletions
+1 -1
View File
@@ -65,7 +65,7 @@ export const Model = Schema.Struct({
name: Schema.String,
family: Schema.optional(Schema.String),
release_date: Schema.String,
attachment: Schema.Boolean,
attachment: Schema.optional(Schema.Boolean),
reasoning: Schema.Boolean,
reasoning_options: Schema.optional(Schema.Array(ReasoningOption)),
temperature: Schema.optional(Schema.Boolean),
+5 -4
View File
@@ -185,11 +185,12 @@ function applyModel(
draft.family = model.family ? ModelV2.Family.make(model.family) : undefined
draft.package = model.provider?.npm ? ProviderV2.aisdk(model.provider.npm) : undefined
draft.settings = model.provider?.api ? { ...draft.settings, baseURL: model.provider.api } : draft.settings
draft.capabilities = {
const capabilities = ModelV2.Capabilities.defaults({
tools: model.tool_call,
input: [...(model.modalities?.input ?? [])],
output: [...(model.modalities?.output ?? [])],
}
input: model.modalities?.input ?? (model.attachment === false ? ["text"] : undefined),
output: model.modalities?.output,
})
draft.capabilities = { ...capabilities, input: [...capabilities.input], output: [...capabilities.output] }
mergeVariants(draft, input.variants ?? [])
draft.time.released = released(model.release_date)
draft.cost = (input.cost ?? cost(model.cost)).map((item) => ({
+5 -4
View File
@@ -237,9 +237,10 @@ const layer = Layer.effect(
const entries = yield* SessionHistory.entriesForRunner(db, session.id, checkpoint.baselineSeq)
const context = entries.map((entry) => entry.message)
const isLastStep = agent.info?.steps !== undefined && currentStep >= agent.info.steps
const toolMaterialization = isLastStep
? undefined
: yield* tools.materialize({ permissions: agent.info?.permissions, model })
const toolMaterialization =
isLastStep || !resolved.capabilities.tools
? undefined
: yield* tools.materialize({ permissions: agent.info?.permissions, model })
const promptCacheKey = /^ses_[0-9a-f]{64}$/.test(session.id) ? session.id.slice(4) : session.id
const request = LLM.request({
model,
@@ -251,7 +252,7 @@ const layer = Layer.effect(
.filter((part): part is string => part !== undefined && part.length > 0)
.map(SystemPart.make),
messages: [
...toLLMMessages(context, resolved.ref, providerMetadataKey),
...toLLMMessages(context, resolved.ref, providerMetadataKey, resolved.capabilities),
...(isLastStep ? [Message.assistant(MAX_STEPS_PROMPT)] : []),
],
tools: toolMaterialization?.definitions ?? [],
+10 -1
View File
@@ -82,6 +82,8 @@ export interface Resolved {
readonly model: Model
/** Selected catalog identity. Durable records and displays must use this, never the API model id. */
readonly ref: ModelV2.Ref
/** Capabilities of the selected catalog model. */
readonly capabilities: ModelV2.Capabilities
/** Catalog pricing in dollars per million tokens. */
readonly cost: ModelV2.Info["cost"]
}
@@ -96,13 +98,19 @@ export class Service extends Context.Service<Service, Interface>()("@opencode/v2
export const layerWith = (resolve: Interface["resolve"]) => Layer.succeed(Service, Service.of({ resolve }))
/** Builds a Resolved whose catalog identity mirrors the route model. Test or embedding seam. */
export const resolved = (model: Model, variant?: ModelV2.VariantID, cost: ModelV2.Info["cost"] = []): Resolved => ({
export const resolved = (
model: Model,
variant?: ModelV2.VariantID,
cost: ModelV2.Info["cost"] = [],
capabilities = ModelV2.Capabilities.defaults(),
): Resolved => ({
model,
ref: ModelV2.Ref.make({
id: ModelV2.ID.make(model.id),
providerID: ProviderV2.ID.make(model.provider),
...(variant === undefined ? {} : { variant }),
}),
capabilities,
cost,
})
@@ -344,6 +352,7 @@ const layer = Layer.effect(
providerID: selected.providerID,
...(session.model?.variant === undefined ? {} : { variant: session.model.variant }),
}),
capabilities: selected.capabilities,
cost: selected.cost,
}
}),
@@ -11,8 +11,6 @@ import type { ModelV2 } from "../../model"
import { SessionMessage } from "../message"
import type { FileAttachment } from "@opencode-ai/schema/prompt"
const imageMimes = new Set(["image/png", "image/jpeg", "image/gif", "image/webp"])
const media = (file: FileAttachment): ContentPart => ({
type: "media",
mediaType: file.mime,
@@ -21,6 +19,23 @@ const media = (file: FileAttachment): ContentPart => ({
metadata: file.description === undefined ? undefined : { description: file.description },
})
const modality = (mime: string) => {
if (mime.startsWith("image/")) return "image"
if (mime.startsWith("audio/")) return "audio"
if (mime.startsWith("video/")) return "video"
if (mime === "application/pdf") return "pdf"
return undefined
}
const attachment = (file: FileAttachment, capabilities?: ModelV2.Capabilities): ContentPart => {
const type = modality(file.mime)
if (!type || (capabilities?.input ?? ["text", "image"]).includes(type)) return media(file)
return {
type: "text",
text: `ERROR: Cannot read ${file.name ? `"${file.name}"` : type} (this model does not support ${type} input). Inform the user.`,
}
}
const textAttachment = (file: FileAttachment) =>
Message.make({
role: "user",
@@ -168,7 +183,12 @@ const assistant = (message: SessionMessage.Assistant, model: ModelV2.Ref, provid
]
}
function toLLMMessage(message: SessionMessage.Info, model: ModelV2.Ref, providerMetadataKey: string): Message[] {
function toLLMMessage(
message: SessionMessage.Info,
model: ModelV2.Ref,
providerMetadataKey: string,
capabilities?: ModelV2.Capabilities,
): Message[] {
switch (message.type) {
case "agent-switched":
case "model-switched":
@@ -183,7 +203,9 @@ function toLLMMessage(message: SessionMessage.Info, model: ModelV2.Ref, provider
role: "user",
content: [
{ type: "text", text: message.text },
...files.filter((file) => imageMimes.has(file.mime)).map(media),
...files
.filter((file) => file.mime !== "text/plain" && file.mime !== "application/x-directory")
.map((file) => attachment(file, capabilities)),
],
metadata: {
...message.metadata,
@@ -236,4 +258,5 @@ export const toLLMMessages = (
messages: readonly SessionMessage.Info[],
model: ModelV2.Ref,
providerMetadataKey: string = model.providerID,
) => messages.flatMap((message) => toLLMMessage(message, model, providerMetadataKey))
capabilities?: ModelV2.Capabilities,
) => messages.flatMap((message) => toLLMMessage(message, model, providerMetadataKey, capabilities))
+10 -2
View File
@@ -7,6 +7,7 @@ import { ConfigMCPV1 } from "./mcp"
import { ConfigPermissionV1 } from "./permission"
import { ConfigProviderV1 } from "./provider"
import { ConfigProviderOptionsV1 } from "./provider-options"
import { ModelV2 } from "../../model"
import { ProviderV2 } from "../../provider"
const keys = new Set([
@@ -245,8 +246,15 @@ function migrateModel(info: typeof ConfigProviderV1.Model.Type) {
: []),
]
const capabilities =
info.tool_call !== undefined || info.modalities?.input !== undefined || info.modalities?.output !== undefined
? { tools: info.tool_call ?? false, input: info.modalities?.input ?? [], output: info.modalities?.output ?? [] }
info.tool_call !== undefined ||
info.attachment !== undefined ||
info.modalities?.input !== undefined ||
info.modalities?.output !== undefined
? ModelV2.Capabilities.defaults({
tools: info.tool_call,
input: info.modalities?.input ?? (info.attachment === false ? ["text"] : undefined),
output: info.modalities?.output,
})
: undefined
return {
modelID: info.id,
+15
View File
@@ -680,9 +680,17 @@ describe("Config", () => {
options: { apiKey: "secret" },
models: {
model: {
attachment: true,
options: { reasoningEffort: "high" },
variants: { fast: { temperature: 0.2 } },
},
text: {
attachment: false,
},
audio: {
attachment: true,
modalities: { input: ["audio"], output: ["audio"] },
},
},
},
openai: {
@@ -758,9 +766,16 @@ describe("Config", () => {
settings: { apiKey: "secret" },
models: {
model: {
capabilities: { tools: true, input: ["text", "image"], output: ["text"] },
settings: { reasoningEffort: "high" },
variants: [{ id: "fast", settings: { temperature: 0.2 } }],
},
text: {
capabilities: { tools: true, input: ["text"], output: ["text"] },
},
audio: {
capabilities: { tools: true, input: ["audio"], output: ["audio"] },
},
},
})
expect(documents[0]?.info.providers?.openai).toMatchObject({
@@ -237,6 +237,7 @@ describe("ConfigProviderPlugin.Plugin", () => {
const provider = required(yield* catalog.provider.get(providerID))
const model = required(yield* catalog.model.get(providerID, modelID))
const defaultModel = required(yield* catalog.model.get(providerID, ModelV2.ID.make("default")))
expect((yield* catalog.model.default())?.id).toBe(ModelV2.ID.make("default"))
expect(provider.name).toBe("Renamed")
expect((yield* integrations.get(Integration.ID.make("custom")))?.methods).toContainEqual({
@@ -252,6 +253,7 @@ describe("ConfigProviderPlugin.Plugin", () => {
expect(model.modelID).toBe(ModelV2.ID.make("api-chat"))
expect(model.name).toBe("Last")
expect(model.capabilities).toEqual({ tools: true, input: ["text"], output: ["text"] })
expect(defaultModel.capabilities).toEqual({ tools: true, input: ["text", "image"], output: ["text"] })
expect(model.enabled).toBe(false)
expect(model.limit).toEqual({ context: 100, output: 75 })
expect(model.cost).toEqual([
+15
View File
@@ -165,6 +165,21 @@ describe("ModelsDev Service", () => {
}),
)
it.effect("allows models.dev entries without legacy attachment metadata", () =>
Effect.sync(() => {
const result = Schema.decodeUnknownSync(ModelsDev.Model)({
id: "no-attachment-model",
name: "No Attachment Model",
release_date: "2026-01-01",
reasoning: false,
tool_call: true,
limit: { context: 128000, output: 8192 },
})
expect(result.attachment).toBeUndefined()
}),
)
it.live("get() returns providers from disk when cache file exists", () =>
Effect.gen(function* () {
yield* writeCache(fixture)
@@ -85,6 +85,24 @@ describe("ModelsDevPlugin", () => {
},
},
},
default: {
id: "default",
name: "Default",
release_date: "2026-01-01",
reasoning: false,
tool_call: false,
limit: { context: 128_000, output: 8_192 },
},
explicit: {
id: "explicit",
name: "Explicit",
release_date: "2026-01-01",
attachment: true,
reasoning: false,
tool_call: false,
modalities: { input: ["audio"], output: ["audio"] },
limit: { context: 128_000, output: 8_192 },
},
},
},
} satisfies Record<string, ModelsDev.Provider>),
@@ -101,9 +119,14 @@ describe("ModelsDevPlugin", () => {
const providerID = ProviderV2.ID.make("acme")
const base = yield* catalog.model.get(providerID, ModelV2.ID.make("gpt-5.4"))
const fast = yield* catalog.model.get(providerID, ModelV2.ID.make("gpt-5.4-fast"))
const defaults = yield* catalog.model.get(providerID, ModelV2.ID.make("default"))
const explicit = yield* catalog.model.get(providerID, ModelV2.ID.make("explicit"))
expect(base?.variants).toEqual([])
expect(base?.body).toEqual({})
expect(base?.capabilities).toEqual({ tools: true, input: ["text"], output: ["text"] })
expect(defaults?.capabilities).toEqual({ tools: false, input: ["text", "image"], output: ["text"] })
expect(explicit?.capabilities).toEqual({ tools: false, input: ["audio"], output: ["audio"] })
expect(fast).toMatchObject({
id: "gpt-5.4-fast",
modelID: "gpt-5.4",
@@ -181,6 +204,9 @@ describe("ModelsDevPlugin", () => {
connections: [],
}),
])
expect(yield* catalog.model.get(ProviderV2.ID.opencode, ModelV2.ID.make("gpt-5.5"))).toMatchObject({
capabilities: { tools: true, input: ["text", "image"], output: ["text"] },
})
}).pipe(Effect.provide(AppNodeBuilder.build(ModelsDev.node))),
(previous) =>
Effect.sync(() => {
@@ -240,7 +240,7 @@ Recent work
expect(messages[1]?.content).toEqual([{ type: "text", text: "Review this directory" }])
})
test("uses materialized image data as provider media and drops unsupported attachments", () => {
test("defaults missing model capabilities to text and image input", () => {
const data = Base64.make("AAECAw==")
const messages = toLLMMessages(
[
@@ -266,6 +266,113 @@ Recent work
expect(messages[0]?.content).toEqual([
{ type: "text", text: "Inspect this image" },
{ type: "media", mediaType: "image/png", data, filename: "image.png" },
{
type: "text",
text: 'ERROR: Cannot read "document.pdf" (this model does not support pdf input). Inform the user.',
},
])
})
test("uses explicit model input capabilities instead of attachment defaults", () => {
const data = Base64.make("AAECAw==")
const messages = toLLMMessages(
[
SessionMessage.User.make({
id: id("user-unsupported-pdf"),
type: "user",
text: "Inspect these files",
files: [
FileAttachment.make({ data, mime: "image/png", source: { type: "inline" }, name: "image.png" }),
FileAttachment.make({
data: Base64.make("JVBERg=="),
mime: "application/pdf",
source: { type: "inline" },
name: "document.pdf",
}),
],
time: { created },
}),
],
model,
model.providerID,
{ tools: true, input: ["text", "pdf"], output: ["text"] },
)
expect(messages[0]?.content).toEqual([
{ type: "text", text: "Inspect these files" },
{
type: "text",
text: 'ERROR: Cannot read "image.png" (this model does not support image input). Inform the user.',
},
{ type: "media", mediaType: "application/pdf", data: "JVBERg==", filename: "document.pdf" },
])
})
test("treats explicit empty input capabilities as authoritative", () => {
const messages = toLLMMessages(
[
SessionMessage.User.make({
id: id("user-empty-capabilities"),
type: "user",
text: "Inspect this image",
files: [
FileAttachment.make({
data: Base64.make("AAECAw=="),
mime: "image/png",
source: { type: "inline" },
name: "image.png",
}),
],
time: { created },
}),
],
model,
model.providerID,
{ tools: false, input: [], output: [] },
)
expect(messages[0]?.content).toEqual([
{ type: "text", text: "Inspect this image" },
{
type: "text",
text: 'ERROR: Cannot read "image.png" (this model does not support image input). Inform the user.',
},
])
})
test("classifies audio and video MIME families through explicit capabilities", () => {
const messages = toLLMMessages(
[
SessionMessage.User.make({
id: id("user-media-capabilities"),
type: "user",
text: "Inspect this media",
files: [
FileAttachment.make({
data: Base64.make("AAECAw=="),
mime: "audio/mpeg",
source: { type: "inline" },
name: "audio.mp3",
}),
FileAttachment.make({
data: Base64.make("AAECAw=="),
mime: "video/webm",
source: { type: "inline" },
name: "video.webm",
}),
],
time: { created },
}),
],
model,
model.providerID,
{ tools: false, input: ["text", "audio", "video"], output: ["text"] },
)
expect(messages[0]?.content).toEqual([
{ type: "text", text: "Inspect this media" },
{ type: "media", mediaType: "audio/mpeg", data: "AAECAw==", filename: "audio.mp3" },
{ type: "media", mediaType: "video/webm", data: "AAECAw==", filename: "video.webm" },
])
})
+20
View File
@@ -235,12 +235,15 @@ const echo = Layer.effectDiscard(
const echoNode = makeLocationNode({ name: "test/session-runner-tools", layer: echo, deps: [ToolRegistry.node] })
let modelResolveHook = Effect.void
let currentModel = model
let modelCapabilities = ModelV2.Capabilities.defaults()
const models = SessionRunnerModel.layerWith((session) =>
modelResolveHook.pipe(
Effect.as(
SessionRunnerModel.resolved(
session.model?.id === "replacement" ? replacementModel : currentModel,
session.model?.variant,
[],
modelCapabilities,
),
),
),
@@ -412,6 +415,7 @@ const setup = Effect.gen(function* () {
systemLoadHook = Effect.void
modelResolveHook = Effect.void
currentModel = model
modelCapabilities = ModelV2.Capabilities.defaults()
skillBaselines.clear()
responses = undefined
streamFailure = undefined
@@ -787,6 +791,22 @@ describe("SessionRunnerLLM", () => {
}),
)
it.effect("does not advertise tools to a model without tool capability", () =>
Effect.gen(function* () {
yield* setup
modelCapabilities = ModelV2.Capabilities.defaults({ tools: false })
const session = yield* SessionV2.Service
yield* session.prompt({ sessionID, prompt: PromptInput.Prompt.make({ text: "No tools" }), resume: false })
requests.length = 0
response = []
yield* session.resume(sessionID)
expect(requests).toHaveLength(1)
expect(requests[0]?.tools).toEqual([])
}),
)
it.effect("retries the first provider turn after system context becomes available", () =>
Effect.gen(function* () {
const session = yield* setup
+19 -2
View File
@@ -26,7 +26,24 @@ export const Capabilities = Schema.Struct({
tools: Schema.Boolean,
input: Schema.Array(Schema.String),
output: Schema.Array(Schema.String),
}).annotate({ identifier: "Model.Capabilities" })
})
.annotate({ identifier: "Model.Capabilities" })
.pipe(
statics((schema) => ({
defaults: (
input: {
readonly tools?: boolean
readonly input?: ReadonlyArray<string>
readonly output?: ReadonlyArray<string>
} = {},
) =>
schema.make({
tools: input.tools ?? true,
input: input.input === undefined ? ["text", "image"] : [...input.input],
output: input.output === undefined ? ["text"] : [...input.output],
}),
})),
)
export interface Cost extends Schema.Schema.Type<typeof Cost> {}
export const Cost = Schema.Struct({
@@ -80,7 +97,7 @@ export const Info = Schema.Struct({
modelID: id,
providerID,
name: id,
capabilities: { tools: false, input: [], output: [] },
capabilities: Capabilities.defaults(),
variants: [],
time: { released: 0 },
cost: [],