Skip to content

Latest commit

 

History

History

codemode

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

README.md

@opencode/codemode

This is our take on code mode: a lightweight, pure interpreter for a JavaScript-like language built around calling tools. The language itself is psychoscript; see its interpreter support checklist for what is supported.

Rather than trying to sandbox arbitrary JavaScript, CodeMode only runs the language features we implement. Programs cannot directly access the network, filesystem, processes, or application APIs. They can interact with the outside world only through tools provided by the host, which can also limit execution time, tool calls, output size, and data.

The idea of code mode was originally introduced by Cloudflare. See their post to learn more about the concept and their isolate-based approach.

How it differs from JavaScript

  • Only supported APIs are available. Programs can use the provided tools, supported JavaScript built-ins, and the globals the host adds through extensions. Timers, process, filesystem access, packages, and export are unavailable.
  • Unfinished work is interrupted. Tool calls and async functions start when called. When the program finishes, anything still running is interrupted. Unhandled rejections from un-awaited promises are returned as warnings.
  • REPL-style results. Without an explicit return, the final top-level expression becomes the result. undefined becomes null.

Unsupported syntax returns an UnsupportedSyntax diagnostic with a source location. Current gaps are tracked in psychoscript's interpreter support checklist.

Layout

The language (compiler, VM, runtime, standard library, snapshots, and the Effect driver) is the psychoscript package. src/codemode/ is Code Mode, built on it: tools and their schemas, the catalog and namespaces, search, scanTools, OpenAPI, hooks, maxToolCalls, CodeMode.make, and Extension.make.

Quick Start

import { CodeMode, Namespace, Tool } from "@opencode/codemode"
import { Effect, Schema } from "effect"

const lookupOrder = Tool.make({
  description: "Look up an order by ID",
  input: Schema.Struct({ id: Schema.String }),
  output: Schema.Struct({ id: Schema.String, status: Schema.String }),
  execute: ({ id }) => Effect.succeed({ id, status: "open" }),
})

const runtime = CodeMode.make({
  tools: { orders: { lookup: lookupOrder } },
})

const result = await Effect.runPromise(
  runtime.execute(`
    const order = await tools.orders.lookup({ id: "order_42" })
    return { id: order.id, needsAttention: order.status !== "complete" }
  `),
)

result is always a CodeMode.Result.

API

Tool.make

input and output accept either an Effect Schema or a render-only JSON Schema document. Effect Schema input is decoded before execute; Effect Schema output is decoded and safely copied before the program sees it. JSON Schemas only shape the model-visible signature. Without output, the signature uses Promise<void>.

Descriptions and schemas are model-visible contracts. Authorization belongs in execute.

Set freeform: true on a tool whose input is one required string field to let programs pass that string directly. For { patchText: string }, the catalog shows tools.patch(patchText: string), and tools.patch("*** Begin Patch...") delivers { patchText: "*** Begin Patch..." } to validation, hooks and the executor. The object form still works.

Nested records are the shorthand for ordinary namespaces. Use Namespace.make when a namespace needs a description:

const runtime = CodeMode.make({
  tools: {
    orders: Namespace.make({
      description: "Purchases, fulfillment, and shipment tracking",
      tools: { lookup: lookupOrder },
    }),
  },
})

Namespace descriptions are optional and participate in search matching for every descendant tool. Names still come from record keys, so the wrapper does not repeat orders. Dots in keys create nested paths; other characters use bracket notation, such as tools.context7["resolve-library-id"](...).

CodeMode.execute and CodeMode.make

CodeMode.execute({ ...options, code }) runs once. CodeMode.make(options) creates a reusable runtime:

const runtime = CodeMode.make({ tools, limits: { timeoutMs: 30_000 } })

runtime.catalog() // structured tool descriptions
runtime.execute(source) // Effect<CodeMode.Result, never, ToolServices>

The Effect environment is inferred from the supplied tools. hooks surround every call the program makes into the host: tool.before/tool.after receive { name, input, key } with decoded input and the call's <runId>:<callId> idempotency key, and extension.before/extension.after receive { extension, name, args }. An after hook also receives how the call ended (success with its value, failure with its error, or interrupted). A before hook runs before the call reaches the host, and a failing one denies it: the host function is never called, and the program catches the failure as a thrown error. A denied call fails the way the call itself would: a tool call or asynchronous global rejects its promise, and a synchronous global throws.

Tool arguments cross into the host the way extension arguments do, as copies (plain data, Date, Map, Set, URL, Headers, errors), and Code Mode then turns them into JSON next to the tool's schema: a Date is its ISO string, a Set its items, a URLSearchParams its query string, an Error { name, message, ...own fields }, and Map and RegExp values {}. A Uint8Array asks to be encoded as text first. A program's own toJSON is not called, but its getters are: each runs once, before the host function does, and its value crosses as data; a property with only a setter is left out. A promise or function inside an argument fails the call: await it, or pass its data.

Suspending and resuming a run

CodeMode.make returns a CodeMode.Runtime, which can also run a program step by step: runtime.program(code, bindings?, { runId }?) makes a CodeMode.Program, and nothing runs until program.run(). Every host call gets a stable key <runId>:<callId> (using a generated UUID when runId is omitted), saved in the snapshot and passed to tools as execute(input, { key }) and to hooks as call.key. run() runs the program in the current fiber until it ends or is suspended, and gives a CodeMode.Outcome: a Result, or:

type Suspended = {
  ok: "suspended"
  state: CodeMode.Snapshot // JSON; CodeMode.toBytes / CodeMode.fromBytes convert it
}

const runtime = CodeMode.make({ tools })

// inline
const outcome = yield* runtime.program(code).run()

// in the background, suspended from another fiber
const program = runtime.program(code, undefined, { runId: "run-1" })
const fiber = yield* Effect.forkChild(program.run())
const suspended = yield* program.suspend() // the same Suspended that run() returns
const outcome = yield* Fiber.join(fiber)

// later, in this process or another one built from the same version
yield* runtime.resume(CodeMode.toBytes(suspended.state)).run()

A program runs once: running it again is a defect, so make another with runtime.program or runtime.resume. run() owns the run's host calls: interrupting it interrupts the program and every call in progress (and cancels any started durable tool work).

Idempotent and durable tools, and safe suspension

A tool's execute is either a plain function or wrapped with Tool.idempotent or Tool.durable:

// plain (default): a suspend is refused while it runs
Tool.make({
  description: "...",
  input: ...,
  output: ...,
  execute: (input, { key }) => compute(input),
})

// idempotent: safe to run again; interrupted on suspend, rerun on resume with the same key
Tool.make({
  description: "...",
  input: ...,
  output: ...,
  execute: Tool.idempotent((input, { key }) => Db.get(input.name)),
})

// durable: keeps running across a suspend, reconnected through its JSON data on resume
Tool.make({
  description: "...",
  input: ...,
  output: ...,
  execute: Tool.durable({
    start: (input, { key }) => Jobs.submit(input, key).pipe(Effect.map((id) => ({ id }))),
    wait: ({ id }, { key }) => Jobs.wait(id),
    cancel: ({ id }) => Jobs.kill(id),
  }),
})
  • Plain (execute: (input, { key }) => ...): not durable. Suspend is refused while the call is running.

  • Tool.idempotent(execute): idempotent/read-only work safe to run again. If a suspend lands while the call is running, the VM detaches and interrupts it without firing tool.after, saves the call in the snapshot as queued, and re-runs it on resume with the same key without double-counting it against maxToolCalls.

  • Tool.durable({ start, wait, cancel }): work that outlives a suspended run:

    • start(input, { key }) => Effect<Data /* Json */> begins the work and returns JSON data (inferred into wait and cancel);
    • wait(data, { key }) => Effect<Output> connects to the running work and waits for its result;
    • cancel(data) => Effect<void> stops the work when the run is interrupted for real or ends, never when it is suspended.

    While start is running, the call blocks suspend. Once start returns its JSON data, the call is in the started state with its data saved in the snapshot, and suspend is allowed. wait is not tied to the program's JavaScript await: the VM calls wait(data) right after start returns, and calls wait(data) again for every started call right when resume starts, whatever the program does with the call's promise (whether it awaits it immediately, passes it to Promise.all, chains .then, or keeps the unawaited promise in a variable and awaits it later).

program.suspend(options?) saves the run's state and stops it at once, and run() returns the same Suspended. Suspend is allowed only when every in-flight call is queued, Tool.idempotent, or Tool.durable in started. When any other host call is in progress — a plain tool, a durable tool still in start, search, or an async or synchronous extension call — suspend fails with a CodeMode.SuspendError naming each blocking call in running by its key and path (tools.x.y, search, fetch), and the run goes on untouched. Tool calls queued but not started don't block: they start on resume. Outcomes already waiting in the inbox are part of the state. A suspend after the run ended gives its result; one issued before run() has started waits for it to start (and for ever if it never runs).

The suspend is immediate: the program shares the thread with the caller, so it is at a step boundary. Its checks, saving the state, and checking it against options.maxBytes all happen before anything about the run changes, so a state that can't be saved or is too large fails the suspend the same way.

A resumed run keeps the limits it was saved with: timeoutMs counts active time across every resume, and maxToolCalls counts every admitted call of the run (Code Mode keeps them in the run's metadata, saved with the state). A state resumes only on the same build of psychoscript and with the same host globals (the same extension globals, with the same snapshot hooks). The state is checked when the resumed program runs: its run() fails with a CodeMode.SnapshotError when it can't be resumed here or is malformed. Built-ins are saved by name and only their changes are copied.

Extension.make

Extensions are host functions a program calls directly as globals, such as fetch. Unlike tools they are not in the catalog, not counted against maxToolCalls, and not described to the model; the host decides what they mean.

const web = Extension.make({ name: "web", globals: { fetch: async (url: string) => globalThis.fetch(url) } })
const runtime = CodeMode.make({ tools, extensions: [web] })

A global is synchronous unless it is an async function or is marked with Extension.async. Mark a plain function that returns a Promise or an Effect, so a denied call rejects its promise like any other failure:

Extension.make({ name: "web", globals: { fetch: Extension.async((url: string) => globalThis.fetch(url)) } })

A global may return a value, a Promise, or an Effect. A synchronous global's Effect runs to completion before the call returns, so the host can offer blocking calls such as fs.readFileSync. An Effect's requirements join the runtime's environment the same way a tool's do, so the host can provide per-execution services such as the calling session:

const web = Extension.make({
  name: "web",
  globals: {
    fetch: Extension.async((url: string) =>
      Effect.gen(function* () {
        const session = yield* Session
        yield* session.authorize(url)
        return yield* Effect.promise(() => globalThis.fetch(url))
      }),
    ),
  },
})
const result = yield* CodeMode.make({ extensions: [web] }).execute(code).pipe(Effect.provideService(Session, session))

Every value crossing in either direction is converted, never shared: arguments come in as copies (a program getter runs and crosses as its value; a setter-only property is left out), results go out as copies, and a function inside a result is callable the same way. When the program passes such a function back to an extension, it arrives as the host function itself, so copy(dest, file(src)) can tell what it was given. Globals and the program's own functions never cross back: they are left out, as in tool arguments. A global that shadows a built-in or another extension throws at CodeMode.make.

Mark a global with Extension.constructible to let programs call it with new, as they would a class. new calls it the same way a plain call does, and the program gets what it returns:

Extension.make({ name: "glob", globals: { Glob: Extension.constructible((pattern: string) => makeGlob(pattern)) } })

Functions inside a result have no data form, so a suspended run loses them by default: after a resume, calling one throws a TypeError. An extension can preserve them across suspend and resume with snapshot hooks. encode receives a value one of the extension's functions returned, as the host function returned it, and returns plain data (JSON values and Uint8Arrays); on resume, decode receives that data once per value and returns a host value of the same shape, whose functions the restored program copies call from then on:

const saved = Symbol("saved")
const makeResponse = (data: { url: string; body: Uint8Array }) => ({
  [saved]: data,
  url: data.url,
  text: async () => new TextDecoder().decode(data.body),
  bytes: async () => data.body,
})
const web = Extension.make({
  name: "web",
  globals: {
    fetch: Extension.async(async (url: string) =>
      makeResponse({ url, body: await (await globalThis.fetch(url)).bytes() }),
    ),
  },
  snapshot: {
    encode: (value) => (value as { [saved]: { url: string; body: Uint8Array } })[saved],
    decode: (data) => makeResponse(data as { url: string; body: Uint8Array }),
  },
})

A suspend fails (and the run continues) when encode throws or returns non-plain data. A snapshot saved with an extension's hooks resumes only in a runtime whose globals of the same names also have snapshot hooks.

A global can also be an object of functions, such as fs. Every function reachable from it is an entry point like a global, so calls fire the extension hooks with their path as the name (fs.readFile, fs.promises.readFile).

Programs load globals the way Node code loads modules. import and require take a global by its Node module name, with or without node:; fs/promises, url, util, and buffer resolve to the globals that provide them, and an extension can add its own names (modules: { bun: "Bun" } makes import { file } from "bun" load Bun). There are no packages, so anything else throws.

import fs, { readFile } from "node:fs" // the `fs` global, if the host provides one
import { URL } from "url"
const crypto = require("crypto")
const { default: same } = await import("fs")

OpenAPI tools

OpenAPI.fromSpec converts an OpenAPI 3.x document into one tool per supported operation. Dotted operationId values create namespaces:

const api = OpenAPI.fromSpec({ spec, auth: { resolve } })
const runtime = CodeMode.make({ tools: { opencode: api.tools } })

The synchronous result is { tools, skipped }. Operations with unsupported parameter encodings, request bodies without JSON content, WebSocket or SSE semantics, or binary responses are reported in skipped.

Authentication is resolved by the host and never shown to the model. Generated tools require HttpClient.HttpClient. Request signatures omit readOnly properties; response signatures omit writeOnly properties. These JSON Schemas shape model-visible signatures but do not filter runtime values: nested JSON body properties and decoded server responses pass through unchanged. See src/codemode/openapi/types.ts for option details.

The language layer

Code Mode is built on psychoscript, which runs programs over any host globals given through Host and Runtime.make. See its README for that API.

Results

Every execution returns:

type Result =
  | {
      readonly ok: true
      readonly value: CodeMode.DataValue
      readonly warnings?: ReadonlyArray<CodeMode.Diagnostic>
      readonly logs?: ReadonlyArray<string>
      readonly truncated?: boolean
      readonly toolCalls: ReadonlyArray<CodeMode.ToolCall>
    }
  | {
      readonly ok: false
      readonly error: CodeMode.Diagnostic
      readonly logs?: ReadonlyArray<string>
      readonly truncated?: boolean
      readonly toolCalls: ReadonlyArray<CodeMode.ToolCall>
    }

value is JSON-safe. warnings are non-fatal diagnostics, logs contain program console output, and truncated indicates that retained output was cut by maxOutputBytes. toolCalls retains admitted calls in order, including after failure.

Diagnostic kinds:

Kind Meaning
ParseError Source is empty or cannot be parsed.
UnsupportedSyntax Parsed JavaScript is outside the supported subset.
UnknownTool The program referenced an unavailable tool.
InvalidToolInput Tool input failed schema decoding or safe-data copying.
InvalidToolOutput Tool output failed schema decoding or safe-data copying.
InvalidDataValue Program data violated the plain-data contract.
ToolCallLimitExceeded The program exceeded maxToolCalls.
TimeoutExceeded Execution timed out; as a warning, background work was interrupted after the program returned.
ToolFailure A tool refused or failed.
ExecutionFailure The program threw or another execution error occurred.
Truncated Warning only: additional warnings were omitted by maxOutputBytes.

Host failures and defects report their messages and underlying causes. Invalid outputs include the validation or copying error. Interruption propagates without becoming an error diagnostic.

Discovery

runtime.catalog() returns structured descriptors — exact path, description, and generated TypeScript signature — for every visible tool. Hosts render their own model-facing instructions from these descriptors; CodeMode.searchSignature and CodeMode.toolExpression(path) supply the exact callable forms.

The synchronous search(...) built-in is always available. It supports exact-path lookup, namespace-scoped search, empty-query browsing, and pagination, and returns callable paths with full signatures. Its result can be used directly, awaited, or chained with then, catch, and finally, so programs work whichever way they treat it. Search counts toward maxToolCalls. Search also matches descriptions from enclosing Namespace values. A plain string is the query: search("issues") is search({ query: "issues" }).

Execution Limits

Limit Default Controls
timeoutMs unlimited Total execution time.
maxToolCalls unlimited Admitted tool calls.
maxOutputBytes unlimited Retained result value and logs.

Execution limits have no default values.

Invalid limit configuration throws RangeError. Warnings receive a separate budget equal to maxOutputBytes. Truncation does not fail execution; an oversized value becomes a string with an in-band marker. Timeouts interrupt tool calls and busy loops, while a result returned before cleanup times out remains successful with a TimeoutExceeded warning. Tool-call concurrency is unrestricted. Boundary data is limited to 32 nested levels.