feat(core): improve read model output (#39146)

This commit is contained in:
Aiden Cline 2026-07-27 14:52:50 -05:00 committed by GitHub
parent 1b39d364bd
commit 6da2f3c38f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 147 additions and 54 deletions

View file

@ -1,7 +1,7 @@
export * as ReadTool from "./read"
import type { Context as PluginContext } from "@opencode-ai/plugin/effect/plugin"
import { dirname } from "path"
import { basename, dirname, join } from "path"
import { ToolFailure } from "@opencode-ai/ai"
import { Effect, Schema } from "effect"
import { FSUtil } from "@opencode-ai/util/fs-util"
@ -16,12 +16,12 @@ export const name = "read"
const FILENAME = "AGENTS.md"
const SUPPORTED_MEDIA_MIMES = new Set(["image/jpeg", "image/png", "image/gif", "image/webp", "application/pdf"])
const LocationInput = Schema.Struct({
path: Schema.String,
path: Schema.String.annotate({ description: "File or directory to read" }),
offset: ReadToolFileSystem.PageInput.fields.offset.annotate({
description: "The 1-based directory entry or text line offset to start reading from",
description: "The line or directory entry to start reading from (1-based)",
}),
limit: ReadToolFileSystem.PageInput.fields.limit.annotate({
description: "The maximum number of directory entries or text lines to read",
description: "The maximum number of lines or directory entries to read (defaults to 2000)",
}),
})
const Input = LocationInput
@ -48,7 +48,7 @@ export const Plugin = {
name,
options: { codemode: false },
description:
"Read a text file or supported image, page through a large UTF-8 text file by line offset, or list a directory page. Relative paths resolve from the current location; absolute paths inside it are accepted, while external absolute paths require external_directory approval.",
"Read the contents of a file or directory. Supports text files, images, and PDFs. Images and PDFs are presented directly to the model. Each text line is prefixed by its 1-based line number as <line>: <content>. The prefix is for reference and is not part of the file content. Directory entries are returned one per line. Use offset and limit to read large files or directories in sections. Prefer one larger read over many small slices, and use grep to find specific content in large files.",
input: Input,
output: Output,
execute: (input, context) => {
@ -69,7 +69,6 @@ export const Plugin = {
})
const resource = target.resource
const absolute = AbsolutePath.make(target.canonical)
const type = yield* reader.inspect(absolute)
yield* permission.assert({
action: name,
resources: [resource],
@ -78,6 +77,9 @@ export const Plugin = {
agent: context.agent,
source,
})
const type = yield* reader.inspect(absolute).pipe(
Effect.catchReason("PlatformError", "NotFound", () => missing(input.path, target.canonical)),
)
const content =
type === "directory"
? yield* reader.list(absolute, { offset: input.offset, limit: input.limit })
@ -114,30 +116,12 @@ export const Plugin = {
return yield* Effect.fail(new ReadToolFileSystem.BinaryFileError({ resource }))
return content
}).pipe(
Effect.map((output) => {
// Image base64 reaches the model through content items; avoid a second
// unresized copy in model text.
const content =
output.type === "file" && output.encoding === "base64"
? SUPPORTED_MEDIA_MIMES.has(output.mime)
? ([
{
type: "text",
text:
output.mime === "application/pdf" ? "PDF read successfully" : "Image read successfully",
},
{
type: "file",
uri: `data:${output.mime};base64,${output.content}`,
mime: output.mime,
name: input.path,
},
] as const)
: JSON.stringify({ ...output, content: "" }, null, 2)
: JSON.stringify(output, null, 2)
return { output, content }
}),
Effect.map((output) => ({
output,
content: toModelContent(input.path, input.offset, output),
})),
Effect.mapError((error) => {
if (error instanceof ToolFailure) return error
const message =
error instanceof ReadToolFileSystem.BinaryFileError ||
error instanceof ReadToolFileSystem.MediaIngestLimitError ||
@ -154,5 +138,62 @@ export const Plugin = {
),
)
.pipe(Effect.orDie)
const missing = Effect.fn("ReadTool.missing")(function* (input: string, canonical: string) {
const base = basename(input).toLowerCase()
const suggestions = yield* fs.readDirectory(dirname(canonical)).pipe(
Effect.map((entries) =>
entries
.filter((entry) => {
const candidate = entry.toLowerCase()
return candidate.includes(base) || base.includes(candidate)
})
.map((entry) => join(dirname(input), entry))
.slice(0, 3),
),
Effect.catch(() => Effect.succeed([] as string[])),
)
const message =
suggestions.length === 0
? `File not found: ${input}`
: `File not found: ${input}\n\nDid you mean one of these?\n${suggestions.join("\n")}`
return yield* new ToolFailure({ message })
})
}),
}
export const toModelContent = (path: string, offset: number | undefined, output: typeof Output.Type) => {
if (output.type === "file" && output.encoding === "base64")
return [
{ type: "text", text: output.mime === "application/pdf" ? "PDF read successfully" : "Image read successfully" },
{
type: "file",
uri: `data:${output.mime};base64,${output.content}`,
mime: output.mime,
name: path,
},
] as const
if (output.type === "list-page") {
const start = offset ?? 1
const content = [
output.entries.length === 0
? `Read directory ${path}, 0 entries`
: `Read directory ${path}, entries ${start}-${start + output.entries.length - 1}`,
]
output.entries.forEach((entry) => content.push(entry.path))
if (output.truncated && output.next !== undefined)
content.push(`[Output truncated. Continue reading with offset: ${output.next}]`)
return content.join("\n")
}
const start = output.type === "text-page" ? output.offset : 1
const lines = output.content === "" ? [] : output.content.replace(/\n$/, "").split("\n")
const content = [
lines.length === 0 ? `Read file ${path}, 0 lines` : `Read file ${path}, lines ${start}-${start + lines.length - 1}`,
]
lines.forEach((line, index) => content.push(`${start + index}: ${line}`))
if (output.type === "text-page" && output.truncated && output.next !== undefined)
content.push(`[Output truncated. Continue reading with offset: ${output.next}]`)
return content.join("\n")
}

View file

@ -5,12 +5,13 @@ import { Config } from "@opencode-ai/core/config"
import { ConfigAttachments } from "@opencode-ai/core/config/attachments"
import { AppNodeBuilder } from "@opencode-ai/core/effect/app-node-builder"
import { LayerNode } from "@opencode-ai/util/effect/layer-node"
import { FileSystem } from "@opencode-ai/core/filesystem"
import { FSUtil } from "@opencode-ai/util/fs-util"
import { Location } from "@opencode-ai/core/location"
import { Image } from "@opencode-ai/core/image"
import { Permission } from "@opencode-ai/core/permission"
import { Session } from "@opencode-ai/core/session"
import { AbsolutePath } from "@opencode-ai/core/schema"
import { AbsolutePath, RelativePath } from "@opencode-ai/core/schema"
import { Global } from "@opencode-ai/util/global"
import { LocationMutation } from "@opencode-ai/core/location-mutation"
import { location } from "./fixture/location"
@ -40,13 +41,23 @@ const readToolNode = makeLocationNode({
const assertions: Permission.AssertInput[] = []
const missingPath = "__missing_read_target__.txt"
const missingAbsolutePath = path.join(process.cwd(), missingPath)
const notFound = (target: string) =>
PlatformError.systemError({
_tag: "NotFound",
module: "FileSystem",
method: "stat",
pathOrDescriptor: target,
})
const readCalls: {
input: AbsolutePath
page: ReadToolFileSystem.PageInput
}[] = []
const listCalls: ReadToolFileSystem.PageInput[] = []
let listResult = new ReadToolFileSystem.ListPage({ type: "list-page", entries: [], truncated: false })
let resolvedType: "file" | "directory" = "file"
let resolveFailure: unknown
let inspectFailure: ReadToolFileSystem.InspectError | undefined
let directoryEntries: string[] = []
let readResult: ReadToolFileSystem.FileContent | ReadToolFileSystem.TextPage = {
type: "file",
uri: "file:///README.md",
@ -59,7 +70,12 @@ let readFailure: ReadToolFileSystem.ReadError | undefined
const reader = Layer.succeed(
ReadToolFileSystem.Service,
ReadToolFileSystem.Service.of({
inspect: () => (resolveFailure === undefined ? Effect.succeed(resolvedType) : Effect.die(resolveFailure)),
inspect: () =>
resolveFailure !== undefined
? Effect.die(resolveFailure)
: inspectFailure !== undefined
? Effect.fail(inspectFailure)
: Effect.succeed(resolvedType),
read: (input, _resource, page = {}) => {
readCalls.push({ input, page })
if (readFailure !== undefined) return Effect.fail(readFailure)
@ -68,7 +84,7 @@ const reader = Layer.succeed(
list: (_path, input = {}) =>
Effect.sync(() => {
listCalls.push(input)
return new ReadToolFileSystem.ListPage({ type: "list-page", entries: [], truncated: false })
return listResult
}),
}),
)
@ -107,6 +123,7 @@ const testFileSystem = Layer.effect(
Effect.succeed(
FSUtil.Service.of({
...fs,
readDirectory: () => Effect.succeed(directoryEntries),
realPath: (path) =>
path === missingAbsolutePath
? Effect.fail(
@ -130,8 +147,6 @@ const mutation = Layer.succeed(
LocationMutation.Service,
LocationMutation.Service.of({
resolve: (input) => {
if (input.path === missingPath)
return Effect.fail(new LocationMutation.PathError({ path: input.path, reason: "non_directory_ancestor" }))
const canonical = path.resolve(process.cwd(), input.path)
const external = path.isAbsolute(input.path) && !FSUtil.contains(process.cwd(), canonical)
const resource = external ? canonical.replaceAll("\\", "/") : path.relative(process.cwd(), canonical) || "."
@ -183,6 +198,8 @@ describe("ReadTool", () => {
allow = true
resolvedType = "file"
resolveFailure = undefined
inspectFailure = undefined
directoryEntries = []
readResult = {
type: "file",
uri: "file:///README.md",
@ -192,6 +209,7 @@ describe("ReadTool", () => {
mime: "text/plain",
}
readFailure = undefined
listResult = new ReadToolFileSystem.ListPage({ type: "list-page", entries: [], truncated: false })
})
it.effect("registers, authorizes, and reads through the location filesystem", () =>
@ -219,6 +237,7 @@ describe("ReadTool", () => {
encoding: "utf8",
mime: "text/plain",
})
expect(execution.content).toEqual([{ type: "text", text: "Read file README.md, lines 1-1\n1: hello" }])
expect(assertions).toMatchObject([{ sessionID, action: "read", resources: ["README.md"], save: ["*"] }])
expect(readCalls).toEqual([
{
@ -658,6 +677,14 @@ describe("ReadTool", () => {
it.effect("returns missing paths as model-visible tool failures", () =>
Effect.gen(function* () {
inspectFailure = notFound(missingAbsolutePath)
directoryEntries = [
"__missing_read_target__.txt.bak",
"copy___missing_read_target__.txt",
"old___missing_read_target__.txt",
"other___missing_read_target__.txt",
"unrelated.txt",
]
const registry = yield* Tool.Service
expect(
@ -666,10 +693,14 @@ describe("ReadTool", () => {
...toolIdentity,
call: { type: "tool-call", id: "call-missing-path", name: "read", input: { path: missingPath } },
}),
// The message-less PathError cause must not erase the tool's curated
// failure message; the canonical error is the sole authority.
).toEqual({ status: "error", error: { type: "tool.execution", message: `Unable to read ${missingPath}` } })
expect(assertions).toEqual([])
).toEqual({
status: "error",
error: {
type: "tool.execution",
message: `File not found: ${missingPath}\n\nDid you mean one of these?\n__missing_read_target__.txt.bak\ncopy___missing_read_target__.txt\nold___missing_read_target__.txt`,
},
})
expect(assertions).toMatchObject([{ sessionID, action: "read", resources: [missingPath], save: ["*"] }])
expect(readCalls).toEqual([])
}),
)
@ -677,10 +708,18 @@ describe("ReadTool", () => {
it.effect("lists a bounded directory page through read", () =>
Effect.gen(function* () {
resolvedType = "directory"
listResult = new ReadToolFileSystem.ListPage({
type: "list-page",
entries: [
FileSystem.Entry.make({ path: RelativePath.make("components/"), type: "directory" }),
FileSystem.Entry.make({ path: RelativePath.make("index.ts"), type: "file" }),
],
truncated: true,
next: 4,
})
const registry = yield* Tool.Service
expect(
yield* executeTool(registry, {
const result = yield* executeTool(registry, {
sessionID,
...toolIdentity,
call: {
@ -689,8 +728,15 @@ describe("ReadTool", () => {
name: "read",
input: { path: "src", offset: 2, limit: 10 },
},
}),
).toMatchObject({ status: "completed", output: { entries: [], truncated: false } })
})
expect(result).toMatchObject({ status: "completed", output: { entries: listResult.entries, truncated: true, next: 4 } })
if (result.status !== "completed") return
expect(result.content).toEqual([
{
type: "text",
text: "Read directory src, entries 2-3\ncomponents/\nindex.ts\n[Output truncated. Continue reading with offset: 4]",
},
])
expect(assertions).toMatchObject([{ sessionID, action: "read", resources: ["src"], save: ["*"] }])
expect(listCalls).toEqual([{ offset: 2, limit: 10 }])
}),
@ -744,21 +790,27 @@ describe("ReadTool", () => {
})
const registry = yield* Tool.Service
expect(
yield* executeTool(registry, {
sessionID,
...toolIdentity,
call: {
type: "tool-call",
id: "call-large",
name: "read",
input: { path: "large.txt", offset: 2, limit: 1 },
},
}),
).toMatchObject({
const result = yield* executeTool(registry, {
sessionID,
...toolIdentity,
call: {
type: "tool-call",
id: "call-large",
name: "read",
input: { path: "large.txt", offset: 2, limit: 1 },
},
})
expect(result).toMatchObject({
status: "completed",
output: { type: "text-page", content: "hello", mime: "text/plain", offset: 2, truncated: true, next: 3 },
})
if (result.status !== "completed") return
expect(result.content).toEqual([
{
type: "text",
text: "Read file large.txt, lines 2-2\n2: hello\n[Output truncated. Continue reading with offset: 3]",
},
])
expect(readCalls).toEqual([
{ input: AbsolutePath.make(path.join(process.cwd(), "large.txt")), page: { offset: 2, limit: 1 } },
])