EXT-617: Refactor read_file tool with pagination support

packages/types/src/tool-params.ts

@@ -12,6 +12,57 @@ export interface FileEntry {
 	lineRanges?: LineRange[]
 }
 
+/**
+ * read_file tool input parameters (new spec)
+ */
+export interface ReadFileInput {
+	file_path: string
+	offset?: number
+	limit?: number
+	format?: "cat_n"
+	max_chars_per_line?: number
+}
+
+/**
+ * read_file tool success output
+ */
+export interface ReadFileSuccess {
+	ok: true
+	file_path: string
+	resolved_path: string
+	mime_type: string
+	encoding: string | null
+	line_offset: number
+	lines_returned: number
+	reached_eof: boolean
+	truncated: boolean
+	truncation_reason?: "limit" | "max_chars_per_line" | "max_total_chars" | "binary_policy"
+	next_offset: number | null
+	content: string
+	warnings: string[]
+}
+
+/**
+ * read_file tool error output
+ */
+export interface ReadFileError {
+	ok: false
+	error: {
+		code:
+			| "file_not_found"
+			| "permission_denied"
+			| "outside_workspace"
+			| "is_directory"
+			| "unsupported_mime_type"
+			| "decode_failed"
+			| "io_error"
+		message: string
+		details?: Record<string, unknown>
+	}
+}
+
+export type ReadFileOutput = ReadFileSuccess | ReadFileError
+
 export interface Coordinate {
 	x: number
 	y: number

src/core/prompts/tools/native-tools/index.ts

@@ -11,7 +11,7 @@ import fetchInstructions from "./fetch_instructions"
 import generateImage from "./generate_image"
 import listFiles from "./list_files"
 import newTask from "./new_task"
-import { createReadFileTool, type ReadFileToolOptions } from "./read_file"
+import { createReadFileTool } from "./read_file"
 import runSlashCommand from "./run_slash_command"
 import searchAndReplace from "./search_and_replace"
 import searchReplace from "./search_replace"
@@ -23,35 +23,21 @@ import writeToFile from "./write_to_file"
 
 export { getMcpServerTools } from "./mcp_server"
 export { convertOpenAIToolToAnthropic, convertOpenAIToolsToAnthropic } from "./converters"
-export type { ReadFileToolOptions } from "./read_file"
 
 /**
  * Options for customizing the native tools array.
+ * Currently empty but reserved for future tool configuration.
  */
-export interface NativeToolsOptions {
-	/** Whether to include line_ranges support in read_file tool (default: true) */
-	partialReadsEnabled?: boolean
-	/** Maximum number of files that can be read in a single read_file request (default: 5) */
-	maxConcurrentFileReads?: number
-	/** Whether the model supports image processing (default: false) */
-	supportsImages?: boolean
-}
+// eslint-disable-next-line @typescript-eslint/no-empty-object-type
+export interface NativeToolsOptions {}
 
 /**
- * Get native tools array, optionally customizing based on settings.
+ * Get native tools array.
  *
- * @param options - Configuration options for the tools
+ * @param options - Configuration options for the tools (currently unused)
  * @returns Array of native tool definitions
  */
 export function getNativeTools(options: NativeToolsOptions = {}): OpenAI.Chat.ChatCompletionTool[] {
-	const { partialReadsEnabled = true, maxConcurrentFileReads = 5, supportsImages = false } = options
-
-	const readFileOptions: ReadFileToolOptions = {
-		partialReadsEnabled,
-		maxConcurrentFileReads,
-		supportsImages,
-	}
-
 	return [
 		accessMcpResource,
 		apply_diff,
@@ -65,7 +51,7 @@ export function getNativeTools(options: NativeToolsOptions = {}): OpenAI.Chat.Ch
 		generateImage,
 		listFiles,
 		newTask,
-		createReadFileTool(readFileOptions),
+		createReadFileTool(),
 		runSlashCommand,
 		searchAndReplace,
 		searchReplace,

src/core/prompts/tools/native-tools/read_file.ts

@@ -1,97 +1,41 @@
 import type OpenAI from "openai"
 
 /**
- * Generates the file support note, optionally including image format support.
+ * Creates the read_file tool definition following the paginated read spec.
  *
- * @param supportsImages - Whether the model supports image processing
- * @returns Support note string
- */
-function getReadFileSupportsNote(supportsImages: boolean): string {
-	if (supportsImages) {
-		return `Supports text extraction from PDF and DOCX files. Automatically processes and returns image files (PNG, JPG, JPEG, GIF, BMP, SVG, WEBP, ICO, AVIF) for visual analysis. May not handle other binary files properly.`
-	}
-	return `Supports text extraction from PDF and DOCX files, but may not handle other binary files properly.`
-}
-
-/**
- * Options for creating the read_file tool definition.
- */
-export interface ReadFileToolOptions {
-	/** Whether to include line_ranges parameter (default: true) */
-	partialReadsEnabled?: boolean
-	/** Maximum number of files that can be read in a single request (default: 5) */
-	maxConcurrentFileReads?: number
-	/** Whether the model supports image processing (default: false) */
-	supportsImages?: boolean
-}
-
-/**
- * Creates the read_file tool definition, optionally including line_ranges support
- * based on whether partial reads are enabled.
+ * Single-file reads with line-based pagination, stable line numbering,
+ * and bounded output to stay within context budgets.
  *
- * @param options - Configuration options for the tool
  * @returns Native tool definition for read_file
  */
-export function createReadFileTool(options: ReadFileToolOptions = {}): OpenAI.Chat.ChatCompletionTool {
-	const { partialReadsEnabled = true, maxConcurrentFileReads = 5, supportsImages = false } = options
-	const isMultipleReadsEnabled = maxConcurrentFileReads > 1
+export function createReadFileTool(): OpenAI.Chat.ChatCompletionTool {
+	const description = `Request to read a file with line-based pagination. Returns at most 2000 lines per call (configurable via limit parameter). Use offset parameter to read subsequent chunks.
 
-	// Build description intro with concurrent reads limit message
-	const descriptionIntro = isMultipleReadsEnabled
-		? `Read one or more files and return their contents with line numbers for diffing or discussion. IMPORTANT: You can read a maximum of ${maxConcurrentFileReads} files in a single request. If you need to read more files, use multiple sequential read_file requests. `
-		: "Read a file and return its contents with line numbers for diffing or discussion. IMPORTANT: Multiple file reads are currently disabled. You can only read one file at a time. "
+Path Resolution and Sandbox:
+- file_path is required and must be relative to workspace root
+- Paths are resolved to absolute and canonicalized
+- Access is restricted to workspace root (sandbox enforcement)
+- Directories are rejected
 
-	const baseDescription =
-		descriptionIntro +
-		"Structure: { files: [{ path: 'relative/path.ts'" +
-		(partialReadsEnabled ? ", line_ranges: [[1, 50], [100, 150]]" : "") +
-		" }] }. " +
-		"The 'path' is required and relative to workspace. "
+Pagination:
+- offset (default: 0): 0-based line offset. offset=0 starts at file line 1
+- limit (default: 2000): Maximum lines per call (hard cap)
+- When reached_eof=false in response, continue with offset=next_offset
+- Line numbers are file-global and stable across chunks
 
-	const optionalRangesDescription = partialReadsEnabled
-		? "The 'line_ranges' is optional for reading specific sections. Each range is a [start, end] tuple (1-based inclusive). "
-		: ""
+Output Format:
+- format (default: "cat_n"): Returns cat -n style with right-aligned line numbers
+- max_chars_per_line (default: 2000): Truncates lines exceeding this limit
+- Binary files return error with code "unsupported_mime_type"
 
-	const examples = partialReadsEnabled
-		? "Example single file: { files: [{ path: 'src/app.ts' }] }. " +
-			"Example with line ranges: { files: [{ path: 'src/app.ts', line_ranges: [[1, 50], [100, 150]] }] }. " +
-			(isMultipleReadsEnabled
-				? `Example multiple files (within ${maxConcurrentFileReads}-file limit): { files: [{ path: 'file1.ts', line_ranges: [[1, 50]] }, { path: 'file2.ts' }] }`
-				: "")
-		: "Example single file: { files: [{ path: 'src/app.ts' }] }. " +
-			(isMultipleReadsEnabled
-				? `Example multiple files (within ${maxConcurrentFileReads}-file limit): { files: [{ path: 'file1.ts' }, { path: 'file2.ts' }] }`
-				: "")
+Example: Read first chunk of file:
+{ "file_path": "src/main.ts" }
 
-	const description =
-		baseDescription + optionalRangesDescription + getReadFileSupportsNote(supportsImages) + " " + examples
+Example: Read next chunk using pagination:
+{ "file_path": "src/main.ts", "offset": 2000 }
 
-	// Build the properties object conditionally
-	const fileProperties: Record<string, any> = {
-		path: {
-			type: "string",
-			description: "Path to the file to read, relative to the workspace",
-		},
-	}
-
-	// Only include line_ranges if partial reads are enabled
-	if (partialReadsEnabled) {
-		fileProperties.line_ranges = {
-			type: ["array", "null"],
-			description:
-				"Optional line ranges to read. Each range is a [start, end] tuple with 1-based inclusive line numbers. Use multiple ranges for non-contiguous sections.",
-			items: {
-				type: "array",
-				items: { type: "integer" },
-				minItems: 2,
-				maxItems: 2,
-			},
-		}
-	}
-
-	// When using strict mode, ALL properties must be in the required array
-	// Optional properties are handled by having type: ["...", "null"]
-	const fileRequiredProperties = partialReadsEnabled ? ["path", "line_ranges"] : ["path"]
+Example: Read specific section with custom limit:
+{ "file_path": "src/main.ts", "offset": 100, "limit": 50 }`
 
 	return {
 		type: "function",
@@ -102,23 +46,33 @@ export function createReadFileTool(options: ReadFileToolOptions = {}): OpenAI.Ch
 			parameters: {
 				type: "object",
 				properties: {
-					files: {
-						type: "array",
-						description: "List of files to read; request related files together when allowed",
-						items: {
-							type: "object",
-							properties: fileProperties,
-							required: fileRequiredProperties,
-							additionalProperties: false,
-						},
-						minItems: 1,
+					file_path: {
+						type: "string",
+						description: "Path to the file to read, relative to workspace root (required)",
+					},
+					offset: {
+						type: ["integer", "null"],
+						description: "0-based line offset. offset=0 starts at file line 1 (default: 0)",
+					},
+					limit: {
+						type: ["integer", "null"],
+						description: "Maximum number of lines to return (default: 2000, hard cap enforced)",
+					},
+					format: {
+						type: ["string", "null"],
+						description: 'Output format, currently only "cat_n" supported (default: "cat_n")',
+						enum: ["cat_n", null],
+					},
+					max_chars_per_line: {
+						type: ["integer", "null"],
+						description: "Maximum characters per line before truncation (default: 2000)",
 					},
 				},
-				required: ["files"],
+				required: ["file_path", "offset", "limit", "format", "max_chars_per_line"],
 				additionalProperties: false,
 			},
 		},
 	} satisfies OpenAI.Chat.ChatCompletionTool
 }
 
-export const read_file = createReadFileTool({ partialReadsEnabled: false })
+export const read_file = createReadFileTool()

src/core/task/build-tools.ts

@@ -109,18 +109,8 @@ export async function buildNativeToolsArrayWithRestrictions(options: BuildToolsO
 		modelInfo,
 	}
 
-	// Determine if partial reads are enabled based on maxReadFileLine setting.
-	const partialReadsEnabled = maxReadFileLine !== -1
-
-	// Check if the model supports images for read_file tool description.
-	const supportsImages = modelInfo?.supportsImages ?? false
-
-	// Build native tools with dynamic read_file tool based on settings.
-	const nativeTools = getNativeTools({
-		partialReadsEnabled,
-		maxConcurrentFileReads,
-		supportsImages,
-	})
+	// Build native tools.
+	const nativeTools = getNativeTools()
 
 	// Filter native tools based on mode restrictions.
 	const filteredNativeTools = filterNativeToolsForMode(