Initial commit

Signed-off-by: Ilan Bigio <ilan@openai.com>
This commit is contained in:
Ilan Bigio
2025-04-16 12:56:08 -04:00
commit 59a180ddec
163 changed files with 30587 additions and 0 deletions

View File

@@ -0,0 +1,92 @@
import type { SafetyAssessment } from "./approvals";
import { canAutoApprove } from "./approvals";
import { describe, test, expect } from "vitest";
describe("canAutoApprove()", () => {
const env = {
PATH: "/usr/local/bin:/usr/bin:/bin",
HOME: "/home/user",
};
const writeablePaths: Array<string> = [];
const check = (command: ReadonlyArray<string>): SafetyAssessment =>
canAutoApprove(command, "suggest", writeablePaths, env);
test("simple safe commands", () => {
expect(check(["ls"])).toEqual({
type: "auto-approve",
reason: "List directory",
group: "Searching",
runInSandbox: false,
});
expect(check(["cat", "file.txt"])).toEqual({
type: "auto-approve",
reason: "View file contents",
group: "Reading files",
runInSandbox: false,
});
expect(check(["pwd"])).toEqual({
type: "auto-approve",
reason: "Print working directory",
group: "Navigating",
runInSandbox: false,
});
});
test("simple safe commands within a `bash -lc` call", () => {
expect(check(["bash", "-lc", "ls"])).toEqual({
type: "auto-approve",
reason: "List directory",
group: "Searching",
runInSandbox: false,
});
expect(check(["bash", "-lc", "ls $HOME"])).toEqual({
type: "auto-approve",
reason: "List directory",
group: "Searching",
runInSandbox: false,
});
expect(check(["bash", "-lc", "git show ab9811cb90"])).toEqual({
type: "auto-approve",
reason: "Git show",
group: "Using git",
runInSandbox: false,
});
});
test("bash -lc commands with unsafe redirects", () => {
expect(check(["bash", "-lc", "echo hello > file.txt"])).toEqual({
type: "ask-user",
});
// In theory, we could make our checker more sophisticated to auto-approve
// This previously required approval, but now that we consider safe
// operators like "&&" the entire expression can be autoapproved.
expect(check(["bash", "-lc", "ls && pwd"])).toEqual({
type: "auto-approve",
reason: "List directory",
group: "Searching",
runInSandbox: false,
});
});
test("true command is considered safe", () => {
expect(check(["true"])).toEqual({
type: "auto-approve",
reason: "Noop (true)",
group: "Utility",
runInSandbox: false,
});
});
test("commands that should require approval", () => {
// Should this be on the auto-approved list?
expect(check(["printenv"])).toEqual({ type: "ask-user" });
expect(check(["git", "commit"])).toEqual({ type: "ask-user" });
expect(check(["pytest"])).toEqual({ type: "ask-user" });
expect(check(["cargo", "build"])).toEqual({ type: "ask-user" });
});
});

View File

@@ -0,0 +1,542 @@
import type { ParseEntry, ControlOperator } from "shell-quote";
import {
identify_files_added,
identify_files_needed,
} from "../utils/agent/apply-patch";
import * as path from "path";
import { parse } from "shell-quote";
export type SafetyAssessment = {
/**
* If set, this approval is for an apply_patch call and these are the
* arguments.
*/
applyPatch?: ApplyPatchCommand;
} & (
| {
type: "auto-approve";
/**
* This must be true if the command is not on the "known safe" list, but
* was auto-approved due to `full-auto` mode.
*/
runInSandbox: boolean;
reason: string;
group: string;
}
| {
type: "ask-user";
}
/**
* Reserved for a case where we are certain the command is unsafe and should
* not be presented as an option to the user.
*/
| {
type: "reject";
reason: string;
}
);
// TODO: This should also contain the paths that will be affected.
export type ApplyPatchCommand = {
patch: string;
};
export type ApprovalPolicy =
/**
* Under this policy, only "known safe" commands as defined by
* `isSafeCommand()` that only read files will be auto-approved.
*/
| "suggest"
/**
* In addition to commands that are auto-approved according to the rules for
* "suggest", commands that write files within the user's approved list of
* writable paths will also be auto-approved.
*/
| "auto-edit"
/**
* All commands are auto-approved, but are expected to be run in a sandbox
* where network access is disabled and writes are limited to a specific set
* of paths.
*/
| "full-auto";
/**
* Tries to assess whether a command is safe to run, though may defer to the
* user for approval.
*
* Note `env` must be the same `env` that will be used to spawn the process.
*/
export function canAutoApprove(
command: ReadonlyArray<string>,
policy: ApprovalPolicy,
writableRoots: ReadonlyArray<string>,
env: NodeJS.ProcessEnv = process.env,
): SafetyAssessment {
try {
if (command[0] === "apply_patch") {
return command.length === 2 && typeof command[1] === "string"
? canAutoApproveApplyPatch(command[1], writableRoots, policy)
: {
type: "reject",
reason: "Invalid apply_patch command",
};
}
const isSafe = isSafeCommand(command);
if (isSafe != null) {
const { reason, group } = isSafe;
return {
type: "auto-approve",
reason,
group,
runInSandbox: false,
};
}
if (
command[0] === "bash" &&
command[1] === "-lc" &&
typeof command[2] === "string" &&
command.length === 3
) {
const applyPatchArg = tryParseApplyPatch(command[2]);
if (applyPatchArg != null) {
return canAutoApproveApplyPatch(applyPatchArg, writableRoots, policy);
}
const bashCmd = parse(command[2], env);
// bashCmd could be a mix of strings and operators, e.g.:
// "ls || (true && pwd)" => [ 'ls', { op: '||' }, '(', 'true', { op: '&&' }, 'pwd', ')' ]
// We try to ensure that *every* command segment is deemed safe and that
// all operators belong to an allowlist. If so, the entire expression is
// considered autoapprovable.
const shellSafe = isEntireShellExpressionSafe(bashCmd);
if (shellSafe != null) {
const { reason, group } = shellSafe;
return {
type: "auto-approve",
reason,
group,
runInSandbox: false,
};
}
}
return policy === "full-auto"
? {
type: "auto-approve",
reason: "Full auto mode",
group: "Running commands",
runInSandbox: true,
}
: { type: "ask-user" };
} catch (err) {
if (policy === "full-auto") {
return {
type: "auto-approve",
reason: "Full auto mode",
group: "Running commands",
runInSandbox: true,
};
} else {
return { type: "ask-user" };
}
}
}
function canAutoApproveApplyPatch(
applyPatchArg: string,
writableRoots: ReadonlyArray<string>,
policy: ApprovalPolicy,
): SafetyAssessment {
switch (policy) {
case "full-auto":
// Continue to see if this can be auto-approved.
break;
case "suggest":
return {
type: "ask-user",
applyPatch: { patch: applyPatchArg },
};
case "auto-edit":
// Continue to see if this can be auto-approved.
break;
}
if (isWritePatchConstrainedToWritablePaths(applyPatchArg, writableRoots)) {
return {
type: "auto-approve",
reason: "apply_patch command is constrained to writable paths",
group: "Editing",
runInSandbox: false,
applyPatch: { patch: applyPatchArg },
};
}
return policy === "full-auto"
? {
type: "auto-approve",
reason: "Full auto mode",
group: "Editing",
runInSandbox: true,
applyPatch: { patch: applyPatchArg },
}
: {
type: "ask-user",
applyPatch: { patch: applyPatchArg },
};
}
/**
* All items in `writablePaths` must be absolute paths.
*/
function isWritePatchConstrainedToWritablePaths(
applyPatchArg: string,
writableRoots: ReadonlyArray<string>,
): boolean {
// `identify_files_needed()` returns a list of files that will be modified or
// deleted by the patch, so all of them should already exist on disk. These
// candidate paths could be further canonicalized via fs.realpath(), though
// that does seem necessary and may even cause false negatives (assuming we
// allow writes in other directories that are symlinked from a writable path)
//
// By comparison, `identify_files_added()` returns a list of files that will
// be added by the patch, so they should NOT exist on disk yet and therefore
// using one with fs.realpath() should return an error.
return (
allPathsConstrainedTowritablePaths(
identify_files_needed(applyPatchArg),
writableRoots,
) &&
allPathsConstrainedTowritablePaths(
identify_files_added(applyPatchArg),
writableRoots,
)
);
}
function allPathsConstrainedTowritablePaths(
candidatePaths: ReadonlyArray<string>,
writableRoots: ReadonlyArray<string>,
): boolean {
return candidatePaths.every((candidatePath) =>
isPathConstrainedTowritablePaths(candidatePath, writableRoots),
);
}
/** If candidatePath is relative, it will be resolved against cwd. */
function isPathConstrainedTowritablePaths(
candidatePath: string,
writableRoots: ReadonlyArray<string>,
): boolean {
const candidateAbsolutePath = path.resolve(candidatePath);
return writableRoots.some((writablePath) =>
pathContains(writablePath, candidateAbsolutePath),
);
}
/** Both `parent` and `child` must be absolute paths. */
function pathContains(parent: string, child: string): boolean {
const relative = path.relative(parent, child);
return (
// relative path doesn't go outside parent
!!relative && !relative.startsWith("..") && !path.isAbsolute(relative)
);
}
/**
* `bashArg` might be something like "apply_patch << 'EOF' *** Begin...".
* If this function returns a string, then it is the content the arg to
* apply_patch with the heredoc removed.
*/
function tryParseApplyPatch(bashArg: string): string | null {
const prefix = "apply_patch";
if (!bashArg.startsWith(prefix)) {
return null;
}
const heredoc = bashArg.slice(prefix.length);
const heredocMatch = heredoc.match(
/^\s*<<\s*['"]?(\w+)['"]?\n([\s\S]*?)\n\1/,
);
if (heredocMatch != null && typeof heredocMatch[2] === "string") {
return heredocMatch[2].trim();
} else {
return heredoc.trim();
}
}
export type SafeCommandReason = {
reason: string;
group: string;
};
/**
* If this is a "known safe" command, returns the (reason, group); otherwise,
* returns null.
*/
export function isSafeCommand(
command: ReadonlyArray<string>,
): SafeCommandReason | null {
const [cmd0, cmd1, cmd2, cmd3] = command;
switch (cmd0) {
case "cd":
return {
reason: "Change directory",
group: "Navigating",
};
case "ls":
return {
reason: "List directory",
group: "Searching",
};
case "pwd":
return {
reason: "Print working directory",
group: "Navigating",
};
case "true":
return {
reason: "Noop (true)",
group: "Utility",
};
case "echo":
return { reason: "Echo string", group: "Printing" };
case "cat":
return {
reason: "View file contents",
group: "Reading files",
};
case "rg":
return {
reason: "Ripgrep search",
group: "Searching",
};
case "find":
return {
reason: "Find files or directories",
group: "Searching",
};
case "grep":
return {
reason: "Text search (grep)",
group: "Searching",
};
case "head":
return {
reason: "Show file head",
group: "Reading files",
};
case "tail":
return {
reason: "Show file tail",
group: "Reading files",
};
case "wc":
return {
reason: "Word count",
group: "Reading files",
};
case "which":
return {
reason: "Locate command",
group: "Searching",
};
case "git":
switch (cmd1) {
case "status":
return {
reason: "Git status",
group: "Versioning",
};
case "branch":
return {
reason: "List Git branches",
group: "Versioning",
};
case "log":
return {
reason: "Git log",
group: "Using git",
};
case "diff":
return {
reason: "Git diff",
group: "Using git",
};
case "show":
return {
reason: "Git show",
group: "Using git",
};
default:
return null;
}
case "cargo":
if (cmd1 === "check") {
return {
reason: "Cargo check",
group: "Running command",
};
}
break;
case "sed":
if (
cmd1 === "-n" &&
isValidSedNArg(cmd2) &&
typeof cmd3 === "string" &&
command.length === 4
) {
return {
reason: "Sed print subset",
group: "Reading files",
};
}
break;
case "oai":
switch (cmd1) {
case "show-lines":
return {
reason: "OAI show lines",
group: "Reading files",
};
case "find-files":
return {
reason: "OAI find files",
group: "Searching",
};
case "file-outline":
return {
reason: "OAI file outline",
group: "Reading files",
};
case "rg":
return {
reason: "OAI ripgrep",
group: "Searching",
};
default:
return null;
}
default:
return null;
}
return null;
}
function isValidSedNArg(arg: string | undefined): boolean {
return arg != null && /^(\d+,)?\d+p$/.test(arg);
}
// ---------------- Helper utilities for complex shell expressions -----------------
// A conservative allowlist of bash operators that do not, on their own, cause
// side effects. Redirections (>, >>, <, etc.) and command substitution `$()`
// are intentionally excluded. Parentheses used for grouping are treated as
// strings by `shellquote`, so we do not add them here. Reference:
// https://github.com/substack/node-shell-quote#parsecmd-opts
const SAFE_SHELL_OPERATORS: ReadonlySet<string> = new Set([
"&&", // logical AND
"||", // logical OR
"|", // pipe
";", // command separator
]);
/**
* Determines whether a parsed shell expression consists solely of safe
* commands (as per `isSafeCommand`) combined using only operators in
* `SAFE_SHELL_OPERATORS`.
*
* If entirely safe, returns the reason/group from the *first* command
* segment so callers can surface a meaningful description. Otherwise returns
* null.
*/
function isEntireShellExpressionSafe(
parts: ReadonlyArray<ParseEntry>,
): SafeCommandReason | null {
if (parts.length === 0) {
return null;
}
try {
// Collect command segments delimited by operators. `shellquote` represents
// subshell grouping parentheses as literal strings "(" and ")"; treat them
// as unsafe to keep the logic simple (since subshells could introduce
// unexpected scope changes).
let currentSegment: Array<string> = [];
let firstReason: SafeCommandReason | null = null;
const flushSegment = (): boolean => {
if (currentSegment.length === 0) {
return true; // nothing to validate (possible leading operator)
}
const assessment = isSafeCommand(currentSegment);
if (assessment == null) {
return false;
}
if (firstReason == null) {
firstReason = assessment;
}
currentSegment = [];
return true;
};
for (const part of parts) {
if (typeof part === "string") {
// If this string looks like an open/close parenthesis or brace, treat as
// unsafe to avoid parsing complexity.
if (part === "(" || part === ")" || part === "{" || part === "}") {
return null;
}
currentSegment.push(part);
} else if (isParseEntryWithOp(part)) {
// Validate the segment accumulated so far.
if (!flushSegment()) {
return null;
}
// Validate the operator itself.
if (!SAFE_SHELL_OPERATORS.has(part.op)) {
return null;
}
} else {
// Unknown token type
return null;
}
}
// Validate any trailing command segment.
if (!flushSegment()) {
return null;
}
return firstReason;
} catch (_err) {
// If there's any kind of failure, just bail out and return null.
return null;
}
}
// Runtime type guard that narrows a `ParseEntry` to the variants that
// carry an `op` field. Using a dedicated function avoids the need for
// inline type assertions and makes the narrowing reusable and explicit.
function isParseEntryWithOp(
entry: ParseEntry,
): entry is { op: ControlOperator } | { op: "glob"; pattern: string } {
return (
typeof entry === "object" &&
entry != null &&
// Using the safe `in` operator keeps the check propertysafe even when
// `entry` is a `string`.
"op" in entry &&
typeof (entry as { op?: unknown }).op === "string"
);
}

View File

@@ -0,0 +1,21 @@
import { formatCommandForDisplay } from "./format-command";
import { describe, test, expect } from "vitest";
describe("formatCommandForDisplay()", () => {
test("ensure empty string arg appears in output", () => {
expect(formatCommandForDisplay(["echo", ""])).toEqual("echo ''");
});
test("ensure special characters are properly escaped", () => {
expect(formatCommandForDisplay(["echo", "$HOME"])).toEqual("echo \\$HOME");
});
test("ensure quotes are properly escaped", () => {
expect(formatCommandForDisplay(["echo", "I can't believe this."])).toEqual(
'echo "I can\'t believe this."',
);
expect(
formatCommandForDisplay(["echo", 'So I said, "No ma\'am!"']),
).toEqual('echo "So I said, \\"No ma\'am\\!\\""');
});
});

View File

@@ -0,0 +1,53 @@
import { quote } from "shell-quote";
/**
* Format the args of an exec command for display as a single string. Prefer
* this to doing `args.join(" ")` as this will handle quoting and escaping
* correctly. See unit test for details.
*/
export function formatCommandForDisplay(command: Array<string>): string {
// The model often wraps arbitrary shell commands in an invocation that looks
// like:
//
// ["bash", "-lc", "'<actual command>'"]
//
// When displaying these back to the user, we do NOT want to show the
// boilerplate "bash -lc" wrapper. Instead, we want to surface only the
// actual command that bash will evaluate.
// Historically we detected this by first quoting the entire command array
// with `shellquote` and then using a regular expression to peel off the
// `bash -lc '…'` prefix. However, that approach was brittle (it depended on
// the exact quoting behavior of `shell-quote`) and unnecessarily
// inefficient.
// A simpler and more robust approach is to look at the raw command array
// itself. If it matches the shape produced by our exec helpers—exactly three
// entries where the first two are «bash» and «-lc»—then we can return the
// third entry directly (after stripping surrounding single quotes if they
// are present).
try {
if (
command.length === 3 &&
command[0] === "bash" &&
command[1] === "-lc" &&
typeof command[2] === "string"
) {
let inner = command[2];
// Some callers wrap the actual command in single quotes (e.g. `'echo foo'`).
// For display purposes we want to drop those outer quotes so that the
// rendered command looks exactly like what the user typed.
if (inner.startsWith("'") && inner.endsWith("'")) {
inner = inner.slice(1, -1);
}
return inner;
}
return quote(command);
} catch (err) {
return command.join(" ");
}
}

View File

@@ -0,0 +1,45 @@
import { parseApplyPatch } from "./parse-apply-patch";
import { expect, test, describe } from "vitest";
// Helper function to unwrap a nonnull result in tests that expect success.
function mustParse(patch: string) {
const parsed = parseApplyPatch(patch);
if (parsed == null) {
throw new Error(
"Expected patch to be valid, but parseApplyPatch returned null",
);
}
return parsed;
}
describe("parseApplyPatch", () => {
test("parses create, update and delete operations in a single patch", () => {
const patch = `*** Begin Patch\n*** Add File: created.txt\n+hello\n+world\n*** Update File: updated.txt\n@@\n-old\n+new\n*** Delete File: removed.txt\n*** End Patch`;
const ops = mustParse(patch);
expect(ops).toEqual([
{
type: "create",
path: "created.txt",
content: "hello\nworld",
},
{
type: "update",
path: "updated.txt",
update: "@@\n-old\n+new",
added: 1,
deleted: 1,
},
{
type: "delete",
path: "removed.txt",
},
]);
});
test("returns null for an invalid patch (missing prefix)", () => {
const invalid = `*** Add File: foo.txt\n+bar\n*** End Patch`;
expect(parseApplyPatch(invalid)).toBeNull();
});
});

View File

@@ -0,0 +1,112 @@
export type ApplyPatchCreateFileOp = {
type: "create";
path: string;
content: string;
};
export type ApplyPatchDeleteFileOp = {
type: "delete";
path: string;
};
export type ApplyPatchUpdateFileOp = {
type: "update";
path: string;
update: string;
added: number;
deleted: number;
};
export type ApplyPatchOp =
| ApplyPatchCreateFileOp
| ApplyPatchDeleteFileOp
| ApplyPatchUpdateFileOp;
const PATCH_PREFIX = "*** Begin Patch\n";
const PATCH_SUFFIX = "\n*** End Patch";
const ADD_FILE_PREFIX = "*** Add File: ";
const DELETE_FILE_PREFIX = "*** Delete File: ";
const UPDATE_FILE_PREFIX = "*** Update File: ";
const END_OF_FILE_PREFIX = "*** End of File";
const HUNK_ADD_LINE_PREFIX = "+";
/**
* @returns null when the patch is invalid
*/
export function parseApplyPatch(patch: string): Array<ApplyPatchOp> | null {
if (!patch.startsWith(PATCH_PREFIX)) {
// Patch must begin with '*** Begin Patch'
return null;
} else if (!patch.endsWith(PATCH_SUFFIX)) {
// Patch must end with '*** End Patch'
return null;
}
const patchBody = patch.slice(
PATCH_PREFIX.length,
patch.length - PATCH_SUFFIX.length,
);
const lines = patchBody.split("\n");
const ops: Array<ApplyPatchOp> = [];
for (const line of lines) {
if (line.startsWith(END_OF_FILE_PREFIX)) {
continue;
} else if (line.startsWith(ADD_FILE_PREFIX)) {
ops.push({
type: "create",
path: line.slice(ADD_FILE_PREFIX.length).trim(),
content: "",
});
continue;
} else if (line.startsWith(DELETE_FILE_PREFIX)) {
ops.push({
type: "delete",
path: line.slice(DELETE_FILE_PREFIX.length).trim(),
});
continue;
} else if (line.startsWith(UPDATE_FILE_PREFIX)) {
ops.push({
type: "update",
path: line.slice(UPDATE_FILE_PREFIX.length).trim(),
update: "",
added: 0,
deleted: 0,
});
continue;
}
const lastOp = ops[ops.length - 1];
if (lastOp?.type === "create") {
lastOp.content = appendLine(
lastOp.content,
line.slice(HUNK_ADD_LINE_PREFIX.length),
);
continue;
}
if (lastOp?.type !== "update") {
// Expected update op but got ${lastOp?.type} for line ${line}
return null;
}
if (line.startsWith(HUNK_ADD_LINE_PREFIX)) {
lastOp.added += 1;
} else if (line.startsWith("-")) {
lastOp.deleted += 1;
}
lastOp.update += lastOp.update ? "\n" + line : line;
}
return ops;
}
function appendLine(content: string, line: string) {
if (!content.length) {
return line;
}
return [content, line].join("\n");
}

View File

@@ -0,0 +1,852 @@
/* eslintdisable no-bitwise */
export type Direction =
| "left"
| "right"
| "up"
| "down"
| "wordLeft"
| "wordRight"
| "home"
| "end";
// Simple helper for wordwise ops.
function isWordChar(ch: string | undefined): boolean {
if (ch === undefined) {
return false;
}
return !/[\s,.;!?]/.test(ch);
}
export interface Viewport {
height: number;
width: number;
}
function clamp(v: number, min: number, max: number): number {
return v < min ? min : v > max ? max : v;
}
/*
* -------------------------------------------------------------------------
* Unicodeaware helpers (work at the codepoint level rather than UTF16
* code units so that surrogatepair emoji count as one "column".)
* ---------------------------------------------------------------------- */
function toCodePoints(str: string): Array<string> {
// [...str] or Array.from both iterate by UTF32 code point, handling
// surrogate pairs correctly.
return Array.from(str);
}
function cpLen(str: string): number {
return toCodePoints(str).length;
}
function cpSlice(str: string, start: number, end?: number): string {
// Slice by codepoint indices and rejoin.
const arr = toCodePoints(str).slice(start, end);
return arr.join("");
}
/* -------------------------------------------------------------------------
* Debug helper enable verbose logging by setting env var TEXTBUFFER_DEBUG=1
* ---------------------------------------------------------------------- */
// Enable verbose logging only when requested via env var.
const DEBUG =
process.env["TEXTBUFFER_DEBUG"] === "1" ||
process.env["TEXTBUFFER_DEBUG"] === "true";
function dbg(...args: Array<unknown>): void {
if (DEBUG) {
// eslint-disable-next-line no-console
console.log("[TextBuffer]", ...args);
}
}
/* ────────────────────────────────────────────────────────────────────────── */
export default class TextBuffer {
private lines: Array<string>;
private cursorRow = 0;
private cursorCol = 0;
private scrollRow = 0;
private scrollCol = 0;
/**
* When the user moves the caret vertically we try to keep their original
* horizontal column even when passing through shorter lines. We remember
* that *preferred* column in this field while the user is still travelling
* vertically. Any explicit horizontal movement resets the preference.
*/
private preferredCol: number | null = null;
/* a single integer that bumps every time text changes */
private version = 0;
/* ------------------------------------------------------------------
* History & clipboard
* ---------------------------------------------------------------- */
private undoStack: Array<{ lines: Array<string>; row: number; col: number }> =
[];
private redoStack: Array<{ lines: Array<string>; row: number; col: number }> =
[];
private historyLimit = 100;
private clipboard: string | null = null;
constructor(text = "") {
this.lines = text.split("\n");
if (this.lines.length === 0) {
this.lines = [""];
}
}
/* =====================================================================
* External editor integration (gitstyle $EDITOR workflow)
* =================================================================== */
/**
* Opens the current buffer contents in the users preferred terminal text
* editor ($VISUAL or $EDITOR, falling back to "vi"). The method blocks
* until the editor exits, then reloads the file and replaces the inmemory
* buffer with whatever the user saved.
*
* The operation is treated as a single undoable edit we snapshot the
* previous state *once* before launching the editor so one `undo()` will
* revert the entire change set.
*
* Note: We purposefully rely on the *synchronous* spawn API so that the
* calling process genuinely waits for the editor to close before
* continuing. This mirrors Gits behaviour and simplifies downstream
* controlflow (callers can simply `await` the Promise).
*/
async openInExternalEditor(opts: { editor?: string } = {}): Promise<void> {
// Deliberately use `require()` so that unit tests can stub the
// respective modules with `vi.spyOn(require("node:child_process"), …)`.
// Dynamic `import()` would circumvent those CommonJS stubs.
// eslint-disable-next-line @typescript-eslint/no-var-requires
const pathMod = require("node:path");
// eslint-disable-next-line @typescript-eslint/no-var-requires
const fs = require("node:fs");
// eslint-disable-next-line @typescript-eslint/no-var-requires
const os = require("node:os");
// eslint-disable-next-line @typescript-eslint/no-var-requires
const { spawnSync } = require("node:child_process");
const editor =
opts.editor ??
process.env["VISUAL"] ??
process.env["EDITOR"] ??
(process.platform === "win32" ? "notepad" : "vi");
// Prepare a temporary file with the current contents. We use mkdtempSync
// to obtain an isolated directory and avoid name collisions.
const tmpDir = fs.mkdtempSync(pathMod.join(os.tmpdir(), "codex-edit-"));
const filePath = pathMod.join(tmpDir, "buffer.txt");
fs.writeFileSync(filePath, this.getText(), "utf8");
// One snapshot for undo semantics *before* we mutate anything.
this.pushUndo();
// The child inherits stdio so the user can interact with the editor as if
// they had launched it directly.
const { status, error } = spawnSync(editor, [filePath], {
stdio: "inherit",
});
if (error) {
throw error;
}
if (typeof status === "number" && status !== 0) {
throw new Error(`External editor exited with status ${status}`);
}
// Read the edited contents back in normalise line endings to \n.
let newText = fs.readFileSync(filePath, "utf8");
newText = newText.replace(/\r\n?/g, "\n");
// Update buffer.
this.lines = newText.split("\n");
if (this.lines.length === 0) {
this.lines = [""];
}
// Position the caret at EOF.
this.cursorRow = this.lines.length - 1;
this.cursorCol = cpLen(this.line(this.cursorRow));
// Reset scroll offsets so the new end is visible.
this.scrollRow = Math.max(0, this.cursorRow - 1);
this.scrollCol = 0;
this.version++;
}
/* =======================================================================
* Geometry helpers
* ===================================================================== */
private line(r: number): string {
return this.lines[r] ?? "";
}
private lineLen(r: number): number {
return cpLen(this.line(r));
}
private ensureCursorInRange(): void {
this.cursorRow = clamp(this.cursorRow, 0, this.lines.length - 1);
this.cursorCol = clamp(this.cursorCol, 0, this.lineLen(this.cursorRow));
}
/* =====================================================================
* History helpers
* =================================================================== */
private snapshot() {
return {
lines: this.lines.slice(),
row: this.cursorRow,
col: this.cursorCol,
};
}
private pushUndo() {
dbg("pushUndo", { cursor: this.getCursor(), text: this.getText() });
this.undoStack.push(this.snapshot());
if (this.undoStack.length > this.historyLimit) {
this.undoStack.shift();
}
// once we mutate we clear redo
this.redoStack.length = 0;
}
/**
* Restore a snapshot and return true if restoration happened.
*/
private restore(
state: { lines: Array<string>; row: number; col: number } | undefined,
): boolean {
if (!state) {
return false;
}
this.lines = state.lines.slice();
this.cursorRow = state.row;
this.cursorCol = state.col;
this.ensureCursorInRange();
return true;
}
/* =======================================================================
* Scrolling helpers
* ===================================================================== */
private ensureCursorVisible(vp: Viewport) {
const { height, width } = vp;
if (this.cursorRow < this.scrollRow) {
this.scrollRow = this.cursorRow;
} else if (this.cursorRow >= this.scrollRow + height) {
this.scrollRow = this.cursorRow - height + 1;
}
if (this.cursorCol < this.scrollCol) {
this.scrollCol = this.cursorCol;
} else if (this.cursorCol >= this.scrollCol + width) {
this.scrollCol = this.cursorCol - width + 1;
}
}
/* =======================================================================
* Public readonly accessors
* ===================================================================== */
getVersion(): number {
return this.version;
}
getCursor(): [number, number] {
return [this.cursorRow, this.cursorCol];
}
getVisibleLines(vp: Viewport): Array<string> {
// Whenever the viewport dimensions change (e.g. on a terminal resize) we
// need to reevaluate whether the current scroll offset still keeps the
// caret visible. Calling `ensureCursorVisible` here guarantees that mere
// rerenders even when not triggered by user input will adjust the
// horizontal and vertical scroll positions so the cursor remains in view.
this.ensureCursorVisible(vp);
return this.lines.slice(this.scrollRow, this.scrollRow + vp.height);
}
getText(): string {
return this.lines.join("\n");
}
getLines(): Array<string> {
return this.lines.slice();
}
/* =====================================================================
* History public API undo / redo
* =================================================================== */
undo(): boolean {
const state = this.undoStack.pop();
if (!state) {
return false;
}
// push current to redo before restore
this.redoStack.push(this.snapshot());
this.restore(state);
this.version++;
return true;
}
redo(): boolean {
const state = this.redoStack.pop();
if (!state) {
return false;
}
// push current to undo before restore
this.undoStack.push(this.snapshot());
this.restore(state);
this.version++;
return true;
}
/* =======================================================================
* Editing operations
* ===================================================================== */
/**
* Insert a single character or string without newlines. If the string
* contains a newline we delegate to insertStr so that line splitting
* logic is shared.
*/
insert(ch: string): void {
// Handle pasted blocks that may contain newline sequences (\n, \r or
// Windowsstyle \r\n). Delegate to `insertStr` so the splitting logic is
// centralised.
if (/[\n\r]/.test(ch)) {
this.insertStr(ch);
return;
}
dbg("insert", { ch, beforeCursor: this.getCursor() });
this.pushUndo();
const line = this.line(this.cursorRow);
this.lines[this.cursorRow] =
cpSlice(line, 0, this.cursorCol) + ch + cpSlice(line, this.cursorCol);
this.cursorCol += ch.length;
this.version++;
dbg("insert:after", {
cursor: this.getCursor(),
line: this.line(this.cursorRow),
});
}
newline(): void {
dbg("newline", { beforeCursor: this.getCursor() });
this.pushUndo();
const l = this.line(this.cursorRow);
const before = cpSlice(l, 0, this.cursorCol);
const after = cpSlice(l, this.cursorCol);
this.lines[this.cursorRow] = before;
this.lines.splice(this.cursorRow + 1, 0, after);
this.cursorRow += 1;
this.cursorCol = 0;
this.version++;
dbg("newline:after", {
cursor: this.getCursor(),
lines: [this.line(this.cursorRow - 1), this.line(this.cursorRow)],
});
}
backspace(): void {
dbg("backspace", { beforeCursor: this.getCursor() });
if (this.cursorCol === 0 && this.cursorRow === 0) {
return;
} // nothing to delete
this.pushUndo();
if (this.cursorCol > 0) {
const line = this.line(this.cursorRow);
this.lines[this.cursorRow] =
cpSlice(line, 0, this.cursorCol - 1) + cpSlice(line, this.cursorCol);
this.cursorCol--;
} else if (this.cursorRow > 0) {
// merge with previous
const prev = this.line(this.cursorRow - 1);
const cur = this.line(this.cursorRow);
const newCol = cpLen(prev);
this.lines[this.cursorRow - 1] = prev + cur;
this.lines.splice(this.cursorRow, 1);
this.cursorRow--;
this.cursorCol = newCol;
}
this.version++;
dbg("backspace:after", {
cursor: this.getCursor(),
line: this.line(this.cursorRow),
});
}
del(): void {
dbg("delete", { beforeCursor: this.getCursor() });
const line = this.line(this.cursorRow);
if (this.cursorCol < this.lineLen(this.cursorRow)) {
this.pushUndo();
this.lines[this.cursorRow] =
cpSlice(line, 0, this.cursorCol) + cpSlice(line, this.cursorCol + 1);
} else if (this.cursorRow < this.lines.length - 1) {
this.pushUndo();
const next = this.line(this.cursorRow + 1);
this.lines[this.cursorRow] = line + next;
this.lines.splice(this.cursorRow + 1, 1);
}
this.version++;
dbg("delete:after", {
cursor: this.getCursor(),
line: this.line(this.cursorRow),
});
}
/* ------------------------------------------------------------------
* Wordwise deletion helpers exposed publicly so tests (and future
* keybindings) can invoke them directly.
* ---------------------------------------------------------------- */
/** Delete the word to the *left* of the caret, mirroring common
* Ctrl/Alt+Backspace behaviour in editors & terminals. Both the adjacent
* whitespace *and* the word characters immediately preceding the caret are
* removed. If the caret is already at column0 this becomes a noop. */
deleteWordLeft(): void {
dbg("deleteWordLeft", { beforeCursor: this.getCursor() });
if (this.cursorCol === 0 && this.cursorRow === 0) {
return;
} // Nothing to delete
// When at column0 but *not* on the first row we merge with the previous
// line matching the behaviour of `backspace` for uniform UX.
if (this.cursorCol === 0) {
this.backspace();
return;
}
this.pushUndo();
const line = this.line(this.cursorRow);
const arr = toCodePoints(line);
// Step 1 skip over any separators sitting *immediately* to the left of
// the caret so that consecutive deletions wipe runs of whitespace first
// then words.
let start = this.cursorCol;
while (start > 0 && !isWordChar(arr[start - 1])) {
start--;
}
// Step 2 now skip the word characters themselves.
while (start > 0 && isWordChar(arr[start - 1])) {
start--;
}
this.lines[this.cursorRow] =
cpSlice(line, 0, start) + cpSlice(line, this.cursorCol);
this.cursorCol = start;
this.version++;
dbg("deleteWordLeft:after", {
cursor: this.getCursor(),
line: this.line(this.cursorRow),
});
}
/** Delete the word to the *right* of the caret, akin to many editors'
* Ctrl/Alt+Delete shortcut. Removes any whitespace/punctuation that
* follows the caret and the next contiguous run of word characters. */
deleteWordRight(): void {
dbg("deleteWordRight", { beforeCursor: this.getCursor() });
const line = this.line(this.cursorRow);
const arr = toCodePoints(line);
if (
this.cursorCol >= arr.length &&
this.cursorRow === this.lines.length - 1
) {
return;
} // nothing to delete
// At endofline ➜ merge with next row (mirrors `del` behaviour).
if (this.cursorCol >= arr.length) {
this.del();
return;
}
this.pushUndo();
let end = this.cursorCol;
// Skip separators *first* so that consecutive calls gradually chew
// through whitespace then whole words.
while (end < arr.length && !isWordChar(arr[end])) {
end++;
}
// Skip the word characters.
while (end < arr.length && isWordChar(arr[end])) {
end++;
}
this.lines[this.cursorRow] =
cpSlice(line, 0, this.cursorCol) + cpSlice(line, end);
// caret stays in place
this.version++;
dbg("deleteWordRight:after", {
cursor: this.getCursor(),
line: this.line(this.cursorRow),
});
}
move(dir: Direction): void {
const before = this.getCursor();
switch (dir) {
case "left":
this.preferredCol = null;
if (this.cursorCol > 0) {
this.cursorCol--;
} else if (this.cursorRow > 0) {
this.cursorRow--;
this.cursorCol = this.lineLen(this.cursorRow);
}
break;
case "right":
this.preferredCol = null;
if (this.cursorCol < this.lineLen(this.cursorRow)) {
this.cursorCol++;
} else if (this.cursorRow < this.lines.length - 1) {
this.cursorRow++;
this.cursorCol = 0;
}
break;
case "up":
if (this.cursorRow > 0) {
if (this.preferredCol == null) {
this.preferredCol = this.cursorCol;
}
this.cursorRow--;
this.cursorCol = clamp(
this.preferredCol,
0,
this.lineLen(this.cursorRow),
);
}
break;
case "down":
if (this.cursorRow < this.lines.length - 1) {
if (this.preferredCol == null) {
this.preferredCol = this.cursorCol;
}
this.cursorRow++;
this.cursorCol = clamp(
this.preferredCol,
0,
this.lineLen(this.cursorRow),
);
}
break;
case "home":
this.preferredCol = null;
this.cursorCol = 0;
break;
case "end":
this.preferredCol = null;
this.cursorCol = this.lineLen(this.cursorRow);
break;
case "wordLeft": {
this.preferredCol = null;
const regex = /[\s,.;!?]+/g;
const slice = cpSlice(
this.line(this.cursorRow),
0,
this.cursorCol,
).replace(/[\s,.;!?]+$/, "");
let lastIdx = 0;
let m;
while ((m = regex.exec(slice)) != null) {
lastIdx = m.index;
}
const last = cpLen(slice.slice(0, lastIdx));
this.cursorCol = last === 0 ? 0 : last + 1;
break;
}
case "wordRight": {
this.preferredCol = null;
const regex = /[\s,.;!?]+/g;
const l = this.line(this.cursorRow);
let moved = false;
let m;
while ((m = regex.exec(l)) != null) {
const cpIdx = cpLen(l.slice(0, m.index));
if (cpIdx > this.cursorCol) {
// We want to land *at the beginning* of the separator run so that a
// subsequent move("right") behaves naturally.
this.cursorCol = cpIdx;
moved = true;
break;
}
}
if (!moved) {
// No boundary to the right jump to EOL.
this.cursorCol = this.lineLen(this.cursorRow);
}
break;
}
}
if (DEBUG) {
dbg("move", { dir, before, after: this.getCursor() });
}
/*
* If the user performed any movement other than a consecutive vertical
* traversal we clear the preferred column so the next vertical run starts
* afresh. The cases that keep the preference already returned earlier.
*/
if (dir !== "up" && dir !== "down") {
this.preferredCol = null;
}
}
/* =====================================================================
* Higherlevel helpers
* =================================================================== */
/**
* Insert an arbitrary string, possibly containing internal newlines.
* Returns true if the buffer was modified.
*/
insertStr(str: string): boolean {
dbg("insertStr", { str, beforeCursor: this.getCursor() });
if (str === "") {
return false;
}
// Normalise all newline conventions (\r, \n, \r\n) to a single '\n'.
const normalised = str.replace(/\r\n/g, "\n").replace(/\r/g, "\n");
// Fast path: resulted in singleline string ➜ delegate back to insert
if (!normalised.includes("\n")) {
this.insert(normalised);
return true;
}
this.pushUndo();
const parts = normalised.split("\n");
const before = cpSlice(this.line(this.cursorRow), 0, this.cursorCol);
const after = cpSlice(this.line(this.cursorRow), this.cursorCol);
// Replace current line with first part combined with before text
this.lines[this.cursorRow] = before + parts[0];
// Middle lines (if any) are inserted verbatim after current row
if (parts.length > 2) {
const middle = parts.slice(1, -1);
this.lines.splice(this.cursorRow + 1, 0, ...middle);
}
// Smart handling of the *final* inserted part:
// • When the caret is midline we preserve existing behaviour merge
// the last part with the text to the **right** of the caret so that
// inserting in the middle of a line keeps the remainder on the same
// row (e.g. "he|llo" → paste "x\ny" ⇒ "he x", "y llo").
// • When the caret is at column0 we instead treat the current line as
// a *separate* row that follows the inserted block. This mirrors
// common editor behaviour and avoids the unintuitive merge that led
// to "cd"+"ef" → "cdef" in the failing tests.
// Append the last part combined with original after text as a new line
const last = parts[parts.length - 1] + after;
this.lines.splice(this.cursorRow + (parts.length - 1), 0, last);
// Update cursor position to end of last inserted part (before 'after')
this.cursorRow += parts.length - 1;
// `parts` is guaranteed to have at least one element here because
// `split("\n")` always returns an array with ≥1 entry. Tell the
// compiler so we can pass a plain `string` to `cpLen`.
this.cursorCol = cpLen(parts[parts.length - 1]!);
this.version++;
return true;
}
/* =====================================================================
* Selection & clipboard helpers (minimal)
* =================================================================== */
private selectionAnchor: [number, number] | null = null;
startSelection(): void {
this.selectionAnchor = [this.cursorRow, this.cursorCol];
}
endSelection(): void {
// noop for now, kept for API symmetry
// we rely on anchor + current cursor to compute selection
}
/** Extract selected text. Returns null if no valid selection. */
private getSelectedText(): string | null {
if (!this.selectionAnchor) {
return null;
}
const [ar, ac] = this.selectionAnchor;
const [br, bc] = [this.cursorRow, this.cursorCol];
// Determine ordering
if (ar === br && ac === bc) {
return null;
} // empty selection
const topBefore = ar < br || (ar === br && ac < bc);
const [sr, sc, er, ec] = topBefore ? [ar, ac, br, bc] : [br, bc, ar, ac];
if (sr === er) {
return cpSlice(this.line(sr), sc, ec);
}
const parts: Array<string> = [];
parts.push(cpSlice(this.line(sr), sc));
for (let r = sr + 1; r < er; r++) {
parts.push(this.line(r));
}
parts.push(cpSlice(this.line(er), 0, ec));
return parts.join("\n");
}
copy(): string | null {
const txt = this.getSelectedText();
if (txt == null) {
return null;
}
this.clipboard = txt;
return txt;
}
paste(): boolean {
if (this.clipboard == null) {
return false;
}
return this.insertStr(this.clipboard);
}
/* =======================================================================
* High level "handleInput" receives what Ink gives us
* Returns true when buffer mutated (=> rerender)
* ===================================================================== */
handleInput(
input: string | undefined,
key: Record<string, boolean>,
vp: Viewport,
): boolean {
if (DEBUG) {
dbg("handleInput", { input, key, cursor: this.getCursor() });
}
const beforeVer = this.version;
const [beforeRow, beforeCol] = this.getCursor();
if (key["escape"]) {
return false;
}
/* new line — Ink sets either `key.return` *or* passes a literal "\n" */
if (key["return"] || input === "\r" || input === "\n") {
this.newline();
} else if (
key["leftArrow"] &&
!key["meta"] &&
!key["ctrl"] &&
!key["alt"]
) {
/* navigation */
this.move("left");
} else if (
key["rightArrow"] &&
!key["meta"] &&
!key["ctrl"] &&
!key["alt"]
) {
this.move("right");
} else if (key["upArrow"]) {
this.move("up");
} else if (key["downArrow"]) {
this.move("down");
} else if ((key["meta"] || key["ctrl"] || key["alt"]) && key["leftArrow"]) {
this.move("wordLeft");
} else if (
(key["meta"] || key["ctrl"] || key["alt"]) &&
key["rightArrow"]
) {
this.move("wordRight");
} else if (key["home"]) {
this.move("home");
} else if (key["end"]) {
this.move("end");
}
/* delete */
// In raw terminal mode many frameworks (Ink included) surface a physical
// Backspace keypress as the single DEL (0x7f) byte placed in `input` with
// no `key.backspace` flag set. Treat that byte exactly like an ordinary
// Backspace for parity with textarea.rs and to make interactive tests
// feedable through the simpler `(ch, {}, vp)` path.
else if (
(key["meta"] || key["ctrl"] || key["alt"]) &&
(key["backspace"] || input === "\x7f")
) {
this.deleteWordLeft();
} else if ((key["meta"] || key["ctrl"] || key["alt"]) && key["delete"]) {
this.deleteWordRight();
} else if (
key["backspace"] ||
input === "\x7f" ||
(key["delete"] && !key["shift"])
) {
// Treat unmodified "delete" (the common Mac backspace key) as a
// standard backspace. Holding Shift+Delete continues to perform a
// forward deletion so we don't lose that capability on keyboards that
// expose both behaviours.
this.backspace();
}
// Forward deletion (Fn+Delete on macOS, or Delete key with Shift held after
// the branch above) remove the character *under / to the right* of the
// caret, merging lines when at EOL similar to many editors.
else if (key["delete"]) {
this.del();
} else if (input && !key["ctrl"] && !key["meta"]) {
this.insert(input);
}
/* printable */
/* clamp + scroll */
this.ensureCursorInRange();
this.ensureCursorVisible(vp);
const cursorMoved =
this.cursorRow !== beforeRow || this.cursorCol !== beforeCol;
if (DEBUG) {
dbg("handleInput:after", {
cursor: this.getCursor(),
text: this.getText(),
});
}
return this.version !== beforeVer || cursorMoved;
}
}