feat: allow read tool to handle images (#3052)

This commit is contained in:
Aiden Cline
2025-10-09 09:05:11 -05:00
committed by GitHub
parent eb4b5721cd
commit 225adc46ba
7 changed files with 159 additions and 83 deletions

View File

@@ -28,6 +28,12 @@ export namespace ModelsDev {
context: z.number(),
output: z.number(),
}),
modalities: z
.object({
input: z.array(z.enum(["text", "audio", "image", "video", "pdf"])),
output: z.array(z.enum(["text", "audio", "image", "video", "pdf"])),
})
.optional(),
experimental: z.boolean().optional(),
options: z.record(z.string(), z.any()),
provider: z.object({ npm: z.string() }).optional(),

View File

@@ -279,6 +279,11 @@ export namespace Provider {
context: 0,
output: 0,
},
modalities: model.modalities ??
existing?.modalities ?? {
input: ["text"],
output: ["text"],
},
provider: model.provider ?? existing?.provider,
}
parsed.models[modelID] = parsedModel

View File

@@ -17,71 +17,6 @@ export namespace MessageV2 {
}),
)
export const ToolStatePending = z
.object({
status: z.literal("pending"),
})
.meta({
ref: "ToolStatePending",
})
export type ToolStatePending = z.infer<typeof ToolStatePending>
export const ToolStateRunning = z
.object({
status: z.literal("running"),
input: z.any(),
title: z.string().optional(),
metadata: z.record(z.string(), z.any()).optional(),
time: z.object({
start: z.number(),
}),
})
.meta({
ref: "ToolStateRunning",
})
export type ToolStateRunning = z.infer<typeof ToolStateRunning>
export const ToolStateCompleted = z
.object({
status: z.literal("completed"),
input: z.record(z.string(), z.any()),
output: z.string(),
title: z.string(),
metadata: z.record(z.string(), z.any()),
time: z.object({
start: z.number(),
end: z.number(),
compacted: z.number().optional(),
}),
})
.meta({
ref: "ToolStateCompleted",
})
export type ToolStateCompleted = z.infer<typeof ToolStateCompleted>
export const ToolStateError = z
.object({
status: z.literal("error"),
input: z.record(z.string(), z.any()),
error: z.string(),
metadata: z.record(z.string(), z.any()).optional(),
time: z.object({
start: z.number(),
end: z.number(),
}),
})
.meta({
ref: "ToolStateError",
})
export type ToolStateError = z.infer<typeof ToolStateError>
export const ToolState = z
.discriminatedUnion("status", [ToolStatePending, ToolStateRunning, ToolStateCompleted, ToolStateError])
.meta({
ref: "ToolState",
})
const PartBase = z.object({
id: z.string(),
sessionID: z.string(),
@@ -134,17 +69,6 @@ export namespace MessageV2 {
})
export type ReasoningPart = z.infer<typeof ReasoningPart>
export const ToolPart = PartBase.extend({
type: z.literal("tool"),
callID: z.string(),
tool: z.string(),
state: ToolState,
metadata: z.record(z.string(), z.any()).optional(),
}).meta({
ref: "ToolPart",
})
export type ToolPart = z.infer<typeof ToolPart>
const FilePartSourceBase = z.object({
text: z
.object({
@@ -228,6 +152,83 @@ export namespace MessageV2 {
})
export type StepFinishPart = z.infer<typeof StepFinishPart>
export const ToolStatePending = z
.object({
status: z.literal("pending"),
})
.meta({
ref: "ToolStatePending",
})
export type ToolStatePending = z.infer<typeof ToolStatePending>
export const ToolStateRunning = z
.object({
status: z.literal("running"),
input: z.any(),
title: z.string().optional(),
metadata: z.record(z.string(), z.any()).optional(),
time: z.object({
start: z.number(),
}),
})
.meta({
ref: "ToolStateRunning",
})
export type ToolStateRunning = z.infer<typeof ToolStateRunning>
export const ToolStateCompleted = z
.object({
status: z.literal("completed"),
input: z.record(z.string(), z.any()),
output: z.string(),
title: z.string(),
metadata: z.record(z.string(), z.any()),
time: z.object({
start: z.number(),
end: z.number(),
compacted: z.number().optional(),
}),
attachments: FilePart.array().optional(),
})
.meta({
ref: "ToolStateCompleted",
})
export type ToolStateCompleted = z.infer<typeof ToolStateCompleted>
export const ToolStateError = z
.object({
status: z.literal("error"),
input: z.record(z.string(), z.any()),
error: z.string(),
metadata: z.record(z.string(), z.any()).optional(),
time: z.object({
start: z.number(),
end: z.number(),
}),
})
.meta({
ref: "ToolStateError",
})
export type ToolStateError = z.infer<typeof ToolStateError>
export const ToolState = z
.discriminatedUnion("status", [ToolStatePending, ToolStateRunning, ToolStateCompleted, ToolStateError])
.meta({
ref: "ToolState",
})
export const ToolPart = PartBase.extend({
type: z.literal("tool"),
callID: z.string(),
tool: z.string(),
state: ToolState,
metadata: z.record(z.string(), z.any()).optional(),
}).meta({
ref: "ToolPart",
})
export type ToolPart = z.infer<typeof ToolPart>
const Base = z.object({
id: z.string(),
sessionID: z.string(),
@@ -531,7 +532,25 @@ export namespace MessageV2 {
},
]
if (part.type === "tool") {
if (part.state.status === "completed")
if (part.state.status === "completed") {
if (part.state.attachments?.length) {
result.push({
id: Identifier.ascending("message"),
role: "user",
parts: [
{
type: "text",
text: `Tool ${part.tool} returned an attachment:`,
},
...part.state.attachments.map((attachment) => ({
type: "file" as const,
url: attachment.url,
mediaType: attachment.mime,
filename: attachment.filename,
})),
],
})
}
return [
{
type: ("tool-" + part.tool) as `tool-${string}`,
@@ -542,6 +561,7 @@ export namespace MessageV2 {
callProviderMetadata: part.metadata,
},
]
}
if (part.state.status === "error")
return [
{

View File

@@ -457,6 +457,10 @@ export namespace SessionPrompt {
abort: options.abortSignal!,
messageID: input.processor.message.id,
callID: options.toolCallId,
extra: {
modelID: input.modelID,
providerID: input.providerID,
},
agent: input.agent.name,
metadata: async (val) => {
const match = input.processor.partFromToolCall(options.toolCallId)
@@ -989,6 +993,7 @@ export namespace SessionPrompt {
start: match.state.time.start,
end: Date.now(),
},
attachments: value.output.attachments,
},
})
delete toolcalls[value.toolCallId]

View File

@@ -7,6 +7,8 @@ import { FileTime } from "../file/time"
import DESCRIPTION from "./read.txt"
import { Filesystem } from "../util/filesystem"
import { Instance } from "../project/instance"
import { Provider } from "../provider/provider"
import { Identifier } from "../id/id"
const DEFAULT_READ_LIMIT = 2000
const MAX_LINE_LENGTH = 2000
@@ -23,6 +25,8 @@ export const ReadTool = Tool.define("read", {
if (!path.isAbsolute(filepath)) {
filepath = path.join(process.cwd(), filepath)
}
const title = path.relative(Instance.worktree, filepath)
if (!ctx.extra?.["bypassCwdCheck"] && !Filesystem.contains(Instance.directory, filepath)) {
throw new Error(`File ${filepath} is not in the current working directory`)
}
@@ -48,12 +52,45 @@ export const ReadTool = Tool.define("read", {
throw new Error(`File not found: ${filepath}`)
}
const limit = params.limit ?? DEFAULT_READ_LIMIT
const offset = params.offset || 0
const isImage = isImageFile(filepath)
if (isImage) throw new Error(`This is an image file of type: ${isImage}\nUse a different tool to process images`)
const supportsImages = await (async () => {
if (!ctx.extra?.["providerID"] || !ctx.extra?.["modelID"]) return false
const providerID = ctx.extra["providerID"] as string
const modelID = ctx.extra["modelID"] as string
const model = await Provider.getModel(providerID, modelID).catch(() => undefined)
if (!model) return false
return model.info.modalities?.input?.includes("image") ?? false
})()
if (isImage) {
if (!supportsImages) {
throw new Error(`Failed to read image: ${filepath}, model may not be able to read images`)
}
const mime = file.type
const msg = "Image read successfully"
return {
title,
output: msg,
metadata: {
preview: msg,
},
attachments: [
{
id: Identifier.ascending("part"),
sessionID: ctx.sessionID,
messageID: ctx.messageID,
type: "file",
mime,
url: `data:${mime};base64,${Buffer.from(await file.bytes()).toString("base64")}`,
},
],
}
}
const isBinary = await isBinaryFile(filepath, file)
if (isBinary) throw new Error(`Cannot read binary file: ${filepath}`)
const limit = params.limit ?? DEFAULT_READ_LIMIT
const offset = params.offset || 0
const lines = await file.text().then((text) => text.split("\n"))
const raw = lines.slice(offset, offset + limit).map((line) => {
return line.length > MAX_LINE_LENGTH ? line.substring(0, MAX_LINE_LENGTH) + "..." : line
@@ -76,7 +113,7 @@ export const ReadTool = Tool.define("read", {
FileTime.read(ctx.sessionID, filepath)
return {
title: path.relative(Instance.worktree, filepath),
title,
output,
metadata: {
preview,

View File

@@ -7,6 +7,6 @@ Usage:
- You can optionally specify a line offset and limit (especially handy for long files), but it's recommended to read the whole file by not providing these parameters
- Any lines longer than 2000 characters will be truncated
- Results are returned using cat -n format, with line numbers starting at 1
- This tool cannot read binary files, including images
- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.
- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.
- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.
- You can read image files using this tool.

View File

@@ -1,9 +1,11 @@
import z from "zod/v4"
import type { MessageV2 } from "../session/message-v2"
export namespace Tool {
interface Metadata {
[key: string]: any
}
export type Context<M extends Metadata = Metadata> = {
sessionID: string
messageID: string
@@ -25,6 +27,7 @@ export namespace Tool {
title: string
metadata: M
output: string
attachments?: MessageV2.FilePart[]
}>
}>
}