@xaccefy/pi-casefile 0.10.1 → 0.11.0
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/README.md +2 -3
- package/package.json +1 -2
- package/src/confirmation.ts +45 -351
- package/src/evidence.ts +59 -220
- package/src/harness-verify.ts +36 -206
- package/src/index.ts +44 -366
- package/src/ledger.ts +24 -71
- package/src/poc-runner.ts +3 -3
- package/src/oob-oracle.ts +0 -279
package/README.md
CHANGED
|
@@ -15,9 +15,8 @@ A structured ledger for offensive-security work — bug bounties, CTFs, audits
|
|
|
15
15
|
|
|
16
16
|
Cases move `hypothesis → investigating → confirmed → reported`. Promotion between phases is gated:
|
|
17
17
|
|
|
18
|
-
- **Zero exit is necessary but never proof** — direct-response findings require nonce-bound body evidence plus a DNS-pinned, conclusive `target_only` replay against
|
|
19
|
-
- **Differential confirmation** —
|
|
20
|
-
- **Blind/OOB classes** confirm through an operator-run oracle with per-run tokens and source-separation attestation
|
|
18
|
+
- **Zero exit is necessary but never proof** — direct-response findings require nonce-bound body evidence plus a DNS-pinned, conclusive `target_only` attack-vs-baseline replay against the case target (recorded at promote; the bundle's baseline binding and differential are re-validated at confirm)
|
|
19
|
+
- **Differential confirmation** — the attack request must satisfy the claimed predicate while a legitimate same-host baseline request must not, so "it worked" means *the discriminator fired*, not "the agent said so"
|
|
21
20
|
- Only the main agent makes the semantic decision and commits phase transitions
|
|
22
21
|
|
|
23
22
|
Designed for **human + AI workflows**: every confirmed finding carries a reproducible evidence trail a human can audit.
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@xaccefy/pi-casefile",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.11.0",
|
|
4
4
|
"description": "Offensive security case tracker for Pi Agent \u2014 bug bounties, CTFs, security audits",
|
|
5
5
|
"keywords": [
|
|
6
6
|
"pi-package",
|
|
@@ -35,7 +35,6 @@
|
|
|
35
35
|
"src/index.ts",
|
|
36
36
|
"src/evidence.ts",
|
|
37
37
|
"src/harness-verify.ts",
|
|
38
|
-
"src/oob-oracle.ts",
|
|
39
38
|
"src/confirmation.ts",
|
|
40
39
|
"src/ledger-internal.ts",
|
|
41
40
|
"src/safe-state.ts",
|
package/src/confirmation.ts
CHANGED
|
@@ -6,14 +6,15 @@
|
|
|
6
6
|
* one readable module. ledger.ts re-exports every public symbol here, so
|
|
7
7
|
* callers and tests are unchanged.
|
|
8
8
|
*
|
|
9
|
-
* The gate's contract
|
|
9
|
+
* The gate's contract:
|
|
10
10
|
* - Zero exit + complete output capture = run integrity, never proof.
|
|
11
11
|
* - Evidence must be nonce-bound, schema-valid, carry a discriminating
|
|
12
12
|
* response-body predicate, and survive durable-hash re-verification.
|
|
13
|
-
* - Target-dependence requires a machine differential:
|
|
14
|
-
*
|
|
15
|
-
*
|
|
16
|
-
*
|
|
13
|
+
* - Target-dependence requires a machine differential: the attack request
|
|
14
|
+
* must satisfy the predicate while a legitimate same-host baseline request
|
|
15
|
+
* (declared in the evidence) must not.
|
|
16
|
+
* - Only the main agent commits the phase-2 verdict; only the ledger can
|
|
17
|
+
* transition a case to confirmed.
|
|
17
18
|
*/
|
|
18
19
|
|
|
19
20
|
import { createHash } from "node:crypto";
|
|
@@ -25,11 +26,9 @@ import {
|
|
|
25
26
|
evidenceNonceMatches,
|
|
26
27
|
type MainAgentVerdict,
|
|
27
28
|
normalizeEvidence,
|
|
28
|
-
panelQuorumReached,
|
|
29
29
|
parsePoCEvidence,
|
|
30
30
|
scanArtifactForSecrets,
|
|
31
31
|
validateMainAgentVerdict,
|
|
32
|
-
validatePanelVotes,
|
|
33
32
|
} from "./evidence.ts";
|
|
34
33
|
import { type HarnessVerifyResult, sameRequest, verifyUrlBindingError } from "./harness-verify.ts";
|
|
35
34
|
import type {
|
|
@@ -37,7 +36,6 @@ import type {
|
|
|
37
36
|
CaseUpdateResult,
|
|
38
37
|
EvidenceItem,
|
|
39
38
|
MainAgentVerdictRecord,
|
|
40
|
-
MainAgentVerification,
|
|
41
39
|
NormalizedCaseInput,
|
|
42
40
|
PendingConfirmation,
|
|
43
41
|
PocEvidenceRun,
|
|
@@ -271,56 +269,21 @@ function validateRunEvidence(run: PocEvidenceRun, label: string): void {
|
|
|
271
269
|
}
|
|
272
270
|
}
|
|
273
271
|
|
|
274
|
-
/** Determinism
|
|
275
|
-
function assertEvidenceDifferential(bundle: PendingConfirmation
|
|
272
|
+
/** Determinism on normalized evidence (nonce/observations stripped). */
|
|
273
|
+
function assertEvidenceDifferential(bundle: PendingConfirmation): void {
|
|
276
274
|
const [r1, r2] = bundle.targetRuns;
|
|
277
275
|
if (normalizeEvidence(r1.evidence) !== normalizeEvidence(r2.evidence)) {
|
|
278
276
|
throw new Error(
|
|
279
277
|
"Target runs produced inconsistent evidence — the exploit did not reproduce deterministically",
|
|
280
278
|
);
|
|
281
279
|
}
|
|
282
|
-
// Intra-target target-dependence is proven by the harness attack-vs-baseline
|
|
283
|
-
// replay (same host), not by comparing a target run to a separate control run.
|
|
284
|
-
if (isIntra) return;
|
|
285
|
-
// OOB-only bundles prove target-dependence via the oracle token differential
|
|
286
|
-
// (assertMachineConfirmation judges callbackVerified); no control run exists.
|
|
287
|
-
if (!bundle.controlRun && bundle.callbackVerified?.attempted) return;
|
|
288
|
-
if (!bundle.controlRun) {
|
|
289
|
-
throw new Error("inter-host confirmation requires a control run");
|
|
290
|
-
}
|
|
291
|
-
if (normalizeEvidence(r1.evidence) === normalizeEvidence(bundle.controlRun.evidence)) {
|
|
292
|
-
throw new Error(
|
|
293
|
-
"Control run produced identical evidence to the target — the claimed impact is not target-dependent",
|
|
294
|
-
);
|
|
295
|
-
}
|
|
296
280
|
}
|
|
297
281
|
|
|
298
282
|
function assertMachineConfirmation(bundle: PendingConfirmation): void {
|
|
299
|
-
const oob = bundle.callbackVerified;
|
|
300
|
-
if (oob?.attempted) {
|
|
301
|
-
if (oob.targetHits === 0) {
|
|
302
|
-
throw new Error(
|
|
303
|
-
`OOB VERIFY FAILED: no interaction with the target-run callback token. ${oob.note}`,
|
|
304
|
-
);
|
|
305
|
-
}
|
|
306
|
-
if (oob.controlHits > 0) {
|
|
307
|
-
throw new Error(
|
|
308
|
-
`OOB VERIFY FAILED: the control-run callback token received ${oob.controlHits} interaction(s) — the callback is not target-dependent. ${oob.note}`,
|
|
309
|
-
);
|
|
310
|
-
}
|
|
311
|
-
if (oob.sourceSeparated !== true) {
|
|
312
|
-
throw new Error(
|
|
313
|
-
"OOB VERIFY FAILED: callback source separation was not established. " +
|
|
314
|
-
"A loopback listener reachable by the PoC is diagnostic telemetry, not proof that the target caused the interaction.",
|
|
315
|
-
);
|
|
316
|
-
}
|
|
317
|
-
return;
|
|
318
|
-
}
|
|
319
|
-
|
|
320
283
|
assertHarnessTargetOnly(
|
|
321
284
|
bundle.harnessVerified,
|
|
322
285
|
"HARNESS DIFFERENTIAL FAILED",
|
|
323
|
-
"no machine-owned
|
|
286
|
+
"no machine-owned attack/baseline replay was recorded",
|
|
324
287
|
);
|
|
325
288
|
}
|
|
326
289
|
|
|
@@ -340,79 +303,6 @@ function assertHarnessTargetOnly(
|
|
|
340
303
|
}
|
|
341
304
|
}
|
|
342
305
|
|
|
343
|
-
function assertHarnessCanary(
|
|
344
|
-
harness: HarnessVerifyResult | undefined,
|
|
345
|
-
required: boolean,
|
|
346
|
-
label: string,
|
|
347
|
-
): void {
|
|
348
|
-
if (!required) return;
|
|
349
|
-
if (
|
|
350
|
-
harness?.canary?.attempted !== true ||
|
|
351
|
-
harness.canary.pass !== true ||
|
|
352
|
-
harness.canary.targetObserved !== true ||
|
|
353
|
-
harness.canary.controlObserved !== false ||
|
|
354
|
-
harness.proofStrength !== "canary_differential"
|
|
355
|
-
) {
|
|
356
|
-
throw new Error(`${label}: ${harness?.canary?.note ?? "required canary transcript missing"}`);
|
|
357
|
-
}
|
|
358
|
-
}
|
|
359
|
-
|
|
360
|
-
function assertMainAgentVerification(
|
|
361
|
-
bundle: PendingConfirmation,
|
|
362
|
-
verification: MainAgentVerification | undefined,
|
|
363
|
-
isIntra = false,
|
|
364
|
-
): asserts verification is MainAgentVerification {
|
|
365
|
-
if (!verification) {
|
|
366
|
-
throw new Error(
|
|
367
|
-
"MAIN-AGENT REPLAY REQUIRED: ConfirmFinding must produce a fresh harness-owned target/control transcript",
|
|
368
|
-
);
|
|
369
|
-
}
|
|
370
|
-
const at = Date.parse(verification.at);
|
|
371
|
-
const bundleAt = Date.parse(bundle.ranAt);
|
|
372
|
-
const now = Date.now();
|
|
373
|
-
if (
|
|
374
|
-
!Number.isFinite(at) ||
|
|
375
|
-
!Number.isFinite(bundleAt) ||
|
|
376
|
-
at < bundleAt ||
|
|
377
|
-
at > now + 30_000 ||
|
|
378
|
-
now - at > 5 * 60 * 1000
|
|
379
|
-
) {
|
|
380
|
-
throw new Error(
|
|
381
|
-
"MAIN-AGENT REPLAY FAILED: transcript timestamp must be valid, newer than phase 1, and no more than 5 minutes old",
|
|
382
|
-
);
|
|
383
|
-
}
|
|
384
|
-
assertHarnessTargetOnly(
|
|
385
|
-
verification.result,
|
|
386
|
-
"MAIN-AGENT REPLAY FAILED",
|
|
387
|
-
"no fresh phase-2 target/control replay was recorded",
|
|
388
|
-
);
|
|
389
|
-
assertHarnessCanary(
|
|
390
|
-
verification.result,
|
|
391
|
-
bundle.targetRuns[0].evidence.verify.canary !== undefined,
|
|
392
|
-
"MAIN-AGENT CANARY FAILED",
|
|
393
|
-
);
|
|
394
|
-
// OOB-only bundles bind by TOKEN identity (enforced at store time on
|
|
395
|
-
// evidence.verify.url); there is no control host to bind a transcript to.
|
|
396
|
-
if (bundle.callbackVerified?.attempted && bundle.oobTokens) return;
|
|
397
|
-
const targetUrl = verification.result.target?.url;
|
|
398
|
-
const controlUrl = verification.result.control?.url;
|
|
399
|
-
const targetIdentity = bundle.targetRuns[0].target;
|
|
400
|
-
if (!targetUrl || verifyUrlBindingError(targetUrl, targetIdentity)) {
|
|
401
|
-
throw new Error("MAIN-AGENT REPLAY FAILED: target transcript is not bound to the case target");
|
|
402
|
-
}
|
|
403
|
-
// Intra-target: the "control" transcript is the legitimate baseline request,
|
|
404
|
-
// which is bound to the SAME case target. Inter-host: it is bound to the
|
|
405
|
-
// distinct control target.
|
|
406
|
-
const controlBindTarget = isIntra ? targetIdentity : bundle.controlTarget;
|
|
407
|
-
if (!controlUrl || !controlBindTarget || verifyUrlBindingError(controlUrl, controlBindTarget)) {
|
|
408
|
-
throw new Error(
|
|
409
|
-
isIntra
|
|
410
|
-
? "MAIN-AGENT REPLAY FAILED: baseline transcript is not bound to the case target"
|
|
411
|
-
: "MAIN-AGENT REPLAY FAILED: control transcript is not bound to control_target",
|
|
412
|
-
);
|
|
413
|
-
}
|
|
414
|
-
}
|
|
415
|
-
|
|
416
306
|
/**
|
|
417
307
|
* Gate for phase 1 of promotion: case must exist, be investigating, and have
|
|
418
308
|
* poc/evidence/impact/severity/target. The disconfirmation is provided by the
|
|
@@ -460,27 +350,16 @@ export function assertPromotable(id: string): CaseRecord {
|
|
|
460
350
|
}
|
|
461
351
|
|
|
462
352
|
/**
|
|
463
|
-
* Phase 1
|
|
464
|
-
*
|
|
465
|
-
*
|
|
466
|
-
*
|
|
353
|
+
* Phase 1: validate a same-host attack-vs-baseline bundle. The differential is
|
|
354
|
+
* proven by the harness replay (attack matched, baseline did not, both against
|
|
355
|
+
* the case target) — the discriminating variable is the request's identity or
|
|
356
|
+
* a parameter, not the host.
|
|
467
357
|
*/
|
|
468
|
-
function
|
|
469
|
-
|
|
470
|
-
id: string,
|
|
471
|
-
bundle: PendingConfirmation,
|
|
472
|
-
): CaseRecord {
|
|
473
|
-
if (bundle.targetRuns.length !== 2) {
|
|
474
|
-
throw new Error("Intra-target confirmation requires two target runs");
|
|
475
|
-
}
|
|
476
|
-
if (bundle.controlRun || bundle.controlTarget) {
|
|
477
|
-
throw new Error(
|
|
478
|
-
"Intra-target confirmation must not carry a control run or control target — the baseline is a same-host request inside the evidence",
|
|
479
|
-
);
|
|
480
|
-
}
|
|
358
|
+
function validateBundle(current: CaseRecord, id: string, bundle: PendingConfirmation): CaseRecord {
|
|
359
|
+
if (bundle.targetRuns.length !== 2) throw new Error("Confirmation requires two target runs");
|
|
481
360
|
const targetRunTarget = bundle.targetRuns[0]?.target;
|
|
482
361
|
if (!targetRunTarget || bundle.targetRuns.some((r) => r.target !== targetRunTarget)) {
|
|
483
|
-
throw new Error("
|
|
362
|
+
throw new Error("Confirmation requires both runs against the same case target");
|
|
484
363
|
}
|
|
485
364
|
let pocHash: string | undefined;
|
|
486
365
|
try {
|
|
@@ -494,35 +373,20 @@ function validateIntraTargetBundle(
|
|
|
494
373
|
for (const run of bundle.targetRuns) {
|
|
495
374
|
validateRunEvidence(run, `${run.mode} run`);
|
|
496
375
|
const ev = run.evidence;
|
|
497
|
-
if (ev.verify.mode !== "intra_target") {
|
|
498
|
-
throw new Error(
|
|
499
|
-
"INTRA-TARGET FAILED: each run's evidence.verify.mode must be 'intra_target'",
|
|
500
|
-
);
|
|
501
|
-
}
|
|
502
|
-
if (!ev.baseline) {
|
|
503
|
-
throw new Error(
|
|
504
|
-
"INTRA-TARGET FAILED: evidence.baseline (a legitimate same-host request) is required",
|
|
505
|
-
);
|
|
506
|
-
}
|
|
507
376
|
const attackBinding = verifyUrlBindingError(ev.verify.url, targetRunTarget);
|
|
508
377
|
if (attackBinding) throw new Error(`ATTACK BINDING FAILED: ${attackBinding}`);
|
|
509
378
|
const baselineBinding = verifyUrlBindingError(ev.baseline.url, targetRunTarget);
|
|
510
379
|
if (baselineBinding) throw new Error(`BASELINE BINDING FAILED: ${baselineBinding}`);
|
|
511
380
|
if (ev.baseline && sameRequest(ev.verify, ev.baseline)) {
|
|
512
381
|
throw new Error(
|
|
513
|
-
"
|
|
382
|
+
"BASELINE CHECK FAILED: attack and baseline requests are identical — vary identity or a parameter",
|
|
514
383
|
);
|
|
515
384
|
}
|
|
516
385
|
}
|
|
517
386
|
if (bundle.caseId !== id) throw new Error("Pending confirmation caseId mismatch");
|
|
518
|
-
assertEvidenceDifferential(bundle
|
|
387
|
+
assertEvidenceDifferential(bundle);
|
|
519
388
|
// Machine floor: attack matched, baseline did not, both against the case target.
|
|
520
389
|
assertMachineConfirmation(bundle);
|
|
521
|
-
assertHarnessCanary(
|
|
522
|
-
bundle.harnessVerified,
|
|
523
|
-
bundle.targetRuns[0].evidence.verify.canary !== undefined,
|
|
524
|
-
"PHASE-1 CANARY FAILED",
|
|
525
|
-
);
|
|
526
390
|
const next = buildRecord({ pendingConfirmation: bundle }, current);
|
|
527
391
|
validateCase(next);
|
|
528
392
|
return next;
|
|
@@ -530,9 +394,9 @@ function validateIntraTargetBundle(
|
|
|
530
394
|
|
|
531
395
|
/**
|
|
532
396
|
* Phase 1: record the harness-observed evidence bundle on the case. The whole
|
|
533
|
-
* contract is validated here —
|
|
534
|
-
*
|
|
535
|
-
*
|
|
397
|
+
* contract is validated here — nonce binding, run completion, determinism
|
|
398
|
+
* across the two target runs, and the attack/baseline differential — so a
|
|
399
|
+
* bundle that cannot promote is rejected before the
|
|
536
400
|
* main agent performs phase-2 review.
|
|
537
401
|
*/
|
|
538
402
|
export function storePendingConfirmation(id: string, bundle: PendingConfirmation): CaseRecord {
|
|
@@ -546,127 +410,22 @@ export function storePendingConfirmation(id: string, bundle: PendingConfirmation
|
|
|
546
410
|
);
|
|
547
411
|
}
|
|
548
412
|
if (bundle.caseId !== id) throw new Error("Pending confirmation caseId mismatch");
|
|
549
|
-
// Panel vote shape is machine-checked at store time — a malformed panel
|
|
550
|
-
// must never silently count toward a quorum later.
|
|
551
|
-
if (bundle.panelVotes !== undefined) {
|
|
552
|
-
const votes = validatePanelVotes(bundle.panelVotes);
|
|
553
|
-
if (!votes.ok) throw new Error(`Pending confirmation panel invalid: ${votes.error}`);
|
|
554
|
-
}
|
|
555
|
-
if (bundle.mode === "intra_target") {
|
|
556
|
-
const next = validateIntraTargetBundle(current, id, bundle);
|
|
557
|
-
upsertCase(db, next);
|
|
558
|
-
appendCaseEvent(db, {
|
|
559
|
-
actor: "harness",
|
|
560
|
-
caseId: id,
|
|
561
|
-
eventType: "promotion_pending",
|
|
562
|
-
payload: { mode: "intra_target", evidence_sha256: bundle.targetRuns[0].evidenceSha256 },
|
|
563
|
-
});
|
|
564
|
-
return next;
|
|
565
|
-
}
|
|
566
|
-
// Control-run requirements key off controlRun PRESENCE, not the OOB flag:
|
|
567
|
-
// an OOB-only bundle has no control run (token differential instead), but
|
|
568
|
-
// an OOB+control bundle still carries one and gets the full checks.
|
|
569
|
-
if (!bundle.controlRun && !bundle.callbackVerified) {
|
|
570
|
-
throw new Error("Pending confirmation requires two target runs and one control run");
|
|
571
|
-
}
|
|
572
|
-
if (bundle.controlRun && (!bundle.pocPath || !bundle.controlPath || !bundle.controlTarget)) {
|
|
573
|
-
throw new Error("Pending confirmation requires pocPath, controlPath, and controlTarget");
|
|
574
|
-
}
|
|
575
|
-
// Control-target binding (machine-verified here, not just in the tool
|
|
576
|
-
// layer): the control run must actually have targeted the declared
|
|
577
|
-
// control_target, that target must differ from the target runs' target,
|
|
578
|
-
// and the control target must differ from the case's target — otherwise
|
|
579
|
-
// "the control demonstrated nothing on the vulnerable target" passes.
|
|
580
|
-
const targetRunTarget = bundle.targetRuns[0]?.target;
|
|
581
|
-
if (!targetRunTarget || bundle.targetRuns.some((r) => r.target !== targetRunTarget)) {
|
|
582
|
-
throw new Error(
|
|
583
|
-
"Pending confirmation requires both target runs against the same case target",
|
|
584
|
-
);
|
|
585
|
-
}
|
|
586
|
-
if (bundle.controlRun) {
|
|
587
|
-
if (!bundle.controlRun.target || bundle.controlRun.target !== bundle.controlTarget) {
|
|
588
|
-
throw new Error(
|
|
589
|
-
"CONTROL BINDING FAILED: controlRun.target must equal control_target — a control run " +
|
|
590
|
-
"against a different host than the one declared proves nothing.",
|
|
591
|
-
);
|
|
592
|
-
}
|
|
593
|
-
if (bundle.controlRun.target === targetRunTarget) {
|
|
594
|
-
throw new Error(
|
|
595
|
-
"CONTROL BINDING FAILED: the control run targeted the same host as the target runs — " +
|
|
596
|
-
"the claimed impact is not target-dependent.",
|
|
597
|
-
);
|
|
598
|
-
}
|
|
599
|
-
}
|
|
600
|
-
if (bundle.controlTarget && bundle.controlTarget === current.target) {
|
|
601
|
-
throw new Error(
|
|
602
|
-
"CONTROL BINDING FAILED: control_target must differ from the case target; a control run " +
|
|
603
|
-
"against the vulnerable target proves nothing.",
|
|
604
|
-
);
|
|
605
|
-
}
|
|
606
|
-
// Same-file contract re-checked at store time (the tool already checked).
|
|
607
|
-
// OOB-only bundles carry no separate control script — the PoC hash alone
|
|
608
|
-
// is re-verified against the file on disk.
|
|
609
413
|
let pocHash: string | undefined;
|
|
610
|
-
let controlHash: string | undefined;
|
|
611
414
|
try {
|
|
612
415
|
pocHash = createHash("sha256").update(readFileSync(bundle.pocPath)).digest("hex");
|
|
613
|
-
controlHash = bundle.controlPath
|
|
614
|
-
? createHash("sha256").update(readFileSync(bundle.controlPath)).digest("hex")
|
|
615
|
-
: pocHash;
|
|
616
416
|
} catch {
|
|
617
417
|
pocHash = undefined;
|
|
618
|
-
controlHash = undefined;
|
|
619
418
|
}
|
|
620
|
-
if (!pocHash ||
|
|
621
|
-
throw new Error(
|
|
622
|
-
"CONTROL CHECK FAILED: control_path must be the SAME script as poc_path " +
|
|
623
|
-
"(sha256 mismatch). A separately written control file proves nothing.",
|
|
624
|
-
);
|
|
625
|
-
}
|
|
626
|
-
if (bundle.pocSha256 && bundle.pocSha256 !== pocHash) {
|
|
419
|
+
if (!pocHash || (bundle.pocSha256 && bundle.pocSha256 !== pocHash)) {
|
|
627
420
|
throw new Error("pocSha256 does not match the PoC file on disk");
|
|
628
421
|
}
|
|
629
|
-
|
|
630
|
-
// OOB+control bundles validate all three.
|
|
631
|
-
const runsToValidate = bundle.controlRun
|
|
632
|
-
? [...bundle.targetRuns, bundle.controlRun]
|
|
633
|
-
: [...bundle.targetRuns];
|
|
634
|
-
for (const run of runsToValidate) {
|
|
635
|
-
validateRunEvidence(run, `${run.mode} run`);
|
|
636
|
-
}
|
|
637
|
-
assertEvidenceDifferential(bundle);
|
|
638
|
-
// Target binding applies to EVERY mode — an OOB bundle's verify.url must
|
|
639
|
-
// still belong to the case target, or the PoC could anchor its evidence on
|
|
640
|
-
// an unrelated host while the callback alone carries the proof.
|
|
641
|
-
for (const run of bundle.targetRuns) {
|
|
642
|
-
const bindingError = verifyUrlBindingError(run.evidence.verify.url, targetRunTarget);
|
|
643
|
-
if (bindingError) throw new Error(`TARGET BINDING FAILED: ${bindingError}`);
|
|
644
|
-
}
|
|
645
|
-
if (bundle.controlRun) {
|
|
646
|
-
const controlBindingError = verifyUrlBindingError(
|
|
647
|
-
bundle.controlRun.evidence.verify.url,
|
|
648
|
-
bundle.controlTarget!,
|
|
649
|
-
);
|
|
650
|
-
if (controlBindingError) {
|
|
651
|
-
throw new Error(`CONTROL BINDING FAILED: ${controlBindingError}`);
|
|
652
|
-
}
|
|
653
|
-
}
|
|
654
|
-
// A clean exit and model-authored evidence are necessary inputs, never the
|
|
655
|
-
// proof. Promotion requires a harness-observed target/control differential
|
|
656
|
-
// or a harness-owned OOB interaction differential.
|
|
657
|
-
assertMachineConfirmation(bundle);
|
|
658
|
-
|
|
659
|
-
const next = buildRecord({ pendingConfirmation: bundle }, current);
|
|
660
|
-
validateCase(next);
|
|
422
|
+
const next = validateBundle(current, id, bundle);
|
|
661
423
|
upsertCase(db, next);
|
|
662
424
|
appendCaseEvent(db, {
|
|
663
425
|
actor: "harness",
|
|
664
426
|
caseId: id,
|
|
665
427
|
eventType: "promotion_pending",
|
|
666
|
-
payload: {
|
|
667
|
-
mode: bundle.callbackVerified?.attempted ? "oob" : "inter_host",
|
|
668
|
-
evidence_sha256: bundle.targetRuns[0].evidenceSha256,
|
|
669
|
-
},
|
|
428
|
+
payload: { mode: "intra_target", evidence_sha256: bundle.targetRuns[0].evidenceSha256 },
|
|
670
429
|
});
|
|
671
430
|
return next;
|
|
672
431
|
});
|
|
@@ -678,16 +437,14 @@ export function storePendingConfirmation(id: string, bundle: PendingConfirmation
|
|
|
678
437
|
* CONFIRMED requires the full bundle to still hold (completion, nonce,
|
|
679
438
|
* determinism, differential), the PoC script to be unchanged since the runs
|
|
680
439
|
* (pocSha256 — otherwise the main agent reviewed different bytes), and a
|
|
681
|
-
* verdict accompanied by a
|
|
682
|
-
*
|
|
683
|
-
* disproved)
|
|
684
|
-
*
|
|
685
|
-
* manual review rather than dropping it.
|
|
440
|
+
* verdict accompanied by a concrete review note and a disconfirmation attempt.
|
|
441
|
+
* NOT_CONFIRMED (positively disproved) and INCONCLUSIVE (neither reproduced
|
|
442
|
+
* nor disproved) both record the verdict and keep the case investigating —
|
|
443
|
+
* INCONCLUSIVE preserves it for manual review rather than dropping it.
|
|
686
444
|
*/
|
|
687
445
|
export function applyConfirmationResult(
|
|
688
446
|
id: string,
|
|
689
447
|
verdictInput: MainAgentVerdict,
|
|
690
|
-
phase2Verification?: MainAgentVerification,
|
|
691
448
|
authority: { startedAsSubagent: boolean } = {
|
|
692
449
|
startedAsSubagent: PROCESS_STARTED_AS_SUBAGENT || process.env.PI_SUBAGENT_CHILD === "1",
|
|
693
450
|
},
|
|
@@ -720,42 +477,11 @@ export function applyConfirmationResult(
|
|
|
720
477
|
const parsed = validateMainAgentVerdict(verdictInput);
|
|
721
478
|
if (!parsed.ok) throw new Error(`Invalid main-agent confirmation verdict: ${parsed.error}`);
|
|
722
479
|
const verdict = parsed.verdict;
|
|
723
|
-
const canaryRequested = bundle.targetRuns[0].evidence.verify.canary !== undefined;
|
|
724
|
-
if (verdict.verdict === "CONFIRMED") {
|
|
725
|
-
if (canaryRequested && verdict.canary_assessment !== "verified") {
|
|
726
|
-
throw new Error(
|
|
727
|
-
"CONFIRMED canary mismatch: evidence requested a harness canary, so canary_assessment must be verified",
|
|
728
|
-
);
|
|
729
|
-
}
|
|
730
|
-
if (!canaryRequested && verdict.canary_assessment !== "not_applicable") {
|
|
731
|
-
throw new Error(
|
|
732
|
-
"CONFIRMED canary mismatch: this evidence has no canary template; record canary_assessment=not_applicable and explain why",
|
|
733
|
-
);
|
|
734
|
-
}
|
|
735
|
-
// Quorum panel pre-gate: CONFIRMED needs a 2/3 exploit panel or an
|
|
736
|
-
// explicit override note recording why the panel was skipped (or
|
|
737
|
-
// overruled). Votes are advisory — the main agent still commits — but a
|
|
738
|
-
// non-quorum CONFIRMED without a note is refused.
|
|
739
|
-
const quorum = panelQuorumReached(bundle.panelVotes);
|
|
740
|
-
if (!quorum.quorum && !verdict.panel_override_note?.trim()) {
|
|
741
|
-
throw new Error(
|
|
742
|
-
`PANEL QUORUM REQUIRED: CONFIRMED needs either a 2/3 exploit panel (got ${quorum.exploit} exploit / ${quorum.total} vote(s)) ` +
|
|
743
|
-
"or an explicit panel_override_note recording why the panel was skipped or overruled. " +
|
|
744
|
-
"Re-run PromoteFinding with panel_votes, or justify the solo confirmation in panel_override_note.",
|
|
745
|
-
);
|
|
746
|
-
}
|
|
747
|
-
}
|
|
748
480
|
const recorded: MainAgentVerdictRecord = {
|
|
749
481
|
...verdict,
|
|
750
482
|
at: new Date().toISOString(),
|
|
751
483
|
reviewer: "main_agent",
|
|
752
|
-
|
|
753
|
-
proofStrength:
|
|
754
|
-
verdict.verdict === "CONFIRMED"
|
|
755
|
-
? canaryRequested
|
|
756
|
-
? "canary_differential"
|
|
757
|
-
: "predicate_differential"
|
|
758
|
-
: undefined,
|
|
484
|
+
proofStrength: verdict.verdict === "CONFIRMED" ? "predicate_differential" : undefined,
|
|
759
485
|
};
|
|
760
486
|
|
|
761
487
|
if (verdict.verdict !== "CONFIRMED") {
|
|
@@ -792,16 +518,11 @@ export function applyConfirmationResult(
|
|
|
792
518
|
|
|
793
519
|
// CONFIRMED — re-validate the whole bundle (defense in depth; the case may
|
|
794
520
|
// have been touched between phase 1 and the verdict).
|
|
795
|
-
const
|
|
796
|
-
const allRuns = isIntra
|
|
797
|
-
? [...bundle.targetRuns]
|
|
798
|
-
: [...bundle.targetRuns, ...(bundle.controlRun ? [bundle.controlRun] : [])];
|
|
799
|
-
for (const run of allRuns) {
|
|
521
|
+
for (const run of bundle.targetRuns) {
|
|
800
522
|
validateRunEvidence(run, `${run.mode} run`);
|
|
801
523
|
}
|
|
802
|
-
assertEvidenceDifferential(bundle
|
|
524
|
+
assertEvidenceDifferential(bundle);
|
|
803
525
|
assertMachineConfirmation(bundle);
|
|
804
|
-
assertHarnessCanary(bundle.harnessVerified, canaryRequested, "PHASE-1 CANARY FAILED");
|
|
805
526
|
let pocHash: string | undefined;
|
|
806
527
|
try {
|
|
807
528
|
pocHash = createHash("sha256").update(readFileSync(bundle.pocPath)).digest("hex");
|
|
@@ -813,9 +534,8 @@ export function applyConfirmationResult(
|
|
|
813
534
|
"PoC script changed since the runs — re-run PromoteFinding (the main agent must review the exact bytes that ran)",
|
|
814
535
|
);
|
|
815
536
|
}
|
|
816
|
-
// The case target must still be the host the PoC ran against
|
|
817
|
-
//
|
|
818
|
-
// target the case adopted after the runs.
|
|
537
|
+
// The case target must still be the host the PoC ran against. The
|
|
538
|
+
// evidence proves nothing about a target the case adopted after the runs.
|
|
819
539
|
const targetRun = bundle.targetRuns[0];
|
|
820
540
|
if (!current.target || current.target !== targetRun.target) {
|
|
821
541
|
throw new Error(
|
|
@@ -823,12 +543,6 @@ export function applyConfirmationResult(
|
|
|
823
543
|
`(bundle target: ${targetRun.target}, case target: ${current.target ?? "(none)"}).`,
|
|
824
544
|
);
|
|
825
545
|
}
|
|
826
|
-
if (!isIntra && current.target === bundle.controlTarget) {
|
|
827
|
-
throw new Error(
|
|
828
|
-
"Case target now equals the control target — the claimed impact is not target-dependent; " +
|
|
829
|
-
"re-run PromoteFinding with a distinct control_target.",
|
|
830
|
-
);
|
|
831
|
-
}
|
|
832
546
|
|
|
833
547
|
// The observation must predate the repro (provenance guard).
|
|
834
548
|
const observation = current.evidenceItems?.find(
|
|
@@ -841,11 +555,6 @@ export function applyConfirmationResult(
|
|
|
841
555
|
);
|
|
842
556
|
}
|
|
843
557
|
|
|
844
|
-
// Phase 1 proves the evidence floor. Phase 2 must freshly replay that same
|
|
845
|
-
// request inside the main agent's ConfirmFinding call; a caller-provided
|
|
846
|
-
// boolean is not accepted as proof of re-execution.
|
|
847
|
-
assertMainAgentVerification(bundle, phase2Verification, isIntra);
|
|
848
|
-
|
|
849
558
|
const reproductionItem: EvidenceItem = {
|
|
850
559
|
id: `ev_${stableShortId(`${id}\nreproduction\n${targetRun.ranAt}`)}`,
|
|
851
560
|
caseId: id,
|
|
@@ -855,7 +564,7 @@ export function applyConfirmationResult(
|
|
|
855
564
|
// exists, so the item stays artifact-backed and re-verifiable.
|
|
856
565
|
artifactPath: targetRun.evidencePath ? basename(targetRun.evidencePath) : "evidence.json",
|
|
857
566
|
sha256: targetRun.evidenceSha256,
|
|
858
|
-
summary: `PoC evidence accepted (2 target runs +
|
|
567
|
+
summary: `PoC evidence accepted (2 target runs + same-host baseline; ${recorded.proofStrength}) — main agent semantic confirmation${verdict.model ? ` (${verdict.model})` : ""}`,
|
|
859
568
|
createdAt: targetRun.ranAt,
|
|
860
569
|
};
|
|
861
570
|
// Defense in depth: the run's evidence.json may embed secrets in
|
|
@@ -897,30 +606,17 @@ export function applyConfirmationResult(
|
|
|
897
606
|
mode: "poc",
|
|
898
607
|
target: targetRun.target,
|
|
899
608
|
},
|
|
900
|
-
controlVerified:
|
|
901
|
-
|
|
902
|
-
|
|
903
|
-
|
|
904
|
-
|
|
905
|
-
|
|
906
|
-
|
|
907
|
-
|
|
908
|
-
|
|
909
|
-
|
|
910
|
-
|
|
911
|
-
target: targetRun.target,
|
|
912
|
-
}
|
|
913
|
-
: {
|
|
914
|
-
path: bundle.controlPath ?? bundle.pocPath,
|
|
915
|
-
exitCode: bundle.controlRun.exitCode,
|
|
916
|
-
ranAt: bundle.controlRun.ranAt,
|
|
917
|
-
output: bundle.controlRun.output,
|
|
918
|
-
sandbox: bundle.controlRun.sandbox,
|
|
919
|
-
completed: true,
|
|
920
|
-
outputComplete: true,
|
|
921
|
-
mode: "control",
|
|
922
|
-
target: bundle.controlRun.target,
|
|
923
|
-
},
|
|
609
|
+
controlVerified: {
|
|
610
|
+
path: bundle.pocPath,
|
|
611
|
+
exitCode: targetRun.exitCode,
|
|
612
|
+
ranAt: targetRun.ranAt,
|
|
613
|
+
output: `same-host baseline: ${bundle.harnessVerified?.control?.note ?? "baseline did not satisfy the attack predicate"}`,
|
|
614
|
+
sandbox: targetRun.sandbox,
|
|
615
|
+
completed: true,
|
|
616
|
+
outputComplete: true,
|
|
617
|
+
mode: "baseline",
|
|
618
|
+
target: targetRun.target,
|
|
619
|
+
},
|
|
924
620
|
disconfirmation: verdict.disconfirmation_attempt,
|
|
925
621
|
confirmerVerdict: recorded,
|
|
926
622
|
pendingConfirmation: undefined,
|
|
@@ -940,8 +636,6 @@ export function applyConfirmationResult(
|
|
|
940
636
|
verdict: "CONFIRMED",
|
|
941
637
|
proof_strength: recorded.proofStrength ?? null,
|
|
942
638
|
model: verdict.model ?? null,
|
|
943
|
-
panel: panelQuorumReached(bundle.panelVotes),
|
|
944
|
-
override: verdict.panel_override_note ? true : false,
|
|
945
639
|
reproduction_evidence_id: reproductionItem.id,
|
|
946
640
|
},
|
|
947
641
|
});
|