fix(codemode): stop leaking undefined into tool arguments (#37652)

This commit is contained in:
Aiden Cline 2026-07-18 13:42:16 -05:00 committed by GitHub
commit 33f1b269e9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 97 additions and 20 deletions

View file

@ -19,7 +19,7 @@ ultimate source of truth.
- [x] Top-level `await` and `return` through the program's implicit async-function scope.
- [x] Explicit `return`, final top-level expression as a REPL-style result, and `null` when no value is produced.
- [x] Program results use JSON-like boundaries, with `undefined` and non-finite numbers normalized to `null`. Tool
arguments remain subject to their schema and the outbound-handling gap listed below.
arguments follow JSON serialization semantics before their schema applies (see the tools section).
- [x] Live Date, RegExp, Map, Set, URL, and URLSearchParams values inside CodeMode.
- [x] Tool calls through the host-provided `tools` tree only.
- [x] The global `search(...)` built-in: synchronous tool discovery that counts as an admitted tool call and is
@ -157,8 +157,11 @@ ultimate source of truth.
- [x] Dotted tool names are canonicalized into namespace paths; a path can be both callable and a namespace, and the
last definition supplied for a canonical path wins.
- [x] Tool path segments may be named `constructor`, `prototype`, or `__proto__` because paths use inert Map keys.
- [ ] Reject `undefined` and non-finite numbers in outbound tool arguments before render-only and OpenAPI tools run;
retain null normalization for program results and JSON serialization.
- [x] Outbound tool arguments follow JSON serialization semantics, like `JSON.stringify`: object properties with
`undefined` values are dropped, `undefined` array elements and non-finite numbers become `null`, and sparse
arrays densify. Tools never receive `undefined` inside their input object, though a bare `tools.t(undefined)`
argument still reaches schema decoding as `undefined`. Program results keep the stricter
normalization where every `undefined` becomes `null`.
- [ ] Tokenize and case-fold non-ASCII tool paths, descriptions, and queries for tool search.
## Objects and properties

View file

@ -44,7 +44,7 @@ export const normalizeError = (error: unknown): Diagnostic => {
message = (value as { message: string }).message
} else {
try {
message = JSON.stringify(copyOut(value)) ?? String(value)
message = JSON.stringify(copyOut(value, "json")) ?? String(value)
} catch {
message = String(value)
}

View file

@ -52,7 +52,7 @@ export const executeWithLimits = <const Provided extends Record<string, unknown>
logs,
)
const value = yield* interpreter.run(program)
const result = copyOut(copyIn(value, "Execution result"), true) as DataValue
const result = copyOut(copyIn(value, "Execution result"), "nullify") as DataValue
returned = { value: result, promises }
const warnings = yield* promises.interrupt()
return {

View file

@ -89,7 +89,7 @@ const formatConsoleTable = (value: unknown, columnsArgument: unknown): string =>
const consoleTableColumns = (value: unknown): ReadonlyArray<string> | undefined => {
if (value === undefined) return undefined
if (containsRuntimeReference(value)) return undefined
const columns = copyOut(copyIn(value, "console.table columns"), true)
const columns = copyOut(copyIn(value, "console.table columns"), "nullify")
return Array.isArray(columns) ? columns.map((column) => String(column)) : undefined
}

View file

@ -18,7 +18,7 @@ export const invokeJsonMethod = (name: string, args: Array<unknown>, node: AstNo
}
const space = args[2]
const indent = typeof space === "number" || typeof space === "string" ? space : undefined
return JSON.stringify(copyOut(copyIn(args[0], "JSON.stringify value")), null, indent)
return JSON.stringify(copyOut(copyIn(args[0], "JSON.stringify value"), "json"), null, indent)
}
case "parse": {
const text = args[0]

View file

@ -118,8 +118,7 @@ export class ToolRuntimeError extends Error {
}
}
const isDefinition = <R>(value: Definition<R> | Tools<R>): value is Definition<R> =>
isToolDefinition<R>(value)
const isDefinition = <R>(value: Definition<R> | Tools<R>): value is Definition<R> => isToolDefinition<R>(value)
const runHost = <A, E, R>(effect: Effect.Effect<A, E, R>): Effect.Effect<A, ToolError, R> =>
effect.pipe(
@ -257,18 +256,31 @@ const copyBounded = (
return copied
}
export const copyOut = (value: unknown, undefinedAsNull = false): unknown => {
if (value === undefined && undefinedAsNull) return null
// "json" mirrors JSON.stringify (undefined object values drop, undefined array elements become
// null, a bare undefined passes through): use it wherever data leaves as JSON, like tool
// arguments and stringify-style formatting. "nullify" turns every undefined, including a bare
// one, into null: use it for program results, where the consumer must never see undefined.
export type CopyOutMode = "json" | "nullify"
export const copyOut = (value: unknown, mode: CopyOutMode): unknown => {
if (value === undefined && mode === "nullify") return null
if (typeof value === "number" && !Number.isFinite(value)) {
return null
}
if (Array.isArray(value)) {
// Array.from densifies holes so sparse arrays normalize at the boundary like JSON does.
return Array.from(value, (item) => copyOut(item, undefinedAsNull))
return Array.from(value, (item) => {
const copied = copyOut(item, mode)
return copied === undefined && mode === "json" ? null : copied
})
}
if (value !== null && typeof value === "object" && !(value instanceof ToolReference)) {
return Object.fromEntries(Object.entries(value).map(([key, item]) => [key, copyOut(item, undefinedAsNull)]))
return Object.fromEntries(
Object.entries(value)
.map(([key, item]) => [key, copyOut(item, mode)] as const)
.filter(([, item]) => !(item === undefined && mode === "json")),
)
}
return value
@ -696,13 +708,13 @@ export const make = <R>(
invokeDefinition(
"search",
searchTool,
args.map((arg) => copyOut(copyIn(arg, "Arguments for tool 'search'"))),
args.map((arg) => copyOut(copyIn(arg, "Arguments for tool 'search'"), "json")),
),
),
invoke: (path, args) =>
Effect.gen(function* () {
const name = canonicalSegments(path).join(".")
const externalArgs = args.map((arg) => copyOut(copyIn(arg, `Arguments for tool '${name}'`)))
const externalArgs = args.map((arg) => copyOut(copyIn(arg, `Arguments for tool '${name}'`), "json"))
const tool = resolve(root, path)
return yield* invokeDefinition(name, tool, externalArgs)
}),

View file

@ -453,6 +453,56 @@ describe("CodeMode schema flexibility", () => {
expect(observed).toStrictEqual([{ id: 42 }])
})
test("outbound tool arguments follow JSON serialization semantics", async () => {
const observed: Array<unknown> = []
const call = Tool.make({
description: "Observe raw input",
input: { type: "object" },
run: (input) =>
Effect.sync(() => {
observed.push(input)
return "ok"
}),
})
const runtime = CodeMode.make({ tools: { adapter: { call } } })
const result = await Effect.runPromise(
runtime.execute(
`return await tools.adapter.call({ q: undefined, limit: 0 / 0, rate: 1 / 0, items: [1, undefined, 2], holes: [1, , 3] })`,
),
)
expect(result.ok).toBe(true)
const received = observed[0] as Record<string, unknown>
expect(received).toStrictEqual({ limit: null, rate: null, items: [1, null, 2], holes: [1, null, 3] })
// The undefined-valued property is dropped like JSON.stringify, not delivered as undefined.
expect(Object.hasOwn(received, "q")).toBe(false)
})
test("dropping undefined values lets optionalKey schemas accept conditional arguments", async () => {
const observed: Array<unknown> = []
const find = Tool.make({
description: "Find things",
input: Schema.Struct({ query: Schema.optionalKey(Schema.String), limit: Schema.optionalKey(Schema.Number) }),
run: (input) =>
Effect.sync(() => {
observed.push(input)
return "ok"
}),
})
const runtime = CodeMode.make({ tools: { things: { find } } })
// The `cond ? value : undefined` idiom: optionalKey rejects a present undefined, so the
// JSON boundary must drop the key before the schema decodes.
const result = await Effect.runPromise(
runtime.execute(`return await tools.things.find({ query: undefined, limit: 5 })`),
)
expect(result.ok).toBe(true)
expect(observed).toStrictEqual([{ limit: 5 }])
const search = await Effect.runPromise(runtime.execute(`return (await search({ query: undefined })).items.length`))
expect(search.ok).toBe(true)
})
test("renders JSON Schema outputs and $defs references", async () => {
const lookup = Tool.make({
description: "Look up a user",

View file

@ -258,11 +258,23 @@ describe("H1: NaN/Infinity flow as intermediates and normalize to null at the bo
test("copyOut normalizes non-finite numbers to null (the shared return + tool-arg boundary)", () => {
// Tool-call arguments funnel through copyOut too, so this one function pins both boundaries.
expect(ToolRuntime.copyOut(NaN)).toBeNull()
expect(ToolRuntime.copyOut(Infinity)).toBeNull()
expect(ToolRuntime.copyOut(-Infinity)).toBeNull()
expect(ToolRuntime.copyOut(42)).toBe(42)
expect(ToolRuntime.copyOut({ a: NaN, b: [Infinity, 1] })).toEqual({ a: null, b: [null, 1] })
expect(ToolRuntime.copyOut(NaN, "json")).toBeNull()
expect(ToolRuntime.copyOut(Infinity, "json")).toBeNull()
expect(ToolRuntime.copyOut(-Infinity, "nullify")).toBeNull()
expect(ToolRuntime.copyOut(42, "json")).toBe(42)
expect(ToolRuntime.copyOut({ a: NaN, b: [Infinity, 1] }, "json")).toEqual({ a: null, b: [null, 1] })
})
})
describe("copyOut undefined handling per boundary mode", () => {
test("json mode mirrors JSON.stringify for undefined", () => {
expect(ToolRuntime.copyOut({ q: undefined, keep: 1 }, "json")).toStrictEqual({ keep: 1 })
expect(ToolRuntime.copyOut([1, undefined, 2], "json")).toStrictEqual([1, null, 2])
expect(ToolRuntime.copyOut({ nested: { a: undefined, b: [undefined] } }, "json")).toStrictEqual({
nested: { b: [null] },
})
expect(ToolRuntime.copyOut(undefined, "json")).toBeUndefined()
expect(ToolRuntime.copyOut({ a: undefined }, "nullify")).toStrictEqual({ a: null })
})
})