@xaccefy/pi-casefile 0.10.0 → 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 +6 -6
- package/package.json +1 -2
- package/src/confirmation.ts +239 -323
- package/src/evidence.ts +101 -135
- package/src/harness-verify.ts +54 -247
- package/src/index.ts +188 -745
- package/src/ledger-internal.ts +118 -4
- package/src/ledger.ts +459 -175
- package/src/poc-runner.ts +43 -14
- package/src/scratchpad.ts +88 -143
- package/src/workflow.ts +6 -2
- package/src/oob-oracle.ts +0 -279
package/src/confirmation.ts
CHANGED
|
@@ -6,18 +6,19 @@
|
|
|
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";
|
|
20
|
-
import { readFileSync } from "node:fs";
|
|
21
|
+
import { lstatSync, readFileSync, statSync } from "node:fs";
|
|
21
22
|
|
|
22
23
|
import { basename } from "node:path";
|
|
23
24
|
|
|
@@ -26,6 +27,7 @@ import {
|
|
|
26
27
|
type MainAgentVerdict,
|
|
27
28
|
normalizeEvidence,
|
|
28
29
|
parsePoCEvidence,
|
|
30
|
+
scanArtifactForSecrets,
|
|
29
31
|
validateMainAgentVerdict,
|
|
30
32
|
} from "./evidence.ts";
|
|
31
33
|
import { type HarnessVerifyResult, sameRequest, verifyUrlBindingError } from "./harness-verify.ts";
|
|
@@ -34,16 +36,17 @@ import type {
|
|
|
34
36
|
CaseUpdateResult,
|
|
35
37
|
EvidenceItem,
|
|
36
38
|
MainAgentVerdictRecord,
|
|
37
|
-
MainAgentVerification,
|
|
38
39
|
NormalizedCaseInput,
|
|
39
40
|
PendingConfirmation,
|
|
40
41
|
PocEvidenceRun,
|
|
41
42
|
} from "./ledger.ts";
|
|
42
43
|
import { getCaseById, readWorkspaceArtifact } from "./ledger.ts";
|
|
43
44
|
import {
|
|
45
|
+
appendCaseEvent,
|
|
44
46
|
buildRecord,
|
|
45
47
|
getDb,
|
|
46
48
|
insertEvidenceItem,
|
|
49
|
+
stableShortId,
|
|
47
50
|
upsertCase,
|
|
48
51
|
validateCase,
|
|
49
52
|
withImmediateTransaction,
|
|
@@ -58,10 +61,159 @@ const PROCESS_STARTED_AS_SUBAGENT = process.env.PI_SUBAGENT_CHILD === "1";
|
|
|
58
61
|
/** Pending confirmation expires after 1h — re-run PromoteFinding for a fresh bundle. */
|
|
59
62
|
export const PENDING_CONFIRM_TTL_MS = 60 * 60 * 1000;
|
|
60
63
|
|
|
61
|
-
//
|
|
62
|
-
|
|
63
|
-
|
|
64
|
+
// ── Report contract gate (confirmed → reported) ──────────────────────
|
|
65
|
+
|
|
66
|
+
/** Typed, fail-closed error for an invalid report contract. */
|
|
67
|
+
export class ReportContractError extends Error {
|
|
68
|
+
readonly code = "REPORT_CONTRACT_INVALID";
|
|
69
|
+
readonly violations: string[];
|
|
70
|
+
|
|
71
|
+
constructor(violations: string[]) {
|
|
72
|
+
super(
|
|
73
|
+
`report contract invalid (${violations.length} violation(s)):\n- ${violations.join("\n- ")}`,
|
|
74
|
+
);
|
|
75
|
+
this.name = "ReportContractError";
|
|
76
|
+
this.violations = violations;
|
|
77
|
+
}
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
/** The closed-schema contract companion path for a report markdown path. */
|
|
81
|
+
export function reportContractPathFor(reportPath: string): string {
|
|
82
|
+
return reportPath.replace(/\.md$/i, ".contract.json");
|
|
64
83
|
}
|
|
84
|
+
|
|
85
|
+
/** Hard cap on the contract document — it is metadata, not a report carrier. */
|
|
86
|
+
const REPORT_CONTRACT_MAX_BYTES = 64 * 1024;
|
|
87
|
+
|
|
88
|
+
/** Keys the closed schema accepts; anything else is a violation. */
|
|
89
|
+
const REPORT_CONTRACT_KEYS = new Set([
|
|
90
|
+
"case_id",
|
|
91
|
+
"title",
|
|
92
|
+
"severity",
|
|
93
|
+
"summary",
|
|
94
|
+
"impact",
|
|
95
|
+
"remediation",
|
|
96
|
+
"steps",
|
|
97
|
+
"evidence_ids",
|
|
98
|
+
"coverage_refs",
|
|
99
|
+
]);
|
|
100
|
+
|
|
101
|
+
/**
|
|
102
|
+
* Validate the closed-schema report contract for a confirmed case:
|
|
103
|
+
* - a regular, non-symlink JSON file of bounded size exists at contractPath;
|
|
104
|
+
* - only schema keys are present, and the required text fields are non-empty;
|
|
105
|
+
* - evidence_ids reference ONLY evidence items that exist on this case, and
|
|
106
|
+
* include at least one observation and one reproduction item;
|
|
107
|
+
* - coverage_refs reference ONLY (asset, class) cells recorded on this case.
|
|
108
|
+
*
|
|
109
|
+
* Throws ReportContractError (fail closed) on any violation.
|
|
110
|
+
*/
|
|
111
|
+
export function validateReportContract(record: CaseRecord, contractPath: string): void {
|
|
112
|
+
const violations: string[] = [];
|
|
113
|
+
let stat: ReturnType<typeof statSync>;
|
|
114
|
+
try {
|
|
115
|
+
stat = statSync(contractPath);
|
|
116
|
+
} catch {
|
|
117
|
+
throw new ReportContractError([
|
|
118
|
+
`report contract not found: ${basename(contractPath)} (write the closed-schema JSON contract next to the report, then retry status='reported')`,
|
|
119
|
+
]);
|
|
120
|
+
}
|
|
121
|
+
if (!stat.isFile()) violations.push("report contract path is not a regular file");
|
|
122
|
+
if (lstatSync(contractPath).isSymbolicLink()) {
|
|
123
|
+
violations.push("report contract must not be a symbolic link");
|
|
124
|
+
}
|
|
125
|
+
if (stat.size > REPORT_CONTRACT_MAX_BYTES) {
|
|
126
|
+
violations.push(
|
|
127
|
+
`report contract too large (${stat.size} bytes; max ${REPORT_CONTRACT_MAX_BYTES})`,
|
|
128
|
+
);
|
|
129
|
+
}
|
|
130
|
+
if (violations.length > 0) throw new ReportContractError(violations);
|
|
131
|
+
|
|
132
|
+
let doc: unknown;
|
|
133
|
+
try {
|
|
134
|
+
doc = JSON.parse(readFileSync(contractPath, "utf8"));
|
|
135
|
+
} catch (e) {
|
|
136
|
+
throw new ReportContractError([`report contract is not valid JSON: ${(e as Error).message}`]);
|
|
137
|
+
}
|
|
138
|
+
if (typeof doc !== "object" || doc === null || Array.isArray(doc)) {
|
|
139
|
+
throw new ReportContractError(["report contract must be a JSON object"]);
|
|
140
|
+
}
|
|
141
|
+
const contract = doc as Record<string, unknown>;
|
|
142
|
+
for (const key of Object.keys(contract)) {
|
|
143
|
+
if (!REPORT_CONTRACT_KEYS.has(key)) {
|
|
144
|
+
violations.push(`unknown key "${key}" — the report contract schema is closed`);
|
|
145
|
+
}
|
|
146
|
+
}
|
|
147
|
+
for (const required of ["case_id", "title", "severity", "summary", "impact", "remediation"]) {
|
|
148
|
+
const v = contract[required];
|
|
149
|
+
if (typeof v !== "string" || v.trim().length === 0) {
|
|
150
|
+
violations.push(`"${required}" must be a non-empty string`);
|
|
151
|
+
}
|
|
152
|
+
}
|
|
153
|
+
if (contract.case_id !== record.id) {
|
|
154
|
+
violations.push(`"case_id" must be ${record.id} (got ${String(contract.case_id)})`);
|
|
155
|
+
}
|
|
156
|
+
const SEVERITIES = ["info", "low", "medium", "high", "critical"];
|
|
157
|
+
if (
|
|
158
|
+
typeof contract.severity === "string" &&
|
|
159
|
+
!(record.severity
|
|
160
|
+
? contract.severity === record.severity
|
|
161
|
+
: SEVERITIES.includes(contract.severity))
|
|
162
|
+
) {
|
|
163
|
+
violations.push(
|
|
164
|
+
`"severity" must match the case severity (${record.severity ?? "unset"}) or be a valid severity`,
|
|
165
|
+
);
|
|
166
|
+
}
|
|
167
|
+
if (!Array.isArray(contract.steps) || contract.steps.length === 0) {
|
|
168
|
+
violations.push('"steps" must be a non-empty array of reproduction steps');
|
|
169
|
+
} else if (!contract.steps.every((s: unknown) => typeof s === "string" && s.trim().length > 0)) {
|
|
170
|
+
violations.push('"steps" entries must be non-empty strings');
|
|
171
|
+
}
|
|
172
|
+
|
|
173
|
+
const items = record.evidenceItems ?? [];
|
|
174
|
+
const knownIds = new Set(items.map((i) => i.id));
|
|
175
|
+
const evidenceIds = contract.evidence_ids;
|
|
176
|
+
if (!Array.isArray(evidenceIds) || evidenceIds.length === 0) {
|
|
177
|
+
violations.push('"evidence_ids" must be a non-empty array of evidence item ids');
|
|
178
|
+
} else {
|
|
179
|
+
if (!evidenceIds.every((id: unknown) => typeof id === "string" && knownIds.has(id))) {
|
|
180
|
+
violations.push('"evidence_ids" references evidence items that do not exist on this case');
|
|
181
|
+
}
|
|
182
|
+
const referenced = items.filter((i) => evidenceIds.includes(i.id));
|
|
183
|
+
if (!referenced.some((i) => i.role === "observation")) {
|
|
184
|
+
violations.push('"evidence_ids" must include at least one observation item');
|
|
185
|
+
}
|
|
186
|
+
if (!referenced.some((i) => i.role === "reproduction")) {
|
|
187
|
+
violations.push('"evidence_ids" must include at least one reproduction item');
|
|
188
|
+
}
|
|
189
|
+
}
|
|
190
|
+
|
|
191
|
+
const coverageRefs = contract.coverage_refs;
|
|
192
|
+
if (coverageRefs !== undefined && !Array.isArray(coverageRefs)) {
|
|
193
|
+
violations.push('"coverage_refs" must be an array of { asset, class } objects');
|
|
194
|
+
} else if (Array.isArray(coverageRefs)) {
|
|
195
|
+
const cells = new Set((record.coverageItems ?? []).map((c) => `${c.asset}\n${c.class}`));
|
|
196
|
+
for (const [index, ref] of coverageRefs.entries()) {
|
|
197
|
+
if (
|
|
198
|
+
typeof ref !== "object" ||
|
|
199
|
+
ref === null ||
|
|
200
|
+
typeof (ref as Record<string, unknown>).asset !== "string" ||
|
|
201
|
+
typeof (ref as Record<string, unknown>).class !== "string"
|
|
202
|
+
) {
|
|
203
|
+
violations.push(`"coverage_refs[${index}]" must be an { asset, class } object`);
|
|
204
|
+
} else if (
|
|
205
|
+
!cells.has(`${(ref as { asset: string }).asset}\n${(ref as { class: string }).class}`)
|
|
206
|
+
) {
|
|
207
|
+
violations.push(
|
|
208
|
+
`"coverage_refs[${index}]" references a coverage cell not recorded on this case`,
|
|
209
|
+
);
|
|
210
|
+
}
|
|
211
|
+
}
|
|
212
|
+
}
|
|
213
|
+
|
|
214
|
+
if (violations.length > 0) throw new ReportContractError(violations);
|
|
215
|
+
}
|
|
216
|
+
|
|
65
217
|
function validateRunEvidence(run: PocEvidenceRun, label: string): void {
|
|
66
218
|
if (!run.completed) {
|
|
67
219
|
throw new Error(`${label} did not complete; a crash is not evidence`);
|
|
@@ -117,56 +269,21 @@ function validateRunEvidence(run: PocEvidenceRun, label: string): void {
|
|
|
117
269
|
}
|
|
118
270
|
}
|
|
119
271
|
|
|
120
|
-
/** Determinism
|
|
121
|
-
function assertEvidenceDifferential(bundle: PendingConfirmation
|
|
272
|
+
/** Determinism on normalized evidence (nonce/observations stripped). */
|
|
273
|
+
function assertEvidenceDifferential(bundle: PendingConfirmation): void {
|
|
122
274
|
const [r1, r2] = bundle.targetRuns;
|
|
123
275
|
if (normalizeEvidence(r1.evidence) !== normalizeEvidence(r2.evidence)) {
|
|
124
276
|
throw new Error(
|
|
125
277
|
"Target runs produced inconsistent evidence — the exploit did not reproduce deterministically",
|
|
126
278
|
);
|
|
127
279
|
}
|
|
128
|
-
// Intra-target target-dependence is proven by the harness attack-vs-baseline
|
|
129
|
-
// replay (same host), not by comparing a target run to a separate control run.
|
|
130
|
-
if (isIntra) return;
|
|
131
|
-
// OOB-only bundles prove target-dependence via the oracle token differential
|
|
132
|
-
// (assertMachineConfirmation judges callbackVerified); no control run exists.
|
|
133
|
-
if (!bundle.controlRun && bundle.callbackVerified?.attempted) return;
|
|
134
|
-
if (!bundle.controlRun) {
|
|
135
|
-
throw new Error("inter-host confirmation requires a control run");
|
|
136
|
-
}
|
|
137
|
-
if (normalizeEvidence(r1.evidence) === normalizeEvidence(bundle.controlRun.evidence)) {
|
|
138
|
-
throw new Error(
|
|
139
|
-
"Control run produced identical evidence to the target — the claimed impact is not target-dependent",
|
|
140
|
-
);
|
|
141
|
-
}
|
|
142
280
|
}
|
|
143
281
|
|
|
144
282
|
function assertMachineConfirmation(bundle: PendingConfirmation): void {
|
|
145
|
-
const oob = bundle.callbackVerified;
|
|
146
|
-
if (oob?.attempted) {
|
|
147
|
-
if (oob.targetHits === 0) {
|
|
148
|
-
throw new Error(
|
|
149
|
-
`OOB VERIFY FAILED: no interaction with the target-run callback token. ${oob.note}`,
|
|
150
|
-
);
|
|
151
|
-
}
|
|
152
|
-
if (oob.controlHits > 0) {
|
|
153
|
-
throw new Error(
|
|
154
|
-
`OOB VERIFY FAILED: the control-run callback token received ${oob.controlHits} interaction(s) — the callback is not target-dependent. ${oob.note}`,
|
|
155
|
-
);
|
|
156
|
-
}
|
|
157
|
-
if (oob.sourceSeparated !== true) {
|
|
158
|
-
throw new Error(
|
|
159
|
-
"OOB VERIFY FAILED: callback source separation was not established. " +
|
|
160
|
-
"A loopback listener reachable by the PoC is diagnostic telemetry, not proof that the target caused the interaction.",
|
|
161
|
-
);
|
|
162
|
-
}
|
|
163
|
-
return;
|
|
164
|
-
}
|
|
165
|
-
|
|
166
283
|
assertHarnessTargetOnly(
|
|
167
284
|
bundle.harnessVerified,
|
|
168
285
|
"HARNESS DIFFERENTIAL FAILED",
|
|
169
|
-
"no machine-owned
|
|
286
|
+
"no machine-owned attack/baseline replay was recorded",
|
|
170
287
|
);
|
|
171
288
|
}
|
|
172
289
|
|
|
@@ -186,79 +303,6 @@ function assertHarnessTargetOnly(
|
|
|
186
303
|
}
|
|
187
304
|
}
|
|
188
305
|
|
|
189
|
-
function assertHarnessCanary(
|
|
190
|
-
harness: HarnessVerifyResult | undefined,
|
|
191
|
-
required: boolean,
|
|
192
|
-
label: string,
|
|
193
|
-
): void {
|
|
194
|
-
if (!required) return;
|
|
195
|
-
if (
|
|
196
|
-
harness?.canary?.attempted !== true ||
|
|
197
|
-
harness.canary.pass !== true ||
|
|
198
|
-
harness.canary.targetObserved !== true ||
|
|
199
|
-
harness.canary.controlObserved !== false ||
|
|
200
|
-
harness.proofStrength !== "canary_differential"
|
|
201
|
-
) {
|
|
202
|
-
throw new Error(`${label}: ${harness?.canary?.note ?? "required canary transcript missing"}`);
|
|
203
|
-
}
|
|
204
|
-
}
|
|
205
|
-
|
|
206
|
-
function assertMainAgentVerification(
|
|
207
|
-
bundle: PendingConfirmation,
|
|
208
|
-
verification: MainAgentVerification | undefined,
|
|
209
|
-
isIntra = false,
|
|
210
|
-
): asserts verification is MainAgentVerification {
|
|
211
|
-
if (!verification) {
|
|
212
|
-
throw new Error(
|
|
213
|
-
"MAIN-AGENT REPLAY REQUIRED: ConfirmFinding must produce a fresh harness-owned target/control transcript",
|
|
214
|
-
);
|
|
215
|
-
}
|
|
216
|
-
const at = Date.parse(verification.at);
|
|
217
|
-
const bundleAt = Date.parse(bundle.ranAt);
|
|
218
|
-
const now = Date.now();
|
|
219
|
-
if (
|
|
220
|
-
!Number.isFinite(at) ||
|
|
221
|
-
!Number.isFinite(bundleAt) ||
|
|
222
|
-
at < bundleAt ||
|
|
223
|
-
at > now + 30_000 ||
|
|
224
|
-
now - at > 5 * 60 * 1000
|
|
225
|
-
) {
|
|
226
|
-
throw new Error(
|
|
227
|
-
"MAIN-AGENT REPLAY FAILED: transcript timestamp must be valid, newer than phase 1, and no more than 5 minutes old",
|
|
228
|
-
);
|
|
229
|
-
}
|
|
230
|
-
assertHarnessTargetOnly(
|
|
231
|
-
verification.result,
|
|
232
|
-
"MAIN-AGENT REPLAY FAILED",
|
|
233
|
-
"no fresh phase-2 target/control replay was recorded",
|
|
234
|
-
);
|
|
235
|
-
assertHarnessCanary(
|
|
236
|
-
verification.result,
|
|
237
|
-
bundle.targetRuns[0].evidence.verify.canary !== undefined,
|
|
238
|
-
"MAIN-AGENT CANARY FAILED",
|
|
239
|
-
);
|
|
240
|
-
// OOB-only bundles bind by TOKEN identity (enforced at store time on
|
|
241
|
-
// evidence.verify.url); there is no control host to bind a transcript to.
|
|
242
|
-
if (bundle.callbackVerified?.attempted && bundle.oobTokens) return;
|
|
243
|
-
const targetUrl = verification.result.target?.url;
|
|
244
|
-
const controlUrl = verification.result.control?.url;
|
|
245
|
-
const targetIdentity = bundle.targetRuns[0].target;
|
|
246
|
-
if (!targetUrl || verifyUrlBindingError(targetUrl, targetIdentity)) {
|
|
247
|
-
throw new Error("MAIN-AGENT REPLAY FAILED: target transcript is not bound to the case target");
|
|
248
|
-
}
|
|
249
|
-
// Intra-target: the "control" transcript is the legitimate baseline request,
|
|
250
|
-
// which is bound to the SAME case target. Inter-host: it is bound to the
|
|
251
|
-
// distinct control target.
|
|
252
|
-
const controlBindTarget = isIntra ? targetIdentity : bundle.controlTarget;
|
|
253
|
-
if (!controlUrl || !controlBindTarget || verifyUrlBindingError(controlUrl, controlBindTarget)) {
|
|
254
|
-
throw new Error(
|
|
255
|
-
isIntra
|
|
256
|
-
? "MAIN-AGENT REPLAY FAILED: baseline transcript is not bound to the case target"
|
|
257
|
-
: "MAIN-AGENT REPLAY FAILED: control transcript is not bound to control_target",
|
|
258
|
-
);
|
|
259
|
-
}
|
|
260
|
-
}
|
|
261
|
-
|
|
262
306
|
/**
|
|
263
307
|
* Gate for phase 1 of promotion: case must exist, be investigating, and have
|
|
264
308
|
* poc/evidence/impact/severity/target. The disconfirmation is provided by the
|
|
@@ -306,27 +350,16 @@ export function assertPromotable(id: string): CaseRecord {
|
|
|
306
350
|
}
|
|
307
351
|
|
|
308
352
|
/**
|
|
309
|
-
* Phase 1
|
|
310
|
-
*
|
|
311
|
-
*
|
|
312
|
-
*
|
|
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.
|
|
313
357
|
*/
|
|
314
|
-
function
|
|
315
|
-
|
|
316
|
-
id: string,
|
|
317
|
-
bundle: PendingConfirmation,
|
|
318
|
-
): CaseRecord {
|
|
319
|
-
if (bundle.targetRuns.length !== 2) {
|
|
320
|
-
throw new Error("Intra-target confirmation requires two target runs");
|
|
321
|
-
}
|
|
322
|
-
if (bundle.controlRun || bundle.controlTarget) {
|
|
323
|
-
throw new Error(
|
|
324
|
-
"Intra-target confirmation must not carry a control run or control target — the baseline is a same-host request inside the evidence",
|
|
325
|
-
);
|
|
326
|
-
}
|
|
358
|
+
function validateBundle(current: CaseRecord, id: string, bundle: PendingConfirmation): CaseRecord {
|
|
359
|
+
if (bundle.targetRuns.length !== 2) throw new Error("Confirmation requires two target runs");
|
|
327
360
|
const targetRunTarget = bundle.targetRuns[0]?.target;
|
|
328
361
|
if (!targetRunTarget || bundle.targetRuns.some((r) => r.target !== targetRunTarget)) {
|
|
329
|
-
throw new Error("
|
|
362
|
+
throw new Error("Confirmation requires both runs against the same case target");
|
|
330
363
|
}
|
|
331
364
|
let pocHash: string | undefined;
|
|
332
365
|
try {
|
|
@@ -340,35 +373,20 @@ function validateIntraTargetBundle(
|
|
|
340
373
|
for (const run of bundle.targetRuns) {
|
|
341
374
|
validateRunEvidence(run, `${run.mode} run`);
|
|
342
375
|
const ev = run.evidence;
|
|
343
|
-
if (ev.verify.mode !== "intra_target") {
|
|
344
|
-
throw new Error(
|
|
345
|
-
"INTRA-TARGET FAILED: each run's evidence.verify.mode must be 'intra_target'",
|
|
346
|
-
);
|
|
347
|
-
}
|
|
348
|
-
if (!ev.baseline) {
|
|
349
|
-
throw new Error(
|
|
350
|
-
"INTRA-TARGET FAILED: evidence.baseline (a legitimate same-host request) is required",
|
|
351
|
-
);
|
|
352
|
-
}
|
|
353
376
|
const attackBinding = verifyUrlBindingError(ev.verify.url, targetRunTarget);
|
|
354
377
|
if (attackBinding) throw new Error(`ATTACK BINDING FAILED: ${attackBinding}`);
|
|
355
378
|
const baselineBinding = verifyUrlBindingError(ev.baseline.url, targetRunTarget);
|
|
356
379
|
if (baselineBinding) throw new Error(`BASELINE BINDING FAILED: ${baselineBinding}`);
|
|
357
380
|
if (ev.baseline && sameRequest(ev.verify, ev.baseline)) {
|
|
358
381
|
throw new Error(
|
|
359
|
-
"
|
|
382
|
+
"BASELINE CHECK FAILED: attack and baseline requests are identical — vary identity or a parameter",
|
|
360
383
|
);
|
|
361
384
|
}
|
|
362
385
|
}
|
|
363
386
|
if (bundle.caseId !== id) throw new Error("Pending confirmation caseId mismatch");
|
|
364
|
-
assertEvidenceDifferential(bundle
|
|
387
|
+
assertEvidenceDifferential(bundle);
|
|
365
388
|
// Machine floor: attack matched, baseline did not, both against the case target.
|
|
366
389
|
assertMachineConfirmation(bundle);
|
|
367
|
-
assertHarnessCanary(
|
|
368
|
-
bundle.harnessVerified,
|
|
369
|
-
bundle.targetRuns[0].evidence.verify.canary !== undefined,
|
|
370
|
-
"PHASE-1 CANARY FAILED",
|
|
371
|
-
);
|
|
372
390
|
const next = buildRecord({ pendingConfirmation: bundle }, current);
|
|
373
391
|
validateCase(next);
|
|
374
392
|
return next;
|
|
@@ -376,9 +394,9 @@ function validateIntraTargetBundle(
|
|
|
376
394
|
|
|
377
395
|
/**
|
|
378
396
|
* Phase 1: record the harness-observed evidence bundle on the case. The whole
|
|
379
|
-
* contract is validated here —
|
|
380
|
-
*
|
|
381
|
-
*
|
|
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
|
|
382
400
|
* main agent performs phase-2 review.
|
|
383
401
|
*/
|
|
384
402
|
export function storePendingConfirmation(id: string, bundle: PendingConfirmation): CaseRecord {
|
|
@@ -392,107 +410,23 @@ export function storePendingConfirmation(id: string, bundle: PendingConfirmation
|
|
|
392
410
|
);
|
|
393
411
|
}
|
|
394
412
|
if (bundle.caseId !== id) throw new Error("Pending confirmation caseId mismatch");
|
|
395
|
-
if (bundle.mode === "intra_target") {
|
|
396
|
-
const next = validateIntraTargetBundle(current, id, bundle);
|
|
397
|
-
upsertCase(db, next);
|
|
398
|
-
return next;
|
|
399
|
-
}
|
|
400
|
-
// Control-run requirements key off controlRun PRESENCE, not the OOB flag:
|
|
401
|
-
// an OOB-only bundle has no control run (token differential instead), but
|
|
402
|
-
// an OOB+control bundle still carries one and gets the full checks.
|
|
403
|
-
if (!bundle.controlRun && !bundle.callbackVerified) {
|
|
404
|
-
throw new Error("Pending confirmation requires two target runs and one control run");
|
|
405
|
-
}
|
|
406
|
-
if (bundle.controlRun && (!bundle.pocPath || !bundle.controlPath || !bundle.controlTarget)) {
|
|
407
|
-
throw new Error("Pending confirmation requires pocPath, controlPath, and controlTarget");
|
|
408
|
-
}
|
|
409
|
-
// Control-target binding (machine-verified here, not just in the tool
|
|
410
|
-
// layer): the control run must actually have targeted the declared
|
|
411
|
-
// control_target, that target must differ from the target runs' target,
|
|
412
|
-
// and the control target must differ from the case's target — otherwise
|
|
413
|
-
// "the control demonstrated nothing on the vulnerable target" passes.
|
|
414
|
-
const targetRunTarget = bundle.targetRuns[0]?.target;
|
|
415
|
-
if (!targetRunTarget || bundle.targetRuns.some((r) => r.target !== targetRunTarget)) {
|
|
416
|
-
throw new Error(
|
|
417
|
-
"Pending confirmation requires both target runs against the same case target",
|
|
418
|
-
);
|
|
419
|
-
}
|
|
420
|
-
if (bundle.controlRun) {
|
|
421
|
-
if (!bundle.controlRun.target || bundle.controlRun.target !== bundle.controlTarget) {
|
|
422
|
-
throw new Error(
|
|
423
|
-
"CONTROL BINDING FAILED: controlRun.target must equal control_target — a control run " +
|
|
424
|
-
"against a different host than the one declared proves nothing.",
|
|
425
|
-
);
|
|
426
|
-
}
|
|
427
|
-
if (bundle.controlRun.target === targetRunTarget) {
|
|
428
|
-
throw new Error(
|
|
429
|
-
"CONTROL BINDING FAILED: the control run targeted the same host as the target runs — " +
|
|
430
|
-
"the claimed impact is not target-dependent.",
|
|
431
|
-
);
|
|
432
|
-
}
|
|
433
|
-
}
|
|
434
|
-
if (bundle.controlTarget && bundle.controlTarget === current.target) {
|
|
435
|
-
throw new Error(
|
|
436
|
-
"CONTROL BINDING FAILED: control_target must differ from the case target; a control run " +
|
|
437
|
-
"against the vulnerable target proves nothing.",
|
|
438
|
-
);
|
|
439
|
-
}
|
|
440
|
-
// Same-file contract re-checked at store time (the tool already checked).
|
|
441
|
-
// OOB-only bundles carry no separate control script — the PoC hash alone
|
|
442
|
-
// is re-verified against the file on disk.
|
|
443
413
|
let pocHash: string | undefined;
|
|
444
|
-
let controlHash: string | undefined;
|
|
445
414
|
try {
|
|
446
415
|
pocHash = createHash("sha256").update(readFileSync(bundle.pocPath)).digest("hex");
|
|
447
|
-
controlHash = bundle.controlPath
|
|
448
|
-
? createHash("sha256").update(readFileSync(bundle.controlPath)).digest("hex")
|
|
449
|
-
: pocHash;
|
|
450
416
|
} catch {
|
|
451
417
|
pocHash = undefined;
|
|
452
|
-
controlHash = undefined;
|
|
453
418
|
}
|
|
454
|
-
if (!pocHash ||
|
|
455
|
-
throw new Error(
|
|
456
|
-
"CONTROL CHECK FAILED: control_path must be the SAME script as poc_path " +
|
|
457
|
-
"(sha256 mismatch). A separately written control file proves nothing.",
|
|
458
|
-
);
|
|
459
|
-
}
|
|
460
|
-
if (bundle.pocSha256 && bundle.pocSha256 !== pocHash) {
|
|
419
|
+
if (!pocHash || (bundle.pocSha256 && bundle.pocSha256 !== pocHash)) {
|
|
461
420
|
throw new Error("pocSha256 does not match the PoC file on disk");
|
|
462
421
|
}
|
|
463
|
-
|
|
464
|
-
// OOB+control bundles validate all three.
|
|
465
|
-
const runsToValidate = bundle.controlRun
|
|
466
|
-
? [...bundle.targetRuns, bundle.controlRun]
|
|
467
|
-
: [...bundle.targetRuns];
|
|
468
|
-
for (const run of runsToValidate) {
|
|
469
|
-
validateRunEvidence(run, `${run.mode} run`);
|
|
470
|
-
}
|
|
471
|
-
assertEvidenceDifferential(bundle);
|
|
472
|
-
// Target binding applies to EVERY mode — an OOB bundle's verify.url must
|
|
473
|
-
// still belong to the case target, or the PoC could anchor its evidence on
|
|
474
|
-
// an unrelated host while the callback alone carries the proof.
|
|
475
|
-
for (const run of bundle.targetRuns) {
|
|
476
|
-
const bindingError = verifyUrlBindingError(run.evidence.verify.url, targetRunTarget);
|
|
477
|
-
if (bindingError) throw new Error(`TARGET BINDING FAILED: ${bindingError}`);
|
|
478
|
-
}
|
|
479
|
-
if (bundle.controlRun) {
|
|
480
|
-
const controlBindingError = verifyUrlBindingError(
|
|
481
|
-
bundle.controlRun.evidence.verify.url,
|
|
482
|
-
bundle.controlTarget!,
|
|
483
|
-
);
|
|
484
|
-
if (controlBindingError) {
|
|
485
|
-
throw new Error(`CONTROL BINDING FAILED: ${controlBindingError}`);
|
|
486
|
-
}
|
|
487
|
-
}
|
|
488
|
-
// A clean exit and model-authored evidence are necessary inputs, never the
|
|
489
|
-
// proof. Promotion requires a harness-observed target/control differential
|
|
490
|
-
// or a harness-owned OOB interaction differential.
|
|
491
|
-
assertMachineConfirmation(bundle);
|
|
492
|
-
|
|
493
|
-
const next = buildRecord({ pendingConfirmation: bundle }, current);
|
|
494
|
-
validateCase(next);
|
|
422
|
+
const next = validateBundle(current, id, bundle);
|
|
495
423
|
upsertCase(db, next);
|
|
424
|
+
appendCaseEvent(db, {
|
|
425
|
+
actor: "harness",
|
|
426
|
+
caseId: id,
|
|
427
|
+
eventType: "promotion_pending",
|
|
428
|
+
payload: { mode: "intra_target", evidence_sha256: bundle.targetRuns[0].evidenceSha256 },
|
|
429
|
+
});
|
|
496
430
|
return next;
|
|
497
431
|
});
|
|
498
432
|
}
|
|
@@ -503,16 +437,14 @@ export function storePendingConfirmation(id: string, bundle: PendingConfirmation
|
|
|
503
437
|
* CONFIRMED requires the full bundle to still hold (completion, nonce,
|
|
504
438
|
* determinism, differential), the PoC script to be unchanged since the runs
|
|
505
439
|
* (pocSha256 — otherwise the main agent reviewed different bytes), and a
|
|
506
|
-
* verdict accompanied by a
|
|
507
|
-
*
|
|
508
|
-
* disproved)
|
|
509
|
-
*
|
|
510
|
-
* 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.
|
|
511
444
|
*/
|
|
512
445
|
export function applyConfirmationResult(
|
|
513
446
|
id: string,
|
|
514
447
|
verdictInput: MainAgentVerdict,
|
|
515
|
-
phase2Verification?: MainAgentVerification,
|
|
516
448
|
authority: { startedAsSubagent: boolean } = {
|
|
517
449
|
startedAsSubagent: PROCESS_STARTED_AS_SUBAGENT || process.env.PI_SUBAGENT_CHILD === "1",
|
|
518
450
|
},
|
|
@@ -545,30 +477,11 @@ export function applyConfirmationResult(
|
|
|
545
477
|
const parsed = validateMainAgentVerdict(verdictInput);
|
|
546
478
|
if (!parsed.ok) throw new Error(`Invalid main-agent confirmation verdict: ${parsed.error}`);
|
|
547
479
|
const verdict = parsed.verdict;
|
|
548
|
-
const canaryRequested = bundle.targetRuns[0].evidence.verify.canary !== undefined;
|
|
549
|
-
if (verdict.verdict === "CONFIRMED") {
|
|
550
|
-
if (canaryRequested && verdict.canary_assessment !== "verified") {
|
|
551
|
-
throw new Error(
|
|
552
|
-
"CONFIRMED canary mismatch: evidence requested a harness canary, so canary_assessment must be verified",
|
|
553
|
-
);
|
|
554
|
-
}
|
|
555
|
-
if (!canaryRequested && verdict.canary_assessment !== "not_applicable") {
|
|
556
|
-
throw new Error(
|
|
557
|
-
"CONFIRMED canary mismatch: this evidence has no canary template; record canary_assessment=not_applicable and explain why",
|
|
558
|
-
);
|
|
559
|
-
}
|
|
560
|
-
}
|
|
561
480
|
const recorded: MainAgentVerdictRecord = {
|
|
562
481
|
...verdict,
|
|
563
482
|
at: new Date().toISOString(),
|
|
564
483
|
reviewer: "main_agent",
|
|
565
|
-
|
|
566
|
-
proofStrength:
|
|
567
|
-
verdict.verdict === "CONFIRMED"
|
|
568
|
-
? canaryRequested
|
|
569
|
-
? "canary_differential"
|
|
570
|
-
: "predicate_differential"
|
|
571
|
-
: undefined,
|
|
484
|
+
proofStrength: verdict.verdict === "CONFIRMED" ? "predicate_differential" : undefined,
|
|
572
485
|
};
|
|
573
486
|
|
|
574
487
|
if (verdict.verdict !== "CONFIRMED") {
|
|
@@ -594,21 +507,22 @@ export function applyConfirmationResult(
|
|
|
594
507
|
next.pendingConfirmation = undefined;
|
|
595
508
|
validateCase(next);
|
|
596
509
|
upsertCase(db, next);
|
|
510
|
+
appendCaseEvent(db, {
|
|
511
|
+
caseId: id,
|
|
512
|
+
actor: "main_agent",
|
|
513
|
+
eventType: "confirmation_verdict",
|
|
514
|
+
payload: { verdict: verdict.verdict, model: verdict.model ?? null },
|
|
515
|
+
});
|
|
597
516
|
return { record: next, changed: true };
|
|
598
517
|
}
|
|
599
518
|
|
|
600
519
|
// CONFIRMED — re-validate the whole bundle (defense in depth; the case may
|
|
601
520
|
// have been touched between phase 1 and the verdict).
|
|
602
|
-
const
|
|
603
|
-
const allRuns = isIntra
|
|
604
|
-
? [...bundle.targetRuns]
|
|
605
|
-
: [...bundle.targetRuns, ...(bundle.controlRun ? [bundle.controlRun] : [])];
|
|
606
|
-
for (const run of allRuns) {
|
|
521
|
+
for (const run of bundle.targetRuns) {
|
|
607
522
|
validateRunEvidence(run, `${run.mode} run`);
|
|
608
523
|
}
|
|
609
|
-
assertEvidenceDifferential(bundle
|
|
524
|
+
assertEvidenceDifferential(bundle);
|
|
610
525
|
assertMachineConfirmation(bundle);
|
|
611
|
-
assertHarnessCanary(bundle.harnessVerified, canaryRequested, "PHASE-1 CANARY FAILED");
|
|
612
526
|
let pocHash: string | undefined;
|
|
613
527
|
try {
|
|
614
528
|
pocHash = createHash("sha256").update(readFileSync(bundle.pocPath)).digest("hex");
|
|
@@ -620,9 +534,8 @@ export function applyConfirmationResult(
|
|
|
620
534
|
"PoC script changed since the runs — re-run PromoteFinding (the main agent must review the exact bytes that ran)",
|
|
621
535
|
);
|
|
622
536
|
}
|
|
623
|
-
// The case target must still be the host the PoC ran against
|
|
624
|
-
//
|
|
625
|
-
// 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.
|
|
626
539
|
const targetRun = bundle.targetRuns[0];
|
|
627
540
|
if (!current.target || current.target !== targetRun.target) {
|
|
628
541
|
throw new Error(
|
|
@@ -630,12 +543,6 @@ export function applyConfirmationResult(
|
|
|
630
543
|
`(bundle target: ${targetRun.target}, case target: ${current.target ?? "(none)"}).`,
|
|
631
544
|
);
|
|
632
545
|
}
|
|
633
|
-
if (!isIntra && current.target === bundle.controlTarget) {
|
|
634
|
-
throw new Error(
|
|
635
|
-
"Case target now equals the control target — the claimed impact is not target-dependent; " +
|
|
636
|
-
"re-run PromoteFinding with a distinct control_target.",
|
|
637
|
-
);
|
|
638
|
-
}
|
|
639
546
|
|
|
640
547
|
// The observation must predate the repro (provenance guard).
|
|
641
548
|
const observation = current.evidenceItems?.find(
|
|
@@ -648,11 +555,6 @@ export function applyConfirmationResult(
|
|
|
648
555
|
);
|
|
649
556
|
}
|
|
650
557
|
|
|
651
|
-
// Phase 1 proves the evidence floor. Phase 2 must freshly replay that same
|
|
652
|
-
// request inside the main agent's ConfirmFinding call; a caller-provided
|
|
653
|
-
// boolean is not accepted as proof of re-execution.
|
|
654
|
-
assertMainAgentVerification(bundle, phase2Verification, isIntra);
|
|
655
|
-
|
|
656
558
|
const reproductionItem: EvidenceItem = {
|
|
657
559
|
id: `ev_${stableShortId(`${id}\nreproduction\n${targetRun.ranAt}`)}`,
|
|
658
560
|
caseId: id,
|
|
@@ -662,9 +564,25 @@ export function applyConfirmationResult(
|
|
|
662
564
|
// exists, so the item stays artifact-backed and re-verifiable.
|
|
663
565
|
artifactPath: targetRun.evidencePath ? basename(targetRun.evidencePath) : "evidence.json",
|
|
664
566
|
sha256: targetRun.evidenceSha256,
|
|
665
|
-
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 ? ` (${verdict.model})` : ""}`,
|
|
666
568
|
createdAt: targetRun.ranAt,
|
|
667
569
|
};
|
|
570
|
+
// Defense in depth: the run's evidence.json may embed secrets in
|
|
571
|
+
// observations/claim text — flag it like any other artifact.
|
|
572
|
+
if (targetRun.evidencePath) {
|
|
573
|
+
try {
|
|
574
|
+
const secretFindings = scanArtifactForSecrets(
|
|
575
|
+
readWorkspaceArtifact(targetRun.evidencePath).bytes,
|
|
576
|
+
);
|
|
577
|
+
if (secretFindings.length > 0) {
|
|
578
|
+
reproductionItem.containsSecret = true;
|
|
579
|
+
reproductionItem.secretFindings = secretFindings;
|
|
580
|
+
}
|
|
581
|
+
} catch {
|
|
582
|
+
// validateRunEvidence already proved the artifact readable; a scan
|
|
583
|
+
// failure never blocks the confirmation itself.
|
|
584
|
+
}
|
|
585
|
+
}
|
|
668
586
|
|
|
669
587
|
const newEvidence =
|
|
670
588
|
(current.evidence ? `${current.evidence}\n\n` : "") +
|
|
@@ -688,30 +606,17 @@ export function applyConfirmationResult(
|
|
|
688
606
|
mode: "poc",
|
|
689
607
|
target: targetRun.target,
|
|
690
608
|
},
|
|
691
|
-
controlVerified:
|
|
692
|
-
|
|
693
|
-
|
|
694
|
-
|
|
695
|
-
|
|
696
|
-
|
|
697
|
-
|
|
698
|
-
|
|
699
|
-
|
|
700
|
-
|
|
701
|
-
|
|
702
|
-
target: targetRun.target,
|
|
703
|
-
}
|
|
704
|
-
: {
|
|
705
|
-
path: bundle.controlPath ?? bundle.pocPath,
|
|
706
|
-
exitCode: bundle.controlRun.exitCode,
|
|
707
|
-
ranAt: bundle.controlRun.ranAt,
|
|
708
|
-
output: bundle.controlRun.output,
|
|
709
|
-
sandbox: bundle.controlRun.sandbox,
|
|
710
|
-
completed: true,
|
|
711
|
-
outputComplete: true,
|
|
712
|
-
mode: "control",
|
|
713
|
-
target: bundle.controlRun.target,
|
|
714
|
-
},
|
|
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
|
+
},
|
|
715
620
|
disconfirmation: verdict.disconfirmation_attempt,
|
|
716
621
|
confirmerVerdict: recorded,
|
|
717
622
|
pendingConfirmation: undefined,
|
|
@@ -723,6 +628,17 @@ export function applyConfirmationResult(
|
|
|
723
628
|
validateCase(next);
|
|
724
629
|
insertEvidenceItem(db, reproductionItem);
|
|
725
630
|
upsertCase(db, next);
|
|
631
|
+
appendCaseEvent(db, {
|
|
632
|
+
caseId: id,
|
|
633
|
+
actor: "main_agent",
|
|
634
|
+
eventType: "case_confirmed",
|
|
635
|
+
payload: {
|
|
636
|
+
verdict: "CONFIRMED",
|
|
637
|
+
proof_strength: recorded.proofStrength ?? null,
|
|
638
|
+
model: verdict.model ?? null,
|
|
639
|
+
reproduction_evidence_id: reproductionItem.id,
|
|
640
|
+
},
|
|
641
|
+
});
|
|
726
642
|
next.evidenceItems = [...(next.evidenceItems ?? []), reproductionItem];
|
|
727
643
|
return { record: next, changed: true };
|
|
728
644
|
});
|