@vellumai/assistant 0.11.2-dev.202608071521.2d2f93e → 0.11.2-dev.202608071719.8f7a9f1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/package.json +1 -1
- package/src/__tests__/conversation-process-app-control-preactivation.test.ts +123 -0
- package/src/__tests__/conversation-queue.test.ts +92 -0
- package/src/__tests__/oauth-provider-seed-logos.test.ts +65 -0
- package/src/__tests__/tool-grant-request-escalation.test.ts +40 -1
- package/src/__tests__/trusted-contact-inline-approval-integration.test.ts +68 -1
- package/src/approvals/guardian-request-resolvers.ts +6 -3
- package/src/config/schemas/timeouts.ts +1 -1
- package/src/daemon/conversation-agent-loop.ts +13 -3
- package/src/daemon/conversation-messaging.ts +21 -1
- package/src/daemon/conversation-process.ts +37 -3
- package/src/daemon/conversation-queue-manager.ts +10 -0
- package/src/daemon/conversation.ts +6 -0
- package/src/daemon/trust-context-types.ts +39 -0
- package/src/inbound/__tests__/oauth-callback-url.test.ts +83 -0
- package/src/inbound/oauth-callback-url.ts +61 -0
- package/src/mcp/__tests__/mcp-auth-orchestrator.test.ts +0 -1
- package/src/mcp/__tests__/mcp-oauth-provider.test.ts +6 -9
- package/src/mcp/client.ts +4 -7
- package/src/mcp/mcp-auth-orchestrator.ts +8 -20
- package/src/mcp/mcp-oauth-provider.ts +49 -233
- package/src/oauth/AGENTS.md +4 -2
- package/src/oauth/seed-providers.ts +6 -3
- package/src/security/oauth2.ts +6 -22
- package/src/subagent/manager.ts +9 -2
- package/src/tools/execution-timeout.ts +9 -2
- package/src/tools/tool-approval-handler.ts +45 -4
package/package.json
CHANGED
|
@@ -86,6 +86,7 @@ import {
|
|
|
86
86
|
MessageQueue,
|
|
87
87
|
type QueuedMessage,
|
|
88
88
|
} from "../daemon/conversation-queue-manager.js";
|
|
89
|
+
import type { TrustContext } from "../daemon/trust-context-types.js";
|
|
89
90
|
|
|
90
91
|
// ---------------------------------------------------------------------------
|
|
91
92
|
// Fake context — captures preactivation calls, satisfies the bare minimum
|
|
@@ -172,6 +173,7 @@ function makeQueuedMessage(opts: {
|
|
|
172
173
|
content?: string;
|
|
173
174
|
turnInterfaceContext?: TurnInterfaceContext;
|
|
174
175
|
sourceActorPrincipalId?: string;
|
|
176
|
+
trustContext?: TrustContext;
|
|
175
177
|
}): QueuedMessage {
|
|
176
178
|
return {
|
|
177
179
|
content: opts.content ?? "follow up",
|
|
@@ -185,6 +187,7 @@ function makeQueuedMessage(opts: {
|
|
|
185
187
|
authContext: opts.sourceActorPrincipalId
|
|
186
188
|
? ({ actorPrincipalId: opts.sourceActorPrincipalId } as never)
|
|
187
189
|
: undefined,
|
|
190
|
+
trustContext: opts.trustContext,
|
|
188
191
|
};
|
|
189
192
|
}
|
|
190
193
|
|
|
@@ -371,6 +374,126 @@ describe("drainQueue preactivation re-add for host-proxy interfaces", () => {
|
|
|
371
374
|
expect(ctx.currentTurnSourceActorPrincipalId).toBe("trusted-contact-user");
|
|
372
375
|
});
|
|
373
376
|
|
|
377
|
+
test("drainSingleMessage runs the turn under the queued sender's trust, not the live slot", async () => {
|
|
378
|
+
// The conversation-level slot holds whichever actor sent most recently.
|
|
379
|
+
// A message that waited while someone else sent must still run as its own
|
|
380
|
+
// sender: trust decides `trustClass`, `executionChannel`, and
|
|
381
|
+
// `requesterExternalUserId`, so reading the slot hands the whole
|
|
382
|
+
// tool-approval path the wrong identity in both directions -- a contact
|
|
383
|
+
// inheriting guardian self-approval, or a guardian's own call escalating
|
|
384
|
+
// back to her as a contact's grant request.
|
|
385
|
+
const contactTrust: TrustContext = {
|
|
386
|
+
trustClass: "trusted_contact",
|
|
387
|
+
sourceChannel: "slack",
|
|
388
|
+
requesterExternalUserId: "U-contact",
|
|
389
|
+
};
|
|
390
|
+
const queue = new MessageQueue();
|
|
391
|
+
queue.push(
|
|
392
|
+
makeQueuedMessage({
|
|
393
|
+
requestId: "req-contact",
|
|
394
|
+
trustContext: contactTrust,
|
|
395
|
+
}),
|
|
396
|
+
);
|
|
397
|
+
const ctx = makeFakeContext({ queue });
|
|
398
|
+
// Someone else sent after this message was queued, moving the slot.
|
|
399
|
+
ctx.trustContext = {
|
|
400
|
+
trustClass: "guardian",
|
|
401
|
+
sourceChannel: "vellum",
|
|
402
|
+
requesterExternalUserId: "guardian-principal",
|
|
403
|
+
};
|
|
404
|
+
|
|
405
|
+
await drainQueue(ctx);
|
|
406
|
+
|
|
407
|
+
expect(ctx.currentTurnTrustContext?.trustClass).toBe("trusted_contact");
|
|
408
|
+
expect(ctx.currentTurnTrustContext?.requesterExternalUserId).toBe(
|
|
409
|
+
"U-contact",
|
|
410
|
+
);
|
|
411
|
+
expect(ctx.currentTurnTrustContext?.sourceChannel).toBe("slack");
|
|
412
|
+
// The slot itself is left alone; only the turn's view is corrected.
|
|
413
|
+
expect(ctx.trustContext?.trustClass).toBe("guardian");
|
|
414
|
+
});
|
|
415
|
+
|
|
416
|
+
test("buildPassthroughBatch refuses to coalesce two channel senders", async () => {
|
|
417
|
+
// Channel senders carry no principal, so the `sourceActorPrincipalId`
|
|
418
|
+
// boundary sees `undefined === undefined` and would batch two different
|
|
419
|
+
// Slack contacts into one turn running under the head's trust. The batch
|
|
420
|
+
// must split on the sender's trust identity instead.
|
|
421
|
+
const ifCtx: TurnInterfaceContext = {
|
|
422
|
+
userMessageInterface: "web",
|
|
423
|
+
assistantMessageInterface: "web",
|
|
424
|
+
};
|
|
425
|
+
const queue = new MessageQueue();
|
|
426
|
+
queue.push(
|
|
427
|
+
makeQueuedMessage({
|
|
428
|
+
requestId: "req-contact-a",
|
|
429
|
+
content: "from A",
|
|
430
|
+
turnInterfaceContext: ifCtx,
|
|
431
|
+
trustContext: {
|
|
432
|
+
trustClass: "trusted_contact",
|
|
433
|
+
sourceChannel: "slack",
|
|
434
|
+
requesterExternalUserId: "U-alex",
|
|
435
|
+
},
|
|
436
|
+
}),
|
|
437
|
+
);
|
|
438
|
+
queue.push(
|
|
439
|
+
makeQueuedMessage({
|
|
440
|
+
requestId: "req-contact-b",
|
|
441
|
+
content: "from B",
|
|
442
|
+
turnInterfaceContext: ifCtx,
|
|
443
|
+
trustContext: {
|
|
444
|
+
trustClass: "trusted_contact",
|
|
445
|
+
sourceChannel: "slack",
|
|
446
|
+
requesterExternalUserId: "U-blake",
|
|
447
|
+
},
|
|
448
|
+
}),
|
|
449
|
+
);
|
|
450
|
+
const ctx = makeFakeContext({ queue, turnInterfaceContext: ifCtx });
|
|
451
|
+
|
|
452
|
+
await drainQueue(ctx);
|
|
453
|
+
|
|
454
|
+
// B stays queued: it gets its own turn under its own trust.
|
|
455
|
+
expect(queue.length).toBe(1);
|
|
456
|
+
expect(queue.peek(0)?.requestId).toBe("req-contact-b");
|
|
457
|
+
expect(ctx.currentTurnTrustContext?.requesterExternalUserId).toBe("U-alex");
|
|
458
|
+
});
|
|
459
|
+
|
|
460
|
+
test("buildPassthroughBatch still coalesces two messages from the same sender", async () => {
|
|
461
|
+
// Sensitivity check on the split above: identical trust must still batch,
|
|
462
|
+
// otherwise the boundary would be splitting on something incidental and
|
|
463
|
+
// the test above would pass for the wrong reason.
|
|
464
|
+
const ifCtx: TurnInterfaceContext = {
|
|
465
|
+
userMessageInterface: "web",
|
|
466
|
+
assistantMessageInterface: "web",
|
|
467
|
+
};
|
|
468
|
+
const sameTrust: TrustContext = {
|
|
469
|
+
trustClass: "trusted_contact",
|
|
470
|
+
sourceChannel: "slack",
|
|
471
|
+
requesterExternalUserId: "U-alex",
|
|
472
|
+
};
|
|
473
|
+
const queue = new MessageQueue();
|
|
474
|
+
queue.push(
|
|
475
|
+
makeQueuedMessage({
|
|
476
|
+
requestId: "req-a1",
|
|
477
|
+
content: "first",
|
|
478
|
+
turnInterfaceContext: ifCtx,
|
|
479
|
+
trustContext: { ...sameTrust },
|
|
480
|
+
}),
|
|
481
|
+
);
|
|
482
|
+
queue.push(
|
|
483
|
+
makeQueuedMessage({
|
|
484
|
+
requestId: "req-a2",
|
|
485
|
+
content: "second",
|
|
486
|
+
turnInterfaceContext: ifCtx,
|
|
487
|
+
trustContext: { ...sameTrust },
|
|
488
|
+
}),
|
|
489
|
+
);
|
|
490
|
+
const ctx = makeFakeContext({ queue, turnInterfaceContext: ifCtx });
|
|
491
|
+
|
|
492
|
+
await drainQueue(ctx);
|
|
493
|
+
|
|
494
|
+
expect(queue.length).toBe(0);
|
|
495
|
+
});
|
|
496
|
+
|
|
374
497
|
test("drainSingleMessage does NOT re-add 'app-control' for web-sourced message when no capable client is connected", async () => {
|
|
375
498
|
// mockCapabilityClients remains [] (reset by afterEach from prior test)
|
|
376
499
|
const queue = new MessageQueue();
|
|
@@ -556,6 +556,98 @@ describe("Conversation message queue", () => {
|
|
|
556
556
|
await new Promise((r) => setTimeout(r, 10));
|
|
557
557
|
});
|
|
558
558
|
|
|
559
|
+
test("enqueueMessage captures the sender's trust, immune to a later slot change", async () => {
|
|
560
|
+
// Trust must ride with the queued message. The conversation-level slot is
|
|
561
|
+
// rewritten by whoever sends next, so a message that reads it at drain time
|
|
562
|
+
// would run as the wrong actor. Capturing at enqueue is what makes the
|
|
563
|
+
// drain's identity independent of who sent afterwards.
|
|
564
|
+
const conversation = makeConversation();
|
|
565
|
+
await conversation.loadFromDb();
|
|
566
|
+
|
|
567
|
+
conversation.setTrustContext({
|
|
568
|
+
trustClass: "trusted_contact",
|
|
569
|
+
sourceChannel: "slack",
|
|
570
|
+
requesterExternalUserId: "U-contact",
|
|
571
|
+
});
|
|
572
|
+
|
|
573
|
+
const p1 = conversation.processMessage({
|
|
574
|
+
content: "msg-1",
|
|
575
|
+
attachments: [],
|
|
576
|
+
onEvent: () => {},
|
|
577
|
+
requestId: "req-1",
|
|
578
|
+
});
|
|
579
|
+
await waitForPendingRun(1);
|
|
580
|
+
|
|
581
|
+
conversation.enqueueMessage({ content: "msg-2", requestId: "req-2" });
|
|
582
|
+
|
|
583
|
+
// A different actor sends while the message waits, moving the slot.
|
|
584
|
+
conversation.setTrustContext({
|
|
585
|
+
trustClass: "guardian",
|
|
586
|
+
sourceChannel: "vellum",
|
|
587
|
+
requesterExternalUserId: "guardian-principal",
|
|
588
|
+
});
|
|
589
|
+
|
|
590
|
+
const queued = conversation.queue.peek(0);
|
|
591
|
+
expect(queued?.requestId).toBe("req-2");
|
|
592
|
+
expect(queued?.trustContext?.trustClass).toBe("trusted_contact");
|
|
593
|
+
expect(queued?.trustContext?.requesterExternalUserId).toBe("U-contact");
|
|
594
|
+
|
|
595
|
+
await resolveRun(0);
|
|
596
|
+
await p1;
|
|
597
|
+
await new Promise((r) => setTimeout(r, 10));
|
|
598
|
+
});
|
|
599
|
+
|
|
600
|
+
test("a drained turn keeps its sender's trust once the agent loop is running", async () => {
|
|
601
|
+
// Exercises the real drain -> runAgentLoop ordering against a live
|
|
602
|
+
// Conversation. The loop re-initializes the per-turn trust snapshot on
|
|
603
|
+
// entry, so a drain that only stamped the field before calling it would
|
|
604
|
+
// have that stamp silently undone and the turn would execute as whoever
|
|
605
|
+
// sent most recently. Asserting inside the run is the point: checking
|
|
606
|
+
// before the loop starts passes either way.
|
|
607
|
+
const conversation = makeConversation();
|
|
608
|
+
await conversation.loadFromDb();
|
|
609
|
+
|
|
610
|
+
conversation.setTrustContext({
|
|
611
|
+
trustClass: "trusted_contact",
|
|
612
|
+
sourceChannel: "slack",
|
|
613
|
+
requesterExternalUserId: "U-contact",
|
|
614
|
+
});
|
|
615
|
+
|
|
616
|
+
const p1 = conversation.processMessage({
|
|
617
|
+
content: "msg-1",
|
|
618
|
+
attachments: [],
|
|
619
|
+
onEvent: () => {},
|
|
620
|
+
requestId: "req-1",
|
|
621
|
+
});
|
|
622
|
+
await waitForPendingRun(1);
|
|
623
|
+
|
|
624
|
+
conversation.enqueueMessage({ content: "msg-2", requestId: "req-2" });
|
|
625
|
+
|
|
626
|
+
// The guardian sends while the contact's message waits, moving the slot.
|
|
627
|
+
conversation.setTrustContext({
|
|
628
|
+
trustClass: "guardian",
|
|
629
|
+
sourceChannel: "vellum",
|
|
630
|
+
requesterExternalUserId: "guardian-principal",
|
|
631
|
+
});
|
|
632
|
+
|
|
633
|
+
// Finish the first turn so the queue drains into a second run.
|
|
634
|
+
await resolveRun(0);
|
|
635
|
+
await p1;
|
|
636
|
+
await waitForPendingRun(2);
|
|
637
|
+
|
|
638
|
+
// Read while run 2 is in flight: this is what tool setup and the
|
|
639
|
+
// guardian-request producers resolve against.
|
|
640
|
+
expect(conversation.currentTurnTrustContext?.trustClass).toBe(
|
|
641
|
+
"trusted_contact",
|
|
642
|
+
);
|
|
643
|
+
expect(conversation.currentTurnTrustContext?.requesterExternalUserId).toBe(
|
|
644
|
+
"U-contact",
|
|
645
|
+
);
|
|
646
|
+
|
|
647
|
+
await resolveRun(1);
|
|
648
|
+
await new Promise((r) => setTimeout(r, 10));
|
|
649
|
+
});
|
|
650
|
+
|
|
559
651
|
test("[experimental] queued passthrough siblings drain as a single batched run", async () => {
|
|
560
652
|
const conversation = makeConversation();
|
|
561
653
|
await conversation.loadFromDb();
|
|
@@ -1,7 +1,44 @@
|
|
|
1
|
+
import { existsSync, readFileSync } from "node:fs";
|
|
2
|
+
import { dirname, join } from "node:path";
|
|
1
3
|
import { describe, expect, test } from "bun:test";
|
|
2
4
|
|
|
3
5
|
import { PROVIDER_SEED_DATA } from "../oauth/seed-providers.js";
|
|
4
6
|
|
|
7
|
+
const REPO_ROOT = dirname(dirname(dirname(import.meta.dir)));
|
|
8
|
+
const WEB_PUBLIC_DIR = join(REPO_ROOT, "clients/web/public");
|
|
9
|
+
const INTEGRATION_ICON_SOURCE = join(
|
|
10
|
+
REPO_ROOT,
|
|
11
|
+
"clients/web/src/components/integrations/integration-icon.tsx",
|
|
12
|
+
);
|
|
13
|
+
|
|
14
|
+
/**
|
|
15
|
+
* Provider keys the web client draws without a `BUNDLED_LOGO_URLS` entry.
|
|
16
|
+
* Google's mark is multi-colour, so it ships as the `GoogleLogo` component
|
|
17
|
+
* and is returned before the bundled-asset lookup runs.
|
|
18
|
+
*/
|
|
19
|
+
const COMPONENT_RENDERED_PROVIDERS = new Set(["google"]);
|
|
20
|
+
|
|
21
|
+
/** `publicAsset("/images/...")` paths from `BUNDLED_LOGO_URLS`, by key. */
|
|
22
|
+
function readBundledLogoMap(): Map<string, string> {
|
|
23
|
+
const source = readFileSync(INTEGRATION_ICON_SOURCE, "utf8");
|
|
24
|
+
const block = source.match(
|
|
25
|
+
/const BUNDLED_LOGO_URLS: Record<string, string> = \{([\s\S]*?)\n\};/,
|
|
26
|
+
);
|
|
27
|
+
if (!block) {
|
|
28
|
+
throw new Error(
|
|
29
|
+
`Could not find BUNDLED_LOGO_URLS in ${INTEGRATION_ICON_SOURCE}. If it ` +
|
|
30
|
+
`was renamed or restructured, update this test rather than deleting it.`,
|
|
31
|
+
);
|
|
32
|
+
}
|
|
33
|
+
const entries = new Map<string, string>();
|
|
34
|
+
for (const [, key, path] of block[1]!.matchAll(
|
|
35
|
+
/(\w+):\s*publicAsset\("([^"]+)"\)/g,
|
|
36
|
+
)) {
|
|
37
|
+
entries.set(key!, path!);
|
|
38
|
+
}
|
|
39
|
+
return entries;
|
|
40
|
+
}
|
|
41
|
+
|
|
5
42
|
/**
|
|
6
43
|
* Allowed CDN prefixes for the ``logoUrl`` field on ``PROVIDER_SEED_DATA``
|
|
7
44
|
* (``assistant/src/oauth/seed-providers.ts``):
|
|
@@ -41,3 +78,31 @@ describe("PROVIDER_SEED_DATA logo URLs", () => {
|
|
|
41
78
|
expect(invalid).toEqual([]);
|
|
42
79
|
});
|
|
43
80
|
});
|
|
81
|
+
|
|
82
|
+
/**
|
|
83
|
+
* The `logoUrl` above is a fallback for providers registered at runtime. Every
|
|
84
|
+
* provider we seed ourselves must draw from an asset we ship, because an icon
|
|
85
|
+
* CDN can drop a brand at any time and the client then degrades to an initials
|
|
86
|
+
* avatar with nothing in CI to notice: the prefix check above passes just as
|
|
87
|
+
* happily for a URL that 404s. Simple Icons hosts no Microsoft mark and no
|
|
88
|
+
* Slack mark, so the seeded URLs for those brands do not resolve there.
|
|
89
|
+
*/
|
|
90
|
+
describe("PROVIDER_SEED_DATA bundled logo coverage", () => {
|
|
91
|
+
test("every seeded provider renders from a bundled asset", () => {
|
|
92
|
+
const bundled = readBundledLogoMap();
|
|
93
|
+
const uncovered = Object.keys(PROVIDER_SEED_DATA).filter(
|
|
94
|
+
(provider) =>
|
|
95
|
+
!bundled.has(provider) && !COMPONENT_RENDERED_PROVIDERS.has(provider),
|
|
96
|
+
);
|
|
97
|
+
|
|
98
|
+
expect(uncovered).toEqual([]);
|
|
99
|
+
});
|
|
100
|
+
|
|
101
|
+
test("every bundled logo path resolves to a file on disk", () => {
|
|
102
|
+
const missingFiles = [...readBundledLogoMap()]
|
|
103
|
+
.filter(([, path]) => !existsSync(join(WEB_PUBLIC_DIR, path)))
|
|
104
|
+
.map(([provider, path]) => ({ provider, path }));
|
|
105
|
+
|
|
106
|
+
expect(missingFiles).toEqual([]);
|
|
107
|
+
});
|
|
108
|
+
});
|
|
@@ -101,17 +101,20 @@ import {
|
|
|
101
101
|
getRegisteredKinds,
|
|
102
102
|
getResolver,
|
|
103
103
|
} from "../approvals/guardian-request-resolvers.js";
|
|
104
|
+
import { getConfig } from "../config/loader.js";
|
|
104
105
|
import { getDb } from "../persistence/db-connection.js";
|
|
105
106
|
import { getSqlite } from "../persistence/db-connection.js";
|
|
106
107
|
import { initializeDb } from "../persistence/db-init.js";
|
|
107
108
|
import { scopedApprovalGrants } from "../persistence/schema/index.js";
|
|
108
109
|
import {
|
|
110
|
+
TC_GRANT_WAIT_MAX_MS,
|
|
109
111
|
ToolApprovalHandler,
|
|
110
112
|
waitForInlineGrant,
|
|
111
113
|
} from "../tools/tool-approval-handler.js";
|
|
112
114
|
import type { ToolContext } from "../tools/types.js";
|
|
115
|
+
import { setConfig } from "./helpers/set-config.js";
|
|
113
116
|
|
|
114
|
-
/** Short wait config for tests
|
|
117
|
+
/** Short wait config for tests: keeps escalation cases off the real budget. */
|
|
115
118
|
const TEST_INLINE_WAIT_CONFIG = { maxWaitMs: 100, intervalMs: 20 };
|
|
116
119
|
|
|
117
120
|
await initializeDb();
|
|
@@ -490,6 +493,42 @@ describe("inline wait-and-resume", () => {
|
|
|
490
493
|
expect(elapsed).toBeLessThan(500);
|
|
491
494
|
});
|
|
492
495
|
|
|
496
|
+
test("waitForInlineGrant spends timeouts.permissionTimeoutSec when no explicit budget is given", async () => {
|
|
497
|
+
// The guardian's escalation window must track the configured approval
|
|
498
|
+
// budget rather than a compiled-in constant, so an operator raising
|
|
499
|
+
// permissionTimeoutSec actually widens the window a guardian has to
|
|
500
|
+
// answer in. Omitting maxWaitMs is what production does.
|
|
501
|
+
const priorTimeouts = getConfig().timeouts;
|
|
502
|
+
setConfig("timeouts", {
|
|
503
|
+
...priorTimeouts,
|
|
504
|
+
permissionTimeoutSec: 0.15,
|
|
505
|
+
});
|
|
506
|
+
try {
|
|
507
|
+
const req = seedGrantRequest("sha256:configbudget");
|
|
508
|
+
|
|
509
|
+
const start = Date.now();
|
|
510
|
+
const result = await waitForInlineGrant(
|
|
511
|
+
req.id,
|
|
512
|
+
{
|
|
513
|
+
toolName: "bash",
|
|
514
|
+
inputDigest: "sha256:configbudget",
|
|
515
|
+
consumingRequestId: "consume-config-budget",
|
|
516
|
+
},
|
|
517
|
+
{ intervalMs: 20 },
|
|
518
|
+
);
|
|
519
|
+
const elapsed = Date.now() - start;
|
|
520
|
+
|
|
521
|
+
expect(result.outcome).toBe("timeout");
|
|
522
|
+
// Bounded well under TC_GRANT_WAIT_MAX_MS: a regression that ignores
|
|
523
|
+
// config and falls back to the constant blows this ceiling by ~400x.
|
|
524
|
+
expect(elapsed).toBeGreaterThanOrEqual(120);
|
|
525
|
+
expect(elapsed).toBeLessThan(3_000);
|
|
526
|
+
expect(elapsed).toBeLessThan(TC_GRANT_WAIT_MAX_MS);
|
|
527
|
+
} finally {
|
|
528
|
+
setConfig("timeouts", priorTimeouts);
|
|
529
|
+
}
|
|
530
|
+
}, 10_000);
|
|
531
|
+
|
|
493
532
|
test("waitForInlineGrant returns aborted when signal fires during wait", async () => {
|
|
494
533
|
const req = seedGrantRequest("sha256:abortwait");
|
|
495
534
|
|
|
@@ -170,6 +170,7 @@ mock.module("../channels/gateway-guardian-requests.js", () => sim.module);
|
|
|
170
170
|
import { applyGuardianDecision } from "../approvals/guardian-decision-primitive.js";
|
|
171
171
|
import type { ActorContext } from "../approvals/guardian-request-resolvers.js";
|
|
172
172
|
import { getResolver } from "../approvals/guardian-request-resolvers.js";
|
|
173
|
+
import { getConfig } from "../config/loader.js";
|
|
173
174
|
import type { TrustContext } from "../daemon/trust-context-types.js";
|
|
174
175
|
import { getDb } from "../persistence/db-connection.js";
|
|
175
176
|
import { initializeDb } from "../persistence/db-init.js";
|
|
@@ -177,12 +178,14 @@ import { scopedApprovalGrants } from "../persistence/schema/index.js";
|
|
|
177
178
|
import { bridgeConfirmationRequestToGuardian } from "../runtime/confirmation-request-guardian-bridge.js";
|
|
178
179
|
import { resolveRoutingState } from "../runtime/trust-context-resolver.js";
|
|
179
180
|
import {
|
|
181
|
+
resolveInlineGrantWaitMs,
|
|
180
182
|
TC_GRANT_WAIT_MAX_MS,
|
|
181
183
|
ToolApprovalHandler,
|
|
182
184
|
waitForInlineGrant,
|
|
183
185
|
} from "../tools/tool-approval-handler.js";
|
|
184
186
|
import type { ToolContext } from "../tools/types.js";
|
|
185
187
|
import { seedContactChannel } from "./helpers/seed-contact-channel.js";
|
|
188
|
+
import { setConfig } from "./helpers/set-config.js";
|
|
186
189
|
|
|
187
190
|
await initializeDb();
|
|
188
191
|
|
|
@@ -754,7 +757,10 @@ describe("(f) timeout/stale flow: stale guardian decision after inline wait time
|
|
|
754
757
|
test("inline_wait_active staleness guard: expired marker allows retry notification", async () => {
|
|
755
758
|
// Create a guardian request with a stale inline_wait_active marker
|
|
756
759
|
// that simulates a daemon crash during the wait.
|
|
757
|
-
|
|
760
|
+
// Age the marker past the real wait budget, which the resolver's
|
|
761
|
+
// staleness threshold tracks. Deriving it from the same helper the
|
|
762
|
+
// waiter uses keeps this case meaningful if the budget changes.
|
|
763
|
+
const staleTimestamp = Date.now() - resolveInlineGrantWaitMs() - 60_000;
|
|
758
764
|
const req = sim.seedRequest({
|
|
759
765
|
id: `req-stale-${Date.now()}`,
|
|
760
766
|
kind: "tool_grant_request",
|
|
@@ -802,6 +808,67 @@ describe("(f) timeout/stale flow: stale guardian decision after inline wait time
|
|
|
802
808
|
expect(retryNotifications.length).toBeGreaterThan(0);
|
|
803
809
|
});
|
|
804
810
|
|
|
811
|
+
test("inline_wait_active marker older than the fallback constant but inside the real budget still suppresses retry", async () => {
|
|
812
|
+
// The staleness threshold must track the wait budget, not the fallback
|
|
813
|
+
// constant. A marker aged past TC_GRANT_WAIT_MAX_MS + buffer but still
|
|
814
|
+
// well inside the configured budget belongs to a waiter that is very
|
|
815
|
+
// much alive: telling the requester to retry would race a call that is
|
|
816
|
+
// about to resume, and the retry then fails against the one-time grant
|
|
817
|
+
// the live waiter consumes.
|
|
818
|
+
const priorTimeouts = getConfig().timeouts;
|
|
819
|
+
// Seed the budget rather than leaning on the ambient default, so the
|
|
820
|
+
// window this case probes exists no matter what the suite's config holds.
|
|
821
|
+
setConfig("timeouts", { ...priorTimeouts, permissionTimeoutSec: 300 });
|
|
822
|
+
try {
|
|
823
|
+
const budgetMs = resolveInlineGrantWaitMs();
|
|
824
|
+
// Comfortably past the old 90s threshold, comfortably short of the budget.
|
|
825
|
+
const markerAgeMs = TC_GRANT_WAIT_MAX_MS + 45_000;
|
|
826
|
+
const liveTimestamp = Date.now() - markerAgeMs;
|
|
827
|
+
expect(markerAgeMs).toBeLessThan(budgetMs);
|
|
828
|
+
|
|
829
|
+
const req = sim.seedRequest({
|
|
830
|
+
id: `req-live-${Date.now()}`,
|
|
831
|
+
kind: "tool_grant_request",
|
|
832
|
+
sourceType: "channel",
|
|
833
|
+
sourceChannel: "telegram",
|
|
834
|
+
sourceConversationId: "conv-live-1",
|
|
835
|
+
requesterExternalUserId: "requester-1",
|
|
836
|
+
requesterChatId: "requester-chat-1",
|
|
837
|
+
guardianExternalUserId: "guardian-1",
|
|
838
|
+
guardianPrincipalId: "test-principal-id",
|
|
839
|
+
toolName: "bash",
|
|
840
|
+
inputDigest: "sha256:livewait",
|
|
841
|
+
expiresAt: Date.now() + 60_000,
|
|
842
|
+
});
|
|
843
|
+
|
|
844
|
+
await sim.module.updateGuardianRequest(req.id, {
|
|
845
|
+
followupState: `inline_wait_active:${liveTimestamp}`,
|
|
846
|
+
});
|
|
847
|
+
|
|
848
|
+
deliveredReplies.length = 0;
|
|
849
|
+
const approvalResult = await applyGuardianDecision({
|
|
850
|
+
requestId: req.id,
|
|
851
|
+
action: "approve_once",
|
|
852
|
+
actorContext: guardianActor(),
|
|
853
|
+
channelDeliveryContext: {
|
|
854
|
+
replyCallbackUrl: "http://localhost:3000/reply",
|
|
855
|
+
guardianChatId: "guardian-chat-1",
|
|
856
|
+
assistantId: "self",
|
|
857
|
+
},
|
|
858
|
+
});
|
|
859
|
+
expect(approvalResult.applied).toBe(true);
|
|
860
|
+
|
|
861
|
+
const retryNotifications = deliveredReplies.filter(
|
|
862
|
+
(r) =>
|
|
863
|
+
typeof r.payload.text === "string" &&
|
|
864
|
+
(r.payload.text as string).includes("Please retry"),
|
|
865
|
+
);
|
|
866
|
+
expect(retryNotifications.length).toBe(0);
|
|
867
|
+
} finally {
|
|
868
|
+
setConfig("timeouts", priorTimeouts);
|
|
869
|
+
}
|
|
870
|
+
});
|
|
871
|
+
|
|
805
872
|
test("fresh inline_wait_active marker suppresses retry notification", async () => {
|
|
806
873
|
// Create a request with a FRESH inline_wait_active marker
|
|
807
874
|
const freshTimestamp = Date.now();
|
|
@@ -54,7 +54,7 @@ import {
|
|
|
54
54
|
readBatchMetadata,
|
|
55
55
|
resolvePendingQuestion,
|
|
56
56
|
} from "../runtime/question-resolution.js";
|
|
57
|
-
import {
|
|
57
|
+
import { resolveInlineGrantWaitMs } from "../tools/tool-approval-handler.js";
|
|
58
58
|
import { getLogger } from "../util/logger.js";
|
|
59
59
|
import {
|
|
60
60
|
channelCanAddressOneReaderInBand,
|
|
@@ -1587,7 +1587,10 @@ const toolGrantRequestResolver: GuardianRequestResolver = {
|
|
|
1587
1587
|
// outlive the actual waiter if the daemon crashes or restarts during
|
|
1588
1588
|
// the wait. To avoid permanently suppressing the retry notification, we
|
|
1589
1589
|
// treat the marker as stale if the encoded start timestamp is older than
|
|
1590
|
-
// the maximum wait budget plus a 30s buffer.
|
|
1590
|
+
// the maximum wait budget plus a 30s buffer. The budget is read from the
|
|
1591
|
+
// same resolver the waiter itself uses, so a config change moves both
|
|
1592
|
+
// together; sizing this off a constant would let it fall below the real
|
|
1593
|
+
// wait and declare a live waiter dead.
|
|
1591
1594
|
const INLINE_WAIT_STALENESS_BUFFER_MS = 30_000;
|
|
1592
1595
|
const freshRequest = await getGuardianRequestOrNull(request.id);
|
|
1593
1596
|
const followupState = freshRequest?.followupState ?? "";
|
|
@@ -1604,7 +1607,7 @@ const toolGrantRequestResolver: GuardianRequestResolver = {
|
|
|
1604
1607
|
? Date.now() - waitStartMs
|
|
1605
1608
|
: Infinity; // Treat unparseable timestamps as stale for safety.
|
|
1606
1609
|
const stalenessThresholdMs =
|
|
1607
|
-
|
|
1610
|
+
resolveInlineGrantWaitMs() + INLINE_WAIT_STALENESS_BUFFER_MS;
|
|
1608
1611
|
if (markerAgeMs > stalenessThresholdMs) {
|
|
1609
1612
|
log.warn(
|
|
1610
1613
|
{
|
|
@@ -26,7 +26,7 @@ export const TimeoutConfigSchema = z
|
|
|
26
26
|
.positive("timeouts.permissionTimeoutSec must be a positive number")
|
|
27
27
|
.default(300)
|
|
28
28
|
.describe(
|
|
29
|
-
"How long to wait for
|
|
29
|
+
"How long to wait for a human to approve a tool call before timing out (seconds). Spent by both approval paths: the interactive permission prompt, and the inline wait when a contact's sensitive tool call is escalated to the guardian.",
|
|
30
30
|
),
|
|
31
31
|
questionResponseTimeoutSec: z
|
|
32
32
|
.number({ error: "timeouts.questionResponseTimeoutSec must be a number" })
|
|
@@ -366,6 +366,13 @@ export async function runAgentLoopImpl(
|
|
|
366
366
|
* scheduled execute turn attributes its LLM spend to that firing.
|
|
367
367
|
*/
|
|
368
368
|
cronRunId?: string | null;
|
|
369
|
+
/**
|
|
370
|
+
* Trust this turn runs under. Queue drains pass the trust captured from
|
|
371
|
+
* the sender at enqueue; without it the initialization below would reset
|
|
372
|
+
* the turn to the conversation slot, which holds whichever actor sent
|
|
373
|
+
* most recently rather than the one this turn belongs to.
|
|
374
|
+
*/
|
|
375
|
+
turnTrustContext?: TrustContext;
|
|
369
376
|
},
|
|
370
377
|
): Promise<void> {
|
|
371
378
|
if (!ctx.abortController) {
|
|
@@ -376,9 +383,12 @@ export async function runAgentLoopImpl(
|
|
|
376
383
|
// voice-session-bridge, regenerate, etc.) that invoke runAgentLoop directly
|
|
377
384
|
// without going through processMessage/drainQueue. This ensures the system
|
|
378
385
|
// prompt callback always reads a valid snapshot rather than undefined.
|
|
379
|
-
//
|
|
380
|
-
//
|
|
381
|
-
|
|
386
|
+
//
|
|
387
|
+
// A queue drain runs a message whose sender may not be the conversation's
|
|
388
|
+
// most recent one, so it passes that sender's trust explicitly. Falling back
|
|
389
|
+
// to the slot here would undo the drain's stamp and hand the turn back to
|
|
390
|
+
// whoever sent last.
|
|
391
|
+
ctx.currentTurnTrustContext = options?.turnTrustContext ?? ctx.trustContext;
|
|
382
392
|
ctx.currentTurnChannelCapabilities = ctx.channelCapabilities;
|
|
383
393
|
|
|
384
394
|
// Re-resolve the system prompt under the snapshots just set and push it into
|
|
@@ -565,6 +565,12 @@ export interface EnqueueMessageOptions {
|
|
|
565
565
|
sourceActorPrincipalId?: string;
|
|
566
566
|
/** Auth context snapshot captured for queued turn-scoped authorization. */
|
|
567
567
|
authContext?: AuthContext;
|
|
568
|
+
/**
|
|
569
|
+
* Sender's trust, for the drain to run this message under. Defaults to the
|
|
570
|
+
* conversation's trust at enqueue time, which the sending route has just
|
|
571
|
+
* set to this sender.
|
|
572
|
+
*/
|
|
573
|
+
trustContext?: TrustContext;
|
|
568
574
|
}
|
|
569
575
|
|
|
570
576
|
// ── enqueueMessage ───────────────────────────────────────────────────
|
|
@@ -593,6 +599,9 @@ export function enqueueMessage(
|
|
|
593
599
|
options.sourceActorPrincipalId ??
|
|
594
600
|
ctx.currentTurnSourceActorPrincipalId ??
|
|
595
601
|
queuedAuthContext?.actorPrincipalId;
|
|
602
|
+
// Deliberately not falling back to `currentTurnTrustContext`: that is the
|
|
603
|
+
// in-flight turn's actor, which is precisely who this message is not from.
|
|
604
|
+
const queuedTrustContext = options.trustContext ?? ctx.trustContext;
|
|
596
605
|
|
|
597
606
|
if (!ctx.isProcessing()) {
|
|
598
607
|
return { queued: false, requestId };
|
|
@@ -619,6 +628,7 @@ export function enqueueMessage(
|
|
|
619
628
|
isInteractive,
|
|
620
629
|
sourceActorPrincipalId,
|
|
621
630
|
authContext: queuedAuthContext,
|
|
631
|
+
trustContext: queuedTrustContext,
|
|
622
632
|
transport,
|
|
623
633
|
displayContent,
|
|
624
634
|
sentAt: Date.now(),
|
|
@@ -668,6 +678,14 @@ export interface PersistMessageOptions {
|
|
|
668
678
|
metadata?: Record<string, unknown>;
|
|
669
679
|
displayContent?: string;
|
|
670
680
|
clientMessageId?: string;
|
|
681
|
+
/**
|
|
682
|
+
* Trust to attribute the stored row to. Queue drains pass the sender's
|
|
683
|
+
* captured trust so persisted provenance names the same actor the turn
|
|
684
|
+
* executes as; the conversation slot may by then hold someone else.
|
|
685
|
+
* Defaults to the conversation's trust, which is correct for callers
|
|
686
|
+
* persisting a message the current actor just sent.
|
|
687
|
+
*/
|
|
688
|
+
trustContext?: TrustContext;
|
|
671
689
|
/**
|
|
672
690
|
* Persist the row without indexing it (no memory segments, embeddings, or
|
|
673
691
|
* lexical-index entry). For machine-authored prompts that must not enter
|
|
@@ -844,7 +862,9 @@ export async function persistQueuedMessageBody(
|
|
|
844
862
|
extractTurnChannelContext(metadata) ?? ctx.getTurnChannelContext();
|
|
845
863
|
const turnIfCtx =
|
|
846
864
|
extractTurnInterfaceContext(metadata) ?? ctx.getTurnInterfaceContext();
|
|
847
|
-
const provenance = provenanceFromTrustContext(
|
|
865
|
+
const provenance = provenanceFromTrustContext(
|
|
866
|
+
options.trustContext ?? ctx.trustContext,
|
|
867
|
+
);
|
|
848
868
|
const imageSourcePaths = extractImageSourcePaths(attachments);
|
|
849
869
|
|
|
850
870
|
// Strip the transient `slackInbound` carrier key from the persisted
|