mirror of
https://github.com/danny-avila/LibreChat.git
synced 2026-11-16 13:00:55 +00:00
👯 fix: Reject Duplicate Tool Approval Identities (#15933)
* fix: Reject aliased tool approval identities * fix: Keep approval validation in TypeScript * style: Format approval validation
This commit is contained in:
parent
16faadb998
commit
b1f6f6df2f
6 changed files with 251 additions and 25 deletions
|
|
@ -2370,6 +2370,42 @@ describe('ResumeAgentController (POST /agents/chat/resume)', () => {
|
|||
expect(mockGenerationJobManager.approvals.resolve).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('400 when a persisted approval payload aliases two calls to the same id', async () => {
|
||||
const job = makeToolApprovalJob();
|
||||
job.metadata.pendingAction.payload.action_requests = [
|
||||
{ tool_call_id: 'tc1', name: 'write', arguments: { value: 'hidden' } },
|
||||
{ tool_call_id: 'tc1', name: 'write', arguments: { value: 'visible' } },
|
||||
];
|
||||
job.metadata.pendingAction.payload.review_configs = [
|
||||
{ tool_call_id: 'tc1', action_name: 'write', allowed_decisions: ['approve', 'reject'] },
|
||||
{ tool_call_id: 'tc1', action_name: 'write', allowed_decisions: ['approve', 'reject'] },
|
||||
];
|
||||
mockGenerationJobManager.getJob.mockResolvedValue(job);
|
||||
|
||||
const res = await post(approveBody());
|
||||
|
||||
expect(res.status).toBe(400);
|
||||
expect(res.body.error).toMatch(/invalid tool approval payload/i);
|
||||
expect(mockGenerationJobManager.approvals.resolve).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('400 when the client submits duplicate decisions for one tool-call id', async () => {
|
||||
mockGenerationJobManager.getJob.mockResolvedValue(makeToolApprovalJob());
|
||||
|
||||
const res = await post(
|
||||
approveBody({
|
||||
decisions: [
|
||||
{ tool_call_id: 'tc1', decision: 'approve' },
|
||||
{ tool_call_id: 'tc1', decision: 'reject' },
|
||||
],
|
||||
}),
|
||||
);
|
||||
|
||||
expect(res.status).toBe(400);
|
||||
expect(res.body.error).toMatch(/invalid tool approval decisions/i);
|
||||
expect(mockGenerationJobManager.approvals.resolve).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('403 when a decision is not permitted by the tool policy', async () => {
|
||||
const job = makeToolApprovalJob();
|
||||
// Policy restricts tc1 to reject only; an `approve` must be refused.
|
||||
|
|
|
|||
|
|
@ -13,15 +13,12 @@ const {
|
|||
GenerationJobManager,
|
||||
GENERATION_RECOVERY_FAILED_ERROR,
|
||||
isPendingActionStale,
|
||||
mapToolApprovalResolutions,
|
||||
resolveToolApprovalResume,
|
||||
resolveAskUserQuestionResume,
|
||||
buildResolvedAskUserQuestion,
|
||||
appendResolvedAskUserQuestion,
|
||||
attachAskUserQuestionAnswers,
|
||||
findAskUserQuestionContentIndex,
|
||||
findUndecidedToolCalls,
|
||||
findDisallowedDecisions,
|
||||
findIncompleteDecisions,
|
||||
computeAgentRequestFingerprint,
|
||||
computeLegacyAgentRequestFingerprint,
|
||||
captureAgentCheckpointGeneration,
|
||||
|
|
@ -351,27 +348,7 @@ function resolveResumeValue(pendingAction, body) {
|
|||
const payload = pendingAction.payload;
|
||||
if (payload?.type === 'tool_approval') {
|
||||
const resolutions = Array.isArray(body.decisions) ? body.decisions : [];
|
||||
const undecided = findUndecidedToolCalls(payload, resolutions);
|
||||
if (undecided.length > 0) {
|
||||
return { status: 400, error: 'Every paused tool call must be decided', undecided };
|
||||
}
|
||||
// Enforce the policy's per-tool allowed_decisions — a crafted POST must not
|
||||
// approve a tool the policy restricted to (e.g.) reject/respond.
|
||||
const disallowed = findDisallowedDecisions(payload, resolutions);
|
||||
if (disallowed.length > 0) {
|
||||
return { status: 403, error: 'Decision not permitted for one or more tools', disallowed };
|
||||
}
|
||||
// `edit`/`respond` must carry their payload — otherwise toSdkDecision's defensive
|
||||
// defaults ({} / '') would resume with an empty input/result the user didn't approve.
|
||||
const incomplete = findIncompleteDecisions(resolutions);
|
||||
if (incomplete.length > 0) {
|
||||
return {
|
||||
status: 400,
|
||||
error: 'edit requires editedArguments and respond requires responseText',
|
||||
incomplete,
|
||||
};
|
||||
}
|
||||
return { resumeValue: mapToolApprovalResolutions(resolutions) };
|
||||
return resolveToolApprovalResume(payload, resolutions);
|
||||
}
|
||||
if (payload?.type === 'ask_user_question') {
|
||||
return resolveAskUserQuestionResume(payload, body);
|
||||
|
|
|
|||
|
|
@ -9,6 +9,7 @@ import {
|
|||
buildToolApprovalPayload,
|
||||
buildAskUserQuestionPayload,
|
||||
buildPendingAction,
|
||||
isToolApprovalPayloadValid,
|
||||
toClientPendingAction,
|
||||
computeAgentRequestFingerprint,
|
||||
computeLegacyAgentRequestFingerprint,
|
||||
|
|
@ -310,6 +311,54 @@ describe('buildPendingAction', () => {
|
|||
expect(typeof action.createdAt).toBe('number');
|
||||
});
|
||||
|
||||
test('rejects duplicate tool-call ids before the approval is persisted', () => {
|
||||
const duplicatePayload: Agents.ToolApprovalInterruptPayload = {
|
||||
type: 'tool_approval',
|
||||
action_requests: [
|
||||
{ name: 'shell', arguments: { command: 'rm marker' }, tool_call_id: 'duplicate' },
|
||||
{ name: 'shell', arguments: { command: 'ls' }, tool_call_id: 'duplicate' },
|
||||
],
|
||||
review_configs: [
|
||||
{
|
||||
action_name: 'shell',
|
||||
tool_call_id: 'duplicate',
|
||||
allowed_decisions: ['approve', 'reject'],
|
||||
},
|
||||
{
|
||||
action_name: 'shell',
|
||||
tool_call_id: 'duplicate',
|
||||
allowed_decisions: ['approve', 'reject'],
|
||||
},
|
||||
],
|
||||
};
|
||||
|
||||
expect(isToolApprovalPayloadValid(duplicatePayload)).toBe(false);
|
||||
expect(() => buildPendingAction(duplicatePayload, ctx)).toThrow(
|
||||
'Invalid tool approval payload',
|
||||
);
|
||||
});
|
||||
|
||||
test('rejects review policies that do not map one-to-one to the requested calls', () => {
|
||||
const mismatchedPayload: Agents.ToolApprovalInterruptPayload = {
|
||||
type: 'tool_approval',
|
||||
action_requests: [
|
||||
{ name: 'read_file', arguments: { path: 'safe.txt' }, tool_call_id: 'call-1' },
|
||||
],
|
||||
review_configs: [
|
||||
{
|
||||
action_name: 'read_file',
|
||||
tool_call_id: 'call-2',
|
||||
allowed_decisions: ['approve', 'reject'],
|
||||
},
|
||||
],
|
||||
};
|
||||
|
||||
expect(isToolApprovalPayloadValid(mismatchedPayload)).toBe(false);
|
||||
expect(() => buildPendingAction(mismatchedPayload, ctx)).toThrow(
|
||||
'Invalid tool approval payload',
|
||||
);
|
||||
});
|
||||
|
||||
test('wraps an ask_user_question payload with the same envelope', () => {
|
||||
const askPayload: Agents.AskUserQuestionInterruptPayload = {
|
||||
type: 'ask_user_question',
|
||||
|
|
|
|||
|
|
@ -315,6 +315,42 @@ export function buildToolApprovalPayload(
|
|||
};
|
||||
}
|
||||
|
||||
/** Require one uniquely identified review policy for every paused tool call. */
|
||||
export function isToolApprovalPayloadValid(payload: Agents.ToolApprovalInterruptPayload): boolean {
|
||||
if (payload.action_requests.length !== payload.review_configs.length) {
|
||||
return false;
|
||||
}
|
||||
|
||||
const requestedIds = new Set<string>();
|
||||
for (const request of payload.action_requests) {
|
||||
if (
|
||||
typeof request.tool_call_id !== 'string' ||
|
||||
request.tool_call_id.length === 0 ||
|
||||
requestedIds.has(request.tool_call_id)
|
||||
) {
|
||||
return false;
|
||||
}
|
||||
requestedIds.add(request.tool_call_id);
|
||||
}
|
||||
|
||||
const reviewedIds = new Set<string>();
|
||||
for (const config of payload.review_configs) {
|
||||
if (
|
||||
typeof config.tool_call_id !== 'string' ||
|
||||
config.tool_call_id.length === 0 ||
|
||||
reviewedIds.has(config.tool_call_id)
|
||||
) {
|
||||
return false;
|
||||
}
|
||||
if (!requestedIds.has(config.tool_call_id)) {
|
||||
return false;
|
||||
}
|
||||
reviewedIds.add(config.tool_call_id);
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/** Build an ask-user-question interrupt payload. */
|
||||
export function buildAskUserQuestionPayload(
|
||||
question: Agents.AskUserQuestionRequest,
|
||||
|
|
@ -792,6 +828,9 @@ export function buildPendingAction(
|
|||
payload: Agents.HumanInterruptPayload,
|
||||
ctx: PendingActionContext,
|
||||
): Agents.PendingAction {
|
||||
if (payload.type === 'tool_approval' && !isToolApprovalPayloadValid(payload)) {
|
||||
throw new Error('Invalid tool approval payload');
|
||||
}
|
||||
const createdAt = Date.now();
|
||||
const ttlExpiresAt = typeof ctx.ttlMs === 'number' ? createdAt + ctx.ttlMs : undefined;
|
||||
let absoluteExpiresAt: number | undefined;
|
||||
|
|
|
|||
|
|
@ -14,6 +14,8 @@ import {
|
|||
serializeAskUserAnswerVariants,
|
||||
resolveAskUserQuestionResume,
|
||||
findUndecidedToolCalls,
|
||||
hasInvalidToolApprovalResolutions,
|
||||
resolveToolApprovalResume,
|
||||
findDisallowedDecisions,
|
||||
findIncompleteDecisions,
|
||||
createContentIndexOffsetHandlers,
|
||||
|
|
@ -230,6 +232,79 @@ describe('findUndecidedToolCalls', () => {
|
|||
});
|
||||
});
|
||||
|
||||
describe('hasInvalidToolApprovalResolutions', () => {
|
||||
const payload: Agents.ToolApprovalInterruptPayload = {
|
||||
type: 'tool_approval',
|
||||
action_requests: [
|
||||
{ tool_call_id: 'a', name: 'read', arguments: {} },
|
||||
{ tool_call_id: 'b', name: 'write', arguments: {} },
|
||||
],
|
||||
review_configs: [],
|
||||
};
|
||||
|
||||
test('rejects duplicate decisions for one tool-call id', () => {
|
||||
expect(
|
||||
hasInvalidToolApprovalResolutions(payload, [
|
||||
{ tool_call_id: 'a', decision: 'approve' },
|
||||
{ tool_call_id: 'a', decision: 'reject' },
|
||||
]),
|
||||
).toBe(true);
|
||||
});
|
||||
|
||||
test('rejects decisions for a tool call outside the pending batch', () => {
|
||||
expect(
|
||||
hasInvalidToolApprovalResolutions(payload, [
|
||||
{ tool_call_id: 'a', decision: 'approve' },
|
||||
{ tool_call_id: 'foreign', decision: 'approve' },
|
||||
]),
|
||||
).toBe(true);
|
||||
});
|
||||
|
||||
test('accepts one decision for each requested tool call', () => {
|
||||
expect(
|
||||
hasInvalidToolApprovalResolutions(payload, [
|
||||
{ tool_call_id: 'a', decision: 'approve' },
|
||||
{ tool_call_id: 'b', decision: 'reject' },
|
||||
]),
|
||||
).toBe(false);
|
||||
});
|
||||
});
|
||||
|
||||
describe('resolveToolApprovalResume', () => {
|
||||
const payload: Agents.ToolApprovalInterruptPayload = {
|
||||
type: 'tool_approval',
|
||||
action_requests: [
|
||||
{ tool_call_id: 'a', name: 'read', arguments: {} },
|
||||
{ tool_call_id: 'b', name: 'write', arguments: {} },
|
||||
],
|
||||
review_configs: [
|
||||
{ tool_call_id: 'a', action_name: 'read', allowed_decisions: ['approve', 'reject'] },
|
||||
{ tool_call_id: 'b', action_name: 'write', allowed_decisions: ['approve', 'reject'] },
|
||||
],
|
||||
};
|
||||
|
||||
test('fails closed when a persisted payload contains duplicate call identities', () => {
|
||||
const aliased = {
|
||||
...payload,
|
||||
action_requests: [payload.action_requests[0], payload.action_requests[0]],
|
||||
review_configs: [payload.review_configs[0], payload.review_configs[0]],
|
||||
};
|
||||
|
||||
expect(
|
||||
resolveToolApprovalResume(aliased, [{ tool_call_id: 'a', decision: 'approve' }]),
|
||||
).toEqual({ status: 400, error: 'Invalid tool approval payload' });
|
||||
});
|
||||
|
||||
test('returns the SDK decision map only for a complete unambiguous batch', () => {
|
||||
expect(
|
||||
resolveToolApprovalResume(payload, [
|
||||
{ tool_call_id: 'a', decision: 'approve' },
|
||||
{ tool_call_id: 'b', decision: 'reject' },
|
||||
]),
|
||||
).toEqual({ resumeValue: { a: { type: 'approve' }, b: { type: 'reject' } } });
|
||||
});
|
||||
});
|
||||
|
||||
describe('findDisallowedDecisions', () => {
|
||||
const payload: Agents.ToolApprovalInterruptPayload = {
|
||||
type: 'tool_approval',
|
||||
|
|
|
|||
|
|
@ -9,6 +9,7 @@ import type {
|
|||
} from '@librechat/agents';
|
||||
import type { Agents } from 'librechat-data-provider';
|
||||
import { ASK_USER_QUESTION_TOOL_NAME } from './askUserQuestionTool';
|
||||
import { isToolApprovalPayloadValid } from './policy';
|
||||
|
||||
/**
|
||||
* Translate the host-facing approval wire format into the SDK's resume value.
|
||||
|
|
@ -266,6 +267,22 @@ export function findUndecidedToolCalls(
|
|||
return payload.action_requests.map((a) => a.tool_call_id).filter((id) => !decided.has(id));
|
||||
}
|
||||
|
||||
/** Reject ambiguous or foreign decisions before adapting them to the SDK's ID-keyed map. */
|
||||
export function hasInvalidToolApprovalResolutions(
|
||||
payload: Agents.ToolApprovalInterruptPayload,
|
||||
resolutions: readonly Agents.ToolApprovalResolution[],
|
||||
): boolean {
|
||||
const requestedIds = new Set(payload.action_requests.map((request) => request.tool_call_id));
|
||||
const resolutionIds = new Set<string>();
|
||||
for (const resolution of resolutions) {
|
||||
if (resolutionIds.has(resolution.tool_call_id) || !requestedIds.has(resolution.tool_call_id)) {
|
||||
return true;
|
||||
}
|
||||
resolutionIds.add(resolution.tool_call_id);
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
/**
|
||||
* Enforce the policy's per-tool `allowed_decisions`. Returns the `tool_call_id`s
|
||||
* whose submitted decision is NOT one the interrupt's `review_configs` permits for
|
||||
|
|
@ -317,6 +334,39 @@ export function findIncompleteDecisions(
|
|||
.map((r) => r.tool_call_id);
|
||||
}
|
||||
|
||||
/** Validate and translate one complete tool-approval batch for the resume controller. */
|
||||
export function resolveToolApprovalResume(
|
||||
payload: Agents.ToolApprovalInterruptPayload,
|
||||
resolutions: readonly Agents.ToolApprovalResolution[],
|
||||
):
|
||||
| { resumeValue: ToolApprovalDecisionMap }
|
||||
| { status: 400; error: string; undecided?: string[]; incomplete?: string[] }
|
||||
| { status: 403; error: string; disallowed: string[] } {
|
||||
if (!isToolApprovalPayloadValid(payload)) {
|
||||
return { status: 400, error: 'Invalid tool approval payload' };
|
||||
}
|
||||
if (hasInvalidToolApprovalResolutions(payload, resolutions)) {
|
||||
return { status: 400, error: 'Invalid tool approval decisions' };
|
||||
}
|
||||
const undecided = findUndecidedToolCalls(payload, resolutions);
|
||||
if (undecided.length > 0) {
|
||||
return { status: 400, error: 'Every paused tool call must be decided', undecided };
|
||||
}
|
||||
const disallowed = findDisallowedDecisions(payload, resolutions);
|
||||
if (disallowed.length > 0) {
|
||||
return { status: 403, error: 'Decision not permitted for one or more tools', disallowed };
|
||||
}
|
||||
const incomplete = findIncompleteDecisions(resolutions);
|
||||
if (incomplete.length > 0) {
|
||||
return {
|
||||
status: 400,
|
||||
error: 'edit requires editedArguments and respond requires responseText',
|
||||
incomplete,
|
||||
};
|
||||
}
|
||||
return { resumeValue: mapToolApprovalResolutions(resolutions) };
|
||||
}
|
||||
|
||||
/**
|
||||
* Reconcile persisted tool-step indices with the content being seeded into a
|
||||
* rebuilt aggregator.
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue