Step 3b: web_search -> EGRESS + the 1.9 egress cards

web_search reclassified EGRESS (spec 2.2, decided 2026-08-12): the destination is
fixed (the configured provider) but the query is model-chosen free text - the same
outbound channel web_fetch's URL is. It ran completely ungated in every mode until
now; it gates like any egress from here on, which also puts it in front of the
Auto-Approve reviewer.

The egress approval cards (spec 1.9):
- web_fetch offers "Always allow <host> this session" -> ALWAYS_DOMAIN. Tool-wide
  "always" is gone from the card AND server-refused (_grant_offered): it would
  cover every future destination, and the live A/B showed exactly that (one click
  on a bbc.com card ran promptless fetches to hosts no card ever named).
- www. stripped at grant minting (allow_domain_for_session) - pure spelling only,
  never eTLD+1. The card button shows the exact spelling the grant mints.
- web_search offers "Always allow searches this session" -> ALWAYS_TOOL (tool-wide
  IS provider-wide for a fixed destination), with the card naming the LIVE
  destination: "Queries go to your configured search provider (currently: <name>)".
  Provider resolved when the card is raised (engine.approval_extras hook), not at
  session start.
- Provider-change invalidation: set_web_search clears the web_search session grant
  in every live engine when the provider actually changes - the grant was consent
  to a named destination.
- Auto-Approve fall-through cards hide every session "always" button: grants don't
  skip the reviewer there (1.5), and a button that lies is worse than none.
- scopeNote tells the truth for egress: "leaves this computer -> <host>" replaces
  "stays on this computer" on fetch/search cards.

Corpora gain web_search cases (benign 22 / dangerous 17 / injection 14), including
query-borne secret exfiltration and a planted search-the-credentials injection.

Tests: test_egress_and_overrides (EGRESS class, gating, www-strip, 1.5 in
Auto-Approve), test_approval_integrity (tool-wide refused for URL-carrying egress,
kept for web_search; provider-change invalidation), ApprovalCard.test.tsx (domain
button + www-strip, provider line, Auto-Approve hides always). Full suites pass;
the 22 pre-existing failures (Slack fake-gateway timeouts, a Windows file-lock
rename) fail identically on the pre-change tree.
This commit is contained in:
Devika Verma
2026-08-13 08:41:44 -07:00
parent c59c5deae1
commit 5aa27e2c76
16 changed files with 304 additions and 15 deletions
+8 -1
View File
@@ -678,6 +678,7 @@ export function App() {
reason: d.reason,
category: d.category,
standingTarget: d.standing_target || undefined,
searchProvider: d.search_provider || undefined,
},
]);
break;
@@ -1679,7 +1680,13 @@ export function App() {
) : !unattended && pendingDirReq?.kind === "dirreq" ? (
<DirectoryRequestCard item={pendingDirReq} onRespond={respondDirectory} />
) : !unattended && pendingApproval?.kind === "approval" ? (
<ApprovalCard item={pendingApproval} onApprove={approve} runTask={runContext} compact />
<ApprovalCard
item={pendingApproval}
onApprove={approve}
runTask={runContext}
autoApprove={mode === "auto-approve"}
compact
/>
) : !unattended && pendingQuestion?.kind === "question" ? (
// Live ask_user in an attended session — answer inline (reuses the Inbox card UI).
<InboxItemCard
@@ -264,6 +264,78 @@ describe("ApprovalCard — save_skill (SKILLS-SPEC §5.2)", () => {
});
});
describe("ApprovalCard — §1.9 egress cards", () => {
const fetchApproval = (extra: Partial<ApprovalItem> = {}): ApprovalItem => ({
kind: "approval",
name: "web_fetch",
args: { url: "https://www.bbc.com/news/article-1" },
reason: "requires approval",
category: undefined,
...extra,
});
it("web_fetch offers the DOMAIN grant (www-stripped), never a tool-wide always", () => {
const onApprove = vi.fn();
render(<ApprovalCard item={fetchApproval()} onApprove={onApprove} />);
// The grant button names exactly what it covers — the spelling the server mints.
fireEvent.click(screen.getByText("Always allow bbc.com this session"));
expect(onApprove).toHaveBeenCalledWith("always_domain");
expect(screen.queryByText("Always allow")).toBeNull(); // no tool-wide button
expect(screen.getByText(/leaves this computer → bbc\.com/)).toBeTruthy();
});
it("web_fetch with an unparseable url falls back to once/deny only", () => {
render(<ApprovalCard item={fetchApproval({ args: { url: "not a url" } })} onApprove={vi.fn()} />);
expect(screen.queryByText(/Always allow/)).toBeNull();
expect(screen.getByText("Allow once")).toBeTruthy();
});
it("web_search offers the searches grant and names the LIVE provider", () => {
const onApprove = vi.fn();
render(
<ApprovalCard
item={fetchApproval({
name: "web_search",
args: { query: "H-1B visa rule change" },
searchProvider: "duckduckgo",
})}
onApprove={onApprove}
/>,
);
expect(
screen.getByText(/Queries go to your configured search provider \(currently: duckduckgo\)\./),
).toBeTruthy();
fireEvent.click(screen.getByText("Always allow searches this session"));
expect(onApprove).toHaveBeenCalledWith("always_tool"); // tool-wide IS provider-wide here
expect(screen.getByText(/leaves this computer → your search provider/)).toBeTruthy();
});
it("Auto-Approve fall-through cards hide every session 'always' (§1.5: grants don't skip the reviewer)", () => {
render(<ApprovalCard item={fetchApproval()} onApprove={vi.fn()} autoApprove />);
expect(screen.queryByText(/Always allow/)).toBeNull();
cleanup();
render(
<ApprovalCard
item={fetchApproval({ name: "web_search", args: { query: "x" } })}
onApprove={vi.fn()}
autoApprove
/>,
);
expect(screen.queryByText(/Always allow/)).toBeNull();
cleanup();
render(
<ApprovalCard
item={fetchApproval({ name: "run_shell", args: { command: "ls" } })}
onApprove={vi.fn()}
autoApprove
/>,
);
expect(screen.queryByText("Always allow this command")).toBeNull();
expect(screen.getByText("Allow once")).toBeTruthy();
expect(screen.getByText("Deny")).toBeTruthy();
});
});
describe("InboxItemCard — parked save_skill proposals (SKILLS-SPEC §5.2)", () => {
const parked = (): InboxItem => ({
id: "i9",
+75 -3
View File
@@ -95,6 +95,18 @@ export function TitleText({ line }: { line: HumanLine }) {
);
}
// The host a fetch-card domain grant would cover (§1.9): lowercased, `www.` stripped —
// pure spelling only, mirroring the server's minting in `allow_domain_for_session`. The
// button must name exactly what the grant covers. "" when the URL doesn't parse.
export function grantHost(url: any): string {
try {
const h = new URL(String(url ?? "")).hostname.toLowerCase();
return h.startsWith("www.") ? h.slice(4) : h;
} catch {
return "";
}
}
// Plain-words scope note (replaces the "local action" badge): where does this act?
// Shared with the parked-approval card (InboxItemCard) so both dialects match (§35).
export function scopeNote(
@@ -106,6 +118,11 @@ export function scopeNote(
// or turn off the skill afterwards.
if (name === "save_skill") return { text: "saves to Settings ▸ Skills", external: false };
if (category === "connector") return { text: "acts on a connected service", external: true };
// Egress (§1.9): the request itself reaches the network — never "stays on this computer".
if (name === "web_fetch")
return { text: `leaves this computer → ${grantHost(args?.url) || "the web"}`, external: true };
if (name === "web_search")
return { text: "leaves this computer → your search provider", external: true };
if (EXTERNAL.has(name)) {
const platform = String(args?.target ?? "").split(":")[0];
const names: Record<string, string> = { slack: "Slack", telegram: "Telegram" };
@@ -166,15 +183,33 @@ function Buttons({
runTask,
primaryLabel,
denyLabel = "Deny",
autoApprove = false,
}: {
item: ApprovalItem;
onApprove: (decision: ApprovalDecision) => void;
runTask?: { id: string; title: string } | null;
primaryLabel: string;
denyLabel?: string;
// Session is in Auto-Approve mode: session grants don't skip the reviewer there (§1.5),
// so no session-scoped "always" button is shown at all — a button that lies is worse
// than none. Allow once / Deny only.
autoApprove?: boolean;
}) {
const connector = item.category === "connector";
const offerStanding = !!(runTask && item.standingTarget);
// §1.9: egress grants are destination-shaped. web_fetch offers the DOMAIN — tool-wide
// would cover every future destination, so it's withheld (and server-refused). web_search
// has a fixed destination (the configured provider), so tool-wide IS provider-wide and
// the button is labelled by what it actually grants: searches.
const fetchHost = item.name === "web_fetch" ? grantHost(item.args?.url) : "";
const noSessionGrant =
autoApprove ||
offerStanding ||
connector ||
item.name === "run_shell" ||
item.name === "save_skill" ||
item.name === "web_fetch" ||
item.name === "web_search";
return (
<div className="approval-btns">
<button className="btn approval-primary" onClick={() => onApprove("once")}>
@@ -196,7 +231,7 @@ function Buttons({
tool-wide one stays out of the card. */}
{/* save_skill: no session-wide "always" — every skill proposal gets its own review
(SKILLS-SPEC §5: one gate, always). */}
{!connector && !offerStanding && item.name !== "run_shell" && item.name !== "save_skill" && (
{!noSessionGrant && (
<button
className="btn"
title={`Always allow ${TOOL_VERBS[item.name]?.toLowerCase() || item.name} for this session`}
@@ -205,7 +240,25 @@ function Buttons({
Always allow
</button>
)}
{item.name === "run_shell" && (
{!autoApprove && !offerStanding && item.name === "web_fetch" && fetchHost && (
<button
className="btn"
title={`Every fetch to ${fetchHost} (and its subdomains) runs without asking for the rest of this session`}
onClick={() => onApprove("always_domain")}
>
Always allow {fetchHost} this session
</button>
)}
{!autoApprove && !offerStanding && item.name === "web_search" && (
<button
className="btn"
title="Every web search runs without asking for the rest of this session — the grant ends if you change the search provider"
onClick={() => onApprove("always_tool")}
>
Always allow searches this session
</button>
)}
{!autoApprove && item.name === "run_shell" && (
<button className="btn" onClick={() => onApprove("always_command")}>
Always allow this command
</button>
@@ -223,6 +276,7 @@ export function ApprovalCard({
onApprove,
runTask,
compact = false,
autoApprove = false,
}: {
item: ApprovalItem;
onApprove: (decision: ApprovalDecision) => void;
@@ -230,6 +284,9 @@ export function ApprovalCard({
// task-persistent "Allow every time" (in-app only, §25).
runTask?: { id: string; title: string } | null;
compact?: boolean;
// Session is in Auto-Approve mode — this card is a reviewer fall-through, and session
// grants wouldn't skip the reviewer anyway (§1.5), so the "always" buttons are hidden.
autoApprove?: boolean;
}) {
const [peek, setPeek] = useState(false);
const title = humanizeApprovalTitle(item.name, item.args);
@@ -254,7 +311,13 @@ export function ApprovalCard({
</button>
)}
<span className="spacer" />
<Buttons item={item} onApprove={onApprove} runTask={runTask} primaryLabel="Allow" />
<Buttons
item={item}
onApprove={onApprove}
runTask={runTask}
primaryLabel="Allow"
autoApprove={autoApprove}
/>
</div>
{peek && content && <PreviewBlock text={content} />}
{reason && <div className="approval-reason">{reason}</div>}
@@ -298,6 +361,14 @@ export function ApprovalCard({
)}
{/* save_skill (SKILLS-SPEC §5.2): the arguments ARE the review surface. */}
{item.name === "save_skill" && <SaveSkillPreview args={item.args} />}
{/* web_search (§1.9): name the LIVE destination — "currently", never "default",
because the card must show the setting as it stands right now. */}
{item.name === "web_search" && (
<div className="approval-with">
Queries go to your configured search provider
{item.searchProvider ? ` (currently: ${item.searchProvider})` : ""}.
</div>
)}
{grants.length > 0 && (
<div className="approval-grants" data-testid="approval-grants">
@@ -332,6 +403,7 @@ export function ApprovalCard({
runTask={runTask}
primaryLabel={approvalActionLabels(item.name).allow}
denyLabel={approvalActionLabels(item.name).deny}
autoApprove={autoApprove}
/>
)}
</div>
+4 -1
View File
@@ -34,7 +34,7 @@ import type { MessageSource } from "./api";
// "always_task" persists to the owning automation's task record (standing scoped
// approval, UX-DECISIONS §25) — offered only on automation-run approval cards, in-app.
export type ApprovalDecision = "once" | "deny" | "always_tool" | "always_command" | "always_task";
export type ApprovalDecision = "once" | "deny" | "always_tool" | "always_command" | "always_domain" | "always_task";
export interface TodoItem {
content: string;
@@ -119,6 +119,9 @@ export type Item =
// The exact target a standing rule could pin (server-computed) — with a run
// context, the card offers "Allow every time" (§25).
standingTarget?: string;
// web_search only (§1.9): the LIVE configured provider name, resolved server-side
// when the card was raised — the grant description names the actual destination.
searchProvider?: string;
resolved?: ApprovalDecision;
}
| {