fix: 修复 agent 终端工具审查问题(terminal_list 输出契约、跨会话归属、UTF-8 截断边界等),工具注入改为设置面板开关(默认关闭)

This commit is contained in:
Menghuan1918
2026-08-10 21:43:58 +08:00
parent 6f62d19905
commit 8edfbfde44
17 changed files with 581 additions and 130 deletions
+41 -13
View File
@@ -27,17 +27,29 @@ export type AgentTerminalSignal = (typeof ALLOWED_SIGNALS)[number]
/** Default read page size (lines) when the caller omits `count`. */
export const DEFAULT_READ_COUNT = 500
/** Smallest pty dimension the registry accepts (mirrors the tool contract). */
export const TERMINAL_DIM_MIN = 2
/** Largest pty dimension the registry accepts (mirrors the tool contract). */
export const TERMINAL_DIM_MAX = 1024
/** Clamp one cols×rows pair into the supported pty range (flooring decimals). */
export function clampDims(cols: number, rows: number): { cols: number; rows: number } {
const clamp = (value: number): number =>
Math.min(TERMINAL_DIM_MAX, Math.max(TERMINAL_DIM_MIN, Math.floor(value)))
return { cols: clamp(cols), rows: clamp(rows) }
}
/**
* Serializable snapshot of one agent terminal — the shape the model sees
* through `terminal_list` and the sidebar sees through the push endpoint.
* Carries no pty reference and no transcript (those are reached through
* dedicated read/attach paths).
* dedicated read/attach paths), and no sessionId: ownership is registry
* internals, scoped by the caller (list filters by session, the push
* endpoint scopes by its query param), never part of the serialized view.
*/
export interface AgentTerminalSnapshot {
/** Stable opaque handle the model passes back to other terminal_* tools. */
uuid: string
/** The conversation id this terminal belongs to (scope filter for list). */
sessionId: string
/** Display title the model chose at create time. */
title: string
/** The command the model asked to run at create time (verbatim). */
@@ -115,8 +127,6 @@ export interface AgentTerminalReadResult {
lineBegin: number
/** 0-based index of the last line in `text` (exclusive). */
lineEnd: number
/** Whether `text` was truncated to fit the read cap. */
truncated: boolean
}
/** Outcome of {@link AgentPtyRegistry.waitFor}. */
@@ -158,7 +168,6 @@ export type AgentTerminalWaitResult =
export function snapshotOf(handle: AgentTerminalHandle): AgentTerminalSnapshot {
const out: AgentTerminalSnapshot = {
uuid: handle.uuid,
sessionId: handle.sessionId,
title: handle.title,
command: handle.command,
exited: handle.exited,
@@ -201,10 +210,11 @@ export class AgentPtyRegistry {
rows = 24,
): string {
const uuid = randomUUID()
const dims = clampDims(cols, rows)
const pty = nodePty.spawn(this.shell, [], {
name: 'xterm-256color',
cols: Math.max(2, Math.floor(cols)),
rows: Math.max(2, Math.floor(rows)),
cols: dims.cols,
rows: dims.rows,
cwd,
env: { ...process.env },
})
@@ -267,6 +277,20 @@ export class AgentPtyRegistry {
return handle
}
/**
* Resolve a live handle that belongs to `sessionId`, or throw `not-found`.
* The model-facing tools call this before every uuid-keyed operation: a
* uuid from another session is indistinguishable from an unknown one, so a
* model can never reach (or probe) a terminal it does not own.
*/
assertOwned(uuid: string, sessionId: string): AgentTerminalHandle {
const handle = this.expect(uuid)
if (handle.sessionId !== sessionId) {
throw new SidebarError('not-found', `agent terminal "${uuid}" not found`, 404)
}
return handle
}
/** Resolve a handle's snapshot, or undefined if it does not exist. */
snapshot(uuid: string): AgentTerminalSnapshot | undefined {
const handle = this.sessions.get(uuid)
@@ -310,15 +334,19 @@ export class AgentPtyRegistry {
totalLines,
lineBegin: start,
lineEnd: end,
truncated: false,
}
}
/** Resize a terminal's pty (clamped to a 2..1024 sane range). */
resize(uuid: string, cols: number, rows: number): void {
/**
* Resize a terminal's pty, clamped to the 2..1024 sane range.
* @returns the dimensions actually applied (the caller echoes these, so the
* reported value always matches the pty).
*/
resize(uuid: string, cols: number, rows: number): { cols: number; rows: number } {
const handle = this.expect(uuid)
if (handle.exited) return
handle.pty.resize(Math.max(2, Math.floor(cols)), Math.max(2, Math.floor(rows)))
const dims = clampDims(cols, rows)
if (!handle.exited) handle.pty.resize(dims.cols, dims.rows)
return dims
}
/**
+20
View File
@@ -128,6 +128,13 @@ export function SideCardSection({ store }: SideCardSectionProps) {
void commit({ autoOpenSubagent: next }).then(outcome => applyOutcome(previous, outcome))
}
const onToggleTools = (next: boolean): void => {
const previous = prefs
setPrefs({ ...previous, agentTerminalTools: next })
setError(null)
void commit({ agentTerminalTools: next }).then(outcome => applyOutcome(previous, outcome))
}
const commitWidth = (): void => {
const parsed = Number(widthDraft)
if (!Number.isFinite(parsed)) {
@@ -193,6 +200,19 @@ export function SideCardSection({ store }: SideCardSectionProps) {
onChange={event => { onToggleSubagent(event.currentTarget.checked) }}
/>
</label>
<label className={css.row}>
<span className={css.rowText}>
<span className={css.title}>{t('settingsToolsTitle')}</span>
<span className={css.desc}>{t('settingsToolsDesc')}</span>
</span>
<input
type="checkbox"
className={css.toggle}
checked={prefs.agentTerminalTools}
aria-label={t('settingsToolsTitle')}
onChange={event => { onToggleTools(event.currentTarget.checked) }}
/>
</label>
{error !== null && (
<div className={css.error} role="alert">
{error}
+28 -11
View File
@@ -15,7 +15,7 @@ import clsx from 'clsx'
import { IconChevronRightOutline14, IconFullscreenOutline16, IconPanelLeftOutline16, Tooltip } from '@deepseek-ai/dsh-client-ui-primitives'
import type { Context, SidebarConversation, SidebarSessionList } from '../context-types.ts'
import {
allLeaves, closeTab, leafWithTab, mapLeaf, moveTab, moveTabToEdge, openTab,
agentUuidOf, allLeaves, closeTab, isAgentTabId, leafWithTab, mapLeaf, moveTab, moveTabToEdge, openTab,
reconcileAgentTerminals,
resizeSplit, setWidth, toggleExpanded, togglePanel,
TERMINAL_LIMIT, type DropZone, type SidebarState, type SidebarStore, type SidebarTab, type SplitNode,
@@ -36,6 +36,10 @@ import { api } from './api.ts'
import { openSidebarFile } from './intercept.tsx'
import css from './sidebar.module.css'
/** How many consecutive reconnect failures stop the agent-terminals push loop
* (mirror of the terminal view's own cap; the loop restarts on session switch). */
const FAILURE_LIMIT = 3
/** Render the content of one tab (dispatched by type). */
function TabContent(props: {
tab: SidebarTab
@@ -83,19 +87,21 @@ function TabContent(props: {
}
}
/** The + menu options for the current state (terminal cap applied). */
/** The + menu options for the current state (UI-tab terminal cap applied).
* Agent-owned terminals (`agent:` tabs) are the model's, not the user's:
* they never count toward the per-session UI cap. */
function buildNewTabOptions(state: SidebarState): NewTabOption[] {
const terminalCount = allLeaves(state.splits)
const uiTerminalCount = allLeaves(state.splits)
.flatMap(leaf => leaf.tabs)
.filter(tab => tab.type === 'terminal').length
.filter(tab => tab.type === 'terminal' && !isAgentTabId(tab.id)).length
return [
{ id: 'explorer', label: t('openExplorer'), icon: tabTypeIcon('explorer', 16) },
{ id: 'git', label: t('openGit'), icon: tabTypeIcon('git', 16) },
{ id: 'subagent', label: t('openSubagent'), icon: tabTypeIcon('subagent', 16) },
{
id: 'terminal',
label: terminalCount >= TERMINAL_LIMIT ? `${t('newTerminal')} (${t('terminalLimit')})` : t('newTerminal'),
disabled: terminalCount >= TERMINAL_LIMIT,
label: uiTerminalCount >= TERMINAL_LIMIT ? `${t('newTerminal')} (${t('terminalLimit')})` : t('newTerminal'),
disabled: uiTerminalCount >= TERMINAL_LIMIT,
icon: tabTypeIcon('terminal', 16),
},
]
@@ -145,13 +151,16 @@ export function Sidebar(props: { ctx: Context; store: SidebarStore }) {
* create / close / exit; the sidebar reconciles the list into tabs
* (id `agent:<uuid>`, title from the agent). A disconnected socket
* retries with a short backoff so a refresh or transient drop reattaches
* the same shell without losing the agent's work.
* the same shell without losing the agent's work — capped like the
* terminal view's own reconnect loop, so a refused endpoint never spins
* forever (the next session switch restarts the loop).
*/
useEffect(() => {
if (sessionId === undefined) return
let socket: WebSocket | null = null
let retry: number | undefined
let closed = false
let failures = 0
const connect = (): void => {
if (closed) return
const url = new URL('/sidebar/ws/agent-terminals', location.origin)
@@ -169,7 +178,13 @@ export function Sidebar(props: { ctx: Context; store: SidebarStore }) {
}
}
socket.onclose = () => {
if (!closed) retry = window.setTimeout(connect, 2000)
if (closed) return
failures += 1
if (failures >= FAILURE_LIMIT) {
console.error('[dsh-better-sidebar] agent-terminals connection failed; stopping reconnect loop', sessionId)
return
}
retry = window.setTimeout(connect, 2000)
}
socket.onerror = () => { socket?.close() }
}
@@ -265,8 +280,8 @@ export function Sidebar(props: { ctx: Context; store: SidebarStore }) {
const tab = leaf?.tabs.find(candidate => candidate.id === tabId)
store.reduce(s => closeTab(s, paneId, tabId))
if (tab?.type === 'terminal') {
if (tabId.startsWith('agent:')) {
const uuid = tabId.slice('agent:'.length)
if (isAgentTabId(tabId)) {
const uuid = agentUuidOf(tabId)
void api.agentPtyClose(uuid).catch(() => { /* the host may already have released it */ })
} else if (sessionId !== undefined) {
void api.ptyClose({ sessionId, cwd }, tabId).catch(() => { /* the host may already have released it */ })
@@ -350,7 +365,9 @@ export function Sidebar(props: { ctx: Context; store: SidebarStore }) {
return openTab(s, { id: 'subagent', type: 'subagent', title: t('subagent') })
}
if (optionId === 'terminal') {
const count = allLeaves(s.splits).flatMap(leaf => leaf.tabs).filter(tab => tab.type === 'terminal').length
const count = allLeaves(s.splits)
.flatMap(leaf => leaf.tabs)
.filter(tab => tab.type === 'terminal' && !isAgentTabId(tab.id)).length
if (count >= TERMINAL_LIMIT) return s
const tab: SidebarTab = {
id: `terminal:${s.nextTerminal}`,
+2 -15
View File
@@ -25,26 +25,13 @@ import { FitAddon } from '@xterm/addon-fit'
import 'xterm/css/xterm.css'
import { t } from './locales.ts'
import type { SessionScope } from './api.ts'
import type { SidebarStore } from './state.ts'
import { agentUuidOf, isAgentTabId, type SidebarStore } from './state.ts'
import { isDarkScheme, subscribeColorScheme, tokenValue } from './theme.ts'
import css from './sidebar.module.css'
/** How many consecutive unreasoned failures before showing the error banner. */
const FAILURE_LIMIT = 3
/** Prefix marking a tab id as an agent-owned terminal (suffix is the uuid). */
const AGENT_TAB_PREFIX = 'agent:'
/** Whether a tab id refers to an agent-owned terminal. */
function isAgentTab(tabId: string): boolean {
return tabId.startsWith(AGENT_TAB_PREFIX)
}
/** Extract the agent terminal uuid from an `agent:<uuid>` tab id. */
function agentUuidOf(tabId: string): string {
return tabId.slice(AGENT_TAB_PREFIX.length)
}
/**
* Curated ANSI palettes for the terminal. The surface colors (background,
* foreground, cursor, selection) ride the theme tokens so the terminal
@@ -126,7 +113,7 @@ export function TerminalView(props: { scope: SessionScope; tabId: string; store:
// Agent terminals attach by uuid (the host looks them up in the agent
// pty registry); UI-tab terminals attach by sessionId+tab (the host
// uses the UI-tab pty manager). Same upgrade endpoint, different query.
if (isAgentTab(tabId)) {
if (isAgentTabId(tabId)) {
url.search = new URLSearchParams({ uuid: agentUuidOf(tabId) }).toString()
} else {
const params = new URLSearchParams({ sessionId: scope.sessionId, tab: tabId })
+4
View File
@@ -90,6 +90,8 @@ const zh = {
zoomHint: 'Alt + 滚轮',
settingsSubagentTitle: '检测到子代理时自动展开子代理页面',
settingsSubagentDesc: '当前会话产生新的子代理时,自动展开侧边栏并打开子代理页面;关闭后需手动打开',
settingsToolsTitle: '为模型注入终端工具',
settingsToolsDesc: '开启后,模型可通过 terminal_create 等 8 个工具创建并操作侧边栏终端(默认关闭)',
subagent: '子代理',
openSubagent: '子代理',
subagentMainAgent: '主代理',
@@ -192,6 +194,8 @@ const en: Record<keyof typeof zh, string> = {
zoomHint: 'Alt + wheel',
settingsSubagentTitle: 'Auto-open the Subagent page when a subagent appears',
settingsSubagentDesc: 'Expand the side card and open the Subagent page when the current conversation spawns a new subagent; turn off to open it manually',
settingsToolsTitle: 'Inject terminal tools for the model',
settingsToolsDesc: 'When enabled, the model can create and drive sidebar terminals through the 8 terminal_* tools (off by default)',
subagent: 'Subagents',
openSubagent: 'Subagents',
subagentMainAgent: 'Main agent',
+3
View File
@@ -41,6 +41,9 @@ export function parsePrefs(value: unknown): SidebarPrefs {
autoOpenSubagent: typeof record.autoOpenSubagent === 'boolean'
? record.autoOpenSubagent
: SIDEBAR_PREFS_DEFAULTS.autoOpenSubagent,
agentTerminalTools: typeof record.agentTerminalTools === 'boolean'
? record.agentTerminalTools
: SIDEBAR_PREFS_DEFAULTS.agentTerminalTools,
}
}
+19 -4
View File
@@ -399,6 +399,21 @@ export function resizeSplit(node: SplitNode, splitId: string, index: number, del
/** Prefix marking a tab id as an agent-owned terminal (suffix is the uuid). */
export const AGENT_TAB_PREFIX = 'agent:'
/** Whether a tab id refers to an agent-owned terminal. */
export function isAgentTabId(tabId: string): boolean {
return tabId.startsWith(AGENT_TAB_PREFIX)
}
/** Extract the agent terminal uuid from an `agent:<uuid>` tab id. */
export function agentUuidOf(tabId: string): string {
return tabId.slice(AGENT_TAB_PREFIX.length)
}
/** Build the sidebar tab id for one agent terminal uuid. */
export function agentTabId(uuid: string): string {
return `${AGENT_TAB_PREFIX}${uuid}`
}
/**
* Reconcile the sidebar's agent-terminal tabs with the host's live list.
* The host pushes the current list of agent terminals (created by the model
@@ -416,11 +431,11 @@ export function reconcileAgentTerminals(
agentTerminals: ReadonlyArray<{ uuid: string; title: string }>,
): SidebarState {
const existingTabs = allLeaves(state.splits).flatMap(leaf => leaf.tabs)
const existingAgentTabs = existingTabs.filter(tab => tab.id.startsWith(AGENT_TAB_PREFIX))
const existingUuids = new Set(existingAgentTabs.map(tab => tab.id.slice(AGENT_TAB_PREFIX.length)))
const existingAgentTabs = existingTabs.filter(tab => isAgentTabId(tab.id))
const existingUuids = new Set(existingAgentTabs.map(tab => agentUuidOf(tab.id)))
const serverUuids = new Set(agentTerminals.map(t => t.uuid))
const toAdd = agentTerminals.filter(t => !existingUuids.has(t.uuid))
const toRemove = existingAgentTabs.filter(tab => !serverUuids.has(tab.id.slice(AGENT_TAB_PREFIX.length)))
const toRemove = existingAgentTabs.filter(tab => !serverUuids.has(agentUuidOf(tab.id)))
if (toAdd.length === 0 && toRemove.length === 0) return state
// Remove tabs whose uuids vanished from the server list (the agent closed
// them, or the pty exited and was reaped). Reuse closeTab's leaf cleanup.
@@ -436,7 +451,7 @@ export function reconcileAgentTerminals(
let next: SidebarState = { ...state, splits }
for (const terminal of toAdd) {
const tab: SidebarTab = {
id: `${AGENT_TAB_PREFIX}${terminal.uuid}`,
id: agentTabId(terminal.uuid),
type: 'terminal',
title: terminal.title,
}
+1
View File
@@ -79,4 +79,5 @@ export const PrefsSchema: z<SidebarPrefs> = z.object({
openByDefault: z.boolean().default(true),
defaultWidthPercent: z.number().step(1).min(WIDTH_PERCENT_MIN).max(WIDTH_PERCENT_MAX).default(WIDTH_PERCENT_DEFAULT),
autoOpenSubagent: z.boolean().default(true),
agentTerminalTools: z.boolean().default(false),
})
+38 -13
View File
@@ -23,13 +23,14 @@ import {
SIDEBAR_PREFS_NS,
type ResolvedSidebarConfig,
type SidebarConfig,
type SidebarPrefs,
} from './config.ts'
import { isWithin, parentOf, requireAbsolute, listDirectory, rootLabel } from './fs-tree.ts'
import { isTrustedApiRequest } from './trust-fence.ts'
import * as git from './git.ts'
import { SettingsConflictError, settingsNamespace, type SettingsNamespace } from '@deepseek-ai/dsh-settings'
import { defaultShell, ensureSpawnHelper, PtyManager } from './pty-manager.ts'
import { AgentPtyRegistry, type AgentTerminalHandle } from './agent-pty.ts'
import { AgentPtyRegistry, clampDims, type AgentTerminalHandle } from './agent-pty.ts'
import { registerTools } from './tools.ts'
import { readJsonBody, requireString, SidebarError, writeError, writeJson, writeOk } from './wire.ts'
@@ -315,9 +316,34 @@ export function apply(ctx: Context, config?: SidebarConfig): void {
// which call the seam in-process. Deployments without a settings service
// simply never fill the face and the client falls back to the defaults.
let settingsFace: SidebarSettingsFace | undefined
// The model-facing terminal tools are gated on the side-card setting
// `agentTerminalTools` (default off): nothing is injected until the user
// turns the feature on, and turning it off mid-session unregisters the
// tools and releases the agent terminals they created.
let toolsDisposers: (() => void) | null = null
const syncToolsGate = (scope: { get(): SidebarPrefs }): void => {
if (scope.get().agentTerminalTools) {
if (toolsDisposers === null) {
toolsDisposers = registerTools(ctx, agentPtyRegistry, (sessionId) => sessionCwdOf(ctx, sessionId))
}
} else if (toolsDisposers !== null) {
toolsDisposers()
toolsDisposers = null
// The feature is off: release every agent terminal the model created
// while it was on (they are only reachable through the tools). The
// registry change fires the push, so the sidebar reconciles them away.
agentPtyRegistry.disposeAll()
}
}
ctx.inject(['settings'], (sctx) => {
const ns: SettingsNamespace = settingsNamespace(SIDEBAR_PREFS_NS)
sctx.settings.register(ns, PrefsSchema)
// The structural settings mirror types `schema` as unknown, so the
// generic is not inferred here; the real service resolves it from the
// schemastery schema (PrefsSchema) — narrow the owner scope explicitly.
const scope = sctx.settings.register(ns, PrefsSchema) as {
get(): SidebarPrefs
watch(callback: (next: SidebarPrefs, prev: SidebarPrefs) => void): () => void
}
const viewOf = (): { value?: unknown; revision?: number } => {
const descriptor = sctx.settings.describe({ redactSecrets: true }).find(candidate => candidate.ns === ns)
return descriptor === undefined
@@ -331,15 +357,12 @@ export function apply(ctx: Context, config?: SidebarConfig): void {
return viewOf()
},
}
// Register (or unregister) the terminal tools from the current setting,
// and keep them in sync with every settings commit.
syncToolsGate(scope)
scope.watch(() => { syncToolsGate(scope) })
})
// ── Model-facing tools (terminal_create / list / send / read / wait_for / resize / signal / close) ─
// Registered once at activation; the disposer is attached to the plugin
// fiber by the tools registry, so unloading unregisters them. The cwd
// resolver threads the live session cwd (authoritative from the session
// store, falling back to the process cwd) so the model never passes it.
registerTools(ctx, agentPtyRegistry, (sessionId) => sessionCwdOf(ctx, sessionId))
// ── JSON API ────────────────────────────────────────────────────────────
const api = buildApi(ctx, ptyManager, agentPtyRegistry, resolved, () => settingsFace)
ctx.effect(() => ctx.httpServer.register({
@@ -467,6 +490,7 @@ export function apply(ctx: Context, config?: SidebarConfig): void {
}), 'dsh-better-sidebar: agent-terminals push WebSocket')
ctx.effect(() => () => {
toolsDisposers?.()
ptyManager.disposeAll()
agentPtyRegistry.disposeAll()
wss.close()
@@ -530,7 +554,7 @@ async function attachTerminal(
ws.close(1011, `agent terminal "${uuid}" not found`)
return
}
pumpAgentTerminal(agentPtyRegistry, handle, ws, resolved)
pumpAgentTerminal(agentPtyRegistry, handle, ws)
return
}
const sessionId = url.searchParams.get('sessionId')
@@ -577,7 +601,8 @@ async function attachTerminal(
&& control.type === 'resize'
&& typeof control.cols === 'number' && typeof control.rows === 'number'
) {
handle.pty.resize(Math.max(2, Math.floor(control.cols)), Math.max(2, Math.floor(control.rows)))
const dims = clampDims(control.cols, control.rows)
handle.pty.resize(dims.cols, dims.rows)
} else {
handle.pty.write(text)
}
@@ -606,7 +631,6 @@ function pumpAgentTerminal(
registry: AgentPtyRegistry,
handle: AgentTerminalHandle,
ws: WebSocket,
_resolved: ResolvedSidebarConfig,
): void {
if (handle.transcript !== '') ws.send(handle.transcript)
const onData = (data: string): void => {
@@ -642,7 +666,8 @@ function pumpAgentTerminal(
&& control.type === 'resize'
&& typeof control.cols === 'number' && typeof control.rows === 'number'
) {
handle.pty.resize(Math.max(2, Math.floor(control.cols)), Math.max(2, Math.floor(control.rows)))
const dims = clampDims(control.cols, control.rows)
handle.pty.resize(dims.cols, dims.rows)
} else if (control === null) {
// Raw text input (a JSON-looking string the pty would have received
// verbatim is reachable in theory but is exotic for an agent terminal;
+8
View File
@@ -20,6 +20,13 @@ export interface SidebarPrefs {
* Subagent page when the current conversation spawns a new subagent.
*/
autoOpenSubagent: boolean
/**
* Whether the model-facing agent terminal tools (terminal_create / list /
* send / read / wait_for / resize / signal / close) are injected into the
* model's toolset. Off by default: the feature stays dormant until the
* user explicitly enables it in the side card settings.
*/
agentTerminalTools: boolean
}
/** Range contract of {@link SidebarPrefs.defaultWidthPercent}. */
@@ -32,6 +39,7 @@ export const SIDEBAR_PREFS_DEFAULTS: SidebarPrefs = {
openByDefault: true,
defaultWidthPercent: WIDTH_PERCENT_DEFAULT,
autoOpenSubagent: true,
agentTerminalTools: false,
}
/** Clamp one width percent into the contract range (shared by schema and client reads). */
+75 -60
View File
@@ -27,17 +27,22 @@ import {
/** Maximum UTF-8 bytes of one `terminal_read` result text. */
const READ_BYTE_LIMIT = 256 * 1024
/** Bound a string to a byte limit, marking truncation. */
function boundBytes(text: string, maxBytes: number): { text: string; truncated: boolean } {
const bytes = Buffer.byteLength(text, 'utf8')
if (bytes <= maxBytes) return { text, truncated: false }
// Trim by character count proportional to the overage, then walk back to a
// valid UTF-8 boundary (Buffer.subarray with garbage on a trailing multibyte
// sequence would decode to U+FFFD).
const ratio = maxBytes / bytes
const cut = Math.floor(text.length * ratio)
const buf = Buffer.from(text, 'utf8').subarray(0, maxBytes)
return { text: buf.toString('utf8').slice(0, cut), truncated: true }
/**
* Bound a string to a byte limit, marking truncation. Truncation never
* splits a multi-byte UTF-8 sequence: when the byte cap lands inside one,
* the walk-back retreats to the sequence's leading byte so the retained
* prefix decodes cleanly (a split would decode to U+FFFD).
* @internal exported for the unit tests, like {@link snapshotOf}.
*/
export function boundBytes(text: string, maxBytes: number): { text: string; truncated: boolean } {
const buf = Buffer.from(text, 'utf8')
if (buf.byteLength <= maxBytes) return { text, truncated: false }
// The byte at `end` is a continuation byte (10xxxxxx) exactly when the
// character that started before `end` spills past the cap. Stop at the
// first non-continuation byte — the retained prefix is then intact.
let end = maxBytes
while (end > 0 && ((buf[end] ?? 0) & 0xc0) === 0x80) end -= 1
return { text: buf.subarray(0, end).toString('utf8'), truncated: true }
}
/** Pure text projection helper (the canonical value is already structured). */
@@ -53,27 +58,36 @@ function requireAgent(agent: Agent | undefined): Agent {
return agent
}
/** Resolve the calling agent's session id (the registry scope filter). */
function sessionOf(exec: ToolRunContext): { agent: Agent; sessionId: string } {
const agent = requireAgent(exec.agent)
return { agent, sessionId: agent.session.id }
/** Resolve the calling agent's session id (the registry scope + ownership key). */
function sessionIdOf(exec: ToolRunContext): string {
return requireAgent(exec.agent).session.id
}
/**
* Register the seven terminal tools against the host tool registry. The
* Register the eight terminal tools against the host tool registry. The
* `resolveCwd` callback threads the live session cwd (authoritative from the
* session store, falling back to the process cwd) so a freshly-created
* terminal lands in the right directory without the model passing it.
* Every uuid-keyed tool first asserts the terminal belongs to the calling
* session (`registry.assertOwned`), so one agent can never reach another
* session's terminals.
* @param ctx - host plugin context (carries the tools service).
* @param registry - the agent-owned terminal registry.
* @param resolveCwd - live cwd resolver for one session id.
* @returns a disposer that unregisters all eight tools (the caller gates
* registration on the side-card setting and calls this to turn them off).
*/
export function registerTools(
ctx: Context,
registry: AgentPtyRegistry,
resolveCwd: (sessionId: string) => string,
): void {
ctx.tools.register(defineTool({
): () => void {
const disposers: Array<() => void> = []
const register = (tool: ReturnType<typeof defineTool>): void => {
disposers.push(ctx.tools.register(tool))
}
register(defineTool({
name: 'terminal_create',
description:
'Open a persistent terminal in the sidebar and run a command in it. '
@@ -111,14 +125,14 @@ export function registerTools(
},
execute: (args: { title: string; command: string }, exec) => {
exec.signal.throwIfAborted()
const { sessionId } = sessionOf(exec)
const sessionId = sessionIdOf(exec)
const cwd = resolveCwd(sessionId)
const uuid = registry.create(sessionId, args.title, args.command, cwd, 80, 24)
return Promise.resolve({ uuid, title: args.title })
},
}))
ctx.tools.register(defineTool({
register(defineTool({
name: 'terminal_list',
description:
'List every terminal the current agent has opened in this session. Returns each terminal\'s uuid, title, '
@@ -154,19 +168,19 @@ export function registerTools(
},
},
execute: (_args, exec) => {
const { sessionId } = sessionOf(exec)
const sessionId = sessionIdOf(exec)
return Promise.resolve(registry.list(sessionId))
},
}))
ctx.tools.register(defineTool({
register(defineTool({
name: 'terminal_send',
description:
'Send raw text (keystrokes) to a terminal opened with terminal_create — tmux send-keys semantics. '
+ 'The text is written verbatim to the pty stdin. '
+ 'To submit a command, set submit=true (appends an Enter key); do NOT put "\\n" or "\\r" in the text yourself. '
+ 'To send Ctrl+C (interrupt the running command), use the terminal_signal tool with signal="SIGINT" — do NOT try to send the control character "\\u0003" as text. '
+ 'Similarly use terminal_signal for Ctrl+D (EOF), Ctrl+Z (suspend), etc. '
+ 'Use terminal_signal with signal="SIGTSTP" for Ctrl+Z (suspend) as well. '
+ 'This tool does NOT wait for the command to finish or for output to settle — pair with terminal_read to observe the result. '
+ 'Throws if the terminal has exited.',
parameters: {
@@ -200,13 +214,15 @@ export function registerTools(
},
execute: (args: { uuid: string; text: string; submit?: boolean }, exec) => {
exec.signal.throwIfAborted()
const sessionId = sessionIdOf(exec)
registry.assertOwned(args.uuid, sessionId)
const payload = args.submit === true ? `${args.text}\r` : args.text
registry.send(args.uuid, payload)
return Promise.resolve({ uuid: args.uuid, bytes: Buffer.byteLength(payload, 'utf8') })
},
}))
ctx.tools.register(defineTool({
register(defineTool({
name: 'terminal_read',
description:
'Read a bounded page of retained output from an agent terminal without sending input. '
@@ -249,6 +265,8 @@ export function registerTools(
},
execute: (args: { uuid: string; offset?: number; count?: number }, exec) => {
exec.signal.throwIfAborted()
const sessionId = sessionIdOf(exec)
registry.assertOwned(args.uuid, sessionId)
const result = registry.read(args.uuid, args.offset, args.count)
const bounded = boundBytes(result.text, READ_BYTE_LIMIT)
return Promise.resolve({
@@ -261,7 +279,7 @@ export function registerTools(
},
}))
ctx.tools.register(defineTool({
register(defineTool({
name: 'terminal_wait_for',
description:
'Block until a substring appears in a terminal\'s retained transcript, or until the timeout elapses, or until the terminal exits — whichever happens first. '
@@ -298,7 +316,7 @@ export function registerTools(
needle: { type: 'string', required: true },
line: { type: 'integer', required: true, description: '0-based line index in the retained transcript where the needle first appeared.' },
column: { type: 'integer', required: true, description: '0-based column index within that line where the match starts.' },
elapsed_ms: { type: 'integer', required: true, description: 'Wall-clock milliseconds from wait start to match.' },
elapsedMs: { type: 'integer', required: true, description: 'Wall-clock milliseconds from wait start to match.' },
},
},
{
@@ -307,8 +325,8 @@ export function registerTools(
properties: {
kind: { type: 'string', required: true, const: 'timeout' },
needle: { type: 'string', required: true },
timeout_ms: { type: 'integer', required: true, description: 'The configured timeout that elapsed.' },
total_lines: { type: 'integer', required: true, description: 'Total lines retained when the timeout fired. Call terminal_read to inspect the tail.' },
timeoutMs: { type: 'integer', required: true, description: 'The configured timeout that elapsed.' },
totalLines: { type: 'integer', required: true, description: 'Total lines retained when the timeout fired. Call terminal_read to inspect the tail.' },
},
},
{
@@ -317,55 +335,42 @@ export function registerTools(
properties: {
kind: { type: 'string', required: true, const: 'exited' },
needle: { type: 'string', required: true },
exit_code: { type: 'integer', description: 'Exit code, if known.' },
exit_signal: { type: 'string', description: 'Exit signal name, if killed by a signal.' },
exitCode: { oneOf: [{ type: 'integer' }, { type: 'null' }], description: 'Exit code, if known.' },
exitSignal: { oneOf: [{ type: 'string' }, { type: 'null' }], description: 'Exit signal name, if killed by a signal.' },
},
},
],
},
render: (_args, value) => {
const v = value as { kind: 'found' | 'timeout' | 'exited'; needle: string; elapsed_ms?: number; timeout_ms?: number; line?: number; column?: number; exit_code?: number | null; exit_signal?: string | null }
const v = value as { kind: 'found' | 'timeout' | 'exited'; needle: string; elapsedMs?: number; timeoutMs?: number; line?: number; column?: number; exitCode?: number | null; exitSignal?: string | null }
if (v.kind === 'found') {
return [{ type: 'text', text: `Found "${v.needle}" at line ${v.line}, column ${v.column} (after ${v.elapsed_ms}ms).` }]
return [{ type: 'text', text: `Found "${v.needle}" at line ${v.line}, column ${v.column} (after ${v.elapsedMs}ms).` }]
}
if (v.kind === 'timeout') {
return [{ type: 'text', text: `Timed out after ${v.timeout_ms}ms waiting for "${v.needle}". Call terminal_read to inspect the transcript.` }]
return [{ type: 'text', text: `Timed out after ${v.timeoutMs}ms waiting for "${v.needle}". Call terminal_read to inspect the transcript.` }]
}
const exitInfo = v.exit_code !== undefined ? ` (exit code ${v.exit_code})` : ''
const exitInfo = v.exitCode !== undefined && v.exitCode !== null ? ` (exit code ${v.exitCode})` : ''
return [{ type: 'text', text: `Terminal exited before "${v.needle}" appeared${exitInfo}.` }]
},
},
async execute(args: { uuid: string; needle: string; timeout_ms?: number }, exec) {
exec.signal.throwIfAborted()
if (args.needle === '') {
throw new Error('needle must be a non-empty string')
}
const sessionId = sessionIdOf(exec)
registry.assertOwned(args.uuid, sessionId)
// The registry validates the needle (empty → bad-request) and returns
// the camelCase result the schema above declares directly — no
// field-by-field projection.
const timeoutMs = args.timeout_ms ?? 10_000
const result = await registry.waitFor(args.uuid, args.needle, timeoutMs, exec.signal)
// Project the registry's camelCase result into the snake_case canonical
// value declared by the output schema (the model sees snake_case).
if (result.kind === 'found') {
return { kind: 'found' as const, needle: result.needle, line: result.line, column: result.column, elapsed_ms: result.elapsedMs }
}
if (result.kind === 'timeout') {
return { kind: 'timeout' as const, needle: result.needle, timeout_ms: result.timeoutMs, total_lines: result.totalLines }
}
const out: { kind: 'exited'; needle: string; exit_code?: number; exit_signal?: string } = {
kind: 'exited',
needle: result.needle,
}
if (result.exitCode !== null && result.exitCode !== undefined) out.exit_code = result.exitCode
if (result.exitSignal !== null && result.exitSignal !== undefined) out.exit_signal = result.exitSignal
return out
return await registry.waitFor(args.uuid, args.needle, timeoutMs, exec.signal)
},
}))
ctx.tools.register(defineTool({
register(defineTool({
name: 'terminal_resize',
description:
'Resize an agent terminal\'s pty ( cols × rows ). The host clamps both to a 2..1024 sane range. '
+ 'Most shells redraw their prompt and any full-screen TUI on the next output frame. '
+ 'No-op if the terminal has exited.',
+ 'No-op if the terminal has exited. Returns the dimensions actually applied.',
parameters: {
uuid: { type: 'string', required: true, description: 'Terminal uuid from terminal_create or terminal_list.' },
cols: { type: 'integer', required: true, description: 'New column count ( clamped to 2..1024 ).' },
@@ -387,15 +392,17 @@ export function registerTools(
},
execute: (args: { uuid: string; cols: number; rows: number }, exec) => {
exec.signal.throwIfAborted()
registry.resize(args.uuid, args.cols, args.rows)
return Promise.resolve({ uuid: args.uuid, cols: args.cols, rows: args.rows })
const sessionId = sessionIdOf(exec)
registry.assertOwned(args.uuid, sessionId)
const dims = registry.resize(args.uuid, args.cols, args.rows)
return Promise.resolve({ uuid: args.uuid, ...dims })
},
}))
ctx.tools.register(defineTool({
register(defineTool({
name: 'terminal_signal',
description:
'Send a POSIX signal to an agent terminal\'s foreground process — this is how you send Ctrl+C, Ctrl+D, Ctrl+Z, etc. '
'Send a POSIX signal to an agent terminal\'s foreground process — this is how you send Ctrl+C, Ctrl+Z, etc. '
+ 'Use signal="SIGINT" for Ctrl+C (interrupt the running command), signal="SIGTERM" to request termination, '
+ 'signal="SIGKILL" to force-kill the pty, signal="SIGHUP" to hang up (many shells exit), signal="SIGTSTP" for Ctrl+Z (suspend). '
+ 'Do NOT try to send control characters (like "\\u0003") through terminal_send — use this tool instead. '
@@ -425,12 +432,14 @@ export function registerTools(
},
execute: (args: { uuid: string; signal: AgentTerminalSignal }, exec) => {
exec.signal.throwIfAborted()
const sessionId = sessionIdOf(exec)
registry.assertOwned(args.uuid, sessionId)
registry.signal(args.uuid, args.signal)
return Promise.resolve({ uuid: args.uuid, signal: args.signal })
},
}))
ctx.tools.register(defineTool({
register(defineTool({
name: 'terminal_close',
description:
'Close an agent terminal and release its process. The uuid becomes invalid for all subsequent tool calls. '
@@ -455,8 +464,14 @@ export function registerTools(
},
execute: (args: { uuid: string }, exec) => {
exec.signal.throwIfAborted()
const sessionId = sessionIdOf(exec)
registry.assertOwned(args.uuid, sessionId)
const closed = registry.close(args.uuid)
return Promise.resolve({ uuid: args.uuid, closed })
},
}))
return () => {
for (const dispose of disposers) dispose()
}
}